A method for identifying vehicle abnormalities and a vehicle

Abstract

The application discloses a vehicle anomaly identification method and vehicle, and relates to the technical field of data processing. Based on multi-dimensional time sequence data such as driving information, environmental information and chassis mud information, mud interference is quantified, the dynamic state of each component is captured in real time, braking performance stage division and braking performance prediction are performed, and thus the abnormal component is located. The method can effectively deal with the situation that the braking performance frequently changes, realizes risk early warning, and meets the demand for monitoring the state of the braking system.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method for vehicle anomaly identification and a vehicle. Background Technology

[0002] The cooling of a vehicle's braking system is closely related to its braking performance, and the cooling effect of the braking system is often affected by a variety of factors. When the cooling effect is unstable, it can easily lead to frequent changes in braking performance and poor stability. If the current state of the braking system cannot be effectively identified, it will affect the safety of braking control and bring about driving safety hazards.

[0003] In existing technologies, traditional thermal models or fixed threshold triggering strategies are mainly used to monitor the state of the braking system. This single logic cannot accurately identify the causes of changes in braking performance, nor can it cope with situations where braking performance changes frequently. It cannot meet the requirements for high-precision and high-real-time monitoring of the braking system state in off-road scenarios. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a method and vehicle for vehicle anomaly identification, which can identify and predict abnormal states of the braking system, achieving high-precision and high-real-time braking system state monitoring.

[0005] This application discloses a method for vehicle anomaly identification, the method comprising: Collect multi-source time-series data of the vehicle where the target braking system is located; the multi-source time-series data includes driving information, environmental information, and chassis mud and water information; Based on the multi-source time-series data, the target braking system is subjected to component state identification, braking performance stage division, and braking performance prediction to obtain the degree of abnormality of each component of the target braking system.

[0006] Optionally, the step of performing component state identification, braking performance stage division, and braking performance prediction on the target braking system based on the multi-source time-series data to obtain the anomaly degree of each component of the target braking system includes: A graph network is constructed based on the multi-source time-series data and the physical layout of the target braking system to obtain graph-structured time-series data. The graph-structured time-series data is input into a pre-trained abnormal cooling identification model; the abnormal cooling identification model is used to perform component state identification, braking performance stage division, and braking performance prediction. Based on the output of the abnormal cooling identification model, the degree of abnormality of each component of the target braking system is obtained.

[0007] Optionally, the construction of the graph network based on the multi-source time-series data and the physical layout of the target braking system includes: The multi-source time series data is subjected to time alignment and data generalization processing to obtain processed multi-source time series data; Based on the multi-source time-series data and the physical arrangement, each node represents each component, and each edge represents a heat conduction or pressure connection path. The edge weight is defined by temperature difference, hydraulic pressure difference, or thermal conductivity of the connecting material to construct the graph network.

[0008] Optionally, the step of constructing a graph network based on the multi-source time-series data and the physical arrangement of the target braking system to obtain graph-structured time-series data includes: Establish an off-road scenario condition library that includes scenario tags and mud and water cooling event tags; The graph network is labeled based on the off-road scenario condition database to obtain the graph structured time series data.

[0009] Optionally, the abnormal cooling identification model includes an identification sub-model, a stage clustering sub-model, and a performance degradation sub-model. The step of obtaining the degree of abnormality of each component of the target braking system based on the output of the abnormal cooling identification model includes: The identification sub-model identifies the state of each node in the graph structured time series data at each time step, and obtains the original risk score of each node and the multidimensional embedding vector sequence of each node. The stage clustering sub-model analyzes the braking performance state corresponding to each frame in the multidimensional embedding vector sequence to obtain the stage label of braking performance for each frame; the types of stage labels include initial cooling, falsely high performance, normal transition, accelerated decay, and performance failure. The performance decay sub-model predicts the state evolution sequence of each node within a future preset time period based on the braking performance of the stage label and the multidimensional embedded vector sequence over a historical time period. The model fusion layer of the abnormal cooling identification model obtains the risk score and global risk level of each node of the target braking system based on the original risk score, the stage label and the state evolution sequence.

[0010] Optionally, the model fusion layer of the abnormal cooling identification model obtains the risk score and global risk level of each node of the target braking system based on the original risk score, the stage label, and the state evolution sequence, including: The gating attention fusion controller dynamically assigns weights to the recognition sub-model, the stage clustering sub-model, and the performance degradation sub-model. The basic risk item is obtained by weighting and fusing the original risk score and the stage label, and the forward risk item is obtained by weighting the state evolution sequence. The risk score is generated by aggregating the basic risk items and the forward-looking risk items. The risk scores at continuous time points are mapped to discrete level intervals to obtain the global risk level.

[0011] Optionally, the dynamic allocation of weights to the recognition sub-model, the stage clustering sub-model, and the performance degradation sub-model based on the gated attention fusion controller includes: Based on the coupling strength between the output of the identification sub-model and the output of the stage clustering sub-model, a first weight is obtained; the first weight is used to generate the basic risk item. Based on the output of the performance degradation sub-model, a second weight is obtained; the second weight is used to generate the prospective risk term.

[0012] Optionally, before the model fusion layer dynamically assigns weights to the recognition sub-model, the stage clustering sub-model, and the performance degradation sub-model based on the gated attention fusion controller, the method further includes: The model fusion layer pre-stores vehicle configuration features; the vehicle configuration features include brake disc material number, hydraulic system type, tire diameter, and off-road mode identifier; the vehicle configuration features are used to participate in weight allocation.

[0013] Optionally, after obtaining the degree of abnormality of each component of the target braking system, the method further includes: Obtain the confidence curve of braking performance degradation over a future preset time period.

[0014] Based on the above-mentioned method for vehicle anomaly identification, this application also discloses a device for vehicle anomaly identification, including: a data acquisition unit and an identification unit; The acquisition unit is used to acquire multi-source time-series data of the vehicle where the target braking system is located; the multi-source time-series data includes driving information, environmental information, and chassis mud and water information. The identification unit is used to identify the component status, divide the braking performance into stages, and predict the braking performance of the target braking system based on the multi-source time-series data, so as to obtain the degree of abnormality of each component of the target braking system. Optionally, the identification unit includes: A sub-unit is constructed to build a graph network based on the multi-source time-series data and the physical arrangement of the target braking system, thereby obtaining graph-structured time-series data. The input subunit is used to input the graph-structured time-series data into a pre-trained abnormal cooling identification model; the abnormal cooling identification model is used to perform component state identification, braking performance stage division, and braking performance prediction. The output subunit is used to obtain the degree of abnormality of each component of the target braking system based on the output of the abnormal cooling identification model.

[0015] Optionally, the building subunit includes: The processing subunit is used to perform time alignment processing and data generalization processing on the multi-source time series data to obtain processed multi-source time series data. The graph construction subunit is used to construct the graph network based on the multi-source time-series data and the physical arrangement, with each node representing each component, each edge representing a heat conduction or pressure connection path, and the edge weights defined by temperature difference, hydraulic pressure difference, or thermal conductivity of the connecting material.

[0016] Optionally, the building subunit includes: Create a sub-unit to build an off-road scenario condition library containing scene tags and mud and water cooling event tags; The annotation subunit is used to annotate the graph network based on the off-road scenario condition library to obtain the graph structured time-series data.

[0017] Optionally, the abnormal cooling identification model includes an identification sub-model, a stage clustering sub-model, and a performance degradation sub-model, and the output sub-unit includes: The identification subunit is used by the identification submodel to identify the state of each node in the graph structured time series data at each time step, and to obtain the original risk score of each node and the multidimensional embedding vector sequence of each node. A clustering subunit is used by the stage clustering submodel to analyze the state of the braking performance corresponding to each frame in the multidimensional embedding vector sequence, and to obtain the stage label of the braking performance of each frame; the types of the stage label include initial cooling, falsely high performance, normal transition, accelerated decay, and performance failure; The prediction subunit is used by the performance degradation submodel to predict the state evolution sequence of each node in a future preset time period based on the braking performance of the stage label and the multidimensional embedded vector sequence in the historical time period. The fusion subunit, used in the model fusion layer of the abnormal cooling identification model, obtains the risk score and global risk level of each node of the target braking system based on the original risk score, the stage label, and the state evolution sequence.

[0018] Optionally, the fusion subunit includes: The allocation subunit is used to dynamically allocate weights to the recognition sub-model, the stage clustering sub-model, and the performance degradation sub-model based on the gated attention fusion controller. The weighted subunit is used to weight and fuse the original risk score and the stage label to obtain the basic risk item, and to weight the state evolution sequence to obtain the forward risk item; An aggregation subunit is used to aggregate the basic risk items and the forward-looking risk items to generate the risk score; The mapping subunit is used to map the risk scores of continuous time nodes to discrete level intervals to obtain the global risk level.

[0019] Optionally, the allocation subunit includes: The first weight acquisition subunit is used to acquire a first weight based on the coupling strength between the output of the identification submodel and the output of the stage clustering submodel; the first weight is used to generate the basic risk item. The second weight acquisition subunit is used to acquire the second weight based on the output of the performance degradation sub-model; the second weight is used to generate the forward risk item.

[0020] Optionally, the device further includes: The feature configuration unit is used to store vehicle configuration features in the model fusion layer; the vehicle configuration features include brake disc material number, hydraulic system type, tire diameter, and off-road mode identifier; the vehicle configuration features are used to participate in weight allocation.

[0021] Optionally, the device further includes: The confidence output unit is used to obtain the confidence curve of braking performance degradation within a preset time period in the future.

[0022] Based on the above-mentioned method for vehicle anomaly identification, this application also discloses a vehicle, including: a processor, a memory, and a system bus; The processor and the memory are connected via the system bus; The memory is used to store a program, the program including instructions that, when executed by the processor, cause the processor to perform the steps of the method described above.

[0023] This application discloses a method and vehicle for vehicle anomaly identification. Based on multi-dimensional time-series data such as driving information, environmental information, and chassis mud and water information, it quantifies mud and water interference, captures the dynamic state of each component in real time, performs braking performance stage division and braking performance prediction, and obtains the degree of anomaly of each component with high accuracy and high real-time performance, thereby locating abnormal components. It can effectively cope with situations where braking performance changes frequently, realize risk warning, and meet the needs of braking system status monitoring. Attached Figure Description

[0024] 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 embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating a method for vehicle anomaly identification disclosed in an embodiment of this application; Figure 2 This is a flowchart illustrating another method for vehicle anomaly identification disclosed in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a vehicle anomaly identification device disclosed in an embodiment of this application. Detailed Implementation

[0026] 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 skilled in the art without creative effort are within the scope of protection of this application.

[0027] Example 1: This application discloses a method for vehicle anomaly identification.

[0028] For details, please refer to Figure 1 The method for vehicle anomaly identification disclosed in this embodiment includes the following steps: Step 101: Collect multi-source time-series data of the vehicle where the target braking system is located.

[0029] In the method of this embodiment, multi-source time-series data can be collected by relying on the existing CAN bus and extended sensor system in the vehicle where the target braking system is located. Specifically, this data can include vehicle driving information (such as brake disc temperature, wheel speed sensor signal, brake pressure, brake pedal travel, vehicle attitude, etc.), environmental information (off-road terrain level, ambient temperature, ambient humidity, mud and water depth, terrain roughness level, airflow rate, etc.), and most importantly, chassis mud and water information. This chassis mud and water information can be calculated by the wheel arch water droplet recognition algorithm based on the images collected by the chassis camera and wheel arch mud and water monitoring camera, or it can be obtained by the background and sent to the vehicle.

[0030] In the method of this embodiment, time alignment and data generalization processing can be performed on the collected multi-source time-series data to obtain processed multi-source time-series data. Time alignment processing can employ methods such as timestamp alignment and data interpolation to ensure consistency of multi-source signals of different frequencies at the time of physical event occurrence. For example, fine-grained alignment can be performed on the transient cooling response of the brake temperature sensor to avoid disturbances caused by short-term cooling effects. Data generalization processing can set different enhancement strategies according to actual needs, such as sample expansion for mud-water cooling events, normalization based on different temperature change rates, and reweighting of braking responses based on vehicle speed and inertia. The above-described time alignment and data generalization processing methods are merely examples; the specific algorithms used in these two processing methods are not specifically limited here, as long as they can process multi-source time-series data.

[0031] In the method of this embodiment, a graph network of the target braking system can be constructed based on the processed multi-source time-series data and the physical layout of the target braking system. In this graph network, each node represents a component in the braking system, such as the brake, heat sink, pipeline, brake disc surface, etc. The edges in the graph network represent heat conduction or pressure connection paths, and the edge weights can be defined according to requirements by temperature difference, hydraulic pressure difference, or the thermal conductivity of the connecting materials. In this way, a heterogeneous graph structure with physical semantics can be obtained, which preserves the thermal / mechanical interaction relationships between the components in the braking system, and also benefits from multi-source time-series data, supporting iterative updates based on dynamic time input.

[0032] In the method of this embodiment, an off-road scenario condition library can be pre-established, which includes scenario labels such as wading depth and off-road slope, as well as condition labels such as continuous braking count and mud-water cooling events, as required. After the graph network is labeled with data from the off-road scenario condition library, tagged graph structured time-series data can be obtained.

[0033] Step 102: Based on the multi-source time-series data, perform component state identification, braking performance stage division, and braking performance prediction on the target braking system to obtain the degree of abnormality of each component of the target braking system.

[0034] In the method of this embodiment, a normal cooling identification model can be pre-built and trained, and then graph-structured time-series data can be input into the abnormal cooling identification model for processing. As an feasible solution, the abnormal cooling identification model is a multi-sub-model fusion architecture model, which may include an identification sub-model, a stage clustering sub-model, and a performance degradation sub-model.

[0035] The identification sub-model is used to identify the state of each component of the target braking system at each moment. As a feasible approach, during the model building and training phases, the identification sub-model can achieve component-level modeling based on the GraphTransformer framework. For example, the braking system can be abstracted into a physical graph structure containing four braking units: front left, front right, rear left, and rear right. Each braking unit is further refined into four types of components: brake disc, caliper, heat conduction interface, and hydraulic lines, resulting in approximately 16 nodes. The nodes are connected by weighted edges formed by heat conduction paths and pressure paths, ultimately resulting in a heterogeneous graph containing approximately 16 nodes and 24 edges.

[0036] In the method of this embodiment, the feature vectors of each node and each edge in the heterogeneous graph are determined by the input graph-structured time-series data. Specifically, the node feature vectors may include seven dimensions: brake disc temperature (which can be a floating-point number in degrees Celsius), temperature change rate (°C / s), hydraulic pressure (bar), wheel speed disturbance value (Hz median deviation), mud and water influence flag (which can be 0 / 1 labels) for the most recent preset time (e.g., 5 seconds), acceleration influence coefficient (obtained through an inertial measurement unit), and braking duration (seconds). The edge feature vectors may include three dimensions: material thermal conductivity, fluid pressure direction weight, and structural connection strength scalar.

[0037] In the method of this embodiment, the structure of the recognition sub-model can employ an encoder (such as a Graph Transformer) containing multiple independently learnable attention heads. One attention head is dedicated to introducing a mask for the local cooling signal under muddy water interference, and another attention head can capture the time delay response of the hydraulic system. For example, in one feasible scheme, the recognition sub-model can contain four stacked layers, each with four attention heads, and each attention head independently learns the attention weights of a physical interaction relationship. The node embedding dimension is set to 64, and the edge embedding dimension is set to 16. Finally, each node outputs a 64-dimensional embedding vector and a raw risk score between 0 and 1. This raw risk score reflects the state of the node at the current moment, specifically representing the probability of the node's contribution to the subsequent braking performance degradation trend at the current moment. Taking 16 nodes as an example, the recognition sub-model can finally output the raw risk score vector of each node (16 nodes × 1 dimension) and the embedding vector matrix of each node (16 × 64 dimensions), which serve as the core input of the staged clustering sub-model. The aforementioned attention mechanism enables the sub-model to jointly model structure and time, accurately identifying abnormal cooling behavior caused by mud and water under complex working conditions and the propagation trend of this behavior within the braking system, providing structural prior support for the staged clustering sub-model.

[0038] In the method of this embodiment, the training of the identification sub-model can employ self-supervised training. Specifically, the prediction task can be designed as the future state of a node, using the state difference between two frames as pseudo-labels. The model is trained to learn the dynamic response law of heat and cold by minimizing the prediction error, and the mean squared error can be used as the loss function. As another feasible approach, supervised fine-tuning can be introduced on the basis of self-supervised training. Specifically, a dataset of manually labeled cooling anomaly nodes can be used, and a binary classification cross-entropy loss optimization model can be employed to identify whether each node is in a state of "artificially high performance after cooling anomaly".

[0039] In the method of this embodiment, the stage clustering sub-model is used to obtain the state stage of the braking performance of the target braking system. As an feasible solution, during the model building and training stages, the stage clustering sub-model can adopt a self-supervised temporal clustering (SSTC) architecture. It mainly uses a self-supervised approach to automatically discover the performance changes experienced by the braking system under the influence of mud and water cooling from the multi-dimensional embedding vector sequence output by the identification sub-model, and divides the entire braking process into multiple state stages.

[0040] As a feasible approach, taking 16 nodes as an example, the staged clustering sub-model receives a 16×64-dimensional embedding vector matrix and arranges them into a sliding window in chronological order. The size of each time window can be set to 10 seconds, corresponding to approximately 100 frames of node embedding. Each frame contains 16 nodes, and each node embedding has a dimension of 64. Therefore, the input dimension of a single time window is 100×16×64.

[0041] In this embodiment, the structural aspect of the stage clustering sub-model can first be addressed by setting a lightweight temporal encoder to map the multidimensional embedding vector sequence into a compressed global stage feature vector. Specifically, this encoder can employ a single-layer Temporal Convolutional Network (TCN) and add a deformable convolutional module to enhance the perception of stage fluctuations. For training, a self-supervised clustering mechanism is used to construct pseudo-labels. The state stage vectors of braking performance within two consecutive time windows are compared, and the probability estimate of whether the braking performance within the two time windows belongs to the same state stage is optimized using the Maximum Mutual Information (MMI) criterion.

[0042] The staged clustering sub-model internally maintains a dynamically updated pool of cluster center vectors. Each cluster center represents a state stage. In each training iteration, the affiliation is updated based on the distance between the current state stage vector and the cluster center. Simultaneously, each cluster center is assigned a representative stage label, such as "early cooling down," "artificially high performance," "normal transition," "accelerated decay," or "performance failure." Therefore, the staged clustering sub-model can reliably identify the start time of the "artificially high performance" stage, significantly improving the response speed of the performance decay sub-model.

[0043] In this embodiment, the stage clustering sub-model is trained without manual labels, relying solely on temporal consistency enhancement and intra-stage contrastive loss. The loss function can be designed to minimize intra-stage distance and maximize inter-stage difference. This enables high-quality stage segmentation of complex nonlinear degradation processes and achieves a balance between physical consistency and temporal stability through co-optimization with structure-aware embedding, providing a clear stage prior structure for the subsequent performance degradation sub-model. The final output data of the stage clustering sub-model is the stage label of braking performance within each time window, as well as the generated center vector representation of each stage, used for state initialization of the performance degradation sub-model.

[0044] In this embodiment, the performance degradation sub-model is used to predict the state evolution of braking performance. It primarily describes the nonlinear trajectory of braking performance changes over time after the slurry cools, such as the transition from an "abnormal cooling" state to a "performance degradation" state. As a feasible approach, the performance degradation sub-model can specifically employ a Neural Ordinary Differential Equation (Neural ODE) structure to model the state evolution process in a continuous time space. The core advantage of the Neural ODE model lies in its ability to model state changes on non-uniform timescales. Compared to traditional RNN models, it exhibits higher fitting accuracy and physical plausibility when handling nonlinear degradation curves after abrupt cooling changes. Combined with pre-sequence structure embedding and degradation stage information, it can achieve high-resolution, highly physically consistent performance degradation trend modeling, providing a continuously differentiable predictive foundation for subsequent model fusion.

[0045] A learnable differential function network f(θ) is used, taking the current state of the braking system as input and outputting its derivative (i.e., rate of change). This differential function network can be configured as a 3-layer fully connected network with hidden layer dimensions of 128 or 64, and the activation function is Swish. The differential solution can use a Dopri5 variable-step explicit integrator, with a maximum step size of 0.1s and a minimum step size of 0.01s, to predict the state changes of each node within a preset time period (e.g., 5 seconds) and obtain the state evolution curve.

[0046] As a feasible approach, to improve the fitting ability of the performance decay sub-model to the "sudden decay" segment, a rate-of-change weighting mechanism can be added during the training process of the differential function network to increase the gradient feedback weight for state segments with large temperature drop rates and large braking response time rise rates, thereby achieving enhanced modeling of abrupt decay states.

[0047] As a feasible approach, the loss function for training the performance degradation sub-model can employ a two-stage joint optimization strategy. The first stage uses the trajectory reconstruction error as the main loss, which is the L2 difference (i.e., Euclidean loss) between the predicted value and the actual braking performance parameters (brake disc temperature, response delay, braking deceleration) in the continuous time space. The second stage introduces a degradation rate regularization term to constrain the stability of the rate of change and avoid unreasonable oscillations in the prediction.

[0048] In this embodiment, the input data for the performance degradation sub-model consists of braking performance data from a multi-dimensional embedding vector sequence from the identification sub-model over a historical time period, and stage labels from the stage clustering sub-model. Taking 16 nodes as an example, 50 frames of data from the multi-dimensional embedding vector sequence within a 5-second time window can be extracted, with each frame having a dimension of 16×64. Furthermore, the multi-dimensional embedding vector sequence also includes environmental information such as off-road terrain level, ambient temperature, ambient humidity, mud and water depth, terrain roughness level, and airflow rate. This environmental information serves as an auxiliary input feature vector, and its dimension is set according to actual needs (e.g., 12).

[0049] In the method of this embodiment, the model fusion layer of the abnormal cooling identification model can fuse the identification sub-model, the stage clustering sub-model and the performance degradation sub-model to obtain a model that supports full-process monitoring, feature decoupling and degradation trend prediction of braking performance abnormalities caused by mud and water.

[0050] As a feasible solution, the multi-sub-model fusion framework of the anomaly cooling identification model can be divided into three layers: the bottom layer is the feature acquisition and alignment layer, the middle layer is the unified output representation layer, and the top layer is the fusion strategy scheduling and risk inference layer. The bottom layer is used to uniformly calibrate the output features of the three sub-models in time and space, ensuring that the outputs of each sub-model have fully aligned timestamps and node sequence order. For example, the multi-dimensional embedding vector matrix of each moment output by the identification sub-model can be cached with a preset sampling period (e.g., 100ms). The stage labels output by the stage clustering sub-model are aligned with the same sampling period, and synchronization anchor points are inserted at stage label switching points. The performance degradation sub-model outputs the state evolution of nodes over a future period, which is then filled in to the current moment using a sliding window method and concatenated with historical node states to obtain a complete time-chain state evolution sequence.

[0051] In the method of this embodiment, the middle layer is used to transform the outputs of the three sub-models into fused feature vectors under a unified semantic space. Specifically, a lightweight representation alignment network (RAN) can be designed, taking the outputs of the three sub-models as input data, and mapping the input data into 128-dimensional representation vectors through three different encoders, and then concatenating them to form a 384-dimensional fused representation.

[0052] In this embodiment, the top layer is used to implement the key model fusion strategy, which can employ a Gated Attention Fusion Controller (GAFC). Based on the output data of each sub-model in the fusion representation obtained by the current middle layer, the GAFC dynamically allocates the weight ratio of each sub-model in the final decision output. As one feasible approach, the GAFC can include two gated sub-networks. The first gated network calculates the coupling strength between the original risk score output by the identification sub-model (reflecting the degree of component anomaly) and the stage label output by the stage clustering sub-model (reflecting state stage changes and their intensity), and obtains a first weight based on this coupling strength. The second gated network evaluates the rate of change of the state evolution sequence output by the performance degradation sub-model and outputs a second weight.

[0053] Subsequently, at the signal level, the product of the first weight and the original risk score signal is calculated, and the difference between this product and the stage label signal is used to obtain the basic risk term. The product of the second weight and the state evolution sequence signal is calculated to obtain the prospective risk term. Finally, the basic risk term and the prospective risk term are aggregated to generate the risk score for each node. The risk scores of each node at continuous time nodes are then mapped to discrete level intervals to obtain a global risk level that reflects the global anomaly degree of the braking system within a preset time period. Taking 16 nodes as an example, the abnormal cooling identification model finally outputs a set of node-level comprehensive risk score vectors (16 nodes × 1 dimension) and a global braking system trend risk level (which can be quantized through levels 1 to 5).

[0054] This multi-model fusion framework introduces a collaborative mechanism among three types of models: structure perception, stage classification, and continuous trend prediction. It forms a closed loop from physical response modeling and state evolution understanding to future trend prediction, possessing strong robustness, transferability, and interpretability. It can stably cope with nonlinear abrupt changes and stage jumps in braking performance under complex working conditions.

[0055] In this embodiment, to enhance the adaptability of model fusion to differences among vehicles, the model fusion layer can pre-store vehicle configuration features as additional input conditions. These features include brake disc material number, brake disc size, heat dissipation type (ventilated / non-ventilated), hydraulic system type, hydraulic system model, tire diameter, off-road mode identifier, deceleration threshold setting, vehicle curb weight, off-road driving mode type, and terrain adaptability coefficient. These vehicle configuration features can be used to participate in weight allocation.

[0056] In this embodiment, as an feasible approach, the abnormal cooling identification model can also output a confidence curve of braking performance decay over a preset future time period, where the confidence level can be a floating-point sequence between 0 and 1. This decay confidence curve represents the model's reliability assessment of future risk prediction results. Specifically, it can assess the spatial consistency of risk for the identification sub-model; for example, coordinated risk distribution across nodes and the absence of isolated outliers indicate a more reliable judgment. It can also measure the clarity of the current state stage division for the stage clustering sub-model; for example, higher confidence occurs when samples are far from the cluster boundary. Furthermore, it can calculate the prediction stability for the performance decay sub-model, such as whether the trajectory is smooth, whether there are abnormal oscillations or unreasonable mutations; higher stability equates to stronger confidence. These three pieces of information are weighted and fused, and then calculated hourly on the future timeline to obtain a confidence sequence over a period of time. Subsequently, robust optimization processing, including temporal smoothing (sliding window and exponential decay) and spatial consistency correction (soft pruning of outlier nodes), is applied to eliminate noise and mutation effects. Finally, a continuous confidence curve is formed. This confidence curve can be used in conjunction with the final output of the abnormal cooling identification model. For example, it can trigger a strong warning when the risk is high and the confidence level is also high, and make a cautious decision or wait for more information when the risk is high but the confidence level is low.

[0057] In the method of this embodiment, after the abnormal cooling identification model is trained, a mechanism for prediction robustness optimization and risk score stabilization can be designed to address issues such as perturbation sensitivity, prediction instability, and large score fluctuations that may occur in practical applications, so that the output of the abnormal cooling identification model remains reliable in complex off-road scenarios.

[0058] The robustness optimization component employs a two-layer smoothing process for the model output in the time domain. The first layer uses a sliding window mean filter to smooth each risk score briefly, eliminating interference caused by sudden changes in a single frame. The second layer uses a trend-aware exponential decay filter, automatically adjusting the decay coefficient based on the current rate of change in the score. This maintains responsiveness during sudden trend changes and enhances score stability during gradual trend changes. Simultaneously, a node consistency discrimination mechanism is introduced in the spatial domain. Standard deviation analysis is performed on the risk scores of each node. When the risk score of a node deviates from the global average by more than a set threshold, a confidence adjuster is triggered. This confidence adjuster can soft-prune or reweight excessively deviating node risk scores to prevent local noise amplification from affecting the overall judgment. A score compensation module can also be integrated to prevent misjudgments caused by missing or malfunctioning sensor data. For example, if temperature or pressure data for a node is missing for more than 0.3 seconds, a graph imputation model based on the embedded features of neighboring nodes is activated to perform high-dimensional completion of the missing items and simultaneously correct the risk score of the current frame.

[0059] In this embodiment, the risk score stabilization is performed after robust optimization, primarily generating a phased risk index using a time-weighted approach. A multi-factor scorer can be designed, comprising three parts: the first part is a basic risk score module, which uses the mean of the risk scores to weight the current stage label to form an initial score. The second part is a trend acceleration factor, which estimates the current braking performance degradation rate based on the state evolution sequence and weights it accordingly in the final risk score as a speed factor. The third part is a stage confidence adjustment factor, which determines the score's reliability range based on the confidence level of the current state stage. The final risk score formula is a weighted sum of these three parts, where the weights are dynamically assigned by a fusion strategy learned from vehicle configuration parameters and test data, achieving a personalized scoring system tailored to each vehicle.

[0060] In the method of this embodiment, the scoring results include a single-frame instantaneous score for real-time monitoring of performance fluctuations, a continuous scoring curve showing future degradation trends, and a braking system health level label (which can be set to four levels: green, yellow, orange, and red, where orange and red levels trigger the warning channel and control strategy linkage mechanism).

[0061] In the method of this embodiment, to address the issue of individual vehicle differences in actual engineering deployments, a personalized model fine-tuning and adaptability verification mechanism driven by real-vehicle feedback data can be constructed. This enables the abnormal cooling identification model to have transferable, generalizable, and rapidly convergent deployment capabilities under different configurations, braking structures, geographical environments, and driving behaviors. The overall mechanism can be divided into three parts: personalized configuration vector encoding, a closed-loop acquisition and fine-tuning triggering mechanism for real-vehicle feedback data, and a model accuracy evaluation and adaptability verification process. Among them, personalized configuration encoding can construct a unique vehicle configuration feature descriptor for each vehicle's physical structure and software parameters.

[0062] The real-vehicle feedback data closed-loop acquisition and fine-tuning trigger mechanism continuously collects braking event data during actual off-road operation (especially marking periods containing a complete closed loop of mud and water erosion, abnormal cooling, artificially high response, and performance degradation), and defines two types of trigger conditions. Meeting either trigger condition initiates the personalized fine-tuning phase. One type of trigger condition is scoring deviation triggering, where a significant deviation occurs between the model deceleration curve calculated based on the model output and the measured deceleration curve (e.g., exceeding 0.2g or a prediction deviation lasting more than 3 seconds). The other type of trigger condition is stage identification failure triggering, where multiple consecutive predictions fail to correctly identify the "artificially high performance" stage, indicating that the current abnormal cooling identification model has a blind spot for this vehicle.

[0063] In this personalized fine-tuning mode, the local fine-tuning optimizer can be invoked to make minor adjustments to the weights in the abnormal cooling identification model. Training samples during the fine-tuning process are automatically constructed using the vehicle's own data, requiring no manual intervention. The loss function is automatically calculated and generated from the model's predicted output and actual braking performance. After fine-tuning, the model enters the adaptability verification process. A replay verification mechanism is set up, using recent running data (e.g., 30 minutes) to perform replay inference on the fine-tuned model, verifying prediction lead, scoring stability, and stage identification accuracy. If preset accuracy indicators are met (e.g., prediction lead greater than 5 seconds, scoring standard deviation less than 0.12, and stage identification accuracy greater than 0.9), the fine-tuned model is updated to the current primary version, and the model version number is associated with the configuration vector and archived. If the requirements are not met, the model is rolled back to the previous version.

[0064] In the method of this embodiment, to improve cross-vehicle adaptation efficiency, a central model update platform can be constructed to upload the fine-tuned and verified model parameters and configuration vectors to the server for other vehicles with similar configurations to initialize and load. This avoids redundant training and achieves rapid adaptation within the same platform.

[0065] The method described in this embodiment introduces mud and water information to quantify mud and water interference. Simultaneously, it integrates physical structure and dynamic time-series data by constructing a graph network, transforming the physical relationships such as heat conduction and pressure connectivity of various components in the braking system into a graph network topology. This gives the data clear physical meaning and solves the problem of lost spatial correlation information in traditional time-series data. The abnormal cooling identification model based on the three-model fusion architecture achieves state identification, stage division, and trend prediction, overcoming the shortcomings of single-model functionality and insufficient accuracy. The output results have low error and high reliability, solving problems such as abnormal cooling caused by mud and water erosion in the braking system, difficulty in identifying inflated performance, and lag in predicting degradation trends. It also supports the integration of vehicle configuration coding vectors to achieve personalized fusion strategies. Furthermore, a robust optimization strategy, a data missing compensation module, and a multi-factor dynamic risk scoring mechanism are designed to effectively adapt to complex off-road conditions, enabling the model to work stably and reliably under various disturbances. The scoring results can be linked to the vehicle's safety control strategy, achieving seamless integration of risk prediction and safety control, and meeting the requirements for high-precision, high-real-time braking system state monitoring. An automated, personalized fine-tuning and adaptation mechanism based on real-vehicle feedback was established, and a central model update platform was built to achieve cloud-based parameter sharing and improve cross-vehicle adaptation efficiency.

[0066] Example 2: This application discloses another method for vehicle anomaly identification; please refer to [link / reference]. Figure 2 This embodiment describes the internal workflow of the abnormal cooling identification model.

[0067] Step 201: Collect multi-source time-series data of the target braking system.

[0068] Step 202: Construct a graph network based on multi-source time-series data and the physical layout of the target braking system to obtain graph-structured time-series data.

[0069] Step 203: Input the graph-structured time series data into the abnormal cooling identification model, which is then received by the identification sub-model.

[0070] Step 204: The identification sub-model outputs the original risk score of each node and the multidimensional embedding vector sequence of each node based on the graph structured time series data.

[0071] Step 205: The stage clustering sub-model receives the multidimensional embedding vector sequence from the recognition sub-model and outputs the stage label of braking performance for each frame.

[0072] Step 206: The performance degradation sub-model receives the multidimensional embedding vector sequence from the identification sub-model and the stage labels from the stage clustering sub-model, and outputs the state evolution sequence of each node over a future period.

[0073] Step 207: The model fusion layer of the abnormal cooling identification model calculates the first weight and the second weight based on the original risk score, stage label and state evolution sequence.

[0074] Step 208: The model fusion layer calculates the basic risk items based on the first weight, the original risk score, and the stage label.

[0075] Step 209: The model fusion layer calculates the forward risk term based on the second weight and the state evolution sequence.

[0076] Step 210: The model fusion layer calculates the risk score for each node based on the basic risk items and the forward-looking risk items.

[0077] Step 211: The model fusion layer maps the risk scores of each node at continuous time points to discrete level intervals to obtain the global risk level.

[0078] Step 212: The abnormal cooling identification model outputs the risk score of each node and the global risk level.

[0079] Based on the vehicle anomaly identification method disclosed in the above embodiments, this embodiment correspondingly discloses a vehicle anomaly identification device. Please refer to... Figure 3 The device for identifying vehicle anomalies includes: a data acquisition unit 301 and an identification unit 302; The acquisition unit 301 is used to acquire multi-source time-series data of the vehicle where the target braking system is located; the multi-source time-series data includes driving information, environmental information and chassis mud and water information. The identification unit 302 is used to identify the component status, divide the braking performance into stages, and predict the braking performance of the target braking system based on the multi-source time-series data, so as to obtain the degree of abnormality of each component of the target braking system. Optionally, the identification unit 302 includes: A sub-unit is constructed to build a graph network based on the multi-source time-series data and the physical arrangement of the target braking system, thereby obtaining graph-structured time-series data. The input subunit is used to input the graph-structured time-series data into a pre-trained abnormal cooling identification model; the abnormal cooling identification model is used to perform component state identification, braking performance stage division, and braking performance prediction. The output subunit is used to obtain the degree of abnormality of each component of the target braking system based on the output of the abnormal cooling identification model.

[0080] Optionally, the building subunit includes: The processing subunit is used to perform time alignment processing and data generalization processing on the multi-source time series data to obtain processed multi-source time series data. The graph construction subunit is used to construct the graph network based on the multi-source time-series data and the physical arrangement, with each node representing each component, each edge representing a heat conduction or pressure connection path, and the edge weights defined by temperature difference, hydraulic pressure difference, or thermal conductivity of the connecting material.

[0081] Optionally, the building subunit includes: Create a sub-unit to build an off-road scenario condition library containing scene tags and mud and water cooling event tags; The annotation subunit is used to annotate the graph network based on the off-road scenario condition library to obtain the graph structured time-series data.

[0082] Optionally, the abnormal cooling identification model includes an identification sub-model, a stage clustering sub-model, and a performance degradation sub-model, and the output sub-unit includes: The identification subunit is used by the identification submodel to identify the state of each node in the graph structured time series data at each time step, and to obtain the original risk score of each node and the multidimensional embedding vector sequence of each node. A clustering subunit is used by the stage clustering submodel to analyze the state of the braking performance corresponding to each frame in the multidimensional embedding vector sequence, and to obtain the stage label of the braking performance of each frame; the types of the stage label include initial cooling, falsely high performance, normal transition, accelerated decay, and performance failure; The prediction subunit is used by the performance degradation submodel to predict the state evolution sequence of each node in a future preset time period based on the braking performance of the stage label and the multidimensional embedded vector sequence in the historical time period. The fusion subunit, used in the model fusion layer of the abnormal cooling identification model, obtains the risk score and global risk level of each node of the target braking system based on the original risk score, the stage label, and the state evolution sequence.

[0083] Optionally, the fusion subunit includes: The allocation subunit is used to dynamically allocate weights to the recognition sub-model, the stage clustering sub-model, and the performance degradation sub-model based on the gated attention fusion controller. The weighted subunit is used to weight and fuse the original risk score and the stage label to obtain the basic risk item, and to weight the state evolution sequence to obtain the forward risk item; An aggregation subunit is used to aggregate the basic risk items and the forward-looking risk items to generate the risk score; The mapping subunit is used to map the risk scores of continuous time nodes to discrete level intervals to obtain the global risk level.

[0084] Optionally, the allocation subunit includes: The first weight acquisition subunit is used to acquire a first weight based on the coupling strength between the output of the identification submodel and the output of the stage clustering submodel; the first weight is used to generate the basic risk item. The second weight acquisition subunit is used to acquire the second weight based on the output of the performance degradation sub-model; the second weight is used to generate the forward risk item.

[0085] Optionally, the device further includes: The feature configuration unit is used to store vehicle configuration features in the model fusion layer; the vehicle configuration features include brake disc material number, hydraulic system type, tire diameter, and off-road mode identifier; the vehicle configuration features are used to participate in weight allocation.

[0086] Optionally, the device further includes: The confidence output unit is used to obtain the confidence curve of braking performance degradation within a preset time period in the future.

[0087] Based on the above-mentioned method for vehicle anomaly identification, this application also discloses a vehicle, including: a processor, a memory, and a system bus; The processor and the memory are connected via the system bus; The memory is used to store a program, the program including instructions that, when executed by the processor, cause the processor to perform the steps of the method described above.

[0088] Based on the above-mentioned method for vehicle anomaly identification, this application also discloses a vehicle, including: a processor, a memory, and a system bus; The processor and the memory are connected via the system bus; The memory is used to store a program, the program including instructions that, when executed by the processor, cause the processor to perform the steps of the method described above.

[0089] Based on the above-described method for vehicle anomaly identification, this application also discloses a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the above-described method.

[0090] Based on the above-described method for vehicle anomaly identification, this application also discloses a storage medium for storing computer program instructions, which, when executed by a central processing unit, are used to implement the steps of the above-described method.

[0091] The embodiments in this specification are described in a progressive manner. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant details can be found in the method section.

[0092] 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 comprises 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 phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0093] 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.

[0094] The features described in the embodiments of this specification can be substituted for or combined with each other, so that those skilled in the art can implement or use this application.

[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 method for vehicle anomaly identification, characterized in that, include: Collect multi-source time-series data of the vehicle where the target braking system is located; The multi-source time-series data includes driving information, environmental information, and chassis mud and water information; Based on the multi-source time-series data, the target braking system is subjected to component state identification, braking performance stage division, and braking performance prediction to obtain the degree of abnormality of each component of the target braking system.

2. The method according to claim 1, characterized in that, The process of identifying component states, dividing braking performance stages, and predicting braking performance of the target braking system based on the multi-source time-series data, to obtain the degree of anomaly of each component of the target braking system, includes: A graph network is constructed based on the multi-source time-series data and the physical layout of the target braking system to obtain graph-structured time-series data. The graph-structured time-series data is input into a pre-trained abnormal cooling identification model; the abnormal cooling identification model is used to perform component state identification, braking performance stage division, and braking performance prediction. Based on the output of the abnormal cooling identification model, the degree of abnormality of each component of the target braking system is obtained.

3. The method according to claim 2, characterized in that, The construction of the graph network based on the multi-source time-series data and the physical layout of the target braking system includes: The multi-source time series data is subjected to time alignment and data generalization processing to obtain processed multi-source time series data; Based on the multi-source time-series data and the physical arrangement, each node represents each component, and each edge represents a heat conduction or pressure connection path. The edge weight is defined by temperature difference, hydraulic pressure difference, or thermal conductivity of the connecting material to construct the graph network.

4. The method according to claim 2, characterized in that, The process of constructing a graph network based on the multi-source time-series data and the physical layout of the target braking system to obtain graph-structured time-series data includes: Establish an off-road scenario condition library that includes scenario tags and mud and water cooling event tags; The graph network is labeled based on the off-road scenario condition database to obtain the graph structured time series data.

5. The method according to claim 2, characterized in that, The abnormal cooling identification model includes an identification sub-model, a stage clustering sub-model, and a performance degradation sub-model. The abnormality level of each component of the target braking system is obtained based on the output of the abnormal cooling identification model, including: The identification sub-model identifies the state of each node in the graph structured time series data at each time step, and obtains the original risk score of each node and the multidimensional embedding vector sequence of each node. The stage clustering sub-model analyzes the braking performance state corresponding to each frame in the multidimensional embedding vector sequence to obtain the stage label of braking performance for each frame; the types of stage labels include initial cooling, falsely high performance, normal transition, accelerated decay, and performance failure. The performance decay sub-model predicts the state evolution sequence of each node within a future preset time period based on the braking performance of the stage label and the multidimensional embedded vector sequence over a historical time period. The model fusion layer of the abnormal cooling identification model obtains the risk score and global risk level of each node of the target braking system based on the original risk score, the stage label and the state evolution sequence.

6. The method according to claim 5, characterized in that, The model fusion layer of the abnormal cooling identification model obtains the risk score and global risk level of each node of the target braking system based on the original risk score, the stage label, and the state evolution sequence, including: The gating attention fusion controller dynamically assigns weights to the recognition sub-model, the stage clustering sub-model, and the performance degradation sub-model. The basic risk item is obtained by weighting and fusing the original risk score and the stage label, and the forward risk item is obtained by weighting the state evolution sequence. The risk score is generated by aggregating the basic risk items and the forward-looking risk items. The risk scores at continuous time points are mapped to discrete level intervals to obtain the global risk level.

7. The method according to claim 6, characterized in that, The dynamic allocation of weights to the recognition sub-model, the stage clustering sub-model, and the performance degradation sub-model based on the gated attention fusion controller includes: Based on the coupling strength between the output of the identification sub-model and the output of the stage clustering sub-model, a first weight is obtained; the first weight is used to generate the basic risk item. Based on the output of the performance degradation sub-model, a second weight is obtained; the second weight is used to generate the prospective risk term.

8. The method according to claim 6, characterized in that, Before the model fusion layer dynamically assigns weights to the recognition sub-model, the stage clustering sub-model, and the performance degradation sub-model based on the gated attention fusion controller, the method further includes: The model fusion layer pre-stores vehicle configuration features; the vehicle configuration features include brake disc material number, hydraulic system type, tire diameter, and off-road mode identifier; the vehicle configuration features are used to participate in weight allocation.

9. The method according to any one of claims 1-8, characterized in that, After obtaining the degree of abnormality of each component of the target braking system, the method further includes: Obtain the confidence curve of braking performance degradation over a future preset time period.

10. A vehicle, characterized in that, include: Processor, memory, system bus; The processor and the memory are connected via the system bus; The memory is used to store a program, the program including instructions that, when executed by the processor, cause the processor to perform the steps of the method according to any one of claims 1-9.

Images