Suspension control device fault diagnosis method and system based on image recognition

Abstract

The application discloses a suspension control equipment fault diagnosis method based on image recognition and belongs to the technical field of industrial control equipment fault diagnosis. The application proposes a new intelligent fault diagnosis method based on image classification, collects fault image information of a new energy automobile suspension control equipment based on a magneto-rheological elastomer, and realizes accurate identification of faults; in addition to the image classification technology, the neural network feature map is reprocessed, the intermediate feature matrix is generated through the reprocessing after flattening, and the neural network modeling is used again to form an enhanced pooling feature, and the corresponding enhancement module can be used as an enhanced model and applied to an image classification task to improve performance.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for industrial control equipment, and specifically to a method and system for fault diagnosis of suspension control equipment based on image recognition. Background Technology

[0002] The suspension system of new energy vehicles, especially active and semi-active suspensions using intelligent materials such as magnetorheological elastomers, is a key component for achieving vehicle ride comfort and handling. Suspension control equipment includes controllers, sensors, actuators (such as magnetorheological dampers), and communication modules. Because suspension control equipment operates under complex conditions such as vibration, temperature differences, and current surges for extended periods, it has a high risk of failure.

[0003] Currently, fault diagnosis of suspension systems mainly relies on the following technical approaches:

[0004] Sensor-based parameter monitoring: This method collects suspension operating parameters using sensors such as accelerometers, displacement sensors, and current sensors, and compares these parameters with preset thresholds to determine faults. However, this method suffers from drawbacks such as high sensor deployment costs and the inability of a single parameter to comprehensively characterize the fault type.

[0005] Fault code reading based on on-board diagnostic (OBD) system: Diagnosis is performed by reading fault codes on the controller area network (CAN) bus. However, OBD systems can usually only identify pre-set electrical faults and lack the ability to effectively distinguish progressive faults such as mechanical degradation, sensor drift, and actuator performance degradation.

[0006] Fault feature extraction methods based on signal processing: These methods extract time-frequency domain features of vibration signals using wavelet transform, empirical mode decomposition, etc., and then combine them with traditional classifiers (such as support vector machines and K-nearest neighbors) for fault identification. However, these methods rely on manual feature engineering, and the effectiveness of feature extraction directly affects diagnostic accuracy, limiting their generalization ability.

[0007] In recent years, deep learning technology has demonstrated tremendous potential in the field of industrial fault diagnosis. The development of artificial intelligence technology and its application in diagnostic reasoning will continuously improve the intelligence level of diagnostic systems. For example, the emergence of neural network reasoning methods has effectively solved the problem of "rule combinatorial explosion" in expert systems, and the recent refinement of neural network algorithms and the emergence of new neural network algorithms will further drive the development of fault diagnosis systems towards a higher level of intelligence. However, existing research mostly focuses on fault mode recognition using one-dimensional time-series data such as vibration signals and current signals, with less emphasis on introducing image recognition technology into the field of suspension fault diagnosis. In fact, the physical state of suspension control equipment—such as leakage marks in magnetorheological dampers, wear or detachment of wiring harnesses, burn marks on circuit boards, and sensor installation position misalignment—can all be intuitively reflected through image information. If image recognition and neural network technologies can be integrated, it is expected to achieve comprehensive diagnosis of the external physical state and internal operating parameters of suspension control equipment, overcoming the shortcomings of existing technologies in terms of single information sources and insufficient fault type discrimination capabilities. Summary of the Invention

[0008] The technical problem to be solved by this invention is: how to achieve comprehensive diagnosis of the external physical state and internal operating parameters of suspension equipment based on image recognition and neural network technology, so as to make up for the shortcomings of the existing technology in terms of single information source and insufficient fault type discrimination capability, and provide a fault diagnosis method for suspension control equipment based on image recognition.

[0009] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:

[0010] S1: Data Acquisition and Preprocessing

[0011] Fault images of the suspension control equipment are acquired and preprocessed to obtain standardized input images;

[0012] S2: Feature Extraction

[0013] A deep convolutional neural network is used to perform hierarchical feature extraction on the standardized input image, and the output neural network feature map is generated.

[0014] S3: Enhanced Feature Modeling

[0015] The neural network feature map is flattened to obtain a flattened feature map vector. The flattened feature map vector is transformed to generate an intermediate feature matrix. Then, the intermediate feature matrix is ​​subjected to feature convolution, row and column max pooling operations and concatenated with the feature vector to obtain enhanced pooling features.

[0016] S4: Fault Classification Output

[0017] The enhanced pooling features are input into the classification module, and the fault diagnosis results of the suspension control device are output.

[0018] Furthermore, in step S1, the image acquisition device includes an industrial camera, an inspection robot, and a mobile phone, which are deployed on the bottom of the vehicle or at the maintenance station; the acquisition perspective includes the appearance of the magnetorheological damper, the wiring harness interface, the controller circuit board, and the sensor installation position.

[0019] Furthermore, in step S1, the preprocessing operations include image denoising, size normalization, and data augmentation, and the batch of fault images are preprocessed to form a standardized input image dataset.

[0020] Furthermore, in step S2, the specific process of hierarchical feature extraction is as follows:

[0021] S21: Extract shallow physical features such as edges and textures using convolutional modules located at the bottom layer of deep convolutional neural networks;

[0022] S22: Extract component-level semantic features of damper body, wire harness interface, and circuit board components using residual modules located in the middle layer of a deep convolutional neural network;

[0023] S23: Enhance the model's ability to focus on fault-related regions by utilizing attention modules located in the high layers of the deep convolutional neural network, and finally output the neural network feature map.

[0024] Furthermore, the deep convolutional neural network is a ResNet-50 network. The bottom convolutional module includes a 7×7 convolutional layer and a 3×3 max pooling layer to extract shallow physical features such as edges and textures. The middle residual module is composed of 16 stacked residual units, each of which includes a bottleneck structure of two 1×1 convolutional layers and one 3×3 convolutional layer. The residual module uses residual connections to achieve identity mapping, thereby extracting component-level semantic features such as damper body, wiring harness interface, and circuit board components. The high-level attention module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule adopts a dual-path design of global average pooling and max pooling to capture both global statistical features and local salient features of the fault area. Then, it performs inter-channel modeling through two fully connected layers to generate channel weight coefficients. The spatial attention submodule uses a 7×7 convolutional kernel to extract cross-regional correlation features. Based on the channel weight coefficients obtained from the channel attention submodule, it establishes semantic relationships between different positions on the feature map and finally outputs the neural network feature map.

[0025] Furthermore, in step S3, the formula for calculating the intermediate feature matrix is ​​as follows:

[0026] ;

[0027] in, It's weight. It is the feature map vector after the pixel matrix is ​​flattened, i.e., the flattened feature map vector. This is the intermediate feature matrix.

[0028] Furthermore, in step S3, the specific formula for feature convolution is as follows:

[0029] Assume the number of convolutional filters is , No. Each convolution operation is as follows:

[0030] ;

[0031] in, , It is a non-linear activation function. Represents two-dimensional convolution. and They are the first The weights and biases of each convolutional kernel. It is a feature map.

[0032] Furthermore, in step S3, the specific formula for the row-column max pooling operation is as follows:

[0033] ;

[0034] in, and These are the max-pooled vectors generated in the row and column directions, where t represents the t-th max-pooled row and column vectors. It is the feature map vector at the j-th row and g-th column. It is the feature map vector at the g-th row and j-th column, d is the size of the 2D convolution kernel, and k is the feature map value. The size.

[0035] Furthermore, in step S3, the feature vector concatenation formula is as follows:

[0036] ;

[0037] in, , Indicates the total number of convolutional filters. This is the concatenated feature vector, i.e., the enhanced pooling feature.

[0038] Furthermore, in step S5, the classification module performs a binary classification task or a multi-classification task; the binary classification task is to determine whether the suspension control device is in a faulty state or a fault-free state, and the multi-classification task is to determine the specific fault type of the suspension control device.

[0039] This invention also provides an image recognition-based suspension control device fault diagnosis system, applied to the above-mentioned method, comprising:

[0040] The acquisition and preprocessing module is used to acquire fault images of the suspension control equipment and preprocess them to obtain standardized input images.

[0041] The feature extraction module is used to perform hierarchical feature extraction on the standardized input image using a deep convolutional neural network and output a neural network feature map.

[0042] The enhanced feature modeling module is used to flatten the neural network feature map to obtain flattened feature map vectors, transform the flattened feature map vectors to generate intermediate feature matrices, and then perform feature convolution, row and column max pooling operations on the intermediate feature matrix and concatenate it with the feature vectors to obtain enhanced pooling features.

[0043] The fault classification output module is used to input enhanced pooling features into the classification module and output the fault diagnosis results of the suspension control device.

[0044] Compared with the prior art, the present invention has the following advantages: This image recognition-based suspension control equipment fault diagnosis method proposes a new intelligent fault diagnosis method based on image classification. It collects fault image information of the suspension control equipment of new energy vehicles based on magnetorheological elastomers and achieves accurate fault identification. In addition to using image classification technology, the present invention also reprocesses the neural network feature map, generates an intermediate feature matrix through flattened reprocessing, and then uses the neural network to model it into enhanced pooling features. The corresponding enhancement module can be used as an enhancement model to improve the performance of image classification tasks. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the model architecture of the image recognition-based suspension control device fault diagnosis method in an embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram illustrating the process of generating an intermediate feature matrix from a neural network feature map in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the enhanced feature modeling and generation process in an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of the processing of the deep convolutional neural network ResNet-50 in an embodiment of the present invention. Detailed Implementation

[0049] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0050] Example 1

[0051] like Figure 1 As shown, this embodiment provides a technical solution: a suspension control device fault diagnosis method based on image recognition, including the following steps:

[0052] Step 1: First, collect images of various faults in the suspension control equipment of the new energy vehicle. Image acquisition devices (including but not limited to industrial cameras, intelligent inspection robots, and smartphones) deployed under the vehicle or at the repair bay are used to collect image data of the suspension control equipment. Acquisition angles include, but are not limited to: the appearance of the magnetorheological damper, wiring harness interfaces, controller circuit boards, and sensor installation locations.

[0053] Step Two: Next, preprocessing operations such as denoising, size normalization, and data augmentation are performed on the original images to form a standardized input image dataset. A deep convolutional neural network, ResNet-50, is used as the feature extraction backbone. A schematic diagram of the process is shown below. Figure 4 As shown, the bottom convolutional module consists of a 7×7 convolutional layer (stride 2, output channels 64) and a 3×3 max pooling layer, responsible for extracting shallow physical features such as edges and textures (e.g., texture anomalies in leakage areas, broken wire harness edges). The middle residual module consists of 16 stacked residual units, each containing a bottleneck structure of two 1×1 convolutional layers and one 3×3 convolutional layer. This module utilizes residual connections to achieve identity mapping, enabling the network to focus on learning the residuals between input and output, thereby effectively extracting component-level semantic features (e.g., damper body, wire harness interface, circuit board components). The high-level attention module introduces channel attention and spatial attention mechanisms to enhance the model's ability to focus on fault-related regions. The channel attention submodule adopts a dual-path design of global average pooling and max pooling, simultaneously capturing global statistical features and local salient features of the fault region, and then uses two fully connected layers to model high-order dependencies between channels, generating channel weight coefficients. The spatial attention submodule uses a 7×7 convolutional kernel to extract cross-regional associated features and establishes semantic associations between different locations on the feature map based on the enhanced results of channel attention.

[0054] Considering that the input of image information in computer vision is a pixel matrix, this invention directly flattens these extracted feature matrices (pixel matrices) to obtain their respective flattened feature map vectors. Then for each Perform the operation shown in formula (1) to obtain the corresponding intermediate feature matrix. ,in These are weights, which can be learned. The process of generating an intermediate feature matrix from a neural network feature map is as follows: Figure 2 As shown. Formula (1) is as follows:

[0055] (1)

[0056] in, It's weights, which can be acquired through learning. It is the feature map vector after the pixel matrix is ​​flattened. This is the intermediate feature matrix.

[0057] Step 3: Finally, this invention combines the intermediate feature matrix with a convolutional neural network using two-dimensional convolutional kernels, and reduces the fault diagnosis of new energy vehicle suspension control equipment to a binary classification task of image recognition and classification, i.e., fault present / no fault. Furthermore, this invention labels the collected dataset and sets fault categories, reducing the fault diagnosis of new energy vehicle suspension control equipment to a multi-class classification task of image recognition and classification. The process of modeling the feature matrix is ​​as follows... Figure 3 As shown.

[0058] In a convolutional neural network with two-dimensional convolutional kernels, assuming the number of convolutional filters is... , No. Each convolution operation is as follows:

[0059] (2)

[0060] in, , It is a non-linear activation function. Represents two-dimensional convolution. and They are the first The weights and biases of each convolutional kernel. It is a feature map.

[0061] After the convolutional layer obtains the feature map, this invention then uses row and column max pooling to generate vectors respectively. as well as .

[0062] (3)

[0063] Connect these vectors as follows:

[0064] (4)

[0065] in The convolution operations described above aim to extract useful similarity patterns, with each convolution kernel corresponding to a feature. The parameters of these convolution kernels can be adjusted during model training.

[0066] In binary classification tasks:

[0067] The classification module uses a single fully connected layer combined with a sigmoid activation function to concatenate the feature vectors. The input to the fully connected layer undergoes a nonlinear transformation to obtain the hidden representation. This process maps high-dimensional features to a low-dimensional feature space using learnable weights and biases, while preserving discriminative information relevant to fault identification. Then, a random deactivation operation is performed to suppress excessive reliance on specific features during training, enhancing generalization ability and obtaining the original binary classification score z. Subsequently, the sigmoid activation function is used to calculate a piecewise binary classification function, mapped to the [0,1] interval, yielding the fault probability value:

[0068]

[0069] Finally, the fault probability value is binarized for decision-making, specifically as follows:

[0070]

[0071] in The preset decision threshold (usually taken as...) (The actual deployment cost can be adjusted based on the cost of false negatives / false positives).

[0072] In multi-classification tasks:

[0073] The multi-classification module is implemented using fully connected layers combined with multi-neuron Softmax output, which concatenates the feature vectors. The input to the fully connected layer undergoes a nonlinear transformation to obtain the hidden representation. This process maps high-dimensional features to a low-dimensional feature space using learnable weights and biases, while preserving discriminative information relevant to fault detection. Then, a random deactivation operation is performed to suppress excessive reliance on specific features during training, enhancing generalization ability and yielding the original score vectors for the five classification classes. Then, the Softmax activation function is used to activate... Mapped to the probability distribution of five types of faults The category with the highest probability is the predicted fault type.

[0074] Example 2

[0075] To verify the solution of this invention, the designed model was first tested on a classic image classification task, and the results are shown in Table 1. The results show that the model designed in this invention is useful as an image classification enhancement model.

[0076] Table 1. Experimental results of the proposed solution as an enhancement model in image classification tasks.

[0077]

[0078] In this embodiment, fault image information was collected using a new energy vehicle suspension control device based on magnetorheological elastomers from a certain company. First, intelligent fault diagnosis was performed to determine whether there was a fault, and then the fault types were classified and labeled into 5 categories, as shown in Table 2.

[0079] Table 2. Summary of Fault Type Classification Labels and Identifiable Image Features

[0080]

[0081] The specific experimental results of this embodiment are shown in Table 3.

[0082] Table 3 Experimental Results of Intelligent Fault Identification for New Energy Vehicle Suspension Control Equipment Based on Magnetorheological Elastomers

[0083]

[0084] The test results in Table 3 show that the present invention can accurately identify faults in the suspension control equipment of new energy vehicles based on image classification.

[0085] In summary, the image recognition-based suspension control equipment fault diagnosis method described in the above embodiments proposes a novel intelligent fault diagnosis method based on image classification. It collects fault image information of the suspension control equipment of new energy vehicles based on magnetorheological elastomers and achieves accurate fault identification. In addition to utilizing image classification technology, this invention also reprocesses the neural network feature maps, generating an intermediate feature matrix through flattened reprocessing, and then uses the neural network to model it as enhanced pooling features. The corresponding enhancement module can be used as an enhancement model to improve the performance of image classification tasks.

[0086] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for fault diagnosis of suspension control equipment based on image recognition, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing Fault images of the suspension control equipment are acquired and preprocessed to obtain standardized input images; S2: Feature Extraction A deep convolutional neural network is used to perform hierarchical feature extraction on the standardized input image, and the output neural network feature map is generated. S3: Enhanced Feature Modeling The neural network feature map is flattened to obtain a flattened feature map vector. The flattened feature map vector is transformed to generate an intermediate feature matrix. Then, the intermediate feature matrix is ​​subjected to feature convolution, row and column max pooling operations and concatenated with the feature vector to obtain enhanced pooling features. S4: Fault Classification Output The enhanced pooling features are input into the classification module, and the fault diagnosis results of the suspension control device are output.

2. The image recognition-based suspension control device fault diagnosis method according to claim 1, characterized in that, In step S1, the image acquisition device includes an industrial camera, an inspection robot, and a mobile phone, which are deployed on the bottom of the vehicle or at the maintenance station; the acquisition perspective includes the appearance of the magnetorheological damper, the wiring harness interface, the controller circuit board, and the sensor installation position.

3. The image recognition-based suspension control device fault diagnosis method according to claim 1, characterized in that, In step S1, the preprocessing operations include image denoising, size normalization, and data augmentation. The batch of fault images are preprocessed to form a standardized input image dataset.

4. The image recognition-based suspension control device fault diagnosis method according to claim 1, characterized in that, In step S2, the specific process of hierarchical feature extraction is as follows: S21: Extract shallow physical features such as edges and textures using convolutional modules located at the bottom layer of deep convolutional neural networks; S22: Extract component-level semantic features of damper body, wire harness interface, and circuit board components using residual modules located in the middle layer of a deep convolutional neural network; S23: Enhance the model's ability to focus on fault-related regions by utilizing attention modules located in the high layers of the deep convolutional neural network, and finally output the neural network feature map.

5. The image recognition-based suspension control device fault diagnosis method according to claim 4, characterized in that, The deep convolutional neural network is a ResNet-50 network. The bottom convolutional module includes a 7×7 convolutional layer and a 3×3 max pooling layer to extract shallow physical features such as edges and textures. The middle residual module consists of 16 stacked residual units, each of which includes a bottleneck structure of two 1×1 convolutional layers and one 3×3 convolutional layer. The residual module uses residual connections to achieve identity mapping, thereby extracting component-level semantic features such as damper body, wiring harness interface, and circuit board components. The high-level attention module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule adopts a dual-path design of global average pooling and max pooling to capture both global statistical features and local salient features of the fault area. Then, it performs inter-channel modeling through two fully connected layers to generate channel weight coefficients. The spatial attention submodule uses a 7×7 convolutional kernel to extract cross-regional correlation features. Based on the channel weight coefficients obtained from the channel attention submodule, it establishes semantic relationships between different positions on the feature map and finally outputs the neural network feature map.

6. The image recognition-based suspension control device fault diagnosis method according to claim 5, characterized in that, In step S3, the formula for calculating the intermediate feature matrix is ​​as follows: ； in, It's weight. It is the feature map vector after the pixel matrix is ​​flattened, i.e., the flattened feature map vector. This is the intermediate feature matrix.

7. The image recognition-based suspension control device fault diagnosis method according to claim 6, characterized in that, In step S3, the specific formula for feature convolution is as follows: Assume the number of convolutional filters is , No. Each convolution operation is as follows: ； in, , It is a non-linear activation function. Represents two-dimensional convolution. and They are the first The weights and biases of each convolutional kernel. It is a feature map.

8. The image recognition-based suspension control device fault diagnosis method according to claim 7, characterized in that, In step S3, the specific formula for the row-column max pooling operation is as follows: ； in, and These are the max-pooled vectors generated in the row and column directions, where t represents the t-th max-pooled row and column vectors. It is the feature map vector at the j-th row and g-th column. It is the feature map vector at the g-th row and j-th column, d is the size of the 2D convolution kernel, and k is the feature map value. The size.

9. The image recognition-based suspension control device fault diagnosis method according to claim 8, characterized in that, In step S3, the feature vector concatenation formula is as follows: ； in, , Indicates the total number of convolutional filters. This is the concatenated feature vector, i.e., the enhanced pooling feature.

10. A suspension control equipment fault diagnosis system based on image recognition, characterized in that, The method applied to any one of claims 1 to 9 includes: The acquisition and preprocessing module is used to acquire fault images of the suspension control equipment and preprocess them to obtain standardized input images. The feature extraction module is used to perform hierarchical feature extraction on the standardized input image using a deep convolutional neural network and output a neural network feature map. The enhanced feature modeling module is used to flatten the neural network feature map to obtain flattened feature map vectors, transform the flattened feature map vectors to generate intermediate feature matrices, and then perform feature convolution, row and column max pooling operations on the intermediate feature matrix and concatenate it with the feature vectors to obtain enhanced pooling features. The fault classification output module is used to input enhanced pooling features into the classification module and output the fault diagnosis results of the suspension control device.

Images