A dual-model closed-loop self-learning control method and device and medium

Abstract

The application discloses a double-model closed-loop self-learning control method and device and a medium, and comprises the following steps: generating a predicted control variable by using a first time sequence neural network according to a current control target and historical time sequence data; determining an actual control variable based on the predicted control variable and a control variable correction value of a previous period and outputting the actual control variable to a controlled system; simultaneously, predicting a next-period output value by using a second time sequence neural network according to the actual control variable and the historical time sequence data; determining a control variable correction value of a current period based on a deviation between the next-period output value and an actual output value of the controlled system in the next period, and storing all time sequence data.

Description

Technical Field

[0001] This invention discloses a dual-model closed-loop self-learning control method, device, and medium, belonging to the field of closed-loop self-learning control technology. Background Technology

[0002] In the field of industrial process control, controlled systems such as large-diameter flow control valves, industrial furnaces, and chemical reactors generally exhibit complex characteristics such as strong nonlinearity, large transmission lag, and multivariable coupling. Traditional PID control is difficult to effectively cope with scenarios with large lag and significant time-varying characteristics, and is prone to continuous oscillations. Although model predictive control can handle multivariable coupling and large lag problems, it is highly dependent on accurate analytical mathematical models. When system characteristics drift due to factors such as equipment wear and medium changes, model identification and parameter tuning need to be re-performed, which is difficult to implement in engineering. Existing deep learning-based control schemes mostly adopt a single-model architecture, which is difficult to meet the dual requirements of industrial control for response speed and control accuracy at the same time, and lacks built-in adaptive capabilities, making it impossible to achieve steady-state convergence operation at the system level.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This invention provides a dual-model closed-loop self-learning control method, device, and medium to at least solve the technical problems of existing control schemes that rely on precise analytical models, have insufficient adaptive capabilities, and cannot simultaneously take into account response speed and control accuracy.

[0005] According to one aspect of the present invention, in order to achieve the above-mentioned objective, a dual-model closed-loop self-learning control method is provided, comprising:

[0006] Based on the current control objective and historical time series data, predictive control quantities are generated using the first time series neural network.

[0007] Based on the predicted control quantity and the control quantity correction value of the previous cycle, the actual control quantity is determined and output to the controlled system. At the same time, based on the actual control quantity and historical time series data, the output value of the next cycle is predicted using a second time series neural network.

[0008] Based on the deviation between the output value of the next cycle and the actual output value of the controlled system in the next cycle, the control quantity correction value for the current cycle is determined, and all timing data is stored.

[0009] Furthermore, based on the current control objective and historical time-series data, a predictive control quantity is generated using a first time-series neural network, including:

[0010] Obtain the current control objective and historical time series data from the time series database. The historical time series data shall include at least: environmental parameter sequence, target value sequence, control variable sequence, and output value sequence.

[0011] Based on the current control objective and historical time series data, a predicted time series is constructed, which includes environmental parameters, target values, control variables, and the current control objective as the last target value at historical moments.

[0012] Based on the predicted time series, a first time series neural network is used to generate predictive control variables.

[0013] Furthermore, the first temporal neural network is constructed using recurrent neural networks and their derivative variants, or liquid neural networks and their optimized variants;

[0014] The weight parameters of the first time-series neural network are used to calculate the predicted time series to obtain the predictive control quantity.

[0015] Furthermore, based on the predicted control quantity and the control quantity correction value of the previous period, the actual control quantity is determined, including:

[0016] When the current cycle is the initial cycle or the system is in a steady state, the control quantity correction value of the previous cycle is zero.

[0017] When the current period is not the initial period, the predicted control quantity is added to the control quantity correction value of the previous period to obtain the actual control quantity.

[0018] Furthermore, based on the actual control quantity and historical time-series data, a second time-series neural network is used to predict the output value of the next cycle, including:

[0019] Retrieve the time series data of environmental parameters, control variables, and outputs from the historical time series database to the current step.

[0020] The actual control quantity is used as the final control quantity to construct a forecast time series;

[0021] The weight parameters of the second time-series neural network are used to calculate the predicted time series to obtain the predicted output value for the next period.

[0022] The second temporal neural network is constructed using recurrent neural networks and their derivative variants, or liquid neural networks and their optimized variants.

[0023] Furthermore, based on the deviation between the output value of the next cycle and the actual output value of the controlled system in the next cycle, the control quantity correction value for the current cycle is determined, including:

[0024] Obtain the next cycle output value predicted by the second temporal neural network;

[0025] Obtain the actual output value of the controlled system in the next cycle after the actual control quantity is executed;

[0026] Calculate the deviation between the actual output value and the predicted output value;

[0027] Obtain the gradient of the output of the second temporal neural network with respect to the control quantity;

[0028] Based on the bias and gradient, the control quantity correction value for the current cycle is determined according to the preset learning rate, where the learning rate is greater than zero.

[0029] Furthermore, the method also includes an online training step:

[0030] Extract at least one period of historical time series from the time series database. The historical time series includes environmental parameter series, control target series, and actual control quantity series. Using the control quantity as a label, train the first time series neural network online and update the weight parameters of the first time series neural network.

[0031] Extract at least one period of historical time series from the time series database. The historical time series includes environmental parameter sequences, control quantity sequences, and actual output sequences. Use the actual output as a label to train the second time series neural network online and update the weight parameters of the second time series neural network.

[0032] The updated first and second temporal neural networks are then used to continue executing the predictive control loop.

[0033] Furthermore, the method also includes fault monitoring steps:

[0034] Obtain the predicted output value of the second temporal neural network in the previous cycle, and calculate the prediction residual of the second neural network with the actual output value in the current cycle;

[0035] Obtain the predicted control quantity and the actual control quantity generated by the first time-series neural network, and calculate the prediction residual of the first neural network;

[0036] When both the prediction residuals of the second neural network and the prediction residuals of the first neural network exceed their respective confidence interval thresholds based on statistical distributions, the system is deemed to have a fault.

[0037] When a fault is detected, the contribution of each environmental parameter feature to the prediction residual of the second neural network is calculated, and all contribution data are stored.

[0038] According to one embodiment of the present invention, a dual-model closed-loop self-learning control device is also provided, comprising:

[0039] The generation module is used to generate predictive control quantities based on the current control objective and historical time series data using a first time series neural network.

[0040] The determination module is used to determine the actual control quantity based on the predicted control quantity and the control quantity correction value of the previous cycle and output it to the controlled system. At the same time, based on the actual control quantity and historical time series data, it uses a second time series neural network to predict the output value of the next cycle.

[0041] The correction module is used to determine the control quantity correction value for the current cycle based on the deviation between the output value of the next cycle and the actual output value of the controlled system in the next cycle, and to store all timing data.

[0042] In this embodiment of the invention, a first temporal neural network directly generates a predictive control quantity as a feedforward control signal based on the control objective, achieving millisecond-level fast response. A second temporal neural network identifies the dynamic characteristics of the system online and performs gradient correction compensation on the control quantity based on the deviation between its predicted output and the actual output, effectively eliminating control errors introduced by model mismatch, system characteristic drift, and external disturbances. At the same time, all time-series data is continuously stored, accumulating real-time running samples for online updates of the dual models, enabling the system to continuously learn the latest dynamic characteristics of the controlled system during operation, gradually reducing the correction amplitude and eventually entering a self-learning convergent steady state. This closed-loop architecture enables the fast feedforward of the inverse model and the precise correction of the forward model to work together, ensuring both response speed and control accuracy without relying on the analytical physical model of the controlled system. When the training data covers all possible operating conditions and the model capacity is sufficient, the closed-loop system is stable from input to state, the output error is eventually bounded, and the final control accuracy can be maintained within the allowable range for a long time. Attached Figure Description

[0043] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0044] Figure 1 This is a flowchart of a dual-model closed-loop self-learning control method according to one embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of the architecture of a dual-model closed-loop self-learning control system according to one embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram illustrating the working principle of the inverse control model in a dual-model closed-loop self-learning control method according to one embodiment of the present invention.

[0047] Figure 4 This is a schematic diagram illustrating the working principle of the forward identification model in a dual-model closed-loop self-learning control method according to one embodiment of the present invention.

[0048] Figure 5 This is a schematic diagram illustrating the working principle of online model training in a dual-model closed-loop self-learning control method according to one embodiment of the present invention.

[0049] Figure 6 This is a schematic diagram illustrating the working principle of a fault monitoring module in a dual-model closed-loop self-learning control method according to one embodiment of the present invention.

[0050] Figure 7 This is a structural block diagram of a dual-model closed-loop self-learning control device according to one embodiment of the present invention. Detailed Implementation

[0051] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0052] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0053] According to embodiments of the present invention, a dual-model closed-loop self-learning control method embodiment is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer device such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0054] This method embodiment can be executed in an electronic device or similar computing device that includes a memory and a processor. Taking operation on an industrial valve as an example, the industrial valve may include one or more processors (processors may include, but are not limited to, central processing units (CPUs), graphics processing units (GPUs), digital signal processing (DSP) chips, microprocessors (MCUs), field-programmable gate arrays (FPGAs), neural network processors (NPUs), tensor processors (TPUs), artificial intelligence (AI) type processors, etc.) and memory for storing data. Optionally, the industrial valve may also include transmission devices, input / output devices, and display devices for communication functions. Those skilled in the art will understand that the above structural description is merely illustrative and does not limit the structure of the industrial valve. For example, the industrial valve may include more or fewer components than described above, or have a different configuration than described above.

[0055] The memory can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the dual-model closed-loop self-learning control method in this embodiment of the invention. The processor executes various functional applications and data processing by running the computer program stored in the memory, thereby realizing the aforementioned dual-model closed-loop self-learning control method. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the mobile terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0056] The transmission device is used to receive or send data via a network. Specific examples of the network mentioned above may include a wireless network provided by the mobile terminal's communication provider. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0057] Display devices can be, for example, touchscreen liquid crystal displays (LCDs) and touch displays (also referred to as "touchscreens" or "touch displays"). The LCD allows users to interact with the user interface of the mobile terminal. In some embodiments, the mobile terminal has a graphical user interface (GUI), which allows users to interact with the GUI through finger contact and / or gestures on a touch-sensitive surface. Optional human-computer interaction functions include: creating web pages, drawing, word processing, creating electronic documents, playing games, video conferencing, instant messaging, sending and receiving emails, call interfaces, playing digital video, playing digital music, and / or web browsing, etc. Executable instructions for performing the above human-computer interaction functions are configured / stored in one or more processor-executable computer program products or readable storage media.

[0058] Figure 1 This is a flowchart of a dual-model closed-loop self-learning control method according to one embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0059] Step S110: Based on the current control objective and historical time series data, generate predictive control quantities using the first time series neural network. The specific steps are as follows:

[0060] In step S110, at the beginning of each control cycle, the bidirectional timing control module first receives the current control objective from the upper-level control system. This control objective serves as the expected output value that the controlled system needs to achieve in this cycle and is the trigger condition for the entire predictive control process. Simultaneously, the timing database continuously stores all historical timing data generated during the past operation of the controlled system, including environmental parameters, control objectives, actual outputs, and actual control variables at each moment. This historical data records the dynamic response characteristics of the controlled system under different operating conditions and at different times, providing fundamental data support for the timing neural network to make predictions.

[0061] The first-order neural network, or inverse control model, is constructed using a time-order neural network. Optional architectures include recurrent neural networks and their derivatives (such as LSTM and GRU) as well as liquid neural networks and their optimized variants. This model undergoes offline pre-training before leaving the factory, learning the basic inverse dynamic characteristics of the controlled system, i.e., the mapping relationship between the desired output and the required control quantity. During actual operation, this model is responsible for inversely calculating the control quantity required to achieve the current control objective.

[0062] When generating predictive control inputs, the first temporal neural network first extracts historical time series data up to the previous cycle from the time series database. This series includes the values ​​of variables such as environmental parameters, control objectives, control inputs, and actual outputs at multiple historical moments. Then, the current control objective is used as the last target value and concatenated with the aforementioned historical time series to construct a complete predictive time series. This predictive time series not only includes the trajectory of the system's past operating state but also clearly defines the control expectation for the current cycle, providing sufficient contextual information for model inference.

[0063] Subsequently, the first temporal neural network uses its current weight parameters to perform calculations and inferences on the constructed predicted time series, directly outputting the predicted control quantity required to bring the controlled system close to the control objective in the current cycle. This predicted control quantity, as an open-loop feedforward control signal, can provide control decisions before the controlled system generates an actual response, thus effectively overcoming the large transmission lag problem commonly found in industrial processes and significantly improving the system's response speed. After calculation, the predicted control quantity is sent to the control output module as the basis for determining the actual control quantity in the current cycle.

[0064] Step S120: Based on the predicted control quantity and the control quantity correction value of the previous cycle, determine the actual control quantity and output it to the controlled system. At the same time, based on the actual control quantity and historical time series data, use the second time series neural network to predict the output value of the next cycle. The specific steps are as follows:

[0065] In step S120, after the predictive control quantity is generated, the control quantity output module immediately performs the operation of determining the actual control quantity. This module receives the predictive control quantity output from the inverse control model and retrieves the control quantity correction value stored in the previous cycle from the time series database. This correction value is calculated by the gradient correction module in the previous cycle based on the prediction deviation of the forward identification model, reflecting the control quantity compensation amount corresponding to the deviation between the actual output and the predicted output in the previous cycle after gradient backpropagation. The control quantity output module sums the predictive control quantity of the current cycle with the control quantity correction value of the previous cycle to obtain the actual control quantity applied to the controlled system in the current cycle. It is worth noting that when the current running cycle is the initial start-up cycle, since there is no running data from the previous cycle, the control quantity correction value of the previous cycle is directly set to zero. Similarly, when the system has entered a steady-state operating state, the gradient correction amplitude generated by the bidirectional time series control has been continuously close to zero, and the control quantity correction value of the previous cycle is also zero. The system mainly relies on the direct predictive output of the inverse control model to maintain control accuracy.

[0066] Once the actual control quantity is determined, the control quantity output module outputs it simultaneously in three paths: the first path is output to the controlled system as a command signal to drive the actuator, causing the valve, heater, or frequency converter and other actuators to adjust their working state according to the control quantity, so that the output of the controlled system gradually approaches the current control target; the second path is output to the time series database as historical control quantity data stored by timestamp, providing new samples for model inference and online training in subsequent cycles; the third path is sent to the second time series neural network as one of the input data for the model to predict the output of the next cycle.

[0067] Simultaneously, the second temporal neural network, the forward identification model, initiates predictive calculations for the controlled system's output in the next cycle. This model is also built upon a temporal neural network, with the same selectable architecture as the first temporal neural network, including recurrent neural networks and their derivatives, as well as liquid neural networks and their optimized variants. The forward identification model has undergone offline pre-training before leaving the factory, initially learning the forward dynamic characteristics of the controlled system, i.e., the mapping relationship between control inputs, environmental parameters, and system output. In each control cycle, the model extracts historical time series up to the current step from the time series database. This series contains the values ​​of variables such as environmental parameters, control inputs, and actual outputs at multiple historical moments. Based on this, the actual control input for the current cycle, just determined by the control input output module, is used as the last control input in this time series to construct a complete predictive time series. Subsequently, the forward identification model uses its current weight parameters to perform calculations and inferences on this predictive time series, outputting the predicted output value of the controlled system for the next cycle.

[0068] The predicted output value is an advance estimate of the controlled system's operating state in the next cycle. Its physical meaning is: if the currently determined actual control variable is applied to the controlled system, the output response the system will produce in the next cycle. This prediction result is then sent to the gradient correction module to compare with the actual output arriving in the next cycle, calculating the prediction deviation and generating a control variable correction value accordingly. Through this closed-loop structure of "control-prediction-correction," the participation of the forward identification model enables the system to predict the control effect in advance and store the gradient information needed for correction before the deviation actually occurs, thereby significantly improving the response accuracy and disturbance rejection capability of the control system.

[0069] Step S140: Based on the deviation between the output value of the next cycle and the actual output value of the controlled system in the next cycle, determine the control quantity correction value for the current cycle and store all timing data. The specific steps are as follows:

[0070] In step S140, after receiving the actual control quantity, the controlled system executes the control command starting from the current cycle k steps, and enters the next cycle, i.e., k+1 steps, after one complete control cycle. At this time, the controlled system generates an actual output value corresponding to the actual control quantity. This actual output value reflects the true response result under the combined effect of the current environmental parameters, the current state of the controlled system, and the applied actual control quantity. The controlled system feeds back this actual output value to the gradient correction module for deviation calculation and correction value generation.

[0071] Meanwhile, during the control output phase of the previous cycle (step k), the forward identification model has already calculated and output the predicted output value for the next cycle (step k+1) based on the historical time series data up to step k and the actual control quantity of the current cycle. This predicted output value is also transmitted to the gradient correction module as a benchmark for evaluating the control effect and calculating the control deviation. At this point, the gradient correction module possesses two key data points: the forward identification model's predicted output value for step k+1 in the previous cycle, and the actual output value for step k+1 fed back by the controlled system after actual operation. The gradient correction module subtracts the predicted output value from the actual output value to calculate the deviation. The sign and magnitude of this deviation directly reflect the degree of deviation between the actual control quantity applied in the current cycle and the ideal control quantity. A deviation of zero indicates precise control; a positive deviation indicates that the actual output is too high, i.e., the control quantity is too large; a negative deviation indicates that the actual output is too low, i.e., the control quantity is too small.

[0072] However, simply knowing the numerical value of the deviation is insufficient to determine the direction and magnitude of the control quantity correction; it is also necessary to know how changes in the control quantity will affect the system output. To this end, the gradient correction module further utilizes the differential properties of the forward identification model to obtain the gradient information of the model's predicted output with respect to the control quantity. The physical meaning of this gradient is that, near the current operating point, each unit change in the control quantity causes a change in the predicted output; its sign represents the direction of the output response when the control quantity is adjusted, and its magnitude reflects the system's sensitivity to changes in the control quantity. The gradient correction module combines the aforementioned deviation with the gradient information and calculates the control quantity correction value for the current cycle according to a preset learning rate. Specifically, when the actual output is higher than the predicted output, the deviation is positive, and the gradient correction module generates a correction value of corresponding polarity based on the gradient direction. This correction value will apply reverse compensation to the control quantity in subsequent control cycles, causing the actual output to move back towards the target value; when the actual output is lower than the predicted output, the deviation is negative, and the correction value exhibits the opposite polarity, driving the control quantity to adjust in the direction of increasing output. The value of the learning rate determines the magnitude of a single correction. Its setting needs to balance convergence speed and stability. If the learning rate is too large, it will cause overshoot or even oscillation. If the learning rate is too small, the convergence process will be too slow.

[0073] After the control quantity correction value is generated, the gradient correction module transmits it to two destinations. First, it is transmitted to the control quantity output module as the control quantity correction value from the previous cycle required to determine the actual control quantity in the next control cycle, enabling this correction value to participate in the calculation of the actual control quantity in the next cycle, thus forming a continuous closed-loop compensation mechanism for control deviation. Second, it is transmitted to the time series database for persistent storage according to the current cycle's timestamp. Simultaneously, all time series data generated during the operation of the bidirectional time series control in this cycle, including but not limited to the environmental parameters of the current cycle, the control objectives issued by the superior, the actual output fed back by the controlled system, the actual control quantity determined by the control quantity output module, the predicted control quantity generated by the inverse control model, the predicted output value generated by the forward identification model, and the control quantity correction value calculated by the gradient correction module, are all uniformly stored in the time series database. These time-series data are strictly arranged in chronological order and persistently stored, forming a complete digital record of the controlled system's operation. This not only provides continuous and reliable historical data support for the prediction and correction of bidirectional time-series control in each subsequent control cycle, but also accumulates the training sample set needed for updating the weight parameters of the dual models for the online model training module. Furthermore, it provides a data foundation for the fault monitoring module to perform residual analysis and contribution calculation. The aforementioned cycle begins with the issuance of the current control objective and sequentially goes through the stages of predictive control quantity generation, actual control quantity determination and output, forward prediction, deviation calculation and correction value generation, and data storage. It runs continuously and cyclically in each control cycle, forming a complete dual-model closed-loop self-learning control loop.

[0074] Based on steps S110 to S140 above, in this embodiment of the invention, a first time-series neural network directly generates a predictive control quantity as a feedforward control signal according to the control objective, achieving millisecond-level fast response; a second time-series neural network identifies the dynamic characteristics of the system online, and performs gradient correction compensation on the control quantity based on the deviation between its predicted output and actual output, effectively eliminating control errors introduced by model mismatch, system characteristic drift, and external disturbances; at the same time, all time-series data are continuously stored, accumulating real-time running samples for online updates of the dual models, enabling the system to continuously learn the latest dynamic characteristics of the controlled system during operation, gradually reducing the correction amplitude and eventually entering a self-learning convergent steady state; this closed-loop architecture enables the fast feedforward of the inverse model and the precise correction of the forward model to work together, ensuring both response speed and control accuracy without relying on the analytical physical model of the controlled system. When the training data covers all possible operating conditions and the model capacity is sufficient, the closed-loop system is stable from input to state, the output error is eventually bounded, and the final control accuracy can be maintained within the allowable range for a long time.

[0075] The method of an embodiment of the present invention generates a predictive control quantity using a first time-series neural network based on a current control target and historical time-series data, including: acquiring the current control target and historical time-series data in a time-series database, wherein the historical time-series data includes at least: an environmental parameter sequence, a target value sequence, a control quantity sequence, and an output value sequence; constructing a predictive time series based on the current control target and historical time-series data, including environmental parameters, target values, control quantities, and the current control target as the last target value at historical moments; and generating a predictive control quantity using the first time-series neural network based on the predictive time series.

[0076] Furthermore, the first time-series neural network is constructed using a recurrent neural network and its derivative variants, or a liquid neural network and its optimized variants; the weight parameters of the first time-series neural network are used to calculate the predicted time series to obtain the predicted control quantity.

[0077] Furthermore, based on the predicted control quantity and the control quantity correction value of the previous cycle, the actual control quantity is determined, including: when the current cycle is the initial cycle or the system is in a steady state, the control quantity correction value of the previous cycle is zero; when the current cycle is not the initial cycle, the predicted control quantity is added to the control quantity correction value of the previous cycle to obtain the actual control quantity.

[0078] Furthermore, based on the actual control quantity and historical time series data, a second time series neural network is used to predict the output value of the next cycle, including: obtaining the time series of environmental parameters, control quantity, and output from the historical time series database to the current step; using the actual control quantity as the control quantity of the last step to construct the prediction time series; and using the weight parameters of the second time series neural network to calculate the prediction time series to obtain the predicted output value of the next cycle; wherein, the second time series neural network is constructed using a recurrent neural network and its derivative variants, or a liquid neural network and its optimized variants.

[0079] Furthermore, based on the deviation between the output value of the next cycle and the actual output value of the controlled system in the next cycle, the control quantity correction value for the current cycle is determined, including: obtaining the output value of the next cycle predicted by the second temporal neural network; obtaining the actual output value of the controlled system in the next cycle after executing the actual control quantity; calculating the deviation between the actual output value and the predicted output value; obtaining the gradient of the output of the second temporal neural network with respect to the control quantity; and determining the control quantity correction value for the current cycle based on the deviation and the gradient, according to a preset learning rate, wherein the learning rate is greater than zero.

[0080] Furthermore, the method also includes an online training step: extracting at least one period of historical time series from a time series database, the historical time series including environmental parameter sequences, control target sequences, and actual control quantity sequences; using the control quantity as a label, training a first time series neural network online and updating the weight parameters of the first time series neural network; extracting at least one period of historical time series from a time series database, the historical time series including environmental parameter sequences, control quantity sequences, and actual output sequences; using the actual output as a label, training a second time series neural network online and updating the weight parameters of the second time series neural network; and continuing to execute the predictive control loop using the updated first and second time series neural networks.

[0081] Furthermore, the method also includes a fault monitoring step: obtaining the predicted output value of the second temporal neural network in the previous cycle and calculating the prediction residual of the second neural network with the actual output value in the current cycle; obtaining the predicted control quantity generated by the first temporal neural network and the actual control quantity, and calculating the prediction residual of the first neural network; when both the prediction residual of the second neural network and the prediction residual of the first neural network exceed their respective confidence interval thresholds based on statistical distributions, it is determined that there is a fault in the system; when a fault is determined to exist, the contribution of each environmental parameter feature to the prediction residual of the second neural network is calculated, and all contribution data are stored.

[0082] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0083] This invention also provides a dual-model closed-loop self-learning control device for implementing the above embodiments and preferred embodiments, which will not be repeated hereafter. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0084] According to one embodiment of the present invention, a dual-model closed-loop self-learning control system, such as Figure 2-6 As shown, the system consists of bidirectional timing control (0), timing database (5), online model training (6), and fault monitoring module (7).

[0085] The bidirectional timing control (0) consists of an inverse control model (1), a forward identification model (3), a control output module (2), and a gradient correction module (4).

[0086] The reverse control model (1) is constructed using a time-series neural network, which includes recurrent neural networks (RNN) and their derivative variants, liquid neural networks (LNN) and their optimized variants;

[0087] The forward identification model (3) is constructed using a time-series neural network, which includes recurrent neural networks (RNNs) and their derivatives, liquid neural networks (LNNs) and their optimized variants. The reverse control model (1) and the forward identification model (3) can dynamically adjust the length of the time window, the model structure and the control strategy according to the time dimension characteristics of the controlled system. The time dimension includes long-time dimensions from minutes to hours and high-speed dimensions from milliseconds to microseconds. When the controlled system is a long-time dimension system, a long time window is configured. Long-time dimension systems include large-diameter flow regulating valves, large storage tank level control systems, industrial furnace temperature control systems and chemical reactor pressure control systems.

[0088] The inverse control model (1) and the forward identification model (3) employ recurrent neural networks (RNNs) or their variants, including LSTM, GRU, or Transformer temporal neural networks, to capture the long-term lag characteristics of the system. The neural network learns the system's impulse response and lag patterns through its internal control structure and cell states, directly compensating for system transmission lag. During system operation, the online model training (6) continuously updates the model parameters, adapting to the slow drift of system characteristics caused by equipment wear and medium changes. The neural network learns the system's impulse response and lag patterns through its internal control structure and cell states, directly compensating for system transmission lag. During system operation, the online model training (6) continuously updates the model parameters, adapting to the slow drift of system characteristics caused by equipment wear and medium changes.

[0089] When the inverse control model (1) and the forward identification model (3) adopt a liquid neural network, the liquid neural network adopts a continuous fluid cell (CfC) architecture, which reduces the computational load through sparse connection structure and continuous time dynamic characteristics, and adapts to the low-latency inference requirements of high-speed dimensional scenarios. The lightweight inverse control model (1) and forward identification model (3) are deployed on hardware acceleration platforms, including FPGA or ASIC, so that the single inference latency of the model is controlled within 1ms, which meets the real-time requirements of millisecond / microsecond-level high-speed control. The inverse control model (1) and the forward identification model (3) adopt an offline pre-training + online fine-tuning mode; most of the training process is completed in the offline environment, and only inference and low learning rate parameter fine-tuning are performed in the online running stage. The online learning rate is less than .

[0090] The two-way timing control (0) is configured to receive signals from the upper-level control system at step k. And transmit it to the time series database (5);

[0091] The inverse control model (1) is configured to extract the time series of environmental parameters x, target y, and control variable u from the time series database (5) up to step k-1. Target As the last target value, construct the predicted time series. ,use calculate get And send it to the control output module (2);

[0092] The control output module (2) is configured to calculate: Obtain actual control The results are then output to the controlled system(s), the forward identification model (3), and the time series database (5). The control quantity is adjusted for the previous cycle. If k=0 or the system is in steady state, the control quantity is adjusted accordingly. ;

[0093] The forward identification model (3) is configured to extract the time series of environmental parameters x, control quantity u, and output y from the time series database (5) up to k steps, and then convert the actual control quantity into the time series data. As the final step, the control variable u is used to construct the predicted time series. ,calculate Obtaining predictive control results And send it to the gradient correction module (4);

[0094] The two-way timing control (0) is configured to use the actual control quantity in the controlled system (s). When executing from step k to step k+1, the actual output of step k+1 will be... The parameters are passed to the gradient correction module (4) and environmental parameters are obtained from the controlled system (S). Passed to the time series database (5);

[0095] The gradient correction module (4) is configured to receive the actual output. Then, calculate and predict the output. Deviation between Calculate the gradient Obtain control quantity correction in k+1 steps. (in: >0 indicates the learning rate. The control quantity is transmitted to the control quantity output module (2) as the control quantity correction value for step k+1, and simultaneously stored in the timing database (5);

[0096] The time-series database (5) is configured to continuously collect storage environment parameters x and control targets. Actual output Actual control quantity Control quantity correction ;

[0097] The bidirectional timing control (0) is configured to repeat the above loop.

[0098] The online training of the model (6) is configured to perform the following online training update steps:

[0099] Step 1: Online training of the model (6) extracts historical time series of length T with at least one past period from the time series database (5). The tag is The inverse control model (1) was trained online to obtain... (Model neuron weights) and update the original ;

[0100] Step 2: Online training of the model (6), extracting historical time series of at least one past period and length T from the time series database (5). The tag is The forward recognition model (3) is trained online to obtain (Model neuron weights) and update the original ;

[0101] Step 3, the inverse control model (1) and forward identification model (3) of the bidirectional timing control (0) part use a new Perform a predictive control loop.

[0102] Step 4: Repeat the above loop to continuously update the weight parameters of the dual models.

[0103] The fault monitoring module (7) is configured to perform the following fault judgment steps:

[0104] Step 1: While the bidirectional timing control (0) part is running, the prediction results of the forward identification model (3) in period k-1 are... , and the actual output of k cycles According to the formula Calculate the predicted residuals ; Reverse control model (1) according to formula Calculate the predicted residuals ;

[0105] Step 2, when the residual and A fault can be identified when all residuals exceed their respective confidence interval thresholds based on statistical distributions; among which... Used to improve the accuracy of fault diagnosis and reduce the false alarm rate, but not involved in subsequent fault source tracing calculations; the confidence interval threshold based on statistical distribution is a 3σ confidence interval threshold;

[0106] Step 3: Only when a fault is determined, the fault monitoring module (7) calculates the contribution of each feature to the predicted residual: in The weight of the influence of the i-th feature on the residual in the k-th period is given, and all Stored in a time-series database (5), contribution Used to locate fault-related features;

[0107] Step four, the fault monitoring module (7) analyzes all features within period T in the time series database (5). Perform differential calculations to obtain the rate of change. The baseline threshold for fault determination is dynamically and statistically generated through a fixed-length historical sliding time window, independent of absolute steady state; from direction The system comprehensively judges the fault type based on magnitude and can adapt to the continuous drift of system characteristics.

[0108] Step 5: Repeat the above steps to continuously monitor the system's operating status.

[0109] Figure 7 A dual-model closed-loop self-learning control device according to one embodiment of the present invention includes:

[0110] The generation module 201 is used to generate predictive control quantities based on the current control target and historical time series data using a first time series neural network.

[0111] The determination module 202 is used to determine the actual control quantity based on the predicted control quantity and the control quantity correction value of the previous cycle and output it to the controlled system. At the same time, based on the actual control quantity and historical time series data, it uses a second time series neural network to predict the output value of the next cycle.

[0112] The correction module 203 is used to determine the control quantity correction value for the current cycle based on the deviation between the output value of the next cycle and the actual output value of the controlled system in the next cycle, and to store all timing data.

[0113] It should be noted that the above modules can be implemented by software or hardware. For the latter, they can be implemented in the following ways, but are not limited to: all the above modules are located in the same processor; or, the above modules are located in different processors in any combination.

[0114] According to one embodiment of the present invention, an electronic device is also provided, comprising: a memory storing an executable program; and a processor for running the program, wherein the program executes the above-described dual-model closed-loop self-learning control method during runtime.

[0115] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0116] Step S1: Based on the current control objective and historical time series data, generate predictive control quantities using the first time series neural network;

[0117] Step S2: Based on the predicted control quantity and the control quantity correction value of the previous cycle, determine the actual control quantity and output it to the controlled system. At the same time, based on the actual control quantity and historical time series data, use the second time series neural network to predict the output value of the next cycle.

[0118] Step S3: Based on the deviation between the output value of the next cycle and the actual output value of the controlled system in the next cycle, determine the control quantity correction value for the current cycle and store all timing data.

[0119] According to one embodiment of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the storage medium is located to execute the above-described dual-model closed-loop self-learning control method.

[0120] Optionally, in this embodiment, the storage medium may be configured to store a computer program for performing the following steps:

[0121] Step S1: Based on the current control objective and historical time series data, generate predictive control quantities using the first time series neural network;

[0122] Step S2: Based on the predicted control quantity and the control quantity correction value of the previous cycle, determine the actual control quantity and output it to the controlled system. At the same time, based on the actual control quantity and historical time series data, use the second time series neural network to predict the output value of the next cycle.

[0123] Step S3: Based on the deviation between the output value of the next cycle and the actual output value of the controlled system in the next cycle, determine the control quantity correction value for the current cycle and store all timing data.

[0124] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0125] According to one embodiment of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the above-described dual-model closed-loop self-learning control method.

[0126] Optionally, in this embodiment, the above-mentioned computer program product can be configured as a computer program that performs the following steps:

[0127] Step S1: Based on the current control objective and historical time series data, generate predictive control quantities using the first time series neural network;

[0128] Step S2: Based on the predicted control quantity and the control quantity correction value of the previous cycle, determine the actual control quantity and output it to the controlled system. At the same time, based on the actual control quantity and historical time series data, use the second time series neural network to predict the output value of the next cycle.

[0129] Step S3: Based on the deviation between the output value of the next cycle and the actual output value of the controlled system in the next cycle, determine the control quantity correction value for the current cycle and store all timing data.

[0130] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0131] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0132] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between units or modules may be electrical or other forms.

[0133] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0134] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0135] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This 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 of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0136] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A dual-model closed-loop self-learning control method, characterized in that, include: Based on the current control objective and historical time series data, predictive control quantities are generated using the first time series neural network. Based on the predicted control quantity and the control quantity correction value of the previous cycle, the actual control quantity is determined and output to the controlled system. At the same time, based on the actual control quantity and historical time series data, the output value of the next cycle is predicted using a second time series neural network. Based on the deviation between the output value of the next cycle and the actual output value of the controlled system in the next cycle, the control quantity correction value for the current cycle is determined, and all timing data is stored.

2. The dual-model closed-loop self-learning control method according to claim 1, characterized in that, Based on the current control objective and historical time-series data, the predicted control quantity is generated using a first time-series neural network, including: Obtain the current control target and historical time series data from the time series database, wherein the historical time series data includes at least: environmental parameter sequence, target value sequence, control variable sequence, and output value sequence; Based on the current control objective and the historical time series data, a predicted time series is constructed, which includes environmental parameters, target values, control variables, and the current control objective as the last target value at historical moments. Based on the predicted time series, the predicted control quantity is generated using the first time-series neural network.

3. The dual-model closed-loop self-learning control method according to claim 2, characterized in that, The first temporal neural network is constructed using a recurrent neural network and its derivative variants, or a liquid neural network and its optimized variants; The predicted control quantity is obtained by calculating the predicted time series using the weight parameters of the first time-series neural network.

4. The dual-model closed-loop self-learning control method according to claim 1, characterized in that, Based on the predicted control quantity and the control quantity correction value of the previous cycle, the actual control quantity is determined, including: When the current cycle is the initial cycle or the system is in a steady state, the control quantity correction value of the previous cycle is zero; When the current period is not the initial period, the predicted control quantity is added to the control quantity correction value of the previous period to obtain the actual control quantity.

5. The dual-model closed-loop self-learning control method according to claim 1, characterized in that, Based on the actual control quantity and historical time-series data, the next cycle output value is predicted using a second time-series neural network, including: Retrieve the time series data of environmental parameters, control variables, and outputs from the historical time series database to the current step. The actual control quantity is used as the final control quantity to construct a prediction time series; The predicted time series is calculated using the weight parameters of the second time-series neural network to obtain the predicted output value for the next period. The second temporal neural network is constructed using a recurrent neural network and its derivative variants, or a liquid neural network and its optimized variants.

6. The dual-model closed-loop self-learning control method according to claim 1, characterized in that, Based on the deviation between the output value of the next cycle and the actual output value of the controlled system in the next cycle, the control quantity correction value for the current cycle is determined, including: Obtain the next cycle output value predicted by the second temporal neural network; Obtain the actual output value of the controlled system in the next cycle after executing the actual control quantity; Calculate the deviation between the actual output value and the predicted output value; Obtain the gradient of the output of the second temporal neural network with respect to the control quantity; Based on the deviation and the gradient, the control quantity correction value for the current cycle is determined according to a preset learning rate, wherein the learning rate is greater than zero.

7. The dual-model closed-loop self-learning control method according to claim 1, characterized in that, The method also includes an online training step: Extract at least one period of historical time series from the time series database. The historical time series includes environmental parameter sequences, control target sequences, and actual control quantity sequences. Using the control quantities as labels, train the first time series neural network online and update the weight parameters of the first time series neural network. Extract at least one period of historical time series from the time series database. The historical time series includes environmental parameter sequences, control quantity sequences, and actual output sequences. Use the actual output as a label to train the second time series neural network online and update the weight parameters of the second time series neural network. The predictive control loop continues to be executed using the updated first and second temporal neural networks.

8. The dual-model closed-loop self-learning control method according to claim 1, characterized in that, The method also includes a fault monitoring step: Obtain the predicted output value of the second temporal neural network in the previous cycle, and calculate the prediction residual of the second neural network with the actual output value in the current cycle; Obtain the predicted control quantity and the actual control quantity generated by the first temporal neural network, and calculate the prediction residual of the first neural network; When both the prediction residual of the second neural network and the prediction residual of the first neural network exceed their respective confidence interval thresholds based on statistical distribution, the system is determined to be faulty. When a fault is determined to exist, the contribution of each environmental parameter feature to the prediction residual of the second neural network is calculated, and all contribution data are stored.

9. A dual-model closed-loop self-learning control device, characterized in that, include: The generation module is used to generate predictive control quantities based on the current control objective and historical time series data using a first time series neural network. The determination module is used to determine the actual control quantity based on the predicted control quantity and the control quantity correction value of the previous cycle and output it to the controlled system. At the same time, based on the actual control quantity and historical time series data, it uses a second time series neural network to predict the output value of the next cycle. The correction module is used to determine the control quantity correction value for the current cycle based on the deviation between the output value of the next cycle and the actual output value of the controlled system in the next cycle, and to store all timing data.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored executable program, wherein, when the executable program is executed, it controls the device where the storage medium is located to perform the dual-model closed-loop self-learning control method according to any one of claims 1 to 8.

Images