Industrial process fault root diagnosis method based on dynamic sample selection and causal relationship mining
By adopting the time-sequence causal mining method of dynamic sample selection in industrial process fault diagnosis, the causal information loss caused by sample imbalance is solved, and a more accurate diagnosis of the root cause of failure is achieved.
Patent Information
- Application Number
- CN202311709484.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-13
- Publication Date
- 2025-06-13
AI Technical Summary
The existing industrial process fault diagnosis methods have deteriorated performance under uneven sample conditions, resulting in the loss of causal information and it is difficult to accurately explore causal relationships.
The time-series causal mining method based on dynamic sample selection is adopted, and the dynamic training set is constructed by dynamic selection of samples by agents, and the training of the time-series causal mining network is optimized to avoid the loss of causal information caused by sample imbalance.
It effectively solves the problem of causal information loss caused by sample imbalance, so that the causal relationship mining network can accurately mine the causal relationship in the data set and improves the accuracy of the root cause diagnosis of faults.
Smart Images
Figure CN120146095A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of industrial process fault diagnosis, and specifically relates to an industrial process fault root cause diagnosis method based on dynamic sample selection and causal relationship mining. Background Art
[0002] With the deep integration of industrialization and informatization, modern industrial processes (such as steel, non-ferrous metals, petrochemical, etc.) are developing towards the direction of high efficiency, intelligence, and integration. The continuous and uninterrupted operation of the above industrial processes makes any failure of a unit or subsystem likely to spread and evolve through material flow, energy flow, and information flow among different system levels, affecting the stable operation of the production process and the final product quality. The raw material composition, operating conditions, and key quality indicators, etc. cannot be measured or intelligently perceived online, making industrial process fault diagnosis a comprehensive and complex problem. Therefore, around the strategic goal of realizing a manufacturing power, ensuring the high-quality and high-efficiency operation of industrial processes through reasonable fault diagnosis technology has become an important part of the sustainable development of the national manufacturing industry. It will have important strategic significance for suppressing the decline of product quality and maximizing the potential of process operation, and has become a research hotspot in the current field of industrial process control, with important theoretical value and broad engineering application prospects.
[0003] Fault root cause diagnosis technology generally first conducts fault detection and fault isolation, and then mines the causal relationship of the fault-related variables isolated. The fault root cause is located according to the obtained causal relationship diagram. At present, there have been relatively rich studies on the causal relationship mining of time series. In this field, the two most common types of methods are the Granger-based method and the Constrained-based method respectively. The former usually assumes that the past of a "cause" is necessary and sufficient for optimally predicting its "effect", while the latter first uses conditional independence to establish the correlation relationship between each factor, and then orients the correlation relationship according to a set of constraint rules defined in the direction.
[0004] However, the above methods usually require rich training data and have high requirements for data quality, and their performance will drop sharply in the case of unbalanced samples. Summary of the Invention
[0005] The present invention provides an industrial process fault root cause diagnosis method based on dynamic sample selection and causal relationship mining. Fault detection and fault isolation are performed on industrial process data, and then causal relationship mining is carried out on fault-related variables, and finally the fault root cause is obtained. For the causal relationship mining of fault-related variables, the present invention proposes a time-series causal relationship mining method based on dynamic sample selection. First, an agent dynamically selects samples to construct a dynamic training set, effectively solving the problem of causal information loss caused by sample imbalance; then, the dynamic training set is used to train the time-series causal relationship mining network; finally, the weight modulus obtained after training the causal relationship mining network is calculated to obtain a causal relationship graph. The present invention can dynamically select appropriate samples for the network to train according to the current state of the causal relationship mining network, thereby avoiding the problem of causal information loss caused by sample imbalance, and enabling the causal relationship mining network to accurately mine the causal relationships in the dataset.
[0006] The technical solution adopted by the present invention to achieve the above object is:
[0007] An industrial process fault root cause diagnosis method based on dynamic sample selection and causal relationship mining, which performs causal relationship mining on fault-related variables of an industrial process to obtain fault root cause variables and prompts them to users; includes the following steps:
[0008] Step 1: Collect industrial process variable data and perform preprocessing;
[0009] Step 2: Perform kernel principal component analysis fault detection on the data to determine whether it is fault data;
[0010] Step 3: If so, perform a fault isolation operation to obtain fault-related variables;
[0011] Step 4: Initialize the network models of the causal relationship mining network and the intelligent agent dynamic sample selector; construct an initial dynamic dataset with the fault-related variables and use it as the input of the network model, and dynamically select data samples through the intelligent agent to construct a dynamic dataset; use the dynamic dataset to optimize and train the time-series causal relationship mining network, calculate the weight modulus obtained after training the causal relationship mining network, and obtain a causal graph;
[0012] Step 5: Select variables without father nodes in the causal graph as root faults for early warning in the industrial scenario.
[0013] The preprocessing is data standardization.
[0014] The determination of whether it is fault data is obtained by comparing the data after kernel principal component analysis fault detection with the fault-free value of the industrial process variable data.
[0015] The fault isolation operation is to obtain the fault correlation between variables through contribution degree calculation.
[0016] The dynamic selection of data samples by the agent to construct a dynamic data set includes the following steps:
[0017] Step 4-1: Define a fault model composed of a causal relationship mining network and a selector network; initialize the parameters of the two networks and the dynamic data set;
[0018] Step 4-2: a. Train the causal relationship mining network using the dynamic data set; b. Obtain the causal graph according to the modulus value of the weights of the first layer of the trained network;
[0019] Step 4-3: a. Use the current parameter state of the causal relationship mining network, the causal graph, and its loss value in the current dynamic data set as the input of the agent dynamic sample selector; b. Sample from the selected data segments according to the sample selection decision output by the sample selector, and update it to the dynamic data set.
[0020] The optimization training of the time-series causal relationship mining network using the dynamic data set includes:
[0021] Step 4-4: Return to step 4-2 for iteration, and store the parameter state of the causal relationship mining network, the action decision and loss value of the agent dynamic sample selector, and the new state of the parameter of the causal relationship mining network obtained in the next round in the experience replay pool;
[0022] Step 4-5: Randomly sample from the experience replay pool, and use the sampled samples and combine with the SAC reinforcement learning algorithm to train the sample selector to meet the agent optimization goal:
[0023]
[0024] where π t represents the sample selection strategy given by the agent after the t-th training of the agent, represents the sub-data set selected under the π t strategy, f is the neural network model, represents the entropy of the sub-data set;
[0025] Step 4-6: a. Repeat steps 4-3 to 4-5 until the sample selector network and the causal relationship mining network converge to the overall algorithm optimization goal:
[0026]
[0027] That is, the algorithm converges to the situation where the MSE of f on any sub-dataset is the minimum. At this time, the dynamic data set selected by the sample selector is the dynamic data set with the largest MSE value for f that may change currently and in the future among all dynamic data sets; b. At this point, after the network model is optimized and trained, the causal graph of the causal relationship mining network parameters, the data selection decision output by the intelligent agent dynamic sample selector, and the weight module value of the first layer of the causal relationship mining network is obtained.
[0028] The causal relationship-related paths are obtained through depth-first search.
[0029] The causal relationship mining network is composed of a long short-term memory recurrent neural network, and its training steps are as follows:
[0030] 1) Given N input time series variables and T consecutive time samples, select the samples of the first T-1 time as the input of the neural network;
[0031] 2) Use the samples at the next T-1 moments as labels; for N variables, construct N neural networks with the same structure but different parameters. The input of each neural network is N time series variables, and the output is the predicted value of a time series variable at the next moment;
[0032] 3) Calculate MSE and L based on the predicted value and label value 0 Norm loss value:
[0033] Loss = MSE(f(X),X)+L 0 (f)
[0034] Where f is the neural network model, L 0 is the norm;
[0035] 4) Perform back propagation based on the loss value to calculate the gradient of each parameter of the neural network;
[0036] 5) Update the neural network parameters using stochastic gradient descent according to the gradient;
[0037] 6) Use proximal gradient descent to update the parameters of the first layer of the network.
[0038] The dynamic sample selector is a reinforcement learning agent, which inputs the current parameter state of the causal mining network, the causal graph, and the loss value of the causal mining network in the current dynamic data set; the action space is the index number of the entire data set; wherein, in each round, the dynamic data set is first divided into M continuous segments, each of which has an index number, and the agent action is M discrete actions, which represent the sequential retrieval of the data pointed to by the index number selected by the agent.
[0039] An industrial process fault root cause diagnosis device based on dynamic sample selection and causal relationship mining, including a front-end interface and a background. The background is provided with a memory and a processor. A program is stored in the processor. When the processor loads the program, it executes the method steps described in any one of claims 1-9, mines the causal relationships of the fault-related variables in the industrial process, thereby obtaining the fault root cause variables, and visually displays them to the user front-end interface of the industrial scenario.
[0040] The present invention has the following beneficial effects and advantages:
[0041] A method for diagnosing the root cause of faults in an industrial process based on dynamic sample selection and causal relationship mining. Compared with the existing optimal method in terms of time-series causal relationship mining, it can dynamically select appropriate samples for the network to train according to the state of the current causal relationship mining network, effectively solve the problem of causal information loss caused by sample imbalance, and enable the causal relationship mining network to accurately mine the causal relationships in the dataset. Description of the Drawings
[0042] Figure 1 It is a flowchart of the method of the present invention.
[0043] Figure 2 It is a flowchart of the method for mining the causal relationships of the fault-related variables of the present invention.
[0044] Figure 3 It is a directed causal graph of the present invention. Detailed Embodiments
[0045] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following describes the specific implementation method of the present invention in detail with reference to the accompanying drawings. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the invention. Therefore, the present invention is not limited by the specific implementations disclosed below.
[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments, and are not intended to limit the present invention.
[0047] As Figure 1 shown, it is a flowchart of the method of the present invention.
[0048] A method for diagnosing the root cause of faults in an industrial process based on dynamic sample selection and causal relationship mining. The programming language used for the program execution steps of the present invention is not limited to MATLAB, Python, etc.
[0049] The specific steps of the present invention are as follows:
[0050] Step 1: Collect the fault variables that may occur in the industrial process and the variable data that may be related to the fault (taking the chemical production process as an example, variables such as water pressure, material input, valve pressure, reactor pressure, reactor temperature, etc. can all be used as variable data), and preprocess the collected data.
[0051] Step 1.1: Subtract the sample mean from each sample and perform variance normalization.
[0052] Step 2: Perform KPCA fault detection on the data to determine whether it is fault data. Determining whether it is fault data is obtained by comparing the preprocessed data after kernel principal component analysis fault detection with the fault-free values of the industrial process variable data. For example, after fault detection, 8 fault variable data are obtained: A, B, C, D, E, F, G, H.
[0053] Step 3: If a fault is detected, perform fault isolation to obtain the fault-related variables. The fault isolation operation is to obtain the fault correlation between variables through contribution degree calculation. For example, the fault-related variables among the fault variable data A, B, C, D, E, F, G, H are 5: A, B, C, D, E.
[0054] Step 4: Mine the causal relationship of the fault-related variables obtained by fault isolation. Assume that the above N (5) fault-related variables are obtained, and the number of samples for each variable is M. Assume that the maximum causal relationship delay is lag_max. lag_max is the maximum value of the causal delay time among the above 5 fault-related variables.
[0055] Step 4-1: a. Initialize the causal relationship mining network.
[0056] b. Initialize the sample selector network.
[0057] c. Initialize the dynamic data set.
[0058] Step 4-2: a. Use the dynamic data set to train the causal relationship mining network. Among them, the causal relationship mining network is composed of a long short-term memory recurrent neural network, and its training steps are as follows:
[0059] a1. Given N time series variables and T consecutive moment samples as input. Select the samples of the first T-1 moments as the input of the neural network.
[0060] a2. Use the sample at the T-1th moment after use as the label. For N variables, construct N neural networks with the same structure but different parameters. The input of each neural network is N time-series variables, and the input is the predicted value of the next moment of a single time-series variable.
[0061] a3. Calculate the MSE and L 0 norm loss value according to the predicted value and the label value. The loss value is as described by the following formula:
[0062] Loss = MSE(f(X), X) + L 0 (f)
[0063] where f is the neural network model and L 0 is the norm.
[0064] a4. Perform backpropagation according to the loss value to calculate the gradients of each parameter of the neural network.
[0065] a5. Update the neural network parameters using the stochastic gradient descent method according to the gradients.
[0066] a6. Use proximal gradient descent to update the parameters of the first layer of the network.
[0067] b. Obtain the causal graph according to the weight modulus value of the first layer of the network in the training result.
[0068] Step 4-3, a. Use the current network parameter state of the causal relationship mining network, the causal graph, and its loss value in the current dynamic dataset as the state input of the sample selector. The dynamic sample selector is a reinforcement learning agent, and its inputs are the current state of the causal relationship mining network, the causal graph, and the loss value of the causal relationship mining network in the current dynamic dataset; the action space is the index number of the overall dataset. Specifically, first divide the dataset into M consecutive segmented fragments, where each fragment has an index number, and the agent's action is M discrete actions, representing the index number selected by the agent.
[0069] b. According to the action output by the sample selector, which is the sample selection decision, then sample in the selected data fragment to obtain the sample selected by this decision, and put this sample into the dynamic dataset to obtain a new dynamic dataset.
[0070] Step 4-4: Repeat Step 4-2, and use the loss value after training the causal relationship mining network as the reward signal. Combine the actions and parameter states in Step 4-3 and the new parameter states obtained in 4-2, and store them in the experience replay pool.
[0071] Step 4-5: Randomly sample from the experience replay pool and use the sampled samples to train the sample selector. The algorithm uses the SAC reinforcement learning algorithm.
[0072]
[0073] where π t represents the sample selection strategy given by the agent after the t-th training of the agent, represents under π t the dynamic data set selected under the strategy, represents the entropy of the dynamic data set.
[0074] Steps 4-6, a. Repeat Steps 4-3 to 4-5 until the sample selector network and the causal relationship mining network converge. The overall algorithm optimization objective is as follows:
[0075]
[0076] That is, the algorithm converges to the case where the MSE of f on any sub-data set is the minimum value. At this time, the sub-data set selected by the sample selector is the sub-data set with the largest MSE value for the current and future possible changing f among all sub-data sets.
[0077] b. Obtain a causal graph based on the weight modulus of the first layer of the causal relationship mining network. The path related to the causal relationship is obtained by depth-first search.
[0078] Step 5. Select the variable without a father node as the root cause of the fault in the final causal graph obtained in Step 4.
[0079] Step 6. The root cause variable prompt can be directly displayed on the operator interface to help the operator adjust the industrial field variables and improve the safety and production efficiency of the industrial process.
[0080] The present invention further provides an industrial process fault root cause diagnosis device based on dynamic sample selection and causal relationship mining, including a front-end interface and a background. The background is provided with a memory and a processor. A program is stored in the processor. When the processor loads the program, it executes the above method steps to mine the causal relationship of the fault-related variables in the industrial process, so as to obtain the root cause variable of the fault, and visually display it to the front-end user interface of the industrial scenario in the form of a graph or 3D data.
[0081] The result is as Figure 3 shown. It can be seen that variable 5 is the root cause variable of the fault. It affects all other variables and causes other variables to fail.
[0082] The embodiments described in the above description 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 for those of ordinary skill in the art, several transformations and improvements can be made without departing from the concept of the present invention. These all fall within the protection scope of the present invention.
Claims
1. An industrial process fault root cause diagnosis method based on dynamic sample selection and causal relationship mining, characterized in that, causal relationship mining is carried out on the fault-related variables of the industrial process to obtain the fault root cause variables and prompt them to the user; the method includes the following steps: Step 1: Collect industrial process variable data and perform preprocessing; Step 2: Perform kernel principal component analysis fault detection on the data to determine whether it is fault data; Step 3: If so, perform a fault isolation operation to obtain fault-related variables; Step 4: Initialize the network models of the causal relationship mining network and the intelligent agent dynamic sample selector; construct an initial dynamic data set with the fault-related variables and use it as the input of the network model, and dynamically select data samples through the intelligent agent to construct a dynamic data set; Use the dynamic data set to optimize and train the time-series causal relationship mining network, calculate the weight modulus obtained after the training of the causal relationship mining network, and obtain a causal diagram; Step 5: Select the variables without father nodes in the causal diagram as the root causes of faults for early warning in the industrial scenario.
2. The industrial process fault root cause diagnosis method based on dynamic sample selection and causal relationship mining according to claim 1, characterized in that, the preprocessing is data standardization.
3. The industrial process fault root cause diagnosis method based on dynamic sample selection and causal relationship mining according to claim 1, characterized in that, the determination of whether it is fault data is obtained by comparing the data after kernel principal component analysis fault detection with the fault-free values of the industrial process variable data.
4. The industrial process fault root cause diagnosis method based on dynamic sample selection and causal relationship mining according to claim 1, characterized in that, the fault isolation operation is to obtain the fault correlation between variables through contribution degree calculation.
5. The industrial process fault root cause diagnosis method based on dynamic sample selection and causal relationship mining according to claim 1, characterized in that, the dynamic selection of data samples by the intelligent agent to construct a dynamic data set includes the following steps: Step 4-1: Define a fault model composed of a causal relationship mining network and a selector network; initialize the two network parameters and the dynamic data set; Step 4-2: a. Use the dynamic data set to train the causal relationship mining network; b. Obtain a causal diagram according to the weight modulus of the first layer of the training result network; Step 4-3: a. Use the current parameter state of the causal relationship mining network, the causal diagram, and its loss value in the current dynamic data set as the input of the intelligent agent dynamic sample selector; b. According to the sample selection decision output by the sample selector, sample in the selected data segment to obtain the sample selected by this decision and update it to the dynamic data set.
6. The industrial process fault root cause diagnosis method based on dynamic sample selection and causal relationship mining according to claim 1, characterized in that, the use of the dynamic data set to optimize and train the time-series causal relationship mining network includes: Step 4-4: Return to step 4-2 for iteration, and store the causal relationship mining network parameter state of the current training round, the action decision and loss value of the agent dynamic sample selector, and the new state of the causal relationship mining network parameters obtained in the next round into the experience replay pool; Step 4-5: Randomly sample the experience replay pool, and use the sampled samples in combination with the SAC reinforcement learning algorithm to train the sample selector to meet the agent optimization goal: where π t represents the sample selection strategy given by the agent after the t-th training of the agent, represents the sub-dataset selected under the π t strategy, f is the neural network model, represents the entropy of the sub-dataset; Step 4-6, a. Repeat steps 4-3 to 4-5 until the sample selector network and the causal relationship mining network converge to the overall algorithm optimization goal: That is, the algorithm converges to the situation where the MSE of f on any sub-dataset is the minimum. At this time, the dynamic data set selected by the sample selector is the dynamic data set with the largest MSE value for the current and possible future changes of f among all dynamic data sets. b. At this point, after the network model is optimized and trained, the causal relationship mining network parameters, the data selection decision output by the intelligent agent dynamic sample selector, and the causal graph obtained by the weight module value of the first layer of the causal relationship mining network are obtained.
7. The industrial process fault root cause diagnosis method based on dynamic sample selection and causal relationship mining according to claim 6, It is characterized in that The causal relationship-related paths are obtained through depth-first search.
8. The industrial process fault root cause diagnosis method based on dynamic sample selection and causal relationship mining according to claim 6, It is characterized in that The causal relationship mining network is composed of a long short-term memory recurrent neural network, and its training steps are as follows: 1) Given N input time series variables and T consecutive time samples, select the samples of the first T-1 time as the input of the neural network; 2) Use the samples at the next T-1 moments as labels; for N variables, construct N neural networks with the same structure but different parameters. The input of each neural network is N time series variables, and the output is the predicted value of a time series variable at the next moment; 3) Calculate the MSE and L 0 norm loss value based on the predicted value and the label value: Loss = MSE(f(X), X) + L 0 (f) where f is the neural network model, and L 0 is the norm; 4) Perform back propagation based on the loss value to calculate the gradient of each parameter of the neural network; 5) Update the neural network parameters using stochastic gradient descent according to the gradient; 6) Use proximal gradient descent to update the parameters of the first layer of the network.
9. The industrial process fault root cause diagnosis method based on dynamic sample selection and causal relationship mining according to claim 6, It is characterized in that The dynamic sample selector is a reinforcement learning agent, which inputs the current parameter state of the causal mining network, the causal graph, and the loss value of the causal mining network in the current dynamic data set; the action space is the index number of the entire data set; wherein, in each round, the dynamic data set is first divided into M continuous segments, each of which has an index number, and the agent action is M discrete actions, which represent the sequential retrieval of the data pointed to by the index number selected by the agent.
10. Industrial process fault root cause diagnosis device based on dynamic sample selection and causal relationship mining, It is characterized in that It includes a front end of the interface and a back end. The back end is provided with a memory and a processor. A program is stored in the processor. When the processor loads the program, it executes the method steps described in any one of claims 1-9 to mine the causal relationship of the fault-related variables in the industrial process, so as to obtain the fault root cause variables and visually display them to the user front end of the industrial scenario.