CN111274737A - Method and system for predicting remaining service life of mechanical equipment - Google Patents

Abstract

The present disclosure discloses a method and system for predicting the remaining service life of mechanical equipment, including: using a time convolution network as a feature extraction algorithm, a long short-term memory network as a regression prediction algorithm, constructing a deep neural network life prediction model, and training the life of the deep neural network Prediction model: According to the model of the device under test and the time sequence of data collection, the collected real-time operation data of the device under test is constructed into a life prediction data set with time series characteristics; the life prediction data set is predicted with a deep neural network life prediction model. , to obtain the remaining service life of the device under test. According to the characteristics of the time series of the state monitoring signal output by the sensor monitoring the mechanical equipment, the time convolution network and the long short-term memory network are combined to establish a deep neural network life prediction model to predict the RUL of the mechanical equipment, and solve the general deep neural network model. The existing problems of over-fitting and vanishing gradients can improve the prediction accuracy.

Description

Method and system for predicting remaining service life of mechanical equipment

technical field

The present disclosure relates to the technical field of mechanical equipment maintenance prediction, and in particular, to a method and system for predicting the remaining service life of mechanical equipment.

Background technique

The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

With the development of modern science and technology, computer technology has been widely used, and the functions of various systems have become more and more perfect. People have put forward higher requirements for the reliability of the system under long-term cycle and high load. For some large-scale system platforms, such as ships , aircraft, etc., it is even more necessary to carry out intelligent maintenance of its key components and perform real-time life prediction tasks. The Remaining Useful Life (RUL) of mechanical equipment generally refers to the time difference between the current operating time and the time when the mechanical equipment fails. Remaining service life prediction technology refers to the long-term condition monitoring of mechanical equipment components under normal operating conditions or working conditions to calculate how long the mechanical equipment components can continue to operate safely and reliably under normal operating conditions. As a result, the health status of the equipment can be judged, and fault prediction can be made, so that maintenance plans can be formulated and spare parts ordered in advance. For large platforms such as ships, this technology can perform remote health management based on the monitoring data sent back during their departure from the port, predict faults in advance, and make optimal maintenance decisions. Reduce maintenance time.

At present, due to the more complex structure of integrated and intelligent equipment, the difficulty of maintenance and prediction also increases. With the continuous advancement of industrialization, the maintenance and support work of mechanical equipment is becoming more and more heavy, and the workload of operation and maintenance is increasing. There are several types of RUL prediction techniques, such as expert systems, statistical methods, and physical model-based methods. However, the inventors found that the above methods have at least the following problems:

(1) The expert system is mainly an intelligent program established by combining the knowledge and experience accumulated by experts in related fields. According to the operating state of the mechanical equipment, it can reason about the equipment failure or remaining life without establishing an accurate mathematical prediction model, nor does it rely on historical data. However, it is very difficult to obtain expert knowledge and requires a long period of accumulation; the mobility of expert systems is poor. Usually, an expert system developed for a certain type of equipment or mechanical parts is only valid for this type of equipment and cannot be applied to other equipment, which makes the expert system development High cost, but low utilization.

(2) Statistical method is to use statistical model to predict RUL, and the commonly used models are exponential distribution model and Weibull distribution model. However, the disadvantage of this method is that it is necessary to accurately analyze the degradation principle of mechanical equipment before establishing the statistical distribution law of degradation influencing factors and life.

(3) Based on the physical model method, the physical characteristics of the equipment are firstly studied, and the physical degradation evaluation standard is established and expressed by physical and mathematical formulas, so as to predict the remaining service life of the mechanical equipment. However, the disadvantage of this method is that the theoretical analysis is very difficult, and it is difficult to obtain an accurate physical degradation model in practice. Generally, there is a certain deviation between the built model and the real failure mechanism, so the research on the RUL prediction method based on the physical model is relatively difficult. In addition, the problem of difficult migration restricts the wide application of this technology.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present disclosure proposes a method and system for predicting the remaining service life of mechanical equipment. According to the characteristics of the time series of the state monitoring signal output by the sensor monitoring the mechanical equipment, the time convolution network and the long short-term memory network are combined. , establish a deep neural network life prediction model for RUL prediction of mechanical equipment, solve the problem of over-fitting and gradient disappearance in general deep neural network models, and improve the prediction accuracy at the same time.

In order to achieve the above object, the present disclosure adopts the following technical solutions:

In a first aspect, the present disclosure provides a method for predicting the remaining service life of mechanical equipment, including:

Using the temporal convolutional network as the feature extraction algorithm and the long short-term memory network as the regression prediction algorithm, a deep neural network lifetime prediction model is constructed, and the historical operation data is used as the training data to train the deep neural network lifetime prediction model;

According to the model of the device under test and the time sequence of data collection, the collected real-time operation data of the device under test is constructed into a life prediction data set with time series characteristics;

Use the trained deep neural network life prediction model to perform prediction processing on the life prediction data set to obtain the remaining service life of the device under test.

In a second aspect, the present disclosure provides a system for predicting the remaining service life of mechanical equipment, including: a historical database, a life prediction model, and a full life cycle database;

The lifespan prediction model includes constructing a deep neural network lifespan prediction model, the deep neural network lifespan prediction model is constructed with a time convolutional network as a feature extraction algorithm, a long short-term memory network as a regression prediction algorithm, and is based on the historical operation in the historical database. Data training deep neural network life prediction model;

The full life cycle database stores the collected real-time operation data of the equipment under test, and constructs the collected real-time operation data of the equipment under test into a life prediction data set with time series characteristics according to the model of the equipment under test and the time sequence of data collection;

The full life cycle database inputs the life prediction data set into the life prediction module, and performs prediction processing on the life prediction data set with the trained deep neural network life prediction model to obtain the remaining service life of the device under test.

Compared with the prior art, the beneficial effects of the present disclosure are:

The prediction method proposed in the present disclosure is based on data-driven, and the advantage of data-driven is that the monitoring data of the equipment is processed by using the established deep neural network life prediction model, and there is no need for complex theoretical research and maintenance personnel to accumulate long-term expertise. The knowledge can then use the model to make RUL predictions.

After the prediction model is established, for the RUL prediction task of different devices, it is only necessary to use the monitoring data of the corresponding device to train the model, adapt to different types of devices, and has good transferability.

At the same time, the present disclosure combines the time convolution network and the long short-term memory network to establish a deep neural network lifetime prediction model for the RUL prediction of the mechanical equipment according to the characteristics of the time series of the state monitoring signal output by the sensor monitoring the mechanical equipment, so as to solve the problem of general Deep neural network model, over-fitting problem and gradient disappearance problem in the training process, improve the prediction accuracy.

Description of drawings

The accompanying drawings that constitute a part of the present disclosure are used to provide further understanding of the present disclosure, and the exemplary embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation of the present disclosure.

Figure 1 is a classification diagram of the current mechanical equipment RUL prediction method;

2 is a schematic flowchart of the method provided in Embodiment 1 of the present disclosure;

FIG. 3(a) is a convolution graph with a time convolutional network convolution kernel size of 3 and an expansion rate of 1 provided in Embodiment 1 of the present disclosure;

FIG. 3(b) is a convolution graph with a temporal convolution network expansion rate of 2 provided in Embodiment 1 of the present disclosure;

FIG. 3(c) is a convolution graph with a time convolutional network expansion rate of 3 provided in Embodiment 1 of the present disclosure;

Figure 4(a) shows the feature extraction process of the existing temporal convolutional network;

Figure 4(b) feature extraction process of the TCN network provided in Embodiment 1 of the present disclosure;

5 is a basic structural diagram of a TCN network provided by Embodiment 1 of the present disclosure;

6 is a basic structural diagram of a long short-term memory LSTM network provided in Embodiment 1 of the present disclosure;

FIG. 7 is a change curve of the predicted value root mean square value provided in Embodiment 1 of the present disclosure;

FIG. 8 is a prediction error curve of the test sample provided in Embodiment 1 of the present disclosure;

FIG. 9 is a structural diagram of a system for predicting the remaining service life of mechanical equipment according to Embodiment 2 of the present disclosure.

Detailed ways:

The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

It should be noted that the following detailed description is exemplary and intended to provide further explanation of the present disclosure. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs.

It should be noted that the terminology used herein is for the purpose of describing specific embodiments only, and is not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural as well, furthermore, it is to be understood that when the terms "comprising" and/or "including" are used in this specification, it indicates that There are features, steps, operations, devices, components and/or combinations thereof.

Example 1

During the maintenance of mechanical equipment, if the remaining service life of the equipment can be accurately predicted, the warning value of the life of the equipment can be known in advance. Equipment maintenance, replacement of parts, replacement of spare equipment, etc., to avoid equipment failure during operation, predict and "cure" the disease before the equipment "falls on." In this way, accidents caused by equipment failures can be avoided, economic losses can be reduced, and personnel safety can be protected; maintenance plans can be formulated in advance according to the results of life prediction, spare parts or spare equipment can be purchased in advance, so as to reduce downtime, improve maintenance efficiency, and reduce transportation costs. Maintenance cost; can coordinate the respective tasks of multiple devices according to the forecast results, formulate corresponding production plans, and improve production efficiency. As shown in FIG. 1, the current RUL prediction method for mechanical equipment is shown. However, as described in the background art, the existing prediction method has many problems. The method in this embodiment has the characteristics of time series according to the state monitoring signal output by the sensor monitoring the mechanical equipment. , Combining temporal convolutional network and long-term and short-term memory network to establish a deep neural network life prediction model for RUL prediction of mechanical equipment, solve the problem of over-fitting and gradient disappearance in general deep neural network models, and improve the prediction accuracy at the same time .

As shown in FIG. 2 , this embodiment provides a method for predicting the remaining service life of mechanical equipment, including:

S1: Use the time convolutional network as the feature extraction algorithm and the long short-term memory network as the regression prediction algorithm to build a deep neural network lifetime prediction model, and use the historical operating data as the training data to train the deep neural network lifetime prediction model;

S2: According to the model of the device under test and the time sequence of data collection, the collected real-time operation data of the device under test is constructed into a life prediction data set with time series characteristics;

S3: Use the trained deep neural network life prediction model to perform prediction processing on the life prediction data set to obtain the remaining service life of the device under test.

In the step S1, collect historical data, use sensors installed on the equipment to be tested to collect operating data, and use the historical operating data of the same type of equipment or components under the same environment and the same working conditions as the model prediction of the equipment to be tested. Training data; through pre-processing of data, data enhancement, etc., the historical operation data is standardized and used as the training data of the life prediction model.

Among them, the device under test is a life cycle from the initial stage of operation to the failure, and a full life cycle database is constructed, and the collected real-time operation data of the device under test is stored in the full life cycle database. When a life cycle of the device under test ends, The operation data of the device under test in a life cycle is regarded as the historical operation data.

The full life cycle database includes inputting the real-time operation data of the device under test as life prediction data into the prediction model; when one life cycle of the device ends, adding the operation data set of the device in one life cycle to the historical database , expand the historical database, so as to continuously optimize and improve the life prediction model.

In the step S1, the construction of the deep neural network life prediction model includes two stages: the first stage uses an improved Temporal Convolutional Network (TCN) for feature extraction, and the second stage uses a long short-term memory network ( Long-short Term Memory, LSTM) for regression prediction. After the model is established, the model is trained using the state signal collected by real-time monitoring of the equipment, so as to construct a complex mapping relationship between the feature state and the remaining life, including:

(1) The time convolution network is different from the general convolution network. When the time convolution network (TCN) performs the convolution operation, it skips part of the input, so that the convolution kernel is applied to the convolution kernel larger than its own size. region, which is equivalent to dilating the original convolution kernel into a larger kernel by zero-padding.

Taking two-dimensional convolution as an example, as shown in Figure 3(a)-3(c); in Figure 3(a), the size of the convolution kernel is s=3, the dilated rate r=1, and the perceptual field of view y= s*s=9; in Fig. 3(b), the expansion rate of the convolution kernel is r=2, and the perceptual field of view y=(s*r+r-1)^2=49; in Fig. 3(c), the convolution kernel is expanded Rate r=3, perceptual field of view y=(s*r+r-1)^2=121;

It can be seen that when performing convolution operations, part of the input data is regularly omitted, so that the perceptual field of view covered by the convolution kernel is increased. For processing time series data, feature information in a wider time range can be perceived. When the number of network layers increases, the perceptual field of view increases exponentially; when the number of network layers deepens to a certain extent, the gradient will gradually disappear during the propagation process, that is, the gradient disappears, resulting in the inability to effectively adjust the weights of the previous network layers. This embodiment solves the above problem by introducing a residual block.

As shown in Figure 4(a): for the input X, the learning target Y, the convolutional layer or the fully connected layer will undergo Relu nonlinear transformation when the information is transmitted, and there will be problems such as information loss and loss, and the learning is difficult. When the number increases, the loss will increase cumulatively, affecting the accuracy of the network.

As shown in Figure 4(b): In this embodiment, a cross-layer connection residual block is added to connect the input X to the output layer across layers. When the target Y remains unchanged, the task of the network becomes learning Y-X=H (x), hence the name residual block. In this way, in the learning target Y=H(x)+X, only H(x) has undergone the nonlinear transformation of Rule, and only H(x) has information loss loss, which reduces feature loss and eliminates the problem of gradient disappearance.

Finally, the basic structure of the TCN network module is shown in Figure 5. In the Dropout process, some neurons in the hidden layer are randomly deleted first, the input x is propagated forward to the output, and then the error is propagated backward to correct the parameters, and a small batch of samples are trained. After missing the network, the retained neurons have their parameters updated, while the deleted neurons do not have their parameters updated. Then restore the previously deleted neurons, and then randomly delete a part, that is, repeat the above steps until the sample training is completed. In the whole process, half of the hidden neurons are randomly deleted, resulting in a different network structure. The entire dropout process is equivalent to taking the average of many different neural networks, and different networks produce different overfitting, some of which are "reverse" to each other. "The fittings cancel each other out to reduce overfitting as a whole. The complex co-adaptive relationship between neurons is reduced, the hidden layer nodes appear randomly together, and the weight update does not depend on the interaction of the hidden layer nodes with inherent relationship, and the robustness is better.

Stacking multi-layer TCN networks constitutes the feature extraction part of the deep learning model.

(2) The long-term and short-term memory network LSTM performs regression prediction. LSTM is an improved network based on RNN. It has a more complex structure and can remember long-term states, mainly by adding gate control units Input Gate, Output Gate, and Forget Gate. To control the update of variables in the network, it can handle long-term time series; Input Gate, OutputGate and Forget Gate are all learned by learning by themselves, and their structure is shown in Figure 6:

Taking X=(x ₀ , x ₁ , x ₂ ,...,x _t-1 ,x _t ,...,x _n ) as input, the update of parameters and variables in LSTM includes:

f _t =σ(W _f x _t +U _f h _t-1 +b _f ),

i _t =σ(W _i x _t +U _i h _t-1 +b _i ),

o _t =σ(W _o x _t +U _o h _t-1 +b _o ),

h _t =o _t *tanh(c _t ),

Among them, W, U, b are the weight matrix and bias vector of x, h respectively; the activation function σ, tanh are the sigmoid function and the hyperbolic tangent function, respectively.

As shown in Figure 6, the leftmost path determines whether the current cell state has content that needs to be forgotten according to the input and the output of the previous moment. After the output passes through the sigmoid function, the closer it is to 0, the more forgotten, and the closer it is to the sigmoid function. 1 The less is forgotten; the middle path depends on the sigmoid function to decide what should be remembered; the rightmost path uses the sigmoid function as a gate, and filters the result after tanh on the state obtained in the second step to get the final prediction result.

According to the above technical ideas, this embodiment uses 4 layers of TCN and 2 layers of LSTM to construct a deep neural network lifetime prediction model.

In the step S2, in the normal working state of the equipment, use the acceleration sensor and other types of sensors to collect real-time operating signals; it should be noted that the data used for life prediction is the entire life cycle of the equipment from the start of work to the failure. Carry out collection to ensure the regularity and integrity of the collection process.

For the data used for life prediction, first check whether the data format is correct, and then preprocess the data, including data fusion, standardization and other operations, and make a preliminary judgment on the obvious abnormal working status of the equipment.

From the beginning of equipment operation, according to the model of the equipment under test and the time sequence of data collection, the collected real-time operation data of the equipment under test is constructed into a life prediction data set with time series characteristics; The prediction data set is used for the task of life prediction, and according to the prediction results, the health status of the equipment is evaluated and the remaining life is given.

At the same time, according to the remaining service life of the device under test, predict the location and cause of the failure of the device under test, and formulate a maintenance plan. When the predicted service life obtained is short, predict the possible failure locations and possible parts of the equipment, focus on monitoring the predicted failure locations or components, and formulate maintenance plans in advance according to the actual situation, such as replacing lubricating oil, replacing bearings or more Sliding spare equipment, etc., so as to avoid system downtime or other significant loss of personnel and property.

Experimental verification

In order to verify the effectiveness of the method in this embodiment, the public data set C-MAPSS is used as the input data for model training. NASA released the CMAPSSData dataset for the PHM-themed competition. This dataset only uses historical data (run-to-failure) to predict equipment remaining life (RUL) without considering the underlying physical factors. Each engine equipment is installed with 21 sensors to monitor its running status. As time goes on, each engine equipment will have some faults, and the signal acquisition will end at the faulty node. The data includes engine number markers, time-series markers, three operating conditions, and data read from 21 sensors.

After the above neural network is iteratively trained using the C-MAPSS data set, the above network is tested using the test set. The results are shown in Figure 7 and Figure 8. In Figure 7, "cost" is the root mean square value of the model prediction result. , as the number of training increases, the number of samples increases, and the error tends to zero; in Figure 8, "errors" is the difference between the predicted lifespan y^' and the actual lifespan y. The RUL prediction method of the method in this embodiment can effectively perform RUL prediction, and the prediction accuracy on the test data set reaches more than 99%.

Example 2

As shown in FIG. 9 , this embodiment provides a system for predicting the remaining service life of mechanical equipment, including: a historical database, a life prediction model, and a full life cycle database;

The historical database is to store all the state monitoring data accumulated by the equipment of the same model and under the same working conditions, and use it as the training data to train the life prediction model;

The lifespan prediction model includes constructing a deep neural network lifespan prediction model, and the deep neural network lifespan prediction model is constructed using a time convolutional network as a feature extraction algorithm and a long short-term memory network as a regression prediction algorithm;

The full life cycle database stores the collected real-time operation data of the equipment under test, and constructs the collected real-time operation data of the equipment under test into a life prediction data set with time series characteristics according to the model of the equipment under test and the time sequence of data collection;

The full life cycle database inputs the life prediction data set into the life prediction module, and performs prediction processing on the life prediction data set with the trained deep neural network life prediction model to obtain the remaining service life of the device under test.

The full life cycle database stores the device model, device operation configuration, real-time data collected by sensors, etc. The device under test from the initial stage of operation to failure is a life cycle, and the collected real-time operation data of the device under test is stored in the Full life cycle database, when a life cycle of the device under test ends, the operation data of the device under test in a life cycle is added to the historical database.

The system further includes an operation monitoring module, which collects the operation data of the device under test in real time through the state monitoring sensor and the acceleration sensor installed on the device under test, and stores it in the database of the whole life cycle.

In this embodiment, a life prediction model database is established by collecting historical data and data enhancement; combining TCN and LSTM algorithm to build a remaining service life prediction model, and using the established historical database to train the model; and other parameters are measured, and after the data is preprocessed, the trained prediction model is used for life prediction. According to the prediction results, the health status of the equipment is judged, the possible faults of the equipment are discovered in time, the location and severity of the faults are analyzed, and different maintenance plans are proposed according to the analysis results. It greatly reduces the complexity of equipment maintenance, improves the timeliness and scientificity of equipment maintenance, reduces equipment maintenance costs, and ensures safe and reliable operation of equipment.

The above are only preferred embodiments of the present disclosure, and are not intended to limit the present disclosure. For those skilled in the art, the present disclosure may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included within the protection scope of the present disclosure.

Although the specific embodiments of the present disclosure have been described above in conjunction with the accompanying drawings, they do not limit the protection scope of the present disclosure. Those skilled in the art should understand that on the basis of the technical solutions of the present disclosure, those skilled in the art do not need to pay creative efforts. Various modifications or variations that can be made are still within the protection scope of the present disclosure.

Claims (10)

1. A method for predicting the remaining useful life of mechanical equipment, comprising: Using the temporal convolutional network as the feature extraction algorithm and the long short-term memory network as the regression prediction algorithm, a deep neural network lifetime prediction model is constructed, and the historical operation data is used as the training data to train the deep neural network lifetime prediction model; According to the model of the device under test and the time sequence of data collection, the collected real-time operation data of the device under test is constructed into a life prediction data set with time series characteristics; Use the trained deep neural network life prediction model to perform prediction processing on the life prediction data set to obtain the remaining service life of the device under test. 2. A method for predicting the remaining service life of mechanical equipment as claimed in claim 1, characterized in that, using sensors installed on the equipment under test to collect operating data, collecting the equipment under test and the equipment under test and having the same model as the equipment under test. The historical operating data of the equipment under the same working conditions are preprocessed for data enhancement and normalization, and the preprocessed historical operating data is used as training data. 3. The method for predicting the remaining service life of a mechanical equipment as claimed in claim 1, wherein the equipment under test is a life cycle from the initial stage of operation to failure, and a full life cycle database is constructed, and the collected tested equipment is a life cycle. The real-time operation data of the equipment is stored in the full life cycle database. When a life cycle of the device under test ends, the operation data of the device under test in a life cycle is used as the historical operation data. 4. The method for predicting the remaining service life of mechanical equipment according to claim 1, wherein a cross-layer connection residual block is added to the time convolutional network, and the cross-layer connection residual block is: The input data X is connected to the output layer across layers, and when the learning target Y remains unchanged, the learning task H(x) of the temporal convolutional network becomes learning H(x)=Y-X. 5. The method for predicting the remaining useful life of mechanical equipment according to claim 1, wherein in the long short-term memory network, whether the current state needs to be forgotten is determined according to the input data and the output data of the previous moment. The content, according to the sigmoid function, determines the content that needs to be remembered, and filters through the tanh function to obtain the final prediction result. 6 . The method for predicting the remaining service life of mechanical equipment according to claim 1 , wherein, according to the remaining service life of the equipment under test, the location and cause of the failure of the equipment under test are predicted, and a maintenance plan is formulated. 7 . 7. A system for predicting the remaining service life of mechanical equipment, comprising: a historical database, a life prediction model and a full life cycle database; The lifespan prediction model includes building a deep neural network lifespan prediction model. The deep neural network lifespan prediction model uses a time convolutional network as a feature extraction algorithm and a long short-term memory network as a regression prediction algorithm. Data training deep neural network life prediction model; The full life cycle database stores the collected real-time operation data of the equipment under test, and constructs the collected real-time operation data of the equipment under test into a life prediction data set with time series characteristics according to the model of the equipment under test and the time sequence of data collection; The full life cycle database inputs the life prediction data set into the life prediction module, and performs prediction processing on the life prediction data set with the trained deep neural network life prediction model to obtain the remaining service life of the device under test. 8 . The system for predicting the remaining service life of mechanical equipment according to claim 7 , wherein the system further comprises an operation monitoring module, and the operation monitoring module adopts the condition monitoring sensor and Acceleration sensor collects the operating data of the device under test in real time and stores it in the full life cycle database. 9. The system for predicting the remaining service life of mechanical equipment according to claim 7, wherein the historical database stores the historical operation data of the equipment with the same model and the same working condition as the equipment under test, so as to This is the training data. 10 . The system for predicting the remaining service life of mechanical equipment according to claim 7 , wherein, in the full life cycle database, the equipment under test is a life cycle from the initial stage of operation to failure, and a full life cycle is constructed. 11 . The cycle database stores the collected real-time operation data of the device under test in the full life cycle database. When a life cycle of the device under test ends, the operation data of the device under test in a life cycle is used as the historical operation data.

Images