An industrial system identification method and system based on switch mechanism and conjugate gradient
By combining the SWITCH mechanism with conjugate gradients, and utilizing a dynamic threshold criterion switching update strategy based on gradient norm and change, the problems of model oscillation and computational resource waste in conjugate gradient identification methods during industrial processes are solved, achieving efficient and stable system identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- QINGDAO UNIV OF SCI & TECH
- Filing Date
- 2026-06-26
- Publication Date
- 2026-07-24
AI Technical Summary
Existing conjugate gradient identification methods suffer from problems such as model oscillation due to fixed update step size, waste of computational resources, and insufficient robustness in industrial processes, especially in non-stationary noise environments and fluctuating operating conditions.
By combining the SWITCH mechanism with conjugate gradients, and using a dynamic threshold criterion of gradient norm and gradient change to switch update strategies, combined with momentum smoothing and conjugate gradient updates, adaptive parameter updates are achieved.
It improves identification accuracy, convergence stability, and computational efficiency, enhances adaptability to changes in operating conditions and external disturbances, reduces computational overhead, and improves the reliability of the identification method.
Smart Images
Figure CN122449967A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial process control and system identification technology, and in particular to an industrial system identification method and system based on the SWITCH mechanism and conjugate gradient. Background Technology
[0002] Industrial system identification is a fundamental technology in the field of process control and optimization. Its core objective is to construct mathematical models that accurately reflect the dynamic characteristics of industrial processes using input and output data. Typical industrial processes exhibit complex characteristics such as nonlinearity, time-varying nature, large time delays, and strong coupling. System identification methods must balance convergence accuracy and computational efficiency. Traditional methods such as recursive least squares and gradient descent inherently conflict between steady-state tracking capability and computational load. The conjugate gradient (CG) algorithm, due to its quadratic convergence characteristics, has been introduced into the field of system identification. It can achieve rapid parameter optimization without directly calculating the Hessian matrix, alleviating computational pressure to some extent, but it still has inherent limitations.
[0003] Existing conjugate gradient identification methods exhibit the following shortcomings in industrial applications: First, the algorithms employ fixed update step size rules, lacking adaptability to non-stationary noise environments and operating condition fluctuations. When industrial processes experience load changes, external disturbances, or setpoint adjustments, the fixed update strategy easily leads to oscillations in weight estimation and frequent jumps in the identification model, directly affecting the stability of the control system and the quality of the final product. Second, the conjugate gradient algorithm has high computational complexity. Each iteration requires recalculating the conjugate direction scaling factor, search step size, and gradient information. Even in the steady-state phase where model parameters are close to convergence, the algorithm continues to run with the same high computational overhead, resulting in a serious waste of controller computing resources and limiting its application in industrial control systems with strict real-time requirements. Third, existing methods lack effective convergence state discrimination mechanisms, making it difficult to distinguish between true convergence states and pseudo-convergence caused by noise interference, resulting in insufficient algorithm robustness.
[0004] How to solve the above-mentioned technical problems is the challenge facing this invention. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides an industrial system identification method and system based on the SWITCH mechanism and conjugate gradient, which achieves a synergistic improvement in identification accuracy, convergence stability, and computational efficiency.
[0006] The technical solution adopted by this invention to solve its technical problem is as follows: This invention provides an industrial system identification method based on the SWITCH mechanism and conjugate gradient, comprising the following steps: S1. Collect the input and output signals of the industrial process model to be identified, and recursively update the correlation matrix of the input signals and the cross-correlation vector between the input and output signals based on the input and output signals. S2. Calculate the gradient vector based on the correlation matrix and cross-correlation vector, and calculate the gradient norm and gradient change based on the gradient vector; S3. Based on the gradient norm and gradient change, combined with a predefined dynamic threshold criterion, determine whether the SWITCH freezing mechanism is triggered, and execute a differentiated update strategy for the model parameters according to the determination result. The differentiated update strategy includes a first update strategy and a second update strategy; the SWITCH freezing mechanism is used to switch between the first update strategy and the second update strategy during iteration. When it is determined that the SWITCH freezing mechanism is not triggered, the first update strategy is adopted; when it is determined that the SWITCH freezing mechanism is triggered, the second update strategy is adopted. S4. Repeat steps S1 to S3 until the preset convergence condition is met to complete the industrial system identification.
[0007] Preferably, the calculation formula for recursively updating the correlation matrix of the input signal and the cross-correlation vector between the input signal and the output signal based on the input signal and the output signal in step S1 is expressed as follows:
[0008] in, For the correlation matrix, For cross-correlation vectors, For the current moment, For input signal, Forgetting factor, For the desired signal, This is the transpose symbol.
[0009] Preferably, the gradient vector formula in step S2 is expressed as follows:
[0010] in, For filter tap weights; The gradient norm is calculated using the Euclidean norm; the gradient change is obtained by subtracting the gradient vectors at adjacent time points.
[0011] Preferably, the determination of whether the SWITCH freeze mechanism is triggered in step S3 specifically includes: The dynamic threshold at the current moment is calculated based on the gradient norm and gradient change, combined with a predefined dynamic threshold criterion. The gradient norm at the current time is compared with the dynamic threshold at the current time, and the gradient change at the current time is compared with the preset minimum change threshold. If the gradient norm at the current time is less than the dynamic threshold at the current time, and the gradient change at the current time is less than the minimum change threshold, the SWITCH freeze mechanism is triggered; otherwise, the SWITCH freeze mechanism is not triggered. The calculation formula for the predefined dynamic threshold criterion is expressed as follows:
[0012] in: This represents the dynamic threshold at the current moment. Represents the smoothing factor. Indicates the threshold scaling factor; This represents the historical mean of the gradient norm at a recent point in time.
[0013] Preferably, the first update strategy is a conjugate gradient update, specifically including: Based on the gradient vector at the current moment and the gradient vector at the previous moment, the conjugate direction scaling factor at the current moment is calculated. The formula for calculating the conjugate direction scaling factor is as follows:
[0014] in, The scaling factor for the conjugate direction. Let be the gradient vector at the current time. Let be the gradient vector from the previous time step. To prevent regularization constants with denominators of zero, It is the transpose symbol; The conjugate direction vector at the current moment is recursively updated based on the conjugate direction scaling factor at the current moment, resulting in the updated conjugate direction vector, as expressed by the following formula:
[0015] in, This represents the conjugate direction vector of the previous time step. This is the conjugate direction vector at the current moment; The search step size is calculated based on the correlation matrix and the updated conjugate direction vector, and a constraint is imposed on the step size. The formula is as follows:
[0016] in, The preset maximum step size factor, To prevent regularization constants with denominators of zero, This is the search step size; The filter tap weights are updated using the updated conjugate direction vector and search step size, as expressed by the following formula:
[0017] in, The filter tap weights for the next time step; This represents the filter tap weight at the current moment.
[0018] Preferably, the second update strategy is: pausing the conjugate gradient update and fine-tuning the filter tap weights using a momentum smoothing formula; the momentum smoothing formula is shown below:
[0019] in, This is the momentum factor.
[0020] Preferably, the preset convergence condition includes: If the gradient norm at the current time is less than the dynamic threshold at the current time, and the gradient change at the current time is less than the minimum change threshold, and this condition is met for a preset number of consecutive times, then the model parameters are determined to have entered the steady-state convergence interval.
[0021] This invention also provides an industrial system identification system based on the SWITCH mechanism and conjugate gradient, comprising: The recursive update module is used to collect the input and output signals of the industrial process model to be identified, and recursively update the correlation matrix of the input signal and the cross-correlation vector between the input and output signals based on the input and output signals. The computation execution module is used to calculate the gradient vector based on the correlation matrix and cross-correlation vector, and to calculate the gradient norm and gradient change based on the gradient vector; The trigger determination and model update module is used to determine whether to trigger the SWITCH freeze mechanism based on the gradient norm and gradient change amount, combined with predefined dynamic threshold criteria, and to execute a differentiated update strategy for the model parameters according to the determination result. The iterative execution module is used to repeatedly execute steps S1 to S3 until the preset convergence condition is met, thus completing the industrial system identification.
[0022] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described industrial system identification method based on the SWITCH mechanism and conjugate gradient.
[0023] The present invention also provides a computer storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described industrial system identification method based on the SWITCH mechanism and conjugate gradient.
[0024] The beneficial effects of this invention are as follows: it achieves a synergistic improvement in identification accuracy, convergence stability, and computational efficiency. A SWITCH freezing mechanism is set up, and the system switches between a first update strategy and a second update strategy based on the determination results of the gradient norm and gradient change. This allows the system to perform full iterative updates of the conjugate gradient during the non-convergent parameter phase and fine-tuning updates during the near-convergent parameter phase, achieving adaptive adjustment of the parameter update process and improving the adaptability of industrial process identification to changes in operating conditions and external disturbances. Simultaneously, the introduction of the gradient norm and gradient change as dual determination criteria effectively distinguishes between the true convergence state and the pseudo-convergence state caused by noise interference, avoiding update strategy switching errors caused by misjudgment of a single indicator, and further improving the reliability of the identification method in industrial high-noise environments. In the first update strategy, a conjugate direction scaling factor with a max correction formula is introduced for adaptive correction, combined with a search step size with a saturation constraint mechanism, effectively suppressing the excessive amplification of the step size by large noise in the industrial environment and avoiding parameter oscillations caused by load changes or external disturbances. In the second update strategy, a momentum smoothing update strategy is adopted to fine-tune the model parameters, which reduces the repeated calculation process in the convergence phase and reduces the computational cost of the online identification process. Attached Figure Description
[0025] Figure 1 This is a diagram illustrating the method steps of the present invention.
[0026] Figure 2 This is a system module diagram of the present invention.
[0027] Figure 3 This is a comparison curve of the mean square deviation changes of different identification methods under high noise or large disturbance conditions in Embodiment 3 of the present invention.
[0028] Figure 4 This is a comparison curve of the mean square deviation changes of different identification methods in Embodiment 3 of the present invention under low noise or stable operating conditions.
[0029] Figure 5 This is a diagram of the internal structure of a computer device according to Embodiment 4 of the present invention. Detailed Implementation
[0030] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.
[0031] Example 1: See Figure 1 As shown, this embodiment is an industrial system identification method based on the SWITCH mechanism and conjugate gradient, including the following steps: S1. Collect the input and output signals of the industrial process model to be identified, and recursively update the correlation matrix of the input signals and the cross-correlation vector between the input and output signals based on the input and output signals. It should be noted that this step is used to construct the statistical information required for online identification of industrial process models, providing a data foundation for subsequent gradient vector calculation, dynamic threshold determination, and model parameter updates. The industrial process model to be identified can be a distillation column, reaction vessel, heat exchanger, distillation unit, or other industrial process objects with dynamic response characteristics.
[0032] Specifically, during the operation of an industrial process, process input signals and process output signals are acquired in real time. The input signals are used to characterize the control quantities acting on the industrial process object, and the output signals are used to characterize the response results of the industrial process object.
[0033] Furthermore, based on the input signal vector acquired at the current moment, the correlation matrix of the input signal and the cross-correlation vector between the input and output signals are recursively updated. The correlation matrix characterizes the degree of correlation between different dimensions of the input signal and reflects the statistical characteristics of the industrial process input data. The cross-correlation vector characterizes the relationship between the input and output signals, reflecting the degree of influence of changes in the industrial process input on the output response. As new input data continuously enters, the conjugate gradient method is used to recursively update the correlation matrix and cross-correlation vector, as shown in the following formula:
[0034] in, For the correlation matrix, For cross-correlation vectors, For the current moment, For input signal, Forgetting factor, This is the desired signal.
[0035] It should be noted that the forgetting factor The forgetting factor is determined through industrial system identification and simulation testing. In this embodiment, the forgetting factor is... The value is set to 0.95. This value primarily considers the continuity of industrial process data, requiring both current sampling data and recent historical data to be included in the identification calculation. For the exponentially weighted recursive structure, the forgetting factor... A value of 0.95 corresponds to an equivalent memory length of approximately 20 sampling periods, thus simultaneously preserving the statistical value of recent data and enhancing the ability to track changes in operating conditions. If If the value is too small, the recursive result is easily affected by single noise or instantaneous disturbance; if If the value is too large, the historical data will have too high a weight, and the model parameter adjustment speed will be reduced when the system operating conditions change.
[0036] Furthermore, after obtaining new input and output signals at each sampling moment, the correlation matrix and cross-correlation vector are updated synchronously, so that the correlation matrix continuously records the statistical characteristics of the input data and the cross-correlation vector continuously records the correlation characteristics between the input and output, thereby forming a statistical data basis suitable for online system identification.
[0037] It should be noted that by continuously updating the correlation matrix and cross-correlation vector, dynamic maintenance of industrial process statistics can be achieved without storing all historical sampling data, thereby reducing data storage requirements and online computing load. Furthermore, since the correlation matrix and cross-correlation vector can reflect the data characteristics of the industrial process under current operating conditions in real time, they can provide a reliable basis for the accurate calculation of gradient vectors in subsequent steps, improving the adaptability of the industrial system's identification process to operating condition fluctuations and noise interference.
[0038] S2. Calculate the gradient vector based on the correlation matrix and cross-correlation vector, and calculate the gradient norm and gradient change based on the gradient vector; After updating the correlation matrix and cross-correlation vector in step S1, this step is performed. It should be noted that this step calculates the gradient information at the current time. This gradient information reflects the sensitivity of the parameters of the industrial process model to the output error under the current estimate, and is used for subsequent dynamic threshold determination and parameter update strategy selection, providing a core basis for the SWITCH mechanism trigger judgment.
[0039] Specifically, the input signal correlation matrix obtained by recursively updating in step S1 is... and the cross-correlation vector between the input and output signals This data serves as a representation of the system's statistical information at the current moment. It is combined with filter tap weights. Calculate the gradient vector Its definition is as follows:
[0040] It should be noted that the direction of the gradient vector reflects the steepest descent direction of the parameter adjustment, and the magnitude of the gradient vector reflects the degree to which the current model parameters deviate from the ideal identification state.
[0041] Furthermore, to obtain the overall magnitude information of the gradient vector, the gradient norm is calculated. The Euclidean norm (norm 2) is typically used. It should be noted that the gradient norm is used to quantitatively characterize the degree to which the current parameters deviate from the convergence state. The larger the norm, the more severe the deviation; the smaller the norm, the more the parameters gradually tend towards a steady state.
[0042] It should also be noted that, in order to reflect the trend of gradient change over time and to calculate the amount of gradient change, it is defined as the gradient vector at the current moment. gradient vector from the previous time step The difference norm:
[0043] The gradient change is used to reflect the characteristics of the convergence speed during the parameter update process of an industrial process model. It can reveal whether the system is in a steady state or a rapid adjustment phase, thus avoiding misjudgments caused by relying solely on the instantaneous gradient norm.
[0044] Furthermore, to support the construction of subsequent dynamic threshold criteria, the gradient norm sequence of the most recent time points should be stored, and its mean should be calculated as a reference for the adaptive threshold. It should be noted that the sequence length can be set according to the dynamic characteristics of the industrial process, typically ranging from 5 to 20 sampling times.
[0045] Furthermore, the gradient vector, gradient norm, gradient change, and historical average gradient output in this step will be directly used in step S3 to construct a dynamic threshold criterion, thereby determining the current model parameter update status: when both the gradient norm and gradient change are large, the system is in a rapid adjustment state; when the gradient norm decreases and the gradient change is stable, the system tends to converge and can trigger the SWITCH freezing mechanism.
[0046] It should be noted that this step only relies on the correlation matrix and cross-correlation vector maintained in step S1, as well as the gradient information of the previous time step. It does not require saving all historical input and output data, has low computational complexity, and can meet the real-time requirements of online identification of industrial processes.
[0047] S3. Based on the gradient norm and gradient change, combined with a predefined dynamic threshold criterion, determine whether the SWITCH freezing mechanism is triggered, and execute a differentiated update strategy for the model parameters according to the determination result. The differentiated update strategy includes a first update strategy and a second update strategy; the SWITCH freeze mechanism is used to switch between the first update strategy and the second update strategy during iteration. When it is determined that the SWITCH freeze mechanism is not triggered, the first update strategy is used; when it is determined that the SWITCH freeze mechanism is triggered, the second update strategy is used. After calculating the gradient vector, gradient norm, gradient change, and gradient history information at the current time in step S2, proceed to this step.
[0048] It should be noted that this step is used to adaptively determine whether to continue full iterative updates at the current moment based on the convergence state of the industrial process model parameters, thereby reducing unnecessary computational overhead and improving the stability and real-time performance of the online identification process while ensuring the system's identification accuracy.
[0049] Specifically, based on the gradient norm sequence stored in step S2 for the most recent time moments, the historical mean of the gradient at the current time moment is calculated, and a dynamic threshold is constructed based on the historical mean of the gradient. The dynamic threshold is used to characterize the allowable gradient fluctuation range of the current industrial process model under the corresponding operating conditions, and its calculation method is as follows:
[0050] in: This represents the dynamic threshold at the current moment. Represents the smoothing factor. Indicates the threshold scaling factor; This represents the historical mean of the gradient norm at a recent point in time.
[0051] It should be noted that the smoothing factor With threshold scaling factor All were determined through industrial system identification and simulation tests. In this embodiment, the smoothing factor... Set to 0.98, threshold scaling factor Set to 0.001. The dynamic threshold is determined by the threshold at the previous time step and the average gradient norm of the most recent 10 sampling times. Smoothing factor. A value of 0.98 is used to make the threshold change more gradual, avoiding frequent threshold changes due to measurement noise or single gradient fluctuations, thereby improving the stability of the freeze criterion. Threshold scaling factor Used to control the trigger sensitivity of the SWITCH freeze mechanism. Experiments show that when When the value is too large, it is easy to enter the frozen state prematurely. If the value is too small, it will delay the freeze trigger and increase computational overhead. Therefore, this embodiment selects... =0.001 is a typical value, which keeps the freeze criterion relatively strict, and the SWITCH freeze mechanism is only triggered when the gradient norm has dropped to a low level.
[0052] It should be noted that the historical average gradient is obtained by statistically analyzing the gradient norm within a preset time window, and it reflects the overall convergence level of the industrial process at the current operating stage. By introducing the historical average gradient to construct a dynamic threshold, the threshold can be dynamically adjusted as the system identification process changes, avoiding the problem that a fixed threshold is difficult to adapt to different operating conditions.
[0053] Furthermore, after obtaining the dynamic threshold, the gradient norm at the current moment is compared with the dynamic threshold, and the gradient change at the current moment is compared with the preset minimum change threshold to determine whether the current system has entered a convergent state.
[0054] The specific judgment criteria are as follows:
[0055] and
[0056] in: This represents the minimum change threshold.
[0057] It should be noted that the gradient norm is used to characterize the degree of deviation between the current model parameters and the target identification result, while the gradient change is used to characterize the degree of change in the parameter adjustment trend between consecutive sampling times. When the gradient norm is less than the dynamic threshold, it indicates that the current model parameters are close to the optimal identification result; when the gradient change is less than the minimum change threshold, it indicates that the model parameter changes tend to stabilize. In this embodiment, the minimum gradient change threshold... Take 3 × 10^ 4. At 30dB, the noise power is approximately 10^30 of the effective signal power. 3. The corresponding noise amplitude ratio is 0.0316. To avoid misjudging noise disturbances as valid gradient changes, the minimum gradient change threshold is set to 0.01 times the noise amplitude ratio. When the current gradient norm is less than the dynamic threshold, and the gradient change in adjacent time steps is less than 3 × 10^ At time 4, it is considered that the changes in model parameters have become relatively small, and the SWITCH freeze mechanism can be triggered; otherwise, the conjugate gradient method is used for full update. This setting does not rely solely on the gradient magnitude for judgment, but considers both the gradient magnitude and the gradient change trend, which can reduce false triggering caused by noise.
[0058] It should also be noted that, to avoid misjudgments caused by measurement noise, operating condition fluctuations, or instantaneous disturbances during industrial process operation, this invention introduces both the gradient norm and gradient change as dual judgment criteria. The gradient norm reflects the degree of convergence at the current moment, while the gradient change reflects the stability of the convergence trend. The combined judgment of these two factors effectively distinguishes between true convergence and false convergence caused by noise, thereby improving the accuracy and reliability of the SWITCH freeze mechanism trigger judgment.
[0059] Furthermore, this step selects to implement different differential parameter update strategies: If any of the above conditions is not met, the switch freeze mechanism is deemed not to have been triggered, and a full iterative update of the industrial process model parameters is performed using a preset first update strategy. Specifically, the first update strategy is a complete conjugate gradient (CG) adaptive full iterative update to ensure that the parameters can quickly approximate the ideal model. Its specific execution method is as follows: First, the conjugate direction scaling factor at the current moment is calculated using the gradient vector at the current moment and the conjugate search direction vector at the previous moment. Its calculation formula is strictly limited to:
[0060] in, To prevent the regularization constant (which is a small positive number) from having a denominator of zero, we introduce... The modified formula in this embodiment constructs a conjugate gradient adaptive adjustment parameter, which can preserve the conjugateness of the direction and automatically provide direction correction in non-stationary noise environment, effectively preventing the oscillation of weight direction.
[0061] Subsequently, the current conjugate search direction is updated recursively based on the conjugate direction scaling factor. :
[0062] Among them, at the initial moment of the first execution of the algorithm ( The initial direction is set to the steepest descent direction, i.e. .
[0063] Next, using the correlation matrix Perform a quadratic projection calculation with the current direction to adaptively update the search step size. To prevent excessive step size amplification under high industrial noise conditions from causing estimation divergence, a saturation constraint mechanism is introduced for step size updates:
[0064] in, The preset maximum step size factor, To prevent regularization constants with denominators of zero.
[0065] It should be noted that the maximum step size factor The value was determined through industrial system identification and simulation testing, and in this embodiment, it is set to 0.1. This parameter is used to limit the maximum step size during the conjugate gradient update process, preventing the filter tap weights from oscillating due to excessively large search step sizes when there is high noise or significant gradient fluctuations.
[0066] Finally, based on the updated conjugate search direction and the constrained step size, the filter tap weights are updated in time sequence:
[0067] in, The filter tap weights for the next time step; This represents the filter tap weight at the current moment.
[0068] When the above conditions are met simultaneously, the SWITCH freeze mechanism is triggered, and the industrial process model parameters are fine-tuned using a preset second update strategy. Specifically, when the SWITCH freeze state is entered, it indicates that the system is close to convergence. In order to completely skip the highly complex conjugate gradient matrix operations and step size recalculation, the system suspends the execution of the first update strategy and instead adopts a low-computational-overhead momentum smoothing update strategy to update only the filter tap weights. The specific mathematical recursive equation is as follows:
[0069] in, The momentum factor, typically ranging from 0.9 to 0.98, is used to establish a fine balance between parameter smoothing and stability and drift tracking capability under slow operating conditions. In this embodiment, the momentum factor... The momentum factor is set to 0.95, and the search step size is decreased by 0.9. At this point, the system no longer repeats the complete calculation of the conjugate direction and step size, but instead performs small, smooth updates based on the existing parameters. The momentum factor is determined through industrial system identification simulation tests. Setting the step size decay factor to 0.95 retains 95% of the historical parameter information while introducing 5% new updates, thus reducing computational complexity while maintaining the ability to track slow-moving parameters. Setting the step size decay factor to 0.9 allows the fine-tuning amount during the freezing phase to gradually decrease, reducing parameter jitter during the steady-state phase.
[0070] It should be noted that the second update strategy does not recalculate the conjugate direction scaling factor, recursively derive new conjugate search directions, or re-execute step size optimization calculations. Instead, it utilizes the direction and step size information already obtained before triggering the SWITCH freeze mechanism, combined with the momentum smoothing mechanism, to complete the fine-tuning of model parameters.
[0071] It should be noted that this step ultimately outputs the state determination result corresponding to the SWITCH freezing mechanism, and passes the state determination result to the subsequent parameter update process to determine whether the first update strategy or the second update strategy is used to complete the industrial process model parameter update at the current moment, so as to provide a basis for state switching for realizing online identification of industrial systems.
[0072] S4. Repeat steps S1 to S3 until the preset convergence condition is met to complete the industrial system identification.
[0073] After updating the model parameters at the current moment in step S3, this step is initiated. It should be noted that this step controls the iterative execution logic of the entire system's identification process, determines whether the preset convergence conditions are met, and decides whether to continue iterating or terminate the identification process based on the determination result, ultimately outputting the online identification model of the industrial process.
[0074] Specifically, let the time index Increasing, that is The process then repeats steps S1 to S3. It should be noted that the above iterative process is driven by the sampling period. At each new sampling moment, new input and output signals are acquired, statistical information is updated recursively, gradients are calculated, the switch state is determined, and model parameters are updated, thereby achieving continuous tracking of the dynamic characteristics of the industrial process.
[0075] Furthermore, the preset convergence conditions include: if the gradient norm at the current time is less than the dynamic threshold at the current time, and the gradient change at the current time is less than the minimum change threshold, and this condition is met for a preset number of consecutive times, then the model parameters are determined to have entered the steady-state convergence interval.
[0076] Furthermore, when the preset convergence condition is met, the iteration terminates, and the model parameter vector at the current moment is output as the final online identification model. This model can be directly used for industrial process control tasks such as predictive control, state estimation, and fault diagnosis. If the convergence condition is not yet met, the time index continues to increment, and steps S1 to S3 are repeated until the convergence condition is triggered. Throughout the online operation, the system continuously collects new process data and recursively updates the model parameters, enabling the identification model to adaptively track the time-varying characteristics of the industrial process.
[0077] It should be noted that through the closed-loop iteration of steps S1 to S4, this invention can achieve online, adaptive, and low-computational-overhead industrial system identification. Specifically, the SWITCH freezing mechanism in step S3 effectively avoids redundant calculations after convergence, and the convergence condition in step S4 ensures that the identification process terminates in a timely manner after reaching the expected accuracy, thereby minimizing computational resource consumption while ensuring model accuracy.
[0078] At this point, step S4 completes the loop control and output of the entire identification process, and the industrial system identification method of this embodiment of the invention has been fully executed.
[0079] Example 2: See Figure 2 As shown, this embodiment is an industrial system identification system based on the SWITCH mechanism and conjugate gradient, including: The recursive update module is used to collect the input and output signals of the industrial process model to be identified, and recursively update the correlation matrix of the input signal and the cross-correlation vector between the input and output signals based on the input and output signals. The computation execution module is used to calculate the gradient vector based on the correlation matrix and cross-correlation vector, and to calculate the gradient norm and gradient change based on the gradient vector; The trigger determination and model update module is used to determine whether to trigger the SWITCH freeze mechanism based on the gradient norm and gradient change amount, combined with predefined dynamic threshold criteria, and to execute a differentiated update strategy for the model parameters according to the determination result. The iterative execution module is used to repeatedly execute steps S1 to S3 until the preset convergence condition is met, thus completing the industrial system identification.
[0080] Example 3: See Figure 3 and Figure 4 This embodiment illustrates the identification performance of the industrial system identification method based on the SWITCH mechanism and conjugate gradient described in this invention under different noise or disturbance conditions. Using industrial processes with dynamic response characteristics, such as distillation columns, as the objects to be identified, and employing the same input signals, output signals, and initial parameter conditions, simulations are performed to compare the method of this invention (SWITCH-CG method) with the LMS method, the improved CG method, and the PCG method. In the figure, the horizontal axis represents the number of iterations, and the vertical axis represents the mean square deviation (MSD) in decibels. The lower the value of the curve, the smaller the parameter estimation error of the identification model, and the higher the system identification accuracy.
[0081] In this embodiment, the method of the present invention is executed according to steps S1 to S4 of Embodiment 1. Specifically, the input signal correlation matrix and input-output cross-correlation vector are updated recursively first, then the gradient vector, gradient norm, and gradient change are calculated, and a dynamic threshold criterion is used to determine whether to trigger the SWITCH freezing mechanism. When the system is not close to convergence, the first update strategy is used to perform a full update of the conjugate gradient; when the gradient norm and gradient change simultaneously meet the threshold condition, the SWITCH freezing mechanism is triggered, and the second update strategy is used for momentum smoothing fine-tuning. In the comparison methods, the LMS method uses a fixed step size update, the improved CG method uses a conventional improved conjugate gradient update but does not introduce the SWITCH freezing criterion, and the PCG method uses a preconditional conjugate gradient update.
[0082] Depend on Figure 3 It is evident that under conditions of high noise or large disturbances, the error of each method decreases rapidly with increasing iterations in the initial stage, indicating that each method possesses a certain degree of identification capability. However, the LMS method stabilizes at a relatively high MSD level after convergence, exhibiting significant error fluctuations. The PCG method shows improvement over the LMS method, but its steady-state curve remains higher than that of the method described in this invention. The improved CG method further reduces steady-state error, but still exhibits some oscillations under disturbance conditions. The SWITCH-CG method of this invention maintains a relatively fast convergence speed in the initial stage, and its curve remains at a lower position after entering steady state with smaller fluctuations, indicating that it has better anti-interference capability and steady-state identification accuracy under conditions of high noise or large disturbances.
[0083] Depend on Figure 4 As can be seen, under low-noise or stable operating conditions, all methods can converge, but their steady-state performance differs. The LMS method, limited by a fixed step size, has a relatively high steady-state MSD; the PCG method has better steady-state performance than the LMS method, but the curve still exhibits periodic fluctuations; the improved CG method achieves lower steady-state error, but it still exhibits oscillations in local iteration intervals. The SWITCH-CG method of this invention maintains the lowest MSD level in the steady-state stage and has minimal curve fluctuations, indicating that the dynamic threshold criterion and freezing mechanism can suppress invalid updates after the model is close to convergence, avoiding repeated parameter adjustments due to minor noise disturbances.
[0084] comprehensive Figure 3 and Figure 4 The comparison results show that the method of this invention achieves adaptive identification of the convergence state by jointly determining the gradient norm and the gradient change; by switching between full conjugate gradient update and momentum smoothing fine-tuning through the SWITCH freezing mechanism, it balances the initial convergence speed with the disturbance resistance stability in the steady state stage. Compared with the LMS method, the improved CG method, and the PCG method, the method of this invention exhibits lower steady-state MSD, smaller error fluctuations, and a more stable identification process under different noise conditions, thus verifying the effectiveness and practicality of this invention in online identification scenarios in industrial systems.
[0085] Example 4: This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0086] This computer device can be a server, and its internal structure diagram can be as follows: Figure 5 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores server data. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements an industrial system identification method based on the SWITCH mechanism and conjugate gradients.
[0087] Those skilled in the art will understand that Figure 5The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0088] Example 5: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0089] If the functions implemented by the method are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art or the current technical solution, can be embodied in the form of a software product. This current computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0090] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.
[0091] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.
[0092] The technical features of this invention not described can be implemented by or using existing technology, and will not be repeated here. Of course, the above description is not a limitation of this invention, and this invention is not limited to the examples above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention should also be within the protection scope of this invention.
Claims
1. An industrial system identification method based on the SWITCH mechanism and conjugate gradient, characterized in that, Includes the following steps: S1. Collect the input and output signals of the industrial process model to be identified, and recursively update the correlation matrix of the input signals and the cross-correlation vector between the input and output signals based on the input and output signals. S2. Calculate the gradient vector based on the correlation matrix and cross-correlation vector, and calculate the gradient norm and gradient change based on the gradient vector; S3. Based on the gradient norm and gradient change, combined with a predefined dynamic threshold criterion, determine whether the SWITCH freezing mechanism is triggered, and execute a differentiated update strategy for the model parameters according to the determination result. The differentiated update strategy includes a first update strategy and a second update strategy; The SWITCH freezing mechanism is used to switch between the first update strategy and the second update strategy during the iteration. When it is determined that the SWITCH freezing mechanism is not triggered, the first update strategy is adopted; when it is determined that the SWITCH freezing mechanism is triggered, the second update strategy is adopted. S4. Repeat steps S1 to S3 until the preset convergence condition is met to complete the industrial system identification.
2. The industrial system identification method based on the SWITCH mechanism and conjugate gradient as described in claim 1, characterized in that, The calculation formulas for recursively updating the correlation matrix of the input signal and the cross-correlation vector between the input and output signals in step S1 are expressed as follows: in, For the correlation matrix, For cross-correlation vectors, For the current moment, For input signal, Forgetting factor, For the desired signal, This is the transpose symbol.
3. The industrial system identification method based on the SWITCH mechanism and conjugate gradient as described in claim 2, characterized in that, The gradient vector formula in step S2 is expressed as follows: in, For filter tap weights; The gradient norm is calculated using the Euclidean norm; the gradient change is obtained by subtracting the gradient vectors at adjacent time points.
4. The industrial system identification method based on the SWITCH mechanism and conjugate gradient as described in claim 3, characterized in that, The determination of whether the SWITCH freeze mechanism is triggered in step S3 specifically includes: The dynamic threshold at the current moment is calculated based on the gradient norm and gradient change, combined with a predefined dynamic threshold criterion. The gradient norm at the current time is compared with the dynamic threshold at the current time, and the gradient change at the current time is compared with the preset minimum change threshold. If the gradient norm at the current time is less than the dynamic threshold at the current time, and the gradient change at the current time is less than the minimum change threshold, the SWITCH freeze mechanism is triggered; otherwise, the SWITCH freeze mechanism is not triggered. The calculation formula for the predefined dynamic threshold criterion is expressed as follows: in: This represents the dynamic threshold at the current moment. Represents the smoothing factor. Indicates the threshold scaling factor; This represents the historical mean of the gradient norm at a recent point in time.
5. The industrial system identification method based on the SWITCH mechanism and conjugate gradient according to claim 4, characterized in that, The first update strategy is conjugate gradient update, specifically including: Based on the gradient vector at the current moment and the gradient vector at the previous moment, the conjugate direction scaling factor at the current moment is calculated. The formula for calculating the conjugate direction scaling factor is as follows: in, The scaling factor for the conjugate direction. Let be the gradient vector at the current time. Let be the gradient vector from the previous time step. To prevent regularization constants with denominators of zero, It is the transpose symbol; The conjugate direction vector at the current moment is recursively updated based on the conjugate direction scaling factor at the current moment, resulting in the updated conjugate direction vector, as expressed by the following formula: in, This represents the conjugate direction vector of the previous time step. This is the conjugate direction vector at the current moment; The search step size is calculated based on the correlation matrix and the updated conjugate direction vector, and a constraint is imposed on the step size. The formula is as follows: in, The preset maximum step size factor, To prevent regularization constants with denominators of zero, This is the search step size; The filter tap weights are updated using the updated conjugate direction vector and search step size, as expressed by the following formula: in, The filter tap weights for the next time step; This represents the filter tap weight at the current moment.
6. The industrial system identification method based on the SWITCH mechanism and conjugate gradient as described in claim 5, characterized in that, The second update strategy is to pause the conjugate gradient update and fine-tune the filter tap weights using a momentum smoothing formula; the momentum smoothing formula is shown below: in, This is the momentum factor.
7. The industrial system identification method based on the SWITCH mechanism and conjugate gradient as described in claim 6, characterized in that, The preset convergence conditions include: If the gradient norm at the current time is less than the dynamic threshold at the current time, and the gradient change at the current time is less than the minimum change threshold, and this condition is met for a preset number of consecutive times, then the model parameters are determined to have entered the steady-state convergence interval.
8. An industrial system identification system based on the SWITCH mechanism and conjugate gradient, characterized in that, The steps for performing the industrial system identification method based on the SWITCH mechanism and conjugate gradient as described in any one of claims 1 to 7 include: The recursive update module is used to collect the input and output signals of the industrial process model to be identified, and recursively update the correlation matrix of the input signal and the cross-correlation vector between the input and output signals based on the input and output signals. The computation execution module is used to calculate the gradient vector based on the correlation matrix and cross-correlation vector, and to calculate the gradient norm and gradient change based on the gradient vector; The trigger determination and model update module is used to determine whether to trigger the SWITCH freeze mechanism based on the gradient norm and gradient change amount, combined with predefined dynamic threshold criteria, and to execute a differentiated update strategy for the model parameters according to the determination result. The iterative execution module is used to repeatedly execute steps S1 to S3 until the preset convergence condition is met, thus completing the industrial system identification.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the industrial system identification method based on the SWITCH mechanism and conjugate gradient as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the industrial system identification method based on the SWITCH mechanism and conjugate gradient as described in any one of claims 1 to 7.