Vehicle surrounding structure fault prediction maintenance method and system based on machine learning

Abstract

The invention discloses a vehicle surrounding structure fault prediction maintenance method and system based on machine learning, and relates to the technical field of vehicle intelligent monitoring and prediction maintenance. Through fusion of vibration, strain and environment corrosion data and a finite element simulation stress map, an environment-load collaborative damage effect is quantified through a physical degradation model, and a multi-modal feature vector is generated. Constructing a dynamic graph topology network, adjusting edge weights in real time, and synchronously capturing spatio-temporal mechanics characteristics by using spatio-temporal graph convolution; and further identifying key risk nodes through a graph attention mechanism, simulating a damage propagation path in combination with a corrosion attenuation coefficient and a graph diffusion model, and generating a probabilistic fault thermodynamic diagram. The fault prediction precision of the vehicle surrounding structure is improved, an adversarial generative network is constructed to optimize the sample weight, and a quantile maintenance list is output in combination with a physical constraint loss function. Fault prediction and maintenance optimization of the vehicle surrounding structure can be effectively carried out, the service life of the structure is prolonged, and the maintenance cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle intelligent monitoring and predictive maintenance, and specifically to a vehicle enclosure structure fault predictive maintenance method and system based on machine learning. Background Art

[0002] With the development of intelligent manufacturing and intelligent operation and maintenance technologies, modern automobile structures are becoming increasingly complex. Vehicles operate at high speeds and in harsh environments. As key protective components, the structural safety and service reliability of their surrounding structures have a significant impact on vehicle performance, driving safety, and user experience. To achieve risk management and preventive maintenance throughout the vehicle lifecycle, the industry has placed higher demands on structural health monitoring (SHM), multi-source data-based fault prediction, and remaining life assessment. In particular, intelligent methods that integrate machine learning and physical models have gradually become a research hotspot and development direction for predictive maintenance of intelligent connected vehicles.

[0003] Existing fault prediction methods for vehicle structures primarily include threshold alarm models based on stress testing, finite element simulation comparative analysis methods, and traditional machine learning algorithms based on a single data source. These methods suffer from the following shortcomings: First, they rely on static simulation models, making it difficult to reflect in real time the actual stress state of the structure and the dynamic impact of environmental changes on the fatigue life of the material. Second, they lack a multimodal feature fusion mechanism and fail to comprehensively consider the coupling effects of vibration, strain, and environmental factors. Third, they ignore the spatial connectivity between structural components and the damage propagation path, making it difficult to accurately predict the initiation location and evolution trend of cracks in key areas. Fourth, prediction models generally lack physical constraints, resulting in deviations from the actual material strength evolution patterns. Therefore, a fault prediction maintenance method is urgently needed that can integrate multi-source sensory data, introduce physical degradation models, and possess structural topology modeling and dynamic learning capabilities to achieve accurate health assessment of the surrounding structure and optimize maintenance decisions. Summary of the Invention

[0004] In response to the deficiencies of the prior art, the present invention provides a vehicle enclosure structure fault prediction and maintenance method and system based on machine learning, which solves the problems of the above-mentioned background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a vehicle enclosure structure fault prediction and maintenance method based on machine learning, comprising the following steps: S1. Acquire the time-frequency domain characteristics, local strain data and environmental corrosion parameters of the vibration signal of the vehicle enclosure structure, and spatially align them with the pre-stored finite element simulation stress distribution map, quantify the influence weight of environmental factors on the fatigue life of the vehicle enclosure structure material through the physical degradation model, and generate a multimodal feature vector; S2. Construct a dynamic graph topology network based on the physical connection relationship of the vehicle enclosure structure components, use each structural component as a node in the graph, and embed a multimodal feature vector in each node, dynamically adjust the edge weight according to the real-time strain correlation between the nodes, and synchronously capture the vibration signal through the spatiotemporal graph convolution network. S3. Calculate the spatiotemporal fusion feature matrix based on the temporal fluctuation characteristics of the signal and the spatial mechanical coupling relationship between the components; S4. Build an adversarial generative network, use the discriminator to perform confidence assessment on the fault propagation heat map and the real crack data, dynamically adjust the training sample weights to enhance the learning of early weak signals, jointly optimize the remaining life regression model and the failure probability classification model, introduce a physical constraint loss function to ensure that the predicted value conforms to the material strength attenuation law, and finally output a maintenance priority list under the quantile confidence interval.

[0006] Furthermore, the specific process of obtaining the time-frequency domain characteristics, local strain data and environmental corrosion parameters of the vibration signal of the vehicle enclosure structure and spatially aligning them with the pre-stored finite element simulation stress distribution map is as follows: according to the geometric model of the vehicle enclosure structure, the stress distribution map generated by the finite element simulation is mapped to the actual physical coordinate system, and the position coordinates of the vibration sensor, strain gauge and environmental sensor are aligned through the spatial grid matching algorithm; the vibration signal is subjected to short-time Fourier transform to extract the time-frequency domain energy distribution characteristics, and the local strain area is divided according to the spatial layout of the strain gauge. Combined with the spatial distribution of environmental corrosion parameters, the correlation mapping relationship between multi-source data and the simulated stress map is established to generate a spatially aligned multi-dimensional data matrix.

[0007] Furthermore, the physical degradation model is used to quantify the influence weight of environmental factors on the fatigue life of vehicle enclosure structure materials. The specific process of generating a multimodal eigenvector is as follows: based on the spatially aligned multidimensional data matrix, the environmental corrosion parameters, vibration time-frequency domain energy distribution and simulated stress values corresponding to each structural region are extracted; the environmental parameters are input into the physical degradation model, the material surface corrosion rate is calculated through the corrosion kinetics equation, and the equivalent cyclic stress amplitude of the vibration signal is coupled to generate an environment-load joint attenuation factor; the environment-load joint attenuation factor is nonlinearly weighted fused with the simulated stress value to quantify the loss weight of the fatigue life of materials in different regions, and finally a multimodal eigenvector containing physical mechanism constraints is constructed.

[0008] Furthermore, a dynamic graph topology network is constructed, with each structural component as a node in the graph, and a multimodal eigenvector is embedded in each node. The specific process of dynamically adjusting the edge weights based on the real-time strain correlation between the nodes is as follows: the initial graph topology is defined according to the assembly relationship of the vehicle enclosure structure, with the nodes being structural components, including longitudinal beams, cross beams, and suspension brackets, and the edges being physical connection relationships, including welding and bolt connections; a multimodal eigenvector is assigned to each node, containing physical degradation indicators, local strain mean, and vibration spectrum entropy values; the Pearson correlation coefficient of the strain data of adjacent nodes is calculated in real time, and a dynamic edge weight update rule is set: when the correlation coefficient exceeds the threshold, the edge weight is enhanced to characterize the activity of the load transfer path; otherwise, the weight is reduced to weaken invalid connections, thereby realizing adaptive adjustment of the graph topology with the mechanical state.

[0009] Furthermore, the spatiotemporal graph convolutional network is used to synchronously capture the temporal fluctuation characteristics of the vibration signal and the spatial mechanical coupling relationship between components. The specific process of outputting the spatiotemporal fusion feature matrix is as follows: a layered spatiotemporal graph convolution module is designed, and dilated convolution is used to extract the long-period trend characteristics and short-term transient impact characteristics of the vibration signal in the time dimension; in the spatial dimension, the adjacent node features are weighted and aggregated according to the dynamic edge weights to capture the coupling effect of stress transfer between components; the output features of the spatiotemporal convolution are spliced along the channel dimension, and the key feature channels are screened through the attention gating mechanism to generate a spatiotemporal fusion feature matrix. Each row vector in the matrix represents the implicit health state of the corresponding component under the action of spatiotemporal coupling.

[0010] Furthermore, the graph attention mechanism is calculated on the spatiotemporal fusion feature matrix to identify key risk nodes and associated edges. The specific process is as follows: based on the spatiotemporal fusion feature matrix, a multi-head graph attention network is constructed, and the dynamic edge weight is used as the initial bias term of the attention distribution; the corrosion attenuation coefficient and the vibration transient energy ratio are fused in the node feature dimension to generate a corrosion-vibration collaborative attention factor; the adjacent node features are nonlinearly transformed through a learnable parameter matrix, and the normalized attention distribution is calculated in combination with the attention factor. Nodes and associated edges with attention scores higher than the preset threshold are screened, marked as key risk nodes and high-risk connection paths, and the risk topology subgraph is output.

[0011] Furthermore, the propagation path of fatigue damage in the dynamic graph is simulated by the graph diffusion model, and the crack initiation probability in the high-risk area is predicted in combination with the corrosion attenuation coefficient. The specific process of generating the fault propagation heat map is as follows: the key risk node is used as the damage diffusion source, and the diffusion intensity is initialized as the corrosion-vibration collaborative attention factor of the node; the damage transfer efficiency along the connecting edge is calculated according to the dynamic edge weight and the finite element simulation stress gradient, and the corrosion attenuation coefficient is introduced as a diffusion damping term to suppress the damage accumulation in the low-corrosion area; the cumulative damage value of each node is updated by iterative diffusion equation, where the increment is nonlinearly positively correlated with the diffusion intensity, transfer efficiency and current damage value; the cumulative damage value is compared with the material fatigue threshold, and the crack initiation probability is predicted in combination with the corrosion attenuation coefficient to generate a spatial dimension fault propagation heat map, marking the potential crack initiation area and propagation path.

[0012] Furthermore, an adversarial generative network is constructed, and the discriminator is used to perform confidence assessment on the fault propagation heat map and the real crack data. The specific process of dynamically adjusting the training sample weights to enhance the learning of early weak signals is as follows: a generator is constructed to simulate the heat map distribution of normal and early fault states, and the discriminator performs confidence scoring on the generated heat map and the real crack data; the sample weights are dynamically calculated according to the discriminator output, and the fuzzy state samples with confidence below the threshold are weighted and strengthened; the generator and discriminator are optimized through adversarial training, so that the generator gradually approaches the real damage propagation mode, thereby improving the model's sensitivity to weak features.

[0013] Furthermore, the remaining life regression model and the failure probability classification model are jointly optimized, and a physical constraint loss function is introduced to ensure that the predicted value conforms to the law of material strength attenuation. The specific process of finally outputting the maintenance priority list under the quantile confidence interval is as follows: a dual-branch network is designed to output the remaining life regression value and the failure probability classification result respectively, and the collaborative update of parameters is achieved through the shared feature layer; a physical constraint term is introduced into the loss function to force the remaining life prediction value to monotonically decrease with the increase of the corrosion attenuation coefficient and the vibration equivalent stress amplitude; the life intervals under different confidence levels are calculated through quantile regression, and the high-risk components are ranked according to the failure probability classification results to generate a list graded by maintenance urgency.

[0014] The vehicle enclosure structure fault prediction and maintenance system based on machine learning includes the following modules: a multimodal fusion module, a dynamic graph network module, a damage propagation module, and a maintenance optimization module; the multimodal fusion module is used to obtain the time-frequency domain characteristics, local strain data and environmental corrosion parameters of the vibration signal of the vehicle enclosure structure, and spatially align them with the pre-stored finite element simulation stress distribution map, and quantify the influence weight of environmental factors on the fatigue life of the vehicle enclosure structure material through the physical degradation model; the dynamic graph network module is used to construct a dynamic graph topology network according to the physical connection relationship of the vehicle enclosure structure components, take each structural component as a node in the graph, and embed a multimodal feature vector in each node, dynamically adjust the edge weight according to the real-time strain correlation between the nodes, and synchronously capture the vibration signal through the spatiotemporal graph convolution network. The temporal fluctuation characteristics of the components are coupled with the spatial mechanics of the components to output a spatiotemporal fusion feature matrix; the damage propagation module is used to perform graph attention mechanism calculation on the spatiotemporal fusion feature matrix, identify key risk nodes and associated edges, simulate the propagation path of fatigue damage in the dynamic graph through the graph diffusion model, and predict the crack initiation probability in the high-risk area in combination with the corrosion attenuation coefficient to generate a fault propagation heat map; the maintenance optimization module is used to construct an adversarial generative network, perform confidence assessment on the fault propagation heat map and the real crack data through the discriminator, dynamically adjust the training sample weights to enhance the learning of early weak signals, jointly optimize the remaining life regression model and the failure probability classification model, introduce a physical constraint loss function to ensure that the predicted value conforms to the material strength attenuation law, and finally output a maintenance priority list under the quantile confidence interval.

[0015] The present invention has the following beneficial effects:

[0016] (1) A machine learning-based vehicle enclosure structure fault prediction and maintenance method quantifies the environment-load synergistic damage effect through multimodal data fusion and physical degradation models, combines dynamic graph topology networks to capture the mechanical coupling relationship between components in real time, uses spatiotemporal graph convolution and graph diffusion models to simulate damage propagation paths, and enhances early weak signal learning through adversarial generative networks to achieve high-precision prediction of vehicle enclosure structure faults. This method breaks through the limitations of traditional single sensors or static models, embeds the corrosion attenuation coefficient into the damage accumulation calculation, generates probabilistic thermal maps and quantile maintenance lists, reduces false alarm rates and maintenance costs, and supports damage source location and propagation path visualization, significantly improving the interpretability and engineering applicability of maintenance decisions.

[0017] (2) The vehicle enclosure structure fault prediction and maintenance system based on machine learning realizes the intelligentization of the entire process from data collection to maintenance decision-making through the closed-loop design of multimodal fusion module, dynamic graph network module, damage propagation module and maintenance optimization module. Based on existing sensor data and finite element simulation priors, the dynamic graph network module updates the edge weights in real time to adapt to changes in vehicle operating conditions. The damage propagation module quantifies the probability of crack initiation through graph attention and diffusion models. The maintenance optimization module combines the adversarial generative network with the physical constraint loss function to output a confidence interval-driven hierarchical list. The system does not require hardware modification and is compatible with on-board edge computing devices. It successfully warned of hidden cracks and discovered risks in advance in actual logistics vehicle measurements, reducing unplanned downtime. It has the core advantages of high precision, low cost and strong scalability.

[0018] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the vehicle enclosure structure fault prediction and maintenance method based on machine learning of the present invention.

[0020] Figure 2 This is a flow chart of the vehicle enclosure structure fault prediction and maintenance system based on machine learning of the present invention. DETAILED DESCRIPTION

[0021] The embodiment of the present application solves the technical bottlenecks of traditional methods in multi-physics field coupling modeling, damage propagation path prediction and early weak signal detection through a vehicle enclosure structure fault prediction and maintenance method and system based on machine learning. By integrating multi-source data such as vibration, strain, and environmental corrosion with finite element simulation stress maps, a physical degradation model is constructed to quantify the environment-load collaborative damage effect; based on a dynamic graph topology network, the mechanical coupling relationship between components is captured in real time, and the damage propagation path is simulated by combining spatiotemporal graph convolution and graph diffusion models; an adversarial generative network is introduced to optimize sample weights, strengthen early crack signal learning, and ensure that the prediction results comply with the material attenuation law through a physical constraint loss function. Finally, a probabilistic fault propagation heat map and a quantile maintenance list are generated to achieve high-precision early warning and low-cost maintenance.

[0022] The overall idea of the solution in the embodiments of this application is as follows:

[0023] Through multimodal data fusion, dynamic graph network construction, and adversarial optimization algorithms, accurate prediction of vehicle enclosure structure failures and maintenance decisions are achieved. Specifically, vibration, strain, and environmental corrosion data are spatially aligned with finite element simulation stress maps to generate physically constrained multimodal feature vectors.

[0024] A dynamic graph topology network is constructed based on the physical connection relationship of components, and spatiotemporal graph convolution is used to simultaneously extract spatiotemporal features and simulate damage propagation paths;

[0025] Combining the corrosion attenuation coefficient with the graph diffusion model to quantify the crack probability in high-risk areas and generate a visual heat map;

[0026] Finally, early signal learning is enhanced through adversarial generative networks, and a quantile maintenance list is output in conjunction with the physical constraint loss function, forming a closed-loop solution from data fusion, damage inference to decision output.

[0027] See also Figure 1, an embodiment of the present invention provides a technical solution: a vehicle enclosure structure fault prediction and maintenance method based on machine learning, comprising the following steps: S1. obtaining the time-frequency domain characteristics, local strain data and environmental corrosion parameters of the vibration signal of the vehicle enclosure structure, and spatially aligning them with the pre-stored finite element simulation stress distribution map, quantifying the influence weight of environmental factors on the fatigue life of the vehicle enclosure structure material through the physical degradation model, and generating a multimodal feature vector; S2. constructing a dynamic graph topology network based on the physical connection relationship of the vehicle enclosure structure components, taking each structural component as a node in the graph, and embedding a multimodal feature vector in each node, dynamically adjusting the edge weight according to the real-time strain correlation between the nodes, and synchronously capturing the time series of the vibration signal through the spatiotemporal graph convolution network. The spatial mechanical coupling relationship between the fluctuation characteristics and the components is analyzed to output the spatiotemporal fusion feature matrix; S3. The graph attention mechanism is calculated on the spatiotemporal fusion feature matrix to identify key risk nodes and associated edges. The propagation path of fatigue damage in the dynamic graph is simulated through the graph diffusion model. The crack initiation probability in high-risk areas is predicted in combination with the corrosion attenuation coefficient to generate a fault propagation heat map; S4. A generative adversarial network is constructed to perform confidence assessment on the fault propagation heat map and the real crack data through the discriminator, and the training sample weights are dynamically adjusted to enhance the learning of early weak signals. The remaining life regression model and the failure probability classification model are jointly optimized. The physical constraint loss function is introduced to ensure that the predicted value conforms to the material strength attenuation law, and finally a maintenance priority list under the quantile confidence interval is output.

[0028] In this embodiment, step S1: This step is mainly used for data collection and feature extraction. The health status of the vehicle enclosure structure is affected by many factors, including vibration, strain, environmental corrosion and other factors. By obtaining these data, the dynamic state of the structure can be fully captured, providing accurate input data for the subsequent fault prediction model. Time-frequency domain characteristics of vibration signals: Through time-frequency analysis (such as short-time Fourier transform), the frequency characteristics of the structural vibration signal are extracted. This feature helps to identify key information such as the resonant frequency and damage location of the structure. Local strain data: Strain sensors are arranged in the vehicle enclosure structure to monitor the strain conditions of different parts in real time. This helps to identify the stress changes of components and determine whether there are cracks or fatigue damage. Environmental corrosion parameters: Environmental corrosion (humidity, salt spray) is a key factor affecting the life of materials. By monitoring these environmental factors, the impact of corrosion on the fatigue life of materials can be quantified. Finite element simulation stress distribution map: Through the finite element analysis (FEA) method, the stress distribution of the vehicle enclosure structure under various loads is simulated in the computer. This map can be used to compare with actual data to improve the accuracy of fault prediction. Multimodal feature vectors: Combine multiple data types, such as vibration signals, strain data, corrosion parameters, and stress distribution, to form a comprehensive feature vector. These features can be used in subsequent machine learning models. Step S2: This step simulates the physical connections and dynamic changes between components in the vehicle's surrounding structure by constructing a dynamic graph topology network. The spatiotemporal graph convolutional network (ST-GCN) not only captures the temporal characteristics of the vibration signal but also accurately captures the mechanical coupling relationships between components, thereby enabling efficient prediction of the structural health status. A dynamic graph topology network is a graph structure in which each node represents a structural component, and the edges between nodes represent the physical connections between components. The topology is "dynamic" because the vehicle's surrounding structure changes over time (e.g., changes in component status). The spatiotemporal graph convolutional network (ST-GCN) is a graph neural network model that can simultaneously process spatial features (connectivity between components) and time series data (vibration signal changes over time) in graph data. This network learns the features of each node and edge in the graph through convolution operations, effectively capturing spatiotemporal dynamics. Spatiotemporal fusion feature matrix: The output matrix processed by the spatiotemporal graph convolutional network integrates the temporal and spatial features of each node, and can effectively characterize the health status of the vehicle surrounding structure. Step S3: This step automatically identifies key risk components in the structure through the graph attention mechanism, and simulates the propagation of damage through the graph diffusion model to help predict which areas may have cracks. The generated fault propagation heat map can be used to guide maintenance decisions and identify high-risk areas that require the most attention. Graph Attention Mechanism (GAT): By assigning different weights to different edges and nodes in the graph, it automatically focuses on the most relevant parts. This mechanism can enhance the model's learning ability for important nodes (such as severely damaged components).Graph Diffusion Model: Used to simulate the propagation of damage from one component to other components. For example, fatigue damage may spread from one component to connected components, a process that can be predicted using a diffusion model. Corrosion Attenuation Coefficient: This coefficient takes into account the impact of corrosion on material strength and quantifies the attenuation effect of corrosion on the fatigue life of the structure. Fault Propagation Heatmap: This heatmap displays the fault propagation risk of different locations in the structure, with the color depth in the heatmap representing the likelihood of crack initiation. Step S4: This step uses a Generative Adversarial Network (GAN) and a physically constrained optimization model to enhance learning of weak fault signals, improve model accuracy, and ultimately output a maintenance priority list based on the predicted maintenance priorities. This step provides a scientific basis for maintenance decisions, ensuring that maintenance work is targeted at high-risk areas. Generative Adversarial Network (GAN): A network consisting of a generator and a discriminator. The generator generates a fault propagation heatmap, while the discriminator determines whether the generated heatmap is consistent with actual crack data. Through adversarial training, the two mutually improve model performance. Discriminator: In the GAN, the discriminator determines the degree of match between the generated heatmap and actual crack data. Through feedback, the discriminator helps the generator improve the generated heat map. Training sample weight adjustment: During the training process, the weights of different training samples are dynamically adjusted so that the model pays more attention to weak, difficult-to-detect early fault signals. Remaining life regression model: used to predict the remaining service life of structural components and help decide which components need maintenance as soon as possible. Failure probability classification model: used to predict the probability of component failure and guide maintenance decisions. Physical constraint loss function: The laws of physics are added during model training, such as the attenuation law of material strength, to ensure that the model output meets actual physical limitations. Quantile confidence interval: Based on the results of the model output, the fault prediction range at different confidence levels is given. Use quantiles to evaluate maintenance priorities and risks.

[0029] Specifically, the specific process of obtaining the time-frequency domain characteristics, local strain data and environmental corrosion parameters of the vibration signal of the vehicle enclosure structure and spatially aligning them with the pre-stored finite element simulation stress distribution map is as follows: according to the geometric model of the vehicle enclosure structure, the stress distribution map generated by the finite element simulation is mapped to the actual physical coordinate system, and the position coordinates of the vibration sensor, strain gauge and environmental sensor are aligned through the spatial grid matching algorithm; the vibration signal is subjected to short-time Fourier transform to extract the time-frequency domain energy distribution characteristics, and the local strain area is divided according to the spatial layout of the strain gauge. Combined with the spatial distribution of environmental corrosion parameters, the correlation mapping relationship between multi-source data and the simulation stress map is established to generate a spatially aligned multi-dimensional data matrix.

[0030] In this embodiment, step 1: based on the geometric model of the vehicle surrounding structure, the stress distribution map generated by the finite element simulation is mapped to the actual physical coordinate system. Geometric model mapping: based on the actual geometric model of the vehicle surrounding structure, the stress distribution map obtained by the finite element simulation is mapped to the actual physical coordinate system. The purpose of this step is to align the simulated stress map with the physical coordinates of the actual measurement point to ensure that the simulation data can be effectively compared and integrated with the data of the real environment. Spatial grid matching: a spatial grid matching algorithm is used to align the position coordinates of the vibration sensor, strain gauge and environmental sensor. Through this algorithm, the sampling point corresponding to each sensor can be ensured to be accurately located in the finite element simulation map, so that multi-source data can be effectively connected. Spatial grid matching algorithm: through grid subdivision, the spatial distance between each data point and the sensor point is calculated and matched within a certain error range. This ensures that the sensor data can be synchronized with the simulation results. Step 2: Short-time Fourier transform is performed on the vibration signal to extract the time-frequency domain energy distribution characteristics. Short-time Fourier transform: a short-time Fourier transform (STFT) is performed on the vibration signal to extract the time-frequency domain energy distribution characteristics. The Short-Time Fourier Transform (STFT) transforms signals from the time domain to the frequency domain. Due to its sliding window properties, it can simultaneously capture the signal's time-varying characteristics. Time-Frequency Domain Characteristics: The STFT generates the signal's time-frequency spectrum, extracting the energy distribution of each frequency component over different time periods. These energy distribution characteristics reflect the vibration behavior of the structure and are important for assessing fatigue damage and changes in material properties. Step 3: Divide the local strain regions based on the spatial layout of the strain gauges. Spatial Layout Division: Based on the strain gauge layout, appropriate spatial regions are selected for local strain analysis. Strain gauges are typically installed at key locations within the vehicle's enveloping structure to monitor material stress. Local Strain Regions: Multiple local strain regions are identified based on the strain gauge location and the structure's geometric characteristics. The strain data from these regions can be used to further analyze the material fatigue condition in the local region and predict the potential location of fatigue failure. Strain Data Acquisition: Strain gauges are used to collect strain data from the vehicle's enveloping structure under different loading conditions. Regional Division: The stress conditions of the local region are determined using the finite element model, and the local strain data is correlated with the simulated stress spectrum. Step 4: Combine the spatial distribution of environmental corrosion parameters to establish a correlation mapping relationship between multi-source data and simulated stress maps. Environmental corrosion parameters: Environmental corrosion parameters (temperature, humidity, salt spray) affect the corrosion rate of materials, thereby affecting the fatigue life of materials. By measuring environmental corrosion parameters, the impact of the environment on structural performance can be obtained. Combination of spatial distribution: By fusing the spatial distribution of environmental corrosion parameters with strain data and simulated stress maps, a comprehensive spatial correlation mapping relationship is established. This relationship can reflect the combined impact of environmental factors, vibration loads, and stress on material fatigue life.Spatial data fusion: Through multi-source data fusion technology, environmental corrosion parameters, vibration time-frequency domain characteristics, local strain data and simulation stress maps are combined to form a comprehensive spatial data matrix. Step 5: Generate a spatially aligned multi-dimensional data matrix. Multi-dimensional data matrix: Through the above steps, a spatially aligned multi-dimensional data matrix is finally formed. The matrix contains a variety of data such as the time-frequency domain energy distribution of the vibration signal, strain data and environmental corrosion parameters. These data can be used for further fatigue analysis, life prediction, etc. Application of data matrix: The data matrix can be used as an input data source and applied to machine learning models or physics-based fatigue analysis models for material life prediction and structural health monitoring. Data matrix construction: By aligning and standardizing multi-source data, a multi-dimensional data matrix in a unified format is generated.

[0031] Specifically, the physical degradation model is used to quantify the influence weight of environmental factors on the fatigue life of vehicle enclosure structure materials. The specific process of generating a multimodal eigenvector is as follows: based on the spatially aligned multidimensional data matrix, the environmental corrosion parameters, vibration time-frequency domain energy distribution and simulated stress values corresponding to each structural area are extracted; the environmental parameters are input into the physical degradation model, the material surface corrosion rate is calculated through the corrosion kinetics equation, and the equivalent cyclic stress amplitude of the vibration signal is coupled to generate an environment-load joint attenuation factor; the environment-load joint attenuation factor is nonlinearly weighted fused with the simulated stress value to quantify the loss weight of the fatigue life of materials in different regions, and finally a multimodal eigenvector containing physical mechanism constraints is constructed.

[0032] In this implementation scheme, step 1: extract the environmental corrosion parameters, vibration time-frequency domain energy distribution and simulation stress values of each structural area. Environmental corrosion parameters: through the multi-dimensional data matrix established above, extract the environmental corrosion parameters of each structural area, such as temperature, humidity, and salt spray. These factors will affect the corrosion rate of the material surface, thereby affecting the fatigue life. Vibration time-frequency domain energy distribution: through the aforementioned time-frequency domain analysis, extract the time-frequency domain energy distribution of the vibration signal of each area, reflecting the vibration characteristics of the structure at different frequencies during operation. These characteristics are closely related to the fatigue life of the material. Simulation stress value: Based on finite element analysis, extract the stress value of each structural part. These stress values can reflect the stress distribution of the material under actual working conditions and have a direct impact on fatigue damage. Step 2: Input the environmental parameters into the physical degradation model to calculate the corrosion rate. Corrosion kinetics equation: Environmental factors such as temperature, humidity, and salt spray can describe the corrosion rate of the material through the corrosion kinetics equation. The corrosion rate is a key parameter for material aging and degradation, and is crucial to the prediction of fatigue life. The formula is as follows: in: Represents the corrosion rate (unit is corrosion amount / time). k(T, H, P) is a corrosion kinetic constant related to temperature, humidity and pollutant concentration (T, H, P). f(C) is a function of the relationship between material type and environment, which represents the interaction between material and environmental factors. Through this formula, the corrosion rate of the material under specific environmental conditions can be calculated. Equivalent cyclic stress amplitude of coupled vibration signal: The fatigue life of the material under vibration load is directly related to the magnitude of the stress amplitude. By analyzing the vibration signal, the cyclic amplitude of the vibration stress can be obtained. This amplitude will be combined with the corrosion rate to form an environment-load joint attenuation factor. Step 3: Generate environment-load joint attenuation factor: Environment-load joint attenuation factor: This factor integrates the impact of environmental corrosion on the material and the stress impact of external loads (such as vibration signals) on the material. The joint attenuation factor represents the joint influence of environmental factors and loads on the fatigue life of the material, and is a key parameter for fatigue life prediction. The formula is as follows: Where: λ represents the environment-load joint attenuation factor. α is a weighting factor related to the corrosion rate, which represents the effect of the corrosion rate on the fatigue life of the material. is the corrosion rate obtained from the corrosion kinetics equation. β is a weighting factor related to the stress amplitude, which represents the effect of vibration load on the fatigue life of the material. σ eq is the equivalent cyclic stress amplitude extracted from the vibration signal. Step 4: Perform a nonlinear weighted fusion of the environmental load joint attenuation factor and the simulated stress value: Nonlinear weighted fusion: By performing a nonlinear weighted fusion of the environmental load joint attenuation factor and the simulated stress value, the fatigue life loss of the material in different regions can be quantified. This fusion method considers the combined effects of factors such as stress in different regions and environmental corrosion on fatigue life. The formula is as follows: Where: total represents the comprehensive fatigue loss factor. λ is the environmental load combined attenuation factor. σ sim is the stress value obtained from the finite element simulation. γ1 and γ2 are parameters for adjusting the attenuation factor and the stress value weight. η is a nonlinear exponent used to adjust the influence of the stress value on fatigue loss. Through this process, the fatigue loss factor of each structural area is obtained as the input for subsequent fatigue life prediction. Step 5: Construct a multimodal eigenvector containing physical mechanism constraints: Multimodal eigenvector: The multimodal eigenvector is a comprehensive feature representation constructed by integrating multiple types of data (such as environmental parameters, vibration signals, strain data and simulated stress values). These features can comprehensively reflect the fatigue state of the structure. Physical mechanism constraints: In the process of constructing the eigenvector, physical mechanism constraints (such as the strength attenuation law of the material, the relationship between corrosion and load, etc.) are introduced to ensure that the generated eigenvector is consistent with the actual physical phenomenon. The formula is expressed as follows: X = [X env , Xvib , X stress , X corr ]Where: X represents the final generated multimodal feature vector. X env 、X vib 、X stress 、X corr They represent the characteristic vectors corresponding to environmental parameters, vibration time-frequency domain characteristics, stress values and corrosion factors respectively.

[0033] Specifically, a dynamic graph topology network is constructed, with each structural component as a node in the graph, and a multimodal eigenvector is embedded in each node. The specific process of dynamically adjusting the edge weights based on the real-time strain correlation between the nodes is as follows: the initial graph topology is defined according to the assembly relationship of the vehicle enclosure structure, where the nodes are structural components, including longitudinal beams, cross beams, and suspension brackets, and the edges are physical connection relationships, including welding and bolt connections; a multimodal eigenvector is assigned to each node, containing physical degradation indicators, local strain mean, and vibration spectrum entropy values; the Pearson correlation coefficient of the strain data of adjacent nodes is calculated in real time, and a dynamic edge weight update rule is set: when the correlation coefficient exceeds the threshold, the edge weight is enhanced to characterize the activity of the load transfer path; otherwise, the weight is reduced to weaken invalid connections, thereby realizing adaptive adjustment of the graph topology with the mechanical state.

[0034] In this embodiment, the initial graph topology is defined as follows: Node definition: Based on the assembly relationship of the vehicle's surrounding structure, each structural component, such as a longitudinal beam, a cross beam, a suspension bracket, etc., is regarded as a node in the graph. These nodes represent different components of the vehicle structure and have their own characteristics and behaviors. Edge definition: Edges represent the physical connection relationship between nodes, including welding, bolting, etc. These edges represent the structural connection between components, reflecting how the components interact and transfer loads through physical connections. Assigning multimodal eigenvectors: Each node (i.e., structural component) is assigned a multimodal eigenvector, which includes the following parts: Physical degradation index: reflects the degradation state of the material or structure, such as degradation caused by environmental corrosion, fatigue, etc. Local strain mean: represents the strain distribution on a certain node or structural component, reflecting the deformation of the component after being subjected to force. Vibration spectrum entropy: reflects the complexity and energy distribution of the structure during vibration. The higher the entropy value, the more complex the structural vibration mode, which may be related to potential faults or abnormalities. Real-time calculation of strain correlation of adjacent nodes: For adjacent nodes (i.e., directly connected structural components), the Pearson correlation coefficient between their strain data is calculated in real time. The Pearson correlation coefficient is used to measure the linear relationship between the strain data of two nodes: if the strain data of adjacent nodes have a strong correlation, it means that there may be a strong load transfer relationship between these components. The Pearson correlation coefficient formula is as follows: Among them, x and y are the strain data of the two nodes respectively, and n is the number of data points. Dynamic edge weight update rule: Edge weight enhancement and reduction: According to the calculated Pearson correlation coefficient, the dynamic edge weight update rule is set: When the Pearson correlation coefficient of adjacent nodes exceeds the set threshold, it means that there is a strong load transfer correlation between the two nodes, and the edge weight is enhanced to characterize the activity of the path. On the contrary, when the correlation coefficient is low, it means that the load transfer between the two nodes is weak, and the edge weight is reduced, weakening its role in the graph. Achieve adaptive adjustment of graph topology: Through the above-mentioned dynamic edge weight update mechanism, the graph topology can be adaptively adjusted according to the actual mechanical state of the structure. When the load transfer relationship between structural components changes (for example, a component is deformed or damaged), the graph topology and edge weights will promptly reflect these changes, so that the topology of the entire vehicle enclosure structure more accurately represents the current mechanical state.

[0035] Specifically, the spatiotemporal graph convolutional network is used to synchronously capture the temporal fluctuation characteristics of the vibration signal and the spatial mechanical coupling relationship between components. The specific process of outputting the spatiotemporal fusion feature matrix is as follows: a layered spatiotemporal graph convolution module is designed, and dilated convolution is used to extract the long-period trend characteristics and short-term transient impact characteristics of the vibration signal in the time dimension; in the spatial dimension, the adjacent node features are weighted and aggregated according to the dynamic edge weights to capture the coupling effect of stress transfer between components; the output features of the spatiotemporal convolution are spliced along the channel dimension, and the key feature channels are screened through the attention gating mechanism to generate a spatiotemporal fusion feature matrix. Each row vector in the matrix represents the implicit health status of the corresponding component under the action of spatiotemporal coupling.

[0036] In this embodiment, the process of synchronously capturing the temporal fluctuation characteristics of the vibration signal and the spatial mechanical coupling relationship between components through the spatiotemporal graph convolution network and outputting the spatiotemporal fusion feature matrix is as follows: Designing a layered spatiotemporal graph convolution module: The design goal of the spatiotemporal graph convolution network is to capture the temporal fluctuation characteristics of the vibration signal and the spatial mechanical coupling relationship between components. By extending the convolution operation to the spatiotemporal domain, the network can simultaneously capture the dynamic changes of the signal in time and the interaction of structural components in space. = Dilated convolution in the time dimension: In the time dimension, dilated convolution is used to extract the long-period trend characteristics and short-term transient impact characteristics of the vibration signal. Dilated convolution increases the receptive field of the convolution kernel by introducing a dilation factor, thereby capturing features of a longer time span in fewer convolution layers. Dilated convolution can not only capture the low-frequency trend (long period) of the vibration signal, but also capture high-frequency transient impacts (short-term transients). The specific operation of dilated convolution: For each moment t of the signal, the dilated convolution operation is to convolve the signal and select time points with larger intervals in the time domain for processing. The formula is as follows: Where: y(t) is the output signal after convolution, x(t) is the input signal, w(k) is the weight of the convolution kernel, r is the expansion factor (controls the size of the receptive field), and K is the size of the convolution kernel. Weighted aggregation in the spatial dimension: In the spatial dimension, the features of adjacent nodes are weightedly aggregated according to the dynamic edge weights. Each node represents a structural component, and the characteristics of each component are determined by its vibration signal and mechanical response. Since the stress transfer between structural components has spatial coupling, the features of adjacent nodes should be aggregated according to certain weights. The process of weighted aggregation: For each node v, its aggregated feature vector h v ' can be expressed by the following formula: v '=∑ u∈N ( v (α vu Wh u ; Among them: h v ' is the aggregate feature of node v, N(v( is the set of adjacent nodes of node v, α vu is the dynamic edge weight between node v and node u, W is the learnable weight matrix, h u is the feature of node u. Splicing of spatiotemporal convolution features and attention gating mechanism: The features output by spatiotemporal convolution are spliced along the channel dimension to form a comprehensive feature matrix. This matrix contains the temporal features of the vibration signal and the spatial features between components, and can comprehensively characterize the spatiotemporal coupling of the system. The attention gating mechanism is used to screen out key feature channels. By assigning different weights to feature channels, the attention mechanism can automatically focus on the feature channels that contribute most to the prediction of health status. The specific operation is to select and weight feature channels through the learned attention weights to obtain a more focused feature representation. Mathematical expression of the attention mechanism: For each feature channel c, the attention score a c The calculation is: Where: q is the query vector, indicating the importance of the feature channel, W c is the learnable weight matrix, h cis the eigenvector of feature channel c, and C is the set of all feature channels. Generating a spatiotemporal fusion feature matrix: Through the above steps, the resulting spatiotemporal fusion feature matrix contains the implicit health status of each component under spatiotemporal coupling. Each row vector of the matrix corresponds to the health status of a component, while the column vectors represent the vibration and mechanical characteristics of the component at different temporal and spatial locations. This feature matrix can serve as input for subsequent tasks such as health prediction and fault diagnosis, helping to analyze the health status of the structure and providing a basis for decision-making. By designing a layered spatiotemporal graph convolution module, we can simultaneously capture the temporal fluctuation characteristics of the vibration signal and the spatial mechanical coupling between structural components. Dilated convolution extracts the long-period and short-duration impact characteristics of the signal, while dynamic edge weights effectively reflect the stress transfer between components. Combining an attention mechanism, weighted aggregation, and feature screening, the generated spatiotemporal fusion feature matrix provides the health status of each component under spatiotemporal coupling, providing effective technical support for health monitoring and fault prediction.

[0037] Specifically, the graph attention mechanism is calculated on the spatiotemporal fusion feature matrix to identify key risk nodes and associated edges. The specific process is as follows: based on the spatiotemporal fusion feature matrix, a multi-head graph attention network is constructed, and the dynamic edge weight is used as the initial bias term of the attention distribution; the corrosion attenuation coefficient and the vibration transient energy ratio are fused in the node feature dimension to generate a corrosion-vibration collaborative attention factor; the adjacent node features are nonlinearly transformed through a learnable parameter matrix, and the normalized attention distribution is calculated in combination with the attention factor. Nodes and associated edges with attention scores higher than the preset threshold are screened, marked as key risk nodes and high-risk connection paths, and the risk topology subgraph is output.

[0038] In this implementation, a multi-head graph attention network is constructed. This is a mechanism based on graph neural networks (GNNs). It processes graph data by computing multiple attention heads in parallel, enhancing the model's learning ability and robustness. Each attention head independently learns different feature representations, and ultimately, by aggregating the outputs of multiple heads, it captures more complex relationships. In this network, dynamic edge weights serve as an initial bias for the attention distribution, providing preliminary direction for the allocation of attention weights between adjacent nodes. Larger edge weights indicate stronger connections between nodes, and the model will pay more attention to these node pairs. Node feature fusion: Corrosion attenuation coefficient and vibration transient energy fraction: Node feature fusion combines features from different sources to generate a more comprehensive node representation. Here, node features are enhanced using two features: the corrosion attenuation coefficient and the vibration transient energy fraction. The corrosion attenuation coefficient (δcorrosion) reflects the material performance degradation caused by corrosion damage and is typically calculated from historical environmental data, corrosion models, and other factors. The vibration transient energy fraction (ηvibration) represents the short-term energy characteristics of the vibration signal within a specific time window. This feature can reflect the transient response of the structure under a specific load. The fusion of the two features is expressed as the total feature of the node, generating the corrosion-vibration collaborative attention factor (αcorrosion-vibration), which is used to adjust the influence of the node in the graph. The formula is as follows: α corrosion-vibration =f(δ corrosion , η vibration ); where: f(·) is the fusion function, which can be a weighted sum, concatenation or other nonlinear functions. Nonlinear transformation and attention factor calculation: Nonlinear transformation is a common operation in graph neural networks. Its purpose is to map and increase the dimension of node features to make node features richer and more expressive. In this process, a learnable parameter matrix (W node ) transforms the node features. After nonlinear transformation, the feature hv of each node can be expressed as: h' v =σ(W node h v ); where: h' v is the transformed node feature, σ(·) is the activation function (Sigmoid), W node is the learnable transformation weight matrix, h v is the original node feature. Calculate the normalized attention distribution: The attention distribution is calculated based on the similarity of node features and the correlation between nodes. For each pair of adjacent nodes v and u, use their transformed features and the erosion-vibration collaborative attention factor α corrosion-vibration To calculate its attention weight α vuThe normalized attention distribution is normalized using the softmax function to ensure that the sum of all attention scores is 1. The formula is: Where: α vu : This is the attention weight between node v and node u. It reflects the importance relationship or correlation between node v and node u. The larger this value is, the closer the relationship between node v and node u is, and the stronger the attention is. a: This is a learnable parameter used to calculate attention. This parameter can be optimized through training to capture the correlation characteristics between nodes. It is a vector, usually represented as Where d is the dimension of the feature. v ∥Wh u ]: This is the concatenation of the features of node v and node u. W is a learnable weight matrix used to perform a linear transformation on the node features; h v and h u are the feature vectors of node v and node u. After the features of the nodes are transformed by the matrix W, they are concatenated (i.e., concatenated along the feature dimension, represented as ∥). LeakyReLU(·): This is a LeakyReLU activation function. The function is to perform nonlinear activation on the transformed features and introduce a small negative slope. The specific form is: LeakyReLU(x) = max(α, x(; where α is a very small constant (e.g., 0.01) to prevent the gradient from disappearing. α corrosion-vibration : This is the corrosion-vibration collaborative attention factor, which reflects the correlation between node v and node u under the corrosion and vibration characteristics. It combines the corrosion attenuation coefficient and the vibration energy ratio to indicate the degree of influence between nodes under these factors. N(v(): This is the neighbor set of node v. That is, the set of all nodes directly connected to node v. For each node v, its neighbor node set N(v( includes all nodes connected to it. Filtering key nodes and high-risk edges: Based on the calculated attention weights, nodes and their associated edges with attention scores above the preset threshold are filtered out. These nodes and edges are considered to have strong correlations in the graph and should be given special attention in structural health management. Key risk nodes are those whose associated characteristics (such as corrosion level, vibration characteristics, etc.) and strong correlations with surrounding nodes significantly affect the overall structural health. High-risk connection paths are component connection paths that may cause systemic failures under strong coupling. Output risk topology subgraph: By filtering key risk nodes and high-risk connection paths, a risk topology subgraph is ultimately obtained. This subgraph only contains nodes and edges that may have failure risks, which can provide engineers with clear high-risk areas for subsequent maintenance and preventive measures.

[0039] Specifically, the propagation path of fatigue damage in the dynamic graph is simulated by the graph diffusion model, and the crack initiation probability in the high-risk area is predicted in combination with the corrosion attenuation coefficient. The specific process of generating the fault propagation heat map is as follows: the key risk node is taken as the damage diffusion source, and the diffusion intensity is initialized as the corrosion-vibration collaborative attention factor of the node; the damage transfer efficiency along the connecting edge is calculated according to the dynamic edge weight and the finite element simulation stress gradient, and the corrosion attenuation coefficient is introduced as a diffusion damping term to suppress the damage accumulation in the low-corrosion area; the cumulative damage value of each node is updated through the iterative diffusion equation, where the increment is nonlinearly positively correlated with the diffusion intensity, transfer efficiency and current damage value; the cumulative damage value is compared with the material fatigue threshold, and the crack initiation probability is predicted in combination with the corrosion attenuation coefficient to generate a spatial dimension fault propagation heat map, marking the potential crack initiation area and propagation path.

[0040] In this implementation, the graph diffusion model is used to simulate the propagation path of fatigue damage in the dynamic graph, and the corrosion attenuation coefficient is combined to predict the crack initiation probability in high-risk areas to generate the specific process of the fault propagation heat map: Initialization diffusion intensity: In the graph diffusion model, the key risk nodes are the source nodes of damage diffusion. Usually, these nodes represent components in the structure that have been highly corroded or affected by vibration. The initial diffusion intensity depends on the corrosion-vibration synergistic attention factor (α corrosion-vibration ), which is calculated based on the corrosion attenuation coefficient and vibration energy ratio. This factor determines the initial damage propagation capability of the node. The formula is: S v (0) = α corrosion-vibration (v); where: S v (0 is the initial diffusion strength of node v, α corrosion-vibration (v) is the corrosive vibration cooperative attention factor corresponding to node v. Calculating damage transfer efficiency: Damage transfer efficiency refers to the ability of damage to propagate from one node to adjacent nodes along the connecting edge. It depends on the dynamic edge weight (W e ) and finite element simulation stress gradient Among them: the dynamic edge weight reflects the strength of the connection between nodes, for example, the contact strength or connection mode between two parts; the stress gradient reflects the stress difference between nodes, which is usually obtained through finite element simulation results. The efficiency of damage propagation along the edge e (T e ) can be expressed as: Where: T e is the efficiency of transmission along edge e, W e is the weight of edge e, is the stress gradient. The corrosion attenuation coefficient is introduced as the diffusion damping term: the corrosion attenuation coefficient (δ corrosion(v)) is used to suppress the accumulation of damage in low-corrosion areas. During the damage diffusion process, for areas with less corrosion, the corrosion attenuation coefficient will reduce the accumulated amount of damage, thereby reducing the diffusion capacity of the area. Therefore, the diffusion damping term of each node is related to the corrosion attenuation coefficient, and the dynamic equation of damage transfer needs to introduce this coefficient for correction. The iterative diffusion equation updates the cumulative damage value of each node: In the graph diffusion model, the cumulative damage value (Dv(t)) represents the damage state of node v at time step t. The diffusion equation is usually expressed as: D v (t+1)=D v (t)+λ·∑ u∈N ( v )T vu ·S u (t)·(1-δ corrosion )v)); where: D v (t) is the damage value of node v at time step t, T vu is the damage transfer efficiency from node u to node v, Su(t) is the diffusion intensity of node u at time step t, λ is the diffusion coefficient, which controls the speed of damage diffusion, and δ corrosion (v) is the corrosion attenuation coefficient of node v. This equation describes the damage growth of each node in one time step, which depends on the damage of adjacent nodes and the damage transfer efficiency. Comparison of damage with material fatigue threshold predicts the probability of crack initiation: Material fatigue threshold (σ fatigue ) refers to the maximum damage value of the material under long-term repeated stress without crack initiation. Generally, if the damage value exceeds the fatigue threshold, cracks may begin to initiate. At each node, if the cumulative damage value D v (t) exceeds the fatigue threshold σfatigue, crack initiation may occur. Crack initiation probability P crack (v) can be calculated by the following formula: Otherwise, P crack (v) = 0 (i.e., no cracks have initiated). Generate a fault propagation heatmap: This fault propagation heatmap visualizes the cumulative damage values of all nodes, showing which areas are most severely damaged and where crack initiation is most likely. The cumulative damage value and crack initiation probability of each node serve as the basis for the heatmap, which uses color to indicate the risk level of different areas.

[0041] Specifically, an adversarial generative network is constructed, and the discriminator is used to perform confidence assessment on the fault propagation heat map and the real crack data. The specific process of dynamically adjusting the training sample weights to enhance the learning of early weak signals is as follows: a generator is constructed to simulate the heat map distribution of normal and early fault states, and the discriminator performs confidence scoring on the generated heat map and the real crack data; the sample weights are dynamically calculated according to the discriminator output, and the fuzzy state samples with confidence below the threshold are weighted and strengthened; the generator and discriminator are optimized through adversarial training, so that the generator gradually approaches the real damage propagation mode, thereby improving the model's sensitivity to weak features.

[0042] In this implementation, a generative adversarial network (GAN) is constructed to simulate fault propagation heatmaps. A discriminator is used to assess the confidence of the generated heatmaps and real crack data, dynamically adjusting the weights of training samples to enhance the model's ability to learn early, weak signals. The specific process can be divided into the following key steps: Step 1: Construct a generator. The generator's goal is to simulate the distribution of heatmaps under normal and early-stage fault conditions. The generator uses noise data (typically random noise or historical data) as input to generate heatmaps similar to real crack propagation patterns. The goal is to make the generated heatmaps as close as possible to the actual damage propagation pattern, thereby helping the discriminator distinguish between real and generated samples. The generator is trained based on adversarial learning. During training, it continuously learns how to generate more realistic heatmaps. Specifically, the generator adjusts its parameters through network optimization to minimize the difference between the generated heatmaps and real crack data. Step 2: The discriminator's role is to assign confidence scores to the generated heatmaps and real crack data. Through a binary classification task, the discriminator distinguishes whether the heatmaps are from real data or the generator. The discriminator outputs a confidence score, indicating the probability that the current sample represents real crack data. The discriminator's goal is to distinguish between real and generated data as accurately as possible. The training of the generator and discriminator is adversarial. The generator improves the quality of its generated heatmaps to make it difficult for the discriminator to distinguish between real and fake data, while the discriminator continuously improves its discrimination ability to improve its accuracy. Step 3: During training, the discriminator assigns a confidence score to each sample. Based on this score, the generator's sample weights are dynamically adjusted. The model assigns higher weights to samples whose discriminator output confidence is below a threshold, typically those with ambiguous or weak states. By weighting these weak signal samples, the generator pays more attention to these difficult-to-distinguish samples during training, thereby enhancing its ability to learn early weak signals. Step 4: Through adversarial training, the generator and discriminator compete with each other to continuously optimize their respective objectives. The generator's optimization goal is to make the generated heatmaps increasingly resemble real crack data, thereby deceiving the discriminator. The optimization goal of the discriminator is to improve the accuracy of the discrimination so that it can better distinguish between real crack data and generated data. The optimization process of the generator and the discriminator is usually carried out alternately. The generator improves the quality of its output by generating more realistic heat maps, while the discriminator improves the accuracy of its assessment of sample confidence through training. Through multiple iterations, the generator can gradually approach the real damage propagation pattern and enhance its sensitivity to weak features. Enhance the ability to learn weak signals: By weighting and reinforcing fuzzy or low-confidence samples, the generator pays more attention to early fault signals during training, thereby improving its ability to learn and capture weak signals.Generate realistic fault propagation heatmaps: The generator gradually generates heatmaps that match the actual fault propagation to help make more accurate fault predictions. Improve the generalization ability of the model: By dynamically adjusting sample weights, the model can learn more effective features under different fault conditions, thereby improving the generalization ability of the model in practical applications.

[0043] Specifically, the remaining life regression model and the failure probability classification model are jointly optimized, and a physical constraint loss function is introduced to ensure that the predicted value conforms to the law of material strength attenuation. The final maintenance priority list under the quantile confidence interval is output. The specific process is as follows: a dual-branch network is designed to output the remaining life regression value and the failure probability classification result respectively, and the collaborative update of parameters is achieved through the shared feature layer; a physical constraint term is introduced into the loss function to force the remaining life prediction value to decrease monotonically with the increase of the corrosion attenuation coefficient and the vibration equivalent stress amplitude; the life intervals under different confidence levels are calculated through quantile regression, and the high-risk components are ranked according to the failure probability classification results to generate a list graded by maintenance urgency.

[0044] In this implementation scheme, step 1: design a dual-branch network. The dual-branch network realizes collaborative updating through a shared feature layer and is divided into two sub-networks: Remaining life regression branch: This branch outputs the predicted remaining life (RUL). This value reflects the time that the equipment or component can continue to be used under the current conditions, and is usually a continuous value. Fault probability classification branch: This branch outputs the probability of equipment failure. The output is usually a category probability value, which indicates the possibility of the equipment being in different fault states. The two branches share the input feature layer, which ensures that learning is carried out through common features, thereby better capturing the correlation between features. By sharing the feature layer, the model can optimize the fault probability classification while learning the remaining life, thereby achieving joint optimization of the two goals. Step 2: Introduce physical constraint loss function. In order to ensure that the prediction of the remaining life conforms to the law of material strength attenuation, it is necessary to introduce physical constraints: Physical constraint: The remaining life of the material should decrease monotonically with the increase of the corrosion attenuation coefficient (indicating the degree of material aging) and the vibration equivalent stress amplitude (indicating the load or external influence of the equipment). This attenuation process can be achieved by introducing a physical constraint loss function. The physical constraint loss function is specifically defined as follows: Let RUL pred represents the predicted remaining life, α corrosion and σ vibration are the corrosion attenuation coefficient and the vibration equivalent stress amplitude respectively. Physical constraint requirements: As α corrosion and σ vibration Increase, RUL pred Should be monotonically decreasing. The physical constraint loss can be expressed as: Among them, RUL pred，i and RULpred，i+1 are the predicted remaining life of samples i and i+1, respectively. This term ensures that the remaining life decreases with the increase of the corrosion attenuation coefficient and the vibration equivalent stress amplitude, which conforms to the law of material strength attenuation. Step 3: Joint loss function The overall loss function of the network is the sum of the remaining life regression loss and the failure probability classification loss, plus the physical constraint loss: in: is the remaining life regression loss, usually the mean squared error (MSE) or other regression loss functions. is the failure probability classification loss, usually the cross entropy loss. is the physical constraint loss. λ1 and λ2 are hyperparameters used to adjust the weights of regression loss, classification loss, and physical constraint loss. Step 4: Calculate the quantile confidence interval and calculate the remaining life interval at different confidence levels through quantile regression. Quantile regression allows the model to directly predict the life distribution at different confidence levels instead of a single average value prediction. For each sample, calculate a confidence interval [q low ,q high ], where q low and q high The remaining life values for the low and high quantiles are given in the quantile regression model. For example, a 95% confidence interval corresponds to the 2.5% and 97.5% quantiles. These intervals are defined using multiple quantile values output by the quantile regression model to ensure a certain level of confidence in the life prediction. Step 5: Failure Probability and Maintenance Priority Ranking. The failure probability classification results are combined with the remaining life intervals to assess and prioritize the risk of each component. Components with a higher probability of failure and a shorter remaining life are prioritized. Each component's failure probability is combined with its remaining life to assign a comprehensive risk score. Based on this score, all components are ranked to generate a maintenance priority list. This list helps the maintenance team determine which components require immediate inspection and repair and which can be deferred. Step 6: Output the maintenance priority list. The final output is a list ranked by maintenance urgency, including each component's remaining life interval (based on the quantile confidence interval), each component's failure probability, and a comprehensive risk score (combining remaining life and failure probability). The maintenance priority list ranks components based on their risk scores, prioritizing those with high risk. By designing a two-branch network to jointly optimize the remaining life regression model and the failure probability classification model, and introducing a physical constraint loss function to ensure that the life prediction conforms to the law of material strength decay, the model can simultaneously output the life interval and failure probability with quantile confidence intervals. Combining these outputs, it can rank high-risk components and generate a maintenance urgency list, helping decision makers optimize maintenance plans and improve equipment reliability and maintenance efficiency.

[0045] See also Figure 2 The vehicle enclosure structure fault prediction and maintenance system based on machine learning includes the following modules: multimodal fusion module, dynamic graph network module, damage propagation module, and maintenance optimization module; the multimodal fusion module is used to obtain the time-frequency domain characteristics, local strain data and environmental corrosion parameters of the vibration signal of the vehicle enclosure structure, and spatially align them with the pre-stored finite element simulation stress distribution map, and quantify the influence weight of environmental factors on the fatigue life of the vehicle enclosure structure material through the physical degradation model; the dynamic graph network module is used to construct a dynamic graph topology network according to the physical connection relationship of the vehicle enclosure structure components, take each structural component as a node in the graph, and embed a multimodal feature vector in each node, dynamically adjust the edge weight according to the real-time strain correlation between the nodes, and synchronously capture the vibration signal through the spatiotemporal graph convolution network. The temporal fluctuation characteristics of the components are coupled with the spatial mechanical coupling relationship between the components, and the spatiotemporal fusion feature matrix is output; the damage propagation module is used to calculate the graph attention mechanism on the spatiotemporal fusion feature matrix, identify key risk nodes and associated edges, simulate the propagation path of fatigue damage in the dynamic graph through the graph diffusion model, and predict the crack initiation probability in the high-risk area in combination with the corrosion attenuation coefficient to generate a fault propagation heat map; the maintenance optimization module is used to construct an adversarial generative network, use the discriminator to perform confidence assessment on the fault propagation heat map and the real crack data, dynamically adjust the training sample weights to enhance the learning of early weak signals, jointly optimize the remaining life regression model and the failure probability classification model, introduce the physical constraint loss function to ensure that the predicted value conforms to the material strength attenuation law, and finally output the maintenance priority list under the quantile confidence interval.

[0046] In this implementation scheme, the multimodal fusion module is mainly used to obtain various input data of the vehicle enclosure structure, including the time-frequency domain characteristics of the vibration signal, local strain data, and environmental corrosion parameters, and spatially align them with the pre-stored finite element simulation stress distribution map. This module quantifies the influence weight of environmental factors on the fatigue life of the vehicle enclosure structure material through the physical degradation model. Creativity: Multimodal data fusion: Combining multimodal signals such as vibration signals, strain data, and corrosion parameters can more comprehensively reflect the status of the vehicle enclosure structure than the traditional method that relies only on a single sensor data. Combining finite element simulation stress distribution with physical degradation model: By spatially aligning the dynamic changes of finite element simulation stress distribution with actual environmental factors, the fatigue life impact is quantified, which enhances the model's adaptability to actual working conditions and prediction accuracy. Dynamic graph network module: The main function of the dynamic graph network module is to construct a dynamic graph topology network based on the physical connection relationship of the vehicle enclosure structure, taking each structural component as a node in the graph, and embedding each node into a multimodal feature vector. By dynamically adjusting edge weights based on real-time strain correlations between nodes, a spatiotemporal graph convolutional network (ST-GCN) is used to simultaneously capture the temporal fluctuations of vibration signals and the spatial mechanical coupling between components, outputting a dynamic graph topology structure with a spatiotemporal fusion feature matrix. Unlike traditional static graph networks, dynamic graph networks dynamically update edge weights based on real-time strain data, allowing the network to adapt to changes in actual use and capture more complex dynamic relationships. The spatiotemporal graph convolutional network (ST-GCN) combines temporal and spatial modeling to simultaneously capture the temporal variations in vibration signals and the spatial mechanical coupling between components. This design, compared to traditional graph neural networks, can better reveal dynamic and complex structural behavior. The damage propagation module utilizes a graph attention mechanism to calculate the spatiotemporal fusion feature matrix, identifying key risk nodes and associated edges. It then simulates the propagation path of fatigue damage in the dynamic graph using a graph diffusion model. Combined with the corrosion attenuation coefficient, it predicts the probability of crack initiation in high-risk areas and generates a failure propagation heatmap. Graph Attention Mechanism (GAT): Unlike traditional graph convolutional networks, the graph attention mechanism can more accurately capture the complex relationships between nodes in the graph by assigning different weights to different nodes. This is especially important for identifying key risk nodes. Graph Diffusion Model: By simulating the path of damage propagation through a diffusion model, the system can identify potential fault paths, thereby providing early warnings and avoiding unpredictable failures. Combined with the corrosion attenuation coefficient, the model can be dynamically adjusted to adapt to different environmental changes. Maintenance Optimization Module: This module uses a generative adversarial network (GAN) to perform confidence assessments on the fault propagation heat map and real crack data through a discriminator. Through adversarial training, the weights of training samples are dynamically adjusted to enhance the learning ability of early weak signals.The remaining life regression model and the failure probability classification model are jointly optimized, and a physical constraint loss function is introduced to ensure that the predicted values conform to the laws of material strength degradation. Ultimately, the system outputs a maintenance priority list based on quantile confidence intervals. Generative Adversarial Network (GAN): Through adversarial training, the generator and discriminator compete with each other, enabling the generator to produce fault propagation heatmaps that more closely resemble real crack data. The discriminator helps assess the confidence levels of different heatmaps and dynamically adjusts the weights of training samples to optimize the model's sensitivity to early, weak signals. A joint optimization model: The remaining life regression model and the failure probability classification model are jointly optimized. This not only predicts lifespan but also estimates the probability of failure, providing more comprehensive maintenance decision support. A physical constraint loss function: This ensures that the predicted remaining life conforms to the laws of material strength degradation, enhancing the model's physical rationality and giving it a more realistic engineering context and constraints. Quantile confidence interval output: By calculating quantile intervals at different confidence levels, maintenance decisions are made with greater reliability and accuracy, enabling the generation of precise maintenance priority lists based on different risk levels.

[0047] In summary, this application has at least the following effects:

[0048] A machine learning-based vehicle enclosure structure fault prediction and maintenance method and system utilizes a multimodal fusion module to capture comprehensive characteristics of vibration, strain, and environmental parameters, spatially aligning them with finite element simulation stress maps. This improves the comprehensiveness and accuracy of vehicle enclosure structure fault prediction. The dynamic graph network module dynamically adjusts the network structure based on real-time strain data, accurately capturing the temporal fluctuation characteristics and spatial mechanical coupling relationships between components, effectively modeling the dynamic behavior of the vehicle enclosure structure. The damage propagation module utilizes a graph attention mechanism and a graph diffusion model to accurately identify risk nodes and simulate fatigue damage propagation paths, generating crack initiation probability maps for high-risk areas and enabling early detection of potential failure areas. The maintenance optimization module utilizes a generative adversarial network (GAN) for training. A discriminator performs confidence assessments on the fault propagation heatmap and real crack data, dynamically adjusting the weights of training samples to enhance learning and recognition of early, weak signals. A physically constrained loss function ensures that the remaining life regression model's predictions conform to the material strength decay law, improving the model's physical interpretability and engineering applicability. Quantile regression is used to calculate lifespan intervals at different confidence levels. Combined with the failure probability classification results, a maintenance urgency checklist is generated, helping decision-makers optimize resource allocation and maintenance priorities, thereby improving the reliability and safety of vehicle enclosures. This system provides comprehensive, multi-dimensional condition monitoring and fault prediction for vehicle enclosures, enabling full lifecycle management of vehicle structures, ensuring fault detection and repair at the most appropriate time, extending service life and reducing maintenance costs.

[0049] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0050] The present invention is described with reference to flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0051] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0052] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0053] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0054] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A vehicle enclosure structure fault prediction and maintenance method based on machine learning, characterized in that: The following steps are involved: S1. Obtain the time-frequency domain characteristics of the vehicle enclosure structure's vibration signal, local strain data, and environmental corrosion parameters, spatially align these with pre-stored finite element simulation stress distribution maps, and quantify the weight of the environmental factors' influence on the fatigue life of the vehicle enclosure structure material using a physical degradation model to generate a multimodal feature vector. S2. Based on the physical connections of the vehicle's surrounding structural components, a dynamic graph topology network is constructed. Each structural component is treated as a node in the graph, and a multimodal feature vector is embedded in each node. Based on the real-time strain correlation between nodes, the edge weights are dynamically adjusted. A spatiotemporal graph convolutional network is used to simultaneously capture the temporal fluctuation characteristics of the vibration signal and the spatial mechanical coupling relationship between components, outputting a spatiotemporal fusion feature matrix. S3. Calculate the graph attention mechanism on the spatiotemporal fusion feature matrix to identify key risk nodes and associated edges. Use the graph diffusion model to simulate the propagation path of fatigue damage in the dynamic graph. Combined with the corrosion attenuation coefficient, predict the crack initiation probability in high-risk areas and generate a fault propagation heat map. S4. Construct a generative adversarial network and use a discriminator to perform confidence assessment on the fault propagation heat map and real crack data. Dynamically adjust the weights of training samples to enhance learning of early weak signals. Jointly optimize the remaining life regression model and the failure probability classification model. Introduce a physical constraint loss function to ensure that the predicted value conforms to the material strength attenuation law. Finally, output a maintenance priority list under the quantile confidence interval.

2. The vehicle enclosure structure fault prediction and maintenance method based on machine learning according to claim 1 is characterized by: The specific process of obtaining the time-frequency domain characteristics of the vibration signal, local strain data, and environmental corrosion parameters of the vehicle enclosure structure and spatially aligning them with the pre-stored finite element simulation stress distribution map is as follows: Based on the geometric model of the vehicle's surrounding structure, the stress distribution map generated by the finite element simulation is mapped to the actual physical coordinate system, and the position coordinates of the vibration sensor, strain gauge, and environmental sensor are aligned using a spatial grid matching algorithm. Short-time Fourier transform is performed on the vibration signal to extract the time-frequency domain energy distribution characteristics. At the same time, the local strain area is divided according to the spatial layout of the strain gauge. Combined with the spatial distribution of environmental corrosion parameters, the correlation mapping relationship between multi-source data and simulated stress maps is established to generate a spatially aligned multi-dimensional data matrix.

3. The vehicle enclosure structure fault prediction and maintenance method based on machine learning according to claim 2, characterized in that: The specific process of quantifying the influence weight of environmental factors on the fatigue life of vehicle surrounding structure materials through the physical degradation model and generating multimodal eigenvectors is as follows: Based on the spatially aligned multi-dimensional data matrix, the environmental corrosion parameters, vibration time-frequency domain energy distribution, and simulated stress values corresponding to each structural area are extracted; The environmental parameters are input into the physical degradation model, the corrosion rate of the material surface is calculated through the corrosion kinetics equation, and the equivalent cyclic stress amplitude of the vibration signal is coupled to generate the environment-load joint attenuation factor; The environment-load joint attenuation factor is nonlinearly weighted fused with the simulated stress value to quantify the loss weight of fatigue life of materials in different regions, and finally a multimodal feature vector containing physical mechanism constraints is constructed.

4. The vehicle enclosure structure fault prediction and maintenance method based on machine learning according to claim 3 is characterized by: The dynamic graph topology network is constructed, each structural component is used as a node in the graph, and a multimodal feature vector is embedded in each node. The specific process of dynamically adjusting the weight of the edge according to the real-time strain correlation between the nodes is as follows: The initial graph topology is defined based on the assembly relationship of the vehicle enclosure structure. Nodes represent structural components, including longitudinal beams, cross beams, and suspension brackets, and edges represent physical connections, including welds and bolts. Assign a multimodal feature vector to each node, including physical degradation indicators, local strain mean and vibration spectrum entropy value; The Pearson correlation coefficient of the strain data of adjacent nodes is calculated in real time, and a dynamic edge weight update rule is set: when the correlation coefficient exceeds the threshold, the edge weight is enhanced to characterize the activity of the load transfer path; otherwise, the weight is reduced to weaken invalid connections, thereby achieving adaptive adjustment of the graph topology according to the mechanical state.

5. The vehicle enclosure structure fault prediction and maintenance method based on machine learning according to claim 4 is characterized in that: The specific process of synchronously capturing the temporal fluctuation characteristics of the vibration signal and the spatial mechanical coupling relationship between components through the spatiotemporal graph convolutional network and outputting the spatiotemporal fusion feature matrix is as follows: A layered spatiotemporal graph convolution module is designed, and dilated convolution is used in the time dimension to extract the long-term trend characteristics and short-term transient impact characteristics of the vibration signal. In the spatial dimension, the adjacent node features are weighted and aggregated according to the dynamic edge weights to capture the coupling effect of stress transfer between components; The output features of the spatiotemporal convolution are spliced along the channel dimension, and the key feature channels are screened through the attention gating mechanism to generate a spatiotemporal fusion feature matrix. Each row vector in the matrix represents the implicit health status of the corresponding component under the action of spatiotemporal coupling.

6. The vehicle enclosure structure fault prediction and maintenance method based on machine learning according to claim 5, characterized in that: The specific process of calculating the graph attention mechanism on the spatiotemporal fusion feature matrix and identifying key risk nodes and associated edges is as follows: Based on the spatiotemporal fusion feature matrix, a multi-head graph attention network is constructed, and the dynamic edge weight is used as the initial bias term of the attention distribution; The corrosion attenuation coefficient and the vibration transient energy ratio are integrated in the node feature dimension to generate the corrosion-vibration collaborative attention factor; The features of adjacent nodes are nonlinearly transformed through a learnable parameter matrix, and the normalized attention distribution is calculated in combination with the attention factor. Nodes and associated edges with attention scores higher than the preset threshold are screened, marked as key risk nodes and high-risk connection paths, and the risk topology subgraph is output.

7. The vehicle enclosure structure fault prediction and maintenance method based on machine learning according to claim 6, characterized in that: The graph diffusion model is used to simulate the propagation path of fatigue damage in the dynamic graph. The corrosion attenuation coefficient is combined to predict the crack initiation probability in high-risk areas. The specific process of generating the fault propagation heat map is as follows: Taking the key risk nodes as the damage diffusion source, the diffusion intensity is initialized as the corrosion-vibration collaborative attention factor of the node; The damage transfer efficiency along the connecting edge is calculated based on the dynamic edge weight and the finite element simulation stress gradient. The corrosion attenuation coefficient is introduced as a diffusion damping term to suppress the damage accumulation in the low corrosion area. The cumulative damage value of each node is updated through the iterative diffusion equation, where the increment is nonlinearly positively correlated with the diffusion intensity, transfer efficiency and current damage value; The cumulative damage value is compared with the material fatigue threshold, and the crack initiation probability is predicted by combining the corrosion attenuation coefficient. A spatial fault propagation heat map is generated to mark the potential crack initiation area and propagation path.

8. The vehicle enclosure structure fault prediction and maintenance method based on machine learning according to claim 7 is characterized in that: The specific process of building a generative adversarial network, using a discriminator to perform confidence assessment on the fault propagation heat map and real crack data, and dynamically adjusting the weights of training samples to enhance the learning of early weak signals is as follows: A generator is built to simulate the thermal map distribution of normal and early fault states, and a discriminator performs confidence scoring on the generated thermal map and the real crack data. Dynamically calculate sample weights based on the discriminator output, and perform weighted reinforcement on fuzzy state samples whose confidence level is lower than the threshold; By optimizing the generator and discriminator through adversarial training, the generator gradually approaches the real damage propagation pattern and improves the model's sensitivity to weak features.

9. The vehicle enclosure structure fault prediction and maintenance method based on machine learning according to claim 8, characterized in that: The specific process of jointly optimizing the remaining life regression model and the failure probability classification model, introducing a physical constraint loss function to ensure that the predicted value conforms to the material strength attenuation law, and finally outputting the maintenance priority list under the quantile confidence interval is as follows: A dual-branch network is designed to output the remaining life regression value and the failure probability classification result respectively, and the parameters are collaboratively updated through the shared feature layer. A physical constraint term is introduced into the loss function to force the remaining life prediction value to decrease monotonically with the increase of the corrosion attenuation coefficient and the vibration equivalent stress amplitude. The life intervals at different confidence levels are calculated through quantile regression. The high-risk components are ranked according to the failure probability classification results, and a maintenance urgency list is generated.

10. A vehicle enclosure structure fault prediction and maintenance system based on machine learning, applying the vehicle enclosure structure fault prediction and maintenance method based on machine learning according to any one of claims 1 to 9, characterized in that: Includes the following modules: multimodal fusion module, dynamic graph network module, damage propagation module, maintenance optimization module; The multimodal fusion module is used to obtain the time-frequency domain characteristics of the vibration signal of the vehicle enclosure structure, local strain data and environmental corrosion parameters, and spatially align them with the pre-stored finite element simulation stress distribution map, and quantify the influence weight of environmental factors on the fatigue life of the vehicle enclosure structure material through the physical degradation model; The dynamic graph network module is used to construct a dynamic graph topology network based on the physical connection relationship of the vehicle's surrounding structural components. Each structural component is used as a node in the graph, and a multimodal feature vector is embedded in each node. The edge weights are dynamically adjusted based on the real-time strain correlation between each node. The spatiotemporal graph convolutional network synchronously captures the temporal fluctuation characteristics of the vibration signal and the spatial mechanical coupling relationship between the components, and outputs a spatiotemporal fusion feature matrix. The damage propagation module is used to calculate the graph attention mechanism on the spatiotemporal fusion feature matrix, identify key risk nodes and associated edges, simulate the propagation path of fatigue damage in the dynamic graph through the graph diffusion model, predict the crack initiation probability in high-risk areas based on the corrosion attenuation coefficient, and generate a fault propagation heat map; The maintenance optimization module is used to build an adversarial generative network, use a discriminator to perform confidence assessment on fault propagation heat maps and real crack data, dynamically adjust training sample weights to enhance learning of early weak signals, jointly optimize the remaining life regression model and the failure probability classification model, introduce a physical constraint loss function to ensure that the predicted value conforms to the law of material strength attenuation, and finally output a maintenance priority list within the quantile confidence interval.

Images