Mechanical Fault Diagnosis Method Based on Binary Aspect Ratio Network

Abstract

This invention discloses a mechanical fault diagnosis method based on binary aspect ratio networks. First, a binary neural network fault diagnosis model is constructed. Then, a performance gain evaluation model for fault diagnosis is established. Finally, a fault diagnosis model based on the binary aspect ratio network is established, and network reshaping is performed using pruning techniques to obtain the most suitable backbone network. Under comparable computational cost, the diagnostic accuracy is improved. Compared with full-precision networks, the proposed method significantly reduces computational complexity while maintaining a high level of diagnostic accuracy. Comprehensive experiments demonstrate that the proposed binary aspect ratio network fault diagnosis method has superior diagnostic performance and broad application prospects in mechanical fault diagnosis.

Description

Technical Field

[0001] This invention relates to artificial intelligence-based mechanical fault diagnosis technology, and in particular to a mechanical fault diagnosis method based on binary aspect ratio networks. Background Technology

[0002] Deep neural networks have led to significant innovation and breakthroughs in many fields, especially in the diagnosis of mechanical system faults, where many neural network models have been applied. However, due to the large number of parameters in deep neural network models, optimization and retraining become extremely difficult, greatly limiting their application in high real-time scenarios. As a method for compressing structures and accelerating computation, model quantization is considered a direct and effective technique for solving deep neural network problems. Quantization significantly reduces model size and computational complexity by quantizing full-precision (e.g., 32-bit) parameters to low-precision (e.g., 4-bit and 2-bit) parameters. Binary Neural Networks (BNNs), as a classic version of model quantization, quantize 32-bit floating-point numbers into 1-bit binary values, offering advantages such as low storage requirements, low computational load, and low latency. Although significant progress has been made in network binarization, these methods are mainly based on existing full-precision network backbones. It is well known that full-precision networks tend to have deep and narrow backbones; increasing network depth greatly enhances the network's representational capabilities. However, the quantization and gradient errors introduced by BNNs accumulate layer by layer with increasing network depth, leading to a decline in network performance. Therefore, the gains from a deep and narrow backbone network are reduced in BNNs and may not be suitable for them. Research has found that, without considering overfitting, increasing either the network depth or width enhances its representational power. Increasing depth polynomially is as effective as increasing width exponentially. Therefore, it is necessary to reshape the BNN network structure to find a balance between depth and width to achieve optimal network performance. Summary of the Invention

[0003] Based on the aforementioned technical background, this invention provides a mechanical fault diagnosis method based on binary aspect ratio networks. By constructing a performance gain evaluation model and combining it with pruning techniques to reshape the network structure, the network structure of the model is further optimized, thereby improving diagnostic accuracy, reducing computational complexity, and enhancing diagnostic performance.

[0004] The present invention achieves the above objectives through the following technical solution: A mechanical fault diagnosis method based on binary aspect ratio networks, the specific steps of which are as follows:

[0005] 1) Establish a binary neural network fault diagnosis model: By binarizing the floating-point weights and activations of the network to "+1" or "-1", and replacing floating-point multiplication and accumulation with bitwise operations, storage space is saved and computational complexity is reduced;

[0006] 2) Establish a performance gain evaluation model for fault diagnosis, and find the balance point between depth and width to obtain the optimal network diagnostic performance. The formula is as follows:

[0007] G=f(d,w) (1)

[0008] In the formula: G represents the performance gain of the model, d represents the depth of the network, and w represents the width of the network; the reasonable range of depth is d∈[d min ,d max The reasonable range for the width is w∈[w min ,w max ];

[0009] When f(d,w)>0 and At this time, both increasing network depth and increasing network width improve diagnostic performance; generally, the gain from increasing depth is greater than the gain from increasing width; the performance gain slows down as d and w increase, therefore, there exists an equilibrium point d = d0, such that...

[0010] 3) Establish a fault diagnosis model based on a binary aspect ratio network to overcome the problem that the quantization error and gradient error introduced in the constructed BNN fault diagnosis model accumulate with increasing depth, leading to a decrease in model diagnostic performance; by introducing a scaling factor s to adjust the ratio of depth d and width w, a new depth d / s and width w ratio is found. By combining pruning techniques with network reshaping, unimportant parameters are removed, reducing model complexity, and the backbone is ultimately adjusted to... Without exceeding the computational cost of the original model, a more suitable BNN backbone is obtained by adjusting the network's depth and width to meet the demands for fast response and high accuracy in fault diagnosis; the formula is as follows:

[0011]

[0012] In the formula: This represents the actual performance gain of BNN, B b This represents the computational budget of the BNN; by restricting the reshaping strategy, equation (2) can be simplified to

[0013] This invention addresses the problem of drastically declining model performance caused by the reduction in the representation space of binary neural networks' output values, which limits model capacity and representational ability. An improved solution is proposed: an aspect ratio network is introduced, a performance gain evaluation model is established, and a fault diagnosis model based on a binary aspect ratio network is constructed. Experimental studies using this model are conducted, and the results are compared with the original BNN fault diagnosis method and the improved BNN fault diagnosis method. Under comparable computational cost, the diagnostic accuracy is improved. Compared with full-precision networks, the proposed method significantly reduces computational complexity while maintaining a high level of diagnostic accuracy. Comprehensive experiments demonstrate that the proposed binary aspect ratio network-based fault diagnosis method has superior diagnostic performance and broad application prospects in mechanical fault diagnosis. Attached Figure Description

[0014] Figure 1 This is an example diagram of the BNN model network architecture in this invention;

[0015] Figure 2 This is a structural diagram of the DWO method in this invention;

[0016] Figure 3 This is a 3D visualization of the classification results of the ResNet-18-None-DWO model in this invention;

[0017] Figure 4 This is a 3D visualization of the classification results of the ResNet-18-SD-BNN-DWO model in this invention;

[0018] Figure 5 This is a three-dimensional visualization of the classification results of the ResNet-18-None-DWO model in this invention under SNR=-6dB noise conditions;

[0019] Figure 6 This is a three-dimensional visualization of the classification results of the ResNet-18-SD-BNN-DWO model in this invention under SNR=-6dB noise conditions;

[0020] Figure 7 This is the confusion matrix diagram of the ResNet-18-None-DWO model in this invention;

[0021] Figure 8 This is the confusion matrix diagram of the ResNet-18-SD-BNN-DWO model in this invention;

[0022] Figure 9 This is the confusion matrix diagram of the ResNet-18-None-DWO model in this invention under the SNR=-6dB noise condition;

[0023] Figure 10This is the confusion matrix diagram of the ResNet-18-SD-BNN-DWO model in this invention under the SNR=-6dB noise condition. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. See also Figures 1 to 10 The mechanical fault diagnosis method based on binary aspect ratio networks has the following specific steps:

[0025] 1) Establish a binary neural network (BNN) fault diagnosis model: By binarizing the floating-point weights and activations of the network to "+1" or "-1", and replacing floating-point multiplication and accumulation with bitwise operations, storage space is saved and computational complexity is reduced;

[0026] 2) Establish a performance gain evaluation model for fault diagnosis, and find the balance point between depth and width to obtain the optimal network diagnostic performance; the formula is as follows:

[0027] G=f(d,w) (1)

[0028] In the formula: G represents the performance gain of the model, d represents the depth of the network, and w represents the width of the network; the reasonable range of depth is d∈[d min ,d max The reasonable range for the width is w∈[w min ,w max ];

[0029] When f(d,w)>0 and At this time, both increasing network depth and increasing network width improve diagnostic performance; generally, the gain from increasing depth is greater than the gain from increasing width; the performance gain slows down as d and w increase, therefore, there exists an equilibrium point d = d0, such that...

[0030] 3) A fault diagnosis model based on a binary aspect ratio network is established to overcome the problem of declining diagnostic performance caused by the accumulation of quantization and gradient errors with increasing depth in the constructed BNN fault diagnosis model. Network reshaping (DWO) is performed using pruning techniques to remove unimportant parameters, reduce model complexity, and finally adjust the backbone to... Without exceeding the computational cost of the original model, a more suitable BNN backbone is obtained by adjusting the network's depth and width to meet the demands for fast response and high accuracy in fault diagnosis; the formula is as follows:

[0031]

[0032] In the formula: This represents the actual performance gain of BNN, Bb This represents the computational budget of the BNN; by restricting the reshaping strategy, equation (2) can be simplified to

[0033] The specific implementation of this invention is as follows:

[0034] 1. Construct a mechanical fault diagnosis model based on a binary aspect ratio network:

[0035] Taking ResNet-18 as an example (e.g.) Figure 1 As shown, except for the first convolutional layer, the last layer, and the downsampling layer, all layers are binarized. Floating-point multiplication and accumulation are replaced with bitwise operations, and weights and activations are binarized to "+1" or "-1". First, the input is processed through the first convolutional layer, then through four stages, each consisting of two residual blocks with the same number of output channels, all using binary convolution. Finally, a classifier is connected after the four stages to classify the extracted features and output the result.

[0036] Based on this, a fault diagnosis model based on a binary aspect ratio network is built, and its process is as follows:

[0037] 1) First, a BNN fault diagnosis model is constructed. Except for the first and last layers and the downsampling layer, all layers are binarized to save storage space and reduce computational complexity. Next, a performance gain evaluation model for fault diagnosis is constructed. The optimal network diagnostic performance is obtained by finding a balance between depth and width. Compared to a deep and narrow backbone network, a shallow and wide backbone network is more suitable for BNNs. Finally, a fault diagnosis model based on a binary aspect ratio network is constructed to overcome the quantization and gradient errors introduced into the constructed BNN fault diagnosis model. Pruning techniques are used to reshape the network, removing unimportant parameters and reducing model complexity. Without exceeding the computational cost of the original model, a more suitable BNN backbone is obtained by adjusting the network's depth and width to meet the requirements of fast response and high accuracy in fault diagnosis.

[0038] 2) The network performance gain evaluation model was constructed, and the specific process is as follows:

[0039] Let G be the performance gain of the model, as shown in formula (1):

[0040] G=f(d,w) (1)

[0041] In the formula: d represents the network depth, and w represents the network width. A reasonable range for the depth is d ∈ [d...]. min ,d max The reasonable range for the width is w∈[w min ,w max ].

[0042] When f(d,w)>0 and At the same time, both the gain from increasing network depth and the gain from increasing network width can improve diagnostic performance;

[0043] Typically, given a calculation budget B, That is, the gain from increasing depth is greater than the gain from increasing width;

[0044] The performance gain slows down as d and w increase.

[0045] Therefore, there exists an equilibrium point d = d0 such that... The performance gains obtained by increasing either depth d or width w are equal. The gain from further increasing depth d will not be greater than the gain from increasing width w, that is:

[0046]

[0047] In the formula: τ>0, τ is a tiny increment.

[0048] In BNNs, if we only consider the variation in representation precision, that is, from a full-precision network to a 2... 32 Reduce to 2 1 This significantly reduces the network's representational power and leads to a decrease in task performance, but the backbone of a full-precision network is still suitable for BNNs, where the gain relationship between depth and width should remain unchanged. b (d,w) has properties similar to f(d,w), and can be expressed as:

[0049] G b =f b (d,w)<f(d,w) (3)

[0050] In the formula: G b This represents the performance gain of a binary neural network.

[0051] However, network binarization not only reduces the representation accuracy of each neuron, but also introduces quantization error and gradient error, which accumulate as the network depth increases.

[0052] Forward propagation: Quantization error is propagated from the current layer to the next layer and accumulates along the depth.

[0053]

[0054] In the formula: e i Indicates the real-valued activation at the i-th layer. and binarization activation The residuals between; This represents the binarized activation output of the i-th layer; B represents the binarized activation output of the (i+1)th layer; i+1 E represents the binary operation at level i+1; Q This indicates the quantization error.

[0055] Backpropagation: Gradient errors in deeper layers are also propagated to shallower layers and accumulate along the depth according to the chain rule.

[0056]

[0057] In the formula: E G represents the gradient error; L represents the network loss; Let L represent the gradient of the loss L with respect to the real-valued activations of the i-th layer; Let L represent the gradient of the loss L with respect to the binarized activation of the i-th layer; l represents the total number of layers in the network.

[0058] For convenience, quantization error and gradient error are fused into a single quantization loss function. Since both are related to the network depth, the quantization loss function accumulates along the depth and can therefore be expressed as:

[0059]

[0060] The larger the quantization loss function, the more severe the decline in the model's diagnostic performance. Therefore, the relationship between the quantization loss function and the model's performance gain is as follows:

[0061]

[0062] In the formula: G L This represents the gain resulting from quantization loss.

[0063] Therefore, the actual performance gain of BNN is If the gain from increasing depth is the same as that from increasing width, that is:

[0064]

[0065] Due to quantization loss, the difference in performance gain between increasing depth and width at this point is:

[0066]

[0067] Let lim(τ0→0). From formulas (7) and (8), we know that Δ<0. Therefore, the gain from increasing the width is greater than the gain from increasing the depth, that is:

[0068]

[0069] Compared to full-precision networks, BNNs shift their depth equilibrium point forward, thus creating a new equilibrium point d = d0. n <d0, such that:

[0070]

[0071] This indicates that, given a computational budget, the optimal configuration of the depth and width of a full-precision network is N. r When [d0, w0], the optimal configuration of the corresponding BNN is: In short, the quantization loss caused by binarization partially offsets the performance gain from increasing depth, causing BNNs to tend to have shallow and wide backbones rather than the deep and narrow backbones of full-precision networks.

[0072] 3) Construct a fault diagnosis model based on a binary aspect ratio network. By reshaping the depth and width of the existing full-precision network backbone, a balance point between depth and width is found to obtain a backbone network suitable for a BNN. That is, given a BNN... b The full-precision network N corresponding to [d0, w0] r [d0,w0], and calculate the budget B b (d0, w0), obtain N satisfying the following condition by reshaping the depth and width. b [d n ,w n ].

[0073]

[0074] In theory, it is possible to add or remove any convolutional layers in a network to adjust its depth, and it is also possible to fine-tune the number of channels in each layer to adjust its width. Therefore, it is difficult to accurately find [d]. n ,w n To ensure comparability of the network backbone before and after adjustment, similar to model scaling, all layers must be scaled uniformly at a constant ratio. However, the number of layers and channels in the network can only be positive integers, so absolute uniformity of adjustment cannot be guaranteed. Since the first and last layers in a BNN are usually real-valued layers, modifying their channels would significantly impact computational cost. Therefore, alignment layers are used to keep their channels constant. By restricting the reshaping strategy, equation (12) can be simplified to:

[0075]

[0076] In the formula: s represents the scaling factor.

[0077] Initial network reshaping is straightforward because the network backbone is determined by defining s. Furthermore, the range of s is very limited, resulting in a limited design space for the network. To further improve the backbone and achieve better performance gains, pruning techniques are combined with the initial reshaping to optimize the network's depth and width (e.g., ...). Figure 2(As shown). First, the original trunk N is reshaped through preliminary reconstruction. b [d,w] is reshaped into N b [d / s,sw], making the reshaped network backbone wider and shallower; then, through pruning techniques, clustering is performed first, and then unimportant parameter parts are pruned, limiting the computational cost to below that of the original backbone, while minimizing accuracy degradation, and finally adjusting the backbone to... Furthermore, the computational cost of the final model is comparable to the backbone obtained from the initial network reshaping. The computational cost is comparable, that is, Because the trimming during width scaling is uneven, w is used. * To make distinctions.

[0078] 2. Experimental verification:

[0079] To verify the effectiveness of the fault diagnosis model based on a binary aspect ratio network, this invention utilizes a dataset of intermediate bearings for aero-engines from a university for fault classification. Data collected from the acceleration response stored in column 5 is selected. Five operating conditions are included: two normal operating conditions; inner ring faults with fault dimensions of 0.5mm depth and 0.5mm length, and 0.5mm depth and 1.0mm length, respectively; and outer ring faults with fault dimensions of 0.5mm depth and 0.5mm length. Continuous wavelet transform is used to convert the vibration signal under each operating condition into a time-frequency graph. 504, 450, 504, 504, and 450 samples are extracted from the five operating conditions, totaling 2412 samples. The obtained time-frequency graph samples are divided into training and testing sets in a 7:3 ratio. This study investigates how the fault diagnosis method based on a binary aspect ratio network improves the robustness and generalization of the model, thereby improving absolute accuracy with comparable computational cost.

[0080] The Ops (operations) of a full-precision network are FLOPs (floating-point operations), while the Ops of a BNN include both FLOPs and BOPs (binary operations). The Ops of the BNN are calculated using OPs = BOPs / 64 + FLOPs. Experimental metrics are shown in Table 1: Without additional optimization methods, the DWO backbone improves accuracy by 2.21%; compared to other methods, the proposed method increases Ops by only a small amount but significantly improves accuracy. Furthermore, the optimization of the original backbone in the proposed method is constrained by computational cost, thus the computational cost before and after optimization is comparable. Compared to a full-precision network, the proposed method significantly reduces computational complexity while maintaining a high diagnostic accuracy, demonstrating excellent diagnostic performance.

[0081] Table 1 Comparison of results from different methods

[0082]

[0083] Figure 3 and Figure 4 To visualize the distribution of extracted features in three-dimensional space using the t-SNE algorithm, the model successfully completed the task of feature clustering for five bearing operating conditions and classification of different fault categories. As can be seen from the figure, the features extracted by the network structure obtained by the proposed method have good discriminability, and using certain additional optimization methods is beneficial to improving diagnostic accuracy. Figure 5 and Figure 6 Under noise conditions of SNR = -6dB, the distribution of extracted features in three-dimensional space was visualized using the t-SNE algorithm. Even under strong noise conditions, the model was able to perform feature clustering and classification of different fault categories for five bearing operating conditions very well. As can be seen from the figure, the features extracted by the network structure obtained by the proposed method have good discriminability, indicating that the method can effectively handle strong noise interference and learn more effective fault information from complex environments.

[0084] Figure 7 This is the confusion matrix of the diagnostic model obtained by the DWO method for ResNet-18 on the test set without any additional improvement methods. There are a total of 723 samples, with 11 samples being misclassified. Figure 8 This is the confusion matrix of the diagnostic model obtained by the DWO method for ResNet-18 with additional improvements, on the test set. There are a total of 723 samples, with 0 samples misclassified. Figure 9 The confusion matrix of the diagnostic model obtained by ResNet-18 using the DWO method under the condition of SNR=-6dB noise is the result of the diagnostic results on the test set. There are a total of 723 samples, and 29 samples were misclassified. Figure 10 The given data is the confusion matrix of the ResNet-18 diagnostic model obtained using the DWO method under noise conditions with an SNR of -6dB. Out of 723 samples, 4 were misclassified. The proposed method demonstrates high diagnostic accuracy and strong robustness in noisy environments. Furthermore, combining it with additional optimization methods further enhances its diagnostic performance.

[0085] This invention introduces a BNN-based fault diagnosis method, saving model storage space and reducing computational complexity. While the BNN-based fault diagnosis method accelerates model inference, it introduces quantization and gradient errors, which accumulate with increasing depth, leading to a decline in diagnostic performance. Therefore, a performance gain evaluation model is constructed to find the balance between depth and width to achieve optimal network diagnostic performance. Compared to deep and narrow backbone networks, the constructed BNN tends to have shallower and wider backbone networks. In the constructed binary aspect ratio network-based fault diagnosis model, the proposed DWO method first uniformly scales the depth and width of the full-precision network backbone, and then uses pruning techniques to further optimize the scaled network backbone, obtaining the final diagnostic model. Through comparative experiments on noise resistance and complexity, the experimental results show that accuracy is improved, computational complexity is reduced, and excellent diagnostic performance is achieved.

Claims

1. A mechanical fault diagnosis method based on binary aspect ratio networks, characterized in that, The specific steps are as follows: 1) Establish a binary neural network fault diagnosis model: By binarizing the floating-point weights and activations of the network to "+1" or "-1", and replacing floating-point multiplication and accumulation with bitwise operations, storage space is saved and computational complexity is reduced; 2) Establish a performance gain evaluation model for fault diagnosis, and find the balance point between depth and width to obtain the optimal network diagnostic performance. The formula is as follows: G=f(d,w) (1) In the formula: G represents the performance gain of the model, d represents the depth of the network, and w represents the width of the network; the reasonable range of depth is d∈[d min ,d max The reasonable range for the width is w∈[w min ,w max ]; When f(d,w)>0 and At this time, both increasing network depth and increasing network width improve diagnostic performance; generally, the gain from increasing depth is greater than the gain from increasing width; the performance gain slows down as d and w increase, therefore, there exists an equilibrium point d = d0, such that... 3) Establish a fault diagnosis model based on binary aspect ratio network to overcome the problem that the quantization error and gradient error introduced in the constructed BNN fault diagnosis model accumulate with the increase of depth, which leads to the decline of the model's diagnostic performance. By introducing a scaling factor s to adjust the ratio of depth d and width w, a new depth / s and width ratio can be found. By combining pruning techniques with network reshaping, unimportant parameters are removed, reducing model complexity, and the backbone is ultimately adjusted to... Without exceeding the computational cost of the original model, a more suitable BNN backbone is obtained by adjusting the network's depth and width to meet the demands for fast response and high accuracy in fault diagnosis; the formula is as follows: In the formula: This represents the actual performance gain of BNN, B b This represents the computational budget of the BNN; by restricting the reshaping strategy, equation (2) can be simplified to

Images