Industrial Fault Diagnosis Method and System Based on Prior Knowledge

Abstract

The present invention provides an industrial fault diagnosis method and system based on prior knowledge, including: Step 1: Use a signed directed graph for modeling to construct a topological prior graph of industrial knowledge; Step 2: Extract local features in the graph structure of the input data through a graph convolutional network; Step 3: Construct a statistic for fault diagnosis; Step 4: Based on the graph convolutional network, use a graph convolutional long short-term memory recurrent neural network to learn the dynamic time characteristics in industrial data for fault prediction; Step 5: Combine the graph convolution and the long short-term memory network, and encapsulate them together with the SDG modeling into a microservice device to implement fault diagnosis and fault prediction of the industrial system in the form of microservices. The flexible application of the microservices of the present invention makes the method flexible to use and has a wide range of applications.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic information engineering, and specifically, to an industrial fault diagnosis method and system based on prior knowledge. Background Art

[0002] In the history of industrial production, there have been numerous tragic accidents caused by equipment failures, resulting in immeasurable losses. Therefore, an industrial fault diagnosis system is very important, which helps to avoid major risks caused by failures in the industrial process and has great significance for the safety and stability of the system. The various components of complex industrial systems and the various variables during system operation are often interrelated and tightly coupled. When performing fault prediction, sometimes it is necessary to analyze a specific variable or some variables. Some prediction methods based on basic neural networks often only focus on the temporal characteristics of historical data, thus ignoring the potential topological information between variables in the input data. Traditional fault diagnosis methods based on causal relationships, such as graph theory, fault trees, qualitative physical analysis, etc., often focus on the associations between variables and focus on analyzing the information of the internal topological structure of the system. However, such qualitative inference methods often ignore the information contained in the data itself, so it is difficult to achieve quantitative analysis. Therefore, it is necessary to grasp both the internal associations between variables in the data and the temporal characteristics of the data sequence during prediction.

[0003] Patent document CN111898872A (application number: 202010658415.3) discloses a power grid fault diagnosis method based on a microservice architecture, belonging to the technical field of electronic information engineering. The method includes: an enterprise middle platform with a microservice architecture generates power grid inspection task information, and sends the task information to the PC command center. The PC command center sends the task information to the inspection robot. The inspection robot inspects the power grid and sends the inspection information to the PC command center. The command center analyzes and judges whether a fault occurs and the location where the fault occurs.

[0004] The original data obtained during the production process is the key to fault diagnosis. Traditional industrial data is often limited to a single department or a single workshop, resulting in isolated information and forming information islands. With the rise of intelligent manufacturing and the advent of the industrial 4.0 era, data is the most precious wealth, and data can be fully utilized to create new value. Fault diagnosis under intelligent manufacturing breaks information islands and extracts more value from data. With the rapid development of technology, today's industrial systems are getting larger and more complex in structure. Summary of the Invention

[0005] Aiming at the defects in the prior art, the purpose of the present invention is to provide an industrial fault diagnosis method and system based on prior knowledge.

[0006] The industrial fault diagnosis method based on prior knowledge provided by the present invention includes:

[0007] Step 1: Use a signed directed graph for modeling to construct a topological prior graph of industrial knowledge;

[0008] Step 2: Extract local features in the graph structure of the input data through a graph convolutional network;

[0009] Step 3: Construct a statistic for fault diagnosis;

[0010] Step 4: Based on the graph convolutional network, use a graph convolutional long short-term memory recurrent neural network to learn the dynamic time characteristics in industrial data for fault prediction;

[0011] Step 5: Combine graph convolution and long short-term memory network, and encapsulate them with SDG modeling into a microservice device to implement fault diagnosis and fault prediction of industrial systems in the form of microservices.

[0012] Preferably, the said Step 1 includes:

[0013] First, list the variables and related components of the industrial system, and define the variables as the nodes of the model;

[0014] Then, determine the interaction relationships between variables based on empirical knowledge to obtain an influence equation set;

[0015] Finally, establish an SDG model from the influence equation set, and verify and correct the SDG using industrial data to finally obtain the industrial process SDG graph.

[0016] Preferably, the said Step 2 includes:

[0017] The input of the graph convolutional network is a complete graph. In the first convolutional layer, a convolutional operation is performed on the neighbor nodes of each node in the graph, and the obtained result is used to update the node, and then a non-linear mapping is performed using an activation function; repeat the above process to construct a deep graph convolutional network model;

[0018] The product of the neighborhood matrix A, the input data x t and the trainable weight matrix W is regarded as the graph convolutional operator G for extracting first-order domain features t = Ax t W is used as the local feature of graph convolution.

[0019] Preferably, the said Step 3 includes:

[0020] Define the space spanned by the local features of the graph convolutional network, which is called the feature subspace. Based on the statistical distribution characteristics of the features, based on the local feature G t , construct H under the feature subspace2 Perform a hypothesis test on the statistic, and the expression is:

[0021] H 2 = G t T G t

[0022] For any input sample, H 2 represents the sum of squares of local features, which is used to express the statistical data of the feature space. The corresponding hidden layer features fall within this distribution with a preset high probability; for test data that does not meet the preset conditions, the corresponding hidden layer features will deviate from the feature space, thereby causing H 2 to deviate from this distribution.

[0023] Preferably, the step 4 includes:

[0024] Expand the receptive field of the graph convolution from the first-order neighborhood to the k-order neighborhood, and define a k-order graph convolution operator:

[0025]

[0026] where ⊙ is the Hadamard product operator, that is, the element-wise matrix multiplication operator; x t represents the vector composed of variables at the k-th time step of the industrial process; represents the k-order neighborhood matrix; W gk ∈R N×N is the trainable weight matrix of the k-order graph convolution; G t k represents the extracted k-order graph convolution feature;

[0027] The neighborhoods of different orders correspond to different graph convolution features. The graph convolution features extracted from neighborhoods of different orders are spliced together: as the input of the prediction network.

[0028] According to the industrial fault diagnosis system based on prior knowledge provided by the present invention, it includes:

[0029] Module M1: Use a signed directed graph for modeling to construct a topological prior graph of industrial knowledge;

[0030] Module M2: Extract local features in the graph structure of the input data through a graph convolutional network;

[0031] Module M3: Construct a statistic for fault diagnosis;

[0032] Module M4: Based on the graph convolutional network, use a graph convolutional long short-term memory recurrent neural network to learn the dynamic time characteristics in industrial data for fault prediction;

[0033] Module M5: Combines graph convolution with long short-term memory network, encapsulates it with SDG modeling into a microservice device, and realizes fault diagnosis and fault prediction of industrial systems in the form of microservices.

[0034] Preferably, the module M1 includes:

[0035] First, list the variables and related components of the industrial system, define the variables as the nodes of the model.

[0036] Then, determine the interaction relationships between variables based on empirical knowledge to obtain the influence equations.

[0037] Finally, establish an SDG model from the influence equations, verify and correct the SDG using industrial data, and finally obtain the industrial process SDG graph.

[0038] Preferably, the module M2 includes:

[0039] The input of the graph convolutional network is a complete graph. In the first convolutional layer, a convolutional operation is performed on the neighbor nodes of each node in the graph, and the obtained result is used to update the node. Then, a non-linear mapping is performed using an activation function; repeat the above process to construct a deep graph convolutional network model.

[0040] The product of the neighborhood matrix A, the input data x t and the trainable weight matrix W is regarded as the graph convolutional operator G for extracting first-order domain features t = Ax t W is used as the local feature of graph convolution.

[0041] Preferably, the module M3 includes:

[0042] Define the space spanned by the local features of the graph convolutional network, called the feature subspace. Based on the statistical distribution characteristics of the features, construct H t , under the feature subspace 2 for hypothesis testing of the statistic, and the expression is:

[0043] H 2 = G t T G t

[0044] For any input sample, H 2 represents the sum of squares of the local features, used to express that the hidden layer features corresponding to the statistical data of the feature space fall within this distribution with a preset high probability; for test data that does not meet the preset conditions, the corresponding hidden layer features will deviate from the feature space, resulting in H 2 deviating from this distribution.

[0045] Preferably, the module M4 includes:

[0046] Expand the receptive field of graph convolution from the first-order neighborhood to the k-order neighborhood, and define a k-order graph convolution operator:

[0047]

[0048] where ⊙ is the Hadamard product operator, that is, the element-wise matrix multiplication operator; x t represents the vector composed of variables at the k-th time step of the industrial process; represents the k-order neighborhood matrix; W gk ∈R N×N is the trainable weight matrix of the k-order graph convolution; G t k represents the extracted k-order graph convolution features;

[0049] The neighborhoods of different orders correspond to different graph convolution features. Concatenate the graph convolution features extracted from neighborhoods of different orders: As the input of the prediction network.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] The present invention realizes the fault diagnosis and fault prediction of the industrial system through microservices, interconnects and tightly couples the various components of the industrial system. Using the microservice architecture, the industrial system can be decoupled, and many services can be divided into multiple microservices, and the microservices are relatively independent and easy to manage. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] By reading the following detailed description of the non-limiting embodiments with reference to the accompanying drawings, other features, objects and advantages of the present invention will become more apparent:

[0053] Figure 1 is the GC-LSTM model;

[0054] Figure 2 is the fault diagnosis model based on topological prior;

[0055] Figure 3 is the flowchart of fault diagnosis and fault prediction. DETAILED DESCRIPTION OF THE INVENTION

[0056] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several changes and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0057] Embodiment:

[0058] The present invention provides an industrial fault diagnosis method based on prior knowledge. First, a symbolic directed graph (SDG) modeling method using prior knowledge is used to construct a topological graph of the industrial process. Then, graph convolution is used to extract prior features. Construct the H 2 statistic for fault diagnosis. A long short-term memory (LSTM) network model is used for fault prediction. Finally, graph convolution and the LSTM network are combined and encapsulated with the SDG modeling into a microservice device. Ultimately, the fault diagnosis and fault prediction of the industrial system are realized in the form of microservices.

[0059] Specifically, it includes the following steps:

[0060] 1. Use SDG modeling to construct a topological prior graph of industrial knowledge

[0061] The symbolic directed graph (SDG) is a topological model used to describe the causal relationships between internal nodes of a complex system. It consists of nodes and directed edges between the nodes, and can visually display the complex system in the form of a graph, which provides an important model basis for studying the risks of anomalies and the propagation of faults in complex industrial systems and timely locating the fault sources. SDG has excellent completeness and can encompass a large amount of information. Compared with traditional shallow empirical models - expert systems, the knowledge model of SDG is deeper and contains more comprehensive information.

[0062] The present invention uses a method combining flowcharts and experience to perform SDG modeling on the industrial system. First, important variables (as nodes of the model) and related components involved in the industrial system are listed, and the variables are defined. Then, based on empirical knowledge, the interaction relationships between the variables are determined to obtain an influence equation set. Finally, an SDG model is established from the influence equation set, and the SDG is verified and corrected using industrial data to obtain the final SDG graph of the industrial process.

[0063] 2. Use graph convolution to extract prior features

[0064] The core idea of the convolutional layer of a traditional neural network is to extract local features of the input industrial data. Similar to traditional convolution, the core idea of graph convolution is also to extract local features in the graph structure of the input data. The input of a graph convolutional network is a complete graph. In the first convolutional layer, a convolution operation is performed on the neighbor nodes of each node in the graph, and the resulting result is used to update the node. Then, a non-linear mapping is performed using an activation function; repeating the above process, a deep graph convolutional network model can be constructed. Therefore, the product of the neighborhood matrix A, the input data x t and the trainable weight matrix W can be regarded as a graph convolution operator G for extracting first-order neighborhood features t = Ax t W as the local feature of graph convolution.

[0065] 3. Construct statistics for fault diagnosis

[0066] The space formed by the local features of the graph convolutional network is called the feature subspace. According to the statistical distribution characteristics of the features, based on the local features G t , construct H in the characteristic subspace 2 Statistics for hypothesis testing;

[0067] H 2 =G t T G t

[0068] For any input sample, H 2 Represents the sum of squares of local features and can be used to express the statistical distribution of feature space. The hidden features corresponding to the data should fall into this distribution with a greater probability. However, for abnormal test data, the corresponding hidden features will deviate from the feature space, which will lead to H 2 deviate from this distribution.

[0069] 4. Long Short-Term Memory Network for Fault Prediction

[0070] Based on the graph convolutional network, the graph convolutional long short-term memory (GC-LSTM) recurrent neural network is used to learn the dynamic time characteristics of industrial data and realize fault prediction.

[0071] The method of determining the receptive field using the first-order neighborhood is limited, as it ignores the information contained in nodes in the more distant neighborhood. To avoid this limitation, it is necessary to expand the receptive field of graph convolution from the first-order neighborhood to the k-order neighborhood. Therefore, a k-order graph convolution operator is defined:

[0072]

[0073] Among them, ⊙ is the Hadamard product operator, that is, the element-by-element matrix multiplication operator, x t Represents the vector of variables in the k-time step of the industrial process, W represents the k-order domain matrix. gk ∈R N×N is the trainable weight matrix of k-order graph convolution, G t k Represents the extracted k-order graph convolution features, since is a sparse matrix, so is also sparse. In addition, the trained weight matrix W gk The mutual influence between graph nodes can be measured, thereby improving the interpretability of the model.

[0074] Domains of different orders correspond to different graph convolution features. To enrich the feature space, the graph convolution features extracted from domains of different orders (from 1 to K) are concatenated together, that is: which will be used as the input of the prediction network.

[0075] 5. Fault Diagnosis Microservice Method and Device

[0076] The industrial fault diagnosis method based on topological prior is encapsulated into a microservice implementation device. The microservice device includes an API gateway, a database, a GC-LSTM model, etc.

[0077] Among them, the API gateway provides a user interface, through which real-time data can be input into the API interface, and the fault prediction and diagnosis results can also be obtained from the API interface.

[0078] The database stores the historical data of important parameters in the industrial system, as well as the fault-free real-time data input into the microservice device.

[0079] The GC-LSTM model is the model business module of the fault diagnosis microservice device, which is used to implement fault diagnosis and fault prediction.

[0080] The specific steps are as Figure 3 shown:

[0081] 1. Input the real-time data of the industrial system into the fault diagnosis microservice device through the API interface;

[0082] 2. Select important parameters in the industrial system, obtain their historical data and store them in the database, and construct a standard data set. The standard data set constructs an SDG graph based on prior knowledge and industrial flowcharts;

[0083] 3. Use the SDG graph to train the GC model to obtain the optimal parameters;

[0084] 4. Extract features and calculate as the input of the LSTM prediction model;

[0085] 5. Use to train the LSTM prediction model;

[0086] 6. Estimate the control threshold of the statistic;

[0087] 7. Calculate the statistic and determine whether the statistic threshold is reached;

[0088] 8. If the threshold is reached, a fault occurs, and the fault diagnosis result is output through the API interface; if the threshold is not reached, no fault occurs, and the real-time data is input into the database.

[0089] As Figure 1As shown, the GC-LSTM model extracts spatial features by performing convolution operations on the graph information of industrial processes. Simultaneously, it leverages LSTM to learn the dynamic temporal characteristics of industrial data. The gate structure and hidden layers in this model remain consistent with those of the basic LSTM, while the input is replaced by a graph convolution feature vector. Because the variables corresponding to each node in the SDG graph of an industrial process are related to the historical state of the node itself and the variables corresponding to the domain nodes, the LSTM cell state of each node in the graph is also affected by the cell state of the domain nodes. Therefore, a cell state gate is added to the LSTM unit.

[0090] like Figure 2 As shown, unlike common prediction models, which take as input a vector sequence of variables, this model incorporates topological priors and takes as input a feature map of the relationship between variables, processed through a graph convolution operation. This allows it to extract patterns and features containing extensive information from the spatial domain, improving fault prediction performance.

[0091] The industrial fault diagnosis system based on prior knowledge provided by the present invention includes: module M1: using symbolic directed graphs for modeling to construct a topological prior graph of industrial knowledge; module M2: extracting local features in the input data graph structure through a graph convolutional network; module M3: constructing statistics for fault diagnosis; module M4: based on the graph convolutional network, using a graph convolutional long short-term memory recurrent neural network to learn the dynamic time characteristics in industrial data and perform fault prediction; module M5: combining graph convolution with long short-term memory networks, and encapsulating them together with SDG modeling into a microservice device, to realize fault diagnosis and fault prediction of industrial systems in a microservice manner.

[0092] The module M1 includes: first listing the variables and related components of the industrial system, defining the variables as nodes of the model; then determining the interaction relationship between the variables based on empirical knowledge to obtain a set of influence equations; finally, establishing an SDG model based on the set of influence equations, and using industrial data to verify and modify the SDG, ultimately obtaining an SDG map of the industrial process.

[0093] The module M2 includes: the input of the graph convolutional network is a complete graph, and a convolution operation is performed on the neighboring nodes of each node in the graph in the first convolution layer. The result is used to update the node, and then a nonlinear mapping is performed using the activation function; the above process is repeated to build a deep graph convolutional network model; the neighborhood matrix A, the input data x t The product of the trainable weight matrix W is regarded as the graph convolution operator G that extracts the first-order domain features. t =Ax t W is used as the local feature of graph convolution.

[0094] The module M3 includes: defining the space spanned by the local features of the graph convolutional network, called the feature subspace, based on the local feature G according to the statistical distribution characteristics of the features. t , construct H in the characteristic subspace 2 The statistic is used for hypothesis testing, and the expression is: H 2 =G t T G t , for any input sample, H 2 Represents the sum of squares of local features, which is used to express the statistical distribution of feature space. The hidden features corresponding to the data fall into the distribution with a preset high probability. For test data that does not meet the preset conditions, the corresponding hidden features will deviate from the feature space, which will lead to H 2 deviate from this distribution.

[0095] The module M4 includes: expanding the receptive field of graph convolution from the first-order neighborhood to the k-order neighborhood, and defining a k-order graph convolution operator: Where ⊙ is the Hadamard product operator, i.e., the element-by-element matrix multiplication operator; x t Represents the vector composed of the variables of the industrial process k time steps; represents the k-order domain matrix; W gk ∈R N×N is the trainable weight matrix of k-order graph convolution; G t k Represents the extracted k-order graph convolution features; different order fields correspond to different graph convolution features, and the graph convolution features extracted from different order fields are spliced together: As input to the prediction network.

[0096] Those skilled in the art will appreciate that, in addition to implementing the system, device, and various modules provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, and the like by logically programming the method steps. Therefore, the system, device, and various modules provided by the present invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; the modules for implementing various functions can also be considered both software programs for implementing the method and structures within the hardware component.

[0097] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. An industrial fault diagnosis method based on prior knowledge, characterized in that, Including: Step 1: Use signed directed graph for modeling to construct the topological prior graph of industrial knowledge; Step 2: Extract local features in the graph structure of input data through graph convolutional network; Step 3: Construct statistics for fault diagnosis; Step 4: Based on the graph convolutional network, use graph convolutional long short-term memory recurrent neural network to learn the dynamic time characteristics in industrial data for fault prediction; Step 5: Combine graph convolution and long short-term memory network, and encapsulate them with SDG modeling into a microservice device to realize fault diagnosis and fault prediction of industrial systems in the form of microservices; The said Step 1 includes: First list the variables and related components of the industrial system, and define the variables as the nodes of the model; Then determine the interaction relationship between variables according to empirical knowledge to obtain the influence equations; Finally, establish an SDG model from the influence equations, and use industrial data to verify and correct the SDG, and finally obtain the industrial process SDG graph; The said Step 2 includes: The input of the graph convolutional network is a complete graph. In the first convolutional layer, a convolutional operation is performed on the neighbor nodes of each node in the graph, and the obtained result is used to update the node, and then a non-linear mapping is performed using an activation function; repeat the above process to construct a deep graph convolutional network model; Neighborhood matrix A, input data x t The product of the trainable weight matrix W and the above is regarded as the graph convolution operator G for extracting first-order neighborhood features t = Ax t W serves as the local feature of graph convolution; The said Step 3 includes: Define the space spanned by the local features of the graph convolutional network, which is called the feature subspace. Based on the statistical distribution characteristics of the features and the local feature G t , construct H 2 under the feature subspace for hypothesis testing. The expression is as follows: H 2 = G t T G t For any input sample, H 2 represents the sum of squares of local features, which is used to express the statistical data of the feature space. The corresponding hidden layer features fall into this distribution with a preset high probability; for test data that does not meet the preset conditions, the corresponding hidden layer features will deviate from the feature space, thereby causing H 2 to deviate from this distribution.

2. The industrial fault diagnosis method based on prior knowledge according to claim 1, wherein The said Step 4 includes: Expand the receptive field of graph convolution from the first-order neighborhood to the k-order neighborhood, and define a k-order graph convolution operator: where, ⊙ is the Hadamard product operator, i.e., the element-wise matrix multiplication operator; x t represents the vector composed of variables at the k-th time step of the industrial process; represents the k-th order neighborhood matrix; W gk ∈ R N×N is the trainable weight matrix for the k-th order graph convolution; G t k represents the extracted k-th order graph convolution features; Domains of different orders correspond to different graph convolution features, and the graph convolution features extracted from domains of different orders are spliced together: As the input of the prediction network.

3. An industrial fault diagnosis system based on prior knowledge, characterized in that, Including: Module M1: Use signed directed graph for modeling to construct the topological prior graph of industrial knowledge; Module M2: Extract local features in the graph structure of input data through graph convolutional network; Module M3: Construct statistics for fault diagnosis; Module M4: Based on the graph convolutional network, use graph convolutional long short-term memory recurrent neural network to learn the dynamic time characteristics in industrial data for fault prediction; Module M5: Combine graph convolution and long short-term memory network, and encapsulate them with SDG modeling into a microservice device to realize fault diagnosis and fault prediction of industrial systems in the form of microservices; The said Module M1 includes: First list the variables and related components of the industrial system, and define the variables as the nodes of the model; Then determine the interaction relationship between variables according to empirical knowledge to obtain the influence equations; Finally, establish an SDG model from the influence equations, and use industrial data to verify and correct the SDG, and finally obtain the industrial process SDG graph; The said Module M2 includes: The input of the graph convolutional network is a complete graph. In the first convolutional layer, a convolutional operation is performed on the neighbor nodes of each node in the graph, and the obtained result is used to update the node, and then a non-linear mapping is performed using an activation function; repeat the above process to construct a deep graph convolutional network model; Neighborhood matrix A, input data x t The product of the trainable weight matrix W and the above is regarded as the graph convolution operator G for extracting first-order neighborhood features t = Ax t W is the local feature of graph convolution; The said Module M3 includes: Define the space spanned by the local features of the graph convolutional network, which is called the feature subspace. Based on the statistical distribution characteristics of the features, construct H t under the feature subspace 2 and conduct a hypothesis test on the statistic. The expression is as follows: H 2 = G t T G t For any input sample, H 2 represents the sum of squares of local features, which is used to express the statistical data of the feature space. The corresponding hidden layer features fall within this distribution with a preset high probability; for test data that does not meet the preset conditions, the corresponding hidden layer features will deviate from the feature space, thereby causing H 2 to deviate from this distribution.

4. The industrial fault diagnosis system based on prior knowledge according to claim 3, characterized in that The said Module M4 includes: Expand the receptive field of graph convolution from the first-order neighborhood to the k-order neighborhood, and define a k-order graph convolution operator: where ⊙ is the Hadamard product operator, i.e., the element-wise matrix multiplication operator; x t represents the vector composed of variables at the k-th time step of the industrial process; represents the k-th order neighborhood matrix; W gk ∈R N×N is the trainable weight matrix of the k-th order graph convolution; G t k represents the extracted k-th order graph convolution features; Domains of different orders correspond to different graph convolution features, and the graph convolution features extracted from domains of different orders are spliced together: As the input of the prediction network.

Images