CN109800875A - Chemical industry fault detection method based on particle group optimizing and noise reduction sparse coding machine - Google Patents

Abstract

The invention discloses a kind of chemical industry fault detection method based on particle group optimizing and noise reduction sparse coding machine, this method is mainly sparse from the unsupervised feature learning of code machine progress using the noise reduction of multiple stacks to the training data after standardization and albefaction, the training of Softmax sorter model is carried out by way of having supervision again, finally by the weight parameter for having supervision fine tuning whole network, particle swarm optimization algorithm automated tuning is introduced simultaneously for crucial adjustable hyper parameter, trained chemical process failure detection intelligent model is obtained for the fault detection of process real time data.The present invention uses the layer-by-layer training method of the greediness of deep neural network adaptively intelligence learning chemical process initial data institute tacit knowledge, so as to more accurately extract the information of procedure fault, the method is more intelligent compared with for conventional method, it can be improved the performance of fault detection, and since algorithms of automatic optimization is added, many times are saved than artificial parameter tuning.

Description

Chemical fault detection method based on particle swarm optimization and noise reduction sparse coding machine

technical field

The invention relates to the field of chemical process fault detection and diagnosis, in particular to a chemical fault detection method based on particle swarm optimization and noise reduction sparse coding machine.

Background technique

As one of the most powerful tools for the management of abnormal working conditions in the chemical process, chemical process fault detection provides a certain guarantee for process safety early warning and saves a lot of economic losses. According to estimates by the US National Security Administration, abnormal operating conditions have caused economic losses of at least about 20 billion US dollars per year to the US petroleum and chemical industry. In the United Kingdom, the annual loss caused by abnormal operating conditions is as high as 27 billion US dollars. The chemical process fault detection method is very important in the actual chemical production.

Chemical process data has the characteristics of non-linear, high-dimensional, non-Gaussian distribution, etc., so it will be more complicated to extract fault information in the fault detection process. Many traditional chemical process fault detection methods based on process data have also been developed. The data method does not need to acquire a large amount of expert knowledge in advance, but only needs to acquire the data collected in the chemical process, and by establishing an appropriate fault detection model, the current system status can be predicted. It is widely used in industrial applications. Although the popular traditional methods such as PCA, ICA, KPCA, KICA, and MICA can effectively detect some faults, the detection rate of some disturbance faults is extremely low, indicating that the traditional methods still cannot fully accurately extract It is necessary to develop some new methods to improve the detection rate of these faults. The nonlinear, high noise, non-Gaussian distribution and other characteristics of complex chemical processes make traditional chemical process fault detection methods do not show excellent diagnostic performance. Therefore, it is necessary to develop fault monitoring methods suitable for complex non-linear chemical processes. . The layer-by-layer greedy learning mode of the deep neural network can more accurately learn the features hidden in the raw data of the chemical process, so that the application of an appropriate classification model can realize the monitoring of the chemical process.

With the introduction of deep learning, many scholars in the field of computer began to study deep neural networks. The biggest difference between deep neural networks and shallow neural networks is that they use unsupervised learning to train deep networks layer by layer, which has excellent characteristics. Learning ability, the learned features have a more essential description of the data, which is conducive to classification. The research in the direction of auto-encoder is a popular research direction in the field of deep learning in recent years. Because of its excellent performance in feature learning, it is widely used in handwriting recognition, image classification, audio feature extraction, etc., and gradually in other fields. It is widely used. For example, in mechanical fault diagnosis, some scholars have begun to use automatic encoders or improved automatic encoders for feature learning. In 2006, the famous deep learning scholar Hiton et al. proposed the deep neural network algorithm of Auto Encoder (AE), and applied it to image recognition to obtain a good classification effect. Scholars have proposed improved algorithms. For example, Bengio et al. proposed a sparse auto encoder (Sparse Auto Encoder, SAE), adding sparsity constraints, so that the learned features can more accurately restore the original data; Vincent et al. In 2008, a Denoising Auto Encoder (DAE) was proposed, which made the learned features more robust by adding noise reduction strategies in unsupervised feature learning. Stacked Denoising Sparse Autoencoder (SDSA) is a popular improved autoencoder in the past three years. Stacked means that multiple denoising sparse autoencoders are stacked for layer-by-layer training and Layer learning means. The algorithm is based on the denoising sparse auto-encoder, and builds a deep neural network by stacking multiple similar denoising sparse auto-encoders to train the weight parameters and bias terms of each layer layer by layer in an unsupervised manner. Network, the learned features can deal with high noise data and make the learned features sparse, and its training method is carried out in an unsupervised manner, so the algorithm is more automatic than traditional manual feature extraction in feature learning. At the same time, because its algorithm is based on the theory of minimum information entropy loss in the loss function, the learned features have strong accuracy and can more accurately predict the activation value of the model.

Since the method disclosed in the present invention involves many manually adjustable hyperparameters, the selection of hyperparameters by manual testing is random and time-consuming. Therefore, single-objective discrete optimization should be performed on multiple parameters to realize automatic adjustment of the method. optimization features, so that the method is more intelligent and achieves better performance. Commonly used parameter optimization algorithms include genetic algorithm, artificial fish swarm algorithm, particle swarm algorithm, etc. Particle swarm optimization (PSO) algorithm is a bionic algorithm proposed by Kennedy et al. in 1995 to simulate the foraging behavior of birds. It has been widely used in the engineering field and has become one of the fastest-growing parameter intelligent optimization algorithms.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a chemical fault detection method based on particle swarm optimization and noise reduction sparse coding machine, aiming at the deficiencies of the prior art. The method applies the stack-type noise reduction sparse automatic encoder (SDSA) algorithm in deep learning to the feature learning of chemical process, and then conducts Softmax classifier model training in a supervised manner, and finally fine-tunes the entire network through BP algorithm. weight parameters, and the particle swarm optimization algorithm is introduced for automatic tuning of key adjustable hyperparameters. This method can use the greedy layer-by-layer training method of the deep neural network to adaptively and intelligently learn the implicit knowledge of the original data, so as to extract the fault information. Manual tuning saves a lot of time.

The purpose of the present invention can be realized by following technical scheme:

A chemical fault detection method based on particle swarm optimization and noise reduction sparse coding machine, the method comprises the following steps:

Step 1. Data collection:

The historical time series data collected by the simulation system or DCS system is used as the training sample set X _train , and the real-time chemical process data from the DCS system is used as the test sample set X _test , wherein the collected training sample set X _train includes The time series data under various fault conditions is used to establish the intelligent fault detection model of this method. The test sample set X _test is the real-time operating condition data monitored online, and also includes the time series data under normal conditions and various faults. To verify the diagnostic accuracy of the method or apply the model established by the method in the actual industry to realize fault detection;

Step 2: Data preprocessing:

First, calculate the mean value X _mean and standard deviation X _std of each monitoring variable of the data in the training sample set X _train under _normal working conditions, and _then use the mean value X _mean and the standard The difference X _std is subjected to standardization preprocessing, and the preprocessed training sample set X _trainstd and test sample set X _teststd are subjected to whitening preprocessing to obtain the whitened training sample set X _trainwhite and test sample set X _testwhite , so far the training samples and Preprocessing of the test sample set;

Step 3. Offline training:

The purpose of offline training is to use pre-processed training samples to establish an intelligent fault detection model for chemical processes. The process can be divided into unsupervised pre-training stack noise reduction and sparse auto-encoder, supervised pre-training Softmax classifier, and BP algorithm global fine-tuning. Network parameters and particle swarm optimization can be divided into four major parts; first, unsupervised pre-training stack type noise reduction sparse automatic encoder is carried out, and N noise reduction sparse encoders are used for the preprocessed training set to convert the sample set layer by layer. Coding into feature space, using the method of layer-by-layer training, in the training process of each layer, the loss function of each layer is optimized to minimize, so as to obtain the model parameters of each layer, and finally obtain the final training sample learned by its N hidden layers. Feature h _N ; secondly, supervised pre-training Softmax classifier is performed, h _N is used as the input of the Softmax classifier model, and all the learned training sample features are attached with corresponding working condition labels y ⁱ =1 represents that the sample is normal, y ⁱ =2 represents that the sample is faulty, and the pre-trained Softmax model parameters are obtained by optimizing the cost function of Softmax; then the BP algorithm is used to globally fine-tune the network parameters. All model parameters in the training process are fine-tuned in a supervised manner. The unsupervised pre-trained SDSA model parameters and the supervised pre-trained Softmax model parameters are used as initial values, and the pre-processed training samples are used as the input layer. First, pass these SDSA parameters. Layer-by-layer feature learning is used to obtain the feature matrix of the final hidden layer. The feature matrix is used to calculate its loss function value through the Softmax classifier, and then the feedback propagation (BP) algorithm is used to optimize the global parameters so that the loss function can be minimized and converged. Obtain the fault detection model that has been initially trained; finally, adjust the hyperparameters by particle swarm optimization. Since the artificially adjustable hyperparameters in the entire intelligent fault detection model are not optimal, if these hyperparameters are set improperly, the performance of SDSA may not be the same. Therefore, the particle swarm (PSO) optimization algorithm will be called for hyperparameter tuning, and finally the optimal hyperparameter values and all corresponding model parameters will be determined, so as to obtain the final trained fault detection model, which can be used for Online monitoring of process status;

Step 4. Online monitoring:

For the chemical process of real-time continuous production, the fault detection model trained in the above step 3 can effectively predict whether the current process is in a fault state; using the preprocessed test sample X _testwhite as the input layer, according to the trained fault detection model The model, for the determined neural network structure, adopts the layer-by-layer feature learning method to obtain the final learned test features, and then calculates the membership probability value of its Softmax prediction function through the learned final features. When the obtained membership category is 1, then It indicates that the working condition is normal. When the affiliation category is 2, it means that there is an abnormal working condition, and the connected alarm device will issue an early warning to indicate the fault detection. It can realize fault monitoring of the current working condition.

Further, in step 2, the standardized preprocessing of the training sample set X _train and the test sample set X _test is realized by the following steps:

(1) The training sample set X _train is an n×m matrix, where n is the number of samples and m is the number of observed variables. The standardized training sample set X _trainstd and the test sample set X _teststd are solved by the following formula:

Among them, X _ij represents the value of the jth variable of the ith sample in the training sample set X _train and the test sample set X _test , and X _mean,j is the value of the jth variable of the data in the training sample X _train under normal conditions. Mean, X _std,j The standard deviation of the jth variable of the data under normal conditions in the training sample X _train .

Further, in step 2, the standardization and whitening preprocessing of the training sample set X _train and the test sample set X _test is realized by the following steps:

(1) Perform whitening preprocessing on the standardized training sample set X _trainstd and the test sample set X _teststd , first perform eigenvalue decomposition on the covariance matrix of the training sample by the following formula, and obtain the eigenvector of the covariance matrix. The orthogonal matrix and the diagonal matrix of its eigenvalues, resulting in the whitening matrix W _white :

Cov=VDV ^T (2)

W _white = VD ^-1/2 V ^T (3)

Among them, Cov is the covariance matrix of the standardized training sample set X _trainstd , V is the orthogonal matrix of the eigenvectors of the covariance matrix, and D is the diagonal matrix of the eigenvalues of the covariance matrix;

(2), then the training sample X _trainwhite and the test sample X _testwhite after whitening are calculated by W _white :

Among them, W _white is the whitening matrix of whitening preprocessing, X _trainwhite is the training sample set after whitening processing, and X _testwhite is the testing sample set after whitening processing.

Further, in step 3, the unsupervised pre-training stack-type noise reduction and sparse automatic encoder in the offline training is specifically performed through the following steps:

(1) For the preprocessed training sample X _trainwhite is a matrix of n×m, n is the number of samples, m is the number of observation variables, assuming that the number of hidden layers of the stack-type noise reduction sparse automatic encoder SDSA is N, set its deep neural network structure, that is, mainly determine the number of nodes HL ₁ -HL _N of each hidden layer network, and each layer of its deep neural network is mainly defined as: input layer (training samples after preprocessing), N A hidden layer (that is, learning N-layer features) and a classification layer (that is, the predicted probability of output), so its global network can be regarded as a neural network with N+2 layers; first, initialize the weight matrix parameters of the first hidden layer, and set the X _{trainwhite is} used as the input layer of SDSA to train the noise reduction sparse auto-encoder (DSA ₁ ) of the first hidden layer. The training sample X _{trainwhite is} added with some Gaussian noise that obeys the normal distribution by the following formula, and the training set becomes a mixed Noisy dataset Xc:

Xc=X+le*G (5)

Among them, X is the training set matrix when no noise is added, and X is the preprocessed training set X _trainwhite in the first hidden layer; le is the noise level, which can be set manually (its value is 0-1, generally 0.1 is can); G is the generated Gaussian noise with the same dimension as the training set matrix when no noise is added, and Xc is the data set containing artificial noise;

(2) After adding the interference noise, use Xc as the input to perform coding learning and decoding reconstruction. The coding stage is essentially the stage of feature learning, and the decoding is the reconstruction of the feature data. The coding formula and decoding formula are:

Among them, h is the learned feature, Y is the information reconstructed by the feature h, the closer the value is to the original data X without noise, the better the model parameters are trained; W is the encoding weight parameter matrix, b ₁ , b ₂ is the bias vector, W ^T is the decoding weight parameter matrix, the feature h ₁ of the first hidden layer and the model parameter W ₁ of the first hidden layer can be learned through the above formula, ₁₁ ,b ₂₁ ;

(3) However, it can be noticed that these model parameters are not optimal, so the learned first hidden layer features cannot deeply express the information contained in the original data; just define a reasonable loss function, and then continuously optimize the When the loss function reaches the minimum value, the optimal ability of the model can be exerted. The loss function is as follows:

Among them, L _total (W,W ^T ,b ₁ ,b ₂ ) represents the loss function value of the noise reduction sparse auto-encoder of the current hidden layer, n is the number of samples, and Y _i is the reconstructed information vector of the ith sample value, X _i is the original data vector value of the ith sample without adding noise, is the value of the i-th row and the j-th column of the weight parameter of the l-th layer, λ _r is the regularization weight decay hyperparameter used to adjust the weight of the equation; sl represents the number of nodes in the l-th layer of a single noise reduction auto-encoder , s(l+1) represents the number of nodes in the l+1th layer, β is the weight hyperparameter that controls the sparsity penalty term, and needs to be manually adjusted to a more appropriate value; s ₂ is the number of nodes of the current hidden layer learning feature , that is, the number of features learned by a single sample; is the relative entropy, which represents the difference between the current average activation degree and the sparsity constraint; ρ is the sparsity parameter, which represents the sparsity constraint, which is generally a small value set artificially, for example, ρ takes 0.1; is the average activation of the jth variable in the feature matrix h, Represents the value of the jth variable of the ith sample in the feature matrix h of the current hidden layer;

During offline training, the matlab toolbox minFunc can be called to minimize the loss function of equation (8), and the optimized weight parameter W ₁ after the first hidden layer pre-training can be obtained, and bias terms b ₁₁ , b ₂₁ , and then use the trained model parameters for feature learning of the training samples of the first hidden layer;

(4) The first three steps above are equivalent to completing the offline training of the model of the first hidden layer. Since the stack-type noise reduction sparse automatic encoder of this method contains N hidden layers, it is necessary to repeat the calling step (1- 3) The model parameter training of the multi-hidden layer network is performed until the SDSA pre-training of N hidden layers is completed.

Further, the step (4) specifically includes:

First, after the training of the noise reduction sparse autoencoder (DSA ₁ ) of the first hidden layer is completed, its training sample X _trainwhite is encoded as h ₁ by the following formula:

h ₁ =f(W ₁ X+b ₁₁ ) (11)

Among them, W ₁ , b ₁₁ are the model parameters of the trained DSA ₁ , X is the training set matrix when no noise is added, X is the preprocessed training set X _trainwhite in the first hidden layer, and the second-N hidden layer is represented as the feature matrix learned by the previous layer; then use the learned first hidden layer feature h ₁ as the input of the second hidden layer noise reduction sparse auto-encoder (DSA ₂ ), and repeat the calling step (1 -3) To train DSA ₂ , get the model weight parameters and bias term W ₂ of the second hidden layer, b ₁₂ , b ₂₂ , encode the feature of the second hidden layer into h ₂ through the above formula (11), repeat this process until the DSA _N training is completed, and obtain the final feature h _N learned by its N hidden layers, and set The features learned by pre-training are used for the subsequent classification layer, and the model parameters of the Softmax classifier can be trained.

Further, in step 3, the supervised pre-training Softmax classifier in the offline training is specifically performed through the following steps:

(1) The Nth hidden layer feature hN learned after SDSA training with _N hidden layers is used as the input of the Softmax classifier model, and then the corresponding working condition label is attached y ⁱ = 1 represents that the ith sample is normal, and y ⁱ = 2 represents that the ith sample is faulty; first initialize its model parameter matrix θ, and predict the probability value of its sample belonging to each category according to the following Softmax classifier prediction function:

in, is the probability value of the i-th sample belonging to each category, and θ is the model parameter matrix of the Softmax classifier, which is represented by Vector composition; k is the number of categories defined by the classifier, where k=2; is the feature vector of the Nth hidden layer of the ith sample;

(2) Since the various membership probability values predicted by the prediction function are not accurate, a classification loss function needs to be constructed to obtain the optimal model parameters. The Softmax classifier used considers the regularization term, which can effectively avoid The classification model is trained and overfitted, and the loss function is defined as follows:

Among them, the meaning of the 1{.} part is that if the index function in the _parentheses is true, it will return 1, otherwise it will _return 0; Loss function, you can call the matlab toolbox minFunc to optimize the minimum loss function, so as to obtain the optimal parameter matrix θ of the Softmax classifier model.

Further, in step 3, the global fine-tuning of network parameters of the BP algorithm in the offline training is specifically performed through the following steps:

Take all pre-trained model parameters {(W ₁ ,b ₁₁ ),(W ₂ ,b ₁₂ ),…,(W _N ,b _1N ),θ} as initial values, and take the pre-trained training sample X _trainwhite For the input layer, firstly through the model initial parameters of SDSA, Equation (11) is used to perform layer-by-layer feature learning, so as to obtain the final feature matrix of the hidden layer. The propagation algorithm optimizes the global parameters, and updates all model parameters once in each iteration, so that the loss value of equation (13) can be converged to the minimum, and the fine-tuning process is completed.

Further, in step 3, the particle swarm optimization adjustable hyperparameters in the offline training are specifically performed through the following steps:

(1) The parameters of this method can be divided into two categories: model parameters and adjustable parameters. The method optimizes three key adjustable parameters, namely: regularization weight decay hyperparameter λ _r , control sparsity The weight hyperparameter β of the penalty term, the weight decay term coefficient λ _sm of Softmax; these three parameters constitute a single particle (λ _r , β, λ _sm ) optimized by PSO, so the dimension of each particle is 3 dimensions; define N _p Each particle is optimized at the same time, K iterations are set, the initial position and velocity of each particle are initialized, and the fitness function value is the overall accuracy rate of the training sample. The calculation method uses these three key adjustable hyperparameters as independent variables, By training the above fine-tuned fault detection model, the overall accuracy Accuracy of the training samples can be obtained, and its formula is defined as the following formula:

Among them, pi represents the category of the ⁱ -th training sample predicted by the fault detection model, and ^yi represents the category of the i-th training sample in the artificially given label. If the two are equal, it will return 1. If they are not equal, it means that the current sample is wrongly diagnosed by the model, and the return value is 0; the optimal fitness and position of each particle and the global optimal fitness and position of the particle swarm are calculated through the above formula (14);

(2) Since the global optimal fitness and optimal position after the above one iteration are not the most accurate, the velocity and position of each particle need to be updated, and the position of the particle under the t-th iteration can be expressed as The flight speed of each particle in the t-th iteration can be expressed as The PSO algorithm updates the velocity and position of each particle by the following formulas:

Among them, W _p is the inertia coefficient, C ₁ is the acceleration coefficient of the particle tracking its own historical optimal value, indicating the particle's cognition of itself; C ₂ is the acceleration coefficient of the particle tracking group optimal value, indicating the particle's cognition of the group knowledge , namely social knowledge, usually set C ₁ =C ₂ =2; t is the current number of iterations; ξ and η are uniformly distributed random numbers in the interval [0,1]; is the position of the optimal fitness experienced by particle i up to the t-th iteration; gb ^t is the best position experienced by the particle swarm up to the t-th iteration;

(3) Retrain the unsupervised pre-training SDSA of the above fault detection model, the supervised pre-training Softmax classifier, and the BP algorithm to globally fine-tune the network parameters to obtain a new global optimal fitness value and position, and use the formula (15- 16) Continuously update the position and speed of the particle, and terminate the optimization when the number of iterations exceeds its defined K iterations, so as to obtain the particle position under the global optimal fitness, that is, under the optimal adjustable hyperparameter, its The detection accuracy rate of the training samples is the highest, thus completing the automatic tuning of hyperparameters; then set the optimized tunable hyperparameters as SDSA's optimized tunable hyperparameters, and retrain the determined network structure and the optimal tunable hyperparameters. All model parameters of SDSA under the optimal hyperparameters are obtained to obtain a trained fault detection model, whose model parameters can be used for feature learning and classification of test samples.

Further, in step 4, the feature learning and membership probability prediction of the online monitoring are performed through the following steps:

For the chemical process of real-time continuous production, the fault detection model trained in the above step 3 can effectively predict whether the current process is in a fault state. The preprocessed test sample X _{testwhite is} used as the input layer, and the trained fault detection Model, for the determined neural network structure, the layer-by-layer feature learning method is adopted, and the weight parameters {(W ₁ ,b ₁₁ ),(W ₂ ,b ₁₂ ),…,( W _N , b _1N )}, use formula (11) to perform feedforward learning to obtain the features of the current hidden layer of the test sample, and then use the features learned by the current hidden layer as input, and use the same DSA model parameters in conjunction with the next hidden layer. Formula to learn features, and so on after learning all hidden layers to get the final learned features, and then use formula (12) to predict each membership category. In fault detection, there are only two categories, namely fault state and normal state. Therefore, for a single sample, the predicted probability vector has only two values, and the prediction category with the largest probability value is selected as the membership category predicted by the entire fault detection model. When the calculated membership category is 1, it means that the working condition is normal. If the category is 2, it means that there is an abnormal working condition, and the connected alarm device will issue an early warning to indicate the fault detection, and at the same time notify the technician or engineer to check the system safety and eliminate the fault in time, so as to realize the fault monitoring of the current working condition. .

Compared with the prior art, the present invention has the following advantages and beneficial effects:

1. The present invention uses the stack noise reduction and sparse autoencoder algorithm in the deep learning algorithm combined with the particle swarm parameter optimization algorithm to develop a new fault monitoring method suitable for complex non-linear chemical processes, because there is no need to label data during feature learning. Compared with the traditional shallow neural network method and artificial neural network, it can realize adaptive and intelligent learning of the characteristics of the original data, which saves a lot of energy and time than manual extraction of features and knowledge. The learned features can more deeply distinguish normal and abnormal data, and make the clustering performance of each category of samples more obvious, so the algorithm is more intelligent.

2. The present invention avoids the randomness and time-consuming of manual selection of parameters in the selection of artificially adjustable hyperparameters, and adopts the technical scheme of automatic parameter tuning. Tuning hyperparameters is automatically tuned, which saves a lot of time and makes the method more automated.

3. Compared with the traditional chemical process fault detection technology (such as PCA, ICA, KPCA, MICA, etc.), the present invention adopts the layer-by-layer training method of deep neural network, which can adaptively learn nonlinear process data during feature learning. At the same time, the noise reduction strategy is added to enhance the robustness of the algorithm to high-noise data, so that the fault detection model established by it is also more accurate. The method of the invention has the advantages of significantly improved fault detection rate and false alarm. It has the advantages of low rate, fast fault detection, and can be applied to big data modeling, so it has a good monitoring effect and can improve the performance of fault detection.

Description of drawings

FIG. 1 is a flowchart of a chemical fault detection method based on particle swarm optimization and noise reduction sparse encoder according to an embodiment of the present invention.

2 is a process flow diagram of the Tennessee-Eastman (TE) chemical process adopted in the embodiment of the present invention.

FIG. 3 is a graph of the average fault detection rate and false alarm rate of ten repeated tests according to an embodiment of the present invention.

FIG. 4 is a comparison diagram of the fault detection rate of each method in which the detection rate is greatly improved according to the embodiment of the present invention.

Figure 5(a), Figure 5(b), Figure 5(c), Figure 5(d), Figure 5(e), Figure 5(f), Figure 5(g), Figure 5(h), Figure 5 (i) Fault monitoring result diagrams of fault 1, fault 2, fault 4, fault 6, fault 7, fault 8, fault 12, fault 13, and fault 17, respectively.

Figure 6(a) is a comparison diagram of the first three PCA of the preprocessed test samples, and Figure 6(b) is the first three PCA diagrams of the test features learned by the PSO-SDSA method.

Detailed ways

The present invention will be described in further detail below with reference to the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.

Example:

This embodiment provides a chemical fault detection method based on particle swarm optimization and noise reduction sparse coding machine. The flowchart of the method is shown in Figure 1. The method proposed in this embodiment is applied to Tennessee-Eastman (TEE) ) benchmark chemical process to further illustrate the method of the present embodiment, the TE process is the computer simulation of the actual chemical process published by Downs and Vogel in the SCI journal of "Computers & Chemical Engineering" in 1993, and this process has now been mainly developed into evaluation process monitoring. The performance of the method, the process flow diagram of the process is shown in Figure 2. The TE process mainly includes five operating units, namely: reactor, condenser, vapor-liquid separator, circulating compressor, and stripper. In the simulated data, a total of 41 observed variables are monitored, including 22 continuous process variables and 19 component variables. The TE process also includes 21 preset faults. In this embodiment, the first 20 faults are used for monitoring, and the 20 preset faults are shown in Table 1 below.

Table 1. 20 preset failures of the TE process

The chemical process fault detection method comprises the steps:

Step 1. Data collection:

Collect the data under normal working conditions and 20 faults in the TE process, and divide it into a training sample set X _train and a test sample set X _test . The training sample set contains 13,480 samples under normal working conditions and 480 samples under each fault. The test sample set contains 960 normal samples and 960 fault samples each, but the fault samples are all in the fault state from the 161st sample. The process monitors 41 variables, so the training sample set forms a 23080×41 matrix and the test sample set forms a 20160×41 matrix.

Step 2: Data preprocessing:

First, the mean value X _mean and the standard deviation X _std of each variable are obtained from the 13480 normal samples in the training data. Then use the mean X _mean and the standard deviation X _std to standardize both the training sample set and the test sample set using formula (1), and then use formula (2-3) to obtain the whitening matrix W _white of the training sample set, and use formula ( 4-5) Obtain the whitened training sample set X _trainwhite and test sample set X _testwhite .

Step 3. Offline training:

First, define the number of hidden layers as 2 and the network structure as 41-130-20-2, initialize the weight parameter matrix and bias term of SDSA, and use the whitened training sample set X _trainwhite as the input layer of SDSA, through 2 For the layer-by-layer greedy training of the hidden layer, use the matlab toolbox minFunc to optimize the loss function of SDSA to obtain the optimized model parameters, and use formula (11) feedforward learning to obtain the feature h ₂ of the training sample. Then use the training sample feature h ₂ as the model input as the input of the Softmax classifier model, and attach all the learned training sample features with the corresponding working condition labels y ⁱ =1 represents that the sample is normal, y ⁱ =2 represents that the sample is faulty, and the loss function of Softmax is optimized through the matlab toolbox minFunc to obtain the pre-trained Softmax model parameters; then the BP algorithm is used to globally fine-tune the network. parameters, supervised fine-tuning of all model parameters in the whole training process, using unsupervised pre-trained SDSA model parameters and supervised pre-trained Softmax model parameters as initial values, pre-processed training samples as input layer, using The feedback propagation (BP) algorithm optimizes the global parameters so that the loss function is minimized and converged, so as to obtain a preliminarily trained fault detection model. Finally, particle swarm optimization is carried out for three key adjustable hyperparameters, and finally the values of the optimized hyperparameters and all the corresponding model parameters are determined, so as to obtain the final trained fault detection model, which can be used for online monitoring of process status.

Step 4. Online monitoring:

For the chemical process of real-time continuous production, using the fault detection model trained in the third step above can effectively predict whether the current process is in a fault state. Taking the preprocessed test sample X _testwhite as the input layer, according to the trained fault detection model, for the determined neural network structure, the layer-by-layer feature learning method is adopted, and the weight parameters of the trained noise reduction sparse auto-encoder are passed { (W ₁ ,b ₁₁ ),(W ₂ ,b ₁₂ ),…,(W _N ,b _1N )}, using formula (11) for feedforward learning to obtain the features of the current hidden layer of the test sample, and then the current hidden layer The learned features are used as input, and the DSA model parameters of the next hidden layer are combined with the same formula to learn the features, and so on after learning all hidden layers to obtain the final learned features, and then formula (12) is used for each membership category. Prediction, there are only two categories in fault detection, namely fault state and normal state, so for a single sample, the predicted probability vector has only two values, and the predicted category with the largest probability value is selected as the membership category predicted by the entire fault detection model. When the calculated affiliation category is 1, it means that the working condition is normal; when the affiliation category is 2, it means that there is an abnormal working condition, and finally the detection rate of each fault is counted;

The test sample set is regarded as real-time working condition data, and fault diagnosis can be made on the real-time data collected in the actual chemical process through the above steps.

First, ten repeated experiments were carried out under the selected network structure of 41-130-20-2 and the tuned hyperparameters. The optimal hyperparameters obtained by this method in the TE process are listed as follows: Regularization weight decay hyperparameter λ _r =0.002, the weight hyperparameter β=2.628255 that controls the sparsity penalty term, the coefficient of the weight decay term of Softmax λ _sm =0.0001, and its global optimal fitness value=0.9998. The average test failure detection rate and test average false alarm rate obtained in each test are shown in Figure 3. It can be seen from the figure that the average value of the test sample failure detection rate (FDR) of the ten average tests is 82.42 %, the standard deviation is 0.857%, the average value of false alarm rate (FAR) is 0.64%, and the standard deviation is 1.194%, indicating that the overall FDR of this method is high and the FAR is low. According to the method of fault detection performance evaluation, this method The diagnostic performance is better. In addition, the FDR and FAR of the ten trials are not much different from each other, indicating that this method integrates PSO optimization and has the effect of improving the stability of the model. If the model hyperparameters are not optimized, it is easy to make the parameters inappropriate and lead to training. Difficulty and falling into local optimum, so choosing PSO optimization can not only avoid the occurrence of such problems, but also enable the model to achieve better diagnostic performance under optimal parameters. At the same time, PSO optimization is a relatively mature unconstrained discrete optimization method. It has the advantages of fast convergence speed and easy programming. To sum up, it is necessary and easy to implement PSO parameter optimization.

The first result (the ninth test) with the best performance (the highest fault detection rate and the lowest false alarm rate) of this method is selected to show the detection rate of each fault in the test sample and compare it with other methods as shown in Table 2. From Table 2, it can be seen that the proposed PSO-SDSA method has a higher diagnostic accuracy. It can be seen that compared with the PCA method, the MICA method and the KPCA method, the proposed method (PSO-SDSA) with the average fault detection rate is the best, and the average FDR of all faults reaches 83.48%, and the FAR is only 0.21%. In the methods developed in the past, faults 3, 9, and 15 are the faults that are difficult to detect in the TE process. Because the difference between the fault itself and the normal sample data is not large, many methods appear when detecting these faults. Very low diagnosis rate (mostly lower than 10%), because the modeling data of deep neural network and traditional methods are different from a large number of normal samples and fault samples. Training and adding noise modeling to strictly limit the loss of information, the learned fault features can deeply mine the differences between these small disturbance faults and normal, so that they can be better differentiated from normal features. Therefore, the fault detection rates of faults 3, 9, and 15 have been greatly improved in the proposed method, among which fault 3 reaches 50.5%, fault 5 reaches 44.755%, fault 15 reaches 44.875%, and other methods are basically maintained below 10%, therefore, to a certain extent, the proposed PSO-SDSA method can effectively improve the performance of some difficult-to-detect fault points. Although the PSO-SDSA method is slightly lower than other methods in some faults (fault 1, 14), in general, this method is modeled under the consideration of global errors. Therefore, the model may be biased to reduce the error of the faults that are more difficult to detect, while slightly ignoring some samples of faults that are easier to detect, so the FDR of some faults that are easier to detect may be slightly lower than that of the traditional method. The resulting fault detection rate (96.5% for fault 1 and 90.375% for fault 14) is perfectly acceptable. In general, compared with the other methods above, the performance of this method has been greatly improved, and it is the most advantageous FDD method in the current research. In addition, Fig. 4 shows a comparison chart of the fault detection rate (faults 3, 5, 9, 10, 15, 19, 20) of this method compared with other methods, which can be significantly improved by FDR, which can more intuitively show this method. The superiority of the method performance improvement, it can be seen that each point of the PSO-SDSA curve is above the curve of other methods, indicating that the detection rate of the seven faults given has been greatly improved.

Table 2. Failure detection rates of various methods in the TE process

The early warning speed analysis of online monitoring is carried out on the test results, and 9 faults with a high detection rate are selected for the detection speed analysis (fault 1, 2, 4, 6, 7, 8, 12, 13, 17). The fault samples and predicted probability are respectively drawn in Figure 5 in the above order. The 0.5 warning line in the figure is the classification control limit. When the predicted probability value of the sample exceeds the control limit, the model is regarded as a fault state. As can be seen from Figure 5, for the selected 9 faults, after the fault is detected, most of the samples thereafter can be detected, thus resulting in a high fault detection rate (their FDRs are all higher than 90 %), faults 1, 4, 6, 7, 8, and 12 are all detected in the 161st sample, with a delay of 0 points, indicating that the detection speed of these faults is extremely high. The detection point of faults 2 and 17 is at the 162nd sample. Although it is delayed by 1 point, the detection speed is still very high, and its early warning effect is completely acceptable in practical application. The detection speed of fault 13 is relatively lower, and the fault is only detected at the 167th sample, and the alarm starts after a delay of 6 points. In addition, it can be seen from the fault online monitoring results given that after faults 2, 4, 6, 7, and 12 are detected, there are very few points below the control limit, resulting in a fault detection rate of more than 99%. , indicating that this method has a strong ability to capture these faults. Faults 1, 8, 13, and 17 occasionally fall below the control limit step by step after the fault is detected, resulting in a lower detection rate of these faults, but the detection rate is above 94%, so its detection Performance is also better.

In order to further study the reason why this method has better fault detection rate and false alarm rate, principal component analysis is carried out on the features learned by this method. We selected the first three PCs in the normal state and the data of faults 1, 2, 6, and 7 for comparison. Figure 6(a) is the comparison chart of the first three PCs of the preprocessed test samples, and Figure 6(b) Plots of the first three principal component analyses of the test features learned for the PSO-SDSA method. Comparing Fig. 6(a) and Fig. 6(b), it can be seen that in the pre-processed test data, the data of faults 1, 2, and 7 overlap with the data of normal samples more seriously, showing serious linear inseparability characteristics, which is Due to the complex and nonlinear characteristics of the TE process, it cannot be directly fed to the classification model to achieve correct classification. However, the principal component analysis of the features learned by the PSO-SDSA method shows that the faults 1, 2, 6, and 7 all have relatively high performance. Good discrimination, and each fault itself has good sample aggregation. At the same time, it is noted that the clustering performance between its normal samples is also good, and there is no great dispersion, so these faults and normal samples can be relatively complete It leads to the effect of high fault diagnosis rate and extremely low false alarm rate, which also shows that this method can be used to learn the implicit knowledge in chemical process data. Under the feature learning of the stack noise reduction sparse autoencoder, the fault information and normal information can be deeply distinguished, so that the learned features have excellent clustering performance, thereby improving the performance of fault detection.

The above is only a preferred embodiment of the patent of the present invention, but the protection scope of the patent of the present invention is not limited to this. The technical solution and the invention patent concept of the invention are equivalently replaced or changed, all belong to the protection scope of the invention patent.

Claims (9)

1. a chemical fault detection method based on particle swarm optimization and noise reduction sparse coding machine, is characterized in that, described method comprises the following steps: Step 1. Data collection: The historical time series data collected by the simulation system or DCS system is used as the training sample set X _train , and the real-time chemical process data from the DCS system is used as the test sample set X _test , wherein the collected training sample set X _train includes The time series data under various fault conditions is used to establish the intelligent fault detection model of this method. The test sample set X _test is the real-time operating condition data monitored online, and also includes the time series data under normal conditions and various faults. To verify the diagnostic accuracy of the method or apply the model established by the method in the actual industry to realize fault detection; Step 2: Data preprocessing: First, calculate the mean value X _mean and standard deviation X _std of each monitoring variable of the data in the training sample set X _train under _normal working conditions, and _then use the mean value X _mean and the standard The difference X _std is subjected to standardization preprocessing, and the preprocessed training sample set X _trainstd and test sample set X _teststd are subjected to whitening preprocessing to obtain the whitened training sample set X _trainwhite and test sample set X _testwhite , so far the training samples and Preprocessing of the test sample set; Step 3. Offline training: The pre-processed training samples are used to establish an intelligent fault detection model for chemical processes. The process can be divided into unsupervised pre-training stack noise reduction and sparse automatic encoder, supervised pre-training Softmax classifier, BP algorithm global fine-tuning network parameters, particle swarm The four major parts of the adjustable hyperparameters are optimized. First, the unsupervised pre-training stack-type noise reduction sparse auto-encoder is performed, and N noise reduction sparse encoders are used for the preprocessed training set to encode the sample set into feature space layer by layer. Using the method of layer-by-layer training, in the training process of each layer, the loss function of each layer is optimized to minimize the loss function of each layer, so as to obtain the model parameters of each layer, and finally obtain the final feature h _N of the training samples learned by its N hidden layers; secondly Perform supervised pre-training of the Softmax classifier, use h _N as the input of the Softmax classifier model, and attach all the learned training sample features to the corresponding working condition labels y ⁱ =1 represents that the sample is normal, y ⁱ =2 represents that the sample is faulty, and the pre-trained Softmax model parameters are obtained by optimizing the cost function of Softmax; then the BP algorithm is used to globally fine-tune the network parameters. All model parameters in the training process are fine-tuned in a supervised manner. The unsupervised pre-trained SDSA model parameters and the supervised pre-trained Softmax model parameters are used as initial values, and the pre-processed training samples are used as the input layer. First, pass these SDSA parameters. Layer-by-layer feature learning is used to obtain the feature matrix of the final hidden layer. The feature matrix is used to calculate its loss function value through the Softmax classifier, and then the feedback propagation (BP) algorithm is used to optimize the global parameters so that the loss function can be minimized and converged. Obtain the fault detection model that has been initially trained; finally, adjust the hyperparameters by particle swarm optimization. Since the artificially adjustable hyperparameters in the entire intelligent fault detection model are not optimal, if these hyperparameters are set improperly, the performance of SDSA may not be the same. Therefore, the particle swarm PSO optimization algorithm will be called for hyperparameter tuning, and finally the optimal hyperparameter value and all corresponding model parameters will be determined, so as to obtain the final trained fault detection model for online monitoring of process status. ; Step 4. Online monitoring: Taking the preprocessed test sample X _testwhite as the input layer, according to the trained fault detection model, for the determined neural network structure, the layer-by-layer feature learning method is used to obtain the final learned test features, and then calculate the final learned features. The membership probability value of its Softmax prediction function, when the obtained membership category is 1, it means that the working condition is normal; when the membership category is 2, it means that there is an abnormal working condition, and then the connected alarm device issues an early warning to indicate a fault. It is detected, and at the same time, the technician or engineer is notified to check the safety of the system and eliminate the fault in time, so as to realize the fault monitoring of the current working condition. 2. the chemical industry fault detection method based on particle swarm optimization and noise reduction sparse coding machine according to claim 1, is characterized in that, in step 2, realizes the standardization of training sample set X _train and test sample set X _test by following steps Preprocessing: The training sample set X _train is an n×m matrix, where n is the number of samples and m is the number of observed variables. The standardized training sample set X _trainstd and the test sample set X _teststd are solved by the following formula: Among them, X _ij represents the value of the jth variable of the ith sample in the training sample set X _train and the test sample set X _test , and X _{mean, j} is the value of the jth variable of the data in the training sample X _train under normal conditions. Mean, X _{std, j} The standard deviation of the jth variable of the data under normal conditions in the training sample X _train . 3. the chemical industry fault detection method based on particle swarm optimization and noise reduction sparse coding machine according to claim 2, is characterized in that, in step 2, realizes the whitening of training sample set X _train and test sample set X _test by following steps Preprocessing: (1) Perform whitening preprocessing on the standardized training sample set X _trainstd and the test sample set X _teststd , first perform eigenvalue decomposition on the covariance matrix of the training sample by the following formula, and obtain the eigenvector of the covariance matrix. The orthogonal matrix and the diagonal matrix of its eigenvalues, resulting in the whitening matrix W _white : Cov=VDV ^T (2) W _white = VD ^-1/2 V ^T (3) Among them, Cov is the covariance matrix of the standardized training sample set X _trainstd , V is the orthogonal matrix of the eigenvectors of the covariance matrix, and D is the diagonal matrix of the eigenvalues of the covariance matrix; (2), then the training sample X _trainwhite and the test sample X _testwhite after whitening are calculated by W _white : Among them, W _white is the whitening matrix of whitening preprocessing, X _trainwhite is the training sample set after whitening processing, and X _testwhite is the testing sample set after whitening processing. 4. the chemical industry fault detection method based on particle swarm optimization and noise reduction sparse coding machine according to claim 3, is characterized in that, in step 3, the unsupervised pre-training stack type noise reduction sparse automatic coding in described offline training The machine is carried out through the following steps: (1) For the preprocessed training sample X _trainwhite is a matrix of n × m, n is the number of samples, m is the number of observation variables, assuming the hidden layer of the stack noise reduction sparse automatic encoder (SDSA) The number is N, and its deep neural network structure is set, that is, the number of nodes in each hidden layer network is mainly determined HL ₁ --- HL _N , and each layer of the deep neural network is mainly defined as: input layer, N hidden layers, classification Therefore, its global network can be regarded as a neural network with N+2 layers. First, initialize the weight matrix parameters of the first hidden layer, use X _trainwhite as the input layer of SDSA, and train the noise reduction and sparse automatic of the first hidden layer. The encoder DSA ₁ , adds part of the Gaussian noise that obeys the normal distribution to the training sample X _trainwhite by the following formula, and turns the training set into a data set Xc mixed with noise: Xc=X+le*G (5) Among them, X is the training set matrix when no noise is added, and X is the preprocessed training set X _trainwhite in the first hidden layer; le is the noise level, and its value is 0-1, generally 0.1; G is The generated Gaussian noise with the same dimension as the training set matrix when no noise is added, Xc is the data set containing artificial noise; (2) After adding the interference noise, use Xc as the input to perform coding learning and decoding reconstruction. The coding stage is essentially the stage of feature learning, and the decoding is the reconstruction of the feature data. The coding formula and decoding formula are: Among them, h is the learned feature, Y is the information reconstructed by the feature h, the closer the value is to the original data X without noise, the better the model parameters are trained; W is the encoding weight parameter matrix, b ₁ , b ₂ is the bias vector, W ^T is the decoding weight parameter matrix, the feature h ₁ of the first hidden layer and the model parameter W ₁ of the first hidden layer can be learned through the above formula, b ₁₁ , b ₂₁ ; (3) Define a reasonable loss function, and then continuously optimize the loss function to reach the minimum value, which can exert the optimal ability of the model. The loss function is as follows: Among them, L _total (W, W ^T , b ₁ , b ₂ ) represents the loss function value of the noise reduction sparse auto-encoder of the current hidden layer, n is the number of samples, and Y _i is the reconstructed information vector of the ith sample value, x _i is the original data vector value of the ith sample without adding noise, is the value of the i-th row and the j-th column of the weight parameter of the l-th layer, λ _r is the regularization weight decay hyperparameter used to adjust the weight of the equation; sl represents the number of nodes in the l-th layer of a single noise reduction auto-encoder , s(l+1) represents the number of nodes in the l+1th layer, β is the weight hyperparameter that controls the sparsity penalty term, and needs to be manually adjusted to a more appropriate value; s ₂ is the number of nodes of the current hidden layer learning feature , that is, the number of features learned by a single sample; is the relative entropy, which represents the difference between the current average activation degree and the sparsity constraint; ρ is the sparsity parameter, which represents the sparsity constraint, which is generally a small value set artificially, for example, ρ takes 0.1; is the average activation of the jth variable in the feature matrix h, Represents the value of the jth variable of the ith sample in the feature matrix h of the current hidden layer; During offline training, the matlab toolbox minFunc can be called to minimize the loss function of equation (8), and the optimized weight parameter W ₁ after the first hidden layer pre-training can be obtained, and bias terms b ₁₁ , b ₂₁ , and then use the trained model parameters for feature learning of the training samples of the first hidden layer; (4) Repeat calling steps (1-3) to train the model parameters of the multi-hidden layer network until the SDSA pre-training of N hidden layers is completed. 5. the chemical industry fault detection method based on particle swarm optimization and noise reduction sparse coding machine according to claim 4, is characterized in that, described step (4) specifically comprises: First, after the above-mentioned first hidden layer noise reduction sparse auto-encoder DSA ₁ is trained, its training sample X _trainwhite is encoded as h ₁ by the following formula: h ₁ =f(W ₁ X+b ₁₁ ) (11) Among them, W ₁ , b ₁₁ are the model parameters of the trained DSA ₁ , X is the training set matrix when no noise is added, X is the preprocessed training set X _trainwhite in the first hidden layer, and the second-N hidden layer is represented as the feature matrix learned by the previous layer; Then use the learned first hidden layer feature h ₁ as the input of the second hidden layer noise reduction sparse auto-encoder DSA ₂ , and repeat steps (1-3) to train DSA ₂ to obtain the model weight of the second hidden layer parameter and bias term W ₂ , b ₁₂ , b ₂₂ , encode the feature of the second hidden layer as h ₂ through the above formula (11), repeat this process until the DSA _N training is completed, and obtain the final feature h _N learned by its N hidden layers, The features learned by pre-training are used for the subsequent classification layer, and the model parameters of the Softmax classifier can be trained. 6. the chemical industry fault detection method based on particle swarm optimization and noise reduction sparse coding machine according to claim 5, is characterized in that, in step 3, the supervised pre-training Softmax classifier in described off-line training specifically passes following two steps to proceed: (1) The Nth hidden layer feature hN learned after SDSA training with _N hidden layers is used as the input of the Softmax classifier model, and then the corresponding working condition label is attached y ⁱ =1 represents that the ith sample is normal, and y ⁱ =2 represents that the ith sample is faulty: first initialize its model parameter matrix θ, and predict the probability value of its sample belonging to each category according to the following Softmax classifier prediction function: in, is the probability value of the i-th sample belonging to each category, and θ is the model parameter matrix of the Softmax classifier, which is represented by Vector composition; k is the number of categories defined by the classifier, where k=2; is the feature vector of the Nth hidden layer of the ith sample; (2) Construct a classification loss function to obtain the optimal model parameters. The Softmax classifier used considers the regularization term, which can effectively avoid overfitting of the classification model training. The loss function is defined as follows: Among them, the meaning of the 1{.} part is that if the index function in the _parentheses is true, it will return 1, otherwise it will _return 0; Loss function, call the matlab toolbox minFunc to optimize the minimum loss function, so as to obtain the optimal parameter matrix θ of the Softmax classifier model. 7. the chemical industry fault detection method based on particle swarm optimization and noise reduction sparse coding machine according to claim 6, is characterized in that, in step 3, the BP algorithm in described off-line training global fine-tuning network parameter is specifically carried out by following steps : Take all pre-trained model parameters {(W ₁ , b ₁₁ ), (W ₂ , b ₁₂ ), ..., (W _N , b _1N ), θ} as initial values, and take the pre-trained training samples as initial values X _trainwhite is the input layer. First, use equation (11) to perform layer-by-layer feature learning through the initial parameters of SDSA model, so as to obtain the feature matrix of the final hidden layer, and calculate the loss value of equation (13) through the feature matrix through the classification layer, and then The feedback propagation algorithm is used to optimize the global parameters, and all model parameters are updated once in each iteration, so that the loss value of equation (13) can be converged to the minimum, and the fine-tuning process is completed. 8. The chemical industry fault detection method based on particle swarm optimization and noise reduction sparse coding machine according to claim 7, is characterized in that, in step 3, the particle swarm optimization adjustable hyperparameter in described offline training specifically passes following steps conduct: (1) Optimizing three key adjustable parameters, namely: regularization weight decay hyperparameter λ _r , weight hyperparameter β controlling sparsity penalty term, and Softmax weight decay term coefficient λ _sm . Three adjustable parameters constitute a single particle (λ _r , β, λ _sm ) optimized by PSO, so the dimension of each particle is 3 dimensions; define N _p particles for simultaneous optimization, set K iterations, and initialize the initial value of each particle Position and speed, the fitness function value of which is the overall accuracy of the training samples. The calculation method uses these three key adjustable hyperparameters as independent variables. After training the above fine-tuned fault detection model, the overall training sample can be obtained. Accuracy, its formula is defined as the following formula: Among them, pi represents the category of the ⁱ -th training sample predicted by the fault detection model, and ^yi represents the category of the i-th training sample in the artificially given label. If the two are equal, it will return 1. If they are not equal, it means that the current sample is wrongly diagnosed by the model, and the return value is 0; the optimal fitness and position of each particle and the global optimal fitness and position of the particle swarm are calculated through the above formula (14); (2) Update the velocity and position of each particle, and suppose that the position of the particle under the t-th iteration can be expressed as The flight speed of each particle in the t-th iteration can be expressed as The PSO algorithm updates the velocity and position of each particle by the following formulas: Among them, W _p is the inertia coefficient, C ₁ is the acceleration coefficient of the particle tracking its own historical optimal value, indicating the particle's cognition of itself; C ₂ is the acceleration coefficient of the particle tracking group optimal value, indicating the particle's cognition of the group knowledge , namely social knowledge, usually set C ₁ =C ₂ =2; t is the current number of iterations; ξ and η are uniformly distributed random numbers in the interval [0, 1]; is the position of the optimal fitness experienced by particle i up to the t-th iteration; gb ^t is the best position experienced by the particle swarm up to the t-th iteration; (3) Retrain the unsupervised pre-training SDSA of the above fault detection model, the supervised pre-training Softmax classifier, and the BP algorithm to globally fine-tune the network parameters to obtain a new global optimal fitness value and position, and use the formula (15- 16) Continuously update the position and speed of the particle, and terminate the optimization when the number of iterations exceeds its defined K iterations, so as to obtain the particle position under the global optimal fitness, that is, under the optimal adjustable hyperparameter, its The detection accuracy rate of the training samples is the highest, thus completing the automatic tuning of hyperparameters; then set the optimized tunable hyperparameters as SDSA's optimized tunable hyperparameters, and retrain the determined network structure and the optimal tunable hyperparameters. All model parameters of SDSA under the optimal hyperparameters are obtained to obtain a trained fault detection model, whose model parameters can be used for feature learning and classification of test samples. 9. the chemical industry fault detection method based on particle swarm optimization and noise reduction sparse coding machine according to claim 8, is characterized in that, in step 4, the feature learning of described online monitoring and membership probability prediction are carried out by following steps: Taking the preprocessed test sample X _testwhite as the input layer, according to the trained fault detection model, for the determined neural network structure, the layer-by-layer feature learning method is adopted, and the weight parameters of the trained noise reduction sparse auto-encoder are passed { (W ₁ , b ₁₁ ), (W ₂ , b ₁₂ ), ..., (W _N , b _1n )}, use formula (11) to perform feedforward learning to obtain the features of the current hidden layer of the test sample, and then use the current The features learned by the hidden layer are used as input, and the DSA model parameters of the next hidden layer are combined with the same formula to learn the features, and so on after all hidden layers are learned to obtain the final learned features, and then formula (12) is used for each membership. For category prediction, in fault detection, there are only two categories, namely fault state and normal state. Therefore, for a single sample, the predicted probability vector has only two values, and the predicted category with the largest probability value is selected as the membership predicted by the entire fault detection model. When the calculated affiliation category is 1, it means that the working condition is normal; when the affiliation category is 2, it means that there is an abnormal working condition, and then the connected alarm device will issue an early warning, indicating the fault is detected, and notify the technician at the same time. Or engineers can check the system safety and troubleshoot in time, so that the current working conditions can be monitored for faults.