Industrial process fault control method based on double-layer network and ensemble learning

Abstract

The invention relates to the technical field of industrial production process fault control, and provides an industrial process fault control method based on a double-layer network and ensemble learning, and the technical scheme designed by the invention comprises the steps: constructing an industrial process fault propagation model of the double-layer network in which physical equipment and a controller are propagation nodes; designing a dynamic repair control strategy and a dynamic removal control strategy to respectively control the physical equipment and the controller; constructing a target cost function; and solving the solution of the target cost function based on an EL ensemble learning algorithm to obtain a final control pair. According to the invention, through dynamic diagnosis of the fault propagation path and collaborative optimization of the controller parameters, the reliability and economical efficiency of the industrial system are improved.

Description

Technical Field

[0001] The present invention belongs to the field of industrial production process fault control, and particularly relates to an industrial process fault control method based on a double-layer network and integrated learning. Background Art

[0002] With the in-depth promotion of the intelligent manufacturing and Industry 4.0 strategies, modern industry presents significant characteristics of multi-physical field coupling and deep integration of cyber-physical systems. In typical process industrial fields such as energy, chemical engineering, and high-end equipment manufacturing, the continuity of production processes and the strong coupling characteristics between equipment constitute complex non-linear dynamic networks. While such systems improve production efficiency and resource utilization rate, the research on their fault propagation mechanisms and control strategies faces unprecedented theoretical and technical challenges.

[0003] Modern industrial systems build a multi-dimensional correlation structure through transmission pipelines, control bus networks, and distributed sensor networks. This correlation is manifested as physical connections such as mechanical drive chains and thermal cycle systems at the physical level, and as a data interaction system based on industrial Ethernet and OPC-UA protocols at the information level. When a subsystem fails, the disturbance signal will propagate in multiple directions along the physical connection paths between equipment and the information transfer links in the control loop.

[0004] The current mainstream industrial process fault control methods mainly have the following limitations: one is the technical route based on single-point monitoring, which is difficult to effectively identify early latent faults; the other is that most control models rely on static parameter settings and cannot adapt to the time-varying system characteristics brought about by equipment degradation. In addition, random factors such as environmental temperature and humidity fluctuations and human operation errors commonly existing in industrial sites lead to significant spatio-temporal uncertainty in the fault propagation path. Summary of the Invention

[0005] In view of the above defects of the prior art, the present invention proposes an industrial process fault control method based on a double-layer network and integrated learning, including the following steps:

[0006] S1: Construct an industrial process fault propagation model of a double-layer network with physical devices and controllers as propagation nodes respectively;

[0007] S2: Design a dynamic repair control strategy and a dynamic removal control strategy to control physical devices and controllers respectively;

[0008] S3: Construct an objective cost function;

[0009] S4: Solve the solution of the objective cost function based on the EL integrated learning algorithm to obtain the final control pair.

[0010] Preferably, the S1 includes:

[0011] Construct the first - layer propagation network and the second - layer propagation network, and introduce the first - layer propagation network propagation mechanism and the second - layer propagation network propagation mechanism respectively;

[0012] The construction of the first - layer propagation network includes using physical devices as propagation nodes and propagating fault information through transmission links of physical connections and / or superior - subordinate logical relationships;

[0013] The construction of the second - layer propagation network includes using controllers as propagation nodes and propagating control information through transmission links of signal communication.

[0014] Preferably, the first - layer propagation network propagation mechanism includes:

[0015] Introduce parameters related to the probability of fault infection, namely probability failure rate β1, repair rate γ1, success rate θ1, immigration rate Λ1, and removal rate η i (i = 1, 2, 3), and simulate the propagation of fault information among physical devices;

[0016] The simulation of the propagation of fault information among physical devices includes:

[0017] After a normal device S(t) comes into contact with a faulty device I(t), it turns into a faulty device I(t) with a failure rate of β1; after the faulty device I(t) is processed and repaired, it turns into a repaired device R(t) with a repair rate of γ1; after the repaired device R(t) undergoes a period of testing and use, it turns into a normal device S(t) with a success rate of θ1; for a normal device S(t), there is an immigration rate Λ1, and for a normal device S(t) and a faulty device I(t), there is a removal rate η i (i = 1, 2, 3).

[0018] Preferably, the second - layer network propagation mechanism includes:

[0019] Introduce parameters related to the probability of early - warning infection, namely failure rate β2, repair rate γ2, success rate θ2, immigration rate Λ2, and removal rate η i (i = 4, 5, 6), and simulate the propagation of control information among controllers;

[0020] The simulation of the propagation of control information among controllers includes:

[0021] After the controller N(t) of the normal signal comes into contact with the controller A(t) carrying the alarm signal, it is converted into the controller A(t) carrying the alarm signal at the failure rate β2; after the controller A(t) carrying the alarm signal enters the alarm state, it is converted into the controller C(t) for adjusting parameters at the repair rate γ2; after the controller C(t) for adjusting parameters goes through the steps of optimizing parameters, it will be converted into the controller N(t) of the normal signal at the success rate θ2; for the controller N(t) of the normal signal, there is an immigration rate Λ2, and for the controller N(t) of the normal signal and the controller A(t) carrying the alarm signal, there is a removal rate η i (i = 4, 5, 6).

[0022] Preferably, there is a coupling effect of information interaction between the first-layer propagation network and the second-layer propagation network, including:

[0023] The faulty physical device in the first-layer propagation network sends an alarm signal to the controller of the second-layer propagation network, causing the corresponding controller to fall into a state where early warning and adjustment are required;

[0024] The controller after optimizing the parameters in the second-layer propagation network returns the optimized parameters to the corresponding physical device in the first-layer propagation network.

[0025] Preferably, the S1 further includes:

[0026] In the industrial process fault propagation model, the influence of uncertain factors on the state of propagation nodes is characterized by a random noise term, and the formula is as follows:

[0027] dI(t) / dt → dI(t) / dt + σI(t)dB(t)

[0028] In the formula, I(t) is the faulty device, σ is the white noise intensity, and B(t) is an independent scalar Brownian motion, which is used to characterize uncertain factors.

[0029] Preferably, the dynamic repair control strategy in the S2 has the following formula:

[0030] γ1 → γ1 + u1(t)

[0031] In the formula, u1(t) is the dynamic repair control strategy.

[0032] Preferably, the dynamic removal control strategy in the S2 has the following formula:

[0033] γ2 → γ2 + u2(t)

[0034] In the formula, u2(t) is the dynamic removal control strategy.

[0035] Preferably, the S3 is constructed as follows:

[0036]

[0037] In the formula, J(u) is the objective cost function, the average value of the random variable cost, W1 is the cost coefficient related to I(t), W2 is the cost coefficient related to A(t), D1 is the cost coefficient related to u1(t), D2 is the cost coefficient related to u2(t), and t f is the duration of the simulated industrial process.

[0038] Preferably, the S4 includes:

[0039] Solving the objective cost function based on the optimal control theory method to obtain the theoretical control pair (x * (t), u * (t)) (t ∈ 0, tf), and constructing it into a data set. x*t and u*t are the theoretical optimal trajectories of x(t) and u(t) respectively. The state x*(t) = {S * (t), I * (t), R * (t), N * (t), A * (t), C * (t)}, and the control variable Define x * (t) as the input feature of the industrial process fault propagation model, and u * (t) as the output of the objective cost function;

[0040] Train the industrial process fault propagation model, divide the data set into a training set and a test set, the model architecture is the random forest algorithm. For the control variables and Taking the number of decision trees as the core variable, set four groups of hyperparameter combinations for comparative experiments respectively. Through dual-index quantitative evaluation, use the mean absolute deviation to reflect the average linear error between the predicted value and the true value, and use the root mean square deviation to characterize the dispersion degree of the prediction error and the influence of extreme values. Screen out the optimal sub-model of each control channel through grid search, and integrate the optimal sub-models to construct a complete controller;

[0041] Deploy the trained complete controller to the control application of fault information propagation. Extract the initial state x * (0) from the test set as the input feature, combine with the real-time model parameters, and generate the predicted value by model inference. Substitute the predicted value of u * (t) into the differential equation of the industrial process fault propagation model for one-step state update to obtain the state x *(1), and iterate until the complete time domain. The final output of this process is the state x(t), control variable u(t), and the corresponding target cost function value predicted by the industrial process fault propagation model, realizing closed-loop optimal control.

[0042] Beneficial effects:

[0043] The present application proposes an industrial process fault control method based on a double-layer network and ensemble learning, innovatively constructing a "fault propagation - controller feedback - dynamic control" closed-loop mechanism. By dynamically diagnosing the fault propagation path and co-optimizing the controller parameters, the reliability and economy of the industrial system are improved. Description of the drawings

[0044] Figure 1 is a schematic flowchart of a preferred embodiment of the present invention;

[0045] Figure 2 is a schematic diagram of the model state transition of a preferred embodiment of the present invention;

[0046] Figure 3 is a schematic diagram of the curves of the model state and control variable changes of a preferred embodiment of the present invention;

[0047] Figure 4 is a schematic bar chart of the cost distribution of a preferred embodiment of the present invention;

[0048] Figure 5 is a schematic diagram of the simulation results of the ensemble algorithm of a preferred embodiment of the present invention;

[0049] Figure 6 is a schematic diagram of the cost comparison between the optimal control theory method and the results of the ensemble algorithm of a preferred embodiment of the present invention. Detailed implementation manners

[0050] The following details the embodiments of the present invention. The following embodiments are implemented on the premise of the technical solution of the present invention, and detailed implementation manners and specific operation processes are given. However, the protection scope of the present invention is not limited to the following embodiments.

[0051] The present invention designs an industrial process fault control method based on a double-layer network and ensemble learning, as Figures 1-3 shown, including the following steps:

[0052] S1: Construct an industrial process fault propagation model of a double-layer network with physical devices and controllers as propagation nodes respectively;

[0053] S2: Design a dynamic repair control strategy and a dynamic removal control strategy to control physical devices and controllers respectively;

[0054] S3: Construct a target cost function;

[0055] S4: Solve the solution of the target cost function based on the EL integrated learning algorithm to obtain the final control pair.

[0056] Preferably, S1 includes:

[0057] Construct the first-layer propagation network and the second-layer propagation network, and introduce the propagation mechanism of the first-layer propagation network and the propagation mechanism of the second-layer propagation network respectively;

[0058] Constructing the first-layer propagation network includes using physical devices as propagation nodes and propagating fault information through physical connections and / or transmission links of the superior-subordinate logical relationship;

[0059] Constructing the second-layer propagation network includes using controllers as propagation nodes and propagating control information through the transmission link of signal communication.

[0060] Preferably, the propagation mechanism of the first-layer propagation network includes:

[0061] Introduce the parameters related to the fault infection probability, namely the probability failure rate β1, repair rate γ1, success rate θ1, immigration rate Λ1 and removal rate η i (i = 1, 2, 3), and simulate the propagation of fault information among physical devices;

[0062] Simulating the propagation of fault information among physical devices includes:

[0063] After the normal device S(t) contacts the faulty device I(t), it turns into the faulty device I(t) with the failure rate β1; after the faulty device I(t) is processed and repaired, it turns into the repaired device R(t) with the repair rate γ1; after the repaired device R(t) undergoes a period of testing and use, it will turn into the normal device S(t) with the success rate θ1; for the normal device S(t), there is an immigration rate Λ1, and for the normal device S(t) and the faulty device I(t), there is a removal rate η i (i = 1, 2, 3).

[0064] Preferably, the propagation mechanism of the second-layer network includes:

[0065] Introduce the parameters related to the warning infection probability, namely the failure rate β2, repair rate γ2, success rate θ2, immigration rate Λ2 and removal rate η i (i = 4, 5, 6), and simulate the propagation of control information among controllers;

[0066] Simulating the propagation of control information among controllers includes:

[0067] After the controller N(t) with normal signal contacts the controller A(t) carrying the alarm signal, it turns into the controller A(t) carrying the alarm signal at the failure rate β2; after the controller A(t) carrying the alarm signal enters the alarm state, it turns into the controller C(t) for adjusting parameters at the repair rate γ2; after the controller C(t) for adjusting parameters goes through the steps of optimizing parameters, it will turn into the controller N(t) with normal signal at the success rate θ2; for the controller N(t) with normal signal, there is an immigration rate Λ2, and for the controller N(t) with normal signal and the controller A(t) carrying the alarm signal, there is a removal rate η i (i = 4, 5, 6).

[0068] Specifically, the established industrial process fault propagation model considers the propagation mechanism of fault signals between physical devices and controller information; in the first-layer propagation network (physical device layer network), physical devices in different states have an interactive effect. For example, some normal devices will be affected by faulty devices and fall into the faulty state together; in the second-layer propagation network (network control information layer), analogous to the physical device layer network, there is a similar infection mechanism among nodes of different classifications in the control information layer network. The controller receiving the alarm signal will affect the normal controller connected to it, making it also fall into the alarm state and affecting its use. Among them, physical device layer: In this application, the physical devices in the industrial system are set as the propagation nodes of the first-layer model, and the device layer network is constructed according to the physical connections (such as pipelines, conveyor belts) or logical dependency relationships (such as production process sequence) between devices. The device state is defined as three categories: normal S(t), faulty I(t), and repaired R(t), and the state transition is affected by neighboring devices. Control information layer: In this application, the control information in the industrial system is used as the second-layer network, and the control devices are used as the propagation nodes, classified as the controller N(t) with normal signal, the controller A(t) receiving the alarm signal, and the controller C(t) for adjusting parameters. The controller nodes collect device state data in real time and interact information through the communication link. The control information has a monitoring and feedback effect on the physical device layer, and its data abnormality may indirectly affect the determination of the device state.

[0069] Preferably, there is a coupling effect of information interaction between the first-layer propagation network and the second-layer propagation network, including:

[0070] The faulty physical device in the first-layer propagation network sends the alarm signal to the controller in the second-layer propagation network, making the corresponding controller fall into the state of early warning and needing adjustment and processing;

[0071] The controller in the second-layer propagation network after optimizing the parameters returns the optimized parameters to the corresponding physical device in the first-layer propagation network.

[0072] Specifically, there is a coupling effect of information interaction between the physical device layer network and the control information layer network. Some nodes in the control information layer perform cross-layer information propagation, which affects the nodes in the physical device layer. Conversely, the state of the physical device affects and changes the state of the nodes in the control information layer. The coupling propagation mechanism between the two layers is as follows: The faulty device I(t) in the physical device layer directly triggers an alarm message and propagates it to the corresponding controller with probability α, thereby increasing the number of nodes of the controller A(t) carrying the alarm signal; The controller C(t) that adjusts parameters in the control information layer can accelerate the repair speed of the corresponding physical device during the process of optimizing parameters and propagates the repair information to the physical device layer with probability to increase the repaired physical devices R(t).

[0073] Preferably, S1 further includes:

[0074] In the industrial process fault propagation model, a random noise term is used to characterize the influence of uncertain factors on the state of propagation nodes. The formula is as follows:

[0075] dI(t)dt→dI(t)dt+σI(t)dB(t)

[0076] In the formula, I(t) is the faulty device, σ is the white noise intensity, and B(t) is an independent scalar Brownian motion, which is used to characterize uncertain factors.

[0077] Specifically, considering that in reality, random disturbance factors such as human interference and changes in environmental temperature and humidity often occur in industrial engineering, such random factors are considered in the industrial process fault propagation model based on the double-layer network. In the industrial process fault propagation model, a random noise term is used to characterize the influence of uncertain factors such as equipment aging, human interference, and environmental disturbance on the state of propagation nodes.

[0078] Preferably, the dynamic repair control strategy in S2 has the following formula:

[0079] γ1→γ1+u1(t)

[0080] In the formula, u1(t) is the dynamic repair control strategy.

[0081] Preferably, the dynamic removal control strategy in S2 has the following formula:

[0082] γ2→γ2+u2(t)

[0083] In the formula, u2(t) is the dynamic removal control strategy.

[0084] Specifically, in the industrial process fault propagation model, two propagation nodes are extremely crucial, namely I(t) and A(t). Two types of propagation nodes in the fault state and alarm state will cause the industrial production chain to break, and even shut down the entire production line, which not only has a significant impact on the relevant propagation nodes before and after the production link, but also has a great impact on the overall production and operation, bringing incalculable economic losses to industrial production. To minimize the loss impact caused by faults, we designed two types of continuous control strategies for these two key nodes I(t) and A(t), including dynamic repair control strategy and dynamic removal control strategy, so as to reduce the negative impact of key nodes such as faulty equipment and alarm state controllers on the entire propagation process. In reality, the two control strategy means can be regarded as applying more efficient and advanced repair technologies or control methods to improve the repair rate, or can be regarded as manually removing and replacing equipment or controllers that cannot be used due to faults.

[0085] In addition, for the industrial process fault propagation model, the constructed differential equations are as follows:

[0086] dS(t) = [Λ1 - β1S(t)I(t) + θ1R(t) - η1S(t)]dt

[0087] dI(t) = [β1S(t)I(t) - (γ1 + u1(t))I(t) - η2I(t)]dt + σI(t)dB(t)

[0088]

[0089] dN(t) = [Λ2 - β2N(t)A(t) + θ1C(t) - η4N(t)]dt

[0090] dA(t) = [β2B(t)A(t) - (γ2 + u2(t))A(t) + αI(t) - η5A(t)]dt

[0091] dC(t) = [(γ2 + u2(t))A(t) - θ2C(t) - η6C(t)]dt.

[0092] Preferably, the construction formula of S3 is as follows:

[0093]

[0094] In the formula, J(u) is the objective cost function, the average value of the random variable cost, W1 is the cost coefficient related to I(t), W2 is the cost coefficient related to A(t), D1 is the cost coefficient related to u1(t), D2 is the cost coefficient related to u2(t), and t f is the duration of simulating the industrial process.

[0095] Specifically, for the objective cost function of optimal control, the problem of cost minimization based on minimizing the cost of industrial fault control strategies is considered. It is proposed to achieve a balance between control propagation and control cost, and its goal is to solve for the optimal control pair that minimizes the cost function. To solve for the optimal control pair, this application first proposes an optimal control theory method for solving the objective cost function of optimal control, including:

[0096] Design the Hamilton function:

[0097]

[0098] Based on the industrial process fault propagation model and its differential equations, the objective cost function of this optimal control has the following co-state equation solutions:

[0099] dλ1(t) = [λ1(t)(β1I(t) - η1) - λ2(t)β1I(t)]dt

[0100] dλ2(t) = [-W1 + λ1(t)β1S(t) - λ2(t)(β1S(t) - (γ1 + u1(t)) - η2) - λ3(t)(γ1 + u1(t)) - λ5(t)α - q(t)σ]dt

[0101] dλ3(t) = [-λ1(t)θ1 + λ3(t)(θ1 + η3)]dt

[0102] dλ4(t) = [λ4(t)(β2A(t) - η4) - λ5(t)β2A(t)]dt

[0103] dλ2(t) = [-W2 + λ4(t)β2N(t) - λ5(t)(β2N(t) - (γ2 + u2(t)) - η5) - λ6(t)(γ2 + u2(t))]dt

[0104]

[0105] And there are boundary conditions λ i (t f ) = 0, i = 1, 2,..., 6, thus obtaining the solution of the optimal control pair as:

[0106]

[0107] However, for the above-mentioned optimal control theory method for solving the objective cost function of optimal control, although the optimal control theory method can directly calculate u *(t) solution, but there are defects of complex iterative calculation and weak adaptability, which limit its practical application in high-dimensional, non-linear, and time-varying systems. Specifically, since the solution of the objective cost function of optimal control usually relies on local linearization or iterative optimization, the non-linear dynamic characteristics may lead to slow algorithm convergence speed, multiple adjustments of the step size are required, and it has the characteristic of insufficient robustness. To solve the above problems brought by the optimal control method, the present invention introduces an EL ensemble learning algorithm, based on the framework of the random forest algorithm, to solve the objective cost function of the optimal control, thereby effectively improving the calculation efficiency, and using the calculation method of the optimal control theory to obtain as Figures 3-4 The resulting graph shown.

[0108] Preferably, S4 includes:

[0109] Solving the objective cost function based on the optimal control theory method to obtain the theoretical control pair (x * (t), u * (t)) (t ∈ [0, t f ), and constructing it into a data set, x * (t) and u * (t) are the theoretical optimal trajectories of x(t) and u(t) respectively, and the state x * (t) = {S * (t), I * (t), R * (t), N * (t), A * (t), C * (t)} and the control variable Define x * (t) as the input feature of the industrial process fault propagation model, and u * (t) as the output of the objective cost function;

[0110] Train the industrial process fault propagation model, divide the data set into a training set and a test set, the model architecture is the random forest algorithm, and for the control variables and Taking the number of decision trees as the core variable, set four groups of hyperparameter combinations respectively for comparative experiments. Through dual-index quantitative evaluation, use the mean absolute deviation to reflect the average linear error between the predicted value and the true value, and use the root mean square deviation to characterize the dispersion degree of the prediction error and the influence of extreme values. Screen out the optimal sub-models of each control channel through grid search, and integrate the optimal sub-models to construct a complete controller;

[0111] Deploy the trained complete controller to the control application of fault information propagation, extract the initial state x * (0) from the test set as the input feature, combine the real-time model parameters, and generate by model inference The predicted value of u * Substitute the predicted value of u(t) into the differential equation of the industrial process fault propagation model for one-step state update to obtain the state x at the next time node * (1), and iterate until the complete time domain. This process finally outputs the state x(t), control variable u(t) predicted by the industrial process fault propagation model, and the corresponding target cost function value, realizing closed-loop optimal control.

[0112] Specifically, x(t) and u(t) respectively refer to the state variable and control variable that appear in the industrial process fault propagation model. The finally output state x(t) and control variable u(t) predicted by the industrial process fault propagation model are the final control pair;

[0113] In addition, based on the results of the optimal control theory, the present invention uses the EL algorithm to replace the optimal control theory method. The simulation experiment results obtained are as follows Figures 5-6 shown. In the training stage, the data set is divided into a training set (70%) and a test set (30%) using a stratified sampling strategy. The model architecture selects the random forest algorithm, and its core advantage is to improve the generalization performance through the ensemble learning of multiple decision trees. It can be seen that the comparison of the cost data under the optimal control theory method and the EL algorithm can illustrate that the EL algorithm can well replace the optimal control theory and be applied to fault propagation and control applications.

[0114] Preferably, the method further includes:

[0115] Conduct numerical simulation experiments on the industrial process fault propagation control model and verify it. Use MATLAB to build a simulation platform to simulate the fault propagation process characterized in the industrial process fault propagation model. Adopt the calculation method of the theoretical solution of the target cost function of the optimal control theory method to obtain the numerical simulation experiment parameter results of the theoretical solution. Adopt the calculation method solved by the EL ensemble learning algorithm to obtain the numerical simulation experiment parameter results of the EL ensemble learning algorithm. Through the comparison of the cost data in Table 1 (numerical simulation experiment parameter table) under the theoretical solution and the solution by the EL ensemble learning algorithm, it can be illustrated that the EL algorithm can well replace the theory and be applied to fault propagation and control applications.

[0116] Table 1 Numerical simulation experiment parameter table

[0117]

[0118]

[0119] The experimental results show that the above two methods can reflect that the industrial process fault propagation model can achieve accurate modeling, simultaneously depict the coupling effect of the physical connection of equipment and the interaction of sensor data, and improve the prediction accuracy of fault propagation. Secondly, the EL algorithm is used to make up for the computationally complex system, and an efficient calculation method is used to adapt to the random interference and dynamic changes in the industrial environment. In addition, a cost function is designed as the objective of the optimal control problem to achieve an optimal balance between the fault suppression effect and the control cost, and it has good performance and applications in both the simulation prediction and control of the fault propagation process.

[0120] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative labor. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art shall fall within the protection scope determined by the claims.

Claims

1. An industrial process fault control method based on a double-layer network and ensemble learning, characterized in that: The following steps are involved: S1: Construct an industrial process fault propagation model with a two-layer network where physical devices and controllers are propagation nodes; S2: Design dynamic repair control strategy and dynamic removal control strategy to control physical devices and controllers respectively; S3: construct the target cost function; S4: Solve the solution of the target cost function based on the EL ensemble learning algorithm and obtain the final control pair.

2. The industrial process fault control method based on double-layer network and ensemble learning according to claim 1 is characterized in that: The S1 includes: Construct the first-layer propagation network and the second-layer propagation network, and introduce the propagation mechanism of the first-layer propagation network and the propagation mechanism of the second-layer propagation network respectively; The construction of the first layer propagation network includes using physical devices as propagation nodes to propagate fault information through physical connections and / or transmission links of upper and lower logical relationships; The construction of the second layer propagation network includes using the controller as a propagation node to propagate control information through a transmission link of signal communication.

3. The industrial process fault control method based on double-layer network and ensemble learning according to claim 2 is characterized in that: The first-layer propagation network propagation mechanism includes: Introduce the parameters related to the probability of fault infection: probability failure rate β1, repair rate γ1, success rate θ1, migration rate Λ1 and removal rate η i (i=1,2,3), simulating the propagation of fault information among physical devices; The simulated fault information propagation between physical devices includes: After the normal device S(t) comes into contact with the faulty device I(t), it is converted into the faulty device I(t) with a failure rate β1; after being processed and repaired, the faulty device I(t) is converted into the repaired device R(t) with a repair rate γ1; after a period of testing and use, the repaired device R(t) will be converted into the normal device S(t) with a success rate θ1; for the normal device S(t), there is a migration rate Λ1, and for the normal device S(t) and the faulty device I(t), there is a removal rate η i (i=1,2,3).

4. The industrial process fault control method based on double-layer network and ensemble learning according to claim 3 is characterized in that: The second layer network propagation mechanism includes: Introduce the warning infection probability related parameters failure rate β2, repair rate γ2, success rate θ2, migration rate Λ2 and removal rate η i (i=4,5,6), simulating the propagation of control information among controllers; The propagation of the simulation control information between controllers includes: After the controller N(t) with normal signal contacts the controller A(t) with alarm signal, it is converted into the controller A(t) with alarm signal at the failure rate β2; after the controller A(t) with alarm signal enters the alarm state, it is converted into the controller C(t) with parameter adjustment at the repair rate γ2; after the controller C(t) with parameter adjustment goes through the parameter optimization step, it is converted into the controller N(t) with normal signal at the success rate θ2; for the controller N(t) with normal signal, there is a migration rate Λ2; for the controller N(t) with normal signal and the controller A(t) with alarm signal, there is a removal rate or i (i=4,5,6).

5. The industrial process fault control method based on double-layer network and ensemble learning according to claim 4 is characterized in that: There is a coupling effect of information interaction between the first layer propagation network and the second layer propagation network, including: The faulty physical device in the first layer of the propagation network sends an alarm signal to the controller of the second layer of the propagation network, causing the corresponding controller to fall into a state of early warning and needing adjustment and processing; After optimizing the parameters of the second-layer propagation network, the controller returns the optimized parameters to the physical device corresponding to the first-layer propagation network.

6. The industrial process fault control method based on double-layer network and ensemble learning according to claim 5 is characterized in that: The S1 further comprises: In the industrial process fault propagation model, random noise terms are used to characterize the impact of uncertainty factors on the state of propagation nodes. The formula is as follows: dI(t)dt→dI(t)dt+σI(t)dB(t) Where I(t) is the faulty equipment, σ is the intensity of white noise, and B(t) is the independent scalar Brownian motion, which is used to characterize the uncertainty factors.

7. The industrial process fault control method based on double-layer network and ensemble learning according to claim 6 is characterized in that: The dynamic repair control strategy in S2 is as follows: γ1→γ1+u1(t) Where u1(t) is the dynamic repair control strategy.

8. The industrial process fault control method based on double-layer network and ensemble learning according to claim 7 is characterized in that: The dynamic removal control strategy in S2 is as follows: γ2→γ2+u2(t) Where u2(t) is the dynamic removal control strategy.

9. The industrial process fault control method based on double-layer network and ensemble learning according to claim 8 is characterized in that: The S3 construction formula is as follows: Where J(u) is the target cost function, E is the average value of the random variable cost, W1 is the cost coefficient related to I(t), W2 is the cost coefficient related to A(t), D1 is the cost coefficient related to u1(t), D2 is the cost coefficient related to u2(t), t f Duration of the simulated industrial process.

10. The industrial process fault control method based on double-layer network and ensemble learning according to claim 9 is characterized in that: The S4 includes: Based on the optimal control theory method, the target cost function is solved to obtain the theoretical control pair (x * (t),u * (t))(t∈[0,t f ]) and construct it into a data set, x * (t) and u * (t) are the theoretical optimal trajectories of x(t) and u(t), respectively, with state x * (t) = {S * (T),I * (t),R * (t),N * (t),A * (t),C * (t)} and control variables x * (t) is defined as the input feature of the industrial process fault propagation model, u * (t) as the output of the target cost function; The industrial process fault propagation model is trained, and the data set is divided into a training set and a test set. The model architecture is a random forest algorithm. and Taking the number of decision trees as the core variable, four sets of hyperparameter combinations were set for comparative experiments. The dual indicators were used for quantitative evaluation. The mean absolute deviation was used to reflect the average linear error between the predicted value and the true value. The root mean square deviation was used to characterize the discrete degree and extreme value influence of the prediction error. The optimal sub-model of each control channel was screened out through grid search, and the optimal sub-model was integrated to build a complete controller. Deploy the trained complete controller to the control application of fault information propagation and extract the initial state x from the test set. * (0) As input features, combined with real-time model parameters, generated by model reasoning The predicted value of u * The predicted value of (t) is substituted into the differential equation of the industrial process fault propagation model to perform a state update step, and the state x at the next time node is obtained. * (1) and repeats iteratively until the complete time domain. This process finally outputs the state x(t), the control variable u(t) and the corresponding target cost function value predicted by the industrial process fault propagation model to achieve closed-loop optimization control.

Images