Cooperative control method of multi-motor system and vehicle

Abstract

The application discloses a kind of cooperative control method and vehicle of multi-motor system, it is related to vehicle technical field, the control method includes: obtaining the system state data of multi-motor drive system, system state data is used to characterize the response delay difference between multi-motor;Based on system state data, construct causal graph, wherein, causal graph is used to characterize the causal relationship between delay difference and multi-motor output mismatch;Based on causal graph, determine delay compensation control strategy;Based on the actual output response data of multi-motor drive system after executing delay compensation control strategy, predict the residual synchronization error between multi-motor;According to the residual synchronization error of prediction, delay compensation is carried out to multi-motor.The method of the application, by causal modeling and multistage compensation, the effective inhibition of time-varying delay difference is realized, reduces the transient mismatch of multi-motor output, improves the precision and stability of system synchronization control.

Description

Technical Field

[0001] This application relates to the field of vehicle technology, and in particular to a cooperative control method for a multi-motor system and a vehicle. Background Technology

[0002] Current control strategies for multi-motor drive systems largely rely on a unified controller issuing synchronization commands and assume consistent responses from all motors. However, in real-world operating conditions, due to hardware differences, sensor delays, CAN bus congestion, and temperature variations, microsecond-level response delays exist between motors, leading to transient mismatches in the actual output. This mismatch not only disrupts the coordinated output between motors but also causes slight vibrations in the vehicle's posture during start-up, acceleration, and steering, which can severely reduce vehicle stability and ride comfort. Existing methods often lack the ability to model these minute nonlinear delay errors and lack structured cognitive mechanisms and scalable strategy scheduling frameworks, making it impossible to achieve high-precision delay compensation and synchronization control under dynamic and complex operating conditions. Summary of the Invention

[0003] This application aims to at least partially address one of the technical problems in related technologies. To this end, the first objective of this application is to propose a cooperative control method for multi-motor systems. This method involves acquiring system state data characterizing delay differences, constructing a causal graph based on the system state data to demonstrate the causal relationship between delay differences and output mismatch, determining a preliminary delay compensation control strategy based on the causal graph, predicting residual synchronization errors in the system response after strategy execution, and then performing refined compensation based on these predictions. This achieves a systematic modeling and cancellation of complex time-varying delay differences from causal identification to error compensation, thereby significantly improving the synchronization accuracy and dynamic response performance of multi-motor drives and effectively solving the output mismatch problem.

[0004] To achieve the above objectives, the first aspect of this application proposes a cooperative control method for a multi-motor system, comprising: acquiring system state data of a multi-motor drive system, the system state data including timing data for determining the response delay differences between the multi-motors and vehicle dynamic response data for characterizing the output mismatch of the multi-motors; constructing a causal graph based on the system state data, wherein the causal graph is used to characterize the causal relationship between the delay difference and the output mismatch of the multi-motors; determining a delay compensation control strategy based on the causal graph; predicting the residual synchronization error between the multi-motors based on the actual output response data of the multi-motor drive system after executing the delay compensation control strategy; and compensating the multi-motors according to the predicted residual synchronization error.

[0005] According to the cooperative control method for a multi-motor system according to embodiments of this application, system state data of the multi-motor drive system is acquired. This system state data includes data for determining response delay differences among the motors. Based on the system state data, a causal graph is constructed, whereby the causal graph characterizes the causal relationship between delay differences and multi-motor output mismatch. A delay compensation control strategy is determined based on the causal graph. Then, based on the actual output response data of the multi-motor drive system after implementing the delay compensation control strategy, the residual synchronization error among the motors is predicted. Finally, compensation is applied to the motors based on the predicted residual synchronization error. Therefore, this method effectively suppresses time-varying delay differences through causal modeling and multi-level compensation, reducing transient mismatch in multi-motor output and improving the accuracy and stability of system synchronization control.

[0006] In some embodiments of this application, obtaining system status data of a multi-motor drive system includes: obtaining raw data of the multi-motor system; and performing timestamp alignment processing on the raw data based on a high-precision synchronous clock to obtain system status data.

[0007] In some embodiments of this application, constructing a causal graph based on system state data includes: processing the system state data based on a neural causal graph model to construct a causal graph, wherein the neural causal graph model includes a variational autoencoder or a self-attention encoder; processing the system state data based on the neural causal graph model includes: generating a feature vector for each variable in the system state data based on a variational autoencoder or a self-attention encoder; and processing the feature vector of each variable based on a neural network algorithm to determine the dependencies and causal action paths between variables.

[0008] In some embodiments of this application, determining a delay compensation control strategy based on a causal graph includes: processing the causal graph using a modular reinforcement learning controller to determine the delay compensation control strategy, wherein the modular reinforcement learning controller includes multiple pre-trained sub-policies and a policy selector based on causal structure embedding; processing the causal graph using the modular reinforcement learning controller includes: determining a causal structure embedding vector based on the causal graph; using the policy selector based on causal structure embedding to obtain the semantic similarity between the embedding vectors of multiple pre-trained sub-policies and the causal structure embedding vector, and using the sub-policy corresponding to the maximum value among the multiple semantic similarities as the delay compensation control strategy.

[0009] In some embodiments of this application, predicting residual synchronization error among multiple motors includes: modeling the actual output response data based on a time-series residual alignment network to predict the residual synchronization error; modeling the actual output response data based on the time-series residual alignment network includes: aligning the actual output response data of each motor with the corresponding target reference data in time, and inputting the preliminary residual between the actual output response data of each motor and the corresponding target reference data as an independent channel to a multi-channel parallel encoder; extracting high-frequency and hysteresis timing features based on the multi-channel parallel encoder; inputting multiple high-frequency and hysteresis timing features into a residual connection network layer to obtain a time residual path graph; analyzing the time residual path graph based on a time attention mechanism to obtain target features, wherein the target features are used to characterize the contribution points that cause the synchronization error, and the contribution points are the motor and time point that cause the residual; and outputting the residual synchronization error based on the target features through a fully connected layer.

[0010] In some embodiments of this application, the above-described cooperative control method for multi-motor systems further includes: generating target control commands based on causal graphs, delay compensation control strategies, and residual synchronization errors through a multi-model fusion cooperative optimization mechanism.

[0011] In some embodiments of this application, a target control command is generated through a multi-model fusion collaborative optimization mechanism, including: outputting a causal graph based on a neural causal graph model, wherein the causal graph includes causal path encoding and causal path weights; outputting a delay compensation control strategy and confidence score based on a modular reinforcement learning controller; outputting residual synchronization error and error importance score based on a temporal residual alignment network; using the causal path weights, confidence scores, and error importance scores as inputs, and outputting weight coefficients of the neural causal graph model, the modular reinforcement learning controller, and the temporal residual alignment network based on a preset optimization algorithm; and performing weighted fusion based on the output weight coefficients to generate the target control command.

[0012] In some embodiments of this application, the above-described cooperative control method for multi-motor systems further includes: updating at least one of the neural causal graph model, the modular reinforcement learning controller, and the temporal residual alignment network based on the actual output response data of the multi-motor drive system after delay compensation.

[0013] In some embodiments of this application, the above-described cooperative control method for a multi-motor system includes: obtaining a causal prediction deviation based on a comparison between actual output response data and disturbance prediction by a neural causal graph model, and adjusting the weights of causal action paths in the causal graph based on the causal prediction deviation; adjusting the network parameters of the online reinforcement learning algorithm in the modular reinforcement learning controller based on the deviation between the reward value calculated from the actual output response data and the expected reward value predicted by the modular reinforcement learning controller; and updating the weights of the temporal residual alignment network online using a backpropagation algorithm based on the actual change in residual error calculated from the actual output response data and the error adjustment value predicted by the temporal residual alignment network.

[0014] To achieve the above objectives, a second aspect of this application provides a vehicle including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described cooperative control method for a multi-motor system.

[0015] According to the vehicle in the embodiments of this application, by executing the above-described cooperative control method for a multi-motor system, the time-varying delay difference can be effectively suppressed through causal modeling and multi-level compensation, thereby reducing the transient mismatch of multi-motor output and improving the accuracy and stability of system synchronous control.

[0016] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0017] Figure 1 This is a flowchart of a cooperative control method for a multi-motor system according to some embodiments of this application.

[0018] Figure 2 This is a block diagram of a vehicle according to some embodiments of this application. Detailed Implementation

[0019] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0020] As described in the background section, to achieve dynamic synchronization between motors and suppress vehicle posture vibration caused by output mismatch, related technologies typically employ a control architecture where a unified controller issues synchronization commands. This approach generates and sends identical control commands to each motor driver through a main controller, assuming that each motor has consistent response characteristics to achieve synchronization. In actual operating conditions, the slight differences that may exist between the various execution channels depend on the consistency of hardware performance.

[0021] However, when this solution is applied to actual vehicle dynamics, especially under harsh conditions with drastic transient power demands such as starting, acceleration, or steering, microsecond-level response delays exist between motors due to hardware differences, sensor latency, CAN bus congestion, and temperature variations, resulting in transient mismatches in the actual output. This mismatch not only disrupts the coordinated output between motors but also causes slight vibrations in the vehicle's posture during starting, acceleration, and steering, which can severely reduce vehicle stability and ride comfort. Existing methods often lack the ability to model such minute nonlinear delay errors and lack structured cognitive mechanisms and scalable strategy scheduling frameworks, making it impossible to achieve high-precision delay compensation and synchronization control under dynamic and complex conditions.

[0022] Through in-depth analysis, the applicant discovered that existing technologies lack a preprocessing mechanism for microsecond-level timestamp alignment of multi-source asynchronous data from different controllers and sensors, resulting in a baseline error in the data foundation used to analyze delay differences. Furthermore, control logic is typically based on experience or simple feedback adjustment, failing to consider the causal relationship between delay differences in system response data and output mismatch, thus failing to effectively compensate for the root cause of the mismatch. In addition, residual, even smaller, synchronization errors are not considered after initial compensation.

[0023] To address the aforementioned issues, this application proposes a cooperative control method for multi-motor systems. By integrating a neural causal graph model, a modular reinforcement learning controller, and a temporal residual alignment network, a multi-motor cooperative control system with structural awareness, adaptive control, and error fine-tuning capabilities is constructed. This achieves microsecond-level dynamic synchronization compensation and optimal scheduling of control strategies, solving the technical problem of transient mismatch caused by delay differences in existing multi-motor systems under dynamic conditions. It also realizes high-precision and highly robust microsecond-level synchronization control.

[0024] The following description, with reference to the accompanying drawings, describes the cooperative control method and vehicle for a multi-motor system proposed in this application.

[0025] Figure 1 This is a flowchart of a cooperative control method for a multi-motor system according to an embodiment of this application.

[0026] like Figure 1As shown, the cooperative control method for a multi-motor system according to an embodiment of this application may include the following steps: S101, acquire system status data of the multi-motor drive system. The system status data includes timing data for determining the response delay differences between the multi-motors and vehicle dynamic response data for characterizing the output mismatch of the multi-motors.

[0027] Specifically, the raw data of multiple motors can be acquired first, including: motor command and timestamp data, motor operation data, vehicle operation data, and vehicle dynamic response data. Among them, motor command and timestamp data includes: timestamps of control commands issued through the vehicle controller (VCU) and each motor controller, and timestamps of motor response feedback signals received from the CAN bus or dedicated feedback channel; motor operation data includes: real-time speed, torque, three-phase current, bus voltage, and inverter drive pulse signals of each motor collected by the motor controller or sensors; vehicle operation data includes: actual wheel speeds of each drive wheel collected by wheel speed sensors; and vehicle dynamic response data includes: yaw rate, lateral acceleration, and pitch rate of change collected by the inertial measurement unit (IMU).

[0028] By preprocessing the raw data, such as performing time alignment, system status data can be obtained. For example, by comparing the timestamp of the same control command being issued with the timestamp of the actual response (such as speed change or current increase) of each motor, the response delay of each motor can be directly calculated. By comparing the delay values ​​of different motors, the difference in response delay between multiple motors can be determined.

[0029] S102, Based on system state data, construct a causal graph, whereby the causal graph is used to characterize the causal relationship between delay difference and multi-motor output mismatch.

[0030] Specifically, a causal graph can refer to a directed acyclic graph consisting of nodes and directed edges, where nodes represent system variables, directed edges represent the direction of causal action, and the weights on the edges represent the strength of the causal action. Examples include, but are not limited to: latent variable graphs learned through neural networks, structural equation model graphs obtained through causal discovery algorithms, or Bayesian network graphs obtained by combining expert knowledge with data.

[0031] As an example, a causal graph can be constructed using a pre-trained neural causal graph model. The system state data obtained in the above embodiments (e.g., time-series data used to determine the response delay differences between multiple motors: control commands, speeds, currents, voltages, and their precise timestamps for each motor; and vehicle dynamic response data characterizing output mismatch: yaw rate, lateral acceleration, and pitch rate of change) are used as input. The model internally employs a neural structural causal discovery network based on a combination of variational autoencoders and self-attention mechanisms. Each variable in the input data (e.g., the delay of motor A, the current of motor B, and the yaw rate) is modeled as a graph node. The network learns the directional connections between nodes in a non-linear manner, automatically discovering and quantifying the directed causal dependencies between these nodes, forming a directed acyclic graph as the causal graph. Examples include the indirect impact of the response delay of a specific motor on vehicle pitch disturbances, or the transient deviation coupling caused by the interaction between two motors.

[0032] S103, based on the causal graph, determine the delay compensation control strategy.

[0033] Specifically, a modular reinforcement learning controller can be used to acquire delay compensation control strategies. The controller pre-stores multiple sub-policies trained for different typical causal scenarios, such as single-motor delay-dominated scenarios, dual-motor delay coupling scenarios, and mismatch caused by communication jitter. A policy selector based on causal structure embedding calculates the semantic similarity (e.g., cosine similarity) between the current causal embedding vector (obtained from the causal graph) and the scenario embedding vector associated with each sub-policy, and selects the sub-policy with the highest similarity as the currently executed policy. For example, if the current causal graph shows that the diagonal sway of the vehicle body is mainly caused by delay coupling between the left front motor and the right rear motor, the policy selector will match a sub-policy trained for the diagonal coupling delay scenario. This sub-policy may output a specific set of compensation instructions aimed at decoupling the influence of these two motors. This enables fast and accurate scenario matching of the control strategy, significantly improving the system's adaptability and robustness in dealing with diverse mismatch modes.

[0034] In this process, a causal embedding encoder is used to assign initial features to each node of the causal graph terminal. Through a multi-layer message passing mechanism, the features of each node interact and aggregate with the features of its causal parent-child nodes, child nodes, and neighboring nodes. After multiple iterations, the features of each node contain the structural information of its local causal neighborhood. Finally, through a graph pooling operation (such as global average pooling), the information of all nodes is aggregated into a dense vector of fixed dimension, and a low-dimensional, continuous numerical vector, namely the causal embedding vector, is output. This causal embedding vector indicates which variables are the root causes, which are intermediate variables or the final result, the key causal paths, and the influence (relationship strength) of different causal relationships.

[0035] For example, in a causal scenario dominated by single-motor delay, the causal graph shows that a specific attitude mismatch of the vehicle (such as continuous pitch jitter) is strongly causally related to the delay of a single motor node (e.g., the left front motor), and this delay is much greater than that of other motors. The corresponding strategy could be to issue a fixed or linearly calculated timing advance command to the dominant delay motor (e.g., continuously advancing the execution time of its control commands by 20 microseconds), and issue a small reverse compensation to other motors to counteract any new torque imbalance that may arise after the dominant motor's compensation.

[0036] For causal scenarios involving delay coupling between two motors, the causal graph shows that system mismatch (such as periodic yaw) is not caused by a single motor, but by the combined effect of the delays of two motor nodes (e.g., the left front and right rear motors), with mutually influencing edges between these two nodes in the graph. In this case, coordinated compensation between the two motors is required. The policy network outputs a set of paired compensation commands based on the real-time delay difference, current phase difference, and other states of the two motors. For example, the command might be: advance the phase of motor A by θ degrees, while simultaneously delaying the phase of motor B by (θ-δ) degrees, where δ is a correction amount calculated based on the coupling strength, aiming to counteract their undesirable interaction.

[0037] For causal scenarios where communication jitter causes mismatch, the causal graph shows that the delays of multiple motors exhibit high-frequency, non-periodic common fluctuations (caused by CAN bus congestion, etc.), and these common fluctuations are strongly correlated with the high-frequency micro-jitter of the entire vehicle. In this case, it is necessary to suppress the mismatch caused by random jitter, rather than compensating for a fixed delay difference. A possible strategy is to first use an online filtering module to estimate the spectrum or statistical characteristics of the current communication jitter, and then, based on the filtering results, generate a set of high-frequency fine-tuning commands that are opposite to the jitter trend. For example, when a random increase in delay is detected, the strategy will issue a small-amplitude, rapidly changing dynamic adjustment command to offset the jitter and smooth out the output fluctuations caused by the jitter.

[0038] S104, based on the actual output response data of the multi-motor drive system after the execution of the delay compensation control strategy, predicts the residual synchronization error between the multi-motors.

[0039] Residual synchronization error refers to the minute time or state differences that still exist and are not completely eliminated among the actual output responses of multiple motors after the initial delay compensation control strategy is implemented. For example, on a microsecond time scale, it can refer to the difference in the time when the speed or torque of each motor reaches the target value, or it can refer to the phase difference between the output waveforms of each motor (such as sinusoidal current).

[0040] Specifically, residual synchronization error can be predicted using a temporal residual alignment network. After initial compensation using a delay compensation control strategy, residual synchronization error still exists, causing slight vehicle body vibration. The actual output response data is used as input, while referencing the ideal target response curve. The temporal residual alignment network employs a multi-channel parallel encoder structure to process the residual between the response sequence and the target sequence of each motor, extracting high-frequency jitter features and hysteresis patterns. These features are fused through a residual connection layer to form a temporal residual path map. Then, through a temporal attention mechanism, the network focuses on the key time points and specific motor channels that contribute most to the overall synchronization error. Finally, a vector is output through a fully connected layer, representing the predicted time difference or phase difference between the motors over a future period.

[0041] S105 compensates for multiple motors based on the predicted residual synchronization error.

[0042] Specifically, the predicted time offset of each motor can be directly converted into a fine-tuning of the timing of the control command issuance for the corresponding motor in the next cycle for compensation. For example, if it is predicted that the left front motor needs to be advanced by 1.2 microseconds, then in the next control cycle, the torque command calculation and issuance thread for the left front motor will be triggered 1.2 microseconds earlier. Alternatively, the time compensation can be equivalently converted into the phase advance angle of the phase current, which can be achieved by adjusting the phase angle of the space vector pulse width modulation. For example, for a 400Hz motor current fundamental frequency, a time difference of 1.2 microseconds corresponds to approximately a phase angle of 0.17 degrees, which can be compensated by fine-tuning the phase setpoint of the current loop.

[0043] Therefore, by acquiring and analyzing system state data that can be used to determine microsecond-level delay differences and characterize multi-motor output mismatch, and constructing a causal relationship graph between the data and output mismatch, the cause of mismatch is determined based on the causal graph, and a targeted compensation strategy is then determined. After executing the delay compensation strategy, the residual synchronization error is predicted and compensated, thereby offsetting the impact of time-varying delay differences, effectively reducing transient output mismatch between multiple motors, realizing microsecond-level dynamic synchronization control, and suppressing the resulting vehicle attitude vibration, thus improving vehicle stability and driving smoothness.

[0044] To optimize the initial quality of data input, in some embodiments of this application, system status data of a multi-motor drive system is obtained, including: obtaining raw data of the multi-motor system; and performing timestamp alignment processing on the raw data based on a high-precision synchronous clock to obtain system status data.

[0045] Specifically, a high-precision synchronization clock module (such as a PTP master clock) distributes synchronization signals to the VCU, motor controllers, sensors, and collaborative control computing platform via Ethernet or a dedicated clock line. Each node timestamps its data using the synchronization clock when generating data (such as sending CAN messages or sampling sensor values). Upon receiving the timestamped raw data, the collaborative control computing platform performs alignment processing. For example, using a fixed-frequency time grid as a reference, data from different sources at the same time with slightly different timestamps are aligned to a unified grid time point using interpolation algorithms (such as linear interpolation or spline interpolation), forming a synchronized multivariate time series. This alignment processing can be completed within a microsecond-level time resolution.

[0046] Therefore, by introducing a high-precision synchronous clock to perform unified timestamp alignment on all raw data, it is ensured that data from different controllers and sensors have a consistent and accurate microsecond-level time reference, effectively avoiding analysis errors caused by time asynchrony and improving the accuracy of subsequent processes.

[0047] In some embodiments of this application, constructing a causal graph based on system state data includes: processing the system state data based on a neural causal graph model to construct a causal graph, wherein the neural causal graph model includes a variational autoencoder or a self-attention encoder; processing the system state data based on the neural causal graph model includes: generating a feature vector for each variable in the system state data based on a variational autoencoder or a self-attention encoder; and processing the feature vector of each variable based on a neural network algorithm to determine the dependencies and causal action paths between variables.

[0048] Specifically, a neural causal graph model can be an end-to-end neural network, with input being a time-aligned, multivariate time series. Internally, an encoder (e.g., a one-dimensional convolutional neural network combined with self-attention) first learns a context-aware feature representation for each variable in the entire sequence. Then, a graph generation layer (e.g., a differentiable layer based on the NOTEARS algorithm) outputs an adjacency matrix based on these feature representations, where each element Aij in the adjacency matrix represents the causal strength from variable j to variable i; the matrix represents the structure and weights of the causal graph. During training, the neural causal graph model learns by optimizing an objective function that combines data fitting loss and graph sparsity constraints. Understandably, the encoder is not limited to variational autoencoders or self-attention encoders; sequence encoders such as graph convolutional networks and long short-term memory networks can also be used. The dependencies and causal paths between variables are ultimately reflected in the non-zero elements of the adjacency matrix and their values. By employing the aforementioned neural causal graph model, complex causal chains can be automatically identified, such as: motor temperature rise → inverter switching delay increases → torque response lag → vehicle body yaw.

[0049] Therefore, by employing a neural causal graph model that includes variational autoencoders or self-attention encoders, it is possible to perform deep feature extraction and representation learning on high-dimensional, nonlinear system state data. Then, neural network algorithms can be used to discover and quantify the dependencies and causal directions between variables, thereby constructing a causal structure graph.

[0050] In some embodiments of this application, determining a delay compensation control strategy based on a causal graph includes: processing the causal graph using a modular reinforcement learning controller to determine the delay compensation control strategy, wherein the modular reinforcement learning controller includes multiple pre-trained sub-policies and a policy selector based on causal structure embedding; processing the causal graph using the modular reinforcement learning controller includes: determining a causal structure embedding vector based on the causal graph; using the policy selector based on causal structure embedding to obtain the semantic similarity between the embedding vectors of multiple pre-trained sub-policies and the causal structure embedding vector, and using the sub-policy corresponding to the maximum value among the multiple semantic similarities as the delay compensation control strategy.

[0051] Specifically, causal structure embedding vectors can be obtained by encoding the causal graph, for example, by using a graph neural network to map the graph into a dense vector of fixed length. After training, each pre-trained sub-policy also encodes its behavioral features (such as the pattern of taking which action in which state) into a policy embedding vector through an encoding network. The policy selector based on the causal structure embedding calculates the semantic similarity (such as cosine similarity) between the current causal embedding vector and the policy embedding vector corresponding to each sub-policy in the library, and then selects the sub-policy with the highest similarity as the delay compensation control policy. By adopting a modular reinforcement learning controller and an embedding matching selection mechanism, the most suitable compensation policy can be found. For example, when the system identifies that the current scenario is a global small delay caused by communication jitter, it can quickly switch to a sub-policy that specifically handles this type of problem, instead of using a general policy that may not be effective.

[0052] Therefore, by encoding the real-time constructed causal graph into an embedding vector and semantically matching it with the pre-trained embedding vectors of each control sub-strategy for different causal scenarios, a mechanism for dynamically and quickly selecting the optimal control strategy based on the real-time causal structure characteristics is realized. This enables the control system to adaptively switch to the most suitable compensation strategy for different delay combination modes or disturbance scenarios, thereby significantly improving the robustness and control accuracy of the system in dealing with complex nonlinear mismatch phenomena and avoiding the limitations of using a single fixed strategy.

[0053] In some embodiments of this application, predicting residual synchronization error among multiple motors includes: modeling the actual output response data based on a time-series residual alignment network to predict the residual synchronization error; modeling the actual output response data based on the time-series residual alignment network includes: aligning the actual output response data of each motor with the corresponding target reference data in time, and inputting the preliminary residual between the actual output response data of each motor and the corresponding target reference data as an independent channel to a multi-channel parallel encoder; extracting high-frequency and hysteresis timing features based on the multi-channel parallel encoder; inputting multiple high-frequency and hysteresis timing features into a residual connection network layer to obtain a time residual path graph; analyzing the time residual path graph based on a time attention mechanism to obtain target features, wherein the target features are used to characterize the contribution points that cause the synchronization error, and the contribution points are the motor and time point that cause the residual; and outputting the residual synchronization error based on the target features through a fully connected layer.

[0054] Specifically, the actual response sequence (such as current waveform) newly acquired by each motor after executing the control strategy and the ideal target reference sequence are used as inputs to the timing residual network. First, for each motor, a preliminary residual sequence is calculated. Each preliminary residual sequence is treated as an independent channel and input into a multi-channel parallel encoder. This encoder can be composed of multiple parallel, identically structured one-dimensional causal convolutional neural networks. Each network is responsible for processing the residual sequence of one motor to extract the high-frequency noise components and hysteresis mode features specific to that channel relative to the control command. Next, the features extracted from all channels are concatenated along the feature dimension and input into a network layer containing residual connections. This layer constructs a time residual path graph through cross-channel information interaction and deep feature fusion. This time residual path graph encodes the temporal interaction relationships of the residuals of different motors. Then, the time attention mechanism module analyzes this path graph, calculates the importance weight of the features at each time step, and thus finds the contribution point that contributes the most to the overall synchronization error. For example, it may be found that at a certain time point, the residual of motor 3 suddenly increases and affects the subsequent response of motor 1. Based on the attention-weighted features, the data is fed into a fully connected layer for regression, and finally outputs a vector that directly predicts the microsecond-level time offset that each motor needs to compensate for in the next few time steps.

[0055] Therefore, by using a time-series residual alignment network, especially its multi-channel parallel coding structure, it is possible to finely separate and extract high-frequency and time-delay features that characterize small mismatches from the response sequences of each motor. Furthermore, by utilizing residual connections and time attention mechanisms, the key time points and motor channels that contribute the most to the overall synchronization error are identified, thereby achieving accurate modeling and quantitative prediction of microsecond-level residual synchronization errors. This effectively improves the final accuracy of synchronization control and solves the problem that traditional methods are unable to capture and predict such small residual errors.

[0056] In some embodiments of this application, the above-described cooperative control method for multi-motor systems further includes: generating target control commands based on causal graphs, delay compensation control strategies, and residual synchronization errors through a multi-model fusion cooperative optimization mechanism.

[0057] Specifically, based on the dynamic coupling relationship between motors represented by the causal graph, the advance phase compensation provided by the delay compensation control strategy, and the microsecond-level residual synchronization error predicted by the time-series residual alignment network, the multi-model fusion collaborative optimization mechanism first performs feature fusion. This involves dimensional alignment and normalization of the motor influence weight matrix output by the causal graph, the phase pre-adjustment of each motor generated by the delay compensation control strategy, and the residual synchronization error prediction vector, forming a unified fusion feature space. Then, a confidence-based dynamic weighted fusion strategy is adopted. Specifically, for the current operating conditions, the prediction confidence of each sub-model is evaluated in real time: when the system is in steady-state operation and the load fluctuation is small, the weight of the delay compensation control strategy is increased to fully leverage its accuracy advantage based on the physical model; when the system faces sudden disturbances or operating condition changes, the weight of the causal graph inference results is increased to utilize its rapid adaptability to dynamic coupling relationships; and in the fine-tuning stage of pursuing ultimate synchronization accuracy, the decision weight of the residual synchronization error prediction model is increased. This dynamic weighting mechanism is implemented through a learnable gating network. The input to the gating network includes the current system state characteristics, historical prediction error statistics of each sub-model, and operating condition identification labels. The output is the real-time fusion weight coefficients of the three sub-models. Next, the fused features are input into a multi-objective optimization solver, constructing a constrained optimization problem with the optimization objectives of minimizing synchronization error, optimizing energy consumption, and balancing mechanical stress. The optimization objective function comprehensively considers the weighted sum of squares of the speed tracking errors of each motor, the total system input power, and the variance of the output torque of each motor. Constraints include motor current limits, inverter switching frequency limits, and safe operating boundaries. A model predictive control framework is used to solve the problem in a finite time domain, obtaining the optimal control sequence for each motor in future control cycles. Finally, the optimization solution is converted into directly executable PWM duty cycle instructions by the instruction generation module, forming the target control instructions which are then sent to each motor drive unit.

[0058] Therefore, by designing a multi-model fusion and collaborative optimization mechanism, the model outputs obtained from three different dimensions—causal structure understanding, strategy decision-making, and real-time error prediction—are comprehensively processed and optimized. Through hierarchical and progressive information fusion and decision optimization, deep collaboration between physical model-driven and data-driven methods is achieved, taking into account the speed of control response, the high accuracy of steady-state precision, and the adaptive capability under complex working conditions. This results in optimal collaborative control performance across the entire working range of multi-motor systems.

[0059] In some embodiments of this application, a target control command is generated through a multi-model fusion collaborative optimization mechanism, including: outputting a causal graph based on a neural causal graph model, wherein the causal graph includes causal path encoding and causal path weights; outputting a delay compensation control strategy and confidence score based on a modular reinforcement learning controller; outputting residual synchronization error and error importance score based on a temporal residual alignment network; using the causal path weights, confidence scores, and error importance scores as inputs, and outputting weight coefficients of the neural causal graph model, the modular reinforcement learning controller, and the temporal residual alignment network based on a preset optimization algorithm; and performing weighted fusion based on the output weight coefficients to generate the target control command.

[0060] Specifically, the neural causal graph model not only outputs a causal graph but also assigns a weight value between 0 and 1 to each edge (causal path), representing the strength of the causal relationship or the current activation level. The confidence score can be the highest similarity value calculated by the policy selector or the reward value predicted based on the sub-policy in the current state. The importance score can be the sum of the attention weights assigned to each motor or time point by the internal attention mechanism of the temporal residual alignment network, representing the certainty of the error prediction or the degree of impact on the overall mismatch.

[0061] The fusion mechanism requires unified modeling of information from the three models to establish cross-model information mapping and decision coordination paths. The internal mechanism design adopts a cascaded fusion + collaborative supervision strategy framework. Structurally, a unified control decision representation space is first constructed. Information output from all sub-models is processed by an embedding transformation module and mapped to a unified high-dimensional control space. Based on this, a collaborative optimization network is used to weight and fuse the outputs of each model. This network takes the weights of causal paths, confidence scores, and error importance scores as inputs, dynamically adjusting the contribution ratio of each model to the final control command generation. For example, when the vehicle attitude changes drastically, the weight of the residual alignment network output is increased; when the delayed causal path is clear and the causal relationship of the disturbance is explicit, the guiding role of the causal graph model is strengthened; and in stages of abnormal motor behavior or frequent switching of control strategies, the robustness of the reinforcement learning controller is prioritized.

[0062] Therefore, by introducing a pre-defined optimization algorithm, and based on the internal confidence index output by each sub-model, the fusion weights of each model are dynamically calculated and allocated in the final decision, thus achieving adaptive weighted fusion. This enables the system to intelligently adjust the decision focus under different operating conditions. For example, it emphasizes causal guidance when the causal relationship is clear, and emphasizes error correction when error disturbances are significant. This ensures that the fused control commands always adapt to the most important system characteristics and control requirements, improving the intelligence, flexibility, and scenario adaptability of the fusion decision.

[0063] In some embodiments of this application, the above-described cooperative control method for multi-motor systems further includes: updating at least one of the neural causal graph model, the modular reinforcement learning controller, and the temporal residual alignment network based on the actual output response data of the multi-motor drive system after delay compensation.

[0064] Specifically, updates can be performed periodically or triggered. For example, the model can be fine-tuned and updated every 100 kilometers or every 24 hours using data collected during that period. A trigger condition could be when control performance indicators (such as the average synchronization error) are consistently above a threshold for a period of time. The actual output response data required for the update, the actual synchronization state achieved by the system, and vehicle attitude data are compared with the model's predictions to generate feedback signals that can be used for updates.

[0065] By employing the aforementioned closed-loop update mechanism, the model can continuously track changes in the dynamic characteristics of the system. For example, as the motor bearings wear down, their mechanical time constant may change slightly, causing a drift in delay characteristics. Through updates, the neural causal graph model can learn new causal relationships, the modular reinforcement learning controller can adjust its strategy to cope with new delay patterns, and the temporal residual alignment network can adapt to new error patterns.

[0066] Therefore, by using closed-loop feedback data generated during actual operation to update the core model online or periodically, the causal graph, control strategy, and error prediction model can continuously evolve and optimize in accordance with changes in system dynamic characteristics, component states, or environmental conditions. This overcomes the performance degradation problem that may occur with fixed parameter models and ensures the long-term, stable, and high-precision compensation capability of the control system for microsecond-level delay mismatch throughout its entire life cycle.

[0067] In some embodiments of this application, the above-described cooperative control method for a multi-motor system includes: obtaining a causal prediction deviation based on a comparison between actual output response data and disturbance prediction by a neural causal graph model, and adjusting the weights of causal action paths in the causal graph based on the causal prediction deviation; adjusting the network parameters of the online reinforcement learning algorithm in the modular reinforcement learning controller based on the deviation between the reward value calculated from the actual output response data and the expected reward value predicted by the modular reinforcement learning controller; and updating the weights of the temporal residual alignment network online using a backpropagation algorithm based on the actual change in residual error calculated from the actual output response data and the error adjustment value predicted by the temporal residual alignment network.

[0068] Specifically, the neural causal graph model predicts the expected change trend of each state variable in the next control cycle based on the current system state variables (including the speed, torque, current, temperature of each motor, and vehicle operating parameters) and detected external disturbance signals. Simultaneously, the system collects actual output response data through high-precision sensors, including the actual speed and torque output of each motor, as well as the actual fluctuations in system bus voltage and current. The predicted output of the neural causal graph model is compared dimension-by-dimensionally with the actual output response data, calculating the comprehensive deviation in the time and frequency domains. This comprehensive deviation is the causal prediction deviation. When the causal prediction deviation exceeds a preset first threshold, it indicates that the weights of some causal paths in the current causal graph fail to accurately reflect the true physical coupling relationship. At this point, the system initiates an online adjustment mechanism for the causal graph: using an attention-based weight correction algorithm, gradient descent optimization is performed on the weights of the causal paths between nodes in the causal graph. Specifically, minimizing the causal prediction deviation is used as the objective function; the partial derivatives of each causal path weight with respect to the objective function are calculated, and iterative updates are performed according to a preset learning rate. For causal paths that consistently exhibit high prediction bias, the system will also trigger adaptive adjustments at the structural level, including adding or removing causal connections or introducing new latent variable nodes, to more comprehensively characterize the nonlinear and time-varying coupling characteristics of the multi-motor system. For example, the causal graph model might predict that a 10μs increase in the right rear motor delay will lead to a 0.5° / s increase in yaw rate. After actual compensation, the collected yaw rate change is only 0.3° / s. This bias (0.2° / s) can be used to adjust the weight of the edge from the right rear motor delay node to the yaw rate node, slightly reducing it to better reflect the actual causal strength of the current system.

[0069] A modular reinforcement learning controller predicts a desired cumulative reward (Q-value) for the current state and the action taken. After the action is executed, an actual reward value can be calculated based on the improvement in synchronization accuracy and energy consumption. Using the deviation between these two values ​​(temporal differential error), the parameters of the policy network can be incrementally updated online or offline using reinforcement learning algorithms such as policy gradient. This allows the control policy to be optimized based on actual control performance, continuously approaching the optimal policy.

[0070] The temporal residual alignment network predicts the error adjustment value for the next time step based on historical error sequences and future control command sequences. This prediction is used for feedforward compensation to suppress the impact of microsecond-level delay mismatch. At the end of each control cycle, the system calculates the actual change in residual error, i.e., the difference between the actual tracking error at the current time and the actual tracking error at the previous time. The actual change is compared with the error adjustment value predicted by the network to form a predicted error signal. This predicted error signal drives the weight update of the temporal residual alignment network through a backpropagation algorithm. For example, the network predicts a compensation of +1.2 μs for the left front motor. After this compensation, high-precision sensor measurements show that the actual synchronization error of the left front motor has decreased by 1.0 μs. The deviation (+0.2 μs) between this predicted value (+1.2 μs) and the actual result (+1.0 μs), along with the corresponding input data, can constitute a training sample. This sample is then used to perform a small gradient update on the weights of the temporal residual alignment network through the backpropagation algorithm, making its future predictions more accurate.

[0071] Therefore, by designing differentiated online update algorithms for the causal graph model, reinforcement learning controller, and temporal residual alignment network, tailored to their respective model characteristics and learning mechanisms, precise and efficient fine-tuning of the core parameters or structures of each model was achieved. This differentiated update strategy ensures that each model can be effectively optimized based on its own error feedback and learning rules, thereby achieving continuous improvement and real-time adaptation of the overall control system performance with lower computational cost and targeted adjustments, guaranteeing the effectiveness and efficiency of the closed-loop update mechanism.

[0072] In summary, the cooperative control method for a multi-motor system according to the embodiments of this application acquires system state data of the multi-motor drive system. This system state data includes time-series data for determining the response delay differences between the multiple motors and vehicle dynamic response data for characterizing the output mismatch of the multiple motors. Based on the system state data, a causal graph is constructed, whereby the causal graph characterizes the causal relationship between delay differences and multi-motor output mismatch. Based on the causal graph, a delay compensation control strategy is determined. Then, based on the actual output response data of the multi-motor drive system after implementing the delay compensation control strategy, the residual synchronization error between the multiple motors is predicted, and compensation is applied to the multiple motors according to the predicted residual synchronization error. Therefore, this method can effectively suppress time-varying delay differences through causal modeling and multi-level compensation, reducing transient mismatch in multi-motor output and improving the accuracy and stability of system synchronization control.

[0073] Corresponding to the above embodiments, this application also proposes a vehicle.

[0074] like Figure 2As shown, the vehicle 200 in this embodiment may include: a memory 210, a processor 220, and a program stored in the memory 210 and executable on the processor 220. When the processor 220 executes the program, it implements the above-described cooperative control method for a multi-motor system.

[0075] According to the vehicle in the embodiments of this application, by executing the above-described cooperative control method for a multi-motor system, the time-varying delay difference can be effectively suppressed through causal modeling and multi-level compensation, thereby reducing the transient mismatch of multi-motor output and improving the accuracy and stability of system synchronous control.

[0076] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0077] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0078] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0079] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0080] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0081] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0082] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A cooperative control method for a multi-motor system, characterized in that, The method includes: The system status data of the multi-motor drive system is obtained, including timing data for determining the response delay differences between the multi-motors and vehicle dynamic response data for characterizing the output mismatch of the multi-motors. Based on the system state data, a causal graph is constructed, wherein the causal graph is used to characterize the causal relationship between the delay difference and the multi-motor output mismatch; Based on the causal graph, a delay compensation control strategy is determined; Based on the actual output response data of the multi-motor drive system after executing the delay compensation control strategy, the residual synchronization error between the multi-motors is predicted. The multiple motors are compensated based on the predicted residual synchronization error.

2. The cooperative control method for a multi-motor system according to claim 1, characterized in that, Acquiring the system status data of the multi-motor drive system includes: Obtain the raw data of the multiple motors; The original data is timestamped and aligned using a high-precision synchronous clock to obtain the system status data.

3. The cooperative control method for a multi-motor system according to claim 1, characterized in that, Based on the system state data, a causal graph is constructed, including: The system state data is processed based on a neural causal graph model to construct the causal graph, wherein the neural causal graph model includes a variational autoencoder or a self-attention encoder. The processing of the system state data based on the neural causal graph model includes: Based on the variational autoencoder or self-attention encoder, a feature vector is generated for each variable in the system state data; The feature vectors of each variable are processed using neural network algorithms to determine the dependencies and causal paths between variables.

4. The cooperative control method for a multi-motor system according to claim 1, characterized in that, Based on the causal graph, a delay compensation control strategy is determined, including: The causal graph is processed based on a modular reinforcement learning controller to determine the delay compensation control policy, wherein the modular reinforcement learning controller includes multiple pre-trained sub-policies and a policy selector based on causal structure embedding. The process of processing the causal graph based on the modular reinforcement learning controller includes: The causal structure embedding vector is determined based on the causal graph; A policy selector based on causal structure embedding is used to obtain the semantic similarity between the embedding vectors of multiple pre-trained sub-policies and the causal structure embedding vector, and the sub-policy corresponding to the maximum value among the multiple semantic similarities is used as the delay compensation control policy.

5. The cooperative control method for a multi-motor system according to claim 1, characterized in that, The prediction of the residual synchronization error among the multiple motors includes: The actual output response data is modeled based on a time-series residual alignment network to predict the residual synchronization error; Modeling the actual output response data based on a temporal residual alignment network includes: The actual output response data of each motor is time-aligned with the corresponding target reference data, and the preliminary residual between the actual output response data of each motor and the corresponding target reference data is used as an independent channel and input to the multi-channel parallel encoder. High-frequency and hysteresis timing features are extracted based on the multi-channel parallel encoder; Multiple high-frequency and lag time-series features are input into the residual connection network layer to obtain a time residual path graph; The time residual path graph is analyzed based on the time attention mechanism to obtain target features, wherein the target features are used to characterize the contribution points that cause synchronization error, and the contribution points are the motor and time points that cause the residual. Based on the target features, the residual synchronization error is output through a fully connected layer.

6. The cooperative control method for a multi-motor system according to claim 1, characterized in that, The method further includes: Based on the causal graph, the delay compensation control strategy, and the residual synchronization error, the target control command is generated through a multi-model fusion collaborative optimization mechanism.

7. The cooperative control method for a multi-motor system according to claim 6, characterized in that, Target control commands are generated through a multi-model fusion and collaborative optimization mechanism, including: The causal graph is output based on the neural causal graph model, wherein the causal graph includes causal path encoding and causal path weights; The delay compensation control strategy and confidence score are output based on the modular reinforcement learning controller. The residual synchronization error and error importance score are output based on the temporal residual alignment network; Based on a preset optimization algorithm, the weights of the causal path, the confidence score, and the error importance score are taken as inputs, and the weight coefficients of the neural causal graph model, the modular reinforcement learning controller, and the temporal residual alignment network are output. The target control command is generated by weighting and fusing the output weight coefficients.

8. The cooperative control method for a multi-motor system according to claim 7, characterized in that, The method further includes: Based on the actual output response data of the multi-motor drive system after delay compensation, at least one of the neural causal graph model, the modular reinforcement learning controller, and the temporal residual alignment network is updated.

9. The cooperative control method for a multi-motor system according to claim 8, characterized in that, The method includes: Based on the comparison between the actual output response data and the perturbation prediction of the neural causal graph model, the causal prediction bias is obtained, and the weights of the causal action paths in the causal graph are adjusted based on the causal prediction bias. The network parameters of the online reinforcement learning algorithm in the modular reinforcement learning controller are adjusted based on the deviation between the reward value calculated based on the actual output response data and the expected reward value predicted by the modular reinforcement learning controller. Based on the actual output response data, the actual change in residual error is calculated and the error adjustment value predicted by the temporal residual alignment network is compared with the error adjustment value predicted by the temporal residual alignment network. The weights of the temporal residual alignment network are then updated online using the backpropagation algorithm.

10. A vehicle, characterized in that, include: A memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the cooperative control method for a multi-motor system according to any one of claims 1-9.

Images