Vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion

Abstract

The invention discloses a vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion, and the method comprises the steps: obtaining a plurality of subspaces through projection according to the collected multi-dimensional data of a vehicle engine bearing under an early slight fault degree, and setting weights according to the correlation of the subspaces one by one through a multi-head attention mechanism, and obtaining fusion features; according to the fusion features, energy values of the multiple features are obtained one by one according to the multiple neurons through a parameter-free attention mechanism, and feature processing is performed on the fusion features according to the energy values so as to reduce noise; according to the fusion features after noise reduction, deep features are extracted step by step through multi-layer two-dimensional convolution operation, and the fault type is judged through a full connection layer to obtain a fault diagnosis model. According to the method, a complete fault diagnosis process is formed through multi-source heterogeneous data fusion (vibration and sound), adaptive noise reduction and the improved convolutional neural network, and the problems of incomplete feature coverage, insufficient noise reduction fusion capability and the like in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle engine bearing fault diagnosis, and in particular relates to a vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion. Background Art

[0002] As a core component of the engine, the health status of automobile engine bearings directly affects the operational safety and reliability of the entire vehicle. However, due to the precision characteristics of the engine, it is extremely difficult to directly observe the health status of the bearings. In actual operation, even bearings produced in the same batch can show significantly different remaining service life due to factors such as working environment (such as urban highway driving, mountain driving, and bumpy roads), driving habits (sudden acceleration, sudden braking), and load changes. This uncertainty complicates engine maintenance strategies: for regular inspections, if bearings that have not reached their expected life are frequently inspected, a lot of resources will be consumed and unnecessary downtime will be increased; for delayed inspections, if the bearings are not inspected until they are close to the life limit, they may face the risk of serious failure, resulting in high repair costs and safety hazards.

[0003] The evolution of a bearing's remaining life is closely related to its early, minor faults. Once an early failure occurs, its remaining service life can be drastically reduced. Therefore, timely identification and treatment of early, minor faults are crucial to ensuring stable engine operation.

[0004] Existing technology primarily relies on single-sensor data acquisition for automobile engine bearing fault diagnosis. This involves using a vibration sensor to collect bearing vibration signals and combining them with traditional signal processing methods (such as spectrum analysis and wavelet transform) to extract features. A single sensor (such as a vibration sensor) cannot fully capture bearing fault information. For example, increased friction and temperature caused by lubrication problems cannot be directly reflected in the vibration signal. Early-stage microcracks exhibit extremely subtle changes in the vibration signal, making them difficult to directly identify through vibration analysis.

[0005] Furthermore, the fusion of heterogeneous multi-source data (such as vibration, sound, and temperature) often relies on splicing or weighted averaging, failing to consider the specificities of different data types, resulting in limited fusion effectiveness. Furthermore, early, weak fault signals are highly similar to normal operating signals, making them difficult to distinguish based on simple fusion of multi-source data, leading to a high rate of misdiagnosis.

[0006] Therefore, there is an urgent need for an intelligent fault diagnosis solution for automobile engine bearings to solve key problems in existing technologies such as incomplete feature coverage and insufficient noise reduction and fusion capabilities. Summary of the Invention

[0007] The present invention provides a vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion to solve the problems in the prior art of incomplete bearing fault information collection and difficulty in identifying fault types due to single data collection, as well as insufficient noise reduction and fusion capabilities and high diagnostic misjudgment rate due to insufficient multi-source data fusion.

[0008] The technical solution adopted in the present invention is:

[0009] A vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion, comprising:

[0010] Based on the collected multidimensional data of the vehicle engine bearing at an early stage of a minor fault, multiple subspaces are projected and weighted according to the relevance of the subspaces using a multi-head attention mechanism to obtain fused features, wherein the multidimensional data includes at least vibration data and sound data of the vehicle engine bearing in three orthogonal directions;

[0011] According to the fused features, energy values ​​of multiple features are obtained one by one according to multiple neurons through a parameter-free attention mechanism, and feature processing is performed on the fused features according to the energy values ​​to reduce noise;

[0012] According to the fused features after noise reduction, deep features are gradually extracted through multi-layer two-dimensional convolution operations, and the fault type is determined through the fully connected layer to obtain a fault diagnosis model.

[0013] The vehicle engine bearing fault diagnosis method based on multidimensional data fusion disclosed in the present invention also has the following additional technical features:

[0014] The multi-dimensional data collected for the vehicle engine bearing at the early stage of a minor fault is as follows:

[0015] Collecting multi-dimensional data of the vehicle engine bearing under multiple fault types and multiple health states, wherein the fault types include at least inner race fault, outer race fault, and ball fault, and the health states include at least early minor fault, minor fault, medium fault, and severe fault;

[0016] The fault diagnosis model is trained based on the multidimensional data of the early minor fault.

[0017] The health status includes at least early minor faults, minor faults, medium faults, and major faults, specifically:

[0018] Collecting fault size data of the vehicle engine bearing under different fault types throughout its life cycle, and obtaining the fault sizes corresponding to early minor faults under different fault types;

[0019] According to the plurality of fault sizes and in combination with the corresponding fault type probabilities, a comprehensive fault size corresponding to the early minor fault of the vehicle engine bearing is obtained.

[0020] The specific method for collecting multidimensional data is as follows:

[0021] Collect multi-dimensional data of vehicle engine bearings under various artificial faults in laboratory scenarios to simulate the full life cycle of vehicle engine bearings under different fault types;

[0022] Gaussian white noise is added to the multidimensional data one by one, and the intensity of the Gaussian white noise is set differently according to the intensity of the collected multidimensional data.

[0023] Through the multi-head attention mechanism, the weights are set one by one according to the relevance of the subspaces, specifically:

[0024] Through single-head attention, multiple subspaces are processed one-to-one;

[0025] A query matrix, a key matrix and a value matrix are obtained according to the subspace, and a probability distribution is obtained according to the correlation between the query matrix and the key matrix to set a weight for the value matrix.

[0026] Get the fusion features, specifically:

[0027] Through the multi-head attention mechanism, the weighted subspaces are spliced ​​and the fusion features are obtained through dimension conversion.

[0028] The energy values ​​of multiple features are obtained one by one according to multiple neurons, specifically:

[0029] By combining the features of the neuron input, an energy value corresponding to the feature is obtained;

[0030] An enhancement suppression factor is set according to the energy value for feature processing, and the enhancement suppression factor is negatively correlated with the energy value.

[0031] The training of multi-head attention mechanism and parameter-free attention mechanism is as follows:

[0032] The loss value is obtained by using the mean square loss based on the difference between the data before and after denoising and the residual signal after denoising.

[0033] According to the loss value, the parameters of the multi-head attention mechanism and the parameter-free attention mechanism network are optimized through back propagation.

[0034] Fault diagnosis model, specifically:

[0035] The fault type is determined by the fault diagnosis model based on the multi-dimensional data of the vehicle engine bearing operation collected in real time during the vehicle operation.

[0036] The present invention also provides a processing device, comprising:

[0037] memory for storing computer programs;

[0038] A processor is used to implement the steps of the vehicle engine bearing fault diagnosis method based on multidimensional data fusion when executing the computer program.

[0039] Due to the adoption of the above technical solution, the beneficial effects achieved by the present invention are as follows:

[0040] 1. In this invention, multidimensional data is first collected from a vehicle engine bearing under early-stage minor faults, including vibration data in three orthogonal directions (X / Y / Z) and sound data. Using a multi-head attention mechanism, the multidimensional data is projected into multiple subspaces, and weights are dynamically assigned based on the correlations between the subspaces to generate fused features. The multi-head attention mechanism dynamically suppresses noisy areas (such as random vibration interference) and enhances effective feature areas (such as weak crack signals) by learning information from different subspaces in parallel (such as the strong correlation between sound signals and Z-axis vibration).

[0041] Based on the fused features, a parameter-free attention mechanism (SimAM) calculates the characteristic energy values ​​of multiple neurons and performs noise reduction on the fused features based on the energy values. SimAM uses the energy function to identify high-energy regions (noise) and weaken their impact, while enhancing low-energy regions (effective features). For early-stage minor faults (such as a 0.42mm outer ring crack), SimAM uses energy values ​​to filter effective signals and suppress environmental noise (such as bumps and wind noise), making weak signal mutations easier to identify.

[0042] Based on the denoised fused features, deep features are gradually extracted through multi-layer two-dimensional convolution operations. Fault types are then determined through fully connected layers to generate a fault diagnosis model. Large-size convolution kernels are used to expand the receptive field, capturing local details in high-noise environments (such as signal fluctuations caused by tiny cracks), improving accuracy in noisy data. A multi-layer convolutional architecture (Conv+ReLU+Dropout) gradually extracts deep features. Combined with Dropout (0.12) regularization, this architecture maintains high generalization performance and reduces the risk of overfitting.

[0043] The present invention forms a complete fault diagnosis process through multi-source heterogeneous data fusion (vibration, sound), adaptive noise reduction and improved convolutional neural network, which solves the problems of incomplete feature coverage and insufficient noise reduction and fusion capabilities in the existing technology, and significantly improves the accuracy and robustness of automobile engine bearing fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0045] Figure 1 The figure is a flow chart of the vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion according to one embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to more clearly illustrate the overall concept of the present invention, a detailed description is given below in an exemplary manner in conjunction with the accompanying drawings.

[0047] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0048] like Figure 1 As shown, a vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion includes:

[0049] S100: Based on the collected multidimensional data of the vehicle engine bearing at the early stage of a minor fault, multiple subspaces are projected, and weights are set one by one according to the correlation of the subspaces through a multi-head attention mechanism to obtain fusion features, wherein the multidimensional data at least includes vibration data and sound data of the vehicle engine bearing in three orthogonal directions.

[0050] The core goal of this step is to improve the comprehensiveness of feature coverage and model robustness of early minor fault diagnosis of vehicle engine bearings through multi-source heterogeneous data fusion and adaptive feature weighting.

[0051] Understandably, existing technologies rely on a single sensor (such as a vibration sensor) and are unable to fully capture complex fault characteristics such as lubrication anomalies (temperature changes) and microcracks (subtle fluctuations in the vibration signal). This step integrates vibration data from three orthogonal directions (X / Y / Z) and sound data to cover a more comprehensive range of fault modes, addressing the incomplete data coverage of a single sensor in existing technologies.

[0052] In addition, this step uses the multi-head attention mechanism to dynamically allocate subspace weights, suppress noise areas (such as random vibration interference), enhance effective feature areas (such as weak crack signals), and improve the adaptability of feature fusion.

[0053] This system collects multi-source heterogeneous data, including vibration data. Three accelerometers are used to collect bearing vibration signals in the three orthogonal directions of X, Y, and Z, covering mechanical anomalies (such as cracks and wear). Microphones are used to collect sound signals, reflecting lubrication conditions (such as changes in acoustic signals caused by increased friction). This step significantly improves feature coverage by supplementing vibration and sound data. After integrating this multi-source data, the model's accuracy in identifying early-stage minor faults (such as a 0.42mm outer ring crack) has improved.

[0054] In this step, the multi-head attention mechanism sets weights based on the relevance of each subspace. This means that each single attention head processes the data in one subspace. Therefore, it is necessary to divide the multidimensional data into multiple subspaces in advance. Evenly split the high-dimensional feature space (e.g., 512 dimensions) into multiple low-dimensional subspaces (e.g., 8 subspaces, each with 64 dimensions).

[0055] Each subspace is learned by an independent attention head. Through single-head attention, the corresponding subspace information is learned in parallel, and weights are dynamically assigned by calculating correlation scores. The weighted subspaces output by the multi-head attention mechanism are concatenated and transformed using the transposed matrix to obtain the fused features.

[0056] This step solves the problems of incomplete feature coverage and insufficient noise reduction capabilities in existing technologies through multi-source heterogeneous data fusion and multi-head attention mechanism, significantly improving the accuracy and robustness of early diagnosis of minor faults in automobile engine bearings.

[0057] S200: According to the fused features, energy values ​​of multiple features are obtained one by one according to multiple neurons through a parameter-free attention mechanism, and feature processing is performed on the fused features according to the energy values ​​to reduce noise.

[0058] The core goal of this step is to adaptively denoise the fused features through the parameter-free attention mechanism (SimAM) to further improve the accuracy and robustness of early-stage weak fault diagnosis of automobile engine bearings.

[0059] Understandably, existing noise reduction methods (such as low-pass filtering and Kalman filtering) rely on fixed rules and are difficult to adapt to complex noise environments (such as vehicle bumps, wind noise, and road noise). This step, however, uses the SimAM mechanism to dynamically identify and suppress noisy areas (such as high-energy interference signals) while enhancing effective feature areas (such as weak crack signals).

[0060] This step uses the energy function to filter effective signals, making it easier to identify weak fault feature mutations, preventing early weak fault signals (such as a 0.42mm outer ring crack) from being masked by environmental noise, and improving the model's ability to identify early weak faults.

[0061] The parameter-free attention mechanism (SimAM) directly optimizes feature regions using an energy function, eliminating the need for explicit gradient calculations and reducing computational complexity. The input fusion features are obtained in the previous step. It should be noted that in this step, the noise energy values ​​in high-energy regions are high and need to be suppressed. In low-energy regions, the effective feature energy values ​​are low and need to be enhanced.

[0062] In this step, the energy value of each neuron in the fusion feature is first calculated for subsequent noise reduction. It can be understood that high energy values ​​correspond to noise signals (such as random vibration interference, environmental noise), and their weights are suppressed by the energy function (for example, the energy value is mapped to a value close to 1, indicating low priority). Low energy values ​​correspond to effective features (such as weak crack signals, lubrication abnormality signals), and their weights are enhanced by the energy function (for example, the energy value is mapped to a value close to 0, indicating high priority). In this way, the fusion feature after noise reduction is obtained (the noise area is suppressed and the effective feature is amplified).

[0063] The previous step, using the multi-head attention mechanism (Transformer), initially suppressed some noise, but did not completely eliminate environmental interference (such as audio noise collected by the microphone). This step further reduces noise, and the resulting noise reduction features serve as input for subsequent steps (such as multi-layer convolution operations) to further extract deep features.

[0064] It's important to note that traditional noise reduction methods, such as low-pass filtering and fixed-rule methods, have limited effectiveness in complex noise environments (such as turbulence and wind noise), resulting in minimal improvement in the signal-to-noise ratio (SNR). This method, using the SimAM mechanism, improves the SNR by 15dB and reduces the mean square error (MSE).

[0065] Furthermore, in this step, SimAM uses energy values ​​to filter valid signals, making it easier to identify subtle fault characteristic changes. For example, the Z-axis vibration signal fluctuations caused by an outer ring crack are significantly amplified, while ambient noise (such as wind noise) is suppressed.

[0066] This step uses a parameter-free attention mechanism (SimAM) to adaptively reduce the noise of the fused features, solving problems such as insufficient noise reduction capabilities and difficulty in identifying early weak faults in existing technologies. It significantly improves the accuracy and robustness of automobile engine bearing fault diagnosis under noisy data.

[0067] S300: Based on the fused features after noise reduction, deep features are gradually extracted through multi-layer two-dimensional convolution operations, and the fault type is determined through a fully connected layer to obtain a fault diagnosis model.

[0068] The core goal of this step is to gradually extract deep features through multi-layer two-dimensional convolution operations, and use fully connected layers to complete fault classification and build a high-precision vehicle engine bearing fault diagnosis model.

[0069] Understandably, existing methods (such as traditional convolutional neural networks) struggle to capture local detail features (such as signal fluctuations caused by tiny cracks) in high-noise environments. The improved wide-core convolutional neural network (WDCNN) and multi-layer convolutional architecture significantly enhance feature extraction capabilities, resolving the issue of insufficient deep feature extraction capabilities in existing technologies. Multi-layer convolution gradually extracts cross-modal features (such as the coupling relationship between vibration and sound signals), combined with fully connected layers to complete fault classification, improving diagnostic accuracy and enhancing the model's classification accuracy.

[0070] Specifically, a large-size 31×1 convolution kernel is used to expand the receptive field to capture local detail features in high-noise environments (such as signal mutations caused by tiny cracks).

[0071]

[0072] Where k is the convolution kernel size (31×1), X is the input feature map, and W is the convolution kernel weight.

[0073] Multi-layer convolution gradually extracts deep features through a combination of multi-layer convolution + normalization + pooling. For example: the first layer of convolution extracts local texture features (such as crack mutations in vibration signals). The second layer of convolution fuses local features with global correlations (such as the coupling relationship between vibration signals and sound signals). During the convolution process, Batch Normalization (BN) is used to reduce internal covariate shift and accelerate training. Max Pooling is used to reduce the number of parameters and retain key features (such as the peak area of ​​the crack signal). The multi-layer convolution structure gradually extracts cross-modal features (such as the coupling relationship between vibration signals and sound signals) and increases the feature dimension (from 512 dimensions to 2048 dimensions).

[0074] The feature maps extracted by multi-layer convolution are flattened into one-dimensional vectors through the Flatten layer and input into the fully connected layer.

[0075] Fully Connected (F) = σ (FW + b)

[0076] Where F is the flattened eigenvector, W is the weight matrix, b is the bias term, and σ is the activation function (ReLU).

[0077] The last fully connected layer uses the Softmax function to output the fault probability distribution, the formula is:

[0078]

[0079] where z i is the score of the i-th type of fault, and n is the number of fault categories (such as inner race fault, outer race fault, ball fault).

[0080] In this step, the fused features after noise suppression in the first two steps are directly input into the convolutional network for feature extraction and classification. Multi-layer convolution gradually extracts cross-modal features (such as the coupling relationship between vibration signals and sound signals), and combines with the Softmax classifier to achieve high-precision fault diagnosis, reducing the misdiagnosis rate. The accuracy under noisy data is significantly higher than that of traditional methods.

[0081] As a preferred embodiment of the present invention, the multi-dimensional data of the vehicle engine bearing at the early stage of a minor fault is collected and specifically includes:

[0082] Collecting multi-dimensional data of the vehicle engine bearing under multiple fault types and multiple health states, wherein the fault types include at least inner race fault, outer race fault, and ball fault, and the health states include at least early minor fault, minor fault, medium fault, and severe fault;

[0083] The fault diagnosis model is trained based on the multidimensional data of the early minor fault.

[0084] The core goal of this implementation is to improve the generalization ability of the model and the accuracy of early fault identification by collecting multi-dimensional data under multiple fault types and multiple health states and training the fault diagnosis model based on early minor fault data.

[0085] Existing methods rely on data from a single fault type or health state, making it difficult to cover failure modes under complex operating conditions. This implementation collects data from multiple fault types (inner race, outer race, and ball bearing faults) and multiple health states (from early minor faults to severe faults), ensuring that the model learns comprehensive fault characteristics, addressing the incomplete feature coverage problem in existing technologies.

[0086] Furthermore, early-stage minor fault signals are weak and easily masked by noise, making them difficult to effectively identify using traditional methods. This implementation specifically collects early-stage minor fault data for training, making the model more sensitive to weak signal changes (such as a 0.42mm outer ring crack), thereby enhancing the diagnostic capabilities of early-stage minor faults.

[0087] Bearings generally consist of three parts: the inner ring, outer ring, and ball bearings. Therefore, the fault types include inner ring fault (IR-1), which simulates inner ring cracks; outer ring fault (OR-1), which simulates outer ring cracks; and ball fault (RO-1), which simulates ball surface damage.

[0088] Based on the degree of fault development, the health status of vehicle engine bearings is categorized as early-stage minor faults, minor faults, moderate faults, and severe faults. It is understood that as a vehicle engine bearing fault develops, the strength of the collected signal increases. The present invention uses data corresponding to early-stage minor faults to train a model, improving the model's ability to extract and sensitivity to subtle characteristic signals. This allows the model to identify faults in their early, minor states, minimizing the lifespan reduction caused by fault development.

[0089] This embodiment solves the problems of incomplete feature coverage and insufficient early fault identification capabilities in the existing technology by collecting multi-dimensional data under multiple fault types and multiple health states, combined with special training on early minor fault data.

[0090] As an example of this implementation, the health status includes at least early minor fault, minor fault, medium fault, and major fault, specifically:

[0091] Collecting fault size data of the vehicle engine bearing under different fault types throughout its life cycle, and obtaining the fault sizes corresponding to early minor faults under different fault types;

[0092] According to the plurality of fault sizes and in combination with the corresponding fault type probabilities, a comprehensive fault size corresponding to the early minor fault of the vehicle engine bearing is obtained.

[0093] The core objective of this embodiment is to collect full life cycle data of vehicle engine bearings under different fault types, combine the fault type probability, and calculate the comprehensive fault size corresponding to early minor faults, so as to identify the multi-dimensional data corresponding to early minor faults for model training.

[0094] It should be noted that for the inner ring, inner ring, and ball of a bearing, when the fault size develops to different sizes, the impact on the bearing function is different. This embodiment comprehensively classifies the fault development degree by calculating the comprehensive fault size, thereby improving the accuracy of the fault development degree classification.

[0095] This example collects fault dimension data for various fault types, including inner race faults (IR-1), outer race faults (OR-1), and ball bearing faults (RO-1), over the entire lifecycle (early minor faults, minor faults, moderate faults, and severe faults). This diverse fault type allows the model to adapt to different failure modes (such as the difference between inner race cracks and ball bearing damage), significantly improving generalization capabilities.

[0096] Based on the impact of fault development on bearing performance, the faults are categorized into early-stage minor faults, minor faults, medium faults, and severe faults. The corresponding sizes of early-stage minor faults for different fault types are determined. For example, for early-stage minor faults, the fault size is small (such as a 0.42mm crack), and the signal change is extremely subtle.

[0097] The early minor fault sizes for different fault types (e.g., 0.43mm for the inner ring, 0.42mm for the outer ring, and 0.49mm for the ball) are weighted based on the probability distribution of fault types over the entire life cycle of the bearing (e.g., 40% probability of inner ring failure, 30% probability of outer ring failure, and 30% probability of ball failure).

[0098]

[0099] where p i is the probability of type i failure, d i is the size of the early minor fault corresponding to the fault type. Specifically,

[0100] Comprehensive fault size = 0.4 × 0.43 + 0.3 × 0.42 + 0.3 × 0.49 = 0.443 mm

[0101] The comprehensive fault size (0.443 mm) reflects the typical characteristics of early minor bearing faults and improves the accuracy of fault severity classification.

[0102] This embodiment solves the problem of insufficient early fault identification capability in the prior art by collecting fault size data throughout the entire life cycle and calculating the comprehensive fault size.

[0103] Specifically, the multi-dimensional data collection method is as follows:

[0104] Collect multi-dimensional data of vehicle engine bearings under various artificial faults in laboratory scenarios to simulate the full life cycle of vehicle engine bearings under different fault types;

[0105] Gaussian white noise is added to the multidimensional data one by one, and the intensity of the Gaussian white noise is set differently according to the intensity of the collected multidimensional data.

[0106] The core goal of this step is to construct a fault dataset covering the entire life cycle of vehicle engine bearings by collecting artificial fault data and injecting Gaussian white noise in laboratory scenarios, so as to improve the robustness and generalization ability of the fault diagnosis model in complex noise environments.

[0107] By collecting artificial fault data in laboratory scenarios, we ensure that the model learns comprehensive characteristics of different fault types (inner race, outer race, ball bearing) and health states (early minor to severe faults). By varying the intensity of Gaussian white noise to simulate a real-world noisy environment, we improve the model's diagnostic accuracy in noisy data.

[0108] Collect artificial fault data in laboratory scenarios, and create artificial faults (such as cracks in the inner ring, outer ring, and ball) through electrical discharge machining (EDM).

[0109] For vibration data, three accelerometers were used to collect vibration signals in the three orthogonal directions of X, Y, and Z. For sound data, a microphone was used to capture the sound signals during bearing operation. Three fault types were set: inner race fault (IR-1), outer race fault (OR-1), and ball bearing fault (RO-1), and data was collected for each.

[0110] In the experiment, cracks of a specified size (e.g., a 0.42mm outer ring crack) were created on the bearing surface using EDM. Engine operating conditions (speed range: 3500-4500 rpm) were simulated on a laboratory bench. Vibration and acoustic data were collected, and fault type and health status labels were recorded.

[0111] Based on the impact of artificial fault size on bearing performance, we categorize them into early minor faults (0.42mm crack), minor faults (0.6mm crack), medium faults (1.0mm crack), and severe faults (1.5mm crack). This allows us to select data corresponding to early minor faults for model training.

[0112] Differentially add Gaussian white noise,

[0113]

[0114] Among them, x clean is the original data collected in the laboratory, x noisy is the data after adding Gaussian white noise.

[0115] The noise intensity is differentiated. For vibration signals, σ=0.1 simulates the background vibration interference of bumpy roads. For sound signals, σ=0.2 simulates external interference such as wind noise and road noise.

[0116] In this embodiment, independent white Gaussian noise is added to the vibration signal (X / Y / Z) and the sound signal, and the noise intensity is dynamically adjusted according to the amplitude of the original data (for example, the higher the amplitude of the vibration signal, the greater the noise intensity σ).

[0117] This method improves the signal-to-noise ratio (SNR) in laboratory data containing Gaussian white noise, aligning with real-world interference. Even under complex operating conditions such as bumpy roads and mountain roads, the model can still reliably identify early-stage faults, improving its robustness to complex noise and reducing misdiagnosis rates.

[0118] As a preferred embodiment of the present invention, a multi-head attention mechanism is used to set weights based on the relevance of subspaces, specifically:

[0119] Through single-head attention, multiple subspaces are processed one-to-one;

[0120] Obtaining a query matrix, a key matrix, and a value matrix according to the subspace, and obtaining a probability distribution according to the correlation between the query matrix and the key matrix to set a weight for the value matrix;

[0121]

[0122] Among them, Q, K, and V represent the query matrix, key matrix, and value matrix respectively, d k Represents the dimensions of the bond matrix.

[0123] The core goal of this implementation is to process multiple subspaces through single-head attention in the multi-head attention mechanism and dynamically calculate the correlation weights between subspaces to enhance the model’s ability to integrate and reduce noise on the multi-dimensional data features of vehicle engine bearings.

[0124] Existing methods rely on simple concatenation or fixed weight assignment, making it difficult to capture the dynamic correlations between multidimensional data (such as vibration and sound signals). This invention, however, uses a single-headed attention mechanism to process the query (Q), key (K), and value (V) matrices within each subspace separately, dynamically calculating the correlation weights between subspaces and achieving more refined feature fusion.

[0125] Each subspace is processed by an independent single-head attention mechanism. For the input data of each subspace (such as the X / Y / Z direction data of the vibration signal), a query matrix Q, a key matrix K, and a value matrix V are generated through linear transformation.

[0126] Q=XW Q , K=XW K , V=XW V

[0127] Where X is the input feature matrix, W Q ,W K ,W V is the learnable weight matrix.

[0128] Calculate the similarity (dot product) between Q and K and generate a probability distribution (attention weight) through the softmax function:

[0129]

[0130] Among them, d k is the dimension of the key matrix. Prevent the dot product from being too large and causing the gradient to disappear.

[0131] Perform weighted summation on V according to the attention weight to obtain the output feature of the current subspace.

[0132] Multi-head attention processes multiple subspaces in parallel, dynamically calculating correlation weights to achieve refined feature fusion. Furthermore, multi-head parallel processing of multiple subspaces improves the model's adaptability to complex patterns and computational efficiency.

[0133] This embodiment processes multiple subspaces through single-head attention in the multi-head attention mechanism, solving the problem of insufficient feature fusion in the prior art.

[0134] As a preferred embodiment of this implementation, the fusion features are obtained as follows:

[0135] Through the multi-head attention mechanism, the weighted subspaces are spliced ​​and the fusion features are obtained through dimension conversion;

[0136] MultiHead(Q,K,V)=Concat(h1,h2,...,h h )W O

[0137] Among them, h i represents the weighted subspace, W O Represents the transposed matrix, used for dimension conversion.

[0138] The core goal of this embodiment is to integrate multiple weighted subspace features into high-dimensional fusion features through the splicing and dimensionality conversion operations of the multi-head attention mechanism, so as to improve the model's global correlation modeling ability and diagnostic accuracy for multi-dimensional data of vehicle engine bearings.

[0139] In this embodiment, the weighted subspace (h i ) and apply the transposed matrix W O , achieving efficient integration and dimension mapping of multi-dimensional features, enhancing the model's ability to model the global correlation of multi-dimensional data, and significantly improving feature fusion accuracy. Subspace splicing weighting and dimension conversion operation (W O ) further optimizes the expression ability of the fusion features to adapt them to downstream tasks (such as classification or regression). h1, h2, ...., h h is the weighted subspace feature, W Ois the transposed matrix, which is used to map the concatenated low-dimensional features back to the original dimension.

[0140] All weighted subspaces h i , spliced ​​into a high-dimensional matrix H by column concat , whose dimension is (n, h·d v ), where n is the sequence length and h is the number of heads.

[0141] H concat =Concat(h1,h2,...,h h )

[0142] By transposing the matrix W O For H concat Perform linear transformation to generate the final fusion feature MultiHead (Q, K, V), whose dimension is (n, d model ).

[0143] MultiHead(Q,K,V)=Concat(h1,h2,...,h h )W O

[0144] It should be noted that W O The dimension is (h·d v , d model ), obtained by training the fully connected layer. Specifically, if h = 8, d v =64,d model =512, then W O The dimensions are (8×64,512).

[0145] This embodiment solves the problem of low feature fusion efficiency in the prior art through subspace splicing and dimensionality conversion operations of the multi-head attention mechanism.

[0146] As a preferred embodiment of the present invention, the energy values ​​of multiple features are obtained one by one according to multiple neurons, specifically:

[0147] By combining the features of the neuron input, an energy value corresponding to the feature is obtained;

[0148]

[0149] Among them, w, b are weight vectors and bias terms respectively, and x represents the input features;

[0150] An enhancement suppression factor is set according to the energy value for feature processing, and the enhancement suppression factor is negatively correlated with the energy value.

[0151] The core goal of this implementation is to calculate the energy value of the input feature by combining neurons with the Sigmoid function to quantify the adaptability or importance of the feature, thereby improving the model's ability to identify key features and nonlinear modeling capabilities.

[0152] The Sigmoid function maps the linear combination to the (0,1) interval, generating nonlinear energy values ​​and enhancing the model's sensitivity to key features. The smoothing properties of the Sigmoid function avoid the vanishing gradient problem caused by traditional hard threshold functions (such as step functions), addressing the linearization problem of feature evaluation in existing technologies.

[0153] Combining neurons with the Sigmoid function, the input feature x is linearly combined to generate the intermediate variable z:

[0154] z=w T x+b inputs the linear combination result z into the Sigmoid function to generate the energy value E.

[0155]

[0156] The output range of the Sigmoid function is (0,1), which can be interpreted as the adaptability probability of the input feature (such as the probability of a fault occurring). Specifically, if w T x+b=5, then E≈0.993; if w T If x+b=-5, then E≈0.007.

[0157] The Sigmoid function uses nonlinear mapping to more accurately identify high-value features (such as crack mutation points in vibration signals). It quantifies the adaptive probability of neuron outputs and assigns dynamic energy values ​​to input features, improving the model's ability to identify key features.

[0158] Based on the energy value, an enhancement / suppression factor is set for feature processing. The enhancement / suppression factor is negatively correlated with the energy value. In this step, noise energy values ​​are high in high-energy regions and need to be suppressed. In low-energy regions, effective feature energy values ​​are low and need to be enhanced. Therefore, an enhancement / suppression factor negatively correlated with the energy value is set to suppress noise regions and amplify effective features.

[0159] In this step, the energy value of each neuron in the fusion feature is first calculated for subsequent noise reduction. It can be understood that high energy values ​​correspond to noise signals (such as random vibration interference, environmental noise), and their weights are suppressed by the energy function (for example, the energy value is mapped to a value close to 1, indicating low priority). Low energy values ​​correspond to effective features (such as weak crack signals, lubrication abnormality signals), and their weights are enhanced by the energy function (for example, the energy value is mapped to a value close to 0, indicating high priority). In this way, the fusion feature after noise reduction is obtained by enhancing the suppression factor (the noise area is suppressed and the effective feature is amplified).

[0160] This embodiment calculates the energy value of the input feature by combining neurons with the Sigmoid function, thereby solving the problem of linearization of feature evaluation in the prior art.

[0161] As a preferred embodiment of the present invention, the training of the multi-head attention mechanism and the parameter-free attention mechanism is specifically as follows:

[0162] The loss value is obtained by using the mean square loss based on the difference between the data before and after denoising and the residual signal after denoising.

[0163] According to the loss value, the parameters of the multi-head attention mechanism and the parameter-free attention mechanism network are optimized through back propagation.

[0164] The core goal of this implementation is to improve the accuracy and robustness of the model in denoising tasks by combining the difference and residual signals of the data before and after denoising, designing the loss function using the mean square error (MSE), and optimizing the parameters of the multi-head attention mechanism and the parameter-free attention mechanism through backpropagation.

[0165] By constructing a loss function based on the difference before and after denoising, the parameter optimization of the attention mechanism is directly linked to the signal reconstruction goals (such as noise suppression and feature retention), solving the problem of unclear training objectives of the attention mechanism in the existing technology.

[0166] Through the loss function, the point-by-point difference Δx=x between the input signal and the noise reduction signal is calculated. noisy -x denoised The difference reflects the model's ability to suppress noise (ideally, the difference should be close to the original noise).

[0167] The difference Δx and the residual signal r are combined to construct the total loss function L:

[0168]

[0169] Among them, n true =x noisy -x cleanis the actual noise, and α and β are weight coefficients. Note that if only noise suppression is important, α>β can be set; if both signal fidelity and noise suppression are important, α≈β can be set.

[0170] For the multi-head attention mechanism, during the back propagation process, the gradient is transferred from the loss function L to the multi-head attention layer to update the weight matrix W. q (Query), W k (key), W v (value) to minimize signal error.

[0171] For the parameter-free attention mechanism, the loss function L is back-propagated to adjust the distribution strategy of the attention weight (such as adjusting the sliding window size or the granularity of the position encoding) to improve the signal reconstruction quality.

[0172] As a preferred embodiment of the present invention, the fault diagnosis model is specifically:

[0173] The fault type is determined by the fault diagnosis model based on the multi-dimensional data of the vehicle engine bearing operation collected in real time during the vehicle operation.

[0174] The core goal of this implementation is to achieve accurate identification and classification of fault types by real-time collection of multi-dimensional data on vehicle engine bearing operation, combined with fault diagnosis models, thereby improving the real-time, accuracy and reliability of vehicle fault detection.

[0175] Real-time vibration signal acquisition: The radial and axial vibration signals of the bearing are collected through acceleration sensors (sampling rate ≥ 10kHz). Sound data is collected through a sound receiving device.

[0176] The data input model receives multi-dimensional data (vibration signals, sounds), outputs the failure probability through the model (such as crack probability 85% > threshold 80%), and triggers early warning signals (such as flashing dashboard, voice prompts).

[0177] This implementation solves the problems of lag and high misdiagnosis rate of traditional fault diagnosis methods by real-time collection of multi-dimensional data and combining it with a deep learning model.

[0178] The present invention again provides a processing device, comprising:

[0179] memory for storing computer programs;

[0180] A processor is used to implement the steps of the vehicle engine bearing fault diagnosis method based on multidimensional data fusion when executing the computer program.

[0181] Therefore, any effect of the vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion can be achieved, which will not be elaborated here.

[0182] Anything not described in the present invention can be achieved by adopting or drawing on existing technologies.

[0183] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0184] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion, characterized in that: include: Based on the collected multidimensional data of the vehicle engine bearing at an early stage of a minor fault, multiple subspaces are projected and weighted according to the relevance of the subspaces using a multi-head attention mechanism to obtain fused features, wherein the multidimensional data includes at least vibration data and sound data of the vehicle engine bearing in three orthogonal directions; According to the fused features, energy values ​​of multiple features are obtained one by one according to multiple neurons through a parameter-free attention mechanism, and feature processing is performed on the fused features according to the energy values ​​to reduce noise; According to the fused features after noise reduction, deep features are gradually extracted through multi-layer two-dimensional convolution operations, and the fault type is determined through the fully connected layer to obtain a fault diagnosis model.

2. The vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion according to claim 1, characterized in that: The multi-dimensional data collected for the vehicle engine bearing at the early stage of a minor fault is as follows: Collecting multi-dimensional data of the vehicle engine bearing under multiple fault types and multiple health states, wherein the fault types include at least inner race fault, outer race fault, and ball fault, and the health states include at least early minor fault, minor fault, medium fault, and severe fault; The fault diagnosis model is trained based on the multidimensional data of the early minor fault.

3. The vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion according to claim 2 is characterized in that: The health status includes at least early minor faults, minor faults, medium faults, and major faults, specifically: Collecting fault size data of the vehicle engine bearing under different fault types throughout its life cycle, and obtaining the fault sizes corresponding to early minor faults under different fault types; According to the plurality of fault sizes and in combination with the corresponding fault type probabilities, a comprehensive fault size corresponding to the early minor fault of the vehicle engine bearing is obtained.

4. The vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion according to claim 3 is characterized in that: The specific method for collecting multidimensional data is as follows: Collect multi-dimensional data of vehicle engine bearings under various artificial faults in laboratory scenarios to simulate the full life cycle of vehicle engine bearings under different fault types; Gaussian white noise is added to the multidimensional data one by one, and the intensity of the Gaussian white noise is set differently according to the intensity of the collected multidimensional data.

5. The vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion according to claim 1, characterized in that: Through the multi-head attention mechanism, the weights are set one by one according to the relevance of the subspaces, specifically: Through single-head attention, multiple subspaces are processed one-to-one; A query matrix, a key matrix and a value matrix are obtained according to the subspace, and a probability distribution is obtained according to the correlation between the query matrix and the key matrix to set a weight for the value matrix.

6. The vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion according to claim 5, characterized in that: Get the fusion features, specifically: Through the multi-head attention mechanism, the weighted subspaces are spliced ​​and the fusion features are obtained through dimension conversion.

7. The vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion according to claim 1, characterized in that: The energy values ​​of multiple features are obtained one by one according to multiple neurons, specifically: By combining the features of the neuron input, an energy value corresponding to the feature is obtained; An enhancement suppression factor is set according to the energy value for feature processing, and the enhancement suppression factor is negatively correlated with the energy value.

8. The vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion according to claim 1, characterized in that: The training of multi-head attention mechanism and parameter-free attention mechanism is as follows: The loss value is obtained by using the mean square loss based on the difference between the data before and after denoising and the residual signal after denoising. According to the loss value, the parameters of the multi-head attention mechanism and the parameter-free attention mechanism network are optimized through back propagation.

9. The vehicle engine bearing fault diagnosis method based on multi-dimensional data fusion according to claim 1, characterized in that: Fault diagnosis model, specifically: The fault type is determined by the fault diagnosis model based on the multi-dimensional data of the vehicle engine bearing operation collected in real time during the vehicle operation.

10. A processing device, characterized in that: include: memory for storing computer programs; A processor is configured to implement the steps of the vehicle engine bearing fault diagnosis method based on multidimensional data fusion as described in any one of claims 1 to 9 when executing the computer program.

Images