WO2023071217A1 - Multi-working-condition process industrial fault detection and diagnosis method based on deep transfer learning - Google Patents

Abstract

Provided is a multi-working-condition process industrial fault detection and diagnosis method based on deep transfer learning, comprising: obtaining historical data of a process industry under a plurality of working conditions, the historical data comprising normal operation data and fault data, and constructing a fault library; standardizing the historical data and arranging same into a two-dimensional matrix; in combination with a maximum mean difference (MMD) of deep transfer learning, training a fault detection model and a fault diagnosis model which are designed, and calculating a detection threshold of the fault detection model; acquiring real-time data, performing corresponding standardization and arrangement processing, then inputting the data into the trained fault detection model to calculate a loss function value, and comparing the loss function value with the detection threshold to determine whether a production system has an anomaly; and if the anomaly occurs, inputting the real-time data into the fault diagnosis model to determine a fault type of the production system.

Description

Fault detection and diagnosis method for multi-working condition process industry based on deep transfer learning

Related Application Cross Reference

This application is based on a Chinese patent application with application number 202111260743.9 and a filing date of October 27, 2021, and claims the priority of this Chinese patent application. The entire content of this Chinese patent application is hereby incorporated by reference into this application.

technical field

The present disclosure relates to the technical field of process industry monitoring, in particular to a multi-working-condition process industry fault detection and diagnosis method based on deep transfer learning.

Background technique

At present, people pay more and more attention to the safety production of the process industry. Once a production accident occurs, it may cause serious loss of life and property and environmental damage. With the advent of the Internet of Things era and the development of information technology, the process industry can collect process data through a large number of sensors for fault detection and diagnosis. Among them, fault detection and diagnosis technology is the key basic technology in the field of process industry safety, and fault detection can be detected in real time. Whether an abnormal event occurs in the production system, fault diagnosis is to further judge which type of fault occurred after the abnormal event occurs, and assist the operator to understand the production process information in a timely and effective manner.

In related technologies, fault detection technology usually uses principal component analysis and autoencoder, principal component analysis is a multivariate statistical method, the method is simple and easy to understand; autoencoder is a deep neural network method, with stronger feature extraction ability, suitable for For large-scale data modeling, fault detection technology usually uses normal operating data for modeling to obtain a fault detection model. When monitoring the production process online in real time, the fault detection model is used to detect real-time data and determine whether the current system is abnormal. The fault diagnosis method usually adopts Fault database search and fault classification methods. The fault database search is to compare and match the current fault data with the fault database data to find the most likely fault type. The process of more fault database data requires a long search time; the fault classification is to use A classifier is constructed from the fault database data, and the current fault data data classifier is directly given the fault type.

However, the inventor found that the production process of the modern process industry not only continues to operate under a stable working condition, but also may constantly change the operating conditions and switch between different working conditions according to the influence of raw materials, products, markets, environments and other factors for steady state operation. And because the operating parameters of different working conditions are different, the data distribution of variables is also different. However, the current fault detection and diagnosis models require that the data distribution of the modeling process and the online monitoring process be the same. Therefore, the above scheme can only build a model and application for each working condition separately, and cannot perform a unified combination for multiple working conditions. Fault detection and diagnosis lead to low monitoring efficiency and high monitoring cost in the actual monitoring process.

Contents of the invention

For this reason, the first purpose of this disclosure is to propose a method for detecting and diagnosing industrial faults in multi-working-condition processes based on deep transfer learning, which uses variational autoencoders for fault detection modeling and convolutional neural networks for fault detection Diagnostic modeling, using the deep transfer learning method to carry out joint training and modeling of process industry data in multiple working conditions, by reducing the distance between the data features of different working conditions, so that the feature distribution extracted from the data in multiple working conditions is as similar as possible, A common fault detection and diagnosis model for multiple working conditions is constructed, which improves the monitoring efficiency of the multi-working condition process industry and reduces the monitoring cost.

The second purpose of the present disclosure is to propose a multi-working-condition process industrial fault detection and diagnosis system based on deep transfer learning.

A third object of the present disclosure is to propose a non-transitory computer-readable storage medium.

A fourth object of the present disclosure is to provide an electronic device.

A fifth object of the present disclosure is to provide a computer program product.

A sixth object of the present disclosure is to propose a computer program.

In order to achieve the above purpose, the embodiment of the first aspect of the present disclosure proposes a method for detecting and diagnosing industrial faults in a multi-working-condition process based on deep transfer learning, which includes the following steps:

Obtain the historical data of the production system of the process industry under multiple working conditions, the historical data includes normal operation data and fault data, and mark the fault data to build a fault database;

Standardize the historical data of each working condition and organize them into a two-dimensional matrix;

Designing a variational autoencoder as a fault detection model according to the two-dimensional matrix, and designing a convolutional neural network as a fault diagnosis model;

The normal operation data of each working condition after sorting is divided into a training set and a test set, and based on the training set, the fault detection model is trained in combination with the maximum mean difference MMD of deep transfer learning, and based on the test set calculating a detection threshold of the fault detection model;

Based on the collated fault data, the fault diagnosis model is trained in combination with the maximum mean difference MMD of deep transfer learning;

Collect real-time data in the process industry, standardize the real-time data and organize it into the size of the two-dimensional matrix, and input the two-dimensional matrix corresponding to the real-time data into the trained fault detection model to calculate the loss function value, comparing the loss function value with the detection threshold to determine whether the production system is abnormal;

If an abnormality occurs, the two-dimensional matrix corresponding to the real-time data is input into the trained fault diagnosis model, and the fault type of the production system is determined through fault classification.

In one embodiment of the present disclosure, standardizing the historical data of each working condition includes: calculating the average value and standard deviation of the normal operation data of each working condition, and using the average value and the The above standard deviation is used as the z-score normalization parameter; the z-score normalization is performed on the historical data of each working condition by the following formula:

Among them, σ is the mean value, μ is the standard deviation, and x is the individual data to be standardized.

In one embodiment of the present disclosure, based on the training set, the fault detection model is trained in combination with the maximum mean difference (MMD) of deep transfer learning, including: calculating the Loss function values for each 2D matrix in the training set:

Among them, x represents the input data, z represents the reconstructed data calculated by the fault detection model, x _s represents the data of the first working condition, x _t represents the data of the second working condition, k(·) represents the Gaussian kernel function; through gradient descent The method trains the fault detection model until the average loss function value of the training set converges.

In one embodiment of the present disclosure, calculating the detection threshold of the fault detection model based on the test set includes: calculating each two-dimensional The loss function value of the matrix; based on the loss function value of each two-dimensional matrix data in the test set, the confidence level of the preset value is taken as the detection threshold of the model through kernel density estimation.

In one embodiment of the present disclosure, based on the collated fault data, the fault diagnosis model is trained in combination with the maximum mean difference MMD of deep transfer learning, including: through the collated fault data and the collated fault data In the corresponding fault type label in the fault storehouse, calculate the classification accuracy rate of the classification result that described fault diagnosis model outputs; The training loss function of convolutional neural network is minimized by following formula:

Among them, y represents the real category of the fault data, z represents the category probability output by the fault diagnosis model, x _s represents the data of the first working condition, x _t represents the data of the second working condition, and k(·) represents the Gaussian kernel function; The gradient descent method trains the fault diagnosis model until the average classification accuracy of the sorted fault data converges.

In one embodiment of the present disclosure, after determining the fault type of the production system, it further includes: determining the fault type of the production system, if it is determined that the fault type of the production system is a fault in the fault database If there is a fault type other than the type label and expert knowledge, then update the fault database and the fault diagnosis model.

In order to achieve the above purpose, the embodiment of the second aspect of the present disclosure proposes a multi-working condition process industrial fault detection and diagnosis system based on deep transfer learning, including the following modules:

The acquisition module is used to acquire historical data of the production system of the process industry under multiple working conditions, the historical data includes normal operation data and fault data, and marks the fault data to build a fault database;

A labeling module, configured to standardize the historical data of each working condition and organize them into a two-dimensional matrix;

A design module, for designing a variational autoencoder as a fault detection model according to the two-dimensional matrix, and designing a convolutional neural network as a fault diagnosis model;

The first training module is used to divide the sorted normal operation data of each working condition into a training set and a test set, and based on the training set, combine the maximum mean difference MMD of deep transfer learning to train the fault detection model, calculating the detection threshold of the fault detection model based on the test set;

The second training module is used to train the fault diagnosis model based on the sorted fault data, combined with the maximum mean difference MMD of deep transfer learning;

The fault detection module is used to collect real-time data in the process industry, perform the standardized processing on the real-time data and organize it into the size of the two-dimensional matrix, and input the two-dimensional matrix corresponding to the real-time data into the training completion The fault detection model calculates a loss function value, and compares the loss function value with the detection threshold to determine whether an abnormality occurs in the production system;

The fault diagnosis module is configured to input the two-dimensional matrix corresponding to the real-time data into the trained fault diagnosis model if an abnormality occurs, and determine the fault type of the production system through fault classification.

In an embodiment of the present disclosure, the labeling module is specifically configured to: calculate the average value and standard deviation of the normal operation data of each working condition, and use the average value and the standard deviation as z-score normalization parameters ; Normalize the z-score of the historical data of each working condition by the following formula:

Among them, σ is the mean value, μ is the standard deviation, and x is the individual data to be standardized.

In one embodiment of the present disclosure, the first training module is specifically configured to: calculate the loss function value of each two-dimensional matrix in the training set through the following formula of the training loss function of the variational autoencoder:

Among them, x represents the input data, z represents the reconstructed data calculated by the fault detection model, x _s represents the data of the first working condition, x _t represents the data of the second working condition, k(·) represents the Gaussian kernel function; through gradient descent The method trains the fault detection model until the average loss function value of the training set converges.

In one embodiment of the present disclosure, the first training module is also used for:

Calculate the loss function value of each two-dimensional matrix in the test set by the formula of the training loss function of the variational autoencoder;

Based on the loss function value of each two-dimensional matrix data in the test set, the confidence level of a preset value is taken as the detection threshold of the model through kernel density estimation.

In one embodiment of the present disclosure, the second training module is used for:

Calculate the classification accuracy rate of the classification result output by the fault diagnosis model through the sorted fault data and the corresponding fault type labels of the sorted fault data in the fault database;

The training loss function of the convolutional neural network is minimized by the following formula:

Among them, y represents the true category of the fault data, z represents the category probability output by the fault diagnosis model, x _s represents the data of the first working condition, x _t represents the data of the second working condition, and k(·) represents the Gaussian kernel function;

The fault diagnosis model is trained by a gradient descent method until the average classification accuracy of the collated fault data converges.

In an embodiment of the present disclosure, the fault diagnosis module is also used to determine the fault type of the production system, if it is determined that the fault type of the production system is a fault type other than the fault type label in the fault library and expert knowledge, Then update the fault library and fault diagnosis model.

The technical solutions provided by the embodiments of the present disclosure bring at least the following beneficial effects: the embodiments of the present disclosure use variational autoencoders for fault detection modeling, use convolutional neural networks for fault diagnosis modeling, and use deep transfer learning methods to Process industrial data of multiple working conditions is jointly trained and modeled, and the distance between the features of different working condition data in the neural network layer is reduced through deep transfer learning, so that the feature distribution extracted from the multi-working condition data is as similar as possible, so that multiple working conditions can be constructed. The general-purpose fault detection and diagnosis model for working conditions is more suitable for the monitoring of multi-working-condition process industrial processes, avoiding separate modeling for each working condition, and performing unified and joint fault detection and diagnosis for multiple working conditions, thereby improving multi-working conditions. It improves the monitoring efficiency of the process industry and reduces the monitoring cost.

In order to realize the above-mentioned embodiments, the embodiment of the third aspect of the present disclosure proposes a non-transitory computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the deep transfer learning based on the above-mentioned embodiments is implemented Multi-working condition process industrial fault detection and diagnosis method.

In order to implement the above embodiments, the fourth aspect of the present disclosure provides an electronic device, including: at least one processor; at least one memory storing computer-executable instructions, wherein the computer-executable instructions are executed by the at least When one processor is running, the at least one processor is made to execute the method for detecting and diagnosing industrial faults in a multi-working-condition process based on deep transfer learning as in the above-mentioned embodiments.

In order to implement the above-mentioned embodiments, the embodiment of the fifth aspect of the present disclosure proposes a computer program product, including computer instructions, and when the computer instructions are executed by at least one processor, the multiplexing based on deep transfer learning as in the above-mentioned embodiments is implemented. Fault detection and diagnosis methods in process industry.

In order to realize the above-mentioned embodiments, the embodiment of the sixth aspect of the present disclosure proposes a computer program, the computer program includes computer program code, when the computer program code is run on the computer, it makes the computer perform the above-mentioned embodiment. Multi-condition process industrial fault detection and diagnosis method based on deep transfer learning.

Additional aspects and advantages of the disclosure will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the disclosure.

Description of drawings

The above and/or additional aspects and advantages of the present disclosure will become apparent and understandable from the following description of the embodiments in conjunction with the accompanying drawings, wherein:

FIG. 1 is a flow chart of a method for detecting and diagnosing industrial faults in a multi-working-condition process based on deep transfer learning proposed by an embodiment of the present disclosure;

FIG. 2 is a schematic structural diagram of a specific fault detection model proposed by an embodiment of the present disclosure;

FIG. 3 is a schematic structural diagram of a specific fault diagnosis model proposed by an embodiment of the present disclosure;

FIG. 4 is a schematic flow diagram of a specific deep transfer learning-based method for detecting and diagnosing industrial faults in a multi-working condition process proposed by an embodiment of the present disclosure;

FIG. 5 is a schematic structural diagram of a multi-working-condition process industrial fault detection and diagnosis system based on deep transfer learning proposed by an embodiment of the present disclosure.

Detailed ways

Embodiments of the present disclosure are described in detail below, examples of which are illustrated in the drawings, in which the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the figures are exemplary and are intended to explain the present disclosure and should not be construed as limiting the present disclosure.

A method and system for detecting and diagnosing industrial faults in a multi-working-condition process based on deep transfer learning proposed by an embodiment of the present disclosure will be described below with reference to the accompanying drawings.

Fig. 1 is a flow chart of a method for detecting and diagnosing industrial faults in a multi-working-condition process based on deep transfer learning proposed by an embodiment of the present disclosure. As shown in Fig. 1 , the method includes the following steps: S101-S107.

Step S101, acquiring historical data of the production system of the process industry under multiple working conditions, wherein the historical data includes normal operation data and fault data, and marking the fault data to build a fault database.

Among them, the process industry refers to the industry based on the production through physical and/or chemical changes. The process industry in the embodiments of the present disclosure may include the production industries such as petroleum and chemical industry. The multiple working conditions are the process industries under different operating conditions. working conditions.

In one embodiment of the present disclosure, when obtaining the historical data of the production system of the process industry under multiple working conditions, the pre-stored production system under different working conditions can be read from the preset database of the production device of the process industry Under the historical data, the obtained historical data includes the data of the system in the normal operation state and the data when the fault occurs. Since the normal operation state and the fault state will last for a certain period of time, the historical data obtained in this disclosure has Process data for a certain length of time.

Further, after collecting normal operation data and fault data under multiple working conditions, mark the fault data to determine the fault type of the fault data, so as to obtain the training data for subsequent training of the fault diagnosis model. , experts can mark each fault by manual marking. Then build a fault database based on the marked fault data, and then store labels of different fault types in the fault database.

In step S102, the historical data of each working condition is standardized and organized into a two-dimensional matrix.

Among them, standardization processing is the processing of converting data of different magnitudes into data values of unified measurement. Through standardization processing, data standards under different working conditions are unified, thereby improving data comparability, weakening data interpretation, and facilitating follow-up Do joint modeling.

During specific implementation, an appropriate standardization processing method can be selected according to actual needs. In the embodiment of the present disclosure, the Z-Score standardization method can be used for standardization processing. Specifically, first calculate the average value of the normal operation data of each working condition and standard deviation, using the mean value and standard deviation as z-score normalization parameters, and then performing z-score normalization on the historical data of each working condition by the following formula:

Wherein, σ is the average value, μ is the standard deviation, and x is the individual data to be standardized, which may be the historical data under a single working condition in the embodiment of the present disclosure. That is to say, the normal operation data and fault data of all working conditions are standardized by z-score according to the above formula.

Further, the normalized historical data is sorted into multi-variable time window data, that is, a two-dimensional matrix of corresponding size. The size of the two-dimensional matrix can be set according to actual needs. In the embodiments of the present disclosure, since the collected historical data is process data and corresponds to different working conditions, the multivariate time-series data is sorted into a two-dimensional matrix.

In the embodiment of the present disclosure, the multi-variable time-series data is organized into a two-dimensional matrix of m×t size, where m represents the number of variables, t represents the length of the time window, and the number of variables refers to the number of variables that need to be monitored under each working condition. The number of parameters of the production system, for example, m is the number of all variables such as temperature, pressure, and input and output currents of the production system that need to be monitored. And t is the preset length of the time window selected from the time series data, for example, t can be set to 10 to 60 minutes, that is, the data of 10 to 60 minutes is intercepted from the historical data. Organize into a two-dimensional matrix according to the required size of m and t.

It should be noted that, in the embodiment of the present disclosure, multi-variable time-series data with a working condition may be organized into multiple two-dimensional matrices. For example, if the collected historical data is 60 minutes, when t is 10 minutes, then The data can be organized into six two-dimensional matrices. Therefore, the embodiment of the present disclosure can organize the standardized processing history data of each working condition into multiple two-dimensional matrices.

Step S103, designing a variational autoencoder as a fault detection model according to the two-dimensional matrix, and designing a convolutional neural network as a fault diagnosis model.

Among them, the variational autoencoder can generate hidden vectors containing data information through the model itself, and can generate new data by reconstructing the data. By measuring the difference between the input data and the reconstructed data, it can be judged whether the input data is in normal operation. state. In the embodiment of the present disclosure, a variational autoencoder is designed according to the generated two-dimensional matrix as a fault detection model.

In one embodiment of the present disclosure, as shown in FIG. 2 , the variational autoencoder designed in the present disclosure includes an encoder 10 and a decoder 20 , and input Data, the reconstructed data is output at the output end, the encoder mines potential features according to the mean and standard deviation calculated above, and the reconstructed data is output through the decoder. Among them, the size of m×t is designed according to the two-dimensional matrix, the format size of the input data and output reconstruction data of the variational autoencoder is designed, and the size of the data operated by the variational autoencoder is determined according to the size of the two-dimensional matrix design .

Further, a convolutional neural network is designed as a fault diagnosis model. The convolutional neural network designed in the embodiment of the present disclosure is shown in Figure 3. The last layer of the convolutional neural network is composed of a softmax function for performing classification tasks. Wherein, the number of categories that can be classified by the softmax function is the same as the number of fault types in the fault library generated in step S101.

Step S104, divide the sorted normal operation data of each working condition into a training set and a test set, and based on the training set, combine the maximum mean difference MMD of deep transfer learning to train the fault detection model, and calculate the fault detection model based on the test set detection threshold.

Among them, deep transfer learning relaxes the assumption that the training data must be independent and identically distributed with the test data, and in the embodiments of the present disclosure, the problem of insufficient training data can be solved by transfer learning. The maximum mean discrepancy (MMD) of deep transfer learning can reduce the distance of different working condition data in the model feature space, that is, the loss function of model training is constructed through MMD, so that the extracted feature distribution is as similar as possible .

Wherein, the detection threshold is used to subsequently compare the acquired real-time data with it to determine whether the real-time data exceeds the detection threshold, and then determine whether the current system is abnormal.

Specifically, first divide the historical data standardized and sorted into a two-dimensional matrix into a training set and a test set according to a preset ratio, for example, divide it into a training set and a test set according to 4:1, and the training set is used for training fault detection model, the test set is used to calculate the detection threshold, and the fault data of all working conditions after normalization and sorting are all used to train the fault diagnosis model.

In one embodiment of the present disclosure, based on the training set, the fault detection model is trained in combination with the maximum mean difference MMD of deep transfer learning, which may include first calculating each two-dimensional matrix in the training set through the formula of the training loss function of the variational autoencoder The loss function value of , where the formula of the training loss function of the variational autoencoder is as follows:

Among them, x represents the input data, z represents the reconstructed data calculated by the fault detection model, x _s represents the data of the first working condition, x _t represents the data of the second working condition, k(·) represents the Gaussian kernel function, and the first working condition The condition data and the second working condition data are data of any two groups of different working conditions among the collected multiple working conditions. Then the fault detection model is trained by the gradient descent method until the average loss function value of the training set converges.

In the embodiment of the present disclosure, specifically, using the training set obtained from normal operation data, the loss function is minimized according to the above formula, and the fault detection model is trained using the gradient descent method. When the average loss function value of the training set gradually converges and no longer decreases After that, stop the model training.

Further, in the embodiment of the present disclosure, the detection threshold of the fault detection model is calculated based on the test set, and the loss function value of each two-dimensional matrix in the test set can be calculated through the formula of the training loss function of the above-mentioned variational autoencoder , and then based on the loss function value of each two-dimensional matrix data in the test set, the confidence level of the preset value is taken as the detection threshold of the model through kernel density estimation. Specifically, using the test set obtained from normal operation data, the loss function value of each two-dimensional matrix data is calculated according to the above formula, and the 99.9% confidence level is taken as the detection threshold of the model according to the kernel density estimation method. Of course, the confidence level The preset value of can be set according to actual needs, for example, a confidence level of 99% to 99.99% can also be used as the detection threshold.

It should be noted that in the above formula of the training loss function of the variational autoencoder, the first item ||xi ^{-z i} || represents the reconstruction loss of the variational autoencoder, x represents the input data, and z represents The reconstructed data calculated by the model; the second term

Represents the KL divergence of the variational autoencoder, which is used to measure the difference between the feature space distribution and the standard normal distribution; the third item

Indicates the maximum mean difference of MMD in the deep transfer learning method, and x _s and x _t respectively represent the data of two working conditions. During the training process of the deep transfer learning method, the data of two different working conditions must be input at the same time.

Among them, i represents the number of calculations. In the embodiment of the present disclosure, when inputting the data of two working conditions from the acquired data of various working conditions, two kinds of working conditions can be sequentially selected from all working conditions by traversal to output the The formula can be used for calculation, or two can be randomly selected from all working conditions, and the number of selections can be selected to meet the calculation requirements. The specific selection method can be set according to actual needs, and there is no limit here.

Step S105, based on the sorted fault data, combined with the maximum mean difference MMD of deep transfer learning to train the fault diagnosis model.

In the embodiment of the present disclosure, based on the organized fault data, the fault diagnosis model is trained in combination with the maximum mean difference MMD of deep transfer learning, including the corresponding faults in the fault database through the sorted fault data and the sorted fault data Type labels, calculate the classification accuracy of the classification results output by the fault diagnosis model; and then minimize the training loss function of the convolutional neural network through the following formula of the training loss function of the convolutional neural network:

Among them, y represents the true category of the fault data, z represents the category probability output by the fault diagnosis model, x _s represents the data of the first working condition, x _t represents the data of the second working condition, k(·) represents the Gaussian kernel function; finally The fault diagnosis model is trained by the gradient descent method until the average classification accuracy of the collated fault data converges.

That is to say, the embodiment of the present disclosure uses the fault data and its corresponding fault type labels in the fault library, minimizes the loss function according to the above-mentioned training loss function formula of the convolutional neural network, and uses the gradient descent method to train the fault diagnosis model. When After the average classification accuracy of the fault data no longer increases, stop the model training.

It should be noted that in the formula of the training loss function of the convolutional neural network, the first item represents the classification loss of the convolutional neural network, y represents the true category of the fault data, and z represents the category probability of the model calculation output; the second term represents The MMD maximum mean difference in the deep transfer learning method, x _s and x _t represent the data of the two working conditions respectively. During the training process of the deep transfer learning method, the data of two different working conditions must be input at the same time. Among them, in the training The method of arbitrarily selecting two different working conditions in the convolutional neural network process can refer to the selection method when training the fault detection model in step 104, and will not be repeated here.

Thus, the fault detection model and fault diagnosis model are trained in the offline stage.

Step S106, collect real-time data in the process industry, standardize the real-time data and organize it into the size of a two-dimensional matrix, and input the two-dimensional matrix corresponding to the real-time data into the trained fault detection model to calculate the loss function value, and the loss The function value is compared with the detection threshold to determine whether an abnormality occurs in the production system.

Specifically, when performing real-time monitoring of a multi-working-condition process industry, real-time data is collected first. In the embodiment of the present disclosure, data can be read from the real-time database of the production system of the process industry. t, the process data whose variable dimension is m.

Further, the real-time data is standardized and sorted into the size of a two-dimensional matrix. Among them, the implementation method of standardization processing corresponds to the method of standardization processing in the offline training process, and the size of the two-dimensional matrix organized by real-time data corresponds to the size of the two-dimensional matrix organized by historical data during the offline training process. The specific organization method With reference to the descriptions in the above embodiments, the mean value and standard deviation of each variable in the current working condition can be used to standardize the z-score and arrange it into a two-dimensional matrix of m×t size.

Further, standardize the real-time data and organize the two-dimensional matrix into the fault detection model composed of the trained variational autoencoder, and calculate the loss function value according to the above formula of the training loss function of the variational autoencoder , and compare it with the preset detection threshold to determine whether the production system is abnormal.

In the embodiment of the present disclosure, if the loss function value of the current real-time data is greater than the detection threshold, it indicates that the current system is abnormal, and subsequent fault diagnosis needs to be performed; if the loss function value of the current real-time data is less than or equal to the detection threshold, it indicates that The current system is running normally and continues to collect real-time data.

Step S107, if an abnormality occurs, input the two-dimensional matrix corresponding to the real-time data into the trained fault diagnosis model, and determine the fault type of the production system through fault classification.

Specifically, after judging that an abnormality has occurred, input the two-dimensional matrix formed by the real-time data into the fault diagnosis model formed by the convolutional neural network, and classify and predict the input data through the fault diagnosis model to determine which type of fault has occurred in the current system , and output the judgment results to realize the detection and diagnosis of industrial faults in multi-working conditions.

In an embodiment of the present disclosure, after the fault diagnosis module outputs the diagnosis result, it can also be submitted to experts or operators, for example, the diagnosis result is sent to the corresponding mobile terminal of the system operator through the wireless network, assisting them to understand the current system status Monitor, judge and make decisions.

In other embodiments of the present disclosure, after determining the fault type of the production system, it also includes determining the fault type of the production system. If the fault type is not specified, then update the fault database and the fault diagnosis model. Specifically, if it is determined that the fault type given by the fault diagnosis model is different from the pre-established expert knowledge, or it is found to be a new fault that is not recorded in the fault type in the fault database, it is necessary to return to step S101 to update the fault database, Add the fault to the database, and update the fault database data and convolutional neural network fault diagnosis model. Therefore, the fault diagnosis model is updated according to new cases generated in the actual production process, and the comprehensiveness and accuracy of fault diagnosis are improved.

Therefore, in the online application stage, through the trained fault detection model and fault diagnosis model, it is possible to accurately detect which faults have occurred in the current system under various working conditions.

To sum up, the multi-working-condition industrial fault detection and diagnosis method based on deep transfer learning in the embodiment of the present disclosure uses variational autoencoder for fault detection modeling, convolutional neural network for fault diagnosis modeling, and uses The deep transfer learning method conducts joint training and modeling of multi-working-condition process industry data, and reduces the distance between the characteristics of different working-condition data in the neural network layer through deep transfer learning, so that the feature distribution extracted from multi-working-condition data is as similar as possible , so that a common fault detection and diagnosis model for multiple working conditions can be constructed, which is more suitable for the monitoring of multi-working process industrial processes, avoiding separate modeling for each working condition, and performing unified and joint fault detection and diagnosis for multiple working conditions. This improves the monitoring efficiency of the multi-working condition process industry and reduces the monitoring cost.

In order to more clearly illustrate the method for detecting and diagnosing industrial faults in a multi-working-condition process based on deep transfer learning according to an embodiment of the present disclosure, a specific embodiment will be described in detail below. Fig. 4 is a schematic flowchart of a specific deep transfer learning-based method for detecting and diagnosing industrial faults in a multi-working condition process proposed by an embodiment of the present disclosure. The method includes an offline modeling stage and a real-time detection stage.

As shown in Fig. 4, when performing offline modeling, steps S01-S07 are included, wherein step S01: constructing a data set. The normal operation data and fault data of multiple working conditions are collected from the historical database of a process industry production device, and each fault is marked by experts to determine the type of fault and build a fault database.

Step S02: Data standardization. Calculate the average σ and standard deviation μ of the normal operating data of each working condition, as the z-score normalization parameters. Then the normal operation data and fault data of all working conditions are standardized by z-score according to the following formula:

Step S03: Convert to multivariate time window data. Organize the multivariate time series data into a two-dimensional matrix of m×t size, m represents the number of variables, and t represents the length of the time window, which can be 10 to 60 minutes, so as to effectively capture the dynamic and multivariate characteristics of the process data.

Step S04: Design a fault detection and diagnosis model. According to the variable dimension m and the time window length t, a variational autoencoder is designed as a fault detection model. A convolutional neural network is designed as a fault diagnosis model, and the last layer is composed of a softmax function for classification tasks, and the number of categories is the same as the number of fault types in the fault library.

Step S05: Divide the data set. After normalization, the normal operation data of all working conditions are divided into training set and test set according to 4:1. The training set is used to train the fault detection model, and the test set is used to calculate the detection threshold; after normalization, all the fault data of all working conditions are used for Train a fault diagnosis model.

Step S06: Training the fault detection model. Use the training set obtained from the normal operation data, minimize the loss function according to the formula of the training loss function of the variational autoencoder, and use the gradient descent method to train the fault detection model. When the average loss function value of the training set gradually converges and no longer declines, Stop model training; use the test set obtained from the normal operation data, calculate the loss function value of each two-dimensional matrix data according to the formula of the training loss function of the variational autoencoder, and take the 99.9% confidence level as the kernel density estimation method. The detection threshold for the model.

Step S07: Training the fault diagnosis model. Using the fault data and its corresponding fault type labels in the fault library, the loss function is minimized according to the training loss function formula of the convolutional neural network, and the fault diagnosis model is trained by using the gradient descent method. When the average classification accuracy of the fault data is no longer After rising, stop model training.

The online real-time detection stage includes steps S08-S11.

Step S08: Collect real-time data. Read the process data with a time length of t and a variable dimension of m from the real-time database, use the average value and standard deviation of each variable in the current working condition to standardize the z-score, and organize it into a two-dimensional matrix of m×t size.

Step S09: Real-time fault detection. The real-time data is normalized and sorted into a two-dimensional matrix input into the fault detection model composed of the variational autoencoder, and the loss function value is calculated according to the training loss function formula of the variational autoencoder, and compared with the detection threshold. If the loss function value of the current real-time data is greater than the detection threshold, it indicates that the current system is abnormal, and step S10 needs to be performed for fault diagnosis to determine the fault type; if the loss function value of the current real-time data is less than or equal to the detection threshold, it indicates that the current system is running normally, and continue to Proceed to step S08 for data collection and detection.

Step S10: Real-time fault diagnosis: standardize the real-time data and organize the two-dimensional matrix into the fault diagnosis model formed by the convolutional neural network, and obtain the type of fault that has occurred in the current system through classification, and submit the result to the expert or Operators, assisting them to monitor, judge and make decisions on the current system status.

Step S11: If the fault type given by the fault diagnosis model is different from expert knowledge, or it is found to be a new fault, it is necessary to return to step S01 to update the fault database, and update the fault database data and the convolutional neural network fault diagnosis model.

Therefore, by constructing a common fault detection and diagnosis model for multiple working conditions, and performing unified and joint fault detection and diagnosis for multiple working conditions, this method is more suitable for the monitoring of industrial processes with multiple working conditions.

In order to realize the above-mentioned embodiments, the embodiment of the present disclosure also proposes a multi-working-condition process industrial fault detection and diagnosis system based on deep transfer learning, and FIG. 5 shows a multi-working-condition process based on deep transfer learning proposed by the embodiment of the present disclosure The structural diagram of industrial fault detection and diagnosis system, as shown in Figure 5, the system includes acquisition module 100, labeling module 200, design module 300, first training module 400, second training module 500, fault detection module 600 and fault diagnosis module 700.

The acquisition module 100 is used to acquire historical data of the production system of the process industry under multiple working conditions, the historical data includes normal operation data and fault data, and marks the fault data to build a fault database.

The labeling module 200 is used to standardize the historical data of each working condition and organize them into a two-dimensional matrix.

The design module 300 is used to design a variational autoencoder as a fault detection model according to a two-dimensional matrix, and design a convolutional neural network as a fault diagnosis model.

The first training module 400 is used to divide the sorted normal operation data of each working condition into a training set and a test set, and based on the training set, combined with the maximum mean difference MMD of deep transfer learning to train the fault detection model, based on the test set Compute the detection threshold for the fault detection model.

The second training module 500 is used to train the fault diagnosis model based on the sorted fault data, combined with the maximum mean difference MMD of deep transfer learning;

The fault detection module 600 is used to collect real-time data in the process industry, standardize the real-time data and organize it into the size of a two-dimensional matrix, and input the two-dimensional matrix corresponding to the real-time data into the trained fault detection model to calculate the loss function Value, compare the loss function value with the detection threshold to determine whether the production system is abnormal;

The fault diagnosis module 700 is configured to input the two-dimensional matrix corresponding to the real-time data into the trained fault diagnosis model if an abnormality occurs, and determine the fault type of the production system through fault classification.

In an embodiment of the present disclosure, the labeling module 200 is specifically configured to: calculate the average value and standard deviation of the normal operation data of each working condition, and use the average value and standard deviation as z-score normalization parameters; use the following formula to The historical data of each working condition is standardized by z-score:

Among them, σ is the mean value, μ is the standard deviation, and x is the individual data to be standardized.

In an embodiment of the present disclosure, the first training module 400 is specifically configured to: calculate the loss function value of each two-dimensional matrix in the training set through the following formula of the training loss function of the variational autoencoder:

Among them, x represents the input data, z represents the reconstructed data calculated by the fault detection model, x _s represents the data of the first working condition, x _t represents the data of the second working condition, k(·) represents the Gaussian kernel function; through gradient descent The method trains the fault detection model until the average loss function value of the training set converges.

In an embodiment of the present disclosure, the first training module 400 is also used to: calculate the loss function value of each two-dimensional matrix in the test set through the formula of the training loss function of the variational autoencoder; The loss function value of the dimensional matrix data, and the confidence level of the preset value is taken as the detection threshold of the model through kernel density estimation.

In an embodiment of the present disclosure, the second training module 500 is specifically configured to: calculate the classification of the classification results output by the fault diagnosis model through the sorted fault data and the corresponding fault type labels of the sorted fault data in the fault database Accuracy; the training loss function of the convolutional neural network is minimized by the following formula:

Among them, y represents the real category of the fault data, z represents the category probability output by the fault diagnosis model, x _s represents the data of the first working condition, x _t represents the data of the second working condition, and k(·) represents the Gaussian kernel function; The gradient descent method trains the fault diagnosis model until the average classification accuracy of the sorted fault data converges.

In one embodiment of the present disclosure, the fault diagnosis module 700 is also used to determine the fault type of the production system. If it is determined that the fault type of the production system is a fault type other than the fault type label in the fault library and expert knowledge, Then update the fault library and fault diagnosis model.

To sum up, the multi-working-condition process industrial fault detection and diagnosis system based on deep transfer learning in the embodiment of the present disclosure uses a variational autoencoder for fault detection modeling, a convolutional neural network for fault diagnosis modeling, and uses The deep transfer learning method conducts joint training and modeling of multi-working-condition process industry data, and reduces the distance between the characteristics of different working-condition data in the neural network layer through deep transfer learning, so that the feature distribution extracted from multi-working-condition data is as similar as possible , so that a common fault detection and diagnosis model for multiple working conditions can be constructed, which is more suitable for the monitoring of multi-working process industrial processes, avoiding separate modeling for each working condition, and performing unified and joint fault detection and diagnosis for multiple working conditions. This improves the monitoring efficiency of the multi-working condition process industry and reduces the monitoring cost.

In order to realize the above-mentioned embodiments, the embodiments of the present disclosure also propose a non-transitory computer-readable storage medium, on which a computer program is stored. Multi-condition process industrial fault detection and diagnosis method based on deep transfer learning.

In order to achieve the above embodiments, an embodiment of the present disclosure also proposes an electronic device, including: at least one processor; at least one memory storing computer-executable instructions, wherein the computer-executable instructions are executed by the at least one processor , making the at least one processor execute the method for detecting and diagnosing industrial faults in multi-working-condition process industries based on deep transfer learning as described in any one of the above-mentioned embodiments.

In order to realize the above-mentioned embodiments, the embodiments of the present disclosure also propose a computer program product, including computer instructions. When the computer instructions are executed by at least one processor, the deep transfer learning-based Fault detection and diagnosis method for multi-working condition process industry.

In order to realize the above-mentioned embodiments, the embodiments of the present disclosure also propose a computer program, the computer program includes computer program code, when the computer program code is run on the computer, the computer executes the computer program described in any one of the above-mentioned embodiments. Multi-condition process industrial fault detection and diagnosis method based on deep transfer learning.

It should be noted that the foregoing explanations of the embodiment of the multi-condition process industrial fault detection and diagnosis method based on deep transfer learning are also applicable to the deep transfer learning-based multi-condition process industrial fault detection and diagnosis system of the embodiment of the present disclosure, Non-transitory computer-readable storage media, electronic devices, computer program products, and computer programs will not be described in detail here.

In the description of this specification, descriptions referring to the terms "one embodiment", "some embodiments", "example", "specific examples", or "some examples" mean that specific features described in connection with the embodiment or example , structure, material or characteristic is included in at least one embodiment or example of the present disclosure. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Furthermore, the described specific features, structures, materials or characteristics may be combined in any suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples and features of different embodiments or examples described in this specification without conflicting with each other.

In addition, the terms "first" and "second" are used for descriptive purposes only, and cannot be interpreted as indicating or implying relative importance or implicitly specifying the quantity of indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include at least one of these features. In the description of the present disclosure, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

Any process or method descriptions in flowcharts or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing custom logical functions or steps of a process , and the scope of preferred embodiments of the present disclosure includes additional implementations in which functions may be performed out of the order shown or discussed, including substantially concurrently or in reverse order depending on the functions involved, which shall It is understood by those skilled in the art to which the embodiments of the present disclosure pertain.

The logic and/or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced listing of executable instructions for implementing logical functions, can be embodied in any computer-readable medium, For use with instruction execution systems, devices, or devices (such as computer-based systems, systems including processors, or other systems that can fetch instructions from instruction execution systems, devices, or devices and execute instructions), or in conjunction with these instruction execution systems, devices or equipment used. For the purposes of this specification, a "computer-readable medium" may be any device that can contain, store, communicate, propagate or transmit a program for use in or in conjunction with an instruction execution system, device or device. More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connection with one or more wires (electronic device), portable computer disk case (magnetic device), random access memory (RAM), Read Only Memory (ROM), Erasable and Editable Read Only Memory (EPROM or Flash Memory), Fiber Optic Devices, and Portable Compact Disc Read Only Memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program can be printed, since the program can be read, for example, by optically scanning the paper or other medium, followed by editing, interpretation or other suitable processing if necessary. The program is processed electronically and stored in computer memory.

It should be understood that various parts of the present disclosure may be implemented in hardware, software, firmware or a combination thereof. In the embodiments described above, various steps or methods may be implemented by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented by any one or a combination of the following techniques known in the art: a discrete Logic circuits, ASICs with suitable combinational logic gates, Programmable Gate Arrays (PGA), Field Programmable Gate Arrays (FPGA), etc.

Those of ordinary skill in the art can understand that all or part of the steps carried by the methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. During execution, one or a combination of the steps of the method embodiments is included.

In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing module, each unit may exist separately physically, or two or more units may be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software function modules. If the integrated modules are realized in the form of software function modules and sold or used as independent products, they can also be stored in a computer-readable storage medium.

The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, and the like. Although the embodiments of the present disclosure have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limitations on the present disclosure, and those skilled in the art can understand the above-mentioned embodiments within the scope of the present disclosure. The embodiments are subject to changes, modifications, substitutions and variations.

Claims (16)

1. A method for detecting and diagnosing industrial faults in a multi-working-condition process based on deep transfer learning, characterized in that it includes the following steps:

   Obtain the historical data of the production system of the process industry under multiple working conditions, the historical data includes normal operation data and fault data, and mark the fault data to build a fault database;

   Standardize the historical data of each working condition and organize them into a two-dimensional matrix;

   Designing a variational autoencoder as a fault detection model according to the two-dimensional matrix, and designing a convolutional neural network as a fault diagnosis model;

   The normal operation data of each working condition after sorting is divided into a training set and a test set, and based on the training set, the fault detection model is trained in combination with the maximum mean difference MMD of deep transfer learning, and based on the test set calculating a detection threshold of the fault detection model;

   Based on the collated fault data, the fault diagnosis model is trained in combination with the maximum mean difference MMD of deep transfer learning;

   Collect real-time data in the process industry, standardize the real-time data and organize it into the size of the two-dimensional matrix, and input the two-dimensional matrix corresponding to the real-time data into the trained fault detection model to calculate the loss function value, comparing the loss function value with the detection threshold to determine whether the production system is abnormal;

   If an abnormality occurs, the two-dimensional matrix corresponding to the real-time data is input into the trained fault diagnosis model, and the fault type of the production system is determined through fault classification.
2. The method according to claim 1, wherein said standardizing the historical data of each working condition includes:

   Calculate the mean value and standard deviation of the normal operation data of each working condition, and use the mean value and the standard deviation as z-score normalization parameters;

   The z-score normalization is performed on the historical data of each working condition by the following formula:

   

   Among them, σ is the mean value, μ is the standard deviation, and x is the individual data to be standardized.
3. The method according to claim 1 or 2, wherein, based on the training set, the fault detection model is trained in combination with the maximum mean difference MMD of deep transfer learning, comprising:

   The loss function value of each two-dimensional matrix in the training set is calculated by the following formula of the training loss function of the variational autoencoder:

   

   Among them, x represents the input data, z represents the reconstructed data calculated by the fault detection model, x _s represents the data of the first working condition, x _t represents the data of the second working condition, and k(·) represents the Gaussian kernel function;

   The fault detection model is trained by a gradient descent method until the average loss function value of the training set converges.
4. The method according to any one of claims 1 to 3, wherein the calculation of the detection threshold of the fault detection model based on the test set includes:

   Calculate the loss function value of each two-dimensional matrix in the test set by the formula of the training loss function of the variational autoencoder;

   Based on the loss function value of each two-dimensional matrix data in the test set, the confidence level of a preset value is taken as the detection threshold of the model through kernel density estimation.
5. The method according to any one of claims 1 to 4, wherein the fault diagnosis model is trained based on the fault data after sorting, combined with the maximum mean difference MMD of deep transfer learning, including:

   Calculate the classification accuracy rate of the classification result output by the fault diagnosis model through the sorted fault data and the corresponding fault type labels of the sorted fault data in the fault database;

   The training loss function of the convolutional neural network is minimized by the following formula:

   

   Among them, y represents the true category of the fault data, z represents the category probability output by the fault diagnosis model, x _s represents the data of the first working condition, x _t represents the data of the second working condition, and k(·) represents the Gaussian kernel function;

   The fault diagnosis model is trained by a gradient descent method until the average classification accuracy of the collated fault data converges.
6. The method according to any one of claims 1 to 5, after the determination of the failure type of the production system, further comprising:

   Determining the fault type of the production system, if it is determined that the fault type of the production system is a fault type other than the fault type label and expert knowledge in the fault database, then updating the fault database and the fault diagnosis Model.
7. A multi-working condition process industrial fault detection and diagnosis system based on deep transfer learning, characterized in that it includes:

   The acquisition module is used to acquire historical data of the production system of the process industry under multiple working conditions, the historical data includes normal operation data and fault data, and marks the fault data to build a fault database;

   A labeling module, configured to standardize the historical data of each working condition and organize them into a two-dimensional matrix;

   A design module, for designing a variational autoencoder as a fault detection model according to the two-dimensional matrix, and designing a convolutional neural network as a fault diagnosis model;

   The first training module is used to divide the sorted normal operation data of each working condition into a training set and a test set, and based on the training set, combine the maximum mean difference MMD of deep transfer learning to train the fault detection model, calculating the detection threshold of the fault detection model based on the test set;

   The second training module is used to train the fault diagnosis model based on the sorted fault data, combined with the maximum mean difference MMD of deep transfer learning;

   The fault detection module is used to collect real-time data in the process industry, perform the standardized processing on the real-time data and organize it into the size of the two-dimensional matrix, and input the two-dimensional matrix corresponding to the real-time data into the training completion The fault detection model calculates a loss function value, and compares the loss function value with the detection threshold to determine whether an abnormality occurs in the production system;

   The fault diagnosis module is configured to input the two-dimensional matrix corresponding to the real-time data into the trained fault diagnosis model if an abnormality occurs, and determine the fault type of the production system through fault classification.
8. The system according to claim 7, wherein the labeling module is specifically used for:

   Calculate the mean value and standard deviation of the normal operation data of each working condition, and use the mean value and the standard deviation as z-score normalization parameters;

   The z-score normalization is performed on the historical data of each working condition by the following formula:

   

   Among them, σ is the mean value, μ is the standard deviation, and x is the individual data to be standardized.
9. The system according to claim 7 or 8, wherein the first training module is specifically used for:

   The loss function value of each two-dimensional matrix in the training set is calculated by the following formula of the training loss function of the variational autoencoder:

   

   Among them, x represents the input data, z represents the reconstructed data calculated by the fault detection model, x _s represents the data of the first working condition, x _t represents the data of the second working condition, and k(·) represents the Gaussian kernel function;

   The fault detection model is trained by a gradient descent method until the average loss function value of the training set converges.
10. The system according to any one of claims 7 to 9, wherein the first training module is also used for:

    Calculate the loss function value of each two-dimensional matrix in the test set by the formula of the training loss function of the variational autoencoder;

    Based on the loss function value of each two-dimensional matrix data in the test set, the confidence level of a preset value is taken as the detection threshold of the model through kernel density estimation.
11. The system according to any one of claims 7 to 10, wherein the second training module is used for:

    Calculate the classification accuracy rate of the classification result output by the fault diagnosis model through the sorted fault data and the corresponding fault type labels of the sorted fault data in the fault database;

    The training loss function of the convolutional neural network is minimized by the following formula:

    

    Among them, y represents the true category of the fault data, z represents the category probability output by the fault diagnosis model, x _s represents the data of the first working condition, x _t represents the data of the second working condition, and k(·) represents the Gaussian kernel function;

    The fault diagnosis model is trained by a gradient descent method until the average classification accuracy of the collated fault data converges.
12. The system according to any one of claims 7 to 11, wherein the fault diagnosis module is also used to determine the fault type of the production system, if it is determined that the fault type of the production system is the fault type in the fault library For fault types other than labels and expert knowledge, the fault library and fault diagnosis model are updated.
13. A non-transitory computer-readable storage medium on which a computer program is stored, wherein when the computer program is executed by a processor, the deep migration learning-based Fault detection and diagnosis method for multi-working condition process industry.
14. An electronic device, characterized in that it comprises:

    at least one processor;

    at least one memory storing computer-executable instructions,

    Wherein, when the computer-executable instructions are executed by the at least one processor, the at least one processor executes the multi-working-condition process based on deep transfer learning according to any one of claims 1 to 6 Industrial fault detection and diagnosis methods.
15. A computer program product, comprising computer instructions, characterized in that, when the computer instructions are executed by at least one processor, the multi-working-condition process industry based on deep transfer learning according to any one of claims 1 to 6 is realized Fault detection and diagnosis methods.
16. A kind of computer program, it is characterized in that, described computer program comprises computer program code, when described computer program code runs on computer, makes computer carry out as described in any one of claims 1 to 6 based on deep transfer learning Multi-working condition process industrial fault detection and diagnosis method.

Images