Control method, device and control system based on communication system

Abstract

The invention provides a control method, system and control device based on a communication system, and the method comprises the steps: obtaining a current control state and a communication state of each terminal device, and carrying out the nonlinear transformation and structure splicing, so as to construct a joint state representation; obtaining the scheduling priority of each task, and selecting K tasks with the highest scheduling priority to form a task set; for each task in the task set, using a control trajectory prediction function to calculate a predicted action trajectory of a target task in future T steps; inputting the predicted action trajectory into a pre-constructed auto-encoder, and generating a low-dimensional control intention vector; in response to the control intention vector, decoding the control intention vector by using a decoder, generating a restored predicted motion trajectory, and minimizing a reconstruction error, so that the restored predicted motion trajectory approaches an original predicted motion trajectory as much as possible; and executing a control action corresponding to the restored predicted action track, and performing closed-loop feedback optimization.

Description

Technical Field

[0001] The present invention belongs to the field of communication system control, and in particular relates to a control method, device and control system based on a communication system. Background Art

[0002] With the rapid development of technologies such as intelligent manufacturing, the Internet of Vehicles, remote robotic control, and integrated air-space-ground networks, distributed control systems are increasingly reliant on communication networks to transmit control commands and environmental status data. In these "coupled control-communication systems," both the issuance of control signals and feedback on system status must be relayed and transmitted through the communication system. The advantage of this architecture is that it enables remote and coordinated control, making it particularly suitable for large-scale, geographically distributed, and complex control objects. However, inherent uncertainties in the communication system are increasingly becoming a major bottleneck limiting control performance. For example, in practical applications, due to factors such as limited network bandwidth, fluctuating communication latency, link packet loss, and uneven resource scheduling, control commands may not be delivered to the controlled objects in a timely and complete manner, leading to control instability, delayed response, and reduced system security.

[0003] Existing research efforts attempt to alleviate these issues from two main perspectives. One approach involves introducing QoS scheduling mechanisms at the communication protocol or network layer to prioritize critical control flows or ensure bandwidth resources. However, these approaches are often tied to specific network hardware or transmission protocols, lack universal applicability, and struggle to accurately perceive the real-time demands of control behaviors. Another approach relies on introducing redundancy, adaptability, or fault-tolerance mechanisms into control strategies, such as building predictive control models and setting control fallback actions, to reduce reliance on communication continuity. However, most existing approaches assume static or simplified communication conditions (e.g., constant bandwidth or fixed latency), resulting in significant adaptability issues in dynamic and complex communication environments.

[0004] In addition, traditional control tasks often send control commands to each terminal device at a fixed frequency according to a predetermined frequency, without considering the actual availability of the communication link and the urgency of the control task itself. This "one-size-fits-all" control sending strategy will further aggravate the communication bottleneck in high-concurrency and resource-limited systems, and at the same time cause some critical tasks to be delayed or even interrupted. In some special scenarios with low bandwidth, weak connection or intermittent communication (such as submarine control systems, disaster area robot rescue, satellite remote control tasks, etc.), the existing solutions lack effective control data compression mechanisms and control semantic communication methods, resulting in only relying on simple or static control methods to maintain system operation, which seriously restricts the responsiveness of intelligent control and the execution ability of complex tasks. Therefore, there is an urgent need for a new method that can take into account the status of the communication system and the requirements of the control task, so that the control system can still achieve reliable communication and effective execution of key control behaviors under communication-restricted conditions. Summary of the Invention

[0005] The purpose of the present invention is to propose a control method, device and control system based on a communication system, which can realize priority sorting, adaptive compression transmission, and accurate restoration and execution of control commands at the remote end when communication conditions change dynamically or are limited.

[0006] In a first aspect, an embodiment of the present invention provides a control method based on a communication system, the method comprising:

[0007] Obtain the current control state and communication state of each terminal device and perform nonlinear transformation and structural splicing to construct a joint state representation. The control state includes the current lateral position deviation, the current speed change rate, the change in distance to the leading vehicle, and the magnitude of the control output change. The communication state includes the current average communication delay, the current packet loss rate, the current channel bandwidth estimate, and the availability score of the currently available link.

[0008] Obtain the scheduling priority of each task, select the K tasks with the highest scheduling priority, and form a task set; wherein the scheduling priority is obtained by mapping the joint state representation to the priority space;

[0009] For each task in the task set, a control trajectory prediction function is used to calculate a predicted motion trajectory for the target task T steps in the future; wherein the predicted motion trajectory change between two consecutive steps is constrained to be less than or equal to the maximum amplitude of the control trajectory change, so as to avoid sudden large adjustments;

[0010] The predicted motion trajectory is input into a pre-built autoencoder to generate a low-dimensional control intention vector for data transmission; wherein the change between two consecutive steps of the control intention vector is constrained to a set maximum change amplitude to avoid drastic fluctuations in the control intention;

[0011] In response to the control intention vector, decoding the control intention vector using a decoder to generate a restored predicted motion trajectory, and performing a reconstruction error analysis between the restored predicted motion trajectory and the original predicted motion trajectory to minimize the reconstruction error so that the restored predicted motion trajectory is as close as possible to the original predicted motion trajectory;

[0012] Execute the control action of the restored predicted motion trajectory and perform closed-loop feedback optimization.

[0013] Furthermore, the scheduling priority is adjusted by a penalty term based on the communication state;

[0014] The penalty term based on the communication state is obtained by weighted summing the link delay, packet loss rate and link bandwidth;

[0015] The K tasks with the highest scheduling priority are selected by a Top-K algorithm.

[0016] Further, the control trajectory prediction function is constructed as follows:

[0017] A 3-layer LSTM is used, with 128 hidden units in each layer, which generates a predicted action trajectory for T steps in the future by combining the current control state, communication state, and scheduling priority of the task;

[0018] The control trajectory prediction function also has a regularization term for unstable communication, which helps adjust the prediction result when the communication quality is poor, ensuring that the priority scheduling of the task and the control behavior are not affected by network fluctuations; the regularization term for unstable communication is obtained by superparametric weighted summation of link delay, packet loss rate, and link bandwidth.

[0019] Further, the pre-constructed autoencoder includes an input layer, a hidden layer, and an output layer;

[0020] The input layer is used to input the predicted action trajectory The shape is (T, d), where T is the number of predicted time steps and d is the feature dimension of each step;

[0021] The input layer is composed of multiple fully connected layers, each using a ReLU activation function, with 3 layers and 128 neurons in each layer;

[0022] The output layer is used to control the intent vector, with a dimension of 32.

[0023] Further, the autoencoder also has a regularization term for automatically adjusting the weight of the autoencoder when compressing the predicted action trajectory, represented as:

[0024]

[0025] where smooth(h i (t)) represents the smoothness of the predicted action trajectory, reflecting the stability of the control behavior; variance(h i (t)) measures the dynamic changes of the predicted action trajectory; α1, α2 are adjustment coefficients for balancing the smoothness and change characteristics of the predicted action trajectory.

[0026] Further, the decoder is designed as a multi-layer perceptron or a recurrent neural network.

[0027] The decoder has a structure including an input layer, a hidden layer, and an output layer; wherein,

[0028] The input layer is used to input a control intention vector with a dimension of k;

[0029] The hidden layer includes two fully connected layers, each layer contains 128 neurons, and the activation function is ReLU;

[0030] The output layer is used to output the restored predicted motion trajectory, and the output dimension of this layer is (T, d).

[0031] Furthermore, the reconstruction error is obtained by calculating the mean square error between the restored predicted motion trajectory and the original predicted motion trajectory;

[0032] The decoder also includes a regularization term Predicted motion trajectories for optimizing decoder output,

[0033] Expressed as:

[0034]

[0035] in, It is the change in the predicted action trajectory between adjacent time steps; It is the sum of squares of the predicted motion trajectory changes, which is used to control the smoothness of trajectory changes during the restoration process; is the L2 norm of the predicted action trajectory, which prevents the trajectory value from being too large or too small and ensures that the predicted action trajectory output by the decoder is within a reasonable range; β1 and β2 are adjustment coefficients used to control the influence of the regularization term.

[0036] Furthermore, the control action of executing the restored predicted motion trajectory and performing closed-loop feedback optimization specifically includes:

[0037] Execute the restored predicted motion trajectory and collect feedback information, including: position, speed and environmental parameters;

[0038] The control action is adjusted based on the error between the feedback information and the restored predicted action trajectory.

[0039] According to a second aspect of the present application, a control device based on a communication system is provided, comprising:

[0040] A state acquisition module is used to obtain the current control state and communication state of each terminal device and perform nonlinear transformation and structural splicing to construct a joint state representation. The control state includes the current lateral position deviation, the current speed change rate, the change in distance to the leading vehicle, and the magnitude of the control output change. The communication state includes the current average communication delay, the current packet loss rate, the current channel bandwidth estimate, and the availability score of the currently available link.

[0041] A scheduling priority module is used to obtain the scheduling priority of each task and select the K tasks with the highest scheduling priority to form a task set; wherein the scheduling priority is obtained by mapping the joint state representation to the priority space;

[0042] A prediction module is configured to calculate, for each task in the task set, a predicted motion trajectory of the target task T steps in the future using a control trajectory prediction function; wherein the predicted motion trajectory change between two consecutive steps is constrained to be less than or equal to the maximum amplitude of the control trajectory change, so as to avoid sudden large adjustments;

[0043] A compression module is configured to input the predicted motion trajectory into a pre-built autoencoder to generate a low-dimensional control intention vector for data transmission; wherein the change between two consecutive steps of the control intention vector is constrained to a set maximum change amplitude to avoid drastic fluctuations in the control intention;

[0044] a reconstruction module, configured to, in response to the control intention vector, decode the control intention vector using a decoder to generate a restored predicted motion trajectory, and perform a reconstruction error analysis between the restored predicted motion trajectory and the original predicted motion trajectory to minimize the reconstruction error so that the restored predicted motion trajectory is as close as possible to the original predicted motion trajectory;

[0045] The optimization module is used to execute the control action corresponding to the restored predicted action trajectory and perform closed-loop feedback optimization.

[0046] According to a second aspect of the present application, a control system based on a communication system is provided, comprising:

[0047] a memory configured to store instructions; and

[0048] The processor is configured to call the instructions from the memory and implement any one of the above-mentioned control methods based on the communication system when executing the instructions.

[0049] The beneficial technical effects of the present invention are at least as follows:

[0050] To address the aforementioned issues, the present invention proposes a control method, system, and device specifically designed to address the uncertainty of control task transmission and execution in communication systems. These methods enable prioritized control command sorting, adaptively compressed transmission, and accurate remote restoration and execution even when communication conditions are dynamically changing or limited. By sensing the communication link status and identifying the urgency of control tasks, the present invention constructs a dynamic control command scheduling and selection mechanism within the system, prioritizing the transmission of critical control tasks when resources are limited. Furthermore, to address control expression issues under low bandwidth or discontinuous communication conditions, the present invention designs a compression strategy for the control behavior itself. This strategy extracts control intent rather than the complete instruction sequence for transmission, and, in conjunction with a remote decoding module, fully restores the compressed control semantics. Furthermore, the present invention introduces a predictive control capability that, based on the current communication status, determines future communication feasibility. This allows for dynamic planning of control execution paths, preemptively supplementing control actions in the event of a potential communication interruption, and enhancing system robustness and control consistency. Through these technical approaches, the present invention effectively overcomes existing issues such as delayed control task transmission, uninterrupted execution, and unpredictable behavior under complex and fluctuating communication systems, significantly improving the stability, efficiency, and intelligence of integrated communication and control systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The present invention is further described with reference to the accompanying drawings. However, the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative effort.

[0052] Figure 1 This is a flow chart of a control method based on a communication system of the present invention.

[0053] Figure 2 This is a framework diagram of a control device based on a communication system of the present invention. DETAILED DESCRIPTION

[0054] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0055] In one embodiment, Figure 1 As shown, a control method based on a communication system is provided, comprising the following steps:

[0056] S1. Obtain the current control state and communication state of each terminal device, and perform nonlinear transformation and structural splicing to construct a joint state representation; wherein the control state includes the current lateral position deviation, the current speed change rate, the change in distance to the leading vehicle, and the control output change amplitude; the communication state includes the current average communication delay, the current packet loss rate, the current channel bandwidth estimate, and the availability score of the current available link.

[0057] Specifically, the goal of this step is to construct a set of state representations h for all control terminals in the system for scheduling evaluation at each time t when the control system is running. i (t), where i represents the i-th terminal or controlled object. This state representation must simultaneously include two core aspects of information: the real-time behavior of the controlled object itself (whether it is urgently needed for control, how quickly it changes, etc.); and the current communication link status of the controlled terminal (whether good transmission conditions are available). This fused state representation serves as input to the subsequent scheduling mechanism, allowing control task scheduling to consider not only the priority of the task itself but also whether it can be delivered to the controlled device in a timely and stable manner, thereby achieving optimal resource allocation in communication-constrained scenarios.

[0058] Furthermore, the control state vector x i Construction of (t):

[0059] For each terminal device (such as a mobile robot, remote manipulator or unmanned vehicle, etc.), the system needs to collect its current control state x i (t), this vector should include key indicators describing its “control urgency” and “dynamics”.

[0060] For example, in an unmanned vehicle fleet application scenario, each vehicle acts as a control terminal, and its control state x i (t) may include the following dimensions:

[0061] Current lateral position deviation (indicating whether it deviates from the ideal trajectory); current speed change rate (acceleration); change in distance from the preceding vehicle; control output change amplitude (such as the control instruction a in the previous cycle); i The difference between (t-1) and the current expected instruction).

[0062] Understandably, assuming the state of a car is: deviation 0.6 meters, acceleration 2.1m / s 2 , the distance reduction speed is 0.8m / s, and the control command change is 0.3. These values ​​can be directly read by the vehicle's local sensors (IMU, odometer, controller state cache), forming a i (t) = a vector of [0.6, 2.1, 0.8, 0.3].

[0063] Furthermore, the communication state vector c i Construction of (t):

[0064] At the same time, the system needs to evaluate the current communication status of each terminal in real time. i (t), which determines whether its control command can be delivered reliably. The communication status mainly includes the following elements:

[0065] The current average communication delay (such as RTT, in milliseconds); the current packet loss rate (such as 6% if 3 packets were lost in the past 50); the current channel bandwidth estimate (which can be reported by the communication protocol stack, such as TCPBBR or application layer throughput estimate); the availability score of the currently available link (a weighted stability score can be calculated based on historical changes in link quality).

[0066] It is understandable that, for example, the current communication status of a car is: delay 55ms, packet loss rate 5%, bandwidth 2.5Mbps, link stability score is 0.82, forming c i (t)=[55,0.05,2.5,0.82].

[0067] This data comes from the system's communication stack. For example, packet loss statistics are obtained through UDP connections, RTT is measured through ping tests, and bandwidth is measured through actual transmission rate. Link stability can be calculated using the exponentially weighted moving average algorithm to calculate the quality fluctuation range within the past 5 seconds.

[0068] Furthermore, the joint state representation h i Construction and calculation method of (t):

[0069] In order to control the state x i (t) and communication state c i (t) is fused into a unified vector h that can be input to the scheduling module i (t), this scheme designs a simple but expressive nonlinear transformation and splicing structure. The specific method is as follows:

[0070] h i (t)=σ(W1x i (t))‖σ(W2c i (t)) (1)

[0071] Among them, σ is the ReLU activation function, W1 and W2 are linear mapping matrices (which can be initialized as unit matrices or trained); "‖" represents the vector concatenation operation, which is to merge the two transformed vectors into one. Finally, h is formed i (t), serves as the input for scheduling priority weight calculation in the next step.

[0072] Let’s take a practical numerical example to illustrate:

[0073] Assume x i (t) is 4-dimensional, c i (t) is 4-dimensional, W1, W2 are 3×4 matrices, then the outputs are 3-dimensional control features and 3-dimensional communication features, after splicing It can express the comprehensive characteristics of "whether the task is urgent and whether it can be sent smoothly".

[0074] S2. Obtain the scheduling priority of each task, select K tasks with the highest scheduling priority, and form a task set; wherein the scheduling priority is obtained by mapping the joint state representation to the priority space.

[0075] Specifically, the goal of this step is to generate the task state vector h according to the task state vector h generated in step 1. i (t), calculate the scheduling priority weight α of each control task i (t), and select the most important task set according to priority This allows for subsequent prediction and compression of motion trajectories. The innovation here is that it combines the control requirements of the tasks with the communication status, assigns dynamic scheduling priorities to each task, and uses a communication penalty mechanism to ensure that critical tasks are prioritized when the network is unstable.

[0076] Furthermore, based on the input joint state vector h i (t), the present invention calculates the scheduling priority weight α of each task i (t), which is used to determine the scheduling order of tasks in the current cycle. To cope with limited network resources, the priority calculation not only depends on the control requirements of the task itself, but also introduces the influence of communication status to ensure that critical tasks can be scheduled first when communication is poor. The calculation formula for scheduling priority weight is as follows:

[0077] α i (t)=σ(Wh i (t))+λ·Penalty(c i (t)) (2)

[0078] Where σ is the ReLU activation function and W is the function used to transform the joint state vector h i (t) is the weight matrix mapped to the priority space; λ is a hyperparameter used to adjust the communication state penalty term Penalty (c i (t)) in the priority calculation; Penalty (c i (t)) is the penalty term of the communication state, which is calculated by combining the delay, packet loss rate and bandwidth to ensure that the task priority is improved when the link quality is poor. i(t)) is defined as:

[0079] Penalty(c i (t))

[0080] = α1·delay(c i (t))+ α2·loss(c i (t))- α3

[0081] ·bandwidth(c i (t)) (3)

[0082] where delay(c i (t)) represents the link delay, a higher delay will increase the priority of the task; loss(c i (t)) represents the packet loss rate, when the packet loss rate is high, the priority will be raised to ensure that the task can be completed through unstable links; bandwidth(c i (t)) is the bandwidth of the link, when the bandwidth is low, the priority is moderately reduced.

[0083] This penalty mechanism ensures that when the communication state is poor (such as high delay or severe packet loss), the system can automatically raise the priority of the task to avoid the task being delayed or lost due to network problems.

[0084] Further, after calculating the priority weight α i (t) of each task, the present application sorts the priorities of all tasks and selects the top K tasks with the highest priority. These tasks will be grouped into a task set for subsequent prediction of action trajectory prediction and compression. The screening method is as follows:

[0085]

[0086] where N is the total number of tasks in the current system, K is the preset number of tasks (for example, the top 10 most important tasks are selected). Top-K is a simple sorting and selection operation that selects the top K tasks with the highest priority for processing.

[0087] In practical applications, the number of K can be dynamically adjusted according to the resource limitations such as communication bandwidth and computing power of the system. In this way, even in the case of resource shortage, the most important tasks can be processed first, avoiding waste of resources.

[0088] This solution combines task control requirements with communication status to achieve a more intelligent task scheduling mechanism. Task priorities are automatically adjusted in poor communication environments, ensuring that critical tasks are prioritized even in unstable networks. This dynamic priority adjustment mechanism, based on communication quality, innovates task scheduling methods in traditional control systems, enabling the system to better cope with complex communication environments and providing greater robustness.

[0089] S3. For each task in the task set, use the control trajectory prediction function to calculate the predicted motion trajectory of the target task for the next T steps; wherein the change in the predicted motion trajectory between two consecutive steps is constrained to be less than or equal to the maximum amplitude of the control trajectory change, so as to avoid sudden large adjustments.

[0090] Specifically, in step 2, the present invention has calculated the scheduling priority α for each task i (t) and selected the most important task set These tasks will be used to predict motion trajectories according to their priorities.

[0091] The state vector h corresponding to each task i i (t) (including control state and communication state) and communication state c i (t) (such as delay, packet loss rate, bandwidth, etc.) will be used as the input for the predicted action trajectory prediction to ensure that the predicted action trajectory of each task is adapted to the current environment.

[0092] Furthermore, for each task in the task set The present invention uses the predicted motion trajectory prediction function f predict To calculate the control action trajectory of the task in the next T steps The calculation formula for predicting the motion trajectory is:

[0093]

[0094] f predict It is a multi-step prediction model. The present invention uses a 3-layer LSTM with 128 hidden units in each layer. It combines the control state h of the task i (t), communication status c i (t) and task priority α i (t) to generate the predicted action trajectory for the next T steps h i (t) is the joint state vector of the task (from step 1), which includes the control requirements of the task and the real-time status of the communication link; c i (t) is the communication status of the task (such as delay, packet loss rate, etc.), which will affect the trajectory prediction; αi (t) is the priority of the task, which determines the scheduling priority of the task in a multi-tasking environment.

[0095] As you can understand, the innovation of this solution lies in considering communication status (such as latency and packet loss rate) as a key factor in predicting motion trajectories, rather than relying solely on the control requirements of the task itself. This approach ensures that the predicted motion trajectory of the task remains stable and accurate even in unstable communication or bandwidth-constrained situations.

[0096] The change amount constraint of the predicted action trajectory during the prediction process avoids excessive trajectory adjustments. In multi-step prediction, the present invention maintains the smoothness of the trajectory by constraining the change amount of the predicted action trajectory, preventing sudden changes caused by network problems.

[0097] Furthermore, in order to ensure that the predicted motion trajectory is smooth and reasonable, the present invention imposes constraints on the variation of the control action:

[0098]

[0099] in, Represents the change in the predicted action trajectory between two consecutive steps; It is to predict the maximum amplitude of the change in the motion trajectory, ensure the smoothness of the control instructions, and avoid sudden large adjustments.

[0100] In addition, this invention introduces a regularization term for unstable communication to help adjust the prediction results when the communication quality is poor, ensuring that the task priority and control behavior are not affected by network fluctuations:

[0101]

[0102] in, is a regularization term related to the communication state; β1, β2, β3 are adjustable hyperparameters used to control the impact of different communication state factors on trajectory prediction; delay(c i (t)),loss(c i (t)),bandwidth(c i (t)) represent the status of link delay, packet loss rate and bandwidth, which affect the accuracy of trajectory prediction.

[0103] Furthermore, for each task Predict the predicted action trajectory for the next T steps These trajectories will be used in subsequent steps such as compression and decoding. The prediction results of the predicted action trajectory not only consider the control requirements of the task itself, but also the stability of the communication link, thus ensuring that the task can still be controlled as expected even under poor network conditions.

[0104] S4. Input the predicted motion trajectory into a pre-built autoencoder to generate a low-dimensional control intention vector for data transmission; wherein the change between two consecutive steps of the control intention vector is constrained to a set maximum change amplitude, thereby avoiding drastic fluctuations in the control intention.

[0105] Specifically, the goal is to predict the trajectory of each task Compressed into a low-dimensional control intention vector z i , so that the key information of the task can be transmitted with a smaller amount of data. To achieve this goal, the present invention adopts an autoencoder model, which can effectively extract key features from the input data and compress it.

[0106] Furthermore, the autoencoder model is designed:

[0107] Encoder architecture: The encoder part of the autoencoder compresses the predicted motion trajectory through multiple fully connected layers (FullyConnectedLayers) And map it to a low-dimensional space. Specifically, the encoder includes:

[0108] Input layer: predicting action trajectory The shape is (T, d), where T is the number of time steps to predict and d is the feature dimension of each step.

[0109] Hidden layers: The encoder consists of three fully connected layers, each using the ReLU activation function. Each layer contains 128 neurons. Through the gradual dimensionality reduction of these layers, the network extracts high-level features for predicting motion trajectories.

[0110] Output layer: Generate control intention vector z through the final output layer i , whose dimension is set to 32 (can be adjusted according to the complexity of the task). This low-dimensional vector represents the core information of the predicted motion trajectory, which is convenient for subsequent transmission and task execution.

[0111] The advantage of autoencoders is that they eliminate the need for manual feature selection and instead automatically learn and extract the most representative features for predicting motion trajectories. Unlike traditional linear dimensionality reduction methods (such as PCA), autoencoders can effectively handle nonlinear relationships and adapt to more complex task control behaviors.

[0112] Furthermore, to prevent drastic fluctuations in the control intention vector during compression, especially during poor communication quality, the present invention incorporates a smoothing constraint into the autoencoder model. This constraint ensures that the control intention vector does not fluctuate too dramatically between adjacent time steps, thereby maintaining the stability of the control task.

[0113] Specifically, the change in the control intention vector Δz i is constrained to:

[0114] |Δz i |≤Δz max (8)

[0115] Where Δz i =z i,t -z i,t-1 is the change between two consecutive steps; Δz max The maximum variation is set to avoid drastic fluctuations in the control intent. This constraint ensures that the control intent vector, after compression, still reflects the core characteristics of mission control while preventing drastic changes in control behavior due to compression. This ensures that mission execution is not affected, especially when the communication link is unstable.

[0116] Furthermore, in addition to the smoothing constraint, the present invention introduces a regularization term to improve compression and enhance the robustness of the autoencoder. This regularization term automatically adjusts the weights of the autoencoder model when compressing and predicting motion trajectories, allowing the model to focus more on predicting low-frequency, important parts of the motion trajectory and ignore irrelevant high-frequency changes.

[0117] Specific regularization terms is defined as:

[0118]

[0119] Among them, smooth(h i (t)) represents the smoothness of the predicted motion trajectory, reflecting the stability of the control behavior; variance (h i (t) measures the dynamic changes in the predicted motion trajectory; α1 and α2 are adjustment coefficients used to balance the smoothness and variability of the predicted motion trajectory. The addition of regularization allows the control intention vector to more effectively retain important control information and enhances the model's adaptability across different tasks.

[0120] Finally, after processing by the autoencoder, the predicted action trajectory of each task i is is compressed into the control intention vector z i This vector is a low-dimensional representation that can effectively transmit the control intent of the task while maintaining the key information in predicting the action trajectory.

[0121] S5. In response to the control intention vector, use a decoder to decode the control intention vector to generate a restored predicted action trajectory, and perform reconstruction error analysis on the restored predicted action trajectory and the original predicted action trajectory to minimize the reconstruction error so that the restored predicted action trajectory is as close as possible to the original predicted action trajectory.

[0122] Specifically, in step 4, the present invention uses the autoencoder to predict the motion trajectory Compressed into control intention vector z i This vector contains the key information of mission control and is transmitted to the receiving end for further processing. i It is the predicted action trajectory from the original The compressed low-dimensional representation is typically much smaller than the original trajectory, aiming to reduce data transmission while preserving core mission control information. Compression reduces the communication burden and effectively supports real-time transmission.

[0123] It is understandable that at the receiving end, the present invention needs to obtain the control intention vector z i Recover the complete predicted motion trajectory This allows the task execution system to execute tasks based on the recovered predicted motion trajectory. The key to this process is to use an efficient decoder to restore the low-dimensional control intention vector to a multi-step predicted motion trajectory.

[0124] To achieve this goal, the receiver uses a decoder model. The decoder is based on the input control intention vector z i Recover the predicted motion trajectory The decoder can automatically learn to predict the temporal characteristics of the action trajectory and output a control signal close to the original trajectory based on the control intention vector.

[0125] Reduction formula:

[0126]

[0127] Among them, f decode It is the core part of the decoder, responsible for controlling the intention vector z i Recovering predicted motion trajectory

[0128] The decoder design can be implemented by a multi-layer perceptron (MLP) or a recurrent neural network (RNN), and the specific choice is adjusted according to the task requirements.

[0129] Furthermore, the decoder is designed:

[0130] Decoder architecture: The goal of the decoder is to transform the control intention vector zi Converted into a long-term predicted action trajectory. The decoder consists of multiple fully connected layers (FullyConnectedLayers), each followed by a nonlinear activation function (such as ReLU). The specific design is as follows:

[0131] Input layer: Input is the control intention vector z i , with dimension k (usually 32 or 64), is obtained after compression by the autoencoder in step 4. This vector contains the core information of task control.

[0132] Hidden layer: The decoder consists of two fully connected layers, each containing 128 neurons and using the Reluctant Unit (ReLU) activation function. The hidden layer gradually expands the low-dimensional control intent vector into high-dimensional features, gradually restoring it to the predicted action trajectory.

[0133] Output layer: The output is the restored predicted action trajectory The output dimension of this layer is (T, d), which is consistent with the time steps and feature dimensions of the original predicted motion trajectory.

[0134] As you can understand, the decoder not only recovers the features of the control intent vector through the fully connected layer, but also learns to predict the dynamic changes of the action trajectory based on the timing characteristics (if an RNN structure is used). For example, using GRU (Gated Recurrent Unit) or LSTM (Long Short-Term Memory) as the decoder network can further capture the temporal dependencies in the control task and improve the accuracy of restoring the predicted action trajectory.

[0135] Furthermore, during the decoding process, the present invention needs to ensure that the predicted motion trajectory after restoration is Able to be as close to the original predicted motion trajectory as possible To achieve this goal, the present invention introduces the reconstruction error, which measures the difference between the restored predicted motion trajectory and the original predicted motion trajectory. The reconstruction error formula is:

[0136]

[0137] Among them, L reconstruction is the error in reconstructing the predicted action trajectory, which measures the difference between the restored trajectory and the original trajectory;

[0138] is the t-th step of the predicted action trajectory restored by the decoder; is the tth step of the original predicted action trajectory; the error is calculated by the mean square error (MSE), and the goal is to minimize the reconstruction error so that the restored trajectory is as close to the original trajectory as possible.

[0139] Regularization term: To prevent the decoder from overfitting and ensure smoothness during the restoration process, the present invention also introduces a regularization term to control the amplitude of the variation during the decoding process:

[0140]

[0141] in, It is the change in the predicted action trajectory between adjacent time steps; It is the sum of squares of the predicted motion trajectory changes, which is used to control the smoothness of trajectory changes during the restoration process; is the L2 norm of the predicted motion trajectory, preventing trajectory values ​​from being too large or too small, ensuring that the predicted motion trajectory output by the decoder is within a reasonable range. β1 and β2 are adjustment coefficients used to control the influence of the regularization term. Through the regularization term, the present invention can further optimize the predicted motion trajectory output by the decoder, ensuring its smoothness and stability and avoiding excessive fluctuations during the recovery process.

[0142] After being processed by the decoder, the receiver obtains the control intention vector z i Recover the predicted motion trajectory This restoration process allows the trajectory of the control task to be restored, and the task control system can actually execute according to this restored trajectory.

[0143] S6. Execute the control action of following the restored predicted motion trajectory and perform closed-loop feedback optimization.

[0144] Specifically, in step 5, the present invention restores the control intention vector z through the decoder i To predict the trajectory of the action These predicted action trajectories contain the expected behavior of task control and have taken into account the key requirements of task execution. The predicted motion trajectories are used as input to the control execution module, which executes control actions and monitors real-time feedback based on these predicted motion trajectories.

[0145] Furthermore, the goal is to execute control actions and collect feedback x′:

[0146] Control execution: The control module receives the restored predicted action trajectory After that, it is converted into actual control actions. For example, in a robot control system, the control actions may include adjusting the angle of the robot joint, setting the speed of the equipment, etc.

[0147] Feedback collection: After executing the control action, the task system collects feedback information x′ in real time through sensors. This feedback information reflects the current state of the system, which can be the current position and speed of the robot, or changes in the external environment.

[0148] Furthermore, the feedback information x′ t May include:

[0149] Position, velocity: the physical state of the system;

[0150] Environmental parameters: the impact of external factors on task execution (such as obstacle location, temperature, etc.).

[0151] Furthermore, a closed-loop feedback mechanism adjusts the control action based on the feedback:

[0152] After collecting the feedback x′, the system will predict the action trajectory based on the feedback information and expectations This process forms a closed-loop feedback loop, ensuring real-time adjustments during task execution.

[0153] Error calculation and control adjustment: feedback information x′ t Predicted motion trajectory with expectation By comparing the control errors, the system can adjust the control actions in real time.

[0154] Feedback update formula:

[0155]

[0156] Among them, f feedback is the feedback acquisition function, which performs the control action u t (control signal of the current time step) to generate feedback information x′ t ;u t is the control signal, indicating the control action executed at the current time step; x′ t It is feedback information, indicating the state of the system after execution.

[0157] Furthermore, whenever new feedback x′ t After being collected, the system calculates the feedback error and adjusts the control action. The difference between the error and the predicted action trajectory is updated through the following formula:

[0158]

[0159] Where K is the control gain, which determines the sensitivity of the feedback error to the control action adjustment; x′ t It is the currently collected feedback information, indicating the actual state of the system; is the expected predicted motion trajectory, which represents the control target of the task; Δu t is the control adjustment amount, which represents the control amount required to correct the current control signal.

[0160] The control gain K can be adjusted to ensure that the system responds moderately to the error, thereby ensuring the stability and accuracy of the control process.

[0161] This application also provides a control device based on a communication system, please refer to Figure 2 , applied to control systems based on communication systems, including:

[0162] The state acquisition module 10 is used to obtain the current control state and communication state of each terminal device and perform nonlinear transformation and structural splicing to construct a joint state representation. The control state includes the current lateral position deviation, the current speed change rate, the change in distance to the leading vehicle, and the control output change amplitude. The communication state includes the current average communication delay, the current packet loss rate, the current channel bandwidth estimate, and the availability score of the current available link.

[0163] A scheduling priority module 20 is used to obtain the scheduling priority of each task and select the K tasks with the highest scheduling priority to form a task set; wherein the scheduling priority is obtained by mapping the joint state representation to the priority space;

[0164] A prediction module 30 is configured to calculate, for each task in the task set, a predicted motion trajectory for the target task T steps in the future using a control trajectory prediction function; wherein the predicted motion trajectory change between two consecutive steps is constrained to be less than or equal to the maximum amplitude of the control trajectory change, to avoid sudden large adjustments;

[0165] A compression module 40 is configured to input the predicted motion trajectory into a pre-built autoencoder to generate a low-dimensional control intention vector for data transmission; wherein the change between two consecutive steps of the control intention vector is constrained to a set maximum change amplitude to avoid drastic fluctuations in the control intention;

[0166] a reconstruction module 50 for decoding the control intention vector using a decoder in response to the control intention vector to generate a restored predicted motion trajectory, and performing a reconstruction error analysis between the restored predicted motion trajectory and the original predicted motion trajectory to minimize the reconstruction error so that the restored predicted motion trajectory is as close as possible to the original predicted motion trajectory;

[0167] The optimization module 60 is used to execute the control action of the restored predicted motion trajectory and perform closed-loop feedback optimization.

[0168] The communication system-based control device provided in this application utilizes the communication system-based control method of the aforementioned embodiment to resolve the technical problem of limited communication. Compared to the prior art, the communication system-based control device provided in this application has the same beneficial effects as the communication system-based control method of the aforementioned embodiment. Other technical features of the communication system-based control device are the same as those disclosed in the aforementioned embodiment and are not further described here.

[0169] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the module, program segment or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be implemented by a combination of dedicated hardware and computer instructions.

[0170] The modules described in the embodiments of the present application may be implemented in software or hardware, wherein the name of a module does not necessarily limit the unit itself.

[0171] The above description is only part of the embodiments of the present application and does not limit the patent scope of the present application. All equivalent structural transformations made by using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect application in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A control method based on a communication system, characterized in that: The method comprises: Obtain the current control state and communication state of each terminal device and perform nonlinear transformation and structural splicing to construct a joint state representation. The control state includes the current lateral position deviation, the current speed change rate, the change in distance to the leading vehicle, and the magnitude of the control output change. The communication state includes the current average communication delay, the current packet loss rate, the current channel bandwidth estimate, and the availability score of the currently available link. Obtain the scheduling priority of each task, select the K tasks with the highest scheduling priority, and form a task set; wherein the scheduling priority is obtained by mapping the joint state representation to the priority space; For each task in the task set, a control trajectory prediction function is used to calculate a predicted motion trajectory for the target task T steps in the future; wherein the predicted motion trajectory change between two consecutive steps is constrained to be less than or equal to the maximum amplitude of the control trajectory change, so as to avoid sudden large adjustments; The predicted motion trajectory is input into a pre-built autoencoder to generate a low-dimensional control intention vector for data transmission; wherein the change between two consecutive steps of the control intention vector is constrained to a set maximum change amplitude to avoid drastic fluctuations in the control intention; In response to the control intention vector, decoding the control intention vector using a decoder to generate a restored predicted motion trajectory, and performing a reconstruction error analysis between the restored predicted motion trajectory and the original predicted motion trajectory to minimize the reconstruction error so that the restored predicted motion trajectory is as close as possible to the original predicted motion trajectory; Execute the control action corresponding to the restored predicted action trajectory and perform closed-loop feedback optimization.

2. A control method based on a communication system according to claim 1, characterized in that: The scheduling priority is adjusted by a penalty term based on the communication state; The penalty term based on the communication state is obtained by weighted summing the link delay, packet loss rate and link bandwidth; The K tasks with the highest scheduling priority are selected through the Top-K algorithm.

3. The control method based on a communication system according to claim 1, characterized in that: The control trajectory prediction function is constructed as follows: Using a 3-layer LSTM with 128 hidden units in each layer, it generates a predicted action trajectory for the next T steps by combining the current control state, communication state, and task scheduling priority; The control trajectory prediction function also has a regularization term for unstable communication, which helps adjust the prediction results when the communication quality is poor, ensuring that the priority scheduling and control behavior of tasks are not affected by network fluctuations; the regularization term for unstable communication is obtained by taking the hyperparameter weighted sum of link delay, packet loss rate and link bandwidth.

4. The control method based on a communication system according to claim 1, characterized in that: The pre-built autoencoder includes an input layer, a hidden layer, and an output layer; The input layer is used to input the predicted action trajectory The shape is (T, d), where T is the number of time steps to predict and d is the feature dimension of each step; The input layer consists of multiple fully connected layers, each layer uses the ReLU activation function, the number of layers is 3, and each layer contains 128 neurons; The output layer is used to control the intention vector, and its dimension is set to 32.

5. A control method based on a communication system according to claim 4, characterized in that: The autoencoder also has a regularization term It is used to automatically adjust the weight of the autoencoder when compressing and predicting the motion trajectory, expressed as: Among them, smooth(h i (t)) represents the smoothness of the predicted motion trajectory, reflecting the stability of the control behavior; variance (h i (t)) measures the dynamic changes of the predicted motion trajectory; α1, α2 are adjustment coefficients used to balance the smoothness and change characteristics of the predicted motion trajectory.

6. The control method based on a communication system according to claim 1, characterized in that: The decoder is designed as a multi-layer perceptron or a recurrent neural network; The decoder structure includes: an input layer, a hidden layer and an output layer; wherein, The input layer is used to input a control intention vector with a dimension of k; The hidden layer includes two fully connected layers, each layer contains 128 neurons, and the activation function is ReLU; The output layer is used to output the restored predicted motion trajectory, and the output dimension of this layer is (T, d).

7. A control method based on a communication system according to claim 6, characterized in that: The reconstruction error is obtained by calculating the mean square error between the restored predicted motion trajectory and the original predicted motion trajectory; The decoder also includes a regularization term The predicted motion trajectory used to optimize the decoder output is expressed as: in, It is the change in the predicted action trajectory between adjacent time steps; It is the sum of squares of the predicted motion trajectory changes, which is used to control the smoothness of trajectory changes during the restoration process; is the L2 norm of the predicted action trajectory, which prevents the trajectory value from being too large or too small and ensures that the predicted action trajectory output by the decoder is within a reasonable range; β1 and β2 are adjustment coefficients used to control the influence of the regularization term.

8. The control method based on a communication system according to claim 1, characterized in that: The executing the control action of following the restored predicted motion trajectory and performing closed-loop feedback optimization specifically includes: Execute the restored predicted motion trajectory and collect feedback information, including: position, speed and environmental parameters; The control action is adjusted based on the error between the feedback information and the restored predicted action trajectory.

9. A control device based on a communication system, characterized in that: include: A state acquisition module is used to obtain the current control state and communication state of each terminal device and perform nonlinear transformation and structural splicing to construct a joint state representation. The control state includes the current lateral position deviation, the current speed change rate, the change in distance to the leading vehicle, and the magnitude of the control output change. The communication state includes the current average communication delay, the current packet loss rate, the current channel bandwidth estimate, and the availability score of the currently available link. A scheduling priority module is used to obtain the scheduling priority of each task and select the K tasks with the highest scheduling priority to form a task set; wherein the scheduling priority is obtained by mapping the joint state representation to the priority space; A prediction module is configured to calculate, for each task in the task set, a predicted motion trajectory of the target task T steps in the future using a control trajectory prediction function; wherein the predicted motion trajectory change between two consecutive steps is constrained to be less than or equal to the maximum amplitude of the control trajectory change, so as to avoid sudden large adjustments; A compression module is configured to input the predicted motion trajectory into a pre-built autoencoder to generate a low-dimensional control intention vector for data transmission; wherein the change between two consecutive steps of the control intention vector is constrained to a set maximum change amplitude to avoid drastic fluctuations in the control intention; a reconstruction module, configured to, in response to the control intention vector, decode the control intention vector using a decoder to generate a restored predicted motion trajectory, and perform a reconstruction error analysis between the restored predicted motion trajectory and the original predicted motion trajectory to minimize the reconstruction error so that the restored predicted motion trajectory is as close as possible to the original predicted motion trajectory; The optimization module is used to execute the control action corresponding to the restored predicted action trajectory and perform closed-loop feedback optimization.

10. A control system based on a communication system, characterized in that: include: a memory configured to store instructions; as well as A processor is configured to call the instructions from the memory and implement the communication system-based control method according to any one of claims 1 to 8 when executing the instructions.

Images