CN-122020445-A - Intelligent diagnosis method for light truck faults driven by combination of data and model

Abstract

The invention discloses a data and model combined driving light card fault intelligent diagnosis method which comprises the steps of acquiring and preprocessing vehicle-mounted CAN data, extracting high-dimensional implicit characteristics based on topology data analysis, establishing an integrated model for hybrid reasoning and monitoring the vehicle state in real time. And extracting persistent coherent features through topology data analysis, generating multi-scale continuous image codes, fusing a physical grouping mechanism of a vehicle subsystem and Gaussian kernel density estimation, and capturing a high-dimensional manifold topological structure. The method comprises the steps of constructing a double-channel convolutional neural network, fusing topological image features and time domain statistical features, integrating self-encoder compression, LSTM time sequence modeling and LightGBM decision output, realizing the identification of micro-fault features, and ensuring reasoning delay precision through model parallel calculation. And a dynamic diagnosis rule feedback mechanism and a Word2 Vec-driven fault knowledge graph are adopted, and an edge computing node is combined to monitor the feature vector deviation threshold value in real time, so that data noise and sample imbalance are overcome.

Inventors

• WANG NING
• Nie Liaodong

Assignees

• 同济大学

Dates

Publication Date

20260512

Application Date

20251219

Claims (7)

1. 1. The intelligent diagnosis method for the light truck fault driven by the combination of the data and the model is characterized by comprising the following steps of: step 1, acquiring basic parameters of a vehicle equipped with CAN equipment; step 2, acquiring and preprocessing CAN bus data; step 3, providing a mixed diagnosis framework for analyzing the TDA by fusing topological data, mapping time sequence data into high-dimensional point cloud through persistent coherent feature extraction, calculating topological features of quantized data manifold by utilizing Vietoris-Rips complex form, and generating a persistent graph PERSISTENCE DIAGRAM to record the service life of the topological features by birth-death coordinates; step 4, converting the space-time distribution of the topological features into a multi-scale gray level image which can be processed by a deep learning model by a Gaussian kernel density estimation method, wherein the highlight region corresponds to a strong robustness topological structure; step 5, establishing a lightweight double-channel classification model to combine topological features and traditional statistical features, wherein the topological feature channels excavate space modes in the continuous images through depth separable convolution; Step 6, a vehicle fault risk hybrid diagnosis model is established, the features extracted in the step 5 are used as input, low-dimensional features are extracted through a self-encoder, then time sequence features are extracted based on a long-short time memory network Long Short Term Memory and an LSTM, feature fusion is carried out by combining a predefined fault recognition rule, and finally the light-weight gradient lifting framework LIGHT GRADIENT Boosting Machine and LightGBM are input for processing to obtain final output; and 7, early warning the vehicle faults in real time, and carrying out feedback optimization according to the fault recognition rules and the model results.
2. 2. The intelligent diagnosis method for the fault of the light truck driven by the combination of the data and the model according to claim 1, wherein the method of the step 2 is as follows: 2.1 Acquiring vehicle-mounted CAN bus data; 2. dynamically dividing data into historical data and online data according to sliding window segmentation, namely, an original CAN data stream ( ) Dividing according to a sliding window; 3. dividing the historical data obtained in the step 2.2 into normal data and fault data according to the fault code; 4. Grouping the data obtained in the step 2.3 according to the dynamics characteristics of the vehicle, wherein the data comprise a power system signal, an emission system signal and a chassis system signal; 5. Z-score normalization is performed on the data in each window according to the signal type, and structured input is constructed for topology feature extraction: ; Wherein, the Is normalized to the first The dimension signal is used to determine the dimension of the object, , Is the first The mean and standard deviation of the dimensional signal within the window.
3. 3. The intelligent diagnosis method for the fault of the light truck driven by the combination of the data and the model according to claim 1, wherein the method of the step 3 is as follows: 3.1 Point cloud construction, namely constructing Gao Weidian cloud by time delay embedding for each window data: ; Wherein, the In order to embed the dimensions in-line, In order to delay the step size, 3.2 Vietoris-Rips complex computing, namely computing point cloud Vietoris-Rips complex form, set its maximum distance threshold , Generating continuous coherent group for Euclidean distance standard deviation of point cloud And (3) with , For connectivity In the structure of the ring, the ring is provided with a plurality of grooves, 3.3 Persistent graph generation, namely, extracting birth-death intervals of topological features Construction of persistent graphs Wherein In order to be of a coherent dimension, Is the number of features.
4. 4. The intelligent diagnosis method for the fault of the light truck driven by the combination of the data and the model according to claim 3, wherein the method of the step4 is as follows: 4.1 Kernel density estimation for each point Applying a gaussian kernel function: ; Wherein, the As a gaussian kernel bandwidth parameter, Normalized to image pixel coordinates In the interval of the time period, 4.2 Multiscale fusion, namely respectively generating continuous images for global point cloud and subsystem grouping point cloud And (3) with Spliced into multi-channel input , 4.3 Persistent graph based on step 4.2 results Converted into a 64 x 64 gray scale image.
5. 5. The intelligent diagnosis method for the fault of the light truck driven by the combination of the data and the model according to claim 3, wherein the method of the step5 is as follows: 5.1 Designing a double-channel Convolutional Neural Network (CNN) structure; 2. topology feature channel input The high-dimensional feature vector is output through multi-layer depth separable convolution; 3. extracting statistics in window as auxiliary features Mapping to 128 dimensions through a full connection layer; 4. feature fusion and classification, namely splicing two channel features, and outputting fault probability distribution through a Softmax layer: ; Wherein, the , For the topological feature and the conventional feature vector, As the number of categories of faults, , Is classifier weight and bias.
6. 6. The intelligent diagnosis method for the fault of the light truck driven by the combination of the data and the model according to claim 1, wherein the method of the step 6 is as follows: 6.1 constructing a self-encoder comprising an encoder, a decoder and a loss function, calculated as follows: (1) Encoder (Encoder) for inputting data in high dimension Mapping to a low-dimensional potential spatial representation , ; (2) Decoder (Decoder) from potential representation Reconstructing original input , ; ; (3) The loss function minimizes the mean square error between the input and the reconstruction, ; Wherein, the In order to input the vector(s), For the purpose of a potential representation, In order for the encoder parameters to be the same, In order for the decoder to be parameter-able, In order to activate the function, For the number of samples to be taken, 6.2 An LSTM model comprising an input gate, a forget gate and an output gate is constructed to control the updating of the memory unit, and the method is calculated as follows: (1) A forgetting gate (Forget Gate) to decide which information to discard from the cell state, ; (2) An Input Gate (Input Gate) determines which new information is to be stored to the cell state, ; (3) Candidate cell state (CANDIDATE CELL STATE) by generating a new candidate value, ; (4) Cell status Update (CELL STATE Update) to Update the cell status, ; (5) An Output Gate (Output Gate) for determining which information to Output, ; (6) A hidden state Output (HIDDEN STATE Output) for generating a current hidden state ; Wherein, the Is a time step Is used to determine the vector of the input vector, In the hidden state of the previous time step, The state of the cells being the previous time step, Is a weight matrix , Is a bias vector , Activating a function for Sigmoid , For hyperbolic tangent activation function , For element-by-element multiplication (Hadamard product), For the vector concatenation operation, 6.3 And designing LightGBM a module, improving the training speed of the fault risk identification and early warning model, and calculating as follows: (1) For the regression task LightGBM is calculated by minimizing the Mean Square Error (MSE): ; (2) For classification tasks LightGBM optimize cross entropy loss: ; 6.4 Performing model fusion, aligning the compact representation extracted in the step 6.1 with the time sequence features mined in the step 6.2 in a hidden space, inputting the compact representation into the step 6.3 to perform multi-granularity decision, and calculating as follows: extracting from the encoder features: ; LSTM timing feature extraction: ; Feature fusion: ; LightGBM outputs: 。
7. 7. the intelligent diagnosis method for the fault of the light truck driven by the combination of the data and the model according to claim 1, wherein the method of the step 7 is as follows: 7.1 Collecting diesel light-truck CAN message data in real time, wherein the data comprise engine operation parameters, a transmission line vibration spectrum, a post-processing device state and an on-board diagnostic (OBD) fault code stream, and performing time sequence alignment and noise cleaning on the original data through sliding window filtering and working condition slicing; 7.2 A Word2vec Word embedding model is introduced, a fault description text is converted into a vector representation, and text characteristics are analyzed to help identify keywords and context information related to different fault types, so that a basis is provided for subsequent fault classification; 7.3 Constructing a historical fault type database based on expert knowledge, wherein the database contains descriptions and features of various known fault types and can be used as a reference for model training and verification, so that the model can better understand and classify new fault data, and the database is updated regularly to reflect the new fault types and modes; 7.4 The convolutional neural network is utilized to strengthen the attention to key features, and the historical fault data set is used for supervised learning, so that the model automatically classifies new fault data; 7.5 The method comprises the steps that an edge computing node is deployed to monitor the state of a vehicle in real time, a fault early warning model is combined, when a feature vector deviates from a working condition reference threshold value, fault pre-diagnosis based on knowledge graph reasoning is triggered, fault information is automatically diagnosed and positioned when faults occur, and relevant information is recorded for subsequent analysis and processing; 7.6 After the fault occurs, the system determines the potential cause of the fault by analyzing the historical fault type database and expert knowledge, and timely informs relevant users and maintenance departments after determining the cause, and appropriate maintenance and repair measures are adopted, and the notification comprises suggested maintenance steps, so that the fault processing efficiency is improved.

Description

Intelligent diagnosis method for light truck faults driven by combination of data and model Technical Field The invention relates to a fault diagnosis method, in particular to a light card fault intelligent diagnosis method driven by data and a model in a combined mode. Background With the improvement of the automation level of modern industrial technology, the structure of an automobile system is more and more complex, and various parts are tightly coupled. Any fault can cause chain reaction, influence the normal operation of the whole vehicle, and even can cause serious traffic accidents, so that the fault diagnosis and early warning technology of the vehicle is very important. The existing vehicle fault diagnosis technology is mainly based on a knowledge base and a deep learning model and is mainly used for fault classification and fault cause analysis. Deep learning models typically require a significant amount of computational resources and time, making early prediction and prevention difficult. Expert models based on knowledge bases generally need to preset expert rules or fault entries, have poor generalization under different models and working conditions, and are difficult to directly apply among different systems. In addition, shallow feature extraction methods have difficulty characterizing micro-fault features in vehicle systems, resulting in accuracy in handling the micro-fault features being compromised. The search finds that: The Chinese patent application No. 202311348964.0, namely a vehicle fault diagnosis method, a vehicle fault diagnosis device, a vehicle and a storage medium, obtains a corresponding fault cause by obtaining a vehicle fault phenomenon and matching the vehicle fault phenomenon with expert rules in a knowledge base, and pushes fault information to a user terminal. The defects are that expert rules and specific fault phenomena are required to be preset, the fault types which are not preset cannot be identified, and diagnosis and early warning cannot be carried out before the fault occurs. The China patent application No. 202211639341.4, analyzing method and device, electronic equipment and storage medium for vehicle faults, obtains at least one target fault entry by analyzing a vehicle controller signal and matching with a preset fault entry table, and sends a fault cause and recommended treatment measures to a vehicle. The defect is that a fault entry table needs to be preset, the generalization performance under different models and working conditions is poor, and the fault entry table belongs to post diagnosis. The Chinese patent application No. 202410367317.2, "vehicle fault diagnosis method, device, equipment and storage medium" determines suspected faults of a vehicle according to vehicle fault description and fault dictionary library sent by a user, performs fault location according to a vehicle fault code and suspected faults, and determines fault components by analyzing state quantities of the suspected fault components. The method has the defects that a fault dictionary library needs to be predefined, fault description actively sent by a user is relied on, and data dependence is strong. The Chinese patent application No. 202310810136.8, namely a fault diagnosis method, a device, electronic equipment and a storage medium, solves the probability of each vehicle fault through a regression equation of the fault, and obtains a fault diagnosis result according to the fault probability and an electronic control unit corresponding to a diagnosis fault code. The method has the defects that the regression model belongs to a shallow feature extraction method, tiny fault features in a vehicle system are difficult to characterize, regression equations of different faults are different, and real-time performance of diagnosing a plurality of faults is poor. The Chinese patent application No. 202210962952.6, yun Bian-based cooperative automobile fault diagnosis method, system and intelligent automobile, designs a long-short-time memory stack self-coding neural network with a brand-new architecture by acquiring historical state parameters of an electric drive module, and screens an optimal model through a vehicle fault diagnosis model evaluation algorithm. The method has the disadvantages that the deep learning model requires a large amount of computing resources and time, has higher quality requirement on the training history data, and has strong data dependence. The Chinese patent application No. 202410485042.2, a vehicle fault analysis method, device, equipment and storage medium, identifies the time of occurrence of a fault of a vehicle and sends vehicle operation data of time periods before and after the fault to a vehicle networking platform for analysis. The defect is that the fault diagnosis technology is post diagnosis and can not realize early prediction and prevention. The Chinese patent application No. 202011557530.8, namely a vehicle fault prediction model generation m