Intelligent control system and method

Abstract

The invention relates to the field of computers, and discloses an intelligent control system and method, and the system comprises a sensor module which is used for collecting the system state data of a controlled object and the external disturbance information of the controlled object in real time; the DCRNN module is connected with the sensor module and used for predicting the state of the controlled object according to the system state data and the external disturbance information to obtain the predicted state of the controlled object or a preliminary control decision for controlling the controlled object; performing global optimal solution search on the prediction state or the preliminary control decision through an ant colony optimization algorithm to obtain an optimized control instruction; and the actuator module is used for executing the optimized control instruction so as to control the controlled object through the optimized control instruction. The problem of high-precision robust control of a complex system under the conditions of large lag, nonlinearity and strong disturbance is solved.

Description

Technical Field

[0001] This application relates to the field of computers, and more particularly to an intelligent control system and method. Background Technology

[0002] In complex systems such as intelligent manufacturing, smart agriculture, autonomous driving, and microgrids, the controlled objects generally exhibit characteristics such as high nonlinearity, large time lag, strong coupling of multiple variables, and environmental uncertainty. Traditional PID or model predictive control (MPC) relies on accurate mechanistic models or linearization assumptions, making it difficult to simultaneously ensure modeling accuracy, real-time performance, and robustness. When the system order increases, operating conditions change frequently, or external disturbances are severe, parameter tuning becomes difficult, overshoot / undershoot is serious, and energy consumption remains high.

[0003] In recent years, deep learning methods (CNN, RNN, Transformer, etc.) have been introduced into the field of control, attempting to replace mechanistic models with their powerful nonlinear mapping capabilities. However, single deep networks still have the following shortcomings:

[0004] 1. Limited receptive field makes it difficult to capture long-range spatial and temporal data;

[0005] 2. Training is prone to getting stuck in local minima and is sensitive to hyperparameters;

[0006] 3. Limited online adaptive capability, with performance plummeting when faced with model drift and sudden disturbances.

[0007] On the other hand, metaheuristic optimization algorithms such as Ant Colony Optimization (ACO) and Simulated Annealing (SA) have global search capabilities, but they converge slowly in high-dimensional continuous spaces or dynamic time-series data, and lack effective collaboration with deep learning models. Existing combinations are mostly simple concatenations, making it difficult to achieve a balance between "real-time performance" and "optimal performance".

[0008] Therefore, the industry urgently needs a fusion control architecture that can simultaneously achieve "precise dynamic modeling + global optimization + online adaptation" to solve the problem of high-precision, robust, and low-energy-consumption control of complex systems under nonlinear, large-time-lag, multivariable coupling, and strong-disturbance environments.

[0009] There is still no effective solution to the problem that the modeling and control accuracy of complex systems with nonlinearity, large time delay and multivariable coupling in related technologies is insufficient, which leads to difficulties in parameter tuning and poor robustness and real-time performance. Summary of the Invention

[0010] Therefore, this application provides an intelligent control system and method to overcome the problems of insufficient modeling and control accuracy of complex systems with nonlinearity, large time delay, and multivariable coupling in the prior art, which leads to difficulties in parameter tuning and poor robustness and real-time performance.

[0011] To achieve the above objectives, according to a first aspect of the embodiments of this application, an intelligent control system is provided, comprising: a sensor module for real-time acquisition of system state data of a controlled object and external disturbance information of the controlled object; a DCRNN module connected to the sensor module for predicting the state of the controlled object based on the system state data and the external disturbance information, thereby obtaining a predicted state of the controlled object or a preliminary control decision for controlling the controlled object; an SA-ACO optimizer for performing a global optimal solution search on the predicted state or the preliminary control decision based on a preset optimization objective using an ant colony optimization algorithm, thereby obtaining an optimized control command; and an actuator module for executing the optimized control command to control the controlled object through the optimized control command.

[0012] In an optional embodiment, the system further includes: a feedback module, configured to, after controlling the controlled object through the optimized control command, re-collect new system state data of the controlled object and feed it back to the DCRNN module, so as to instruct the DCRNN module to re-predict the state of the controlled object based on the new system state data, thereby forming closed-loop control; or, the feedback module, configured to, after controlling the controlled object through the optimized control command, record the new system state data, thereby forming open-loop control.

[0013] In an optional embodiment, the DCRNN module includes: an input layer for receiving multidimensional time-series data, wherein the multidimensional time-series data is used to indicate the system state data and the external disturbance information at different time intervals; at least one dilated convolutional layer for capturing the multi-scale spatiotemporal features of the multidimensional time-series data, wherein different dilated convolutional layers have different dilation rates; at least one recurrent neural network layer for extracting the temporal dynamic features of the multidimensional time-series data and determining the dependencies between different states based on the temporal dynamic features; and a fully connected layer for concatenating the multi-scale spatiotemporal features and the dependencies into a high-dimensional vector, performing a linear mapping process on the high-dimensional vector, and outputting the predicted state or the preliminary control decision.

[0014] In an optional embodiment, the SA-ACO optimizer includes: an ant colony optimization component (ACO) for constructing a sequence of control parameters in parallel in the solution space; a simulated annealing component (SA) for determining whether to accept the control parameter sequence based on the Metropolis criterion to avoid getting trapped in local optima; and a fusion mechanism, wherein the simulated annealing component dynamically adjusts the pheromone evaporation rate of the ant colony optimization component until the final obtained control parameter sequence converges, and generates the optimized control command based on the final obtained control parameter sequence.

[0015] According to a second aspect of the embodiments of this application, an intelligent control method is also provided, comprising: using a DCRNN module to predict the state of the controlled object based on the system state data and the external disturbance information, to obtain a predicted state of the controlled object or a preliminary control decision for controlling the controlled object; using an SA-ACO optimizer, based on a preset optimization objective, to perform a global optimal solution search on the predicted state or the preliminary control decision using an ant colony optimization algorithm, to obtain an optimized control command; and using an actuator module to execute the optimized control command, so as to control the controlled object through the optimized control command.

[0016] In an optional embodiment, after the optimized control command is applied to the controlled object through the actuator module, the method further includes: re-acquiring new system state data of the controlled object; feeding back the new system state data to the DCRNN module to instruct the DCRNN module to re-predict the state of the controlled object based on the new system state data, forming closed-loop control; or, recording the new system state data to form open-loop control.

[0017] In an optional embodiment, the method further includes: an offline stage: training the DCRNN module using historical data and initializing the SA-ACO optimizer parameters, wherein the historical data includes: historical system state data and historical external disturbance information; an online stage: collecting the system state data and the external disturbance information in real time, processing the system state data and the external disturbance information through the DCRNN module and the SA-ACO optimizer to obtain the optimized control command; and an adaptive adjustment stage: fine-tuning the DCRNN module online based on the new system state data of the controlled object and dynamically adjusting the SA-ACO optimizer strategy.

[0018] In an optional embodiment, the method further includes: a safety limiting step: limiting the optimized control command to ensure that the optimized control command is within the allowable range; and an anomaly handling step: resetting the parameters of the SA-ACO optimizer and triggering an alarm when the optimized control command is abnormal.

[0019] According to a third aspect of the embodiments of this application, an electronic device is also provided, including a memory and a processor, characterized in that the memory stores a computer program, and the processor is configured to execute the intelligent control method described in the first aspect through the computer program.

[0020] According to a fourth aspect of the embodiments of this application, a computer program product is also provided, including a computer program, characterized in that the computer program, when executed by a processor, implements the intelligent control method described in the first aspect.

[0021] The sensor module of this application collects the system state of the controlled object (including temperature, pressure, position, etc.) and external disturbance information (sudden environmental changes, load changes, etc.) in real time, forming a high-dimensional time-series data stream. The DCRNN module is directly coupled with the sensor, using dilated convolution + recurrent network to simultaneously capture long-range spatiotemporal dependence and dynamic inertia, mapping the raw data to "future state prediction" or "preliminary control decision", providing high-quality priors for subsequent optimization. The SA-ACO optimizer uses preset objectives (minimum error, minimum energy consumption, etc.) as constraints to perform global optimization on the DCRNN output. It quickly converges and avoids local extrema through a simulated annealing-ant colony hybrid mechanism, outputting optimized control commands that can be directly executed. The actuator module acts precisely according to the commands to achieve real-time closed-loop control of the controlled object. At the same time, the new state returns to the sensor, forming a continuously iterative adaptive system. This solution uses a four-level link of "sensor → DCRNN modeling → SA-ACO optimization → actuator action" to integrate "real-time perception, spatiotemporal modeling, global optimization, and precise execution", thus solving the problem of high-precision robust control of complex systems under conditions of large lag, nonlinearity, and strong disturbance.

[0022] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0023] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of the application. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0024] Figure 1 This is a structural block diagram (I) of an optional intelligent control system according to an embodiment of this application;

[0025] Figure 2 This is a structural block diagram (II) of an optional intelligent control system according to an embodiment of this application;

[0026] Figure 3 This is a system architecture diagram of an optional intelligent control system according to an embodiment of this application;

[0027] Figure 4 This is a flowchart of an optional intelligent control method according to an embodiment of this application;

[0028] Figure 5 This is an optional DCRNN structure diagram according to an embodiment of this application;

[0029] Figure 6 This is a flowchart illustrating the closed-loop / open-loop interaction between an optional DCRNN and SA-ACO according to an embodiment of this application.

[0030] Figure 7 This is a flowchart of an optional SA-ACO optimization process according to an embodiment of this application. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0033] To address the technical problems existing in related technologies, this embodiment provides an intelligent control system. Figure 1 This is a structural block diagram of an intelligent control system according to an embodiment of this application. The system program includes:

[0034] Sensor module 12 is used to collect system status data of the controlled object and external disturbance information of the controlled object in real time;

[0035] DCRNN module 14, connected to the sensor module, is used to predict the state of the controlled object based on the system state data and the external disturbance information, and to obtain the predicted state of the controlled object or a preliminary control decision for controlling the controlled object.

[0036] SA-ACO optimizer 16 is used to perform a global optimal solution search on the predicted state or the preliminary control decision using an ant colony optimization algorithm to obtain optimized control commands.

[0037] The actuator module 18 is used to execute the optimized control instructions to control the controlled object through the optimized control instructions.

[0038] In the aforementioned system, the sensor module collects real-time data on the controlled object's own system state (including temperature, pressure, position, etc.) and external disturbance information (sudden environmental changes, load variations, etc.), forming a high-dimensional time-series data stream. The DCRNN module is directly coupled with the sensor, using dilated convolution and recurrent networks to simultaneously capture long-range spatiotemporal dependencies and dynamic inertia, mapping the raw data to "future state prediction" or "preliminary control decisions," providing high-quality priors for subsequent optimization. The SA-ACO optimizer uses preset objectives (minimum error, minimum energy consumption, etc.) as constraints to globally optimize the DCRNN output, quickly converging and avoiding local extrema through a simulated annealing-ant colony hybrid mechanism, outputting directly executable optimized control commands. The actuator module precisely acts according to the commands, achieving real-time closed-loop control of the controlled object. Simultaneously, the new state returns to the sensor, forming a continuously iterative adaptive system. This solution uses a four-level link of "sensor → DCRNN modeling → SA-ACO optimization → actuator action" to integrate "real-time perception, spatiotemporal modeling, global optimization, and precise execution", thus solving the problem of high-precision robust control of complex systems under conditions of large lag, nonlinearity, and strong disturbance.

[0039] Optional, such as Figure 2 As shown, the system further includes: a feedback module 20, used to re-collect new system state data of the controlled object after controlling the controlled object through the optimized control command, and feed it back to the DCRNN module to instruct the DCRNN module to re-predict the state of the controlled object based on the new system state data, forming closed-loop control; or, the feedback module is used to record the new system state data after controlling the controlled object through the optimized control command, forming open-loop control.

[0040] The feedback module 20 immediately pulls back the system state for a new round of processing after the command is issued, sending it directly back to the DCRNN. ​​For example: 1. In a tomato greenhouse, as soon as the motor opens the skylight to 45%, the temperature and humidity sensors push back the new values ​​within 0.1 seconds, and the DCRNN immediately updates its prediction. 2. After the left and right wheel speeds of the autonomous vehicle are updated, the IMU and LiDAR synchronously transmit the position and speed back, and the network can correct the trajectory in the next second. Continuous loopback allows the model and control strategy to evolve in real time according to the operating conditions, reducing errors and decreasing energy consumption.

[0041] It should be noted that open-loop or closed-loop control are two operating modes of intelligent control systems. These two operating modes will be further explained below:

[0042] 1. Closed-loop control: After the actuator performs an action, the sensor immediately sends the new system state back to the DCRNN; the DCRNN uses this latest data to recalculate predictions or decisions, and the next round of optimization continues in sequence. The data stream is continuous, with loops for improvement, errors converge in real time, and the system is constantly self-correcting.

[0043] 2. Open-loop control: After the actuator performs an action, the sensor data is only used for monitoring or recording and is no longer sent back to the DCRNN; the DCRNN still uses the weights or fixed model trained offline last time, and is not updated in subsequent optimizations. The entire chain flows in one direction without loops.

[0044] Optionally, the DCRNN module includes: an input layer for receiving multi-dimensional time-series data, wherein the multi-dimensional time-series data is used to indicate the system state data and external disturbance information at different time intervals; at least one dilated convolutional layer for capturing the multi-scale spatiotemporal features of the multi-dimensional time-series data, wherein different dilated convolutional layers have different dilation rates; at least one recurrent neural network layer for extracting the temporal dynamic features of the multi-dimensional time-series data and determining the dependencies between different states based on the temporal dynamic features; and a fully connected layer for concatenating the multi-scale spatiotemporal features and the dependencies into a high-dimensional vector, performing a linear mapping process on the high-dimensional vector, and outputting the predicted state or the preliminary control decision.

[0045] The DCRNN module employs a four-layer progressive structure: the input layer reads in multi-dimensional temporal data in one go, recording system state variables and external perturbation variables side-by-side on the same time axis, with the dimensionality depending on the scenario; then, it connects to at least one dilated convolutional layer with 3×3 kernels and progressively increasing dilation rate, forming 1-2-4 levels of receptive fields to capture multi-scale spatiotemporal features from local to global; the convolutional output is fed into a recurrent neural network layer, using LSTM or GRU units to encode the temporal dynamics and establish long-short-term dependencies between states; finally, a fully connected layer maps the fused spatiotemporal features to predicted states or preliminary control decisions, with the output dimension matching the downstream task. Through the dual-channel design of dilated convolution and recurrent networks, the module simultaneously completes spatial context expansion and temporal memory retention within a single path, achieving low-latency, high-precision modeling of complex dynamic systems.

[0046] Specifically, DCRNN combines atrous convolution and recurrent neural networks (RNNs, such as LSTM or GRU) to process time-series data and capture the long-range spatiotemporal dependencies of complex systems.

[0047] structure:

[0048] Input layer: Receives multi-dimensional time series data (such as historical state sequences, setpoint sequences, and disturbance information).

[0049] Dilated convolutional layers: Multiple dilated convolutions are used to expand the receptive field to capture long-range spatiotemporal dependencies (such as the influence of long-term states on current control). Different dilation rates (e.g., 1, 2, 4) are set for each layer to balance local and global feature extraction.

[0050] It's important to note that the receptive field is the number of time steps actually covered by the convolutional kernel on the input sequence. A regular convolutional kernel of size 3 can only see 3 adjacent sampling points. Dilated convolution inserts "holes" between kernel elements, with dilation rates r = 1, 2, and 4 respectively increasing the sampling interval by a factor of 1, 2, and 4, effectively extending the effective kernel length to 3, 5, and 9 time steps. After stacking three layers, the network can perceive states up to 1 + 2*(1 + 2 + 4) = 15 time steps ago, allowing it to capture the long-range influence of distant states on current control in a single pass without increasing parameters, while preserving local details.

[0051] RNN layer: Uses LSTM or GRU to process temporal dynamics and capture the short-term and long-term dependencies of the system (such as the dynamic trend of state changes).

[0052] Fully connected layer: Integrates features and outputs predicted state, feature vector, or preliminary control decision.

[0053] Its functions and application scenarios include:

[0054] Scenario A (Predictor): Predict the future state of the system (such as the liquid level and velocity in the next time step).

[0055] Scenario B (Controller): Directly generate control quantities (such as motor speed, valve opening).

[0056] Scenario C (Feature Extractor): Extracts dynamic features of the system for SA-ACO optimization.

[0057] The training process is as follows: offline training is performed using historical data, and the loss function is mean squared error (MSE) or a customized loss that combines the control objective (such as error + energy consumption weighting).

[0058] Optionally, the SA-ACO optimizer includes: an ant colony optimization component (ACO) for constructing a sequence of control parameters in parallel in the solution space; a simulated annealing component (SA) for determining whether to accept the control parameter sequence based on the Metropolis criterion to avoid getting trapped in local optima; and a fusion mechanism, wherein the simulated annealing component dynamically adjusts the pheromone evaporation rate of the ant colony optimization component until the final obtained control parameter sequence converges, and generates the optimized control command based on the final obtained control parameter sequence.

[0059] The SA-ACO optimizer employs a dual-core collaborative architecture of "ant colony-annealing". The ACO component treats the sequence of control parameters as paths, with multiple artificial ants constructing candidate solutions in parallel. Each ant calculates the transition probability based on pheromone and heuristic values, forming a distributed search. The SA component performs a Metropolis judgment on the solutions generated by each ant: if the new solution is better than the current solution, it is directly adopted; otherwise, a inferior solution is accepted with Boltzmann probability, allowing the search to escape local optima. The fusion mechanism maps the annealing temperature to the pheromone evaporation rate p in real time: increasing ρ during high-temperature stages encourages exploration, while decreasing ρ during low-temperature stages strengthens the optimal path, thereby maintaining global and local balance throughout the optimization process. After iteration terminates, the converged sequence of control parameters becomes the final optimization control command.

[0060] Specifically, SA-ACO combines the global exploration capability of simulated annealing with the parallel positive feedback mechanism of ant colony optimization to optimize DCRNN parameters or generate control instructions.

[0061] Its structural components are as follows:

[0062] ACO component: Centered on the control variable u0 output by the DCRNN, a discrete grid (or continuous interval) is established in its neighborhood to form the search space Ω (i.e., the solution space). 30-50 ants simultaneously construct candidate control sequences within Ω; each ant, according to p... ij =τ ij α ·η ij β The probability of choosing the next action is η, where η is the probability of choosing the next action. ij The DCRNN prediction error gradient or energy cost provides heuristic information.

[0063] SA component: Uses the Metropolis criterion to determine whether to accept new solutions (including inferior solutions to escape local optima) and dynamically adjusts the pheromone evaporation rate or search range.

[0064] The SA-ACO component encapsulates "global search + local escape" into two parallel optimization pipelines: one directed towards the DCRNN itself, and the other towards control instructions.

[0065] 1. Optimize DCRNN parameters.

[0066] Search targets: hyperparameters such as weight matrix, kernel size, dilation rate, number of hidden units in LSTM, and learning rate.

[0067] Encoding method: Discretize the hyperparameters into integer or floating-point grids, and each "solution" corresponds to a set of network configurations.

[0068] Evaluation function: Use the validation set MSE or control error plus energy consumption weighted as the fitness.

[0069] Results: SA-ACO returns the hyperparameter combination with optimal fitness, which can directly replace the existing network; the offline phase can be completed in one go, or it can be updated periodically during the adaptive phase.

[0070] 2. Optimize control commands.

[0071] Search targets: discrete or continuous actions such as path point sequences, actuator opening sequences, and motor speed sequences.

[0072] Encoding method: Quantize actions into integer indices or keep them as continuous real numbers, and the ant selects one action at each step.

[0073] Evaluation function: calculated based on the error between the future trajectory predicted by DCRNN and the set trajectory, plus energy consumption cost.

[0074] Output: The converged sequence of actions is the "optimal control command", which is sent to the actuator in real time to achieve minimum energy consumption scheduling or shortest path tracking.

[0075] The two pipelines share the same SA-ACO engine, and can seamlessly switch between "network tuning" and "action tuning" by changing the objective function.

[0076] The data interaction process between the DCRNN module and the SA-ACO optimizer includes:

[0077] DCRNN to SA-ACO: DCRNN output (such as predicted state, feature vector) guides the search of SA-ACO, narrowing the solution space or providing heuristic information.

[0078] SA-ACO to DCRNN: Optimization results (such as control instructions and parameter updates) are fed back to DCRNN to adjust network weights or control strategies.

[0079] Feedback mechanism: The system status is fed back to the DCRNN to form closed-loop or open-loop control and continuously optimize performance.

[0080] To better understand the operating principle of the aforementioned intelligent control system, this application embodiment combines... Figure 3 The structure of the above-mentioned intelligent control system is described below. The system architecture of the intelligent control system is as follows: Figure 3 As shown, it includes the controlled object, sensors, DCRNN module, SA-ACO optimizer, and actuator. Figure 3 The diagram illustrates the data flow between various components in the system.

[0081] 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 this application, in essence, 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 this application.

[0082] This application also provides an intelligent control method, such as... Figure 4 As shown, Figure 4 This is a flowchart of an intelligent control method according to an embodiment of this application, including the following steps:

[0083] Step S402: The system status data of the controlled object and the external disturbance information of the controlled object are collected in real time through the sensor module;

[0084] Step S404: The DCRNN module predicts the state of the controlled object based on the system state data and the external disturbance information to obtain the predicted state of the controlled object or a preliminary control decision for controlling the controlled object.

[0085] Step S406: Using the SA-ACO optimizer, based on a preset optimization objective, the ant colony optimization algorithm is used to search for the global optimal solution of the predicted state or the preliminary control decision to obtain the optimized control command.

[0086] Step S408: The optimized control command is executed by the actuator module to control the controlled object.

[0087] Using the above methods, the sensor module collects the system state of the controlled object (including temperature, pressure, position, etc.) and external disturbance information (sudden environmental changes, load changes, etc.) in real time, forming a high-dimensional time-series data stream. The DCRNN module is directly coupled with the sensor, using dilated convolution + recurrent network to simultaneously capture long-range spatiotemporal dependence and dynamic inertia, mapping the raw data to "future state prediction" or "preliminary control decision", providing high-quality priors for subsequent optimization. The SA-ACO optimizer uses preset objectives (minimum error, minimum energy consumption, etc.) as constraints to perform global optimization on the DCRNN output. It quickly converges and avoids local extrema through simulated annealing-ant colony hybrid mechanism, outputting optimized control commands that can be directly executed. The actuator module acts precisely according to the commands, realizing real-time closed-loop control of the controlled object. At the same time, the new state returns to the sensor, forming a continuously iterative adaptive system. This solution uses a four-level link of "sensor → DCRNN modeling → SA-ACO optimization → actuator action" to integrate real-time perception, spatiotemporal modeling, global optimization and precise execution, thus solving the problem of high-precision robust control of complex systems under conditions of large lag, nonlinearity and strong disturbance.

[0088] Optionally, after the optimized control command is applied to the controlled object through the actuator module, the method further includes: re-acquiring new system state data of the controlled object; feeding back the new system state data to the DCRNN module to instruct the DCRNN module to re-predict the state of the controlled object based on the new system state data, forming closed-loop control; or, recording the new system state data to form open-loop control.

[0089] The feedback module immediately pulls back to the latest system state after the command is issued, sending it directly back to the DCRNN for a new round of processing. For example: 1. In a tomato greenhouse, as soon as the motor opens the skylight to 45%, the temperature and humidity sensors push back the new values ​​within 0.1 seconds, and the DCRNN immediately updates its prediction. 2. After the left and right wheel speeds of the autonomous vehicle are updated, the IMU and LiDAR synchronously transmit the position and speed back, and the network can correct the trajectory in the next second. Continuous loopback allows the model and control strategy to evolve in real time according to operating conditions, reducing errors and lowering energy consumption.

[0090] It should be noted that open-loop or closed-loop control are two operating modes of intelligent control systems. These two operating modes will be further explained below:

[0091] 1. Closed-loop control: After the actuator performs an action, the sensor immediately sends the new system state back to the DCRNN; the DCRNN uses this latest data to recalculate predictions or decisions, and the next round of optimization continues in sequence. The data stream is continuous, with loops for improvement, errors converge in real time, and the system is constantly self-correcting.

[0092] 2. Open-loop control: After the actuator performs an action, the sensor data is only used for monitoring or recording and is no longer sent back to the DCRNN; the DCRNN still uses the weights or fixed model trained offline last time, and is not updated in subsequent optimizations. The entire chain flows in one direction without loops.

[0093] Optionally, the method further includes: an offline stage: training the DCRNN module using historical data and initializing the SA-ACO optimizer parameters, wherein the historical data includes: historical system state data and historical external disturbance information; an online stage: collecting the system state data and the external disturbance information in real time, processing the system state data and the external disturbance information through the DCRNN module and the SA-ACO optimizer to obtain the optimized control command; and an adaptive adjustment stage: fine-tuning the DCRNN module online based on the new system state data of the controlled object and dynamically adjusting the SA-ACO optimizer strategy.

[0094] The overall process is divided into three stages.

[0095] 1. Offline phase: The multi-dimensional time series data consisting of historical system states and historical external disturbances is fed into the DCRNN module to complete weight training. At the same time, the initial temperature, pheromone factor, evaporation rate and other parameters of the SA-ACO optimizer are set using prior statistics.

[0096] The weight training process includes:

[0097] 1) Data preparation: Align historical system states with external disturbances by timestamp to construct sliding window samples;

[0098] 2) Network setup: Stack 3 layers of dilated convolutions (dilation rate 1-2-4) + 2 layers of LSTM (128) + fully connected output header;

[0099] 3) Construct the loss function;

[0100] 4) Build the optimizer;

[0101] 5) In the training loop, perform forward and backward parameter updates, and use the loss function and optimizer to verify the results and optimize the parameters;

[0102] 6) Preserve the optimal weights.

[0103] Online Phase: Sensors acquire the current system state and external disturbances in real time at a fixed frequency. The DCRNN provides predictions or preliminary control inputs immediately. The SA-ACO performs parallel optimization based on these baseline values ​​and outputs the final control command, forming a closed loop. Adaptive Adjustment Phase: New system states generated during online operation are continuously fed back. The last few layers of the DCRNN are fine-tuned in-situ using a low learning rate. Simultaneously, the SA temperature and pheromone evaporation rate are dynamically adjusted according to the error change rate to achieve synchronous evolution of the model and strategy.

[0104] It should be noted that the online fine-tuning process of the DCRNN module includes:

[0105] 1. Freeze the backbone: After loading the offline weights, freeze the dilated convolutional layer and the first LSTM layer, and only open the last 1-2 LSTM layers and the output fully connected layer.

[0106] 2. Every 500-100 new samples accumulated during online operation form a micro-batch, using the same sliding window and normalization table as the offline stage.

[0107] 3. Optimize configuration, adjust optimizer learning rate and weight decay parameters;

[0108] 4. It runs in a background thread, and the weight copy in memory is updated immediately after each fine-tuning, without interruption to the main inference thread.

[0109] Optionally, the method further includes: a safety limiting step: limiting the optimized control command to ensure that the optimized control command is within the allowable range; and an anomaly handling step: resetting the parameters of the SA-ACO optimizer and triggering an alarm when the optimized control command is abnormal.

[0110] Safety limiting procedure: The optimized control command is first trimmed by upper and lower limits, and then limited by the rate of change, ensuring that any executed quantity always falls within the physically permissible range. Anomaly handling procedure: If continuous over-limit, overshoot, or sudden increase in feedback error occurs, the temperature, pheromone, and volatility of SA-ACO are immediately reset to preset values, and an alarm signal is sent to the monitoring terminal. The system then switches to degraded operation or manual intervention mode.

[0111] In an optional embodiment, the specific implementation of the above-mentioned intelligent control method is as follows:

[0112] 1. Initialization:

[0113] 1) DCRNN Initialization: Network Structure: 3 layers of dilated convolutions (kernel size 3, dilation rate [1, 2, 4]), 2 layers of LSTM (128 hidden units), 1 layer of fully connected (output dimension depends on the application, such as state prediction as state dimension, control quantity as actuator dimension), such as... Figure 5 As shown.

[0114] Training data: historical states, control variables, and setpoint data (such as the past 1000 time steps).

[0115] Loss function: Mean squared error (MSE) or multi-objective loss (e.g., error + energy consumption).

[0116] Optimizer: Adam optimizer, learning rate 0.001, batch size 32.

[0117] 2) SA-ACO initialization: Number of ants: 50.

[0118] Pheromones factor (α): 1.0, heuristic factor (β): 2.0, pheromone evaporation rate (ρ): 0.1.

[0119] SA parameters: initial temperature 1000, cooling coefficient 0.95, termination temperature 0.01, maximum number of iterations 1000.

[0120] Optimization objectives: Minimize tracking error (e.g., the MSE between actual output and setpoint) or comprehensive objectives (e.g., error + energy consumption).

[0121] 2. DCRNN processing:

[0122] Inputs: historical state sequence (such as sensor data from the past 10 time steps, including temperature, pressure, velocity, etc.), setpoint sequence (such as target trajectory, liquid level), external disturbances (such as changes in ambient temperature).

[0123] Processing flow:

[0124] Hollow convolutional layers extract multi-scale spatiotemporal features (such as short-term fluctuations and long-term trends).

[0125] LSTM layers handle temporal dynamics and model the dependencies between states.

[0126] The fully connected layer integrates features and outputs predicted states (such as the system state at the next time step), feature vectors, or preliminary control variables (such as motor speed).

[0127] Output: Depending on the application scenario, output the predicted state, feature vector, or control decision.

[0128] 3. SA-ACO optimization:

[0129] Input: DCRNN output (predicted state, feature vector), optimization objective function, constraints (such as control range, energy consumption limit).

[0130] Optimize the process:

[0131] ACO phase: Ants construct solutions in the solution space (such as a sequence of control parameters). Each ant selects a path based on pheromones and heuristic information (such as the error gradient predicted by DCRNN).

[0132] SA Phase: For the solutions generated by the ants, the Metropolis criterion is used to determine whether to accept them, based on the objective function value and the current temperature. At high temperatures, the probability of accepting inferior solutions is higher, while at low temperatures, the solution tends towards the optimal solution.

[0133] Pheromones are updated: SA dynamically adjusts the pheromone evaporation rate (e.g., a higher evaporation rate at high temperatures promotes exploration; a lower evaporation rate at low temperatures strengthens the optimal solution).

[0134] Output: Optimal control instruction sequence or optimized DCRNN parameters.

[0135] 4. Interaction and Execution: such as Figure 6 As shown, Figure 6 Several closed-loop / open-loop interaction processes between DCRNN and SA-ACO are illustrated below. The following sections will describe these interaction processes in specific scenarios A / B / C:

[0136] Scenario A (DCRNN is the predictor):

[0137] DCRNN predicts future states (such as the liquid level at the next time step).

[0138] SA-ACO optimizes control variables (such as valve opening) based on prediction results, with the goal of minimizing tracking error.

[0139] The actuator applies a control signal, the sensor collects the new state, and the feedback updates the DCRNN.

[0140] Scenario B (DCRNN is the controller):

[0141] SA-ACO optimizes the weights or hyperparameters (such as porosity and number of neurons) of DCRNN.

[0142] The optimized DCRNN generates control variables, which are then applied to the system by the actuator.

[0143] The system provides feedback on performance (such as error and energy consumption), triggering the next SA-ACO optimization.

[0144] Scenario C (Hybrid Decision Making):

[0145] DCRNN extracts real-time features (such as the spatiotemporal feature vector of system state).

[0146] SA-ACO generates refined control instructions based on features and global objectives (such as path optimization and minimum energy consumption).

[0147] The actuator applies instructions, and the system updates DCRNN and SA-ACO in response.

[0148] 5. Feedback and Online Operation:

[0149] Sampling period: Collect system status at a fixed period (e.g., 0.1 seconds) and input it into DCRNN.

[0150] Optimization trigger: SA-ACO optimization is initiated when control performance deteriorates (e.g., error exceeds the limit) or periodically (e.g., every 10 time steps).

[0151] Adaptive updates: DCRNN is fine-tuned online based on new data, and SA-ACO dynamically adjusts the optimization strategy.

[0152] Parameter configuration (example):

[0153] DCRNN:

[0154] Diffuse convolutional layers: 3 layers, kernel size 3, void ratio [1, 2, 4], number of output channels 64.

[0155] RNN layer: 2 LSTM layers, 128 hidden units.

[0156] Fully connected layer: The output dimension depends on the application (e.g., state prediction is the state dimension, and control quantity is the actuator dimension).

[0157] Training parameters: learning rate 0.001, batch size 32, training epochs 100.

[0158] SA-ACO:

[0159] Number of ants: 50.

[0160] Pheromones (α): 1.0, heuristic factor (β): 2.0, volatility (p): 0.1.

[0161] SA parameters: initial temperature 1000, cooling coefficient 0.95, termination temperature 0.01, maximum number of iterations 1000.

[0162] Optimization objective: Minimize MSE (such as the error between actual output and set value) or weighted objective (such as 0.7 × error + 0.3 × energy consumption).

[0163] System parameters:

[0164] Sampling frequency: 10Hz (0.1 seconds / s).

[0165] Optimization trigger conditions: Optimize once when the error exceeds 5% or every 10 time steps.

[0166] Example (using robot path tracking as an example)

[0167] Application scenario: Robots move along a designated path and need to cope with nonlinear dynamics and environmental disturbances (such as changes in ground friction).

[0168] initialization:

[0169] DCRNN: The input is the robot's position, velocity, and target path point over the past 10 time steps, and the output is the predicted position or preliminary velocity command for the next time step. The structure consists of 3 dilated convolutional layers + 2 LSTM layers + 1 fully connected layer.

[0170] SA-ACO: The optimization objective is to minimize path tracking error and energy consumption. Initialize 50 ants, pheromone factor 1.0, and initial temperature 1000.

[0171] Operation process:

[0172] Sensors collect the robot's current state (position, velocity, acceleration) and target path points, and input them into the DCRNN.

[0173] DCRNN predicts the position of the next time step or generates preliminary velocity instructions.

[0174] SA-ACO is based on DCRNN to output optimized speed command sequences (such as adjusting motor speed), with the goal of minimizing tracking error and energy consumption.

[0175] The actuator applies an optimized speed command, the sensor collects the new state, and feeds it back to the DCRNN.

[0176] Performance monitoring:

[0177] The status is collected every 0.1 seconds. If the tracking error exceeds 5%, SA-ACO optimization is triggered.

[0178] DCRNN fine-tunes weights online based on new data to maintain model accuracy.

[0179] Results: Compared with traditional PID (overshoot 10%, response time 0.5 seconds), this method has an overshoot of less than 3%, a response time of 0.2 seconds, and stronger robustness.

[0180] Operational Process (Summary)

[0181] Offline phase:

[0182] Use historical data to train DCRNN to build system dynamic models or control strategies.

[0183] Initialize SA-ACO parameters and set optimization objectives and constraints.

[0184] Online phase:

[0185] Real-time system status is collected and input into DCRNN.

[0186] DCRNN processes data and outputs predictions, features, or preliminary control variables.

[0187] SA-ACO optimizes control commands or parameters based on DCRNN output.

[0188] The actuator applies a control input, the system returns a new state, and the process repeats cyclically.

[0189] Adaptive adjustment:

[0190] DCRNN fine-tunes the model based on the new data.

[0191] SA-ACO dynamically adjusts pheromone updates and temperature parameters to maintain optimal efficiency.

[0192] In an optional embodiment, this application combines Figure 7 The optimization process steps of SA-ACO are introduced, such as... Figure 7 As shown, the specific steps include:

[0193] Step 1: SA-ACO initialization;

[0194] Step 2: Ants construct solutions (i.e., control parameter sequences) in the solution space. Each ant selects a path based on pheromones and heuristic information (such as the error gradient predicted by DCRNN).

[0195] Step 3: Calculate the objective function value J;

[0196] Step 4: Perform the SA (Self-Assessment) judgment. For the solution generated by the ants, determine whether to accept it based on the objective function value and the current temperature using the Metropolis criterion. At high temperatures, the probability of accepting a suboptimal solution is higher; at low temperatures, it tends towards the optimal solution. If rejected, return to step 2; if accepted, proceed to step 5.

[0197] Step 5: Accept the new solution;

[0198] Step 6: Pheromone update. SA dynamically adjusts the pheromone evaporation rate (e.g., higher evaporation rate at high temperatures promotes exploration; lower evaporation rate at low temperatures strengthens the optimal solution).

[0199] Step 7: Cool down;

[0200] Step 8: Determine if the termination condition is met. If not, return to step 2. If it is met, proceed to step 9.

[0201] Step 9: Output the optimal solution (optimal control instruction sequence or optimized DCRNN parameters).

[0202] Embodiments of this application also provide a storage medium including a stored program, wherein the program, when executed, performs any of the methods described above. Optionally, in this embodiment, the storage medium may be configured to store program code for performing the steps of the methods described above.

[0203] Optionally, in this embodiment, the storage medium may include, but is not limited to, 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.

[0204] Embodiments of this application also provide an electronic device including a memory and a processor, the memory storing a computer program and the processor being configured to run the computer program to perform the steps in any of the above method embodiments.

[0205] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0206] Embodiments of this application also provide a computer program product, including a non-volatile computer-readable storage medium storing the computer program product, wherein the computer program, when executed by a processor, implements the steps of the methods described in various embodiments of this application.

[0207] Optionally, in this embodiment, the computer program described above can be configured to implement the steps in any of the above method embodiments when executed by a processor.

[0208] 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.

[0209] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0210] Furthermore, it should be understood that since the various modules are only provided to illustrate the functional units of the device described in this application, the physical devices corresponding to these modules may be the processor itself, or a part of the processor's software, hardware, or a combination of both. Therefore, the number of modules shown in the figures is merely illustrative.

[0211] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, and this is not limited herein.

[0212] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An intelligent control system, characterized in that, include: The sensor module is used to collect system status data of the controlled object and external disturbance information of the controlled object in real time; The DCRNN module, connected to the sensor module, is used to predict the state of the controlled object based on the system state data and the external disturbance information, so as to obtain the predicted state of the controlled object or a preliminary control decision for controlling the controlled object. The SA-ACO optimizer is used to search for the global optimal solution of the predicted state or the preliminary control decision based on a preset optimization target and using the ant colony optimization algorithm to obtain the optimized control command. An actuator module is used to execute the optimized control instructions in order to control the controlled object through the optimized control instructions.

2. The intelligent control system according to claim 1, characterized in that, The system further includes a feedback module, which is used to re-collect new system state data of the controlled object after controlling the controlled object through the optimized control command, and feed it back to the DCRNN module to instruct the DCRNN module to re-predict the state of the controlled object based on the new system state data, thereby forming closed-loop control. Alternatively, the feedback module is used to record the new system state data after controlling the controlled object through the optimized control command, thereby forming open-loop control.

3. The intelligent control system according to claim 1, characterized in that, The DCRNN module includes: An input layer is used to receive multi-dimensional time-series data, wherein the multi-dimensional time-series data is used to indicate the system state data and the external disturbance information at different time sequences; At least one dilated convolutional layer is used to capture the multi-scale spatiotemporal features of the multidimensional time-series data, wherein different dilated convolutional layers have different dilation rates; At least one recurrent neural network layer is used to extract the temporal dynamic features of the multidimensional time series data and determine the dependencies between different states based on the temporal dynamic features; The fully connected layer is used to concatenate the multi-scale spatiotemporal features and the dependencies into a high-dimensional vector, and perform a linear mapping process on the high-dimensional vector to output the predicted state or the preliminary control decision.

4. The intelligent control system according to claim 1, characterized in that, The SA-ACO optimizer includes: Ant Colony Optimization Component (ACO) is used to construct a sequence of control parameters in parallel in the solution space; The simulated annealing component (SA) is used to determine whether to accept the control parameter sequence based on the Metropolis criterion, in order to avoid getting trapped in local optima. The fusion mechanism wherein the simulated annealing component dynamically adjusts the pheromone evaporation rate of the ant colony optimization component until the final control parameter sequence converges, and generates the optimized control command based on the final control parameter sequence.

5. An intelligent control method, characterized in that, The intelligent control system applied to any one of claims 1 to 4 comprises: The sensor module collects system status data of the controlled object and external disturbance information of the controlled object in real time. The DCRNN module predicts the state of the controlled object based on the system state data and the external disturbance information, thereby obtaining the predicted state of the controlled object or a preliminary control decision for controlling the controlled object. The SA-ACO optimizer searches for the global optimal solution for the predicted state or the preliminary control decision using an ant colony optimization algorithm based on a preset optimization objective, in order to obtain the optimized control command. The optimized control instructions are executed by the actuator module to control the controlled object.

6. The intelligent control method according to claim 5, characterized in that, After applying the optimized control commands to the controlled object through the actuator module, the method further includes: Re-collect new system status data of the controlled object; The new system state data is fed back to the DCRNN module to instruct the DCRNN module to re-predict the state of the controlled object based on the new system state data, thus forming closed-loop control; Alternatively, record the new system state data to form open-loop control.

7. The intelligent control method according to claim 5, characterized in that, The method further includes: Offline phase: The DCRNN module is trained using historical data, and the SA-ACO optimizer parameters are initialized. The historical data includes: historical system state data and historical external disturbance information. Online phase: The system status data and external disturbance information are collected in real time, and the system status data and external disturbance information are processed by the DCRNN module and SA-ACO optimizer to obtain the optimized control command; Adaptive adjustment phase: Fine-tune the DCRNN module online based on the new system state data of the controlled object, and dynamically adjust the SA-ACO optimizer strategy.

8. The intelligent control method according to any one of claims 5 to 7, characterized in that, The method further includes: Safety limiting step: Limit the optimized control command to ensure that the optimized control command is within the allowable range; Anomaly handling steps: When the optimized control command is abnormal, reset the parameters of the SA-ACO optimizer and trigger an alarm.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the intelligent control method according to any one of claims 5 to 8 through the computer program.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the intelligent control method as described in any one of claims 5 to 8.

Images