CN108960077A - A kind of intelligent failure diagnosis method based on Recognition with Recurrent Neural Network - Google Patents

Abstract

The invention discloses an intelligent fault diagnosis method based on a cyclic neural network, which comprises the following steps: (1) using an acceleration sensor to obtain time-series vibration signals of rotating machinery working in different health states, and dividing the obtained original vibration signals into training sets and Test set; (2) establish a recurrent neural network; (3) train the recurrent neural network; (4) test the trained recurrent neural network, and judge whether the network reaches the expected diagnostic goal according to the classification results, if the accuracy is low If it is lower than the expected value, then repeat step (3) until obtaining a cyclic neural network whose accuracy is higher than the expected value; (5) carry out intelligent fault diagnosis through the cyclic neural network obtained in the step (4). The invention utilizes the modeling ability of the cyclic neural network for the sequence information to directly process the original time series vibration signal, can make full use of less information to accurately diagnose the fault of the rotating machinery, and has a high identification speed.

Description

An Intelligent Fault Diagnosis Method Based on Recurrent Neural Network

technical field

The invention relates to the technical field of processing vibration signals of rotating machinery, in particular to an intelligent fault diagnosis method based on a cyclic neural network.

Background technique

The fault diagnosis method mainly focuses on the operating state of the system, which can detect faults in time and guide maintenance, which plays an important role in improving system reliability. In general, when a rotating component fails, it will be accompanied by a change in the form of vibration, resulting in an instantaneous vibration pulse. Therefore, the vibration signal carries important diagnostic information and is an important basis for equipment status identification. Mechanical equipment usually needs to work in a complex environment with strong background noise, so the mechanical vibration signals acquired on site are usually multi-component, non-stationary signals with background noise. Therefore, it becomes very difficult to extract fault features from complex mechanical vibration signals to diagnose and classify mechanical vibration signals with similar fault modes. The leap-forward improvement in signal acquisition capabilities of modern condition monitoring systems has brought fault diagnosis into the era of big data, which has magnified the defects of traditional fault diagnosis methods. Therefore, modern machinery fault diagnosis requires new signal processing methods.

In recent years, deep learning technology has developed rapidly, and it has shown extremely high recognition accuracy in many traditional recognition tasks such as images and voices, demonstrating its great potential in data processing. Deep learning technology can automatically learn and extract signal features from data for classification, which is an excellent choice to replace traditional artificial signal processing technology. There have been a lot of research and application of deep learning technology in the field of fault diagnosis, including Boltzmann machine, deep belief network (Deep Belief Network, DBN), autoencoder, convolutional neural network (Convolutional Neural Networks, CNN), sparse filter (Sparse Filtering, SF) and so on. However, the current intelligent diagnosis algorithm still has the following two problems: (1) A large amount of training data is required. When the length of the input signal sequence is not enough to store information, these methods are often difficult to maintain a high enough accuracy, resulting in the comparison of diagnostic results. one-sided. (2) The training data needs to be preprocessed, which leads to a large manual workload and affects the diagnosis efficiency. These issues limit the flexibility of diagnostic methods.

Contents of the invention

Purpose of the invention: for overcoming the defect of above-mentioned existing method, the purpose of the present invention is to provide a kind of intelligent fault diagnosis method based on cyclic neural network, introduce Dropout training technology on the basis of gated cyclic unit cyclic neural network, strengthen model for small The ability to identify dimensional signals.

Technical solutions:

A method for intelligent fault diagnosis based on cyclic neural network, comprising the steps of:

(1) Use the acceleration sensor to obtain the time-series vibration signals of the rotating machinery working in different health states, and divide the obtained original vibration signals into two non-overlapping parts, which are respectively used as the training set and the test set;

(2) set up recurrent neural network: the model of described recurrent neural network selects the gated recurrent unit, and the model is input with the vibration signal x _t , and the forward propagation formula of the model is:

z _t = σ(W _z x _t +U _z h _t-1 +b _z ) (1)

r _t ＝σ(W _t x _t +U _t h _t-1 +b _r ) (2)

In the formula, x _t is the input data at the current time t, h _t-1 is the hidden unit information output at the previous time t-1, z _t is the output of the update gate, r _t is the output of the reset gate, is the candidate hidden state, h _t is the hidden state, σ is the Sigmoid activation function, tanh is the hyperbolic tangent activation function, W and U are the neural network weight matrix, b is the bias, "●" means the Hadamard product, that is, two The product of corresponding elements of a matrix;

The softmax function is used to classify the output of the recurrent neural network, and the objective function of each recurrent unit is:

In the formula, L is the objective function, K is the dimension of the health status category, l ⁽ⁱ⁾ is the health status label, is the output of the hidden layer, e represents a natural constant, 1{ } represents an indicator function that outputs 0 or 1 according to the condition; the final objective function is the synthesis of all single-layer objective functions, which is:

In the formula, M is the number of samples and T is the number of time steps;

(3) Input the time series vibration signal training set with the health status label obtained in step (1) into the recurrent neural network established in step (2) for training, and compare the difference between the classification result and the original signal health label. The network parameters are fine-tuned, including the input of the network, to obtain the trained recurrent neural network;

(4) Input the time-series vibration signal test set with the health state label obtained in step (1) into the recurrent neural network trained in step (3) to diagnose and classify the health state, and judge whether the network meets the expected diagnosis according to the classification results Target, if the accuracy is lower than the expected value, then repeat step (3) until obtaining the recurrent neural network whose accuracy is higher than the expected value; the expected value is 95%;

(5) Carry out intelligent fault diagnosis through the recurrent neural network obtained in the step (4).

The time-series vibration signals in the step (1) include the vibration signals of the system in normal state, wear, broken teeth, and combined faults of broken teeth and wear.

In the recurrent neural network training process of described step (3), introduce Dropout technology in the input and output of recurrent neural network hidden layer:

In the hidden layer, the input and output of the hidden unit at each moment are set to zero with a set ratio p, so that each calculation is jointly worked by randomly selected neurons that are mandatory.

Beneficial effect:

1. The present invention utilizes the modeling ability of the gated cyclic unit cyclic neural network for sequence information to directly process the original time-series vibration signal, which can make full use of less information to accurately diagnose rotating machinery faults, and has a high recognition speed.

2. The present invention introduces Dropout technology in the training process of the cyclic neural network to prevent over-fitting and enhance the generalization ability of the model.

Description of drawings

Fig. 1 is a flow chart of the method of the present invention.

Figure 2 is a diagram of a recurrent neural network model.

Figure 3 is a comparison of the recognition accuracy of training samples of different sizes.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

Fig. 1 is the flowchart of the intelligent fault diagnosis method based on cyclic neural network of the present invention, as shown in Fig. 1, this method comprises the steps:

(1) Use the acceleration sensor to obtain the time-series vibration signals of the rotating machinery working in different health states: the vibration signal data is obtained through the engine-rotor test bench, and the test works in the form of motor-gearbox-load. The gear box is a one-stage planetary gear box, including 3 planetary gears, 1 sun gear and 1 fixed ring gear. The acceleration sensor is used to collect the vibration acceleration in the vertical direction in the test as the test processing signal. The test mainly collects the vibration signals of the planetary gear in four states: normal, tooth surface wear, broken tooth, and combined failure of broken tooth and wear. The experiment sets the sampling frequency as f _s =12.8kHz. The obtained original vibration signal is divided into two non-overlapping parts, which are respectively used as training set and test set. Randomly select 400 groups of data sets with sample sizes of 600, 800, 1000 and 1200 respectively.

(2) Establishing a recurrent neural network: the model of the recurrent neural network uses a gated recurrent unit, and the network model includes a three-layer neural network including an input layer, a hidden recurrent layer, and an output layer. For this data set, the input layer dimension and the hidden layer dimension are set to 20 and 100, respectively.

Referring to Figure 2, the first layer is the input layer, which reads the segmented training data and expands it to a higher dimension before inputting it to the next hidden layer. The gated recurrent unit of the hidden recurrent layer includes two control gates, a reset gate and an update gate, which are used to learn the inherent fault characteristics of the signal, and the output layer uses the fault characteristics obtained by the hidden recurrent layer to classify the input signal. The model takes the vibration signal x _t as input, and the forward propagation process is calculated according to the following formula:

z _t ＝σ(W _z x _t +U _z h _t-1 +b _z ) (1)

r _t ＝σ(W _t x _t +U _t h _t-1 +b _r ) (2)

where x _t is the input data at the current time t, h _t-1 is the hidden unit information output at the previous time t-1, z _t is the output of the update gate, r _t is the output of the reset gate, is the candidate hidden state, h _t is the hidden state, σ is the Sigmoid activation function, tanh is the hyperbolic tangent activation function, W and U are the neural network weight matrix, b is the bias, "●" means the Hadamard product, that is, two The product of the corresponding elements of a matrix.

The last layer is the output layer, which is used to diagnose the health status, that is, the classification layer. The output layer accepts the final output of the hidden layer, and the objective function adopts the softmax function.

The objective function formula of each cycle unit is:

In the formula, L is the objective function, K is the dimension of the health status category, l ⁽ⁱ⁾ is the health status label, is the output of the hidden layer, e represents a natural constant, and 1{ } represents an indicator function that outputs 0 or 1 according to the condition. The final objective function is the synthesis of all single-layer objective functions, which is:

In the formula, M is the number of samples and T is the number of time steps;

(3) Network training: input the time-series vibration signal set with the health status label into the recurrent neural network for training. First, the vibration signal is input into the input layer for dimension expansion, and then the reconstructed data is input into the loop layer to extract the fault features of different health states, and finally classify the extracted fault features into the output layer, compare the difference between the classification result and the original signal health label, and fine-tune the network parameters to obtain a trained recurrent neural network; the fine-tuning That is, adjust the parameters of the recurrent neural network after training and observe the results until the diagnostic accuracy does not increase after adjusting to a certain value, which belongs to empirical adjustment. The adjustment parameters of the fine-tuning include the dimension of the input data, the hidden dimension of the input layer, the number of iterations, and the like.

Use the collected training sample set to train the established recurrent neural network model. During the training process, dropout technology is introduced into the input and output of the hidden layer of the cyclic neural network to prevent overfitting. That is, in the hidden layer, the input and output of the hidden unit at each moment are set to zero at a set ratio p, so that each calculation is made to work together with randomly selected neurons that are mandatory to achieve the weakening of the neurons. The joint adaptability between nodes enhances the generalization ability and solves the problem of overfitting. In the present invention, the input and output of the hidden unit at each moment have a set ratio of p=0.5.

(4) To verify the test accuracy of the method, input the vibration signal into the trained recurrent neural network to diagnose and classify the health status, and judge whether the network has reached the expected diagnosis target according to the classification results. If the accuracy is lower than the expected value, repeat step 3 until the obtained A neural network with a sufficiently high accuracy rate.

(5) Test: The test samples prepared before are used to verify the effectiveness of the proposed method, and the method is compared with coefficient filtering (SF) and LSTM network. Each experiment was performed 10 times, and the average value was taken to ensure generalizability. The accuracy rate comparison is shown in Figure 3, it can be seen that this method has the highest recognition accuracy. It shows that this method can maintain the ability of feature learning when providing less information.

The above is only a preferred embodiment of the present invention, it should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, some improvements and modifications can also be made. It should be regarded as the protection scope of the present invention.

Claims (3)

1. a kind of intelligent failure diagnosis method based on Recognition with Recurrent Neural Network, characterized by the following steps:

(1) the timing sequence vibration signal that rotating machinery works under different health status is obtained using acceleration transducer, will obtained Original vibration signal be divided into two nonoverlapping parts, respectively as training set and test set；

(2) establish Recognition with Recurrent Neural Network: the model selection gating cycle unit of the Recognition with Recurrent Neural Network, model is with vibration signal x_tFor input, model propagation formula forward are as follows:

z_t=σ (W_zx_t+U_zh_t-1+b_z) (1)

r_t=σ (W_tx_t+U_th_t-1+b_r) (2)

In formula, x_tIt is current t moment input data, h_t-1For the hidden unit information of previous moment t-1 output, z_tIt is to update door Output, r_tIt is resetting door output,It is candidate hidden state, h_tIt is hidden state, σ is Sigmoid activation primitive, and tanh is double Bent tangent activation primitive, W, U are neural network weight matrix, and b is biasing, and "●" indicates Hadamard product, i.e. two matrixes The product of corresponding element；

Classified using softmax function to the output of Recognition with Recurrent Neural Network, the objective function of each cycling element are as follows:

In formula, L is objective function, K is the dimension of health status type, l⁽ⁱ⁾For health status label,For hidden layer output, e Indicate natural constant, 1 { } indicated the indicator function according to output with conditions 0 or 1；

Final goal function is the synthesis of all single target functions, are as follows:

M is sample number in formula, T is time step number；

(3) the timing sequence vibration signal training set with health status label that step (1) obtains successively is segmented and is input to step (2) it is trained in the Recognition with Recurrent Neural Network established, the difference of comparison-of-pair sorting's result and original signal health label is to network parameter It is finely adjusted, to obtain trained Recognition with Recurrent Neural Network；

(4) the timing sequence vibration signal testing collection with health status label that step (1) obtains step (3) is input to train Recognition with Recurrent Neural Network in carry out the diagnostic classification of health status, judge whether network reaches expected and diagnose mesh according to classification results Mark, if accuracy is lower than desired value, repeatedly step (3) is until obtaining Recognition with Recurrent Neural Network of the accuracy higher than desired value；Institute Stating desired value is 95%；

(5) intelligent trouble diagnosis is carried out by the Recognition with Recurrent Neural Network that the step (4) obtains.

2. intelligent failure diagnosis method according to claim 1, it is characterised in that: the timing sequence vibration in the step (1) Signal includes vibration signal of the system under normal condition, abrasion, broken teeth and broken teeth and the combined failure state of abrasion.

3. intelligent failure diagnosis method according to claim 1, it is characterised in that: the circulation nerve net of the step (3) Dropout technology is introduced in network training process in the input and output of Recognition with Recurrent Neural Network hidden layer:

In hidden layer, the ratio p zero setting to set is inputted and exports to each moment hidden unit, so that calculating every time It is worked together by the random select neuron of mandatory arrangement.