Electric commercial vehicle lateral battery replacement and whole vehicle interface integrated system

Abstract

The present application relates to the electric vehicle manufacturing technology field, specifically relates to a kind of electric commercial vehicle lateral battery swapping and vehicle interface integrated system.System includes the integrated battery swapping interface of lateral integration of vehicle frame, binocular vision acquisition unit, multi-body dynamics deformation pre-compensation algorithm module and vehicle real-time ethernet redundancy control module.The interface high-voltage power, low-voltage communication and thermal management medium docking terminal adopt coaxial nested layout.Algorithm module uses space-time graph convolution network to calculate the spatial pose offset corresponding to vehicle frame deformation, binocular vision acquisition unit extracts feature points through multi-scale fusion sub-pixel edge detection, and completes docking coordinate correction combined with adaptive filtering method.Redundancy control module completes loop on-off time sequence closed-loop control through deep reinforcement learning time sequence decision and dynamic resource scheduling.The present application solves the interface docking interference problem caused by vehicle frame deformation under heavy load conditions, eliminates single visual positioning blind area, and avoids the safety hazards caused by discrete loop time sequence deviation.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle manufacturing technology, specifically to a side-swap battery and vehicle interface integration system for electric commercial vehicles. Background Technology

[0002] Existing side-mounted battery swapping systems for electric commercial vehicles typically have a swapping interface located on the side of the vehicle frame. They employ binocular vision sensors to capture the relative position of the swapping mechanism and the vehicle interface, and use image processing to extract feature points to guide the swapping mechanism to complete the docking. At the electrical connection level, current technology designs the high-voltage power circuit, low-voltage communication circuit, and battery thermal management medium circuit as independent, separate structures, each located at different physical positions on the side of the vehicle frame. At the control level, the on / off control of each circuit is independent, with each circuit's controller performing high-voltage power-on, communication handshake, and medium conduction actions based on preset thresholds.

[0003] The aforementioned existing technical solutions have a core flaw: under heavy-load conditions in commercial vehicles, the frame is subjected to lateral deflection and suspension compression deformation, causing the vehicle interfaces to deviate from their initial calibration positions. The existing single-vision positioning method can only obtain the feature point coordinates of the current frame image and cannot perceive the hidden displacement caused by the force inside the frame. This results in physical interference and jamming of the terminals when the battery swapping mechanism performs docking actions according to the uncorrected coordinates. At the same time, the split circuit layout will cause the docking process of each circuit to be asynchronous under the deformation of the frame. In addition, the on / off control of each circuit is independent, and it is impossible to dynamically adjust the timing of high-voltage power-on and medium conduction according to the actual docking status, which may lead to electrical arcing or medium leakage safety accidents. Summary of the Invention

[0004] The purpose of this invention is to provide a solution that can effectively address the problems described in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A side-mounted battery swapping and vehicle interface integration system for electric commercial vehicles includes an integrated battery swapping interface located on the side of the vehicle frame, a binocular vision acquisition unit, a multibody dynamics deformation pre-compensation algorithm module, and an on-board real-time Ethernet redundancy control module. The integrated battery swapping interface includes docking terminals for a high-voltage power circuit, a low-voltage communication circuit, and a battery thermal management medium circuit, and the docking terminals adopt a coaxial nested layout. The binocular vision acquisition unit is deployed on the side of the vehicle frame and is used to extract feature points of the battery swapping mechanism and the vehicle interface; The multibody dynamics deformation pre-compensation algorithm module is used to collect load, suspension travel and frame lateral deflection data, calculate the spatial pose offset of the battery swapping interface and complete the docking coordinate correction in combination with the feature points; The vehicle-mounted real-time Ethernet redundancy control module is connected to the battery swapping mechanism, high-voltage interlock circuit, and battery management system. It is used to complete the closed-loop control of the circuit on / off timing and the system of automatic configuration throughout the entire process during the interface docking process.

[0006] Preferably, when calculating the spatial pose offset, the multibody dynamics deformation pre-compensation algorithm module uses a spatiotemporal graph convolutional network to construct the frame deformation-interface displacement mapping model, takes the load, suspension travel, and frame lateral deflection data as graph node features, and uses an adaptive graph attention mechanism to aggregate the deformation propagation features of adjacent nodes to output the spatial pose offset.

[0007] Preferably, when extracting the feature points, the binocular vision acquisition unit uses a sub-pixel edge detection algorithm based on multi-scale fusion. It extracts multi-level edge responses by constructing an image pyramid and uses a method combining non-maximum suppression and Zernike moments to fit sub-pixel level edge positions in the edge normal direction to generate the feature points.

[0008] Preferably, in the process of completing the secondary correction of the docking coordinates, an adaptive unscented Kalman filter method based on federated learning is adopted, using the spatial pose offset as the prior state vector and the feature points as the observation vector. The measurement noise covariance is estimated in real time by introducing variational Bayesian inference, and the corrected docking coordinates are output.

[0009] Preferably, in the closed-loop control process of completing the loop on / off timing, a timing decision algorithm based on deep reinforcement learning is adopted to construct a Markov decision process with the loop on / off state and high-voltage interlock state as the state space and the on / off action as the action space, and the optimal on / off timing command is output through a double-delay deep deterministic policy gradient algorithm.

[0010] Preferably, during the fully automated configuration process, a dynamic topology network resource scheduling method is adopted to map high-voltage power-on, communication handshake, and thermal management loop activation into concurrent task flows with different priorities. A combination of a greedy algorithm and a shortest remaining time priority strategy is used to dynamically allocate network bandwidth and computing resources in the vehicle-mounted real-time Ethernet redundancy control module.

[0011] Preferably, the spatiotemporal graph convolutional network employs a meta-learning mechanism during training. By performing second-order gradient optimization on samples under different heavy-load conditions, the initial parameters of the network are updated, enabling the frame deformation-interface displacement mapping model to quickly and adaptively calculate the spatial pose offset with only a few iterations when encountering unknown load distribution changes.

[0012] Preferably, the multi-scale fusion sub-pixel edge detection algorithm introduces an ambient lighting compensation method based on generative adversarial networks during the image pyramid construction process. This method transforms the current lighting image to the standard lighting domain and embeds a spatial feature transformation layer in the generator network. The algorithm then uses the lighting conditional coding vector to adaptively normalize the multi-level edge responses, thereby enhancing the robustness of the feature points under low lighting conditions.

[0013] Preferably, after outputting the corrected docking coordinates, the adaptive unscented Kalman filtering method performs a residual prediction step based on a long short-term memory network. The correction error sequence of historical time moments is input into the long short-term memory network, the temporal error features are extracted through an attention mechanism, the residual compensation value at the current time moment is output, and the residual compensation value is superimposed and fused with the corrected docking coordinates.

[0014] Preferably, in the process of outputting the optimal on / off timing command, the dual-delay deep deterministic strategy gradient algorithm introduces a state representation extraction branch based on contrastive learning into the evaluator network. By forming positive and negative sample pairs between the current circuit on / off state and historical abnormal states, the distribution distance between safe and dangerous states in the latent space is increased, thereby improving the safety of the optimal on / off timing command under extreme electrical conditions.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention solves the docking interference and safety hazards caused by frame deformation under heavy load conditions by setting up a collaborative architecture of a multibody dynamics deformation pre-compensation algorithm module and a binocular vision acquisition unit. The multibody dynamics deformation pre-compensation algorithm module uses a spatiotemporal graph convolutional network and an adaptive graph attention mechanism to aggregate load, suspension travel, and frame lateral deflection data, calculates spatial pose offset, and transforms the internal force deformation of the frame into a priori state vector. Combined with image feature points extracted by a sub-pixel edge detection algorithm based on multi-scale fusion, a secondary correction is performed using an adaptive unscented Kalman filter method based on federated learning, eliminating blind spots that cannot be perceived by single visual positioning, and ensuring that the docking coordinates conform to the actual interface position after deformation. The on-board real-time Ethernet redundancy control module adopts a timing decision method based on a dual-delay deep deterministic strategy gradient algorithm, generates optimal on / off timing commands based on the real-time docking status, and combines a dynamic topology network resource scheduling method to ensure command issuance, eliminating the timing deviation risk caused by discrete control and avoiding terminal interference and electrical arcing.

[0016] 2. Building upon the aforementioned technical effects, this invention enhances the system's operational stability under complex operating conditions by introducing multiple data processing methods. The spatiotemporal graph convolutional network introduces a meta-learning mechanism, enabling the mapping model to update parameters through a small number of iterations when encountering unknown load distributions, maintaining the continuity of offset calculation. The binocular vision acquisition unit incorporates an ambient lighting compensation method based on generative adversarial networks and a spatial feature transformation layer during sub-pixel edge detection, eliminating interference from nighttime or backlighting environments on feature point extraction. The adaptive unscented Kalman filter method, after outputting corrected coordinates, utilizes a long short-term memory network to extract temporal features of historical correction errors and output residual compensation values, reducing the system's cumulative error. The dual-delay deep deterministic strategy gradient algorithm adds a state representation extraction branch based on contrastive learning to the evaluator network, improving the safety of on / off timing commands under extreme electrical conditions by distinguishing the distribution distances of safe and dangerous states in the latent space. Attached Figure Description

[0017] Figure 1 This is the overall control flowchart of the present invention; Figure 2 This is a flowchart illustrating the calculation of spatial pose offset using the spatiotemporal graph convolutional network of this invention. Figure 3 This is a flowchart of the multi-scale fusion sub-pixel edge detection algorithm of the present invention; Figure 4 This is a flowchart of the adaptive unscented Kalman filter docking coordinate correction based on federated learning in this invention; Figure 5 This is a flowchart of the on / off timing decision algorithm based on deep reinforcement learning in this invention; Figure 6 This is a flowchart illustrating the automatic configuration process for the entire dynamic topology network resource scheduling of this invention. Detailed Implementation

[0018] The electric commercial vehicle side-swap and vehicle interface integration system disclosed in this specific embodiment is applied to the electric commercial vehicle side-swap scenario. The technical solution of the present invention will be clearly and completely described below with reference to several embodiments.

[0019] In one embodiment, reference Figure 1 The electric commercial vehicle side-mounted battery swapping and vehicle interface integration system includes an integrated battery swapping interface located on the side of the vehicle frame, a binocular vision acquisition unit, a multi-body dynamics deformation pre-compensation algorithm module, and an on-board real-time Ethernet redundancy control module.

[0020] Specifically, the integrated battery swapping interface is fixedly installed on the lateral mounting surface of the longitudinal beam of the electric commercial vehicle frame, and is rigidly connected to the longitudinal beam of the frame by high-strength bolts. The flatness tolerance of the mounting surface is no more than 0.2mm. The integrated battery swapping interface includes docking terminals for the high-voltage power circuit, the low-voltage communication circuit, and the battery thermal management medium circuit. All docking terminals adopt a coaxial nested layout. The specific structure of the coaxial nested layout is as follows: The innermost layer is the docking terminal of the battery thermal management medium circuit, including a coaxial sleeve-type liquid inlet terminal and liquid return terminal. A high-pressure resistant fluororubber seal is set between the liquid inlet terminal and the liquid return terminal. A guide cone surface is set at the docking end, with a cone surface angle of 15°-30°. The middle layer is the docking terminal of the low-voltage communication circuit, which uses 4 sets of differential signal docking pins and sleeves, nested in a ring array outside the thermal management medium circuit docking terminal. The docking stroke of the pins and sleeves is greater than that of the high-voltage power circuit docking terminal. The outermost layer is the docking terminal of the high-voltage power circuit, including coaxially symmetrically arranged positive and negative docking terminals, nested outside the low-voltage communication circuit docking terminal. An insulating sleeve is set between the positive and negative docking terminals. A guide structure is set at the docking end. The contacts of the high-voltage interlock circuit are integrated inside the high-voltage power circuit docking terminal. The conduction stroke of the high-voltage interlock contacts lags behind the conduction stroke of the low-voltage communication circuit docking terminal and leads the full-stroke conduction position of the high-voltage power circuit docking terminal.

[0021] The binocular vision acquisition unit is deployed on the side of the vehicle frame, on the same rigid mounting reference plane as the integrated battery swapping interface. Specifically, it consists of two global shutter industrial cameras symmetrically arranged along the longitudinal direction of the vehicle frame. The baseline distance between the two cameras is set to 150mm, the camera optical axis is parallel to the docking plane of the integrated battery swapping interface, the camera acquisition frame rate is set to 30fps, the image resolution is 1920×1080, the cameras have built-in distortion correction parameters, and binocular stereo calibration is completed before leaving the factory. The binocular vision acquisition unit is configured to acquire continuous frame images of the battery swapping mechanism's execution end and the integrated battery swapping interface in real time. From the acquired images, feature points of the battery swapping mechanism and the vehicle interface are extracted. Feature points include the center of the positioning hole at the docking end of the battery swapping mechanism and the reference corner point of the docking end face of the integrated battery swapping interface.

[0022] The multibody dynamics deformation pre-compensation algorithm module is integrated into the computing unit of the vehicle domain controller, which communicates with the sensing units arranged on the chassis via the vehicle CAN bus. The module is configured to collect load, suspension travel, and chassis lateral deflection data. Load data comes from four sets of strain gauge load cells evenly distributed on the chassis longitudinal beams; suspension travel data comes from four sets of magnetostrictive displacement sensors installed on the front and rear axle suspensions; and chassis lateral deflection data comes from six sets of MEMS tilt sensors arranged every 1.5m along the chassis longitudinal beams. Based on the collected load, suspension travel, and chassis lateral deflection data, the module calculates the spatial pose offset of the battery swapping interface. This spatial pose offset includes the translational offset of the battery swapping interface docking face in the three orthogonal directions (X, Y, and Z) and the rotational offset around the three orthogonal axes. Combined with feature points extracted by the binocular vision acquisition unit, the module corrects the docking coordinates of the battery swapping mechanism.

[0023] The onboard real-time Ethernet redundancy control module adopts a dual-redundant onboard Ethernet architecture supporting the IEEE 802.1AS time-sensitive networking standard. It is configured with dual redundant controllers and dual redundant switches. One Ethernet link communicates with the execution controller of the battery swapping mechanism, while the other Ethernet link communicates with the controllers of the vehicle's high-voltage interlock circuit, battery management system, and vehicle thermal management system. The onboard real-time Ethernet redundancy control module is configured to complete closed-loop control of the on / off timing of the high-voltage power circuit, low-voltage communication circuit, and thermal management medium circuit, as well as automatic configuration of the entire battery swapping process, during the docking process between the integrated battery swapping interface and the battery swapping mechanism.

[0024] In this embodiment, the complete workflow of the system is as follows: After the battery swapping process is triggered, the multibody dynamics deformation pre-compensation algorithm module collects load, suspension travel, and frame lateral deflection data in real time, and calculates the spatial pose offset of the integrated battery swapping interface; the binocular vision acquisition unit collects real-time images of the battery swapping mechanism and the integrated battery swapping interface, extracts the corresponding feature points, and outputs the three-dimensional coordinates of the feature points; the multibody dynamics deformation pre-compensation algorithm module fuses the calculated spatial pose offset with the three-dimensional coordinates of the feature points to complete the correction of the docking coordinates, and sends the corrected docking coordinates to the execution controller of the battery swapping mechanism; the battery swapping mechanism performs docking action according to the corrected docking coordinates, driving the docking terminal of the battery swapping mechanism to complete the physical docking with the docking terminal of the integrated battery swapping interface; after docking is in place, the vehicle-mounted real-time Ethernet redundancy control module collects the docking signal, the high-voltage interlock status signal, and the status parameters of the battery management system, executes closed-loop control of the circuit on / off timing, and sequentially completes the automatic configuration of the entire process of low-voltage communication circuit handshake, thermal management medium circuit conduction, and high-voltage power circuit power-on.

[0025] In one embodiment, the multibody dynamics deformation pre-compensation algorithm module uses a spatiotemporal graph convolutional network to construct a frame deformation-interface displacement mapping model when calculating the spatial pose offset.

[0026] Specifically, firstly, an undirected graph G=(V,E) of the vehicle frame structure is constructed, where V is the set of nodes with N nodes, each node corresponding to a discrete structural unit of the frame, including longitudinal beam segment nodes, crossbeam nodes, suspension mounting nodes, and battery swapping interface mounting nodes. The granularity of node partitioning is that each segment is no longer than 1m. E is the set of edges, and the connection relationships of the edges are consistent with the physical topology of the frame. Undirected edges are set between adjacent structural unit nodes to represent the deformation transmission relationship between structural components. Each node corresponds to a node feature vector. , In this embodiment, the feature dimension is used as the feature dimension. , The components include the local load component, the local suspension travel component, the local lateral deflection component, and the three components of the node's initial three-dimensional coordinates.

[0027] The spatiotemporal graph convolutional network consists of cascaded spatial graph convolutional layers, temporal convolutional layers, adaptive graph attention layers, and fully connected output layers. The spatial graph convolutional layers are used to extract the spatial deformation correlation features of nodes at the same time point. The mathematical expression for its convolution operation is:

[0028] in, For the first The input feature matrix of the layer has a dimension of ; This is the normalized graph adjacency matrix. , For the image The adjacency matrix, It is the identity matrix. It is a degree matrix; Let L be the trainable weight matrix of the l-th layer; The ReLU function is used as the non-linear activation function. Let be the output feature matrix of the l-th layer.

[0029] The temporal convolutional layer adopts a causal convolutional structure with a kernel size of 3×1. It is used to extract the temporal variation features of node features within a continuous time series and output a node feature matrix that integrates temporal information.

[0030] The adaptive graph attention layer aggregates the deformation propagation features of neighboring nodes and assigns adaptive attention weights to different neighboring nodes. The expression for calculating the attention weights is as follows:

[0031] in, Let j be the attention weight relative to node i. Let be the set of neighboring nodes of node i; W is the feature transformation weight matrix; a is the trainable parameter vector for attention weight calculation; || is the feature concatenation operation; LeakyReLU is the non-linear activation function.

[0032] The output feature expression of the adaptive graph attention layer is:

[0033] in, For nodes The aggregated features of the output.

[0034] The fully connected output layer receives the aggregated features output by the adaptive graph attention layer and maps them to the spatial pose offset of the integrated battery swapping interface mounting node. The spatial pose offset is a 6-dimensional vector, including translation offsets in the X, Y, and Z directions and rotation offsets around the three axes.

[0035] Preferably, the spatiotemporal graph convolutional network employs a meta-learning mechanism during training, updating the network's initial parameters by performing second-order gradient optimization on samples under different heavy-load conditions. Specifically, the meta-learning training process uses a model-independent meta-learning (MAML) framework, dividing the frame deformation conditions with different load distributions and suspension travels into multiple training tasks. Each task corresponds to a set of condition samples, including a support set and a query set. For each training task, first-order gradient descent is performed on the support set to update the network's temporary parameters. The update expression for the temporary parameters is:

[0036] in, These are the initial parameters of the network; The learning rate for the inner loop; For the first One training task; For the first The loss function for each training task is the mean squared error loss, where the loss value is the error between the predicted spatial pose offset and the actual measured offset. Initial parameters The corresponding frame deformation-interface displacement mapping model; For the first Temporary parameters updated for each training task.

[0037] On the query set, based on temporary parameters Calculate the loss function, perform second-order gradient descent, and update the network's initial parameters. The update expression for the initial parameters is:

[0038] in, The learning rate is the outer loop rate.

[0039] The initial parameters θ obtained after training enable the frame deformation-interface displacement mapping model to quickly and adaptively calculate spatial pose offsets with only a few iterations when encountering unknown load distribution changes. (Refer to...) Figure 2 .

[0040] In one embodiment, reference Figure 3 When extracting feature points, the binocular vision acquisition unit adopts a sub-pixel edge detection algorithm based on multi-scale fusion.

[0041] Specifically, the execution process of the multi-scale fusion sub-pixel edge detection algorithm is as follows: First, a Gaussian image pyramid is constructed from the single-frame images acquired by the binocular vision acquisition unit. The number of layers in the image pyramid is set to 5. The 0th layer is the original input image, and the kth layer image is obtained by convolving the (k-1)th layer image with a Gaussian kernel and then downsampling by a factor of 2. The size of the Gaussian kernel is 5×5, and the standard deviation is 1.6. For each layer of the image pyramid, the Canny edge detection operator is used to extract the edge response map of that layer, resulting in 5 sets of edge response maps at different scales. Multi-scale fusion is then performed on the 5 sets of edge response maps at different scales. The fusion method is to perform a weighted summation of the edge response values ​​at different scales at the same pixel location. The weights are positively correlated with the local variance of the edge response at that scale. After fusion, a global edge response map is obtained.

[0042] Non-maximum suppression is performed on the global edge response map to remove non-edge pixels, resulting in an initial edge image with a single pixel width. For each edge pixel in the initial edge image, a Zernike moment fitting method is used to fit the sub-pixel-level edge position along the edge normal direction. Specifically, a 7×7 neighborhood window centered on the edge pixel is selected, and the nth-order m-fold Zernike moment of the image within this neighborhood window is calculated. The expression for calculating the Zernike moment is:

[0043] Where n is the order of the Zernike moment, and in this embodiment n is 4; m is the number of repetitions, satisfying that n-|m| is even and |m|≤n; The gray value of the pixel (x,y) within the neighborhood window; ρ is the conjugate complex number of the Zernike polynomial, ρ is the polar radius of the pixel relative to the center of its neighborhood, and θ is the polar angle.

[0044] Based on the calculated Zernike moments, the sub-pixel offsets of the edges are solved. The expression for calculating the sub-pixel coordinates in the edge normal direction is as follows:

[0045] in, Subpixel-level edge coordinates; The integer pixel coordinates of the initial edge pixels; φ is the sub-pixel offset; φ is the normal direction angle of the edge. Both φ and φ are obtained by solving for the magnitude and phase of the Zernike moment.

[0046] Based on the sub-pixel level edge positions obtained from the fitting, the corner points and center of the edge are extracted as feature points of the interface between the battery swapping mechanism and the vehicle.

[0047] Preferably, the multi-scale fusion sub-pixel edge detection algorithm introduces an ambient lighting compensation method based on generative adversarial networks (GANs) during the image pyramid construction process to transform the current lighting image to the standard lighting domain. The GAN includes a generator network and a discriminator network. The generator network adopts a U-Net structure, taking the original image under the current lighting as input and outputting the compensated image in the standard lighting domain. The discriminator network adopts a fully convolutional network structure to distinguish whether the input image is the compensated image output by the generator or the real image in the standard lighting domain. A spatial feature transformation layer is embedded in the generator network. This spatial feature transformation layer is configured to adaptively normalize the multi-level edge responses using the lighting conditional encoding vector. Specifically, firstly, the lighting conditional encoding vector is extracted from the input original image. This 16-dimensional vector represents the overall brightness, contrast, and lighting uniformity parameters of the image. The adaptive normalization expression of the spatial feature transformation layer is:

[0048] in, The input is a multi-level edge response feature map; μ and σ are respectively... The channel mean and channel standard deviation; γ and β are the affine transformation parameters of the spatial feature transformation layer, which are obtained by mapping the illumination conditional encoding vector through a fully connected layer; This is the normalized edge response feature map of the output.

[0049] After adaptive normalization, multi-scale edge response fusion and sub-pixel edge fitting are performed to enhance the robustness of feature point extraction under low light conditions.

[0050] In one embodiment, reference Figure 4 In the process of completing the secondary correction of the docking coordinates, an adaptive unscented Kalman filter method based on federated learning is adopted.

[0051] Specifically, firstly, the state vector and observation vector of the adaptive unscented Kalman filter are defined. , represents the spatial pose offset output by the multibody dynamics deformation pre-compensation algorithm module at time k, comprising translational offsets in the X, Y, and Z directions and rotational offsets around the three axes; observation vector Let be the 3D coordinates of the M feature points extracted by the binocular vision acquisition unit at time k, where M ≥ 4. The state vector... As the prior state vector, the observation vector As the observation vector, perform the iterative process of unscented Kalman filtering.

[0052] In the iterative process of unscented Kalman filtering, sigma point sampling is performed first, resulting in 2n+1 sigma points, where n is the dimension of the state vector. In this embodiment, n=6. The sampling expression for the sigma points is:

[0053] in, Let k be the prior state estimate at time k; λ is the prior state covariance matrix; λ is the scaling parameter. α is 0.01, κ is 0; The i-th column is the square root of the matrix.

[0054] Perform a UT transformation on the sampled sigma points to obtain prior observation estimates. Observation covariance matrix Cross-covariance matrix of state and observation .

[0055] Calculate Kalman gain Update the state estimate at time k. The state estimate is the corrected docking coordinate.

[0056] During the filtering iteration process, variational Bayesian inference is introduced to estimate the measurement noise covariance in real time. Specifically, the measurement noise covariance Assuming the distribution follows an inverse Wieshard distribution, iterative solutions are obtained using variational Bayesian inference. The posterior distribution is obtained. The optimal estimate is used to replace the fixed measurement noise covariance in the unscented Kalman filter, thereby achieving adaptive adjustment of the filter.

[0057] The federated learning framework is configured as follows: the on-board controller of a single commercial vehicle is used as a local node, and the edge computing server of the battery swapping station is used as an aggregation node. Multiple local nodes encrypt and upload the locally filtered and updated model parameters to the aggregation node without uploading the original collected data. After the aggregation node performs parameter aggregation, it distributes the global model parameters to each local node, updates the parameters of the local adaptive unscented Kalman filter, and improves the generalization ability of the filter model under different operating conditions.

[0058] Preferably, after outputting the corrected docking coordinates, the adaptive unscented Kalman filtering method performs a residual prediction step based on a long short-term memory (LSTM) network. Specifically, an LSM network is constructed, with the network input being a sequence of correction errors from historical time points. The correction error is the deviation between the corrected docking coordinates output by the filtered data at historical time points and the actual docking coordinates after the battery swapping mechanism has completed docking. The input sequence length is set to 20, i.e., the correction error sequence of the first 20 time points is input. The network contains two LSM layers, each with a hidden layer dimension of 64, an attention layer, and a fully connected output layer. The attention layer is configured to assign adaptive weights to the temporal features output by the LSM layer, and the weight calculation expression is:

[0059] in, Let be the attention weight for the t-th temporal feature; This is the t-th temporal feature vector output by the Long Short-Term Memory layer; , represents the trainable weights and biases; T is the length of the input sequence.

[0060] The fully connected output layer receives the weighted temporal features and outputs the residual compensation value at the current time. The residual compensation value is a 6-dimensional vector, which is superimposed and fused with the corrected docking coordinates to obtain the final docking coordinates.

[0061] In one embodiment, reference Figure 5 In the closed-loop control process of completing the on-off timing of the loop, a timing decision algorithm based on deep reinforcement learning is used to construct a Markov decision process.

[0062] Specifically, the state space S, action space A, and reward function R of a Markov decision process are defined as follows: The state space S is a continuous state space with 12 dimensions. Its state components include the on / off state of the high-voltage power circuit, the connection state of the low-voltage communication circuit, the conduction state of the thermal management medium circuit, the conduction state of the high-voltage interlock circuit, the interface docking signal, the high-voltage circuit voltage value, the high-voltage circuit current value, the thermal management medium pressure value, the SOC value of the battery management system, the battery temperature value, the locking state of the battery swapping mechanism, and the insulation detection value; the action space A is a continuous action space. The dimension is 4, and the action components include low-voltage communication circuit conduction action, thermal management medium circuit conduction action, high-voltage circuit pre-charging action, high-voltage main circuit conduction action, and the timing interval between each action; the reward function R is defined as follows: when the entire process of conduction and disconnection control is completed and there is no abnormal state, a positive reward value is given; when dangerous working conditions occur such as high-voltage interlock not conducting, high-voltage power-on is performed, medium conduction is performed when communication is not handshake performed, or power-on action is performed when circuit insulation is abnormal, a negative reward value with a larger absolute value is given; when timing timeout occurs, a negative reward value with a smaller absolute value is given.

[0063] Based on the constructed Markov decision process, the optimal on / off timing command is output through a dual-delay deep deterministic policy gradient algorithm. The dual-delay deep deterministic policy gradient algorithm includes an actor network, two independent evaluator networks, and corresponding target actor and target evaluator networks. The actor network is configured to receive the current state s∈S and output a deterministic action a∈A, i.e., the on / off timing command; the two evaluator networks are configured to receive the current state s and action a, and output the corresponding Q-values ​​to evaluate the value of the action.

[0064] The network update process of the dual-delay deep deterministic policy gradient algorithm is as follows: The target actor network outputs the target action a' based on the next time step state s', and adds truncated normal distribution noise to the target action to obtain the smoothed target action a''; the two target evaluator networks receive s' and a'' respectively, output two target Q values, and take the smaller of the two target Q values ​​to calculate the update expression of the target Q value as follows:

[0065] Where y is the target Q value; r is the reward value at the current time; and γ is the discount factor, with a value of 0.99. , These are the outputs of the two target evaluator networks, respectively.

[0066] Based on the target Q value y, the mean squared error loss of the two evaluator networks is calculated, and the parameters of the two evaluator networks are updated by gradient descent.

[0067] The evaluator network is updated twice, and the actor network parameters are updated once. The actor network update uses the policy gradient method, and the gradient expression is:

[0068] Where φ represents the parameters of the actor network; For the mapping function of the actor network; This is the output of the first evaluator network.

[0069] Meanwhile, the parameters of the target actor network and the target evaluator network are updated using a soft update method.

[0070] Preferably, the dual-delay deep deterministic policy gradient algorithm introduces a state representation extraction branch based on contrastive learning into the evaluator network during the output of the optimal on / off timing command. The state representation extraction branch shares the bottom convolutional layer with the feature extraction layer of the evaluator network, outputting a 64-dimensional state representation vector. During the training of contrastive learning, the safe state of the current circuit on / off state is used as a positive sample, and the dangerous states occurring in the past, such as electrical arcing, dielectric leakage, and insulation anomalies, are used as negative samples, constructing positive and negative sample pairs. The contrastive loss function adopts the InfoNCE loss, expressed as:

[0071] Where z is the representation vector of the current state; The representation vector of the positive sample; Let be the representation vector of the i-th negative sample; is the temperature coefficient, with a value of 0.07; N is the number of negative samples.

[0072] By comparing the training of the loss function, the distribution distance between safe and dangerous states in the latent space is increased, which improves the accuracy of the evaluator network in identifying dangerous states, thereby enhancing the safety of the optimal on / off timing command under extreme electrical conditions.

[0073] In one embodiment, reference Figure 6 During the fully automated configuration process, a dynamic topology network resource scheduling method is adopted to dynamically allocate network bandwidth and computing resources in the vehicle-mounted real-time Ethernet redundancy control module.

[0074] Specifically, the entire battery swapping process—including high-voltage power-on, communication handshake, and thermal management circuit activation—is mapped into concurrent task flows with different priorities. These task flows include: high-voltage interlock detection and insulation monitoring, low-voltage communication handshake and battery parameter interaction, thermal management medium circuit activation and pressure detection, high-voltage pre-charging and main circuit power-on, and battery swapping status synchronization and log reporting. Each task flow is assigned a fixed priority, from highest to lowest: high-voltage interlock detection and insulation monitoring, low-voltage communication handshake and battery parameter interaction, thermal management medium circuit activation and pressure detection, high-voltage pre-charging and main circuit power-on, and battery swapping status synchronization and log reporting.

[0075] The vehicle-mounted real-time Ethernet redundancy control module employs a combination of a greedy algorithm and a shortest remaining time priority strategy to schedule and allocate resources for task flows. Specifically, for the highest-priority task flow, a greedy algorithm prioritizes allocating network bandwidth and computing resources that meet the task latency requirements, ensuring that the end-to-end latency of the highest-priority task does not exceed 1ms. For task flows of the same priority, a shortest remaining time priority strategy is used to schedule the task flow with the shortest remaining execution time first, thereby improving task execution efficiency.

[0076] The expression for dynamic allocation of network bandwidth is:

[0077] in, This represents the total bandwidth allocated to the i-th task stream; This provides the basic guaranteed bandwidth for the i-th task flow; Redundant allocable bandwidth; This represents the priority value of the i-th task flow; the higher the priority, the better. The larger; is the maximum allowed latency for the i-th task flow; N is the total number of concurrent task flows.

[0078] The dynamic allocation of computing resources is based on the CPU utilization and real-time requirements of the task flow. A corresponding time slice is allocated to each task flow, with the time slice length positively correlated with the task flow's priority and negatively correlated with its remaining execution time. The in-vehicle real-time Ethernet redundancy control module, based on the TSN (Time-Sensitive Networking) standard, allocates a corresponding time slot to each task flow, ensuring conflict-free Ethernet frame transmission between different task flows. Simultaneously, redundant data transmission is achieved through dual redundant links. When one link fails, data transmission automatically switches to the other link, ensuring reliable issuance of control commands.

Claims

1. A side-mounted battery swapping and vehicle interface integration system for electric commercial vehicles, characterized in that, It includes an integrated battery swapping interface located on the side of the vehicle frame, a binocular vision acquisition unit, a multibody dynamics deformation pre-compensation algorithm module, and an on-board real-time Ethernet redundancy control module. The integrated battery swapping interface includes docking terminals for a high-voltage power circuit, a low-voltage communication circuit, and a battery thermal management medium circuit, and the docking terminals adopt a coaxial nested layout. The binocular vision acquisition unit is deployed on the side of the vehicle frame and is used to extract feature points of the battery swapping mechanism and the vehicle interface; The multibody dynamics deformation pre-compensation algorithm module is used to collect load, suspension travel and frame lateral deflection data, calculate the spatial pose offset of the battery swapping interface and complete the docking coordinate correction in combination with the feature points; The vehicle-mounted real-time Ethernet redundancy control module is connected to the battery swapping mechanism, high-voltage interlock circuit, and battery management system. It is used to complete the closed-loop control of the circuit on / off timing and the system of automatic configuration throughout the entire process during the interface docking process.

2. The electric commercial vehicle side-mounted battery swapping and vehicle interface integration system according to claim 1, characterized in that, When calculating the spatial pose offset, the multibody dynamics deformation pre-compensation algorithm module uses a spatiotemporal graph convolutional network to construct the frame deformation-interface displacement mapping model. It uses load, suspension travel, and frame lateral deflection data as graph node features, and uses an adaptive graph attention mechanism to aggregate the deformation propagation features of adjacent nodes to output the spatial pose offset.

3. The electric commercial vehicle side-mounted battery swapping and vehicle interface integration system according to claim 1, characterized in that, When extracting the feature points, the binocular vision acquisition unit uses a sub-pixel edge detection algorithm based on multi-scale fusion. It extracts multi-level edge responses by constructing an image pyramid and uses a method combining non-maximum suppression and Zernike moments to fit the sub-pixel level edge position in the edge normal direction to generate the feature points.

4. The electric commercial vehicle side-mounted battery swapping and vehicle interface integration system according to claim 1, characterized in that, In the process of completing the secondary correction of the docking coordinates, an adaptive unscented Kalman filter method based on federated learning is adopted. The spatial pose offset is used as the prior state vector, and the feature points are used as the observation vector. The measurement noise covariance is estimated in real time by introducing variational Bayesian inference, and the corrected docking coordinates are output.

5. The electric commercial vehicle side-mounted battery swapping and vehicle interface integration system according to claim 1, characterized in that, In the closed-loop control process of completing the on / off timing of the circuit, a timing decision algorithm based on deep reinforcement learning is adopted to construct a Markov decision process with the circuit on / off state and high voltage interlock state as the state space and the on / off action as the action space. The optimal on / off timing command is output through the double-delay deep deterministic policy gradient algorithm.

6. The electric commercial vehicle side-mounted battery swapping and vehicle interface integration system according to claim 1, characterized in that, During the fully automated configuration process, a dynamic topology network resource scheduling method is adopted to map high-voltage power-on, communication handshake, and thermal management loop activation into concurrent task flows with different priorities. By combining a greedy algorithm with a shortest remaining time priority strategy, network bandwidth and computing resources are dynamically allocated in the vehicle-mounted real-time Ethernet redundancy control module.

7. The electric commercial vehicle side-mounted battery swapping and vehicle interface integration system according to claim 2, characterized in that, The spatiotemporal graph convolutional network employs a meta-learning mechanism during training. By performing second-order gradient optimization on samples under different heavy-load conditions, the initial parameters of the network are updated. This enables the frame deformation-interface displacement mapping model to quickly and adaptively calculate the spatial pose offset with only a few iterations when encountering unknown load distribution changes.

8. The electric commercial vehicle side-mounted battery swapping and vehicle interface integration system according to claim 3, characterized in that, The multi-scale fusion sub-pixel edge detection algorithm introduces an ambient lighting compensation method based on generative adversarial networks during the image pyramid construction process. This method transforms the current lighting image to the standard lighting domain and embeds a spatial feature transformation layer in the generator network. The algorithm then uses lighting conditional coding vectors to adaptively normalize the multi-level edge responses.

9. The electric commercial vehicle side-mounted battery swapping and vehicle interface integration system according to claim 4, characterized in that, After outputting the corrected docking coordinates, the adaptive unscented Kalman filtering method performs a residual prediction step based on a long short-term memory network. The correction error sequence of historical time moments is input into the long short-term memory network, and the temporal error features are extracted through an attention mechanism. The residual compensation value at the current time moment is output and superimposed and fused with the corrected docking coordinates.

10. The electric commercial vehicle side-mounted battery swapping and vehicle interface integration system according to claim 5, characterized in that, In the process of outputting the optimal on / off timing command, the dual-delay deep deterministic strategy gradient algorithm introduces a state representation extraction branch based on contrastive learning into the evaluator network. By forming positive and negative sample pairs between the current on / off state and historical abnormal states, the distribution distance between safe and dangerous states in the latent space is increased.

Images