An edge side linkage control method and system for power autonomous operation

Abstract

The embodiment of the application provides an edge side linkage control method and system for power autonomous operation, the method comprises the following steps: solving the problem of the edge side linkage control method for power autonomous operation caused by the calculation overload of the limited calculation resources of the edge side and the phase lag of the loop caused by the unstable communication link by using a "perception-computation-compensation" feedback loop based on an edge side small model and a behavior verification extrapolation mechanism, realizing the calculation overload, the execution logic misalignment and the phase lag problem of the existing power edge side under the limited resources and communication disturbance, and realizing the deterministic evolution of the physical control loop by constructing an edge side original reasoning prediction mechanism.

Description

Technical Field

[0001] This invention relates to the field of power automation technology, and in particular to an edge-side linkage control method and system for autonomous power operations. Background Technology

[0002] The current industrial IoT architecture of new power systems is driving the transformation of edge computing nodes from simple "data relay stations" to "intelligent hubs." As the core carrier supporting unmanned power inspections and autonomous robot operations, the edge needs to process massive, multi-dimensional, and highly transient multimodal sensing data within millisecond-level windows. However, existing edge computing and control methods exhibit significant limitations of "computational overload" and "execution offset" when dealing with extreme operating conditions such as resource constraints and unstable communication links.

[0003] At the computational resource management level, mainstream methods still rely on static resource allocation or cloud task offloading models. These models presuppose stable computing power supply and predictable computational loads. However, in distributed power operation scenarios, edge computing resources are extremely scarce. Faced with the concurrent impact of large-scale multimodal data, the edge lacks the ability to dynamically adjust computational depth, easily falling into a "processing flood" caused by task accumulation, making it difficult to provide deterministic output within a defined time window. Because it cannot dynamically compress computational depth based on instantaneous task intensity, the generation speed of edge decision instructions lags far behind the electromagnetic mutation demands of the physical world, resulting in severe logical lag in the control loop. Essentially, this is because existing edge systems lack an adaptive trade-off mechanism between computational accuracy and response speed.

[0004] At the sensing-computing linkage level, traditional edge control logic heavily relies on high-bandwidth, constant-latency deterministic communication links. These methods treat sensing data acquisition, edge computing inference, and physical action execution as decoupled processes, neglecting the phase impact of communication fluctuations on the control loop. In remote operating conditions, instantaneous contraction of uplink transmission bandwidth or frequent physical link disconnections can lead to sensing frame loss and transmission queuing, causing "loop phase shift" in the actuators. Because existing edge-side systems lack physical mechanism-based "behavior verification extrapolation" capabilities, when real-time sensing data cannot arrive due to network congestion, the system often enters a "decision suspension" or unstable state, unable to utilize the predictive capabilities of edge computing to offset the phase lag caused by the physical link. This makes it difficult to ensure the stability of power operations in highly random and highly interference-prone environments.

[0005] To address the problems in the existing technologies mentioned above, the industry has long faced a triple systemic dilemma: First, at the computational level, there is "computing power saturation," and static computing models are unable to cope with the processing overload caused by the surge of multimodal data. Second, at the control level, there is "execution drift," lacking an adaptive hedging mechanism for phase lag caused by communication latency. Third, at the system level, there is "sensing-computing mismatch": the system cannot achieve dynamic alignment between "edge inference depth" and "instruction generation timing" under network bandwidth fluctuations. In other words, the system struggles to automatically adjust the computational accuracy of model inference based on the current communication link's capacity, resulting in compromised determinism of the control loop in highly dynamic environments. Summary of the Invention

[0006] To address the problems existing in the prior art, embodiments of the present invention provide an edge-side linkage control method and system for autonomous power operations.

[0007] Firstly, embodiments of this specification provide an edge-side linkage control method for autonomous power operations, the method comprising:

[0008] The system collects edge computing load, communication link latency, and environmental complexity, and constructs an association mapping model with action deviation as the output target. The association mapping model fits the nonlinear mapping relationship between different computing power depths, latency fluctuations, and action deviations.

[0009] Based on the real-time determinism of the monitoring and control loop using the association mapping model, when the full-link time consumption triggers the preset threshold, the pruning strategy is automatically executed to dynamically compress the computing power depth of the main inference model, improve the instantaneous response speed of instruction generation, and calculate the action deviation increment after pruning in real time through the association mapping model to ensure that the instruction output of the main inference model does not break through the hard constraint boundary of real-time performance.

[0010] Using a pre-built multimodal world model on the edge side, combined with the current physical state, the pruned action vector, and the action deviation increment, the operator of the multimodal world model is used to perform verification extrapolation to generate compensation instructions. The reserved prediction computing power is dynamically allocated according to the communication delay to increase the number of extrapolation steps.

[0011] In response to the compensation command, when a delayed real-perception frame is detected, the pre-stored predicted state value and real state value are aligned according to the acquisition timestamp, the residual is calculated, and the parameters of the association mapping model and the weights of the world model are corrected in real time based on the residual.

[0012] Secondly, embodiments of this specification provide an edge-side linkage control system for autonomous power operations, the system comprising:

[0013] The mapping model module is used to collect edge computing load, communication link latency and environmental complexity, and construct an association mapping model with action deviation as the output target. The association mapping model fits the nonlinear mapping relationship between different computing power depths, latency fluctuations and action deviations.

[0014] The inference module is used to monitor the real-time determinism of the control loop based on the association mapping model. When the full-link time consumption triggers the preset threshold, it automatically executes the pruning strategy, dynamically compresses the computing power depth of the main inference model, improves the instantaneous response speed of instruction generation, and calculates the action deviation increment after pruning in real time through the association mapping model to ensure that the instruction output of the main inference model does not break the hard constraint boundary of real-time performance.

[0015] The compensation module is used to use a pre-set multimodal world model on the edge side, combined with the current physical state, the pruned action vector and the action deviation increment, to perform verification extrapolation through the operators of the multimodal world model, generate compensation instructions, and dynamically allocate reserved prediction computing power according to communication latency to increase the number of extrapolation steps.

[0016] The correction module is used to respond to the compensation command. When a delayed real perception frame is detected, it aligns the pre-stored predicted state value and the real state value according to the acquisition timestamp, calculates the residual, and corrects the correlation mapping model parameters and world model weights in real time based on the residual.

[0017] Thirdly, embodiments of this specification provide an electronic device, including a processor and a memory;

[0018] The processor is connected to the memory;

[0019] The memory is used to store executable program code;

[0020] The processor runs a program corresponding to the executable program code stored in the memory to perform the methods described in one or more embodiments.

[0021] Fourthly, embodiments of this specification provide a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned edge-side linkage control method for autonomous power operations.

[0022] In view of the above, the beneficial effects of the technical solutions provided by some embodiments of this specification include at least the following: In one or more embodiments of this specification, an edge-side linkage control method for autonomous power operation is provided, which utilizes a "perception-computation-compensation" feedback loop and behavior verification extrapolation mechanism based on a small edge-side model to solve the computational overload caused by limited edge-side computing resources and the loop phase lag caused by unstable communication links. This method addresses the problems of computational overload, execution logic inaccuracy, and phase lag caused by limited resources and communication disturbances on the existing power edge side. By constructing an edge-side native inference prediction mechanism, deterministic evolution of the physical control loop is achieved. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of an edge-side linkage control method for autonomous power operation provided in one embodiment of this specification.

[0025] Figure 2 This is a schematic diagram of an edge-side linkage control system for autonomous power operation provided in one embodiment of this specification.

[0026] Figure 3 This is a schematic diagram of the structure of an electronic device provided in one embodiment of this specification. Detailed Implementation

[0027] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed merely to enable those skilled in the art to better understand and implement the subject matter described herein, and are not intended to limit the scope, applicability, or examples set forth in the claims. The function and arrangement of the elements discussed may be changed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the various examples. For example, the described methods may be performed in a different order than described, and steps may be added, omitted, or combined. Furthermore, features described in some examples may be combined in other examples.

[0028] As used herein, the term "comprising" and its variations are open terms meaning "including but not limited to". The term "based on" means "at least partially based on". The terms "one embodiment" and "an embodiment" mean "at least one embodiment". The term "another embodiment" means "at least one other embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other definitions, whether explicit or implicit, may be included below. Unless explicitly indicated by the context, the definition of a term shall remain consistent throughout the specification.

[0029] Please see Figure 1 , Figure 1 This document presents an overall flowchart of an edge-side linkage control method for autonomous power operation, as provided in an embodiment of this specification.

[0030] like Figure 1 As shown, the edge-side linkage control method for autonomous power operation includes at least the following steps:

[0031] Step S102: Collect edge computing load, communication link latency and environmental complexity, and construct an association mapping model with action deviation as the output target. The association mapping model fits the nonlinear mapping relationship between different computing power depths, latency fluctuations and action deviations.

[0032] Specifically, a quantitative correlation is established between the abstract computing power and communication status at the edge and the actual deviations in the actions of specific power actuators, constructing a corresponding correlation mapping model F. This model can determine the risk of accuracy loss due to latency fluctuations at different computing power depths. The process includes:

[0033] Parameters are implemented by simultaneously sampling the edge side through multiple channels and mapping them to a unified feature tensor through encoding actions, including:

[0034] Computing resource vectorization: Three core computing power metrics of edge nodes are collected in real time and a one-dimensional column vector is constructed.

[0035] Get CPU utilization Memory bandwidth and the current model inference time (ms). Construct the computing power state vector .

[0036] Communication Link Tensorization: Real-time detection of the three core links in wireless / wired communication to construct a one-dimensional column vector:

[0037] Round-trip latency (ms) of real-time detection data from sensor to edge node Jitter variance of communication delay and signal-to-interference-to-noise ratio (dB) Construct the communication state vector .

[0038] Feature fusion and tensor coding: Extracting operational environment complexity features from real-time images of the inspection robot / camera using a visual backbone network. ,in, It can be a scalar value ranging from 0 to 1, with higher values ​​representing more obstacles and narrower spaces. The same dynamic bias carries entirely different risks under varying environmental complexities, thus introducing... This allows subsequent deviation predictions to be combined with the scenario's safety level, making them more aligned with the needs of power field operations.

[0039] Then, a learnable linear projection layer is utilized. Heterogeneous features ( , , Mapped to a latent space of uniform dimension, a global state feature tensor is generated through cascading. :

[0040] In the formula, This represents a feature cascade operation. By mapping to a unified latent space through a linear projection layer Wp, the magnitude differences between features can be eliminated; ReLU activation can filter out invalid negative features, reduce redundancy, and make it easier for the model to learn true causal relationships.

[0041] Furthermore, after determining the characteristics of computing power / communication / environment status, an association mapping model is established to determine the nonlinear relationship between the characteristics of computing power / communication / environment status and the post-pruning action deviation, including:

[0042] Constructing based on action deviation The deep association mapping network for the regression target has the following specific architecture design:

[0043] Network Topology: A residual MLP structure with skip connections is adopted, containing 3 fully connected layers, with each layer having a hidden unit dimension of 256. The residual MLP can fit complex nonlinear relationships and avoid gradient vanishing during deep network training through skip connections, ensuring the model's convergence speed and accuracy.

[0044] Mapping function instantiation: Model The internal operation logic is described as follows:

[0045] In the formula, As a quantized weight representing the depth of model inference, It is a scalar value between 0 and 1, where 1 represents a full-precision model and 0.5 represents pruning the computational depth by 50%. The smaller the value, the shallower the computational depth and the faster the inference.

[0046] Loss function definition: The mean squared error between the deviation of the model prediction and the deviation in the real scene is used as the training criterion, so that the model can accurately fit the physical pose deviation risk caused by the loss of inference accuracy under different time delay fluctuations.

[0047] Furthermore, action bias based on dual-path comparison Real-time quantization methods anticipate the risk of accuracy loss due to latency fluctuations at different computing power depths, including:

[0048] Dual-path inference mechanism: Two logical branches are executed in parallel in the system background: comparing the output differences between the full-precision model and different pruning intensities, and then converting the accuracy in the abstract model into specific physical actuator trajectory deviations. This includes:

[0049] Baseline branch: Calls the full-precision, unpruned model for inference and outputs the ideal action command. .

[0050] Test branch: Simulates different pruning intensities (i.e., computation depth) The model output under () yields experimental action instructions. .

[0051] Motion deviation quantization algorithm: Calculates the pose residual of two sets of instructions in the workspace using Euclidean distance, and combines this with the phase drift caused by communication delay. (Phase drift is also an important source of motion error.) Compensation and correction are performed to obtain the total motion error. :

[0052] Physical alignment: Through this formula, the system directly transforms the abstract "model inference accuracy" into "trajectory deviation of the physical actuator".

[0053] Output transmission: Quantized output and its corresponding computational depth Storing these thresholds as prior knowledge in subsequent steps allows for the rapid identification of appropriate pruning initiation. This provides a definite search benchmark for subsequent "precision-for-time" decisions.

[0054] Step S104: Based on the real-time determinism of the monitoring control loop using the association mapping model, when the full-link time consumption triggers a preset threshold, an automatic pruning strategy is executed to dynamically compress the computing power depth of the main inference model, improve the instantaneous response speed of instruction generation, and calculate the incremental action deviation after pruning in real time through the association mapping model to ensure that the instruction output of the main inference model does not exceed the hard constraint boundary of real-time performance.

[0055] Specifically, the real-time determinism of the monitoring control loop is achieved based on the mapping model established by S102. When the full-link latency triggers a preset threshold, the system performs dead-line monitoring through an independent real-time scheduler to ensure the determinism of instruction generation, including:

[0056] Link latency aggregation calculation: The scheduler extracts the measured communication latency value in step S102 in real time. (Time taken for sensor data to be collected and delivered to the edge node), plus the estimated time taken for the current full-precision inference. The aggregate calculation determines the total time consumption of the entire process.

[0057] Deadline Conflict Detection Algorithm: Based on the insurmountable real-time hard constraint boundary of power operation. Execute Boolean decision logic: ,

[0058] In the formula, This is a step function. When the result... A value of 1 indicates that the total time consumed across the entire process is about to or has already exceeded the hard limit, and the system immediately suspends the full-precision inference process, forcibly switching to a precision compromise mode. When the result... A value of 0 indicates that the time taken is within a safe range.

[0059] When the determination result is 1, the scheduler can immediately execute two high-priority actions: first, forcibly suspend the currently running full-precision model inference process to avoid unnecessary computing power consumption, ensuring that the pruned inference process exclusively enjoys core computing power and guarantees timely instruction output; second, send an interrupt signal to instantly switch to the dynamic pruning precision trade-off mode, prioritizing timely instruction output.

[0060] Furthermore, based on dynamic masking operators The model depth pruning method forces a reduction in inference time. This includes:

[0061] Pruning intensity retrieval and computational power compression requirement calculation: using a proportional formula: ,

[0062] Calculate the current computing power compression requirement. (Formula) This is the historical average latency (representing the baseline latency under normal operating conditions). This is the sensitivity coefficient (configurable according to the safety level of the scene; for high-voltage, high-risk scenes, pruning is more aggressive, leaving more safety margin; for open, low-risk scenes, pruning is more conservative, prioritizing accuracy). This is achieved by reducing the inference granularity. By focusing computing resources on refining core semantic features such as target localization and obstacle avoidance, the decision feedback cycle is forcibly compressed to the millisecond range by sacrificing the accuracy of local details.

[0063] Mask operator generation: The system generates the mask operator based on the compression ratio. Retrieve the target inference depth from the preset depth-latency mapping threshold table. For example, the original =1 corresponds to full precision, and the retrieved value is... =0.7 corresponds to pruning by 30% and a 40% reduction in inference time.

[0064] Operator distribution and instantaneous pruning execution: The system distributes a binary weight mask matrix. The hidden layer operates on the edge side of the neural network, where M is a matrix with the same dimensions as the model's hidden layer weight matrix, and its elements are only 0 or 1. The corresponding rule can be set as follows: the weights of core neurons that need to be retained are set to 1, and the weights of non-core neurons that need to be turned off are set to 0. ,

[0065] in, For the output of the network layer, Here, W is the input to the network layer, and W is the original weight matrix of the network layer. This is the Hadamard product operator. Through the Hadamard product, the system momentarily shuts down some neurons in the network, thereby reducing computational time. Force compression to within a safe threshold to ensure that... Output preliminary control commands before arrival.

[0066] In addition, while performing pruning, the resulting risks are quantified and transferred to subsequent steps to facilitate compensation in those steps, including:

[0067] Deviation risk prediction: The currently adjusted depth parameter Compared with the environmental complexity factor collected in step S102 Input to association mapping model middle.

[0068] Output prediction bias :Model Based on the mapping rules learned in step S102, the expected action deviation caused by deep compression is output in real time. .

[0069] Cross-step parameter distribution: The system will It is encapsulated as a compensation guidance tuple and sent in real time to the world model in subsequent steps via a high-speed system bus. This enables the linkage from "pruning due to congestion" to "subsequent compensation".

[0070] Step S106: Using a pre-set multimodal world model on the edge side, combined with the current physical state, the pruned action vector, and the action deviation increment, the multimodal world model's operators are used to perform verification extrapolation to generate compensation instructions. The reserved prediction computing power is dynamically allocated based on the communication delay to increase the number of extrapolation steps.

[0071] Specifically, a lightweight world model incorporating physical mechanism constraints is pre-defined on the edge side. The model consists of: the lightweight world model operator... It is composed of a deterministic kinematic branch and a random neural network branch. The deterministic branch encapsulates prior physical knowledge, including but not limited to the differential driving kinematics formula of the inspection robot, the joint dynamics constraints of the robotic arm, the maximum speed / acceleration boundaries of the actuator, and the safe distance rules of the work scene; the random branch is used to fit environmental disturbances, including but not limited to wheel slippage caused by uneven ground, small actuator deviations caused by high-voltage electromagnetic interference, and the impact of wind resistance on the movement trajectory.

[0072] The operator receives the expected action deviation transmitted from step S104 in real time. The physical state of the actuator at the current moment (Including pose coordinates, wheel speed, joint angles, distance to obstacles, etc.) and the motion vector being executed. Its function is to simulate the evolutionary trend of the physical world by using "logical inertia" when the accuracy is reduced due to frame loss or pruning.

[0073] The system can automatically trigger operators in two types of scenarios. The calls include:

[0074] Pruning leads to a decrease in accuracy: Correction is performed on the coarse instructions after pruning. Lost perception frames are addressed by maintaining the control loop through model extrapolation.

[0075] Furthermore, by leveraging the prior physical knowledge of the world model, the accuracy loss due to pruning in step S104 is compensated, and the coarse instructions are corrected into precise instructions that meet the safety threshold, including:

[0076] Compensation weight calculation: The system calculates the compensation weight based on the data transmitted from S104. Dynamic calculation of compensation increment The confidence weight.

[0077] Incremental superposition operation: First, through the world model operator Based on the current state and action vectors Output the original pose prediction values ​​for the next k sampling periods. Then through the weighted fusion operator Correct the original output of the world model:

[0078] In the formula, For the predicted physical state after k sampling periods, For local multimodal world model operators, and These are the current state and action vectors, respectively. As a compensation increment based on historical trends, according to Dynamic adjustment. Operator The weighted fusion operator is defined as the linear superposition of the prior vector output by the world model and the history bias compensation vector. If If the value is too large, the system will automatically adjust the increase. The weighting of data forces the use of prior knowledge from the world model to smooth and correct the pruned "rough instructions," ensuring that the output pose meets safety thresholds. This weighted fusion of "real-time information from the pruning model" and "physical prior information from the world model": when pruning deviation is small, it relies more on real-time data; when pruning deviation is large, it relies more on physical prior information for correction. This preserves the validity of real-time data while using physical laws to correct for accuracy loss, balancing real-time performance and accuracy.

[0079] Furthermore, to address pain points such as execution drift, phase shift, and decision suspension, edge-local predictive computing power is used to compensate for communication link deficiencies and offset phase lag caused by communication fluctuations, including:

[0080] Prediction intensity function established: Constructing a monotonically increasing prediction intensity function with respect to time delay. The logic is as follows: the worse the communication quality and the greater the latency, the higher the prediction strength, and the more extrapolation steps the world model needs.

[0081] Phase compensation execution: Based on the gradient change of the prediction intensity function with communication delay, the prediction computing power quota reserved on the edge side is dynamically allocated, and the compensation computing power is allocated using the following coupling relationship. : ,

[0082] in, This is the algorithm's delivery coefficient, which can be adjusted according to the scene's security level. This represents the gradient change in prediction strength as communication latency worsens. The faster the communication latency increases, the larger the gradient, the more prediction computing power is deployed, and the higher the computational priority of the world model. When communication latency... When an increase is made, the system automatically increases the extrapolation steps of the world model. This creates a "timeline forward" effect.

[0083] Ultimately, by allocating additional predictive computing power, the future can be generated before the actual perceived data arrives. Predicted physical state values ​​for each cycle This counteracts the phase drift of the actuator caused by communication fluctuations.

[0084] Step S108: In response to the compensation instruction, after detecting a delayed real perception frame, align the pre-stored predicted state value and the real state value according to the acquisition timestamp, calculate the residual, and correct the correlation mapping model parameters and world model weights in real time based on the residual.

[0085] Specifically, based on measured data, the correlation mapping model F in S102 and the world model in S106 are compared. Prior calibration and continuous optimization are performed to eliminate prediction drift and model aging issues caused by long-term operation. Simultaneously, residual detection automatically identifies abrupt changes in environmental complexity and dynamically adjusts the sensitivity coefficient of S104, ensuring the entire system remains robust even under extreme conditions. This includes:

[0086] Time-aligned state residuals Calculation method: that is, to determine the precise matching relationship between the actual data and the corresponding predicted value, and to calculate the accurate residual.

[0087] When a real-world perceived frame, delayed due to network congestion, arrives at the edge, the system stores it in a time-stamped state cache queue and initiates the traceability verification logic:

[0088] Data timescale alignment: Extracting the lagging true physical state Based on its original collection timestamp Retrieve the corresponding [item] from the cache queue Periodic extrapolation forecast (i.e., current prediction) (The state that will appear at any given time), to achieve accurate time-scale alignment between the actual value and the predicted value.

[0089] Residual tensor calculation: Calculating the residual deviation between the predicted value and the actual physical trajectory using the Euclidean norm. : ,

[0090] The residual This comprehensively reflects the accuracy loss and prediction error caused by S104 pruning, as well as the physical mechanism deviation of the S106 world model when performing "logic fill-in".

[0091] Furthermore, the mapping model is corrected using real data. With world model Online weight refresh to eliminate prediction drift, including:

[0092] The system is based on the residual The amplitude is used to execute the parameter update action triggered by the classification:

[0093] Corrected mapping model Prediction accuracy calibration: The system acquires the actual motion deviation generated by the S102 "dual path comparison" in real time. (That is, the actual deviation between the ideal action output by the full-precision model of the baseline branch and the experimental action output by the pruned model of the test branch). If Model output Compared with the true value If the deviation exceeds a set threshold, the backpropagation operator is triggered to update the model. The connection weights make the model The predictions for the next cycle will be more accurate.

[0094] World Model Physical mechanism correction: utilizing residuals World Model The environmental disturbance operator in the process is fine-tuned online. This is achieved by... As input to the loss function, the sensitivity of the world model to the current electromagnetic environment is dynamically adjusted, such as the operator weights for random environmental disturbances like high-voltage electromagnetic interference, uneven ground, and wind resistance, thereby improving the extrapolation accuracy for the next sampling period.

[0095] In addition, for extreme operating conditions, adaptive sensitivity coefficient adjustment based on residual feedback can be implemented, including:

[0096] Adjusting operator logic: If residuals from multiple consecutive periods are detected... All are at high levels, and the system determines the current environment complexity. A step change occurred, meaning that extreme operating conditions were encountered.

[0097] The system automatically triggers the sensitivity coefficient in S104. Reset. Increase. The magnitude of this makes the system more "sensitive" to latency fluctuations in subsequent operations, forcibly suppressing subsequent residuals through more aggressive pruning and stronger extrapolation compensation. , After the adjustment, the computing power compression ratio of S104 It will become larger, pruning will be more aggressive, and more safety margin can be reserved; at the same time, the prediction strength of S106 will increase, the number of extrapolation steps will be more, the compensation will be more powerful, and the subsequent residuals will be forcibly reduced. Ensure the robustness of the control loop under extreme operating conditions.

[0098] This invention provides an edge-side linkage control method for autonomous power operations. It utilizes a "sensing-computation-compensation" feedback loop and behavior verification extrapolation mechanism based on a small edge-side model to solve the problems of computational overload caused by limited edge-side computing resources and loop phase lag caused by unstable communication links. The method achieves:

[0099] Sensing-Computation Linkage Control Closed Loop: Breaking through the limitation of decoupling perception and computation in traditional edge control, by constructing a "perception-computation-compensation" feedback loop, a correlation mapping model between communication delay and action deviation is established, realizing the transformation of the edge-side decision-making logic from "passive response" to "active adaptive hedging" control paradigm.

[0100] Dynamic inference depth adaptation: Improves the traditional static computing mode by introducing a dynamic pruning mechanism for inference accuracy based on edge load. When the latency touches the hard constraint boundary, the model operator mask is dynamically activated to exchange asymmetric computing accuracy for millisecond-level response speed, ensuring the timing determinism of instruction generation under high concurrency.

[0101] Mechanism-Driven Verification Extrapolation: This research investigates a verification extrapolation technique based on fragments of a multimodal power world model. This technique utilizes locally stored physical mechanism knowledge to perform logical prediction and state synthesis during periods of discontinuous sensing data, enabling advance inference of physical evolution trends and compensating for the bottleneck of missing sensing frames caused by communication congestion.

[0102] Control loop phase offsetting: To address the control loop instability problem caused by communication delay, a phase compensation logic based on the derivative of the prediction intensity function is proposed, which is achieved by allocating computing power quotas. The "timeline forward" effect offsets the physical lag, ensuring the smooth operation of the actuator under extreme conditions.

[0103] Please refer to the following. Figure 2 , Figure 2 A schematic diagram of an edge-side linkage control system for autonomous power operation, provided in an embodiment of this specification, is shown. It should be noted that... Figure 2 The edge-side linkage control system shown is used to execute this specification. Figure 1 The methods shown in the embodiments are illustrated for ease of explanation, showing only the parts related to the embodiments of this specification. For specific technical details not disclosed, please refer to this specification. Figure 1 The example shown.

[0104] like Figure 2 As shown, the edge-side linkage control system for autonomous power operations may include at least:

[0105] The mapping model module S202 is used to collect edge computing load, communication link latency and environmental complexity, and construct an association mapping model with action deviation as the output target. The association mapping model fits the nonlinear mapping relationship between different computing power depths, latency fluctuations and action deviations.

[0106] The inference module S204 is used to monitor the real-time determinism of the control loop based on the association mapping model. When the full-link time consumption triggers the preset threshold, it automatically executes the pruning strategy, dynamically compresses the computing power depth of the main inference model, improves the instantaneous response speed of instruction generation, and calculates the action deviation increment after pruning in real time through the association mapping model to ensure that the instruction output of the main inference model does not break the real-time hard constraint boundary.

[0107] The compensation module S206 is used to use a multimodal world model pre-set on the edge side, combined with the current physical state, the pruned action vector and the action deviation increment, to perform verification extrapolation through the operators of the multimodal world model, generate compensation instructions, and dynamically allocate reserved prediction computing power according to the communication delay to increase the number of extrapolation steps.

[0108] The correction module S208 is used to respond to the compensation instruction. When a delayed real perception frame is detected, it aligns the pre-stored predicted state value and the real state value according to the acquisition timestamp, calculates the residual, and corrects the correlation mapping model parameters and world model weights in real time based on the residual.

[0109] Those skilled in the art will clearly understand that the technical solutions of the embodiments in this specification can be implemented by means of software and / or hardware. In this specification, "unit" and "module" refer to software and / or hardware capable of independently performing or cooperating with other components to perform a specific function. The hardware may be, for example, a Field-Programmable Gate Array (FPGA), an Integrated Circuit (IC), etc.

[0110] Each processing unit and / or module in the embodiments of this specification can be implemented by an analog circuit that implements the functions described in the embodiments of this specification, or by software that executes the functions described in the embodiments of this specification.

[0111] See Figure 3 It shows a schematic diagram of the structure of an electronic device according to an embodiment of this specification, which can be used to implement... Figure 1 The method in the illustrated embodiment. (As shown) Figure 3 As shown, the electronic device 300 may include: at least one central processing unit 301, at least one network interface 304, user interface 303, memory 305, and at least one communication bus 302.

[0112] The communication bus 302 is used to enable communication between these components.

[0113] The user interface 303 may include a display screen and a camera. Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0114] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0115] The processor 301 may include one or more processing cores. The processor 301 connects to various parts within the electronic device 300 using various interfaces and lines, and performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 305, and by calling data stored in the memory 305. Optionally, the processor 301 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 301 may integrate one or a combination of several of the following: a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), and a modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display on the screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 301 and may be implemented as a separate chip.

[0116] The memory 305 may include random access memory (RAM) or read-only memory. Optionally, the memory 305 may include a non-transitory computer-readable storage medium. The memory 305 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 305 may also be at least one storage system located remotely from the aforementioned processor 301. Figure 3 As shown, the memory 305, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and program instructions.

[0117] exist Figure 3 In the illustrated electronic device 300, the user interface 303 is mainly used to provide an input interface for the user and to acquire user input data; while the processor 301 can be used to call the image-based interactive application stored in the memory 305 and specifically perform the following operations:

[0118] The system collects edge computing load, communication link latency, and environmental complexity, and constructs an association mapping model with action deviation as the output target. The association mapping model fits the nonlinear mapping relationship between different computing power depths, latency fluctuations, and action deviations.

[0119] Based on the real-time determinism of the monitoring and control loop using the association mapping model, when the full-link time consumption triggers the preset threshold, the pruning strategy is automatically executed to dynamically compress the computing power depth of the main inference model, improve the instantaneous response speed of instruction generation, and calculate the action deviation increment after pruning in real time through the association mapping model to ensure that the instruction output of the main inference model does not break through the hard constraint boundary of real-time performance.

[0120] Using a pre-built multimodal world model on the edge side, combined with the current physical state, the pruned action vector, and the action deviation increment, the operator of the multimodal world model is used to perform verification extrapolation to generate compensation instructions. The reserved prediction computing power is dynamically allocated according to the communication delay to increase the number of extrapolation steps.

[0121] In response to the compensation command, when a delayed real-perception frame is detected, the pre-stored predicted state value and real state value are aligned according to the acquisition timestamp, the residual is calculated, and the parameters of the association mapping model and the weights of the world model are corrected in real time based on the residual.

[0122] As an optional embodiment of this specification, the construction of the association mapping model with action deviation as the output target includes:

[0123] Real-time acquisition of edge computing load, communication link latency, and environmental complexity; mapping of computing load, communication link latency, and environmental complexity to a unified feature tensor through encoding actions; and establishing an association mapping model based on the feature tensor with the post-pruning action deviation as the output target.

[0124] The nonlinear relationship between computing load, communication link latency, environmental complexity and post-pruning action deviation in the correlation mapping model is fitted, and the action deviation under different pruning intensities is compared through dual-path comparison to generate a pruning decision threshold table.

[0125] As an optional embodiment of this specification, the dual-path comparison calculation formula for the pruning decision threshold table includes: ,

[0126] in, For action deviation, This represents the ideal action command output by the association mapping model under full precision and unpruned conditions. The experimental action commands are for different pruning intensities. For communication link delay, This represents the phase offset.

[0127] As an optional embodiment of this specification, the real-time determinism of monitoring and controlling loops based on the association mapping model, and the automatic execution of a pruning strategy when a preset threshold is triggered by the full-link time consumption, dynamically compressing the computing power depth of the main inference model and improving the instantaneous response speed of instruction generation, includes:

[0128] Based on the real-time determinism of the monitoring control loop using the aforementioned correlation mapping model, dead-line monitoring is performed through an independent time-series detector. When the full-link time consumption triggers a preset real-time hard constraint threshold, a pruning strategy is automatically executed. The optimal target pruning intensity and computing power depth are retrieved by combining the pruning decision threshold table.

[0129] A dynamic masking operator is generated based on the computing power depth. The computing power depth of the main inference model is compressed by the dynamic masking operator, thereby improving the instantaneous response speed of instruction generation. The action deviation increment corresponding to this pruning is calculated in real time through the association mapping model.

[0130] As an optional embodiment of this specification, the proportional formula for compressing the computing power depth of the main inference model through the dynamic masking operator includes: ,

[0131] in, For compression ratio, The historical average latency, This is the sensitivity coefficient. For communication link delay, For real-time hard constraint boundaries.

[0132] As an optional embodiment of this specification, the step of performing verification extrapolation through the operator of the multimodal world model, generating compensation instructions, and dynamically allocating reserved prediction computing power according to communication delay to increase the number of extrapolation steps includes:

[0133] A pre-defined multimodal world model on the edge side, wherein the multimodal world model is a fusion of kinematic deterministic branches and neural network stochastic branches;

[0134] Based on the prior physical knowledge of the world model, combined with the current physical state, the pruned action vector, and the action deviation increment, a verification extrapolation is performed to generate a compensation instruction. The pruned coarse instruction is then corrected into a precise instruction that meets the safety threshold through the compensation instruction.

[0135] Based on the degree of communication latency deterioration, the prediction computing power reserved on the edge side is dynamically allocated to increase the number of extrapolation steps of the model and offset the phase drift caused by communication fluctuations.

[0136] As an optional embodiment of this specification, the step of aligning pre-stored predicted state values ​​and actual state values ​​according to the collection timestamp, calculating residuals, and real-time correcting the correlation mapping model parameters and world model weights based on the residuals includes:

[0137] When a lagging real perception frame is detected, the predicted state value corresponding to the acquisition timestamp is retrieved from the time-stamped state buffer queue, the time-stamp alignment between the predicted state value and the real physical state value is completed, and the residual tensor between the two is calculated.

[0138] Based on the residual tensor, a classification-triggered parameter update action is performed, including the calibration of the prediction accuracy of the association mapping model and the correction of the environmental disturbance fitting ability of the world model, and the pruning sensitivity coefficient under extreme conditions is adaptively adjusted based on residual feedback.

[0139] This specification also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0140] 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 specification is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this specification. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this specification.

[0141] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0142] In the embodiments provided in this specification, it should be understood that the disclosed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between systems or units may be electrical or other forms.

[0143] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0144] Furthermore, the functional units in the various embodiments of this specification can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0145] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this specification, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this specification. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0146] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0147] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

Claims

1. An edge-side linkage control method for autonomous power operation, the method comprising: The system collects edge computing load, communication link latency, and environmental complexity, and constructs an association mapping model with action deviation as the output target. The association mapping model fits the nonlinear mapping relationship between different computing power depths, latency fluctuations, and action deviations. Based on the real-time determinism of the monitoring and control loop using the association mapping model, when the full-link time consumption triggers the preset threshold, the pruning strategy is automatically executed to dynamically compress the computing power depth of the main inference model, improve the instantaneous response speed of instruction generation, and calculate the action deviation increment after pruning in real time through the association mapping model to ensure that the instruction output of the main inference model does not break through the hard constraint boundary of real-time performance. Using a pre-built multimodal world model on the edge side, combined with the current physical state, the pruned action vector, and the action deviation increment, the operator of the multimodal world model is used to perform verification extrapolation to generate compensation instructions. The reserved prediction computing power is dynamically allocated according to the communication delay to increase the number of extrapolation steps. In response to the compensation command, when a delayed real-perception frame is detected, the pre-stored predicted state value and real state value are aligned according to the acquisition timestamp, the residual is calculated, and the parameters of the association mapping model and the weights of the world model are corrected in real time based on the residual.

2. The method according to claim 1, characterized in that, The construction of the association mapping model with action deviation as the output target includes: Real-time acquisition of edge computing load, communication link latency, and environmental complexity; mapping of computing load, communication link latency, and environmental complexity to a unified feature tensor through encoding actions; and establishing an association mapping model based on the feature tensor with the post-pruning action deviation as the output target. The nonlinear relationship between computing load, communication link latency, environmental complexity and post-pruning action deviation in the correlation mapping model is fitted, and the action deviation under different pruning intensities is compared through dual-path comparison to generate a pruning decision threshold table.

3. The method according to claim 2, characterized in that, The dual-path comparison calculation formula for the pruning decision threshold table includes: ， in, For movement deviation, This represents the ideal action command output by the association mapping model under full precision and unpruned conditions. The experimental action commands are for different pruning intensities. For communication link delay, This represents the phase offset.

4. The method according to claim 2, characterized in that, The real-time determinism of the monitoring and control loop based on the association mapping model, when a preset threshold is triggered by the full-link time consumption, automatically executes a pruning strategy to dynamically compress the computing power depth of the main inference model and improve the instantaneous response speed of instruction generation, including: Based on the real-time determinism of the monitoring control loop using the aforementioned correlation mapping model, dead-line monitoring is performed through an independent time-series detector. When the full-link time consumption triggers a preset real-time hard constraint threshold, a pruning strategy is automatically executed. The optimal target pruning intensity and computing power depth are retrieved by combining the pruning decision threshold table. A dynamic masking operator is generated based on the computing power depth. The computing power depth of the main inference model is compressed by the dynamic masking operator, thereby improving the instantaneous response speed of instruction generation. The action deviation increment corresponding to this pruning is calculated in real time through the association mapping model.

5. The method according to claim 4, characterized in that, The formula for proportionally compressing the computing power depth of the main inference model through the dynamic masking operator includes: ， in, For compression ratio, The historical average latency, This is the sensitivity coefficient. For communication link delay, For real-time hard constraint boundaries.

6. The method according to claim 4, characterized in that, The step of performing verification extrapolation through the operators of the multimodal world model, generating compensation instructions, and dynamically allocating reserved prediction computing power based on communication latency to increase the number of extrapolation steps includes: A pre-defined multimodal world model on the edge side, wherein the multimodal world model is a fusion of kinematic deterministic branches and neural network stochastic branches; Based on the prior physical knowledge of the world model, combined with the current physical state, the pruned action vector, and the action deviation increment, a verification extrapolation is performed to generate a compensation instruction. The pruned coarse instruction is then corrected into a precise instruction that meets the safety threshold through the compensation instruction. Based on the degree of communication latency deterioration, the prediction computing power reserved on the edge side is dynamically allocated to increase the number of extrapolation steps of the model and offset the phase drift caused by communication fluctuations.

7. The method according to claim 6, characterized in that, The step of aligning pre-stored predicted state values ​​and actual state values ​​according to the collection timestamp, calculating residuals, and real-time correcting the parameters of the correlation mapping model and the weights of the world model based on the residuals includes: When a lagging real perception frame is detected, the predicted state value corresponding to the acquisition timestamp is retrieved from the time-stamped state buffer queue, the time-stamp alignment between the predicted state value and the real physical state value is completed, and the residual tensor between the two is calculated. Based on the residual tensor, a classification-triggered parameter update action is performed, including the calibration of the prediction accuracy of the association mapping model and the correction of the environmental disturbance fitting ability of the world model, and the pruning sensitivity coefficient under extreme conditions is adaptively adjusted based on residual feedback.

8. An edge-side linkage control system for autonomous power operation, characterized in that, The system includes; The mapping model module is used to collect edge computing load, communication link latency and environmental complexity, and construct an association mapping model with action deviation as the output target. The association mapping model fits the nonlinear mapping relationship between different computing power depths, latency fluctuations and action deviations. The inference module is used to monitor the real-time determinism of the control loop based on the association mapping model. When the full-link time consumption triggers the preset threshold, it automatically executes the pruning strategy, dynamically compresses the computing power depth of the main inference model, improves the instantaneous response speed of instruction generation, and calculates the action deviation increment after pruning in real time through the association mapping model to ensure that the instruction output of the main inference model does not break the hard constraint boundary of real-time performance. The compensation module is used to use a pre-set multimodal world model on the edge side, combined with the current physical state, the pruned action vector and the action deviation increment, to perform verification extrapolation through the operators of the multimodal world model, generate compensation instructions, and dynamically allocate reserved prediction computing power according to communication latency to increase the number of extrapolation steps. The correction module is used to respond to the compensation command. When a delayed real perception frame is detected, it aligns the pre-stored predicted state value and the real state value according to the acquisition timestamp, calculates the residual, and corrects the correlation mapping model parameters and world model weights in real time based on the residual.

9. An electronic device, comprising a processor and a memory; The processor is connected to the memory; The memory is used to store executable program code; The processor runs a program corresponding to the executable program code stored in the memory to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any one of claims 1-7.

Images