Optimal Design Method of Residual Generator for Dynamic Control Systems in Process Industries

Abstract

The invention discloses an optimization design method for a residual generator of a dynamic control system in a process industry, in order to solve the problems of a small fault detection time window and low fault detection sensitivity reflected by the residuals of an existing online residual signal generator applicable to a dynamic system. The invention targets two scenarios, namely, a known and an unknown mechanism model of a controlled object, and performs a unified optimization design based on the residuals of the two scenarios. The controlled system can be modeled as a linear dynamic system and its state space model is consistent. The residual signal is written as a parameterized form of a left coprime decomposition correlation matrix with the controlled system. The residual generator is designed offline through an optimization strategy, and is driven by input and output data collected online in real time, thereby realizing online optimal residual generation. The invention is suitable for residual generation of dynamic control systems in process industries.

Description

Technical Field

[0001] The present invention belongs to the technical field of process monitoring and fault diagnosis of process industry systems, and particularly relates to a residual generator optimization design method applicable to a dynamic control system with multiple variables and coupling between variables. Background Art

[0002] The safe and reliable operation of industrial systems is the cornerstone of ensuring efficiency and productivity. The design of residual generators is a crucial core technology in fields such as process monitoring and fault diagnosis in process industries. When monitoring multivariable, highly coupled dynamic control systems, residual signals are often used to quantify the deviation of the actual measured value of a physical quantity or a dimensionally reduced data feature from its expected theoretical value during the dynamic process, thereby comprehensively reflecting the system's fluctuations. Residual generation involves calculating the residual signal through automated techniques, providing crucial information for subsequent diagnostic decision-making.

[0003] Currently, the main methods for residual generation can be divided into model-based methods and data-based methods. Model-based methods are mainly applicable to scenarios where the controlled object and control system mechanism models are known or easy to identify; while data-based methods are mainly applicable to scenarios such as static systems or dynamic systems with a single operating point and in steady state. Implementing a residual generation method that is applicable to both of these scenarios under a unified framework remains an open technical problem that needs to be solved. Solving this integrated design problem will help simplify the monitoring system structure and reduce the cost of deploying and maintaining the residual generation module in actual systems.

[0004] On the other hand, in order to improve the fault detection rate, reduce the false alarm rate and improve key indicators such as the alarm window, many technical methods have been dedicated to improving the characteristics of the residual signal through the "residual post-processing" step. Residual post-processing is the intermediate link between residual generation and residual evaluation. Its essence is to further extract and transform the generated residual signal through techniques such as spatial mapping, subspace and signal filtering, so that the transformed residual is easier to evaluate, and thus can effectively distinguish between normal working conditions and faults in the decision-making stage. However, relying solely on residual post-processing and ignoring residual generation has become a bottleneck that limits the performance of the entire technical process. Specifically, if effective optimization is not performed in the residual generation stage, the residual signal obtained will lose some key information, which will directly affect the detectability and detection sensitivity of various faults in dynamic systems. The impact caused by these information losses will not be compensated by subsequent steps such as post-processing and residual evaluation.

[0005] In view of this, the present invention provides an integrated optimization design method for a residual generator for improving the statistical characteristics of a residual signal in the residual generation stage. Summary of the Invention

[0006] The present invention aims to solve the problems of small fault detection time window and low fault detection sensitivity of the residual of the existing online residual signal generator applicable to dynamic systems.

[0007] The residual generator optimization design method for dynamic control systems in process industries includes the following steps:

[0008] Step 1: Initialize the state variables of the designed residual generator and the optimal value of the cumulative loss function; set the decay coefficient, time window, number of batch updates, balance coefficient, strategy improvement probability, and iteration termination condition;

[0009] Step 2: Collect process data under normal operating conditions and build an offline historical data demonstration dataset:

[0010] Step B1: Collect sensor output data of the controlled system over a period of time, denoted as y(k); y(k) is a column vector, each element of which corresponds to the output value of a sensor, and k represents a discrete moment;

[0011] Step B2: Record the control instruction or actuator drive signal data at the corresponding moment, denoted as u(k); u(k) is a column vector, each element of which corresponds to a control instruction or actuator drive signal value;

[0012] Step 3: Construct two dynamic subsystems to drive the residual signal. The specific process includes the following steps:

[0013] Step C1: Based on whether the actual conditions of the controlled object mechanism model are known, when the controlled system can be modeled as the following linear dynamic system and its state space model (A, B, C, D) is known, execute steps C2 to C4;

[0014]

[0015] y(k)=Cx(k)+Du(k)

[0016] Where x represents the state variable of the controlled system in the process industry; A, B, C, and D represent the system matrix, input matrix, output matrix, and direct transfer matrix in the state space equation, respectively. The elements in the matrix are obtained by calculation using a known mechanism model.

[0017] Step C2: Design the observer feedback matrix L(θ) so that it satisfies |eig(AL(θ)C)| < 1, where θ represents all the undetermined elements in L(θ); |eig(AL(θ)C)| represents the modulus of the eigenvalue of the matrix (AL(θ)C);

[0018] Step C3: Calculate the left coprime decomposition matrix of the controlled object according to the following formula: And construct dynamic system No. 1, with z1(k) representing the state variables of dynamic system No. 1, A1, B1, C1, and D1 representing the system matrix, input matrix, output matrix, and direct transfer matrix of dynamic system No. 1 respectively:

[0019]

[0020] A1=A-LC, B1=-L, C1=C, D1=I

[0021] z1(k+1)=A1z1(k)+B1y(k)

[0022]

[0023] Step C4: Calculate the left coprime decomposition matrix of the controlled object according to the following formula: And construct dynamic system No. 2, with z2(k) representing the state variables of dynamic system No. 2, A2, B2, C2, and D2 representing the system matrix, input matrix, output matrix, and direct transfer matrix of dynamic system No. 2 respectively:

[0024]

[0025] A2=A-LC, B2=B-LD, C2=C, D2=D

[0026] z2(k+1)=A2z2(k)+B2u(k)

[0027]

[0028] Step 4: Use y(k) in the data set to drive the dynamic system No. 1 and calculate the signal And the dynamic system No. 2 is driven by u(k) in the data set to calculate the signal

[0029] According to the formula Calculate the residual signal r s (k);

[0030] Step 5: Offline evaluation of the residuals, including the following steps:

[0031] Step E1: Based on the residual signal r s (k) Calculate the current error function R(k) and the cumulative loss function V(k);

[0032] Step E2, random batch update parameters θ: for j = 1, 2, ..., J, generate θ j =θ+c r ·randn j , where randn jrepresents the jth random number generated by the normal distribution, J is the total number of batch updates, c r is the balance coefficient;

[0033] Step E3, let k←k+1, repeat steps 3, 4 and 5 of E1, and calculate the corresponding V j (k+1) value;

[0034] Step E4: Calculate the optimal cumulative loss function:

[0035] V * (k)=R(k)+γ·min j {V j (k+1)|j∈(1,J)}

[0036] Step 6: Based on probability p e , using strategy improvement and strategy exploration to * =arg min θ V * (k) perform updates;

[0037] Then judge if θ * Are all the eigenvalues ​​of the system matrices A1 and A2 of the corresponding dynamic system No. 1 and dynamic system No. 2 within the unit circle and satisfy If yes, then the following assignment is performed: Store the corresponding optimal parameter θ * Otherwise, keep the original V * (k),θ * The value remains unchanged;

[0038] Repeat steps 3 to 6 until the termination condition is met;

[0039] Furthermore, in step 3, when the state space model of the controlled system is unknown, steps C5 to C7 are executed;

[0040] Step C5: Randomly generate the undetermined elements in the observer feedback matrix The following formula is used to assign initial values ​​to A1, B1, C1, and D1 so that they meet the controllable standard form:

[0041]

[0042] Matrix B1 has elements, is a dimensionally compatible full parameter matrix, that is, all elements in the matrix are unknown parameters;

[0043] C1=[1 0...0]D1=[0...0]

[0044] Then construct dynamic system No. 1 as follows:

[0045] z1(k+1)=A1z1(k)+B1y(k)

[0046]

[0047] Step C6: Randomly generate the undetermined elements in the observer feedback matrix The following formula is used to assign initial values ​​to A2, B2, C2, and D2 so that they meet the controllable standard form:

[0048]

[0049] Matrix B2 has elements, is a dimensionally compatible full parameter matrix, that is, all elements in the matrix are unknown parameters;

[0050] C2=[1 0...0]D2=[0...0]

[0051] Then construct dynamic system No. 2 as follows:

[0052] z2(k+1)=A2z2(k)+B2u(k)

[0053]

[0054] Step C7: Collect all parameters of the No. 1 dynamic system and the No. 2 dynamic system to obtain θ = [θ *，1 θ *，2 ];

[0055] Furthermore, the error function at the current moment

[0056] Furthermore, the cumulative loss function

[0057] Furthermore, we use strategy improvement and strategy exploration to * =arg min θ V * (k) in the process of updating, with probability p e For parameter θ * Perform optimal update θ * =arg min θ V * (k), with probability 1-p e For parameter θ * Perform random updates;

[0058] Furthermore, the specific process of step one includes the following steps:

[0059] Step A1: The residual generator is a dimension reduction observer, and its state variables represent data features with dynamic characteristics; the residual generator is set to have two state variables z1 and z2, and z1(0)=0 and z2(0)=0 are initialized;

[0060] Step A2: Initialize the optimal value of the cumulative loss function

[0061] Step A3: Set the attenuation coefficient γ of the weighted past-time error function in the cumulative loss function to a constant, 0<γ<1;

[0062] Step A4: Set the value of the time window size κ in the cumulative loss function to a positive integer;

[0063] Step A5: Set the total number of parameter batch updates J to a positive integer;

[0064] Step A6: Set the balance coefficient c in parameter update r ;

[0065] Step A7: Set the probability p of strategy improvement e , satisfying 0<p e <1, then the probability of strategy exploration is 1-p e ;

[0066] Step A8: Set the maximum number of loops, and terminate the execution when the maximum number of loops is exceeded.

[0067] A computer storage medium stores at least one instruction, which is loaded and executed by a processor to implement the residual generator optimization design method for a dynamic control system in a process industry.

[0068] The application method of the residual generator of the dynamic control system in the process industry includes the following steps:

[0069] Collect the online data of the controlled system at the current moment, including the real-time measurement data of the sensors in the controlled system online (k) and the control instruction data u output by the controller online (k0;

[0070] The residual generator designed based on the residual generator optimization design method of the dynamic control system in the process industry generates the optimal residual value; the optimal parameter θ is used * , based on the dynamic system No. 1, the output value is obtained Using the optimal parameter θ * , based on the dynamic system No. 2, the output value is obtained Calculate online residual values

[0071] A computer storage medium stores at least one instruction, which is loaded and executed by a processor to implement an application method of a residual generator in a dynamic control system in a process industry.

[0072] An application device for a residual generator of a dynamic control system in a process industry, the device comprising a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement an application method of the residual generator of the dynamic control system in the process industry.

[0073] The present invention has the following advantages:

[0074] 1. The present invention proposes a residual generator optimization design method. Aiming at tasks such as anomaly detection, process monitoring and fault diagnosis of dynamic control systems in process industries, and taking into account the characteristics of multiple variables and high coupling between variables in the system, a novel parameterization method and optimization algorithm are used in the residual generation stage to obtain a residual signal with better statistical characteristics, which helps to improve the fault detection rate, increase the sensitivity of distinguishing normal and abnormal working conditions, extend the effective alarm time, and reduce alarm jitter and false alarms (monitoring indicators cross the alarm threshold multiple times in a short period of time).

[0075] 2. The technical solution combines strategy exploration with strategy optimization to avoid the optimization process falling into a local optimum. At the same time, the cumulative loss function set in the present invention can quantitatively indicate the optimization effect, has a high learning rate during the optimization process, and can quickly converge to near 0.

[0076] 3. Depending on whether the actual conditions of the controlled object's mechanism model are known, the present invention performs system construction operations in two scenarios. Specifically, by switching between data-driven and mechanism-based steps, the residual generator can be optimized and designed under the condition that the mechanism model is known, or the residual generator can be fully data-driven and optimized using only collected data. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a block diagram of the residual generator optimization design system designed in the invention content.

[0078] Figure 2 It is the unoptimized online residual generation result curve in the embodiment.

[0079] Figure 3 ] is the optimized online residual generation result curve in the embodiment.

[0080] Figure 4 It is a curve showing the cumulative loss function value changing with the number of iterations.

[0081] Figure 5 It is a simplified block diagram of the Tennessee Eastman Chemical Process. DETAILED DESCRIPTION

[0082] The present invention relates to the problem of optimizing the design of a residual generator for online monitoring applications in dynamic industrial control systems in process industries (involving multiple variables and coupling between variables), particularly the problem of improving the statistical characteristics of the residual signal itself during the residual generation phase rather than the residual evaluation phase, as well as the problem of integrated optimization design applicable to both scenarios where the mechanism model of the controlled object is known and unknown.

[0083] To address the problem of optimizing the design of an online residual signal generator for dynamic systems, the present invention expresses the residual signal as a parameterized form of the left coprime decomposition correlation matrix with the controlled system. This residual generator is designed offline using an optimization strategy and driven by input and output data collected online in real time, achieving online optimal residual generation. In this specific implementation, steps 1 through 6 belong to the offline design phase of the optimized residual generator, while steps 7 and 8 belong to the online deployment and application phase of the optimized residual generator.

[0084] Compared with the unoptimized residual generator, the residual signal obtained by the present invention has better statistical characteristics. When a fault or failure occurs, the residual value can significantly and continuously deviate from the baseline corresponding to the normal operating condition, which helps to expand the fault detection time window and improve the fault detection rate.

[0085] Specific implementation method 1: Combination Figure 1 To explain this embodiment,

[0086] The residual generator optimization design method for a dynamic control system in a process industry described in this embodiment includes the following steps:

[0087] Offline design phase for optimizing the residual generator

[0088] Step 1: Initialize the state variables of the designed residual generator and the optimal value of the cumulative loss function; set the decay coefficient, time window, number of batch updates, balance coefficient, strategy improvement probability, and iteration termination condition:

[0089] Step A1: The residual generator is a dimensionality reduction observer. Its state variables represent data features with dynamic characteristics (the state variables are abstract feature variables and have no physical meaning corresponding to measurable variables in the monitored system). Let's assume that the residual generator has two state variables z1 and z2, and initialize z1(0) = 0 and z2(0) = 0.

[0090] Step A2: Initialize the optimal value of the cumulative loss function is a constant of the order of 108, such as

[0091] Step A3: Set the attenuation coefficient γ of the weighted past-time error function in the cumulative loss function to a constant, satisfying 0<γ<1.

[0092] Step A4: Set the value of the time window size κ in the cumulative loss function to a positive integer, such as setting κ=50.

[0093] Step A5: Set the total number of parameter batch updates J to 10 2 ~10 3 A positive integer of magnitude, which can be adjusted according to actual conditions.

[0094] Step A6: Set the balance coefficient cr involved in parameter update. By adjusting its value, the speed of iterative learning and the performance of optimization can be adjusted.

[0095] Step A7: Set the probability of "policy improvement" and "policy exploration" in step 6 below, that is, the corresponding probabilities of policy improvement and policy exploration in reinforcement learning. Set the probability of policy improvement p e , satisfying 0<p e <1, then the probability of strategy exploration is 1-p e Specifically, p e The closer the value of p is to 1, the greater the probability of executing strategy improvement and the smaller the probability of executing strategy exploration. e The closer the value is to 0, the smaller the probability of executing strategy improvement and the greater the probability of executing strategy exploration.

[0096] Step A8: Set the iteration termination condition. Set the maximum number of loops, and terminate the execution when the maximum number of loops is exceeded.

[0097] Step 2: Collect process data under normal operating conditions and build an offline historical data demonstration dataset:

[0098] Step B1: Collect sensor output data from the controlled system over a period of time, denoted as y(k). y(k) is a column vector, where each element corresponds to a sensor output value (the variable measured by the sensor), and k represents a discrete moment in time, such as k = 1, 2, ..., 1000.

[0099] Step B2: Record the control instruction or actuator drive signal data at the corresponding moment, denoted as u(k). u(k) is a column vector, each element of which corresponds to a control instruction or actuator drive signal value.

[0100] Step 3: Construct two key dynamic subsystems that drive the residual signal, denoted as dynamic system No. 1 and dynamic system No. 2;

[0101] The input of dynamic system No. 1 is the sensor measurement value y(k), and the output is the first term of the residual generator, that is, Signal.

[0102] The input of dynamic system No. 2 is the control instruction u(k), and the output is the second term of the residual generator, that is, Signal.

[0103] The specific process includes the following steps:

[0104] Step C1: Based on whether the actual conditions of the controlled object mechanism model are known, the system construction operation is performed in two cases: If the controlled system can be modeled as the following linear dynamic system and its state space model (A, B, C, D) is known, then execute steps C2 to C4;

[0105]

[0106] y(k)=Cx(k)+Du(k)

[0107] Among them, x represents the state variables of the controlled system in the process industry, such as liquid volume, flow rate, liquid level, etc. y and u The meanings are the same as above; A, B, C, and D represent the system matrix, input matrix, output matrix, and direct transfer matrix in the state space equation, respectively. The elements in the matrix are obtained by calculating the known mechanism model.

[0108] If the state space model of the controlled system is unknown, then execute steps C5 to C7;

[0109] Step C2, design the observer feedback matrix L(θ), (L=L(θ), indicating that L is a feedback matrix that depends on the parameter θ) so that it satisfies |eig(AL(θ)C)|<1, where θ represents all the undetermined elements (undetermined parameters) in L(θ); |eig(AL(θ)C)| represents the modulus of the eigenvalues ​​of the matrix (AL(θ)C), that is, all the eigenvalues ​​of (AL(θ)C) are within the unit circle, thereby meeting the stability requirements.

[0110] Step C3: Calculate the left coprime decomposition matrix of the controlled object according to the following formula: And construct dynamic system No. 1, with z1(k) representing the state variables of dynamic system No. 1, A1, B1, C1, and D1 representing the system matrix, input matrix, output matrix, and direct transfer matrix of dynamic system No. 1 respectively:

[0111]

[0112] A1=A-LC, B1=-L, C1=C, D1=I

[0113] z1(k+1)=A1z1(k)+B1y(k)

[0114]

[0115] Step C4: Calculate the left coprime decomposition matrix of the controlled object according to the following formula: And construct dynamic system No. 2, with z2(k) representing the state variables of dynamic system No. 2, A2, B2, C2, and D2 representing the system matrix, input matrix, output matrix, and direct transfer matrix of dynamic system No. 2 respectively:

[0116]

[0117] A2=A-LC, B2=B-LD, C2=C, D2=D

[0118] z2(k+1)=A2z2(k)+B2u(k)

[0119]

[0120] Step C5: Randomly generate undetermined elements (undetermined parameters) in the observer feedback matrix The following formula is used to assign initial values ​​to A1, B1, C1, and D1 so that they meet the controllable standard form:

[0121]

[0122] Matrix B1 has elements, is a dimensionally compatible full parameter matrix, that is, all elements in the matrix are unknown parameters.

[0123] C1=[1 0...0]D1=[0...0]

[0124] Then construct dynamic system No. 1 as follows:

[0125] z1(k+1)=A1z1(k)+B1y(k)

[0126]

[0127] Step C6: Randomly generate undetermined elements (undetermined parameters) in the observer feedback matrix The following formula is used to assign initial values ​​to A2, B2, C2, and D2 so that they meet the controllable standard form:

[0128]

[0129] Matrix B2 has elements, is a dimensionally compatible full parameter matrix, that is, all elements in the matrix are unknown parameters.

[0130] C2=[1 0...0]D2=[0...0]

[0131] Then construct dynamic system No. 2 as follows:

[0132] z2(k+1)=A2z2(k)+B2u(k)

[0133]

[0134] Step C7: Collect all parameters of the No. 1 dynamic system and the No. 2 dynamic system to obtain θ = [θ *，1 θ *，2 ].

[0135] Step 4: Generate a residual signal corresponding to time k; the specific process includes the following steps:

[0136] Step D1: Use y(k) in the data set to drive the dynamic system No. 1 and calculate the signal

[0137] Step D2: Use u(k) in the data set to drive the dynamic system No. 2 and calculate the signal

[0138] Step D3: According to the formula Calculate the residual signal r s The value of (k).

[0139] Step 5: Offline residual evaluation is performed. The specific process includes the following steps:

[0140] Step E1: Calculate the current error function R(k) and the cumulative loss function V(k) in sequence:

[0141]

[0142]

[0143] Step E2, random batch update parameters θ: for j = 1, 2, ..., J, generate θ j =θ+c r ·randn j , where randn j represents the jth random number generated by the normal distribution; J is the total number of initialized batch updates, c r is the initialized balance coefficient.

[0144] Step E3, let k←k+1, repeat steps 3, 4 and 5, and step E1 to calculate the corresponding V j (k+1) numerical value, j=1, 2, ..., J.

[0145] Step E4: Calculate the optimal cumulative loss function:

[0146] V * (k)=R(k)+γ·min j {V j (k+1)|j∈(1,J)}

[0147] Step 6: Policy improvement and policy exploration; the specific process includes the following steps:

[0148] Step F1: In order to avoid falling into local extreme values, e For parameter θ * Perform the following optimal update: θ * =argmin θ V * (k); with probability 1-p e For parameter θ * Perform random updates (assign random values).

[0149] Step F2: Check the stability of the dynamic system. * All eigenvalues ​​of the system matrices A1 and A2 of the corresponding dynamic system No. 1 and dynamic system No. 2 are within the unit circle and satisfy Then make the following assignment: Store the corresponding optimal parameter θ * Otherwise, keep the original V * (k), θ * The value remains unchanged.

[0150] Step F3: If the termination condition set in step A8 in step 1 is not triggered, then repeat steps 3 to 6 in F2.

[0151] Optimizing the online deployment and application phase of the residual generator

[0152] Step 7: Collect the online data of the controlled system at the current moment, including the real-time measurement data of the sensors in the controlled system. online (k) and the control instruction data u output by the controller online (k).

[0153] Step 8. Generate the optimal residual value:

[0154] If the state space model of the controlled system is known (C2-C4 has been executed), then the optimal parameter θ * Substitute into L(θ) in step C3 and drive the dynamic system No. 1 constructed by C3 to obtain the output value By the optimal parameter θ * Substitute into L(θ) in step C4 and drive the dynamic system No. 2 constructed by C4 to obtain the output value

[0155] If the state space model of the controlled system is unknown (C5-C7 have been executed), the optimal parameter θ * Substitute the formula in step C5 to obtain the optimal system matrix of dynamic system No. 1 And drive dynamic system No. 1 to obtain output value By the optimal parameter θ * Substitute the formula in step C6 to obtain the optimal system matrix of dynamic system No. 2 And drive the dynamic system No. 2 to obtain the output value

[0156] Calculate online residual values Specific implementation method 2:

[0158] This embodiment is a computer storage medium, which stores at least one instruction. The at least one instruction is loaded and executed by a processor to implement the process corresponding to the offline design stage of optimizing the residual generator in the residual generator optimization design method of the dynamic control system in the process industry.

[0159] This embodiment emphasizes that the instruction emphasizes the offline design stage of implementing the residual generator, that is, it reflects the design process of the residual generator.

[0160] It should be understood that the instructions include computer program products, software, or computerized methods corresponding to any method described in the present invention; the instructions can be used to program a computer system or other electronic device. Computer storage media may include readable media on which instructions are stored, and may include but are not limited to magnetic storage media, optical storage media; magneto-optical storage media include read-only memory ROM, random access memory RAM, erasable programmable memory (e.g., EPROM and EEPROM) and flash memory layers, or other types of media suitable for storing electronic instructions. Specific implementation method three:

[0162] This embodiment is a computer storage medium, which stores at least one instruction. The at least one instruction is loaded and executed by a processor to implement the process corresponding to the online deployment and application stage of the optimized residual generator in the residual generator optimization design method of the dynamic control system in the process industry.

[0163] This embodiment emphasizes that the instructions emphasize the online deployment and application stage of the residual generator, that is, it reflects the designed application process of the residual generator.

[0164] It should be understood that the instructions include computer program products, software, or computerized methods corresponding to any method described in the present invention; the instructions can be used to program a computer system or other electronic device. Computer storage media may include readable media on which instructions are stored, and may include but are not limited to magnetic storage media, optical storage media; magneto-optical storage media include read-only memory ROM, random access memory RAM, erasable programmable memory (e.g., EPROM and EEPROM) and flash memory layers, or other types of media suitable for storing electronic instructions. Specific implementation method four:

[0166] This embodiment is an application device of a residual generator of a dynamic control system in a process industry. The device includes a processor and a memory. It should be understood that the device includes any device including a processor and a memory described in the present invention. The device may also include other units and modules that perform display, interaction, processing, control, and other functions through signals or instructions.

[0167] At least one instruction is stored in the memory, and the at least one instruction is loaded and executed by the processor to implement the application method of the residual generator of the dynamic control system in the process industry.

[0168] This embodiment emphasizes that the instructions emphasize the objects corresponding to the online deployment and application stages of the residual generator, that is, they reflect the objects for which the residual generator is designed to be applied, including but not limited to the monitoring system of the controlled object.

[0169] Example

[0170] The case study of the present invention is based on the Tennessee Eastman Chemical Process, which serves as a benchmark for process control and system monitoring. The computer simulation system simulates a complex chemical reaction flow consisting of modules such as feed, reaction tower, separator, and stripper. Figure 5 As shown in Table 1, Figure 5 This is a simplified block diagram of the Tennessee Eastman Chemical process, where the numbers correspond to the variables in Table 1.

[0171] The sensor measurement variables used in the embodiment, namely y(k), include the feed flow rate, pressure, liquid level, temperature, power, etc. at multiple locations in the above-mentioned modules, corresponding to numbers XMEANS(1) to XMEANS(20); the control command variable output by the controller used, namely u(k), is the opening of the stripping tower valve, corresponding to number XMV(9).

[0172] Table 1 Physical meaning of variables in the controlled system

[0173]

[0174]

[0175] According to the specific implementation of the present invention, the residual generator is optimized and designed, and online simulation verification is performed. The specific steps are as follows:

[0176] Offline design phase for optimizing the residual generator

[0177] Step 1: Set and initialize relevant parameters, including the state variables z1(0)=0, z2(0)=0 of the residual generator; the optimal value of the product loss function The decay coefficient γ in the cumulative loss function for the weighted error function of the past time is 0.8; the time window size k in the cumulative loss function is 50; the total number of parameter batch updates J is 400; the balance coefficient c involved in the parameter update is r =1; probability p of strategy improvement e =0.99; set the maximum number of cycles to 30 times.

[0178] Step 2: Collect process data under normal operating conditions and build a demonstration data set. The specific process includes the following steps:

[0179] Step B1: Collect the output values ​​of 20 sensors (i.e., XMEANS(1) to XMEANS(20)) at 480 moments, denoted as y(k) (k = 1, 2, ..., 480). y(k) is a 20-dimensional column vector, each element of which corresponds to the output value of a sensor.

[0180] Step B2: Record the controller output (i.e., XMV(9) variable) values ​​at the corresponding 480 moments, denoted as u(k). Since there is only one controller output variable, u(k) is a scalar.

[0181] Step 3: Construct two key dynamic subsystems that drive the residual signal, namely dynamic system 1 and dynamic system 2. The specific process includes the following steps:

[0182] Step C1: In this embodiment, the controlled object is a complex chemical reaction, which belongs to a black box situation with an unknown state space model and cannot be accurately modeled. Therefore, it is determined to execute steps C5 to C7 next.

[0183] Step C5: Randomly generate parameter θ *，1 , according to the generated random numbers, the initial values ​​of A1, B1, C1, and D1 are as follows:

[0184]

[0185] C1=[1 0 0]D1=[0 0 0]

[0186] And build dynamic system No. 1:

[0187] z1(k+1)=A1z1(k)+B1y(k)

[0188]

[0189] Step C6: Randomly generate parameter θ *，2 , the initial values ​​of A2, B2, C2, and D2 are assigned as follows based on the generated random numbers:

[0190]

[0191] B2 is a matrix of dimension 3*20, each element of which is a random number.

[0192] C2=[1 0 0]D1=[0 0 0]

[0193] And build dynamic system No. 2:

[0194] z2(k+1)=A2z2(k)+B2u(k)

[0195]

[0196] Step C7: Arrange all parameters θ=[θ *，1 θ *，2 ], is a 75-dimensional row vector.

[0197] Step 4: Generate a residual signal corresponding to time k; the specific process includes the following steps:

[0198] Step D1: Drive the dynamic system No. 1 by y(k) and calculate

[0199] Step D2: Use u(k) to drive the dynamic system No. 2 and calculate

[0200] Step D3: Calculate residuals

[0201] Step 5: Offline evaluation of residuals; the specific process includes the following steps:

[0202] Step E1: Calculate the error function at the current moment And the cumulative loss function

[0203] Step E2: For j = 1, 2, ..., 400, generate 400 update parameters θ j =θ+randn j , where randn j Represents the value of the jth random number generated according to the normal distribution.

[0204] Step E3, let k←k+1, repeat steps 3, 4 and 5, and calculate the corresponding cumulative loss function V j (k+1)(j=1, 2, ..., 400) numerical values.

[0205] Step E4: Calculate the optimal cumulative loss function V * (k)=R(k)+0.8·min j {V j (k+1)|j∈(1,400)}

[0206] Step 6: Policy improvement and policy exploration. The specific process includes the following steps:

[0207] Step F1: with probability p e = 0.99 for parameter θ * Perform optimal update: θ * =arg min θ V * (k); with probability 0.01 parameter θ * Perform random updates.

[0208] Step F2: If θ * All the corresponding eigenvalues ​​of A1 and A2 are within the unit circle and satisfy but Update storage θ * Otherwise, keep the original V * (k),θ * The value remains unchanged.

[0209] Step F3: If the number of cycles is less than or equal to 30, then repeat steps 3 to 6 (step F2).

[0210] Optimizing the online deployment and application phase of the residual generator

[0211] In this example, the online test consists of 960 moments. The first 160 moments are fault-free, and the last 800 moments are faulty. Specifically, the fault is a step fault in the "component feed ratio," a typical fault in publicly available benchmark test datasets.

[0212] Step 7: Collect the online data of the controlled system at the current moment, including the real-time measurement data of the sensor online (k) and control instruction data u online (k).

[0213] Step 8: Generate the optimal residual value. * Substitute the formula in step C5 to obtain the optimal system matrix of dynamic system No. 1 And drive dynamic system No. 1 to obtain output value By the optimal parameter θ * Substitute the formula in step C6 to obtain the optimal system matrix of dynamic system No. 2 And drive the dynamic system No. 2 to obtain the output value Calculate online residual values

[0214] Figure 2 is the unoptimized online residual generation result curve, Figure 3 is the optimized online residual generation result curve. Figure 2 and Figure 3 It can be seen that when the time range of normal operating conditions is 0<k≤160, both the optimized residual generator designed by the present invention and the traditional non-optimized residual generator have small fluctuations, which are caused by process noise and will not cause systematic false alarms.

[0215] However, when a step fault occurs, that is, when the time range is k>160, although the residual value generated by the traditional unoptimized residual generator has a certain amplitude change near the time of the fault, the mean of the residual signal is still near 0, and the amplitude will quickly return to a level equivalent to normal operating conditions, which will result in a small time window for fault detection, low fault detection sensitivity, and low overall fault detection rate.

[0216] In comparison, after a fault occurs, the residual generated by the optimization design method proposed in the present invention will significantly and continuously deviate from the baseline, which is consistent with the expected residual signal characteristics, and helps to expand the time window of fault detection, improve fault detection sensitivity, improve the comprehensive fault detection rate and other key performance.

[0217] Depend on Figure 4It can be seen that the value of the cumulative loss function is continuously optimized and reduced, and finally converges to near 0.

[0218] The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. A residual generator optimization design method for dynamic control systems in process industries, characterized by: The following steps are involved: Step 1: Initialize the state variables of the designed residual generator and the optimal value of the cumulative loss function; Set the decay coefficient, time window, number of batch updates, balance coefficient, strategy improvement probability, and iteration termination condition; Step 2: Collect process data under normal operating conditions and build an offline historical data demonstration dataset: Step B1: Collect sensor output data of the controlled system over a period of time, denoted as y(k); y(k) is a column vector, each element of which corresponds to the output value of a sensor, and k represents a discrete moment; Step B2: Record the control instruction or actuator drive signal data at the corresponding moment, denoted as u(k); u(k) is a column vector, each element of which corresponds to a control instruction or actuator drive signal value; Step 3: Construct two dynamic subsystems to drive the residual signal. The specific process includes the following steps: Step C1: Based on whether the actual conditions of the controlled object mechanism model are known, when the controlled system can be modeled as the following linear dynamic system and its state space model (A, B, C, D) is known, execute steps C2 to C4; y(k)=Cx(k)+Du(k) Where x represents the state variable of the controlled system in the process industry; A, B, C, and D represent the system matrix, input matrix, output matrix, and direct transfer matrix in the state space equation, respectively. The elements in the matrix are obtained by calculation using a known mechanism model. Step C2: Design the observer feedback matrix L(θ) so that it satisfies |eig(AL(θ)C)|<1, where θ represents all the undetermined elements in L(θ); |eig(AL(θ)C)| represents the modulus of the eigenvalue of the matrix (AL(θ)C); Step C3: Calculate the left coprime decomposition matrix of the controlled object according to the following formula: And build dynamic system No. 1, composed of Represents the state variables of dynamic system No. 1, A1, B1, C1, and D1 represent the system matrix, input matrix, output matrix, and direct transfer matrix of dynamic system No. 1, respectively: A1=A-LC,B1=-L,C1=C,D1=I Step C4: Calculate the left coprime decomposition matrix of the controlled object according to the following formula: And build dynamic system No. 2, composed of Represents the state variables of dynamic system No. 2, A2, B2, C2, and D2 represent the system matrix, input matrix, output matrix, and direct transfer matrix of dynamic system No. 2, respectively: A2=A-LC,B2=B-LD,C2=C,D2=D Step 4: Use y(k) in the data set to drive the dynamic system No. 1 and calculate the signal And the dynamic system No. 2 is driven by u(k) in the data set to calculate the signal According to the formula Calculate the residual signal r s (k); Step 5: Offline evaluation of the residuals, including the following steps: Step E1: Based on the residual signal r s (k) Calculate the current error function R(k) and the cumulative loss function V(k); Step E2, random batch update parameters θ: for j = 1, 2, ..., J, generate θ j =θ+c r ·randn j , where randn j represents the jth random number generated by the normal distribution, J is the total number of batch updates, c r is the balance coefficient; Step E3, let k←k+1, repeat steps 3, 4 and 5 of E1, and calculate the corresponding V j (k+1) value; Step E4: Calculate the optimal cumulative loss function: V * (k)=R(k)+γ·min j {V j (k+1)|j∈(1,J)}, γ is the attenuation coefficient; Step 6: Based on probability p e , using strategy improvement and strategy exploration to * =arg min θ V * (k) perform updates; Then judge if θ * Are all the eigenvalues ​​of the system matrices A1 and A2 of the corresponding dynamic system No. 1 and dynamic system No. 2 within the unit circle and satisfy is the optimal value of the cumulative loss function, then the following assignment is performed: Store the corresponding optimal parameter θ * Otherwise, keep the original V * (k),θ * The value remains unchanged; Repeat steps 3 to 6 until the termination condition is met.

2. The residual generator optimization design method for dynamic control systems in process industries according to claim 1 is characterized in that: In step 3, when the state space model of the controlled system is unknown, steps C5 to C7 are executed; Step C5: Randomly generate the undetermined elements in the observer feedback matrix The following formula is used to assign initial values ​​to A1, B1, C1, and D1 so that they meet the controllable standard form: Matrix B1 has elements, is a dimensionally compatible full parameter matrix, that is, all elements in the matrix are unknown parameters; C1=[1 0 … 0] D1=[0 … 0] Then construct dynamic system No. 1 as follows: Step C6: Randomly generate the undetermined elements in the observer feedback matrix The following formula is used to assign initial values ​​to A2, B2, C2, and D2 so that they meet the controllable standard form: Matrix B2 has elements, is a dimensionally compatible full parameter matrix, that is, all elements in the matrix are unknown parameters; C2=[1 0 … 0] D2=[0 … 0] Then construct dynamic system No. 2 as follows: Step C7: Collect all parameters of the No. 1 dynamic system and the No. 2 dynamic system to obtain θ = [θ *,1 θ *,2 ].

3. The residual generator optimization design method for a dynamic control system in a process industry according to claim 1 or 2, characterized in that: Current moment error function 4. The residual generator optimization design method for a dynamic control system in a process industry according to claim 3, characterized in that: Cumulative loss function 5. The residual generator optimization design method for dynamic control systems in process industries according to claim 4, characterized in that: Using strategy improvement and strategy exploration to * =arg min θ V * (k) in the process of updating, with probability p e For parameter θ * Perform optimal update θ * =arg min θ V * (k), with probability 1-p e For parameter θ * Perform random updates.

6. The residual generator optimization design method for dynamic control systems in process industries according to claim 5, characterized in that: The specific process of step 1 includes the following steps: Step A1: The residual generator is a dimension reduction observer, whose state variables represent data features with dynamic characteristics; the residual generator is assumed to have two state variables initialization Step A2: Initialize the optimal value of the cumulative loss function Step A3: Set the attenuation coefficient γ used to weight the past time error function in the cumulative loss function to a constant, 0<γ<1; Step A4: Set the value of the time window size κ in the cumulative loss function to a positive integer; Step A5: Set the total number of parameter batch updates J to a positive integer; Step A6: Set the balance coefficient c in parameter update r ; Step A7: Set the probability p of strategy improvement e , satisfying 0 <p e <1, then the probability of strategy exploration is 1-p e ; Step A8: Set the maximum number of loops, and terminate the execution when the maximum number of loops is exceeded.

7. A computer storage medium, characterized in that The storage medium stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the residual generator optimization design method for a dynamic control system in a process industry as described in any one of claims 1 to 6.

8. An application method of a residual generator for a dynamic control system in a process industry, characterized in that: The following steps are involved: Collect the online data of the controlled system at the current moment, including the real-time measurement data of the sensors in the controlled system online (k) and the control instruction data u output by the controller online (k); Based on the residual generator designed by the residual generator optimization design method of the dynamic control system in the process industry according to any one of claims 1 to 6, an optimal residual value is generated; using the optimal parameter θ * , based on the dynamic system No. 1, the output value is obtained Using the optimal parameter θ * , based on the dynamic system No. 2, the output value is obtained Calculate online residual values 9. A computer storage medium, characterized in that The storage medium stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the application method of the residual generator of the dynamic control system in the process industry as claimed in claim 8.

10. An application device of a residual generator for a dynamic control system in a process industry, characterized in that: The device includes a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the application method of the residual generator of the dynamic control system in the process industry as claimed in claim 8.

Images