Differential lock control method based on multi-model fusion

Abstract

The invention discloses a differential lock control method based on multi-model fusion, and the method comprises the steps: obtaining the multi-source sensing data of a vehicle, generating a standardized time sequence tensor based on the multi-source sensing data, and enabling the standardized time sequence tensor to be a time sequence composed of the multi-dimensional state features of the vehicle in a fixed time window; inputting the time series data into a space-time attention feature extraction model, and outputting a dynamic feature vector representing the steering stability of the vehicle; based on a lightweight differential lock strategy decision model and the dynamic feature vector, a differential lock control instruction is determined, the differential lock control instruction is sent to a differential lock controller, the problems that in the prior art, differential lock control lags behind, erroneous judgment occurs frequently, and cross-vehicle-type adaptability is poor are solved, and the real-time performance, accuracy and robustness of control are improved.

Description

Technical Field

[0001] This application relates to the field of automotive electronics technology, and in particular to a differential lock control method based on multi-model fusion. Background Technology

[0002] As vehicle intelligence continues to improve, differential locks, as key components for enhancing off-road capability and steering stability, directly impact the vehicle's dynamic performance under complex conditions due to the intelligence level of their control strategies. However, current mainstream differential lock control technologies still rely on traditional calibration methods based on engineers' experience, making it difficult to cope with the changing vehicle states and road environments in real-world driving scenarios. This results in significant deficiencies in the adaptability, generalization, and real-time performance of the control strategies.

[0003] Specifically, current intelligent differential lock control strategies deployed in vehicles generally employ open-loop control methods based on static rules or fixed mapping curves. These methods rely on threshold logic or lookup table curves pre-set by engineers, such as triggering locking when the wheel speed difference exceeds a certain fixed threshold, or closing at a fixed ratio within a specific vehicle speed range. However, in real-world vehicles operating under complex conditions such as low-adhesion cornering and hill-start assist, there is a strong dynamic coupling relationship between steering stability, wheel slip, vehicle attitude changes, and driving force distribution. A single threshold or fixed curve cannot respond in real time to the nonlinear changes of multi-dimensional sensing characteristics, leading to control lag or frequent misjudgments. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a differential lock control method based on multi-model fusion.

[0005] The embodiments of this application disclose the following technical solutions: The first aspect of this application provides a differential lock control method based on multi-model fusion, including: Acquire multi-source sensor data of the vehicle, and generate a standardized time series tensor based on the multi-source sensor data. The standardized time series tensor is a time series sequence composed of multi-dimensional state features of the vehicle within a fixed time window. The time-series data is input into the spatiotemporal attention feature extraction model, which outputs a dynamic feature vector characterizing the vehicle's steering stability. The spatiotemporal attention feature extraction model is used to perform forward and backward temporal modeling and dynamic weighting of sensor features on the standardized time-series tensor to generate a dynamic feature vector that integrates spatiotemporal information. Based on the lightweight differential lock strategy decision model and the dynamic feature vector, the differential lock control command is determined. The lightweight differential lock strategy decision model is a lightweight neural network obtained by compressing a cross-vehicle differential lock strategy transfer learning model as the teacher network through a knowledge distillation mechanism. The differential lock control command is sent to the differential lock controller.

[0006] Embodiment two of this application provides a differential lock control system based on multi-model fusion, including: The acquisition unit is used to acquire multi-source sensor data of the vehicle and generate a standardized time series tensor based on the multi-source sensor data. The standardized time series tensor is a time series sequence composed of multi-dimensional state features of the vehicle within a fixed time window. The extraction unit is used to input the time series data into the spatiotemporal attention feature extraction model to extract dynamic feature vectors characterizing vehicle steering stability. The spatiotemporal attention feature extraction model is used to perform forward and backward temporal modeling and dynamic weighting of sensor features on the standardized time series tensor to generate dynamic feature vectors that fuse spatiotemporal information. The determining unit is used to determine the differential lock control command based on the lightweight differential lock strategy decision model and the dynamic feature vector. The lightweight differential lock strategy decision model is a lightweight neural network obtained by compressing a cross-vehicle differential lock strategy transfer learning model as the teacher network through a knowledge distillation mechanism. The sending unit is used to send the differential lock control command to the differential lock controller.

[0007] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the differential lock control method based on multi-model fusion as described in the first aspect above.

[0008] The fourth aspect of this application provides a computer program product, which, when run on a computer, executes the differential lock control method based on multi-model fusion as described in the first aspect above.

[0009] The fifth aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the differential lock control method based on multi-model fusion as described in the first aspect above.

[0010] Compared with the prior art, this application has the following beneficial effects: This application captures the continuous evolution sequence of multi-dimensional state features of a vehicle by acquiring multi-source sensor data from the vehicle and generating a standardized temporal tensor within a fixed time window, thus avoiding judgment bias caused by a single feature from the data source. Subsequently, the temporal tensor is input into a spatiotemporal attention feature extraction model, and a dynamic feature vector that integrates spatiotemporal information and characterizes the vehicle's steering stability is generated by utilizing forward and backward temporal modeling and a dynamic weighting mechanism of sensor features. This effectively captures the dynamic coupling relationship between multi-dimensional parameters under complex operating conditions, solving the problem that traditional methods cannot perceive the dynamic state of the vehicle in real time. Based on a lightweight differential lock strategy decision model obtained by knowledge distillation and compression using a cross-vehicle differential lock strategy transfer learning model as the teacher network, differential lock control commands are generated in real time according to the dynamic feature vector. The lightweight neural network can complete inference at the vehicle end in milliseconds without waiting for a large model or relying on fixed rules, realizing a real-time response to the nonlinear changes of multi-dimensional features under complex operating conditions, and significantly reducing control lag and the probability of misjudgment. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application 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 only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating a differential lock control method based on multi-model fusion provided in an embodiment of this application; Figure 2 A flowchart of another differential lock control method based on multi-model fusion provided in an embodiment of this application; Figure 3 This is a structural diagram of a differential lock control system based on multi-model fusion provided in an embodiment of this application. Detailed Implementation

[0013] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0014] To facilitate understanding of the technical solutions provided in the embodiments of this application, the technical terms involved in the embodiments of this application will be explained below.

[0015] STAM is a core model for extracting dynamic features from multi-source vehicle sensor data. Its full name is Spatio-Temporal Attention Module. This model takes a standardized temporal tensor as input and performs forward and backward temporal modeling through a bidirectional gated recurrent unit to capture the evolution of the vehicle's state over time. Simultaneously, a spatial attention network dynamically weights the multi-dimensional sensor features at each time step, highlighting key variables strongly correlated with steering stability. Finally, the temporal context features and spatially weighted features are fused to output a high-dimensional dynamic feature vector representing the vehicle's steering stability, providing a unified spatio-temporal perception representation for subsequent policy models.

[0016] TransferBERT is the core model for cross-vehicle knowledge transfer in this application. Its full English name is TransferBERT, and its Chinese name is Cross-Vehicle Differential Lock Policy Transfer Learning Model. This model is built on the Transformer architecture, introducing vehicle tokens that encode platform information such as vehicle hardware structure and electric drive type into the input sequence. Feature modeling is performed through a multi-layer Transformer encoder, and a cross-head attention enhancement structure is adopted, allowing some attention heads to focus on dynamic sequence features while others learn the policy mapping relationships between vehicle models. The model employs a two-stage training strategy of supervised pre-training of the source vehicle and fine-tuning with a small number of samples of the target vehicle, achieving efficient transfer of differential lock control strategies between different vehicle platforms and significantly reducing the calibration cost of new vehicle models.

[0017] KD-Net is a lightweight policy model deployed on the vehicle-side edge controller in this application. Its full English name is Knowledge Distillation Network, and its Chinese name is Lightweight Differential Lock Policy Decision Model. This model uses a pre-trained TransferBERT network as the teacher network and performs compressed training through a soft-label knowledge distillation mechanism. The distillation loss function uses mean squared error and introduces policy consistency constraints to ensure that the student's output direction remains consistent with the teacher network under critical boundary conditions. The main body of the model consists of a single-layer unidirectional gated recurrent unit and two layers of multilayer perceptrons. The number of parameters is compressed to the hundreds of thousands level. After quantization and edge deployment optimization such as structural pruning, inference can be completed in milliseconds, meeting the real-time control requirements of the vehicle.

[0018] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0019] As mentioned earlier, traditional control strategies rely on static rules or fixed mapping curves, which cannot handle the multi-dimensional dynamic characteristics of vehicles in scenarios such as low-adhesion turning and hill-start assist. Furthermore, they lack generalization ability for different vehicle structures. In addition, existing models are bulky and difficult to deploy at the edge, resulting in insufficient real-time performance.

[0020] To address the aforementioned issues, this application provides a differential lock control method based on multi-model fusion, see [link to relevant documentation]. Figure 1 , Figure 1 This is a schematic diagram of the differential lock control method based on multi-model fusion provided in the embodiments of this application. First, through vehicle dynamic data acquisition and preprocessing steps, the original sensor data is processed into a standardized temporal tensor, providing a unified data foundation for subsequent models and solving the problem of difficulty in fusing multi-source heterogeneous data. Then, the standardized tensor is input into the STAM spatiotemporal attention feature extraction model. Through forward and backward temporal modeling and spatial attention weighting, the model simultaneously captures the temporal evolution law of the vehicle state and the dynamic coupling relationship between multiple sensor features. This allows for early detection of steering instability trends in complex scenarios such as low-adhesion turning and hill start recovery, effectively solving the problem that traditional methods cannot perceive multi-dimensional dynamic features in real time.

[0021] Building upon this foundation, two sets of control proposals are output in parallel using the TransferBERT cross-vehicle strategy transfer model and the KD-Net lightweight strategy distillation model. The former utilizes the Transformer architecture and employs a two-stage training strategy of pre-training on the source vehicle and fine-tuning on the target vehicle to deeply integrate vehicle structure information into the decision-making process, achieving efficient transfer of control strategies across different platforms and solving the problems of high calibration costs and poor generalization ability for new vehicle models. In this application, the lightweight differential lock strategy decision model remains online in real time under all vehicle operating modes, with its inference time controlled within 5ms. It can independently generate differential lock control commands, thus overcoming the bottleneck of large models being unable to be deployed in real time at the edge. The cross-vehicle differential lock strategy transfer learning model is selectively activated only when vehicle-side resources are idle or under specific triggering conditions. Its output is used to verify, correct, or fuse the decisions of the lightweight differential lock strategy decision model and is not required to be activated in every control cycle. This master-slave architecture, which defaults to real-time control of the lightweight model and enhances the large model when necessary, ensures millisecond-level response speed while fully utilizing the cross-vehicle generalization ability of the large model. The two sets of strategy outputs and vehicle status enter the strategy fusion scheduling mechanism together. Through optimization scoring, confidence re-scoring and consistency verification, reliable differential lock control commands are generated to further avoid the judgment bias of a single model. The commands are converted into multi-domain collaborative control command packages through the differential lock control mapping matrix and interface to realize the linkage adjustment of differential lock, steering, drive, suspension and other systems.

[0022] Finally, by using online model tuning and a closed-loop optimization mechanism, deviation analysis and personalized correction are performed based on execution logs and driving style characteristics. OTA enables continuous iterative upgrades of the strategy, forming a complete technical closed loop from data collection, feature extraction, strategy generation, fusion decision-making, collaborative execution to closed-loop optimization. This systematically solves the core pain points of existing technologies, such as control lag, frequent misjudgments, and poor adaptability.

[0023] It should be noted that the differential lock control method, system, product, device, and medium based on multi-model fusion provided in this application can be applied to the field of computer technology. The above are merely examples and do not limit the application areas of the differential lock control method, system, product, device, and medium based on multi-model fusion provided in this application. Furthermore, the embodiments of this application may not limit the executing entity of the differential lock control based on multi-model fusion. For example, the differential lock control method based on multi-model fusion in the embodiments of this application can be applied to data processing devices such as terminal devices or servers. The terminal device can be an electronic device such as a computer or a personal digital assistant (PDA). The server can be a standalone server, a cloud server, or a cluster server composed of multiple servers.

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0025] The following embodiment illustrates a differential lock control method based on multi-model fusion provided in this application. See also... Figure 1 ,Should Figure 1 A flowchart of a differential lock control method based on multi-model fusion provided in this application embodiment, the method including: S101. Acquire multi-source sensor data of the vehicle and generate a standardized time-series tensor based on the multi-source sensor data.

[0026] The standardized temporal tensor is a standardized temporal sequence composed of multi-dimensional vehicle state features within a fixed time window, after temporal synchronization, anomaly cleaning, and normalization. This step aims to construct a complete data acquisition and preprocessing workflow to provide unified and standardized input data for subsequent STAM, TransferBERT, and KD-Net models. The core objective of the entire workflow is to collect full dynamic data from the real vehicle environment that accurately reflects vehicle steering stability, wheel-end power distribution, and differential lock intervention effects, and to adapt it to the input requirements of subsequent models through a series of processing steps.

[0027] Data acquisition is conducted on the automaker's test vehicles and mass-production verification vehicles. These vehicles are equipped with a vehicle domain controller, differential lock controller, electric drive system controller, EPS steering system, ESC stability system, and IMU inertial navigation unit. All signals are synchronously transmitted back to the central data recording module via high-speed CAN-FD bus and Ethernet. The collected raw signals cover multiple dimensions, including steering-related parameters such as steering wheel angle, steering wheel speed, and EPS steering torque; motion-related parameters such as vehicle speed, four-wheel wheel speed, yaw rate, lateral acceleration, longitudinal acceleration, vehicle heading angle, and body roll angle; power distribution-related parameters such as front and rear axle output torque, left and right half-axle torque difference, motor speed, motor current, and battery output power; differential lock status-related parameters such as closure ratio; stability control-related parameters such as wheel slip ratio and ESC intervention level; and terrain label information, which is provided by the automaker's self-built road surface recognition model and can distinguish between dry asphalt, wet asphalt, ice and snow, gravel, and mixed adhesion road surfaces. All signals are collected with a minimum sampling period of 5ms and written to the buffer at the vehicle end using a unified timestamp to form the original multidimensional time series data stream.

[0028] After the initial data acquisition, multi-source synchronous calibration is performed first. A master clock alignment mechanism is used to interpolate and align data from the CAN bus, IMU, and differential lock controller, ensuring that all sensor channels within the same time slice reflect the vehicle's physical state at the same moment. Next, anomaly removal is performed. For common issues such as wheel speed spikes, motor current spikes, and IMU drift, a sliding window midpoint filtering combined with threshold trimming is used to mark abnormal sampling points that clearly exceed the vehicle's physical limits as invalid and replace them with valid data from nearby time points. After anomaly cleaning, low-pass filtering is applied to continuous signals to eliminate high-frequency noise interference while preserving key trends during the initial steering phase and the moment of slippage, avoiding the loss of useful information due to over-filtering.

[0029] To enhance the model's ability to anticipate impending vehicle instability, first-order differential features and short-term rate-of-change features are generated in addition to the basic sensor data. These features include vehicle speed change rate, yaw acceleration, wheel speed difference gradient, and differential lock closure rate. All basic and derived features are combined into a fixed-dimensional feature vector, with a single-frame feature dimension of 32. Based on this, a sliding time window approach is used to construct a sample sequence, with 100 consecutive frames as a sample unit, corresponding to approximately 0.5 seconds of vehicle dynamic behavior. Each sample forms a 100×32 temporal matrix, which serves as the direct input to the subsequent STAM model.

[0030] It should be noted that the aforementioned features, such as vehicle speed change rate, yaw acceleration, wheel speed difference gradient, and differential lock closure rate, are input into the spatiotemporal attention feature extraction model STAM as key dimensions of the standardized temporal tensor during subsequent processing. STAM learns the evolutionary patterns of these transient features through bidirectional temporal modeling, enabling the model to capture early signs of potential instability signals such as wheel-end slip acceleration and escalating yaw instability. Simultaneously, the spatial attention mechanism automatically assigns higher weights to feature dimensions with abnormally high rates of change, thereby enhancing the model's sensitivity to impending instability. Based on this, the dynamic feature vector output by STAM is further passed to the cross-vehicle differential lock strategy transfer learning model and the lightweight differential lock strategy decision model. The cross-vehicle differential lock strategy transfer learning model can leverage cross-vehicle transfer capabilities to correlate transient trends with the control boundaries of different vehicle platforms, while the lightweight differential lock strategy decision model can prioritize responses to dangerous conditions with sudden increases in the rate of change during lightweight inference. The resulting differential lock control strategy enables advanced prediction and early intervention, significantly improving the timeliness and accuracy of control commands and effectively avoiding control delays or misjudgments caused by perception lag.

[0031] During the normalization phase, automakers, based on their own million-level real-vehicle data, calculate the average and standard deviation of each feature across different vehicle platforms and establish an independent normalization parameter table for each model. Whether processing online or offline, the system automatically matches the corresponding parameter table based on the vehicle's VIN code, performs Z-score standardization on all continuous features, and unifies the data distribution across different models and powertrain systems to a stable range. For discrete features such as terrain labels and ESC intervention levels, one-hot encoding is used to map them into numerical vectors. After normalization, each sample is uniformly encapsulated into a structured tensor with time and feature dimensions, along with metadata labels such as vehicle ID, drive type, and motor type, for subsequent model calls upon as needed. This entire preprocessing workflow ensures the accuracy, consistency, and cross-model comparability of the input data, laying a solid data foundation for the training and deployment of the intelligent differential lock strategy model.

[0032] S102. Input the time series data into the spatiotemporal attention feature extraction model to extract dynamic feature vectors characterizing vehicle steering stability.

[0033] The Spatio-Temporal Attention Module (STAM) is used to perform forward and backward temporal modeling and dynamic weighting of sensor features on the standardized temporal tensor, generating a dynamic feature vector that integrates spatiotemporal information. It aims to extract implicit dynamic features closely related to differential lock adjustment from continuous vehicle motion states. The core value of this model lies in overcoming the limitations of traditional control methods that struggle to simultaneously analyze temporal evolution patterns and the coupling relationships of multiple sensor features, enabling the system to detect potential instability trends caused by wheel slippage, changes in vehicle attitude, and the distribution of driving force in advance.

[0034] In one possible implementation, the spatiotemporal attention feature extraction model specifically includes: first, performing temporal modeling on the standardized temporal tensor from two consecutive time directions to capture the temporal correlation of vehicle state features and output temporal context features; then, inputting the sensor feature vector of each time step into a fully connected attention network, and after activation function processing and normalization, generating importance weight vectors corresponding to each sensor variable at the current time, multiplying the weight vector element-wise with the original sensor features of the corresponding time step to obtain a weighted feature representation highlighting key sensor information; subsequently, concatenating the above temporal context features and weighted feature representations, retaining the original feature information flow through residual connections to avoid feature loss, and then inputting it into a fully connected fusion network for nonlinear mapping, and after normalization layer processing to obtain a dynamic feature vector of fused spatiotemporal information; finally, using a multilayer perceptron to perform feature compression on the dynamic feature vector of fused spatiotemporal information, and outputting a spatiotemporal embedding sequence as the dynamic feature vector input to the subsequent model.

[0035] For example, the STAM model receives a normalized temporal tensor generated in the previous stage as input. This tensor has a size of 100×32, corresponding to the 32-dimensional state features of the vehicle within 0.5 seconds. After a series of processing steps, the model outputs a high-dimensional embedding sequence with a size of 100×256, which will be directly used in the subsequent policy transfer network. The entire feature extraction process begins with temporal modeling, employing a two-layer bidirectional gated recurrent unit to learn the evolutionary patterns of the input sequence in both forward and backward directions. This design enables the model not only to review historical wheel speed and yaw changes but also to capture impending steering instability signals, ultimately outputting a contextual feature matrix containing complete temporal dependencies, with each time step incorporating information from both forward and backward directions.

[0036] After completing temporal modeling, the model enters the spatial attention processing stage. This stage redistributes the importance of the 32-dimensional original features at each time step. Specifically, the feature vector of each frame is input into a two-layer fully connected attention network. After activation and normalization, the importance weight distribution of each sensor variable at the current time is generated. For example, in the early stage of turning on a low-adhesion road surface, the model automatically increases the weight of key features such as wheel speed difference, lateral acceleration, and slip ratio, while decreasing the weight of weakly correlated features such as vehicle speed. Then, this weight vector is multiplied element-wise with the original features to obtain a weighted feature representation, thereby achieving dynamic focusing on key control variables in the current state.

[0037] Next, spatiotemporal feature fusion is performed. The previously generated temporal context features are concatenated with the spatial attention-weighted features, and the original information flow is preserved through residual connections. Then, a fully connected fusion network is used for nonlinear mapping. This design preserves the complete temporal dependencies while strengthening the influence of instantaneous key physical quantities on the control strategy. The fused feature sequence is then processed through a normalization layer to suppress feature scale differences between different vehicle platforms and improve the model's cross-vehicle generalization ability.

[0038] Finally, a lightweight multilayer perceptron network is used to compress the fused features to a fixed dimension, and a random deactivation mechanism is introduced to prevent overfitting, ultimately outputting a 100×256 spatiotemporal embedding sequence. This sequence serves as a unified representation of the vehicle's current steering dynamic state, which is then used by the subsequent policy transfer network.

[0039] The STAM model is trained using supervised learning, with supervision labels derived from real-vehicle differential lock control logs and a stability evaluation system. Each training sample is associated with a real differential lock closure ratio curve and its corresponding vehicle stability score. The training objective is to minimize the prediction error between the model's output embedded features and the actual control response. The loss function is designed to consider two aspects: one constrains the model's sensitivity to the differential lock engagement moment, and the other constrains the predictive consistency of vehicle yaw stability.

[0040] During the supervised training of the STAM model, the input data is a standardized time-series tensor. Each sample consists of 32-dimensional features from 100 consecutive frames. These features include basic sensor parameters such as wheel speed, yaw rate, and steering wheel angle, as well as derived features such as vehicle speed change rate, yaw acceleration, wheel speed difference gradient, and differential lock closure rate. The output data includes two curves: one is the actual differential lock closure ratio curve for that time period, recorded by a high-precision calibrated vehicle under actual operating conditions; the other is the vehicle stability score, calculated by the automaker's internal stability evaluation system based on yaw rate, roll angle, slip ratio, etc. A higher score indicates greater stability. The distinction between positive and negative samples is as follows: when the vehicle engages the differential lock according to the optimal control strategy under specific conditions, such as low-adhesion cornering or hill start, and the stability score is higher than the threshold, the sample is marked as a positive sample; conversely, if the control intervention is too late, too early, or the closure ratio is inappropriate, resulting in significant sideslip, yaw, or inability to escape, and the stability score is lower than the threshold, it is marked as a negative sample. The intrinsic relationship between input and output, and the state change trends contained in the time-series tensor, such as sudden increases in yaw acceleration and rapid increases in wheel speed difference gradients, directly indicate the degree of risk of vehicle instability. The optimal intervention timing and closure ratio of the differential lock need to be predicted in advance based on these trends. At the same time, the vehicle stability score reflects the effectiveness of the control. The STAM model, through comparative learning of a large number of positive and negative samples, forces the embedded features to not only contain current state information but also encode the mapping law from the current state to the ideal control response, thereby enabling timely and accurate differential lock control commands to be output in the subsequent policy model.

[0041] An adaptive optimizer was used during training, and the initial learning rate was carefully set and dynamically adjusted using a cosine annealing strategy. To enhance the model's adaptability to extreme conditions, data augmentation mechanisms were introduced during the training phase, including adding noise perturbations to the wheel speed signal, scaling the yaw signal, and randomly stretching the steering rate, to simulate changes in data distribution under different driving styles and road surface adhesion conditions.

[0042] S103. Based on the lightweight differential lock strategy decision model and the dynamic feature vector, determine the differential lock control command. The lightweight differential lock strategy decision model is a lightweight neural network obtained by compressing a cross-vehicle differential lock strategy transfer learning model as the teacher network through a knowledge distillation mechanism.

[0043] Using a lightweight differential lock strategy decision model as the default control model, meaning that it generally relies only on the vehicle-side model, ensures the real-time performance and reliability of differential lock control under most driving conditions. In practical applications, the generation of differential lock control commands is solely handled by the vehicle-side lightweight strategy decision model. This lightweight strategy decision model is a compressed, lightweight neural network obtained through knowledge distillation, transferring learning from cross-vehicle differential lock strategies. Its parameter count is typically reduced to less than 1% of a large model, with a model size of less than 5MB and a single inference time controlled within 5ms. This design allows the vehicle-side controller to achieve millisecond-level real-time control without relying on a large model, thus truly overcoming the bottleneck of traditional large models being unable to be deployed at the edge due to their large parameter count and high inference latency.

[0044] In one possible implementation, when the vehicle meets preset triggering conditions, such as the lightweight model output confidence level being lower than a threshold, the vehicle entering a low-adhesion road surface or a high-risk operating condition, the system being in calibration mode, or the vehicle-side computing resources being idle, two strategy models can be simultaneously activated for collaborative decision-making, including the following steps A1~A3: Step A1: If the preset triggering conditions are met, input the dynamic feature vector and vehicle model information into the cross-vehicle differential lock strategy transfer learning model and output the first control suggestion.

[0045] The preset triggering conditions are based on at least one of the following: the output confidence of the lightweight differential lock strategy decision model, the current operating environment and driving status of the vehicle, driving style characteristics, and the vehicle-side computing resource usage. These conditions are used to determine whether to enable the cross-model differential lock strategy transfer learning model and trigger the dual-model fusion decision. This allows for the introduction of a large model to enhance and correct the strategy when the independent control of the lightweight model cannot meet the reliability or adaptability requirements.

[0046] For example, the preset triggering conditions include at least one of the following: the second confidence level output by the lightweight differential lock strategy decision model is lower than the first confidence level threshold (e.g., 0.5); the vehicle's current operating environment is a low-adhesion road surface (including ice and snow, wet asphalt, gravel, or mixed adhesion road surface), a slope turn, or an emergency obstacle avoidance turn; the vehicle is in a climbing turn state or the vehicle body roll angle exceeds a preset angle threshold (e.g., 5 degrees); the driving style recognition result is aggressive, and the deviation between the strategy output and the actual response in the historical control cycle exceeds a second deviation threshold; the vehicle-side computing resources are idle, or the system is in offline calibration and data acquisition mode; the vehicle enters a preset high-risk escape scenario, such as the wheel slip rate exceeding a threshold and ESC frequently intervening.

[0047] The cross-vehicle differential lock strategy transfer learning model is built on the Transformer architecture and employs a two-stage training approach: first, it is pre-trained on source vehicles with a large amount of existing calibration data to learn a general differential lock control strategy; then, it is fine-tuned and adapted on target vehicles using a small number of samples, enabling the model to be quickly transferred to different vehicle platforms. This achieves cross-platform reuse of control strategies and reduces manual calibration and testing costs. The model has tens of millions of parameters and is selectively activated only when computing resources are idle or the system is in calibration mode during vehicle deployment; it is not used as the default model for real-time control.

[0048] The Transformer architecture is a deep learning model architecture based on a self-attention mechanism. It abandons the traditional recurrent neural network and convolutional neural network structures and relies entirely on the attention mechanism to capture global dependencies in sequence data, becoming the foundational architecture in the fields of natural language processing and sequence modeling.

[0049] In the training and inference process of the cross-vehicle differential lock strategy transfer learning model, the input data includes two parts: first, the dynamic feature vector output by the STAM model, which is a 100×256 temporal embedding sequence representing the spatiotemporal perception information of vehicle steering stability and wheel-end power distribution within 0.5 seconds; second, the vehicle identifier, i.e., the vehicle token, which encodes key differentiating information such as the differential lock hardware structure, electric drive type, drive architecture, curb weight, torque ratio, and controller version of the current vehicle model. The output data is the first control suggestion, i.e., a three-dimensional vector, corresponding to the differential lock suggested opening threshold, suggested closing ratio, and first confidence level. Among them, the opening threshold and closing ratio are derived from the actual control parameters recorded by the source or target vehicle calibration engineers according to the principle of optimal stability, and the first confidence level is calculated based on the dynamic response indicators such as vehicle yaw rate and roll recovery time after this control.

[0050] The intrinsic relationship between input and output lies in the fact that the dynamic feature vector reflects the real-time instability risk and steering dynamic trend of the vehicle under different operating conditions, while the vehicle model identifier provides physical constraints on the differential lock response of different platforms. For example, heavy vehicles require a higher opening threshold, and electric drive axle vehicles are more sensitive to torque differences. Through a cross-head attention mechanism, TransferBERT deeply integrates the temporal variation patterns in the dynamic features with the platform difference information in the vehicle model token. It learns which opening threshold and closing ratio mapping function should be output for a specific vehicle model, given the current dynamic trend. This allows it to grasp the general rules of what control suggestions should be output under what conditions and for what vehicle model.

[0051] In one possible implementation, the cross-vehicle differential lock policy transfer learning model includes an input encoding layer, a position embedding layer, a vehicle identifier embedding layer, a multi-layer Transformer encoder, and a policy output mapping layer. The multi-layer Transformer encoder is typically referred to directly in English, while its Chinese name can be expressed as "multi-layer transformer encoder."

[0052] Here, the vehicle model identifier is an identifier vector representing the vehicle model platform. For example, the vehicle model identifier can be a vehicle model token, and the vehicle model identifier embedding layer can be a vehicle model token embedding layer. In the Transformer model, token is a standard term that can be translated as vehicle model token. The vehicle model token embedding layer encodes vehicle model identification information into a vehicle model token, inserts the vehicle model token at the beginning of the sequence of the dynamic feature vector, and obtains a model input sequence composed of the dynamic feature vector. The model input sequence is encoded by the input encoding layer, and position features are added by the position embedding layer before being input to the multi-layer Transformer encoder. The vehicle model token encodes at least one of the following: differential lock hardware structure, electric drive type, drive architecture, vehicle curb weight, front and rear axle torque ratio, and controller version information of the vehicle model to which the vehicle model identification information belongs. The multi-layer Transformer encoder adopts a cross-head attention enhancement structure, wherein the first part of the attention head is used to focus on the current dynamic sequence features of the vehicle, and the second part of the attention head is used to introduce cross-attention between vehicle model tokens to learn the policy transfer mapping relationship between different vehicle model platforms. The policy output mapping layer is used to extract the first token output by the multi-layer Transformer encoder and perform mapping processing to output the first control suggestion.

[0053] The model receives the vehicle dynamic embedding sequence output by the spatiotemporal attention feature extraction network from the previous stage as input. This sequence has a dimension of 100×256, representing a comprehensive expression of the vehicle's spatiotemporal perception results within 0.5 seconds. After a series of processing steps, the model outputs a three-dimensional control vector, which includes the differential lock proposal opening threshold, the proposal closing ratio, and the policy credibility score. The credibility score will be used in the subsequent fusion stage to determine the activation priority of this transfer result.

[0054] During the input processing phase, the cross-vehicle differential lock strategy transfer learning model directly feeds the spatiotemporal embedded sequence into the input channel while simultaneously attaching a vehicle model identification tag to each sample. The vehicle model identification tag is a vector of pre-defined dimensions, encoding key platform information such as the differential lock hardware structure, electric drive type, drive architecture, vehicle curb weight, front-to-rear axle torque distribution, and controller version of the current sample's vehicle model. Token vectors for all vehicle models are pre-constructed by the engineering team and automatically matched using the vehicle identification code during vehicle deployment. This token serves as the first input to the sequence and is used throughout the entire forward computation process of the network, ensuring the model maintains a structural understanding of the current vehicle platform.

[0055] The backbone of the cross-vehicle differential lock strategy transfer learning model consists of a multi-layered Transformer encoder, with each layer containing a multi-head attention mechanism, a feedforward network, and residual connections and normalization structures. Unlike the standard Transformer, this model employs a cross-head attention enhancement mechanism for attention computation. Some attention heads focus on modeling the temporal features of the vehicle's current dynamic sequence, while others forcibly introduce cross-attention between tokens from different vehicle models to learn the transfer mapping relationship between control strategies across different platforms. This design allows the model to simultaneously consider the current vehicle state and platform differences when determining differential lock trigger conditions, thus achieving true cross-platform strategy perception capability.

[0056] The policy output section uses a multilayer perceptron network to extract the hidden state corresponding to the first token in the encoder output sequence, mapping it to the final three-dimensional control vector, which corresponds to the differential lock proposal opening threshold, proposal closure ratio, and policy credibility score, respectively. For model training, a two-stage strategy of pre-training on the source vehicle and fine-tuning on the target vehicle is adopted. In the first stage, supervised pre-training is performed on the source vehicle with a large amount of accumulated calibration data. The training objectives include accurately regressing the actual opening and closing behavior of the differential lock and classifying vehicle attitude stability. The loss function is composed of a combination of regression error and distribution difference metric to constrain the consistency between model prediction and actual control behavior. In the second stage, fine-tuning and adaptation are performed on the target vehicle using only a very small number of labeled samples. At this stage, most of the encoder's front layers are frozen, and only the last few layers are unfrozen for parameter updates. A regularization mechanism is introduced to prevent overfitting under small sample conditions, thus ensuring rapid adaptation to new vehicle models while maintaining generalization ability.

[0057] Step A2: Input the dynamic feature vector into the lightweight differential lock strategy decision model and output the second control suggestion.

[0058] The second control recommendation includes a differential lock opening threshold, a closing ratio, and a second confidence level. The lightweight differential lock strategy decision model is a lightweight neural network obtained by compressing the cross-model differential lock strategy transfer learning model as the teacher network through a knowledge distillation mechanism.

[0059] In one possible implementation, the lightweight differential lock strategy decision model consists of a temporal modeling unit and a multilayer perceptron. The temporal modeling unit is used to extract temporal features from the input dynamic feature vector to capture the temporal correlation of vehicle state features. The multilayer perceptron is used to perform nonlinear mapping on the extracted temporal features and finally output the second control suggestion. The lightweight differential lock strategy decision model achieves lightweight compression through a knowledge distillation mechanism. This knowledge distillation mechanism adopts a soft-label distillation method, using the policy triples output by the cross-vehicle differential lock strategy transfer learning model (teacher network) as the distillation training target. The distillation loss function adopts mean squared error, and policy consistency is introduced as a constraint term to ensure that the output direction of the model is consistent with the teacher network under key boundary conditions, thereby ensuring the reliability of the control strategy.

[0060] The lightweight differential lock strategy decision model receives the embedded feature sequence output from the previous stage's spatiotemporal attention feature extraction network as input. This sequence has a dimension of 100×256, representing a comprehensive expression of the vehicle's spatiotemporal perception results over a continuous 0.5 seconds. The lightweight differential lock strategy decision model outputs a three-dimensional control vector, consistent with the output structure of the teacher model, corresponding to the differential lock suggestion opening threshold, suggestion closure ratio, and strategy confidence score, respectively.

[0061] In terms of model architecture design, a compact temporal neural network is adopted. First, a single-layer unidirectional gated recurrent unit is used to extract temporal dynamic features from the input sequence, capturing the evolution of the vehicle state over time. Then, two layers of multilayer perceptrons are connected to gradually map the temporal features into the final three-dimensional control output. This design compresses the overall number of parameters to the hundreds of thousands level while maintaining high prediction accuracy, and keeps the model size within a few megabytes, which meets the hardware limitations of vehicle-side microcontroller-level deployment.

[0062] The distillation training phase employs a soft-label distillation method. After the teacher model completes pre-training, all its weights are frozen, and its output is used as the training target for the student model. The soft label for each input sample is the policy vector output by the teacher model, rather than the original true control label. The loss function in the distillation process is designed as mean squared error to minimize the difference between the outputs of the student and teacher models. Simultaneously, a policy consistency constraint is introduced, mandating that the student model's output direction must remain consistent with the teacher model in critical boundary conditions when determining whether the differential lock should intervene, preventing misjudgments in critical states.

[0063] To improve the robustness of the distilled model under extreme conditions, the training data not only includes the original training dataset of the teacher model but also incorporates sampled data of the target vehicle under extreme conditions such as slippery turns, hill turns, and emergency obstacle avoidance turns, ensuring the stability of the student model's performance under boundary conditions. All data is processed through a spatiotemporal attention feature extraction network to achieve a unified format before being injected into the student model for reverse training. An adaptive optimizer with a reasonable learning rate is used during training, and an early stopping mechanism is introduced to prevent overfitting.

[0064] In one possible implementation, after the lightweight differential lock strategy decision model completes distillation training, it needs to undergo edge deployment optimization to adapt to the real-time control requirements of the vehicle. Edge deployment optimization includes at least two methods: quantization and structural pruning, which can be implemented individually or in combination. Quantization compresses the model's original floating-point parameters into integer form using a specific algorithm, effectively reducing the model size while ensuring no significant loss in control accuracy. Structural pruning filters the temporal feature channels in the model, removing channels whose contribution to strategy judgment is below a preset threshold, retaining only input signal connection channels with high weights for control decisions, such as yaw rate, wheel speed difference, steering torque, and roll angle. After joint optimization of quantization and structural pruning, the model size can be reduced to within a few megabytes, and the single inference time is controlled at the millisecond level, fully meeting the deployment requirements of real-time differential lock control at the vehicle and ensuring rapid response and stable execution of control commands.

[0065] Step A3: Combining the first confidence level, the second confidence level, and the current operating status of the vehicle, perform a fusion judgment on the first control suggestion and the second control suggestion to determine the differential lock control command.

[0066] In the fusion judgment stage, the first and second control suggestions are fused and judged based on the first confidence level, the second confidence level, and the current operating status of the vehicle to determine the differential lock control command. First, based on the vehicle's historical operating data, the deviation between the strategy output and the actual vehicle response of the cross-model differential lock strategy transfer learning model and the lightweight differential lock strategy decision model are statistically analyzed. The action consistency rate and control response lag index are calculated and fused into a time-weighted score. The model optimization score of the two models is determined accordingly. The higher the score, the stronger the reliability of the strategy output of the corresponding model in the current round.

[0067] Subsequently, based on vehicle operating environment labels, operating speed ranges, and historical strategy trigger point distributions, a dynamic confidence adjustment factor is constructed. The first and second confidence scores are then normalized to eliminate the incomparability of confidence scores caused by differences in training data distribution and model structure between different models, resulting in the re-scored first and second confidence scores. Next, the deviation values ​​between the first and second control suggestions in core control indicators such as differential lock opening threshold and closure ratio are detected. If the deviation exceeds a preset threshold, it is determined that there is a key strategy contradiction between the two sets of control suggestions. In this case, the control suggestion with the higher value between the re-scored first and second confidence scores is selected, and the differential lock closure ratio of that suggestion is reduced, thus generating a preliminary fusion result. If the deviation does not exceed the preset threshold, it is determined that the two sets of control suggestions are highly consistent, and the core control parameters of the first and second control suggestions are directly averaged to obtain a preliminary fusion result.

[0068] Finally, combining the model optimization scores, confidence scores after re-scoring, vehicle current operating status, and operating environment labels of the two models, the preliminary fusion result is finalized according to a preset priority logic: if the cross-model differential lock strategy transfer learning model is enabled and its optimization score is higher than that of the lightweight differential lock strategy decision model, then the first control suggestion replaces the preliminary fusion result as the differential lock control command; if the lightweight differential lock strategy decision model has a better response speed in low-adhesion road scenarios than the cross-model differential lock strategy transfer learning model, then the second control suggestion replaces the preliminary fusion result as the control command. Differential lock control command; if the first confidence level and the second confidence level after re-scoring are both lower than the preset threshold, the differential lock closure ratio in the preliminary fusion result is restricted and adjusted to a preset safe range before being used as the differential lock control command; if the vehicle is currently in a high-risk driving state such as climbing, turning, or high roll angle, a safety compensation value is introduced based on the preliminary fusion result to generate the differential lock control command; if the vehicle is currently in a high-speed driving state or a stable strategy range, the first control suggestion and the second control suggestion are weighted and fused through a fuzzy logic controller to obtain the final differential lock control command.

[0069] In one possible implementation, the entire fusion decision mechanism operates on a fixed cycle, supporting dynamic strategy evaluation and real-time control refresh. For example, the entire fusion decision process operates on a fixed cycle of 100ms. Its overall input includes the dynamic perception embedding vector output by the spatiotemporal attention feature extraction model, the strategy suggestion triplet output by the cross-vehicle differential lock strategy transfer learning model, and the same structure triplet output by the lightweight differential lock strategy decision model. It also incorporates the vehicle's current operating state and the road type operating environment label generated by the spatiotemporal attention feature extraction model as reference parameters. During the fusion process, the reliability and resource utilization adaptability of the two strategy outputs are evaluated in real time. The lightweight differential lock strategy decision model is effective in real time under all operating modes, while the cross-vehicle differential lock strategy transfer learning model is only activated when vehicle-side resources are idle or the vehicle is in calibration mode. The final generated differential lock control command includes the control command, target closure ratio, strategy source label, and confidence score of the current strategy. After internal protocol conversion, it is directly sent to the differential lock controller, which drives the electromechanical device to perform the differential lock closure action. The execution status returned by the controller in real time will be used for the closed-loop verification of the next round of fusion judgment. The entire fusion mechanism has hot switching capability, which can ensure that the core control function is not interrupted.

[0070] S104. Send the differential lock control command to the differential lock controller.

[0071] In one possible implementation, after the differential lock control command is sent to the differential lock controller, a preset dynamic control mapping matrix is ​​queried based on the target closure ratio and control confidence in the differential lock control command to determine the coordinated control parameters corresponding to the electronic power steering system, traction control system, vehicle dynamic control system, electric drive system and active suspension system, and the coordinated control parameters are synchronously sent to the corresponding vehicle control system.

[0072] The dynamic control mapping matrix integrates five types of collaborative control mapping relationships, including: the adjustment mapping relationship between the differential lock closure ratio and yaw rate on the electronic power steering system's assist coefficient; the modulation mapping relationship between the differential lock closure ratio and the front-to-rear axle torque distribution, left-to-right drive torque difference, and traction control system slip rate intervention intensity; the adjustment mapping relationship between the differential lock closure ratio change rate and vehicle roll angle on the active suspension system's damping response rate and response frequency; the correction mapping relationship between the differential lock closure ratio and the target yaw rate curve in the vehicle dynamic control system; and the gating mapping relationship between control confidence and the differential lock strategy execution priority, execution amplitude, and vehicle control system intervention authority.

[0073] Specifically, the mapping relationship for the electronic power steering system can include: in low-adhesion road surface curves or slope escaping scenarios, when the differential lock engagement ratio is greater than 40% and the vehicle yaw rate exceeds a preset threshold, the power assist coefficient of the electronic power steering system is reduced to improve steering wheel road feel; while in emergency escaping situations where the differential lock engagement ratio is close to 100%, the power assist coefficient of the electronic power steering system is increased to compensate for the increased steering rigidity of the vehicle and the resulting handling fatigue.

[0074] The modulation mapping relationship between the front and rear axle torque commands and the traction control system can include: dynamically adjusting the difference between the target torque commands of the left and right electric drives of the rear axle according to the differential lock closure ratio; controlling the difference between the target torque commands of the left and right electric drives of the rear axle to not exceed 10% when the closure ratio is 50%; and making the left and right torques tend to be consistent to simulate a rigid connection state when the closure ratio is 100%. At the same time, during the differential lock closure process, dynamically reducing the intervention intensity of the traction control system on the slip ratio to avoid strategy conflicts between different control systems.

[0075] The mapping relationship of the active suspension system is specifically manifested as follows: based on the rate of change of the differential lock closure ratio and the current vehicle roll angle, a mapping table is established to determine the response mode of the active suspension system. During the process of increasing the differential lock closure ratio, the active suspension damping response is switched to a fast response mode to increase the response frequency, reduce damping delay, and prevent vehicle attitude instability caused by control delay.

[0076] The modification mapping relationship for the vehicle dynamic control system can include: in the high lock-up state where the differential lock closure ratio is greater than 60%, reducing the overall amplitude of the target yaw rate adjustment curve in the vehicle dynamic control system, reducing the false trigger threshold of the controller for excessive directional response, so as to tolerate a certain degree of sideslip freedom and improve the vehicle's driving stability and ability to get out of trouble in complex road conditions.

[0077] Meanwhile, a gating mechanism based on control confidence is introduced. When the control confidence score is below 0.3, the differential lock output action enters a wait-and-see state, and the controller only executes 10% of the closing ratio to provide basic protection. When the control confidence score is greater than 0.8, the differential lock strategy is determined to have high reliability. The active control system forcibly executes the control command and does not accept intervention from the electronic stability control system and traction control system to ensure vehicle handling performance in extreme escape scenarios.

[0078] The five types of mapping relationships mentioned above are integrated into a unified dynamic control mapping matrix. Its internal structure is defined as a combination of input action dimension, control target dimension, weight coefficients, and state thresholds. All mapping tables support dynamic updates and calibration optimization. This matrix is ​​deployed within the strategy linkage module, and its parameters are refreshed by the vehicle's main control platform at a fixed period of 50 milliseconds. Based on the road condition information collected by the current sensors, such as dry asphalt or slippery road surfaces, the corresponding matrix parameter group is automatically switched, thereby realizing the coordinated linkage and dynamic adaptation between differential lock control and various chassis control systems of the vehicle.

[0079] In one possible implementation, the actual vehicle response data after the differential lock control command is executed, as well as driving style features, are obtained. The driving style features are used to correct the strategy confidence and tune the strategy output. The actual vehicle response data and driving style features are uploaded to the cloud so that the cloud can perform cluster analysis and deviation comparison. The cross-vehicle differential lock strategy transfer learning model, the lightweight differential lock strategy decision model, and the spatiotemporal attention feature extraction model are dynamically updated based on the cloud analysis results.

[0080] In one possible implementation, the actual vehicle response data and driving style characteristics after the differential lock control command is executed are obtained, whereby the driving style characteristics are used to correct the policy confidence and tune the policy output. A vehicle-mounted behavioral deviation monitoring module is deployed to capture the difference between the differential lock strategy output and the actual vehicle dynamics. Specific monitored variables include the error between the strategy suggestion closure ratio and the executed closure value, the change in vehicle yaw rate within 0.5 seconds after differential lock intervention, and the decreasing trend of the vehicle's sideslip rate at the strategy activation point. After each differential lock action, the system records the vehicle's temporal state before and after the action, forming a tagged behavioral response log. Simultaneously, a driving style perception module is deployed on the vehicle. Based on behavioral characteristics such as steering rate, throttle gradient, braking depth, and frequency of electronic stability control system intervention, driver style tags are extracted and bound to user IDs. These tags can be used for strategy confidence correction and strategy output parameter tuning. For example, aggressive drivers can tolerate a higher wheel slip rate in corners, so the system will increase the upper limit of the strategy output closure ratio, while conservative drivers tend to intervene in stability control earlier, so the system will generate differential lock activation suggestions in advance and increase the confidence threshold.

[0081] Subsequently, the actual vehicle response data, driving style characteristics, and behavioral response logs were uploaded to the cloud for cluster analysis and deviation comparison. A model feedback correction platform was built in the cloud, which divides the sample set by vehicle type, road condition, and driving style. It compares the deviation between the differential lock strategy's predicted behavior and the vehicle's actual response, and then dynamically adjusts the control boundaries of the cross-vehicle differential lock strategy transfer learning model and the lightweight differential lock strategy decision model. At the same time, it updates the spatiotemporal attention feature extraction model based on the deviation.

[0082] For example, the optimization and update methods may include at least one of the following: First, construct a policy correction table based on sample frequency and deploy it to the vehicle to perform lightweight compensation; second, periodically fine-tune the weights of the last two layers of the cross-vehicle differential lock policy transfer learning model using real feedback data and push OTA updates; third, for policy operating points where deviations are concentrated, retrain the embedded feature module of the spatiotemporal attention feature extraction model to improve the feature recognition capability in this scenario. Through the above methods, the dynamic updating and optimization of the three models can be achieved, ensuring the adaptability and reliability of the differential lock control strategy.

[0083] See Figure 2 , Figure 2 The flowchart illustrates another differential lock control method based on multi-model fusion provided in this application embodiment. As shown in the figure, the process first acquires raw vehicle sensor data through data acquisition and preprocessing, and then performs standardization processing. Subsequently, it extracts dynamic feature vectors that fuse spatiotemporal information from the temporal tensor using the Spatiotemporal Attention Feature Extraction (STAM) method. These feature vectors are then input in parallel to the policy transfer modeling (TransferBERT) and the policy distillation model (KD-Net), generating two sets of control proposals respectively. The two sets of proposals enter the policy fusion scheduling mechanism for optimization and consistency verification, and then output differential lock control commands. The commands are converted into multi-domain collaborative control parameters through a control mapping matrix and distributed to the actuators through a real-time control interface. The differential lock control execution feedback transmits the actual response data back for online model tuning and closed-loop optimization, forming a complete technical closed loop from data acquisition to execution feedback and continuous model iteration.

[0084] The above are some specific implementations of the differential lock control method based on multi-model fusion provided in the embodiments of this application. Based on this, this application also provides a corresponding differential lock control system based on multi-model fusion. The system provided in the embodiments of this application will be described below from the perspective of functional modularity. Figure 3 This is a structural diagram of a differential lock control system based on multi-model fusion provided in an embodiment of this application.

[0085] The system includes: The acquisition unit 110 is used to acquire multi-source sensor data of the vehicle and generate a standardized time series tensor based on the multi-source sensor data. The standardized time series tensor is a time series sequence composed of multi-dimensional state features of the vehicle within a fixed time window. Extraction unit 111 is used to input the time series data into the spatiotemporal attention feature extraction model to extract dynamic feature vectors characterizing vehicle steering stability. The spatiotemporal attention feature extraction model is used to perform forward and backward temporal modeling and dynamic weighting of sensor features on the standardized time series tensor to generate dynamic feature vectors that fuse spatiotemporal information. The determining unit 112 is used to determine the differential lock control command based on the lightweight differential lock strategy decision model and the dynamic feature vector. The lightweight differential lock strategy decision model is a lightweight neural network obtained by compressing a cross-vehicle differential lock strategy transfer learning model as the teacher network through a knowledge distillation mechanism. The sending unit 113 is used to send the differential lock control command to the differential lock controller.

[0086] This application also provides corresponding devices and computer storage media for implementing the differential lock control scheme based on multi-model fusion provided in this application.

[0087] The device includes a memory and a processor. The memory is used to store instructions or code, and the processor is used to execute the instructions or code to enable the device to perform the differential lock control method based on multi-model fusion as described in any embodiment of this application.

[0088] The computer storage medium stores code. When the code is executed, the device running the code implements the differential lock control method based on multi-model fusion as described in any embodiment of this application.

[0089] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

[0090] It should be understood that in this application, "at least one" refers to one or more items, and "more" refers to two or more items. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one" or similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, "at least one" of a, b, or c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0091] It should be understood that the terms center, longitudinal, transverse, up, down, front, back, left, right, vertical, horizontal, top, bottom, inside, outside, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0092] It should be noted that, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0093] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the statement "including a…" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0094] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0095] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A differential lock control method based on multi-model fusion, characterized in that, include: Acquire multi-source sensor data of the vehicle, and generate a standardized time series tensor based on the multi-source sensor data. The standardized time series tensor is a time series sequence composed of multi-dimensional state features of the vehicle within a fixed time window. The time series data is input into the spatiotemporal attention feature extraction model, which outputs a dynamic feature vector characterizing the vehicle's steering stability. The spatiotemporal attention feature extraction model is used to perform forward and backward temporal modeling and dynamic weighting of sensor features on the standardized time series tensor to generate a dynamic feature vector that integrates spatiotemporal information. Based on the lightweight differential lock strategy decision model and the dynamic feature vector, the differential lock control command is determined. The lightweight differential lock strategy decision model is a lightweight neural network obtained by compressing a cross-vehicle differential lock strategy transfer learning model as the teacher network through a knowledge distillation mechanism. The differential lock control command is sent to the differential lock controller.

2. The method according to claim 1, characterized in that, The determination of differential lock control commands based on the lightweight differential lock strategy decision model and the dynamic feature vector includes: If the preset triggering conditions are met, the dynamic feature vector and vehicle model information are input into the cross-vehicle differential lock strategy transfer learning model, and a first control suggestion is output. The first control suggestion includes a differential lock opening threshold, a closing ratio, and a first confidence level. The cross-vehicle differential lock strategy transfer learning model is built based on the Transformer architecture and achieves the transfer of differential lock control strategies between different vehicle models through a two-stage training method of pre-training in the source vehicle model and fine-tuning in the target vehicle model. The dynamic feature vector is input into the lightweight differential lock strategy decision model, and a second control suggestion is output. The second control suggestion includes a differential lock opening threshold, a closing ratio, and a second confidence level. By combining the first confidence level, the second confidence level, and the current operating status of the vehicle, the first control suggestion and the second control suggestion are fused and judged to determine the differential lock control command.

3. The method according to claim 1, characterized in that, The processing method of the spatiotemporal attention feature extraction model includes: Temporal modeling is performed on the standardized temporal tensor from two time directions, and temporal context features are output. The sensor feature vector at each time step is input into a fully connected attention network. After activation and normalization, the importance weight vector of each sensor variable at the current time is generated. The weight vector is then multiplied element-wise with the original features to output a weighted feature representation. The temporal context features and the weighted feature representations are concatenated, and after retaining the original information flow through residual connections, they are nonlinearly mapped through a fully connected fusion network and processed by a normalization layer to output a dynamic feature vector that fuses spatiotemporal information. The dynamic feature vector of the fused spatiotemporal information is compressed by a multilayer perceptron, and the spatiotemporal embedding sequence is output as the dynamic feature vector.

4. The method according to claim 2, characterized in that, The cross-vehicle differential lock strategy transfer learning model includes an input encoding layer, a position embedding layer, a vehicle identifier embedding layer, a multi-layer Transformer encoder, and a strategy output mapping layer. The vehicle model identifier embedding layer is used to encode vehicle model identification information into a vehicle model identifier, and insert the vehicle model identifier into the beginning of the sequence of dynamic feature vectors to obtain the model input sequence composed of the dynamic feature vectors; the model input sequence is encoded by the input encoding layer and position features are added by the position embedding layer before being input to the multi-layer Transformer encoder; The vehicle identification code encodes at least one of the following: differential lock hardware structure, electric drive type, drive architecture, vehicle curb weight, front and rear axle torque ratio, and controller version information of the vehicle to which the vehicle identification information belongs. The multi-layer Transformer encoder adopts a cross-head attention enhancement structure. The first part of the attention head is used to focus on the current dynamic sequence features of the vehicle, and the second part of the attention head is used to introduce cross-attention between vehicle model identifiers to learn the policy transfer mapping relationship between different vehicle model platforms. The strategy output mapping layer is used to extract the first vehicle model identifier output by the multi-layer Transformer encoder and perform mapping processing to output the first control suggestion.

5. The method according to claim 1, characterized in that, The lightweight differential lock strategy decision model includes a temporal modeling unit and a multilayer perceptron. The temporal modeling unit is used to extract temporal features from the dynamic feature vector, and the multilayer perceptron is used to map the temporal features to the second control proposal. The knowledge distillation mechanism adopts a soft-label distillation method, using the policy triples output by the cross-vehicle differential lock strategy transfer learning model as the training target. The distillation loss function is the mean squared error, and the policy consistency is used as a constraint term to ensure that the output direction is consistent with the teacher network under key boundary conditions.

6. The method according to claim 5, characterized in that, The method further includes: If the lightweight differential lock strategy decision model completes distillation training, edge deployment optimization is performed on the lightweight differential lock strategy decision model. The edge deployment optimization includes at least one of quantization processing and structural pruning. The quantization processing is used to compress the model parameters from floating-point form to integer form. The structural pruning is used to prune temporal feature channels in the model whose contribution is lower than a preset threshold.

7. The method according to claim 1, characterized in that, The step of combining the first confidence level, the second confidence level, and the current operating state of the vehicle to perform a fusion judgment on the first control suggestion and the second control suggestion to determine the differential lock control command includes: Based on historical vehicle operation data, the model optimization scores of the cross-model differential lock strategy transfer learning model and the lightweight differential lock strategy decision model are calculated respectively. Based on the vehicle operating environment labels, operating speed ranges, and historical policy trigger point distributions, the first confidence level and the second confidence level are normalized respectively to obtain the first confidence level and the second confidence level after rescoring. The deviation between the first control suggestion and the second control suggestion in the control index is detected. If the deviation exceeds a preset threshold, the control suggestion with the higher confidence level between the first confidence level and the second confidence level after re-scoring is selected, and the differential lock suggestion closure ratio is reduced to obtain a preliminary fusion result. If the deviation does not exceed the preset threshold, the core control parameters of the first control suggestion and the second control suggestion are averaged to obtain a preliminary fusion result. Based on the optimal scores of the cross-model differential lock strategy transfer learning model and the lightweight differential lock strategy decision model, the confidence level after rescoring, the current operating status of the vehicle, and the operating environment labels, the differential lock control command is determined according to a preset priority.

8. The method according to claim 7, characterized in that, The step of determining differential lock control commands according to preset priorities based on the optimal scores of the cross-vehicle differential lock strategy transfer learning model and the lightweight differential lock strategy decision model, the confidence level after rescoring, the vehicle's current operating status, and operating environment labels includes: If the cross-model differential lock strategy transfer learning model is enabled and its preferred score is higher than that of the lightweight differential lock strategy decision model, then the first control suggestion is used instead of the preliminary fusion result as the differential lock control command. If the response speed of the lightweight differential lock strategy decision model in a low-adhesion scenario is better than that of the cross-vehicle differential lock strategy transfer learning model, then the second control suggestion shall replace the preliminary fusion result as the differential lock control command. If both the first confidence level and the second confidence level after re-scoring are lower than the preset threshold, the differential lock closure ratio in the preliminary fusion result is restricted and adjusted to a preset safe range before being used as a differential lock control command. If the vehicle is currently operating in a high-risk state, a safety compensation value is introduced based on the preliminary fusion result to generate a differential lock control command; If the vehicle is currently in a high-speed driving state or in a stable policy range, the first control suggestion and the second control suggestion are weighted and fused by the fuzzy logic controller to obtain the differential lock control command.

9. The method according to claim 1, characterized in that, After sending the differential lock control command to the differential lock controller, the method further includes: Based on the target closure ratio and control confidence in the differential lock control command, a preset dynamic control mapping matrix is ​​queried to determine the coordinated control parameters corresponding to the electronic power steering system, traction control system, vehicle dynamic control system, electric drive system, and active suspension system; the coordinated control parameters are then synchronously sent to the corresponding vehicle control system. The dynamic control mapping matrix includes: the adjustment mapping relationship between the differential lock closure ratio and yaw rate on the electronic power steering system's assist coefficient; the modulation mapping relationship between the differential lock closure ratio and the front-to-rear axle torque distribution, left-to-right drive torque difference, and the traction control system's slip rate intervention intensity; the adjustment mapping relationship between the differential lock closure ratio change rate and vehicle roll angle on the active suspension system's damping response rate and response frequency; the correction mapping relationship between the differential lock closure ratio and the target yaw rate curve in the vehicle dynamic control system; and the gating mapping relationship between control confidence and the differential lock strategy execution priority, execution amplitude, and the vehicle control system's intervention authority.

10. The method according to claim 1, characterized in that, The method further includes: The system acquires the actual vehicle response data after the differential lock control command is executed, as well as driving style characteristics. These driving style characteristics are used to correct the strategy confidence and adjust the strategy output parameters. The actual vehicle response data and driving style characteristics are uploaded to the cloud so that the cloud can perform cluster analysis and deviation comparison. Based on the cloud analysis results, at least one of the following is dynamically updated: the cross-model differential lock strategy transfer learning model, the lightweight differential lock strategy decision model, and the spatiotemporal attention feature extraction model.

Images