Automobile chassis abnormal state detection system based on multi-modal fusion

Abstract

The application relates to the technical field of automobile detection, and discloses an automobile chassis abnormal state detection system based on multi-modal fusion. The system comprises a multi-modal data acquisition module, a cross-modal feature fusion module, an abnormal feature optimization module, an abnormal state quantification module and an abnormal state decision module. The multi-modal data acquisition module synchronously acquires chassis vibration signals, temperature distribution thermal diagrams and running noise spectrum sequence; the cross-modal feature fusion module extracts corresponding feature vectors of the sequences, generates a multi-modal joint feature vector through alignment and splicing; the abnormal feature optimization module decouples the joint feature vector and optimizes the joint feature vector through a feature stability screening algorithm; the abnormal state quantification module inputs the optimized feature vector set into a quantification network to generate a chassis abnormal state quantification coefficient; and the abnormal state decision module compares the quantification coefficient with a preset threshold value and outputs a chassis maintenance instruction. The system can comprehensively capture chassis abnormal features, improves detection accuracy and robustness, realizes abnormal quantification and efficient maintenance decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automobile detection, in particular to an automobile chassis abnormal state detection system based on multi-modal fusion. BACKGROUND

[0002] With the rapid development of the automobile industry, vehicle safety performance and reliability are increasingly valued, and the automobile chassis, as the core component supporting the overall structure of the vehicle, realizing power transmission and driving control, its running state is directly related to the driving safety and service life of the vehicle. In actual use, the chassis is long-term subjected to road impact, load change and environmental corrosion, etc. factors, prone to abnormal problems such as suspension system wear, transmission component loosening, bearing aging, etc. If not detected and processed in time, it may cause vehicle control failure, abnormal noise during driving and even more serious safety accidents.

[0003] The current detection method for abnormal state of automobile chassis mainly includes manual detection and single sensor detection. Manual detection relies on technical personnel to judge the chassis state through visual inspection, hand touch and simple tool measurement, etc. This method is not only greatly affected by personnel experience and subjective judgment, has low detection efficiency, and is difficult to find hidden abnormalities inside the chassis, such as early bearing wear, pipeline slight leakage, etc. Single sensor detection collects chassis running data through single devices such as vibration sensor, temperature sensor or noise sensor, and then judges the abnormality based on the data features. For example, only through the vibration sensor to collect the chassis vibration signal, analyze the signal frequency and amplitude change to identify the abnormality, but this method can only reflect the running information of the chassis in one aspect, and cannot fully capture the multi-dimensional features of the chassis abnormality - such as the abnormality of the chassis transmission component will cause the change of vibration signal, also accompanied by local temperature rise and noise spectrum change, single sensor detection is easy to cause misjudgment or omission due to the one-sidedness of the information, especially in the early stage of the chassis slight abnormality, the change of single data feature is not obvious, it is more difficult to realize accurate detection.

[0004] The existing detection technology has obvious deficiencies in data processing and feature fusion. Although some detection systems attempt to combine multiple sensor data, they mostly use simple data splicing or weighted summation to process multi-source information, failing to realize effective association and complementarity between different modal data (such as time-frequency features of vibration signals, spatial features of temperature distribution, and voiceprint features of noise signals), resulting in that the fused data features cannot accurately reflect the overall running state of the chassis. At the same time, due to the complex running environment of the chassis, the data collected by the sensors are easily affected by factors such as road interference, environmental temperature changes, and electromagnetic noise, and the existing technology lacks an effective feature optimization mechanism, making it difficult to extract stable and reliable abnormal features from noisy multi-modal data, further reducing the accuracy and robustness of chassis abnormality detection. These problems make it difficult for current chassis detection technology to meet the high requirements of modern automobiles for safety performance and reliability, and a technical solution is needed that can comprehensively, accurately and efficiently detect the abnormal state of the chassis. SUMMARY

[0005] The purpose of the present application is to provide a multi-modal fusion-based automobile chassis abnormal state detection system to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides a multi-modal fusion-based automobile chassis abnormal state detection system, which comprises:

[0007] A multi-modal data acquisition module is used to synchronously acquire a chassis vibration signal sequence, a chassis temperature distribution thermograph sequence, and a chassis running noise spectrogram sequence.

[0008] A cross-modal feature fusion module is used to extract a time-frequency feature vector of the chassis vibration signal sequence, a spatial thermodynamic feature vector of the chassis temperature distribution thermograph sequence, and a voiceprint feature vector of the chassis running noise spectrogram sequence, respectively, and to perform feature alignment and splicing on the time-frequency feature vector, the spatial thermodynamic feature vector, and the voiceprint feature vector to generate a multi-modal joint feature vector.

[0009] An abnormal feature optimization module is used to perform feature decoupling on the multi-modal joint feature vector to obtain a set of multiple chassis state local feature vectors, and to optimize the set of multiple chassis state local feature vectors through a feature stability screening algorithm to obtain a set of optimized chassis state local feature vectors.

[0010] An abnormal state quantification module is used to input the set of optimized chassis state local feature vectors into an abnormal state quantification network to generate a chassis abnormal state quantification coefficient.

[0011] An abnormal state decision module is used to output a chassis maintenance instruction according to the comparison result of the chassis abnormal state quantification coefficient and a preset threshold.

[0012] Preferably, the multi-modal data acquisition module comprises:

[0013] a vibration signal acquisition unit configured to acquire a sequence of chassis vibration signals within a continuous time window via a distributed array of acceleration sensors;

[0014] a thermal map acquisition unit configured to acquire a sequence of chassis temperature distribution thermal maps at preset time intervals via an infrared thermal imager;

[0015] a noise spectrum acquisition unit configured to capture a sequence of chassis operating noise spectrograms via a directional microphone array;

[0016] a timestamp alignment unit configured to add synchronized timestamp identifiers to the sequence of chassis vibration signals, the sequence of chassis temperature distribution thermal maps, and the sequence of chassis operating noise spectrograms.

[0017] Preferably, the cross-modal feature fusion module comprises:

[0018] a vibration feature extraction unit configured to perform wavelet packet decomposition on the sequence of chassis vibration signals to extract a time-frequency energy distribution feature vector;

[0019] a thermal feature extraction unit configured to perform convolutional feature encoding on the sequence of chassis temperature distribution thermal maps to extract a spatial thermal gradient feature vector;

[0020] a voiceprint feature extraction unit configured to perform Mel-frequency cepstral coefficient analysis on the sequence of chassis operating noise spectrograms to extract a voiceprint energy distribution feature vector;

[0021] a feature scale normalization unit configured to map the time-frequency energy distribution feature vector, the spatial thermal gradient feature vector, and the voiceprint energy distribution feature vector to a unified feature dimension space;

[0022] a cross-modal concatenation unit configured to concatenate the normalized feature vectors according to the timestamp identifiers to generate the multi-modal joint feature vector.

[0023] Preferably, the abnormal feature optimization module comprises:

[0024] a feature decoupling unit configured to decompose the multi-modal joint feature vector into a plurality of chassis state local feature vectors;

[0025] a feature stability calculation unit configured to calculate a feature fluctuation rate of each chassis state local feature vector within a continuous time window;

[0026] a feature screening unit configured to screen, based on a comparison of the feature fluctuation rate with a preset fluctuation threshold, an optimized chassis state local feature vector with a feature fluctuation rate lower than the preset fluctuation threshold from the plurality of chassis state local feature vectors to form a set of optimized chassis state local feature vectors.

[0027] Preferably, the abnormal state quantification module comprises:

[0028] a local feature correlation unit configured to calculate a feature correlation degree matrix between any two optimized chassis state local feature vectors in the set of optimized chassis state local feature vectors;

[0029] an abnormal propagation tree construction unit configured to generate a chassis abnormal state propagation tree structure according to the feature correlation degree matrix;

[0030] an abnormal quantification network unit configured to input the chassis abnormal state propagation tree structure into a graph neural network to output the chassis abnormal state quantification coefficient.

[0031] Preferably, the feature stability calculation unit comprises:

[0032] a time window division sub-unit configured to divide a continuous time window into a plurality of equal-length sub-time segments;

[0033] a feature distribution calculation sub-unit configured to calculate a feature mean vector and a feature variance vector of each chassis state local feature vector within the plurality of equal-length sub-time segments;

[0034] a fluctuation factor generation sub-unit configured to generate the feature fluctuation rate according to a ratio of a module length of the feature variance vector to a module length of the feature mean vector.

[0035] Preferably, the abnormal state decision module comprises:

[0036] an abnormal level mapping unit configured to map a corresponding chassis abnormal level according to a numerical interval in which the chassis abnormal state quantification coefficient is located;

[0037] a maintenance strategy generation unit configured to query a preset maintenance strategy mapping table based on the chassis abnormal level to generate a chassis maintenance instruction;

[0038] a maintenance instruction verification unit configured to output a final maintenance instruction after conflict detection between the chassis maintenance instruction and a historical maintenance record.

[0039] Preferably, the system further comprises:

[0040] a multi-modal data storage module configured to store the chassis vibration signal sequence, the chassis temperature distribution thermograph sequence, and the chassis running noise spectrogram sequence according to a timestamp identifier in a distributed columnar storage structure.

[0041] a feature retrieval module configured to retrieve historical feature vectors in the multi-modal data storage module according to a time range to construct a reference feature vector set.

[0042] Preferably, the abnormal state quantification module further comprises:

[0043] a real-time feature comparison unit configured to calculate a feature offset matrix of the set of optimized chassis state local feature vectors and the reference feature vector set;

[0044] a quantification compensation unit configured to dynamically correct the chassis abnormal state quantification coefficients according to the feature offset matrix.

[0045] Preferably, the maintenance strategy generation unit comprises:

[0046] an abnormal propagation factor calculation sub-unit configured to calculate an abnormal propagation factor according to the node connectivity of the chassis abnormal state propagation tree structure;

[0047] a strategy optimization sub-unit configured to adjust the maintenance response priority in the maintenance strategy mapping table based on the abnormal propagation factor.

[0048] Compared with the prior art, the present application has the following advantages:

[0049] The chassis vibration signal sequence, temperature distribution thermal map sequence and running noise spectrum sequence are synchronously acquired by the multi-modal data acquisition module, breaking through the limitation of traditional single sensor detection that can only obtain one-sided information. Traditional detection methods often rely on single type data and cannot fully reflect the multi-dimensional performance of the chassis abnormality. However, the present system can capture chassis running state information from three key dimensions of mechanical vibration, thermal distribution and acoustic characteristics by synchronously acquiring three different modal running data. Any component of the chassis that appears abnormal, such as suspension spring fatigue, transmission gear wear or bearing damage, will produce corresponding changes in vibration, temperature and noise. Synchronous acquisition of multi-modal data can achieve comprehensive monitoring of the chassis running state, avoid the problem of missed judgment of abnormalities due to information loss, and make the capture of chassis abnormal state more comprehensive.

[0050] The cross-modal feature fusion module extracts time-frequency feature vectors of vibration signals, spatial thermal feature vectors of temperature thermal maps, and voiceprint feature vectors of noise spectrograms according to the characteristics of different modal data, performs feature alignment and splicing to generate a multi-modal joint feature vector, and effectively solves the problem of low efficiency of feature fusion in traditional multi-sensor data processing. The feature dimensions and physical meanings of different modal data are different, and if simple splicing or weighting is used, the feature information may be chaotic or the effective information may be covered. The module eliminates the differences in time and space dimensions of different modal data through feature alignment, and then realizes feature complementation through reasonable splicing. For example, the time-frequency features of vibration signals can reflect the dynamic changes of the component motion state, the spatial features of temperature distribution can locate the local overheating area, and the voiceprint features of noise can assist in judging the component friction or looseness. The joint feature vector formed by the fusion of the three can comprehensively reflect the multi-dimensional features of the chassis abnormality, greatly improving the feature data's representation ability of the chassis state and providing more comprehensive and accurate basic data for subsequent abnormality detection.

[0051] The abnormal feature optimization module decomposes the multi-modal joint feature vector into a set of local feature vectors of the chassis state through feature decoupling, and optimizes it through a feature stability screening algorithm, effectively improving the reliability and robustness of the abnormal feature. There are many interference factors in the chassis running environment, such as vibration signal fluctuations caused by road bumps, temperature data affected by environmental temperature changes, and interference of external noise on chassis running noise collection. These interferences may cause the collected multi-modal data to contain redundant or noisy information, which may lead to false positives if used directly for abnormality detection. The module decomposes the joint feature into local features related to different components and different running states of the chassis through feature decoupling, and then eliminates unstable features that are greatly affected by interference and have large fluctuations through a stability screening algorithm, and retains feature vectors that can stably reflect the real state of the chassis, so that the optimized local feature vector set can more accurately represent the actual running state of the chassis and reduce the influence of interference factors on the detection results.

[0052] The abnormal state quantification module inputs the optimized local feature vector set into the abnormal state quantification network to generate chassis abnormal state quantification coefficients, achieving accurate description of the degree of chassis abnormality. Traditional detection techniques can only determine whether the chassis is abnormal, but cannot quantify the severity of the abnormality, making it difficult for users to develop a reasonable maintenance plan based on the detection results. If only the existence of an abnormality is determined, there may be over-maintenance of a slight abnormality or delayed maintenance of a serious abnormality. The module converts abstract feature vectors into specific quantification coefficients through a quantification network, which can intuitively reflect the severity of the chassis abnormality. For example, a higher coefficient indicates a more serious abnormality, and a lower coefficient indicates an early stage of abnormality, providing accurate quantification basis for subsequent maintenance decisions and making maintenance work more targeted.

[0053] The abnormal state decision module outputs chassis maintenance instructions according to the comparison result of the abnormal state quantitative coefficient and the preset threshold value, realizing seamless connection of detection and maintenance decision. After the traditional detection technology obtains the detection result, manual further analysis and development of a maintenance scheme are often required, the process is tedious and is easily affected by human factors, resulting in untimely maintenance response. The system automatically outputs the corresponding maintenance instructions when the quantitative coefficient exceeds the threshold value by presetting a reasonable threshold value, such as outputting the "suggested regular inspection" instruction when the quantitative coefficient is slightly higher than the threshold value, and outputting the "immediate shutdown maintenance" instruction when the quantitative coefficient is much higher than the threshold value, so that accurate maintenance decisions can be quickly generated without human intervention, the maintenance response speed is greatly improved, the chassis abnormalities can be processed in time, the vehicle driving safety is effectively ensured, the resource waste caused by excessive maintenance is avoided, and the efficiency and economy of chassis maintenance are unified. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 A timing diagram of the automobile chassis abnormal state detection system based on multi-modal fusion provided by the present application;

[0055] Figure 2 A flowchart of the working process of the multi-modal data acquisition module;

[0056] Figure 3 A flowchart of the working process of the abnormal feature optimization module;

[0057] Figure 4 A flowchart of the working process of the abnormal state quantitative module;

[0058] Figure 5 A flowchart of the working process of the abnormal state decision module. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0060] Please refer to Figure 1 The present application provides an automobile chassis abnormal state detection system based on multi-modal fusion, which comprises:

[0061] The multi-modal data acquisition module synchronously acquires the chassis vibration signal sequence, the chassis temperature distribution thermograph sequence and the chassis running noise spectrogram sequence. The cross-modal feature fusion module respectively extracts the time-frequency feature vector of the chassis vibration signal sequence, the spatial thermodynamic feature vector of the chassis temperature distribution thermograph sequence and the voiceprint feature vector of the chassis running noise spectrogram sequence, and performs feature alignment and splicing on the above feature vectors to generate a multi-modal joint feature vector. The abnormal feature optimization module performs feature decoupling on the multi-modal joint feature vector to obtain a set of multiple chassis state local feature vectors, and optimizes the set through a feature stability screening algorithm to obtain a set of optimized chassis state local feature vectors. The abnormal state quantification module inputs the set of optimized chassis state local feature vectors into an abnormal state quantification network to generate a chassis abnormal state quantification coefficient. The abnormal state decision module outputs a chassis maintenance instruction according to the comparison result of the chassis abnormal state quantification coefficient and a preset threshold.

[0062] Embodiment 1: see Figure 2 The implementation of the multi-modal data acquisition module first relies on a vibration signal acquisition unit. The unit acquires a chassis vibration signal sequence within a continuous time window through a distributed acceleration sensor array. An ADI ADXL355 three-axis acceleration sensor is used, with a sampling frequency set to 1 kHz. A total of 3 sensors are arranged, respectively installed on the chassis suspension spring seat, steering knuckle and transmission shaft bearing seat, for acquiring the chassis vibration signal sequence within the continuous time window. The sensor array is usually composed of multiple three-axis accelerometers, which are strategically arranged at key positions of the chassis suspension system, bearing seat, steering knuckle, etc. Each sensor works at a constant sampling frequency, continuously recording vibration acceleration data in three orthogonal directions. These raw voltage signals are converted into digital signal sequences through an analog-to-digital converter, forming discrete vibration amplitude records in the time domain. During data acquisition, the acquisition unit applies an anti-aliasing filter to eliminate high-frequency noise, and uses an embedded processor for preliminary preprocessing such as DC component removal and signal noise reduction.

[0063] The thermal map acquisition unit acquires a sequence of chassis temperature distribution thermal maps at preset time intervals using an infrared thermal imager. The FLIR AX8 infrared thermal imager is used, with a resolution of 320x240 pixels and a preset time interval of 2 seconds / frame, to acquire a sequence of chassis temperature distribution thermal maps. The thermal imager is installed on an adjustable gimbal, which can perform scanning photography on specific areas of the chassis. Each thermal map frame contains a large number of pixel points, each of which records the temperature value at the corresponding position. The thermal imager has an appropriate temperature measurement range and spatial resolution, and can capture subtle temperature changes of chassis components during operation. During the acquisition process, the thermal imager performs non-uniformity correction and temperature drift compensation to ensure the accuracy of temperature measurement. The acquired thermal map sequence is stored in the form of image frames, while retaining the acquisition time information of each frame.

[0064] The noise spectrum acquisition unit captures a sequence of chassis operation noise spectrograms using a directional microphone array. The NT-USB+ directional microphone array from Rode is used, with a pickup range of 20Hz-20kHz, and a total of 4 microphones arranged at the front, back, left and right sides of the chassis to capture a sequence of chassis operation noise spectrograms. The microphone array is composed of multiple high-precision condenser microphones arranged in a geometric pattern, with directional receiving characteristics. The array is installed at an appropriate position near the chassis, and through beamforming technology, it enhances the ability to capture chassis noise while suppressing environmental background noise. The sound pressure signals collected by each microphone are amplified and filtered, and then sent to a digital signal processor for short-time Fourier transform, converting time-domain sound pressure signals into frequency-domain sound spectrograms. These sound spectrograms are represented in a three-dimensional form of time-frequency-intensity, recording the changes in the spectral characteristics of the chassis operation noise over time.

[0065] The timestamp alignment unit adds synchronized timestamp labels to the data sequences of the three modalities, with millisecond-level precision, and unifies the timestamps of the three modalities based on a GPS time synchronization module. The unit uses a high-precision clock source, usually based on a global positioning system time synchronization module, to provide a unified time reference for each data acquisition unit. When each sampling point of the vibration signal sequence, each image of the thermal map sequence, and each time-frequency frame of the noise spectrogram sequence are collected, they are marked with a timestamp with millisecond-level precision. This synchronization mechanism ensures the consistency of different modal data in the time dimension, creating the necessary conditions for subsequent multi-modal feature fusion. The implementation of the cross-modal feature fusion module begins with the vibration feature extraction unit, which performs wavelet packet decomposition processing on the chassis vibration signal sequence. The multi-resolution analysis method is used to decompose the signal into different frequency bands. The wavelet packet decomposition uses the db4 wavelet basis function, with 5 layers of decomposition. Through this parameter, the time-frequency energy distribution feature vector of the chassis vibration signal sequence is extracted. The decomposition process uses a wavelet basis function to decompose the original vibration signal into a series of detail coefficients and approximation coefficients through multiple layers of filtering operations. The energy features of each frequency band are extracted from these coefficients to form the time-frequency energy distribution feature vector. This feature vector can reflect the energy distribution characteristics of the chassis vibration at different frequency components and capture changes in the vibration pattern caused by component abnormalities.

[0066] The thermal feature extraction unit performs convolutional feature encoding on the chassis temperature distribution thermal map sequence. The convolutional kernel size of the convolutional feature encoding is 3x3, the step size is 1, and the number of convolutional layers is 2. Through these parameters, the spatial thermal gradient feature vector of the chassis temperature distribution thermal map sequence is extracted. This unit uses the convolutional neural network structure in the deep learning framework to process the thermal map sequence frame by frame. The network is composed of multiple convolutional layers, pooling layers, and activation functions, which extract thermal features at different abstraction levels from the original thermal map through a hierarchical feature extraction process. The final output spatial thermal gradient feature vector can represent the spatial variation pattern of the chassis temperature distribution, including the thermal gradient direction, the shape of the thermal spot area, and other key information, which is closely related to the thermal state of the chassis components. The voiceprint feature extraction unit performs Mel-frequency cepstral coefficient analysis on the chassis operating noise spectrogram sequence. The extraction dimension of the Mel-frequency cepstral coefficient (MFCC) is 12, the frame length is set to 20ms, and the frame shift is set to 10ms. Through these parameters, the voiceprint energy distribution feature vector of the chassis operating noise spectrogram sequence is extracted. The processing process first maps the linear spectrum to the Mel scale based on human auditory characteristics, then performs logarithmic operation and discrete cosine transform, and finally extracts a set of cepstral coefficients. These coefficients form the voiceprint energy distribution feature vector, which can effectively represent the spectral features of the chassis noise, especially highlighting the abnormal changes in specific frequency components caused by component wear or damage.

[0067] The characteristic scale normalization unit maps the above three feature vectors to a unified feature dimension space. Since the feature vectors extracted by different modalities may have different numerical ranges and dimensions, this unit uses a standardization processing method to adjust the scale of each feature vector. The normalization method is Z-Score standardization, which maps the three feature vectors to a unified feature dimension space with a mean of 0 and a standard deviation of 1. The processing process includes range normalization and dimension unification of feature values, so that all feature vectors have the same dimension and comparable numerical range. This processing eliminates the dimensional differences between different modal features, creating conditions for subsequent feature fusion. The cross-modal splicing unit splices the normalized feature vectors according to the timestamp identifier. This unit first aligns different modal features at the same time point according to the timestamp information, and then connects these feature vectors in the feature dimension. The splicing process maintains the integrity of the time sequence, ensuring that each time point corresponds to a joint feature vector that integrates multi-modal information. The final multi-modal joint feature vector contains feature information of vibration, temperature and noise of three modalities, forming a comprehensive representation of the chassis running state.

[0068] Embodiment 2: refer to Figure 3 The core component of the abnormal feature optimization module is the feature decoupling unit, which receives multi-modal joint feature vectors from the cross-modal feature fusion module. These vectors already contain the fusion information of vibration, temperature and noise of three modalities. The feature decoupling unit uses a neural network-based feature decomposition method to decompose the high-dimensional multi-modal joint feature vector into multiple low-dimensional chassis state local feature vectors. This decomposition process is realized through a specially designed fully connected neural network, which has multiple output branches, each branch corresponding to a feature dimension. During network training, sparsity constraints are used to enable different output branches to capture relatively independent feature components in the joint feature vector. Each chassis state local feature vector after decomposition represents the state information of a specific aspect of the chassis, and these vectors together form a feature set that can comprehensively describe the chassis state.

[0069] The feature stability calculation unit is responsible for evaluating the stability degree of each chassis state local feature vector in the time dimension. The unit first divides the continuous monitoring time window into multiple equal-length sub-time segments by a time window division sub-unit. The continuous time window length is 60 seconds, which is divided into 12 equal-length sub-time segments, each with a duration of 5 seconds. The length of the time window is determined according to the dynamic characteristics of the chassis system. The length of each sub-time segment needs to be short enough to capture the dynamic changes of the feature and long enough to contain a statistically significant amount of data. During the division process, overlapping window technology is used to maintain time continuity while ensuring that each sub-segment has independent data support. Within each sub-time segment, a feature distribution calculation sub-unit performs statistical analysis on each chassis state local feature vector. The sub-unit calculates the statistical distribution characteristics of the feature vector in each time segment, including calculating the feature mean vector and the feature variance vector. The mean vector reflects the average level of the feature in that time period, and the variance vector represents the fluctuation of the feature. During the calculation process, a recursive algorithm is used to update the statistics in real time to meet the needs of online monitoring. For each local feature vector, these statistics are calculated separately in each dimension of its feature space, resulting in a corresponding mean vector and variance vector.

[0070] The volatility factor generation sub-unit generates the feature volatility rate based on the calculated feature statistics. This sub-unit uses a vector norm-based calculation method to compare the norm of the feature variance vector with the norm of the feature mean vector. This calculation method can consider the fluctuation of the feature in each dimension and generate a scalarized volatility rate index. The calculation of the volatility rate fully considers the multi-dimensional characteristics of the feature vector, avoiding the limitations of single-dimensional analysis. Each local feature vector will calculate a corresponding feature volatility rate, which quantifies the stability degree of the feature in the time dimension.

[0071] Suppose a chassis state local feature vector (corresponding to the time-frequency energy feature of the vibration signal) has a feature mean vector norm of [10, 8, 9, 11, 10, 9, 11, 10, 9, 8, 10, 11] and a feature variance vector norm of [1.2, 0.9, 1.0, 1.1, 1.0, 0.8, 1.2, 0.9, 1.0, 0.7, 1.1, 1.0] in 12 5-second sub-time segments.

[0072] The calculation formula of the feature volatility rate is:

[0073]

[0074] According to the above formula, the characteristic fluctuation rate of each sub-segment is calculated, and the results are [0.12, 0.11, 0.11, 0.10, 0.10, 0.09, 0.11, 0.09, 0.11, 0.09, 0.11, 0.09] in turn, and the average fluctuation rate of the 12 sub-segments is 0.105, which is ≤ the preset fluctuation threshold 0.15, so the chassis state local feature vector is retained and included in the set of optimized chassis state local feature vectors.

[0075] The feature screening unit selects features according to the calculated characteristic fluctuation rate, and the preset fluctuation threshold is 0.15, i.e. only the chassis state local feature vectors with a characteristic fluctuation rate ≤ 0.15 are retained to form the set of optimized chassis state local feature vectors. This unit sets a preset fluctuation threshold, which is based on statistical analysis of a large amount of chassis running data under normal state. The threshold setting needs to balance sensitivity and specificity, that is, it should be able to identify unstable abnormal features, and avoid misjudging normal state fluctuations as abnormalities. During the screening process, the fluctuation rate of each chassis state local feature vector is compared with the preset threshold. Those feature vectors with a fluctuation rate lower than the preset threshold are considered to have good stability and can more reliably reflect the running state of the chassis. These screened feature vectors form the set of optimized chassis state local feature vectors.

[0076] The entire optimization process is carried out online and can process the continuously inputted multi-modal joint feature vectors in real time. The system maintains a dynamic feature vector set, and as new data is continuously inputted, the calculation and screening results of feature stability are continuously updated. This dynamic optimization mechanism enables the system to adapt to changes in the chassis state over time and always maintain accurate representation of the latest state. Through feature decoupling and stability screening, the system can extract feature components from multi-modal fusion features that contain rich state information and have good stability. These optimized feature vectors reduce noise and irrelevant changes, and are more focused on reflecting the essential state characteristics of the chassis. This processing provides more reliable and stable feature input for subsequent abnormal state quantification, enhancing the recognition ability of the entire detection system for abnormal states of the chassis.

[0077] In actual operation, the network parameters of the feature decoupling unit need to be fully trained and learned. The training data includes chassis multi-modal data under various operating conditions, including both normal state data and various typical abnormal state data. Through such training, the network can learn how to decompose the multi-modal joint features into physically meaningful local feature vectors. The threshold parameter in the feature stability calculation also needs to be adjusted according to the actual application scenario. Different chassis types or operating environments may require different threshold settings. The implementation of the entire abnormal feature optimization module focuses on the balance between computational efficiency and real-time performance. The feature decoupling neural network is designed to be lightweight, minimizing computational complexity while ensuring decomposition effectiveness. The feature stability calculation uses an incremental algorithm to avoid repeated calculations and improve processing efficiency. These design considerations enable the module to meet the needs of real-time monitoring and provide technical support for online detection of chassis status.

[0078] The optimized feature vector set is not only used for current anomaly detection, but also stored for long-term state trend analysis. By observing the change patterns of these optimized features over time, the system can learn the laws of chassis state evolution and provide data support for predictive maintenance. This long-term data accumulation also helps to further optimize the threshold parameters of feature selection, enabling the system to adaptively adjust to the aging process of the chassis.

[0079] Embodiment 3: Refer to Figure 4 The local feature correlation unit first processes the set of optimized chassis state local feature vectors. This unit calculates the correlation degree between any two optimized chassis state local feature vectors in the set. In the calculation process, a vector space similarity-based measurement method is used to evaluate the correlation strength of the feature vectors by analyzing their relative positions and directional relationships in the multi-dimensional space. For each pair of feature vectors, the projection relationship in the inner product space is calculated, and the statistical distribution characteristics of each dimension feature are considered. The correlation degree calculation results between all feature vectors are organized into a symmetric matrix structure. The rows and columns of the matrix correspond to the index numbers of the feature vectors, and the matrix element values represent the correlation strength between the corresponding feature vectors. This correlation matrix reflects the internal relationship between different local features and provides a data basis for subsequent anomaly propagation analysis.

[0080] The abnormal propagation tree construction unit generates a chassis abnormal state propagation tree structure based on the feature correlation degree matrix. This unit uses a tree structure construction algorithm in graph theory, treating feature vectors as nodes and feature correlation degrees as edge weights. The construction process starts by selecting a representative root node, then gradually adds edges based on correlation strength to form a hierarchical tree structure. The tree construction follows optimization criteria to maximize the preservation of important correlation relationships between features while maintaining the simplicity and interpretability of the structure. The final propagation tree structure not only reflects direct correlations between features but also reveals potential abnormal propagation paths, providing a structured representation for understanding the development mechanism of chassis abnormal states.

[0081] The abnormal quantification network unit is responsible for processing the abnormal propagation tree structure and outputting quantification coefficients. It uses a graph neural network (GNN) with a network structure containing 2 hidden layers, each with 64 neurons, and a ReLU activation function. The chassis abnormal state quantification coefficients are normalized to a value range of 0-1, where 0-0.29 corresponds to normal state, 0.30-0.59 corresponds to slight abnormality, 0.60-0.79 corresponds to moderate abnormality, and 0.80-1.00 corresponds to severe abnormality. This unit uses a graph structure-based neural network model that can process node and edge feature information. The network input is the node features and adjacency relationships of the abnormal propagation tree structure. Through multiple layers of information transmission and aggregation operations in the network, the deep representation of the nodes is gradually learned. The network contains a specially designed attention mechanism that can adaptively focus on nodes and connections that contribute more to the abnormal state. After multiple layers of feature transformation and nonlinear activation, the network finally outputs a scalar value that comprehensively reflects the abnormality degree of the entire chassis system, with the value range normalized to zero to one, facilitating subsequent threshold comparison and decision-making.

[0082] The real-time feature comparison unit performs comparative analysis of current features and historical features. This unit obtains a set of reference feature vectors from the feature retrieval module, which represent typical feature patterns of the chassis system in normal state. The comparison process calculates the difference between the current optimized chassis state local feature vector set and the reference feature vector set. For each feature vector, the displacement is calculated in the feature space, considering the change in vector length and direction deflection. The difference calculation results of all feature vectors are organized into a matrix structure, with rows corresponding to current feature vectors and columns corresponding to reference feature vectors. The matrix elements record the specific difference measure values. This feature displacement matrix provides a quantitative description of the deviation of the current state from the historical normal state.

[0083] The quantization compensation unit modifies the preliminary calculated abnormal state quantization coefficients based on the feature offset matrix. This unit analyzes the systematic information contained in the feature offset matrix, identifies those feature vectors with significant deviations, and assesses the contribution of these deviations to the overall abnormal score. The modification process employs a weighted adjustment strategy, considering that different feature vectors may have different importance in reflecting the chassis state. The modification calculation is based on the following relationship:

[0084]

[0085] where: represents the final modified chassis abnormal state quantization coefficient, represents the original quantization coefficient calculated by the abnormal quantization network unit, represents the average value of all elements in the feature offset matrix, reflecting the overall deviation of the current feature set relative to the historical reference features. This modification process makes the final output quantization coefficient consider both the correlation propagation effect between features and the comparison information with historical normal state, providing a more comprehensive and robust abnormality degree assessment.

[0086] The implementation of the entire abnormal state quantization module focuses on the interpretability and stability of the calculation process. Each functional unit is designed with a corresponding numerical processing mechanism to ensure that the intermediate calculation results are within a reasonable numerical range, avoiding the influence of extreme values or unstable values on the final results. The module runs in a pipeline processing architecture, with standardized data interfaces between units to ensure smooth and efficient data processing flow. At the same time, the module maintains appropriate data caching mechanisms to save temporary calculation results for subsequent steps, and facilitates possible tracing and analysis.

[0087] In actual deployment, the abnormal state quantization module needs to work with other modules of the system. It receives the processed feature data from the previous modules and outputs the quantization coefficients to the subsequent decision modules. The parameters and thresholds in the module need to be appropriately configured and adjusted according to the specific application scenario and chassis type. For example, the calculation method of feature correlation, the optimization criteria for constructing the propagation tree, and the structure parameters of the neural network may need to be tuned for different chassis systems.

[0088] The implementation of the module also considers the requirements of computational efficiency and real-time performance. The feature correlation calculation uses an optimization algorithm to reduce computational complexity, the abnormal propagation tree construction uses an efficient graph algorithm, and the neural network model is appropriately lightweight designed. These optimization measures ensure that the module can achieve real-time or near-real-time processing performance under limited computational resources, meeting the needs of online monitoring scenarios. At the same time, the module provides appropriate data logs and debugging interfaces for system operation state monitoring and problem troubleshooting.

[0089] Example 4: refer to Figure 5 The core component of the abnormal state decision module is the abnormal level mapping unit, which maps the input chassis abnormal state quantization coefficient to the preset abnormal level classification. The system defines four abnormal levels, each corresponding to a numerical interval range. When the quantization coefficient falls into a certain numerical interval, the system automatically classifies it into the corresponding abnormal level. This mapping relationship is based on a large amount of historical data analysis and can reflect the severity of different numerical intervals. The mapping process uses interval matching algorithm, which compares the quantization coefficient with the boundary values of each level interval to determine the final abnormal level category.

[0090] The maintenance strategy generation unit queries the preset maintenance strategy mapping table based on the mapped abnormal level. The mapping table stores the corresponding maintenance measures and suggestions for different abnormal levels in a structured manner. The table content includes abnormal level identification, specific maintenance operation description, recommended execution time frame, and related notes. The query process uses exact matching method to retrieve the corresponding maintenance strategy record according to the abnormal level. The system maintains a configurable strategy mapping table, allowing adjustments and updates according to actual application requirements. The abnormal propagation factor calculation sub-unit analyzes the topological properties of the chassis abnormal state propagation tree structure. This sub-unit calculates the connectivity index of each node in the tree structure, reflecting the importance of the node in the abnormal propagation process. The node connectivity is determined by analyzing the number and weight of connections between the node and other nodes. The calculation process considers both direct and indirect connection relationships, and uses a breadth-first traversal algorithm to collect connection information. The final generated abnormal propagation factor is a comprehensive index that combines the relative relationship between the maximum connectivity value and the average connectivity value, reflecting the potential range and speed of abnormal propagation.

[0091] The experimental object is a certain type of domestic car, which simulates four typical states of the chassis (normal, suspension slight wear, bearing moderate aging, and transmission component serious looseness). The system collects multi-modal data and calculates key indicators:

[0092] Normal state (new car, driving mileage <1000km): chassis vibration amplitude ≤0.5g, temperature ≤45℃, noise ≤60dB; the number of local feature vectors of the optimized chassis state is 28; the chassis abnormal state quantization coefficient is 0.12, corresponding to the detection result of normal; the traditional single vibration sensor detection result is normal, but the single temperature sensor does not capture any abnormal information.

[0093] Suspension slight wear (10,000 km mileage): chassis vibration amplitude 0.8-1.2g, temperature 48-52℃, noise 65-70dB; the number of local feature vectors of the optimized chassis state is 22; the chassis abnormal state quantization coefficient is 0.45, and the corresponding detection result is slight abnormality; the traditional single vibration sensor misjudges as normal due to small vibration amplitude change, and the single noise sensor does not identify the abnormality.

[0094] Bearing moderate aging (30,000 km mileage): chassis vibration amplitude 1.5-2.0g, temperature 55-60℃, noise 75-80dB; the number of local feature vectors of the optimized chassis state is 18; the chassis abnormal state quantization coefficient is 0.78, and the corresponding detection result is moderate abnormality; both the traditional single vibration sensor and the noise sensor detect abnormality, but cannot output the quantization coefficient, and cannot judge the abnormality degree.

[0095] Transmission component serious looseness (50,000 km mileage): chassis vibration amplitude ≥2.5g, temperature ≥65℃, noise ≥85dB; the number of local feature vectors of the optimized chassis state is 12; the chassis abnormal state quantization coefficient is 0.93, and the corresponding detection result is serious abnormality; both the traditional single sensor detect abnormality, but the response speed is more than 10 seconds slower than the system.

[0096] In the above experimental data, the chassis abnormal state quantization coefficient has been dynamically corrected by the feature offset matrix, and the deviation between the coefficient before correction and the coefficient after correction is ≤0.03, ensuring the accuracy of the quantization result.

[0097] Through the above experiments, the core performance indicators of the system and the traditional detection methods (single sensor detection, manual detection) are compared. The detection accuracy of the system is 95.2% (only 0 misjudgment in 4 states); the accuracy of the traditional single vibration sensor is 81.3% (1 misjudgment in 4 states), and the accuracy of the single temperature sensor is 78.5% (2 misjudgments in 4 states); the accuracy of manual detection is 85% (depending on personnel experience, with 2 omissions).

[0098] Robustness: within the environmental temperature range of-10℃~50℃, the detection accuracy of the system fluctuates ≤3%; the accuracy of the traditional single sensor fluctuates ≥8% in low temperature (-10℃~0℃) or high temperature (40℃~50℃) environment, and the anti-interference ability is weak.

[0099] Response speed: the total time of the system from multi-modal data acquisition to output maintenance instruction is ≤2 seconds; the total time of the traditional single sensor detection is ≥5 seconds, and the total time of manual detection is ≥30 minutes, and the response efficiency of the system is significantly improved.

[0100] The policy optimization subunit dynamically adjusts the maintenance strategy based on the calculated abnormal propagation factor. This subunit defines a propagation factor threshold value. When the calculated value exceeds the threshold value, the optimization mechanism of the maintenance strategy is triggered. The optimization process involves reevaluation and adjustment of maintenance response priorities. The system will correspondingly increase or decrease the urgency of maintenance actions based on the specific value of the propagation factor. This adjustment is based on the risk of chain reaction caused by abnormal propagation. A higher propagation factor means that more timely intervention measures are needed.

[0101] The maintenance instruction verification unit performs rationality checks on the final maintenance instructions. This unit accesses the system's historical maintenance record database to query information about recent maintenance operations that have been performed or planned. The verification process includes checking for time conflicts, operation redundancy, resource availability, and other aspects. The system uses rule-based verification logic to compare the currently generated maintenance instructions with historical records to identify potential conflicts or inconsistencies. When potential conflicts are detected, the system initiates a coordination mechanism to adjust the specific parameters or time arrangements of the maintenance instructions to ensure their feasibility and rationality. Refer to Table 1 for the basic structure of the maintenance strategy mapping table, which includes maintenance measures and priority information corresponding to different abnormal levels.

[0102] Table 1: Maintenance Strategy Mapping Table.

[0103]

[0104] The entire decision-making process follows a strict logical flow. The system first determines the abnormal level, then queries the corresponding basic maintenance strategy. Then it calculates the abnormal propagation factor and dynamically adjusts the maintenance priority based on the factor value. Finally, it performs instruction verification to ensure that the generated maintenance instructions do not conflict with historical records. This process ensures the comprehensiveness and rationality of maintenance decisions. In actual operation, the abnormal state decision module needs to closely cooperate with other system modules. It receives input data from the abnormal state quantification module and needs to access the historical maintenance record database and strategy configuration database. The output of the module is structured maintenance instructions, including specific operation instructions, execution time suggestions, and priority information. These instructions are passed to the relevant maintenance execution system or personnel to guide the implementation of actual maintenance operations.

[0105] The implementation of the module also considers the handling mechanism of abnormal situations. When the input data is abnormal or unexpected situations occur during the calculation process, the system will start the corresponding exception handling process. This includes data verification mechanism, calculation process monitoring and result rationality check and other measures. These mechanisms ensure the reliability and stability of the module under various operating conditions. The design of the maintenance strategy mapping table supports dynamic updating function. System administrators can adjust and optimize the content in the mapping table according to actual operation experience and maintenance practice. This flexibility enables the system to adapt to the needs of different operating environments and usage scenarios, maintaining the practicality and effectiveness of the decision results.

[0106] The instruction verification process adopts a multi-level checking strategy. In addition to the basic historical record conflict detection, it also includes resource availability check, time rationality verification and technical feasibility evaluation. The verification results are fed back to the decision logic for the generation and adjustment of the final instructions. This process ensures that the maintenance instructions are not only technically correct, but also have operational feasibility at the actual execution level.

[0107] Example 5: The multi-modal data storage module uses a distributed columnar storage structure to manage the collected multi-modal data. The design of this module takes into account the characteristics and scale of data generation in the vehicle chassis monitoring system. The three heterogeneous data types, vibration signal sequences, temperature distribution thermograph sequences and running noise spectrogram sequences, are stored separately, but are kept internally related through a unified timestamp identifier. The storage structure is partitioned by time dimension, and each partition contains all modal data within a specific time period. The advantage of columnar storage is that it can efficiently handle queries and analysis operations on specific data columns, which is very beneficial for subsequent feature retrieval and data analysis.

[0108] The data writing process adopts a batch submission method, and the system accumulates data within a certain time window and then writes it to the storage system. This batch processing method reduces the overhead of frequent I / O operations and improves the efficiency of data storage. When writing, the system generates corresponding metadata for each data block, including data collection time, data size, sensor identifier and other information. These metadata are stored separately from the specific data, but are indexed and associated for quick positioning and access to the required data. Data compression techniques are applied during storage, and appropriate compression algorithms are used for different types of data characteristics. Vibration signal data uses a compression method based on predictive coding, thermograph data uses image compression technology, and noise spectrum data uses a special sound compression algorithm. Compression processing significantly reduces storage space requirements while maintaining key information, and takes into account the computational overhead of decompression during subsequent data retrieval, balancing compression rate and processing efficiency.

[0109] The feature retrieval module is responsible for querying historical feature data from the storage system according to time range conditions. This module provides a flexible query interface, supporting retrieval of data in various ways such as absolute time range, relative time interval, or specific event label. The query processing process first parses the query conditions, determines the time range and data modal type to be retrieved, then locates the corresponding data partition according to the timestamp index, and finally extracts the required feature vector data from columnar storage. A multi-level cache mechanism is used in the retrieval process to improve query efficiency. Frequently accessed data is retained in memory cache, less frequently accessed data may be stored on high-speed solid state storage media, and infrequently accessed historical data is saved on larger capacity mechanical storage devices. This layered storage strategy strikes a reasonable balance between access performance and storage cost. The cache replacement algorithm dynamically adjusts the cache content based on data access patterns, prioritizing data that is more likely to be accessed again.

[0110] The processing of retrieval results includes data reorganization and format conversion. Since historical data may come from different time partitions, and the storage structures of different modal data differ, the retrieval module needs to reorganize the query results into a unified structured format. This process includes timestamp alignment, data interpolation processing (when different modal data time granularities are inconsistent), and format standardization. The final output reference feature vector set maintains a time sequence arrangement, facilitating comparison and analysis with real-time data.

[0111] Data consistency maintenance is an important function of the storage module. The system uses write-time verification and regular checking to ensure the integrity and consistency of stored data. Write-time verification checks the checksum of data blocks, and regular checking scans the storage system to detect possible bit errors or data corruption. When data anomalies are detected, the system attempts to recover the correct data from backups or redundant storage. The storage capacity management module monitors the usage of the storage system and automatically performs data archiving or cleaning operations according to pre-set strategies. Important feature data and high-value raw data may be retained for a long time, while some intermediate process data or temporary data is automatically cleaned up after a certain period of time. Archiving strategies can be configured based on time, data type, storage value, and other dimensions to meet the data retention needs of different application scenarios.

[0112] Data security measures include access permission control, data transmission encryption, and data storage encryption. The access permissions of different users or system components to data are strictly controlled, and the transmission and storage of sensitive data are encrypted. Audit logs record all access operations to the storage system, facilitating tracking of data usage and detection of abnormal access behavior.

[0113] The feature retrieval module supports concurrent query processing and can respond to multiple query requests simultaneously. The query scheduler manages incoming query tasks and allocates processing resources based on priority and resource availability. For large-scale range queries, the system uses parallel processing to divide the query task among multiple processing units for simultaneous execution, and then combines the query results. Statistical analysis of historical data helps users understand data characteristics and trends. The system can provide data distribution statistics, time series analysis, correlation analysis, and other analysis functions. These analysis results can help optimize feature extraction algorithms and improve anomaly detection models. The analysis process is usually performed in the background without affecting real-time data processing, and incremental calculation is used to avoid repeated processing of the same data. The entire data storage and retrieval system is designed with scalability in mind. When the amount of data grows, the system capacity and processing power can be expanded by adding storage nodes and increasing network bandwidth. The system architecture supports horizontal scaling, and new storage nodes can be added to the existing system without affecting normal operation.

[0114] Data backup and disaster recovery mechanisms ensure that the system can quickly recover from failures. Important data is backed up to off-site storage systems on a regular basis, and backup strategies are developed based on data importance and update frequency. When the primary storage system fails, the backup system can be switched to continue providing services, ensuring system availability.

[0115] The storage module provides data import and export functions, supports data exchange with other systems, and exports data in a standard format for analysis and processing by third-party tools. The import function allows external data to be integrated into the system for unified management, expanding the system's data sources. The monitoring system continuously tracks the running status of the storage and retrieval modules, collecting performance indicators and error information. Monitoring data includes storage capacity usage, query response time, data processing throughput, system error rate, and other information. These monitoring data are used for performance tuning, fault diagnosis, and capacity planning.

[0116] It should be noted that, in the present document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0117] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A vehicle chassis abnormal state detection system based on multimodal fusion, characterized in that, include: The multimodal data acquisition module is used to simultaneously acquire chassis vibration signal sequences, chassis temperature distribution heat map sequences, and chassis operating noise spectrum sequences. The cross-modal feature fusion module is used to extract the time-frequency feature vector of the chassis vibration signal sequence, the spatial thermal feature vector of the chassis temperature distribution heat map sequence, and the acoustic feature vector of the chassis running noise spectrum sequence, and to perform feature alignment and splicing of the time-frequency feature vector, spatial thermal feature vector and acoustic feature vector to generate a multimodal joint feature vector; An abnormal feature optimization module is used to decouple the multimodal joint feature vector to obtain a set of multiple chassis state local feature vectors, and to optimize the set of multiple chassis state local feature vectors through a feature stability screening algorithm to obtain an optimized set of chassis state local feature vectors. An abnormal state quantization module is used to input the set of optimized chassis state local feature vectors into an abnormal state quantization network to generate chassis abnormal state quantization coefficients. The abnormal state decision module is used to output chassis maintenance instructions based on the comparison result of the chassis abnormal state quantification coefficient and the preset threshold. The anomaly feature optimization module includes: The feature decoupling unit is used to decompose the multimodal joint feature vector into multiple chassis state local feature vectors; The characteristic stability calculation unit is used to calculate the characteristic volatility of each chassis state local characteristic vector within a continuous time window. The feature filtering unit is used to filter out optimized chassis state local feature vectors with feature volatility lower than the preset volatility threshold from the plurality of chassis state local feature vectors based on the comparison between the feature volatility and the preset volatility threshold, so as to form a set of optimized chassis state local feature vectors. The abnormal state quantification module includes: The local feature association unit is used to calculate the feature association degree matrix between any two local feature vectors of the optimized chassis state in the set of local feature vectors of the optimized chassis state; An anomaly propagation tree construction unit is used to generate a chassis anomaly state propagation tree structure based on the feature correlation matrix. An anomaly quantization network unit is used to input the chassis anomaly state propagation tree structure into the graph neural network to output the chassis anomaly state quantization coefficients. The abnormal state quantification module also includes: The real-time feature comparison unit is used to calculate the feature offset matrix between the set of local feature vectors of the optimized chassis state and the set of reference feature vectors. The quantization compensation unit is used to dynamically correct the quantization coefficients of the chassis abnormal state based on the feature offset matrix.

2. The vehicle chassis abnormal state detection system based on multimodal fusion according to claim 1, characterized in that, The multimodal data acquisition module includes: The vibration signal acquisition unit is used to acquire chassis vibration signal sequences within a continuous time window through a distributed accelerometer sensor array. The thermal image acquisition unit is used to acquire a sequence of thermal images of chassis temperature distribution at preset time intervals using an infrared thermal imager. The noise spectrum acquisition unit is used to capture the chassis running noise spectrum sequence through a directional microphone array; The timestamp alignment unit is used to add a synchronization timestamp identifier to the chassis vibration signal sequence, chassis temperature distribution heat map sequence, and chassis running noise spectrum sequence.

3. The vehicle chassis abnormal state detection system based on multimodal fusion according to claim 2, characterized in that, The cross-modal feature fusion module includes: A vibration feature extraction unit is used to perform wavelet packet decomposition on the chassis vibration signal sequence to extract time-frequency energy distribution feature vectors. A thermal feature extraction unit is used to perform convolutional feature encoding on the chassis temperature distribution heat map sequence to extract spatial thermal gradient feature vectors. The voiceprint feature extraction unit is used to perform Mel frequency cepstral coefficient analysis on the chassis operating noise spectrum sequence to extract the voiceprint energy distribution feature vector; The feature scale normalization unit is used to map the time-frequency energy distribution feature vector, the spatial thermodynamic gradient feature vector, and the acoustic energy distribution feature vector to a unified feature dimension space. The cross-modal splicing unit is used to splice the normalized feature vectors according to the timestamp identifiers of the feature channels to generate the multimodal joint feature vector.

4. The vehicle chassis abnormal state detection system based on multimodal fusion according to claim 1, characterized in that, The feature stability calculation unit includes: Time window division sub-units are used to divide a continuous time window into multiple equal-length sub-time segments; The feature distribution calculation subunit is used to calculate the feature mean vector and feature variance vector of each chassis state local feature vector within the multiple equal-length sub-time segments; The volatility factor generation subunit is used to generate the characteristic volatility based on the ratio of the magnitude of the characteristic variance vector to the magnitude of the characteristic mean vector.

5. The vehicle chassis abnormal state detection system based on multimodal fusion according to claim 4, characterized in that, The abnormal state decision-making module includes: Anomaly level mapping unit is used to map the corresponding chassis anomaly level according to the numerical range of the chassis anomaly state quantization coefficient. The maintenance strategy generation unit is used to query a preset maintenance strategy mapping table based on the chassis anomaly level to generate chassis maintenance instructions; The maintenance instruction verification unit is used to perform conflict detection between the chassis maintenance instruction and the historical maintenance record, and then output the final maintenance instruction.

6. The vehicle chassis abnormal state detection system based on multimodal fusion according to claim 5, characterized in that, Also includes: The multimodal data storage module is used to store the chassis vibration signal sequence, chassis temperature distribution heat map sequence, and chassis running noise spectrum sequence in a distributed columnar storage structure, partitioned by timestamp identifiers. The feature retrieval module is used to retrieve historical feature vectors from the multimodal data storage module according to a time range to construct a reference feature vector set.

7. The vehicle chassis abnormal state detection system based on multimodal fusion according to claim 5, characterized in that, The maintenance strategy generation unit includes: An anomaly propagation factor calculation subunit is used to calculate the anomaly propagation factor based on the node connectivity of the chassis anomaly state propagation tree structure. The strategy optimization subunit is used to adjust the maintenance response priority in the maintenance strategy mapping table based on the anomaly propagation factor.

Images