Intelligent fault diagnosis method for automobile wire harness detection platform

Abstract

The invention relates to the technical field of automobile wire harness detection, and discloses an intelligent fault diagnosis method for an automobile wire harness detection platform, and the method comprises the following steps: collecting multi-source data through the automobile wire harness detection platform, and carrying out the data enhancement preprocessing; constructing a multi-branch attention mechanism-based multi-source data fusion and parallel lightweight one-dimensional convolutional neural network model; extracting multi-source data features by using a lightweight one-dimensional convolutional neural network in a multi-branch attention mechanism-based multi-source data fusion and parallel lightweight one-dimensional convolutional neural network model; fusing features by using a multi-branch attention mechanism; and outputting a final fault diagnosis result by using a Softmax classifier. According to the invention, voltage, resistance, temperature and air pressure signals are processed through a parallel network structure, feature adaptive weighted fusion is realized by using a channel attention mechanism and a multi-branch attention mechanism, the complementarity of multi-source data is fully mined, and the fault diagnosis accuracy of an automobile wire harness detection platform is improved.

Description

Technical Field

[0001] This invention relates to the field of automotive wiring harness testing technology, specifically to an intelligent fault diagnosis method for an automotive wiring harness testing platform. Background Technology

[0002] Automotive wiring harnesses form the core of a vehicle's electrical network, and their safety and reliability directly determine the operational stability of the entire vehicle system. With the increasing level of automotive electronics, the structure of automotive wiring harnesses is becoming increasingly complex, and the types and quantities of signals they carry are growing exponentially. Automotive wiring harness testing platforms are crucial equipment for ensuring the quality of automotive wiring harnesses. The main task of these platforms is to identify various faults that occur in the wiring harness during production or use in a timely and accurate manner.

[0003] Existing automotive wiring harness testing technologies typically rely on a single type of signal parameter for judgment. Common testing methods primarily determine the continuity of the wiring harness by measuring resistance values ​​or assess its withstand voltage performance by measuring voltage values. However, relying solely on a single type of signal parameter cannot comprehensively reflect the true health status of automotive wiring harnesses under complex environments. In actual operation, automotive wiring harnesses are affected by environmental factors such as temperature and air pressure. Single-dimensional signal data often contains incomplete information and cannot effectively capture early, subtle fault characteristics under multi-physical field coupling, leading to missed detections or misjudgments by automotive wiring harness testing platforms when faced with complex fault modes.

[0004] On the other hand, traditional automotive wiring harness fault diagnosis methods often employ fixed thresholds or shallow machine learning models based on manual feature extraction. Fixed threshold methods lack adaptability to dynamic signal changes and struggle to detect intermittent faults with small signal fluctuations. Manual feature extraction methods rely on expert experience, are cumbersome, and have weak generalization capabilities. While some existing technologies have begun to introduce deep convolutional neural networks for automatic feature extraction, standard convolutional neural network models typically have a large number of parameters and computational demands. This large number of parameters and computational demands places extremely high demands on hardware computing resources, making it difficult to achieve efficient parallel processing and real-time diagnosis on resource-constrained embedded terminals of automotive wiring harness testing platforms.

[0005] Furthermore, in scenarios involving multi-source data processing, existing technologies lack efficient data fusion mechanisms. Current technologies typically employ simple data concatenation or fixed-weight averaging strategies to process multi-source data. These simple concatenation or fixed-weight averaging strategies ignore the heterogeneity of data between different data sources and the differences in their contribution to fault diagnosis. Under different fault categories, the importance of voltage, resistance, temperature, and air pressure signals dynamically changes. The lack of adaptive recalibration of the importance of feature channels and the lack of dynamic analysis of the global correlation of multi-source data prevent existing technologies from fully utilizing the complementary advantages of multi-source data, limiting further improvements in the fault diagnosis accuracy of automotive wiring harness testing platforms.

[0006] Therefore, this invention proposes an intelligent fault diagnosis method for automotive wiring harness testing platforms to address the shortcomings of existing technologies. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an intelligent fault diagnosis method for automotive wiring harness testing platforms. This method solves the problems of existing technologies, such as the difficulty in accurately identifying weak faults under multi-physical field coupling due to reliance on single-type signal parameters, the high computational complexity of standard convolutional neural network models failing to meet the real-time detection requirements of embedded terminals, and the rigidity of multi-source data fusion strategies failing to fully utilize data complementarity, thus limiting the improvement of fault diagnosis accuracy.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] This invention provides an intelligent fault diagnosis method for an automotive wiring harness testing platform, comprising the following steps:

[0010] Multi-source data is collected using an automotive wiring harness inspection platform, and the data undergoes data augmentation preprocessing. A multi-source data fusion and parallel lightweight one-dimensional convolutional neural network model based on a multi-branch attention mechanism is constructed, denoted as the MBAM-LM-1DCNN model. The preprocessed multi-source data is input into the MBAM-LM-1DCNN model. Feature extraction is performed using the lightweight one-dimensional convolutional neural network in the MBAM-LM-1DCNN model, denoted as the LM-1DCNN. The extracted features are then fused using the multi-branch attention mechanism in the MBAM-LM-1DCNN model, denoted as the MBAM. Finally, the Softmax classifier in the MBAM-LM-1DCNN model outputs the fault category probability of the automotive wiring harness inspection platform based on the fused features, and the category with the highest fault category probability is taken as the final fault diagnosis result.

[0011] In the aforementioned intelligent fault diagnosis method for an automotive wiring harness testing platform, multi-source data includes voltage signals, resistance signals, temperature signals, and air pressure signals. The automotive wiring harness testing platform collects multi-source data based on resistance detection, voltage detection, temperature testing, and airtightness testing functions. The resistance detection function is used to detect the continuity of the wiring harness, the voltage detection function is used to detect the withstand voltage value of the wiring harness, the temperature testing function is used to detect the temperature resistance characteristics of the wiring harness, and the airtightness testing function is used to detect the sealing performance of the wiring harness. The MBAM-LM-1DCNN model adopts a parallel network structure, containing a first data source processing branch and a second data source processing branch. The first data source processing branch processes the voltage signal, and the second data source processing branch processes the resistance signal. Both the first and second data source processing branches use LM-1DCNN.

[0012] In the aforementioned intelligent fault diagnosis method for an automotive wiring harness inspection platform, the LM-1DCNN consists of an input layer, convolutional layers, batch normalization layers, activation function layers, pooling layers, and flattening layers. The convolutional layers use convolutional kernels to perform sliding window calculations on the input signal, extracting local features through convolution operations. The batch normalization layer standardizes the output of the convolutional layers. The activation function layer uses the ReLU activation function. The pooling layer uses max pooling to downsample the feature map. The flattening layer, located at the end of the LM-1DCNN, unfolds the multi-dimensional feature map into a one-dimensional feature vector, which serves as the input to the multi-branch attention mechanism.

[0013] In the aforementioned intelligent fault diagnosis method for an automotive wiring harness inspection platform, a channel attention mechanism is introduced into a lightweight one-dimensional convolutional neural network. The channel attention mechanism processes the feature map output by the one-dimensional convolutional neural network. The channel attention mechanism includes compression, activation, and feature recalibration operations. The compression operation uses global average pooling to compress the feature map spatially, reducing each two-dimensional feature channel to a single real number. The activation operation uses two fully connected layers to establish correlations between channels, generating weights for each feature channel. The feature recalibration operation applies the weights output by the activation operation to the original feature map, multiplying them channel-by-channel to the previous feature map.

[0014] In the aforementioned intelligent fault diagnosis method for an automotive wiring harness testing platform, the multi-branch attention mechanism includes a first local branch, a second local branch, and a third global fusion branch. The first local branch processes features from a first data source, the second local branch processes features from a second data source, and the third global fusion branch calculates the global weights of features from different data sources. The first local branch receives a first flattened feature vector output from the first data source processing branch and performs channel attention processing on the first flattened feature vector to obtain a first refined feature vector. The second local branch receives a second flattened feature vector output from the second data source processing branch and performs channel attention processing on the second flattened feature vector to obtain a second refined feature vector.

[0015] In the aforementioned intelligent fault diagnosis method for an automotive wiring harness inspection platform, the third global fusion branch simultaneously receives the first and second refined feature vectors. The third global fusion branch concatenates the first and second refined feature vectors, and uses a fully connected layer to calculate the attention score corresponding to each branch. The multi-branch attention mechanism uses the Softmax function to normalize the attention scores, generating weight coefficients corresponding to different data sources. The multi-branch attention mechanism then uses these weight coefficients to perform a weighted sum of the first and second refined feature vectors, generating the final fused feature vector. The fused feature vector is input into a fully connected layer, which is connected to a Softmax classifier.

[0016] The intelligent fault diagnosis method for automotive wiring harness testing platforms provided by this invention has the following beneficial effects:

[0017] By constructing the MBAM-LM-1DCNN model, an intelligent fault diagnosis method for automotive wiring harness inspection platforms can effectively integrate heterogeneous data from multiple sources, such as voltage, resistance, temperature, and air pressure. The parallel network structure enables the MBAM-LM-1DCNN model to simultaneously process signals with multiple different physical properties, avoiding the problem of incomplete information from a single data source.

[0018] In the feature extraction stage, a lightweight one-dimensional convolutional neural network combined with a channel attention mechanism can automatically learn and emphasize important channel features in the feature map, and suppress invalid or redundant features, thereby improving feature representation ability while reducing computational complexity.

[0019] During the feature fusion stage, the multi-branch attention mechanism dynamically adjusts weight coefficients based on the contribution of different data sources to fault diagnosis through global correlation analysis and adaptive weight allocation. This attention-based fusion strategy ensures that fault diagnosis results fully utilize the complementarity between multi-source data, improving the accuracy and robustness of the automotive wiring harness inspection platform for complex fault mode recognition.

[0020] This invention provides an intelligent fault diagnosis method for an automotive wiring harness testing platform. It offers the following advantages:

[0021] 1. The intelligent fault diagnosis method for automotive wiring harness testing platforms provided by this invention constructs a multi-branch attention mechanism-based multi-source data fusion and parallel lightweight one-dimensional convolutional neural network model. This model employs a parallel network structure to process voltage, resistance, temperature, and air pressure signals separately. The parallel network structure can comprehensively extract deep features from data of different physical dimensions, avoiding misjudgments caused by incomplete information from a single data source. By fusing multi-source heterogeneous data, the intelligent fault diagnosis method for automotive wiring harness testing platforms improves the robustness and reliability of fault diagnosis.

[0022] 2. The intelligent fault diagnosis method for automotive wiring harness inspection platforms provided by this invention embeds a channel attention mechanism into a lightweight one-dimensional convolutional neural network. The channel attention mechanism assigns weights to feature channels through compression, activation, and feature recalibration operations. This weight allocation enables the lightweight one-dimensional convolutional neural network to automatically enhance feature channels containing key fault information and suppress irrelevant noisy feature channels. This mechanism improves the signal-to-noise ratio and accuracy of feature extraction from the automotive wiring harness inspection platform while ensuring model computational efficiency.

[0023] 3. The intelligent fault diagnosis method for automotive wiring harness testing platforms provided by this invention utilizes a multi-branch attention mechanism for feature fusion. This mechanism generates adaptive weight coefficients by concatenating refined feature vectors from different branches and performing global correlation analysis. Weight coefficients are then used to perform a weighted summation of features from different data sources, dynamically adjusting the fusion ratio based on the contribution of each data source. This attention-based fusion strategy fully leverages the complementarity between multi-source data, further optimizing the final fault classification accuracy of the automotive wiring harness testing platform. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention;

[0025] Figure 2 This is a schematic diagram of the key-value pair structure of the attention mechanism of the present invention;

[0026] Figure 3 This is a flowchart illustrating the computational process of the attention mechanism of the present invention.

[0027] Figure 4 This is a schematic diagram of the channel attention mechanism structure of the present invention;

[0028] Figure 5This is a diagram of the MBAM-LM-1DCNN model architecture in this invention;

[0029] Figure 6 This is a structural diagram of the multi-branch attention mechanism in this invention. Detailed Implementation

[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] See attached document Figure 1 This invention provides an intelligent fault diagnosis method for an automotive wiring harness testing platform. The intelligent fault diagnosis method for the automotive wiring harness testing platform is based on multi-source data fusion and convolutional neural networks. The main parameters measured or acquired by the automotive wiring harness testing platform include voltage signals, resistance signals, temperature signals, and air pressure signals.

[0032] To improve the diagnostic accuracy and efficiency of automotive wiring harness inspection platforms using multi-source data fusion and convolutional neural networks, an intelligent fault diagnosis method for these platforms employs a multi-branch attention mechanism-based multi-source data fusion and parallel lightweight one-dimensional convolutional neural network model. This model is denoted as the MBAM-LM-1DCNN model. The MBAM-LM-1DCNN model is used for intelligent fault diagnosis in automotive wiring harness inspection platforms.

[0033] The MBAM-LM-1DCNN model mainly consists of a data preprocessing part, a feature extraction part, a feature fusion part, and a fault classification and recognition part. The intelligent fault diagnosis method for automotive wiring harness testing platforms first performs data augmentation preprocessing on the acquired voltage, resistance, temperature, and air pressure signals from the automotive wiring harness testing platform.

[0034] After data augmentation preprocessing, the preprocessed voltage, resistance, temperature, and air pressure signals are input into a parallel lightweight one-dimensional convolutional neural network (LM-1DCNN). The parallel LM-1DCNN simultaneously extracts important feature information from the four signal sources: voltage, resistance, temperature, and air pressure. Using a parallel LM-1DCNN improves the efficiency of the MBAM-LM-1DCNN model.

[0035] Following feature extraction, the MBAM-LM-1DCNN model introduces a multi-branch attention mechanism, denoted as MBAM. This mechanism achieves full complementarity and fusion among four data sources: voltage, resistance, temperature, and air pressure. MBAM applies attention mechanisms from multiple branches to different data sources, uncovering key features in each source and highlighting the correlations and complementarities between them. Each branch focuses on different aspects or levels of the data, and the outputs of the branches interact and merge to comprehensively represent the important information from the multi-source data.

[0036] Finally, the MBAM-LM-1DCNN model inputs the fused high-level features into a fully connected layer for non-linear learning and mapping. The MBAM-LM-1DCNN model uses the Softmax function to achieve the final fault classification for the automotive wiring harness inspection platform. An intelligent fault diagnosis method for the automotive wiring harness inspection platform is used to detect the platform's operational status.

[0037] See attached document Figure 5 The intelligent fault diagnosis method for automotive wiring harness testing platforms first performs multi-source data acquisition and data preprocessing steps before fault diagnosis.

[0038] The main functions of the automotive wiring harness testing platform include resistance testing, voltage testing, temperature testing, and airtightness testing. Resistance testing checks the continuity of the wiring harness. Voltage testing checks the withstand voltage of the wiring harness. Temperature testing checks the temperature resistance characteristics of the wiring harness. Airtightness testing checks the sealing performance of the wiring harness.

[0039] The automotive wiring harness testing platform collects multi-source data based on resistance detection, voltage detection, temperature testing, and airtightness testing functions. This multi-source data includes four types of parameters: voltage, resistance, temperature, and air pressure. Specifically, the signals collected by the automotive wiring harness testing platform include voltage, resistance, temperature, and air pressure signals.

[0040] The intelligent fault diagnosis method for automotive wiring harness testing platforms performs data augmentation preprocessing on the acquired voltage, resistance, temperature, and air pressure signals. This data augmentation preprocessing enhances the diversity and richness of the data.

[0041] The voltage, resistance, temperature, and pressure signals, after data augmentation preprocessing, are used as input data. These preprocessed signals are then fed into a multi-source data fusion and parallel lightweight one-dimensional convolutional neural network model based on a multi-branch attention mechanism for subsequent feature extraction.

[0042] See attached document Figure 2 and attached Figure 3 The intelligent fault diagnosis method used in automotive wiring harness testing platforms employs an attention mechanism. This attention mechanism primarily uses key-value pairs.

[0043] The input information vector in key-value pair form is represented as:

[0044] ;

[0045] The dictionary Source contains For key-value pairs, the key-value pair pattern consists of a key, a value, and a query vector. When the key equals the value, the key-value pair form is equivalent to the ordinary form. The core idea of ​​the attention mechanism is that, given a query vector Query, the attention distribution between the query vector Query and the key is calculated in the dictionary Source, and then the attention distribution is applied to the elements in the dictionary Source. The values ​​are weighted for the attention distribution.

[0046] The mathematical formula for the attention mechanism is shown below:

[0047] ;

[0048] in: This represents the output of the attention mechanism. This represents the query vector. Represents the dictionary Source. Indicates the length of the dictionary Source. This represents the similarity calculation function. This refers to the first [item] in the dictionary Source. Key. This refers to the first [item] in the dictionary Source. Value.

[0049] The computational steps of the attention mechanism are divided into the following steps:

[0050] The similarity between the query vector (Query) and the key (Key) is calculated to obtain the weight coefficient. Similarity calculation methods include dot product calculation, cosine similarity calculation, and MLP network calculation.

[0051] The formula for calculating the dot product is shown below:

[0052] ;

[0053] in: Represents the query vector and key. The similarity. This represents the query vector. Indicates the first Key.

[0054] The formula for calculating cosine similarity is shown below:

[0055] ;

[0056] in: Represents the query vector With the dictionary Source Key The similarity between the two. The closer the value is to 1, the more similar the two are; This represents the query vector, which is the input vector in the attention mechanism and is used to find matching information in the dictionary; This refers to the first [item] in the dictionary Source. Each key vector (Key); The modulo operation represents a vector and is used to calculate the length or size of a vector. Represents the query vector The model; Indicates the first Key vectors The model.

[0057] The formula for MLP network computation is shown below:

[0058] ;

[0059] in: This represents a multilayer perceptron network.

[0060] The weight coefficients are numerically transformed using the softmax function to obtain directly usable weights. This numerical transformation normalizes the data and highlights the weights of important elements.

[0061] The formula is shown below:

[0062] ;

[0063] in: Indicates the first The normalized weight coefficients corresponding to the nth element. These weight coefficients are the final weights assigned by the attention mechanism to the nth element. The weight of each value, all The sum is 1; This represents the Softmax activation function; the Softmax activation function is used to convert the input numerical value into a probability distribution form. Indicates the first The similarity of each element is the original input value that needs to be normalized. Represents the natural constant; This represents the index variable used in the summation operation to iterate through each element in the dictionary; Indicates the length of the dictionary Source; Indicates the first term in the summation term The similarity of elements.

[0064] The formula is shown below:

[0065] ;

[0066] in: This represents the output of the attention mechanism. Indicates the length of the dictionary Source. This represents the normalized weight coefficients. Indicates the first Value.

[0067] See attached document Figure 4 A smart fault diagnosis method for automotive wiring harness inspection platforms introduces a channel attention mechanism into the feature extraction network. This channel attention mechanism is used to process the feature map output by the one-dimensional convolutional neural network.

[0068] The channel attention mechanism mainly consists of three steps: compression, activation, and feature recalibration. The first step is compression. This compression operation uses global average pooling to compress the feature map spatially. Each two-dimensional feature channel is compressed into a single real number. This real number has a global receptive field.

[0069] The formula for calculating compression is shown below:

[0070] ;

[0071] in: Indicates the first The output value of each channel after global average pooling. This represents the global average pooling function. Indicates the input number of the first... Feature map of each channel. This indicates the length of the time dimension of the feature map. This represents an index along the time dimension. Indicates the first The feature map of the first channel in the first... The value at each position. This indicates the total number of channels.

[0072] Following the compression operation, the channel attention mechanism performs an activation operation. This activation operation aims to fully capture channel dependencies. It utilizes two fully connected layers to establish correlations between channels. The activation operation generates weights for each feature channel.

[0073] The calculation formula for the stimulus operation is as follows:

[0074] ;

[0075] in: This represents the attention weight vector output by the activation operation. The mapping function represents the excitation operation. This represents the Sigmoid activation function. This represents the weight matrix of the second fully connected layer. This represents the ReLU activation function. This represents the weight matrix of the first fully connected layer. This represents the output values ​​of all channels. The input feature vector is composed of these features.

[0076] Following the activation operation, the channel attention mechanism performs a feature recalibration operation. The feature recalibration operation applies the weight vector output from the activation operation to the original feature map. The feature recalibration operation applies weights to the previous features channel by channel through multiplication. The feature recalibration operation completes the recalibration of the channel features.

[0077] The calculation formula for feature recalibration is shown below:

[0078] ;

[0079] in: This represents the set of output feature maps after recalibration. This represents the set of feature maps of the original input. This represents the attention weight vector. The feature maps represent the individual channels of the original input. This represents the attention weight value for each channel.

[0080] See attached document Figure 5 An intelligent fault diagnosis method for automotive wiring harness inspection platforms is constructed based on a multi-branch attention mechanism, multi-source data fusion, and a parallel lightweight one-dimensional convolutional neural network model. This model is denoted as the MBAM-LM-1DCNN model.

[0081] The parameters of the MBAM-LM-1DCNN model are as follows:

[0082] Table 1. Parameters of the MBAM-LM-1DCNN model

[0083]

[0084] Note: "-" in Table 1 indicates that it is not applicable.

[0085] The MBAM-LM-1DCNN model employs a parallel network structure. Parallel network structures can process multiple different types of input signals simultaneously. Specifically, the MBAM-LM-1DCNN model includes a first data source processing branch and a second data source processing branch. Both the first and second data source processing branches utilize lightweight one-dimensional convolutional neural networks. This lightweight one-dimensional convolutional neural network is denoted as LM-1DCNN.

[0086] The first data source processing branch processes voltage signals. The second data source processing branch processes resistance signals. When temperature and air pressure signals are present, the MBAM-LM-1DCNN model extends the corresponding third and fourth data source processing branches.

[0087] The LM-1DCNN mainly consists of an input layer, convolutional layers, batch normalization layers, activation function layers, pooling layers, and flattening layers. LM-1DCNN extracts deep features of a signal by stacking multiple convolutional and pooling layers.

[0088] The first layer of LM-1DCNN is a convolutional layer. The convolutional layer uses convolutional kernels to perform sliding window computation on the input signal. It extracts local features of the input signal through convolution operations. The parameters of the convolutional layer include kernel size, number of kernels, and stride.

[0089] The LM-1DCNN connects a batch normalization layer after the convolutional layers. The batch normalization layer normalizes the output of the convolutional layers. This layer accelerates the convergence of the MBAM-LM-1DCNN model and prevents the vanishing or exploding gradient problems that often occur during training.

[0090] The LM-1DCNN connects to an activation function layer after the batch normalization layer. The activation function layer uses the ReLU activation function. The ReLU activation function adds a non-linear element to the neural network. The ReLU activation function enables the MBAM-LM-1DCNN model to fit complex non-linear relationships.

[0091] The LM-1DCNN connects pooling layers after the activation function layers. The pooling layers use max pooling. Pooling layers downsample the feature maps. Pooling layers reduce the size of the feature maps. Pooling layers reduce the computational cost of the MBAM-LM-1DCNN model. Pooling layers preserve the main feature information in the feature maps.

[0092] After multiple convolutional, batch normalization, activation, and pooling operations, the LM-1DCNN is connected to a flattening layer at the end of the LM-1DCNN. The flattening layer unfolds the multi-dimensional feature map into a one-dimensional feature vector. The one-dimensional feature vector output by the flattening layer serves as the input to the subsequent multi-branch attention mechanism.

[0093] The specific network parameter settings for the MBAM-LM-1DCNN model are shown in Table 1. Table 1 defines the specific configuration of each network layer. The specific configuration includes the size of the convolutional kernels, the number of convolutional kernels, the stride, and the dimension of the output feature map. By reasonably setting the size and number of convolutional kernels, the MBAM-LM-1DCNN model can minimize the number of parameters while ensuring feature extraction capabilities.

[0094] See attached document Figure 6 The intelligent fault diagnosis method for automotive wiring harness inspection platforms employs a multi-branch attention mechanism after feature extraction. This multi-branch attention mechanism, denoted as MBAM, is used for the fusion of multi-source features.

[0095] The multi-branch attention mechanism mainly consists of a first local branch, a second local branch, and a third global fusion branch. The first local branch processes features from the first data source processing branch. The second local branch processes features from the second data source processing branch. The third global fusion branch calculates the global weights of features from different data sources.

[0096] The first local branch receives the first flattened feature vector output by the lightweight one-dimensional convolutional neural network. The first flattened feature vector is denoted as... The first local branch performs channel attention processing on the first flattened feature vector to obtain the first refined feature vector. The first refined feature vector is denoted as... .

[0097] The second local branch receives the second flattened feature vector output by the lightweight one-dimensional convolutional neural network. The second flattened feature vector is denoted as... The second local branch performs channel attention processing on the second flattened feature vector to obtain the second refined feature vector. The second refined feature vector is denoted as... .

[0098] The third global fusion branch simultaneously receives the first refined feature vector. Second refined feature vector The third global fusion branch uses a fully connected layer to perform global correlation analysis on the first and second refined feature vectors. The third global fusion branch then generates an attention score.

[0099] The formula for calculating the attention score is as follows:

[0100] ;

[0101] ;

[0102] in: This represents the attention score for the first branch. This represents the weight matrix of the first fully connected layer. This represents the concatenated vector of the first refined feature vector and the second refined feature vector. This represents the first refined feature vector. This represents the second refined feature vector. This represents the bias vector of the first fully connected layer. This represents the attention score for the second branch. This represents the weight matrix of the fully connected layer in the second branch. This represents the bias vector of the fully connected layer in the second branch.

[0103] After obtaining the attention score, the multi-branch attention mechanism uses the Softmax function to normalize the attention score. The normalization process generates weight coefficients corresponding to different data sources.

[0104] The formula for calculating the weighting coefficient is as follows:

[0105] ;

[0106] ;

[0107] in: This represents the weighting coefficient of the first data source. Represents the natural constant. This represents the attention score for the first branch. This represents the attention score for the second branch. This represents the weighting coefficient of the second data source.

[0108] Finally, the multi-branch attention mechanism performs weighted fusion of the refined feature vectors from each branch based on the generated weight coefficients. The weighted fusion generates the final fused feature vector.

[0109] The formula for calculating the fused feature vector is as follows:

[0110] ;

[0111] in: This represents the final fused feature vector. This represents the first refined feature vector after weighting. This represents the second refined feature vector after weighting. This represents the weighting coefficient of the first data source. This represents the first refined feature vector. This represents the weighting coefficient of the second data source. This represents the second refined feature vector.

[0112] Fusion feature vectors The input is fed into a fully connected layer. The fully connected layer connects to a Softmax classifier. The Softmax classifier outputs the fault category probability for the automotive wiring harness inspection platform. The category with the highest fault category probability is the final fault diagnosis result.

Claims

1. An intelligent fault diagnosis method for an automotive wiring harness testing platform, characterized in that, Includes the following steps: Multi-source data is collected using an automotive wiring harness testing platform, and the multi-source data is then subjected to data augmentation preprocessing. A one-dimensional convolutional neural network model based on multi-branch attention mechanism, multi-source data fusion, and parallel lightweight is constructed. This model is referred to as the MBAM-LM-1DCNN model. The multi-source data, after data augmentation preprocessing, is input into the MBAM-LM-1DCNN model; Feature extraction is performed using the lightweight one-dimensional convolutional neural network in the MBAM-LM-1DCNN model, which is denoted as LM-1DCNN. Feature fusion is performed using the multi-branch attention mechanism in the MBAM-LM-1DCNN model, which is referred to as MBAM. The fault category probability of the automotive wiring harness detection platform is output by the Softmax classifier in the MBAM-LM-1DCNN model, and the category with the highest fault category probability is taken as the final fault diagnosis result.

2. The intelligent fault diagnosis method for an automotive wiring harness testing platform according to claim 1, characterized in that, The multi-source data includes voltage signals, resistance signals, temperature signals, and air pressure signals; The MBAM-LM-1DCNN model adopts a parallel network structure and includes a first data source processing branch and a second data source processing branch. The first data source processing branch is used to process the voltage signal, and the second data source processing branch is used to process the resistance signal; Both the first data source processing branch and the second data source processing branch use the LM-1DCNN; The LM-1DCNN consists of an input layer, a convolutional layer, a batch normalization layer, an activation function layer, a pooling layer, and a flattening layer.

3. The intelligent fault diagnosis method for an automotive wiring harness testing platform according to claim 2, characterized in that, The convolutional layer uses a convolutional kernel to perform sliding window calculations on the input signal, and extracts local features of the input signal through convolution operations; The batch normalization layer standardizes the output of the convolutional layer. The activation function layer uses the ReLU activation function; The pooling layer employs max pooling to downsample the feature map; The flattening layer is located at the end of the LM-1DCNN and unfolds the multidimensional feature map into a one-dimensional feature vector, which serves as the input to the multi-branch attention mechanism.

4. The intelligent fault diagnosis method for an automotive wiring harness testing platform according to claim 1, characterized in that, The intelligent fault diagnosis method of the automotive wiring harness inspection platform introduces a channel attention mechanism into the feature extraction network; The channel attention mechanism is used to process the feature map output by a one-dimensional convolutional neural network. The channel attention mechanism includes compression operations, activation operations, and feature recalibration operations.

5. The intelligent fault diagnosis method for an automotive wiring harness testing platform according to claim 4, characterized in that, The compression operation uses global average pooling to compress the feature map in the spatial dimension, compressing each two-dimensional feature channel into real numbers. The activation operation utilizes two fully connected layers to establish the correlation between channels and generate the weight of each feature channel; The feature recalibration operation applies the weights output by the activation operation to the original feature map, weighting the previous features channel by channel through multiplication.

6. The intelligent fault diagnosis method for an automotive wiring harness testing platform according to claim 2, characterized in that, The multi-branch attention mechanism includes a first local branch, a second local branch, and a third global fusion branch; The first local branch is used to process the features of the first data source, the second local branch is used to process the features of the second data source, and the third global fusion branch is used to calculate the global weights of the features of different data sources.

7. The intelligent fault diagnosis method for an automotive wiring harness testing platform according to claim 6, characterized in that, The first local branch receives the first flattened feature vector output by the lightweight one-dimensional convolutional neural network, and performs channel attention processing on the first flattened feature vector to obtain the first refined feature vector. The second local branch receives the second flattened feature vector output by the lightweight one-dimensional convolutional neural network, and performs channel attention processing on the second flattened feature vector to obtain the second refined feature vector.

8. The intelligent fault diagnosis method for an automotive wiring harness testing platform according to claim 7, characterized in that, The third global fusion branch simultaneously receives the first refined feature vector and the second refined feature vector; The third global fusion branch uses a fully connected layer to perform global correlation analysis on the first refined feature vector and the second refined feature vector to generate an attention score; The multi-branch attention mechanism uses the Softmax function to normalize the attention score and generate weight coefficients corresponding to different data sources.

9. The intelligent fault diagnosis method for an automotive wiring harness testing platform according to claim 8, characterized in that, The multi-branch attention mechanism performs weighted fusion of the refined feature vectors of the first local branch and the second local branch according to the generated weight coefficients to generate the final fused feature vector. The fused feature vector is input into a fully connected layer, which is connected to the Softmax classifier.

10. The intelligent fault diagnosis method for an automotive wiring harness testing platform according to claim 1, characterized in that, The automotive wiring harness testing platform collects multi-source data based on resistance detection, voltage detection, temperature testing, and airtightness testing functions. The resistance detection function is used to detect the continuity of the wire harness, the voltage detection function is used to detect the withstand voltage value of the wire harness, the temperature test function is used to detect the temperature resistance characteristics of the wire harness, and the airtightness test function is used to detect the sealing performance of the wire harness.

Images