Small sample TBM main bearing rolling body fault state identification method and system

Abstract

The invention discloses a small sample TBM main bearing rolling body fault state identification method and system, and relates to the field of fault diagnosis, and the method comprises the steps: obtaining a to-be-identified vibration signal of a TBM main bearing rolling body, dividing the to-be-identified vibration signal through a time window, and obtaining a plurality of time window sequences; inputting any time window sequence into the trained fault identification neural network to obtain a fault state identification result; the fault identification neural network comprises a data embedding module, a stacking module and a fault identification module; the stacking module comprises cascaded feature extraction module groups; the feature extraction module group comprises a feature extraction module and a local feature enhancement module; the feature extraction module introduces a variance modulation mechanism into the sparse attention mechanism so as to determine the weight of the attention score according to the variance of the attention score; and the local feature enhancement module performs feature enhancement by adopting a multi-scale convolution kernel. According to the invention, the fault state identification precision and efficiency of the TBM main bearing in the small sample state are improved.

Description

Technical Field

[0001] The present application relates to the field of fault diagnosis, and in particular to a method and system for identifying the fault status of rolling elements of a small-sample TBM main bearing. Background Art

[0002] Tunnel Boring Machines (TBMs) are core equipment in modern underground engineering. Their main bearings play a crucial role in mechanically decoupling the cutterhead's rotational drive from the fore and aft structures. Their operating status is directly related to project safety and progress. However, current mainstream methods for main bearing fault identification still face significant challenges due to the difficulty in collecting monitoring data under extreme operating conditions, the scarcity of actual damage samples, and the weak energy of early fault signals. On the one hand, diagnostic methods based on empirical features or traditional signal processing, such as wavelet packet decomposition, envelope spectrum analysis, and empirical mode decomposition, rely on explicit prior models and expert knowledge, resulting in limited recognition performance in the presence of strong noise perturbations and signal nonstationarity. On the other hand, while recent emerging deep learning methods (such as deep convolutional neural networks, long short-term memory neural networks, and the Transformer architecture) have demonstrated promising performance on standard bearing datasets, they still struggle to effectively generalize or extract key features when faced with the small sample sizes, long time series, and weak anomaly signals found in the harsh operating conditions of TBMs, resulting in low fault identification accuracy. Despite recent progress in fault diagnosis under small-sample conditions, including meta-learning, metric learning, and transfer learning, these methods are primarily targeted at small-sized, ultra-high-speed main bearings in standard industrial scenarios. These small-sample learning methods are insufficiently adaptable to fault identification in oversized, ultra-low-speed TBM main bearings. First, the extremely harsh operating conditions of TBM main bearings make effective fault features in vibration signals extremely weak, exhibiting a "strong noise drowning out weak features" characteristic. The fault signal energy is extremely low, severely weakening the generalization foundation of the meta-learning model. Second, the time series sparsity caused by the ultra-low speed of TBM main bearings in small-sample scenarios makes it difficult for transfer learning to effectively capture key dynamic patterns. Third, the computational complexity is high: current small-sample fault identification methods often involve complex model structures and a large number of model parameters. In particular, methods based on the standard attention mechanism have extremely high computational overhead when processing long-time series condition monitoring data of TBM main bearings, resulting in long model response times and failing to meet the high response speed requirements for real-time fault identification of TBM main bearings. Summary of the Invention

[0003] The purpose of this application is to provide a small sample TBM main bearing rolling element fault state identification method and system, which can use a cascaded feature extraction module group to efficiently and accurately identify the fault state of the TBM main bearing rolling element in the case of a small number of samples, and when applied to hardware devices such as computers, it can reduce the calculation complexity of the hardware device, reduce the response time of the hardware device, and improve the calculation efficiency of the hardware device.

[0004] To achieve the above objectives, this application provides the following solutions: In a first aspect, the present application provides a method for identifying rolling element fault states of a TBM main bearing in a small sample size, comprising: Acquire a vibration signal to be identified of a rolling element of a TBM main bearing, and divide the vibration signal to be identified using a time window to obtain multiple time window sequences; Input any time window sequence into a trained fault recognition neural network to obtain a fault state recognition result of the TBM main bearing rolling element; the fault recognition neural network is trained using a small sample set; the small sample set is obtained by dividing the historical vibration signals of the TBM main bearing rolling element under different states by time windows; the different states are normal state, abrasion damage state, or scratch damage state; wherein one time window is one sample; The fault recognition neural network includes a data embedding module, a stacking module and a fault identification module connected in sequence; the stacking module includes a cascaded feature extraction module group; the feature extraction module group includes a feature extraction module and a local feature enhancement module connected in sequence; the feature extraction module is used to introduce a variance modulation mechanism in the sparse attention mechanism to determine the weight of the attention score according to the variance of the attention score; the local feature enhancement module is used to use a multi-scale convolution kernel for feature enhancement.

[0005] In a second aspect, the present application provides a small sample TBM main bearing rolling element fault state identification system, comprising: A time window sequence acquisition module is used to obtain the vibration signal to be identified of the rolling element of the TBM main bearing, and divide the vibration signal to be identified into multiple time window sequences by using time windows; The fault state identification module is used to input any time window sequence into a trained fault identification neural network to obtain a fault state identification result of the TBM main bearing rolling element.

[0006] According to the specific embodiments provided in this application, this application has the following technical effects: The present application provides a method and system for identifying the fault state of a rolling element of a TBM main bearing with a small sample size. The method uses a time window to divide a small number of historical vibration signals of a TBM main bearing rolling element in different states as a small sample set, and trains a fault identification neural network so that the fault identification neural network can accurately identify the fault state of the TBM main bearing rolling element in the case of a small sample weak signal. In addition, the feature extraction module of the fault identification neural network introduces a variance modulation mechanism into the sparse attention mechanism, which can dynamically enhance the signal-to-noise ratio of weak signal components and improve the ability to capture weak signals; the local feature enhancement module of the fault identification neural network can achieve differentiated extraction of features of different sizes by constructing a multi-scale convolution kernel, solving the problem of insufficient adaptability of single-scale feature extraction to diversified damage; the stacking module of the fault identification neural network uses a cascaded feature extraction module group to repeatedly extract features, which greatly improves the recognition accuracy of the fault state and the complexity of the calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0008] Figure 1 This is an application environment diagram of a small sample TBM main bearing rolling element fault state identification method in one embodiment of the present application; Figure 2 A flowchart of a method for identifying a rolling element fault state of a small sample TBM main bearing provided in one embodiment of the present application; Figure 3 A structural diagram of a fault identification neural network in a method for identifying a rolling element fault state of a small sample TBM main bearing provided in one embodiment of the present application; Figure 4 A schematic diagram of the functional modules of a small sample TBM main bearing rolling element fault status identification system provided in one embodiment of the present application. DETAILED DESCRIPTION

[0009] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0010] In order to make the purpose, features and advantages of this application more obvious and easy to understand, this application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0011] The small sample TBM main bearing rolling element fault state identification method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown in the figure, the main bearing rolling element vibration signal acquisition system communicates with the intelligent industrial computer through an industrial bus. The data storage system can store the vibration signals that the intelligent industrial computer needs to process. The data storage system can be integrated on the intelligent industrial computer, or it can be independently deployed on a local storage device or a cloud server. The main bearing rolling element vibration signal acquisition system collects the vibration signals of the TBM main bearing rolling elements in real time and transmits the vibration signals to the intelligent industrial computer. After receiving the vibration signal, the intelligent industrial computer divides the vibration signal into a time window of fixed length to obtain the time window data at the current moment; the time window sequence is input into the pre-trained lightweight fault recognition neural network model to obtain the real-time fault status recognition result of the TBM main bearing rolling elements. The recognition result is visualized through the status monitoring system. If the recognition result is a fault, the corresponding early warning mechanism is triggered.

[0012] The vibration signal acquisition system includes but is not limited to: a high-precision accelerometer, a signal conditioning module (anti-interference, amplification, filtering), an analog-to-digital converter (24-bit ADC), etc. Features of the intelligent industrial computer include but are not limited to: an industrial-grade processor (such as the Intel Core i7-1185G7), an embedded AI acceleration module (such as the NVIDIA Jetson AGX Xavier), and a real-time operating system (such as Linux RT-Preempt). The condition monitoring system enables real-time waveform display (time domain / frequency domain), visualization of fault types and confidence levels, and multi-level early warning for different abnormal damages on the rolling elements of the TBM main bearing.

[0013] In an exemplary embodiment, Figure 2 As shown, a method for identifying the rolling element fault state of a small sample TBM main bearing is provided, comprising the following steps 201 to 202. In which: Step 201: Obtain a vibration signal to be identified of a rolling element of a TBM main bearing, divide the vibration signal to be identified into time windows, and obtain multiple time window sequences. The time windows are multivariate long time series time windows.

[0014] In step 202, any time window sequence is input into a trained fault recognition neural network to obtain fault state recognition results for the TBM main bearing rolling element. The fault recognition neural network is trained using a small sample set; the small sample set is obtained by dividing the historical vibration signals of the TBM main bearing rolling element under different states by time windows; the different states are normal state, abrasion damage state, or scratch damage state; where each time window is a sample.

[0015] The fault recognition neural network includes a data embedding module, a stacking module and a fault identification module connected in sequence; the stacking module includes a cascaded feature extraction module group; the feature extraction module group includes a feature extraction module and a local feature enhancement module connected in sequence; the feature extraction module is used to introduce a variance modulation mechanism in the sparse attention mechanism to determine the weight of the attention score according to the variance of the attention score; the local feature enhancement module is used to use a multi-scale convolution kernel for feature enhancement.

[0016] During the training of the fault identification neural network, the historical vibration signals of the rolling elements of the TBM main bearing are analyzed. , using overlapping sliding window segmentation strategy, the continuous historical vibration signal is converted into a set of time windows , where each time window As a sample, we get a small sample set. is the number of time steps, is the number of sensor channels, is the window length, is a matrix. The historical vibration signal is a high-frequency one. Overlapping time window sampling helps preserve the transient characteristics of impact faults and capture the fault evolution process. In this application, the window length is set to 1024 points, and the sliding step size is set to 512 points (50% overlap). Overlapping sampling preserves the continuity of the fault impact characteristics.

[0017] This application constructs a small sample set of 100 time windows from the rolling elements of a TBM main bearing under nine different conditions, including normal conditions and eight simulated damage states. The simulated damage is divided into two categories: one is abrasion damage with diameters of 2mm, 4mm, and 6mm, specifically arc-shaped abrasion damage; the other is scratch damage with a depth of 3mm (widths of 3mm, 5mm, and 7mm) and a width of 3mm (depths of 3mm, 5mm, and 7mm).

[0018] Implementing steps 201 to 202 above can dynamically enhance the signal-to-noise ratio of weak signal components, improve the ability to capture weak signals, and achieve differentiated extraction of features of different sizes, solving the problem of insufficient adaptability of single-scale feature extraction to diverse damages. In addition, it also greatly improves the recognition accuracy of fault conditions and the complexity of calculations.

[0019] In another exemplary embodiment of the present application, Figure 3 As shown, the above step 202 is replaced by the following steps 301 to 304: Step 301: Input the time window sequence into the data embedding module to obtain an embedding vector.

[0020] Step 302: Input the embedding vector into a stacking module to obtain a final feature vector.

[0021] Step 303: Input the final feature vector into the fault identification module to obtain a fault probability vector.

[0022] Step 304 : Obtain a fault state identification result of the TBM main bearing rolling element according to the fault probability vector.

[0023] In another exemplary embodiment of the present application, step 301 is replaced by the following steps 401 to 403: Step 401, using formula The time window sequence is numerically encoded to obtain a numerical encoding result; wherein, is the numerical encoding result, is a one-dimensional convolution, is a time window sequence.

[0024] Specifically, in order to alleviate the problem of insufficient local feature extraction of traditional Transformer in industrial signal processing, each time window sequence One-dimensional convolution with a wider convolution kernel is used for encoding, and the size of the convolution kernel is set to 3 and the output dimension is set to 128.

[0025] Step 402: Considering the characteristics of the coexistence of periodicity and impact of the vibration signal of the TBM main bearing rolling element, in order to adapt to the characteristics of different frequency ranges, a sinusoidal position code is added to obtain .

[0026] Specifically, the following formula is used to encode each position in the time window sequence to obtain the position encoding result: :

[0027]

[0028] in, is the position code value of the lth position in the time window sequence in dimension 2g, is the position code value of the lth position in the time window sequence in dimension 2g+1, is the position index in the time window sequence, is the dimension index, is the output dimension, is time, L=1024,

[0029] Step 403, using formula The feature fusion of the numerical encoding result and the position encoding result is obtained to obtain the embedding vector: It is layer normalization, which aims to solve the problem of inconsistent dimensions of multiple sensors. To increase the adaptability of small samples, is the embedding vector. This realizes the joint representation of the signal's physical characteristics and spatial position.

[0030] In another exemplary embodiment of the present application, in response to the multi-level feature requirements of TBM main bearing rolling elements, a stacking module is designed to include a cascaded feature extraction module group, namely, a stacking architecture of a feature extraction module and a local feature enhancement module, to perform feature abstraction progressively.

[0031] In the present application, the stacking module includes a first feature extraction module group, a second feature extraction module group, a third feature extraction module group and a fourth feature extraction module group connected in sequence.

[0032] The embedding vector is input into the stacking module to obtain the final feature vector, which specifically includes: The embedding vector is input into a first feature extraction module group to obtain a first feature vector.

[0033] The first feature vector is added to the second feature extraction module group to obtain a second feature vector.

[0034] The second feature vector is added to the third feature extraction module group to obtain a third feature vector.

[0035] The third feature vector to the fourth feature extraction module group are combined to obtain a final feature vector.

[0036] Use the following formula to calculate the The first feature extraction module group Eigenvectors: }.

[0037] in, For the eigenvector, For the -1 eigenvector, is the feature extraction module, It is a local feature enhancement module. 1 o'clock, is the embedding vector, The final feature vector is constructed by stacking four layers to achieve progressive feature abstraction. The bottom layer captures local impact features, and the top layer models the system-level fault evolution pattern.

[0038] In another exemplary embodiment of the present application, the embedding vector is input into a first feature extraction module group to obtain a first feature vector, specifically including: The embedding vector is input into the feature extraction module of the first feature extraction module group to obtain an output vector of the feature extraction module.

[0039] The output vector of the feature extraction module is input into the local feature enhancement module of the first feature extraction module group to obtain a first feature vector.

[0040] In another exemplary embodiment of the present application, the feature extraction module includes a sparse attention layer, a normalization layer, a feedforward layer, and a normalization layer connected in sequence; the output of the data embedding module is residually connected to the output of the first normalization layer in the feature extraction module; the input of the feedforward layer is residually connected to the output of the second normalization layer in the feature extraction module. The normalization layer is processed using a standard layer normalization method.

[0041] Inputting the embedding vector into the feature extraction module of the first feature extraction module group to obtain an output vector of the feature extraction module specifically includes: The embedding vector is input into the sparse attention layer to obtain a sparse attention vector.

[0042] The sparse attention vector is input into the first normalization layer in the feature extraction module to obtain a first standard vector.

[0043] Perform residual addition on the first standard vector and the embedded vector to obtain a residual vector.

[0044] The residual vector is input into the feed-forward layer to obtain an extracted vector.

[0045] The extracted vector is input into the second normalization layer in the feature extraction module to obtain a second normalized vector.

[0046] The second standard vector is added to the residual vector residual to obtain an output vector of the feature extraction module.

[0047] In another exemplary embodiment of the present application, the embedding vector is input into a sparse attention layer to obtain a sparse attention vector, specifically including: The query vector, key vector, and value vector of the embedding vector are calculated using the following formulas: .

[0048] in, is the embedding vector, is the query vector, is the key vector, is a value vector, 、 and is the trainable weight matrix.

[0049] Based on the query vector, the set quantity is determined using the following formula: .

[0050] in, To set the quantity, is the sparse factor, is the length of the query vector.

[0051] According to the query vector Q With key vector K The maximum absolute value of the dot product, select u The index set of the most significant queries. That is, based on the query vector, the key vector and the set number, the index set of the top u query vectors is determined using the following formula: .

[0052] in, is the maximum absolute value of the dot product of the query vector and the key vector, is the position number of the key vector ((j=1,2,…, ), traverse all keys and calculate the maximum value), (⋅) is the first u indexes of the maximum value, u is the set number, is the index set of the first u query vectors, The indices in correspond to the local feature locations in the vibration signal that are most correlated with the global bond.

[0053] Based on the index set of the first u query vectors, the following formula is used to determine the first The attention score of the query vector ( ): .

[0054] in, is the index of the query vector in the index collection, The index in the collection query vectors, d is the dimension of the key vector, For the The attention score of the query vector.

[0055] Based on the The attention score of the query vector and the key vector is determined by the following formula The variance of the attention scores of the query vectors: .

[0056] in, For the query vector With the The attention score, For the query vector Attention scores for all keys.

[0057] Based on the The variance of the attention score of the query vector is determined by the following formula The weight of the variance of the attention scores of the query vectors:

[0058] in, is the modulation coefficient, For the The weight of the attention score of the query vector.

[0059] Based on the The weight of the variance of the attention score of the query vector is calculated using the formula , determine the The modulation weight of the query vector .

[0060] Based on the The modulation weights of the query vector and the value vector are calculated using the formula , determine the The sparse attention vector corresponding to the query vector .

[0061] The sparse attention layer of this application uses an improved sparse attention mechanism to obtain a sparse attention vector. The improved sparse attention mechanism introduces a variance modulation mechanism into the sparse attention mechanism to determine the weight of the attention score based on the variance of the attention score. The improved sparse attention mechanism is a single-head sparse attention mechanism.

[0062] In another exemplary embodiment of the present application, in view of the multi-scale characteristics of the features in the vibration signal of the TBM main bearing rolling element (such as the coexistence of local pitting and distributed wear), an adaptively weighted multi-scale convolution layer is introduced in the local feature enhancement module. The local feature enhancement module includes a convolution layer, a normalization layer, an activation layer and a maximum pooling layer connected in sequence; the convolution layer includes a plurality of convolution kernels of different scales, and the sum of the output vectors of the convolution layer is input into a designed distillation path to eliminate the feature disturbance caused by operating condition fluctuations. The distillation path includes a normalization layer, an activation layer and a maximum pooling layer. While retaining more than 95% of the feature energy, the present application compresses the feature dimension by 50%, significantly improving the computational efficiency.

[0063] Based on the output vector of the feature extraction module, the formula , and get the sum of the output vectors of the convolutional layer.

[0064] Among them, a is the size sequence of the convolution kernel, is the size of the convolution kernel of size a (k=3,7,11), To learn adaptive weights, it aims to achieve dynamic feature fusion. is the sum of the output vectors of the convolutional layer, is the output vector of the feature extraction module.

[0065] Based on the sum of the output vectors of the convolutional layer, the formula , determine the first eigenvector.

[0066] Among them, BatchNorm is the batch normalization in the normalization layer, which aims to solve the problem of signal amplitude differences in different tunneling stages. It is the maximum mean pooling in the maximum pooling layer, which aims to retain key features while reducing the amount of calculation. is the first eigenvector, It is the activation function in the activation layer, which aims to retain negative features and solve the problem of traditional ReLU losing negative features.

[0067] In another exemplary embodiment of the present application, after the embedding vector is input into the stacking module to obtain the final feature vector, the method further includes: use The final feature vector is flattened to obtain a flattened feature vector. is the flattened eigenvector, is the final eigenvector, Flatten the feature.

[0068] In another exemplary embodiment of the present application, the fault identification module includes a linear layer, a ReLU activation layer, a linear layer, and a Softmax activation layer connected in sequence. The following formula is used to determine the fault probability vector: .

[0069] in, is the failure probability vector, is the fault status category number, is the probability of belonging to the normal state, Belongs to the fault status category probability.

[0070] According to the fault probability vector, the fault state identification result of the TBM main bearing rolling element is obtained. , specifically:

[0071] .

[0072] This example demonstrates that, with a small sample size of only 100 samples per fault state time window, this method achieves a 90.5% accuracy rate for identifying various degrees of abrasions and scratches on TBM main bearings, confirming its effectiveness in practical engineering applications. In particular, the method achieves 97.1% and 94.8% accuracy for identifying micro-damage (2mm abrasions) and scratches (3mm deep and 3mm wide), respectively, providing reliable technical support for preventive maintenance of TBM main bearings.

[0073] The advantages of this application are: First, under the condition of a small sample, the fault state of the TBM main bearing rolling element is identified with high precision. Compared with the existing technology, this application has achieved an identification accuracy of over 90% for arc-shaped abrasion damage of 2mm, 4mm, and 6mm and scratch damage of 3mm in depth (3mm, 5mm, and 7mm in width) and 3mm in width (3mm, 5mm, and 7mm in depth) under the condition of a small sample of only 100 time window samples for each type of fault state (window length 1024 points, sampling rate 2kHz). Through comparative experiments (compared with traditional DCNN, LSTM, and Transformer methods under the same data set), the recognition accuracy of this application has been improved by 10~30%, verifying the sparse attention mechanism (sparse factor =5) and multi-scale feature distillation fusion (k=3,7,11) for small sample feature extraction.

[0074] Second, the proposed method achieves multi-scale damage fault state identification, particularly for the early-stage micro-damage of main bearing rolling elements. Tests in a simulated noise environment (SNR = 5dB) show that the proposed method improves detection sensitivity to 97.1% for 2mm abrasion signals and 94.8% for 3mm×3mm scratch signals, respectively. This represents a 30% improvement over conventional methods, effectively resolving the issue of existing technologies' insufficient response to early-stage micro-damage (<3mm) in TBM main bearings.

[0075] Third, the real-time response performance of the lightweight model is achieved. This application achieves significant lightweight and real-time improvements while ensuring high precision through the innovative design of sparse attention layers, feature distillation technology, and alternating stacking architecture. Specifically, the dynamic query selection mechanism effectively reduces 70% of redundant calculations, and cooperates with the feature distillation path of batch normalization → ELU → maximum pooling to compress the model parameters to 3.93M (only 18.03% of ResNet34) while retaining 95% of the feature energy. Actual deployment tests show that this application achieves a single inference speed of 12.8ms, which fully meets the real-time requirements of TBM main bearing online monitoring. This lightweight design provides key technical support for the embedded deployment of equipment while maintaining 90.5% recognition accuracy.

[0076] Based on the same inventive concept, an embodiment of the present application also provides a small sample TBM main bearing rolling element fault state identification system for implementing the above-mentioned small sample TBM main bearing rolling element fault state identification method. The implementation solution provided by this system is similar to the implementation solution recorded in the above-mentioned method. Therefore, the specific limitations in one or more small sample TBM main bearing rolling element fault state identification system embodiments provided below can be found in the above limitations on the small sample TBM main bearing rolling element fault state identification method, and will not be repeated here.

[0077] In an exemplary embodiment, Figure 4 As shown, a small sample TBM main bearing rolling element fault state identification system is provided, including: The time window sequence acquisition module 501 is used to acquire the vibration signal to be identified of the rolling element of the TBM main bearing, and divide the vibration signal to be identified into time windows to obtain multiple time window sequences.

[0078] The fault state identification module 502 is used to input any time window sequence into a trained fault identification neural network to obtain a fault state identification result of the TBM main bearing rolling element.

[0079] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0080] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0081] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for identifying rolling element fault status of a TBM main bearing with a small sample size, characterized in that: The method comprises: Acquire a vibration signal to be identified of a rolling element of a TBM main bearing, and divide the vibration signal to be identified using a time window to obtain multiple time window sequences; Input any time window sequence into a trained fault recognition neural network to obtain a fault state recognition result of the TBM main bearing rolling element; the fault recognition neural network is trained using a small sample set; the small sample set is obtained by dividing the historical vibration signals of the TBM main bearing rolling element under different states by time windows; the different states are normal state, abrasion damage state, or scratch damage state; wherein one time window is one sample; The fault recognition neural network includes a data embedding module, a stacking module and a fault identification module connected in sequence; the stacking module includes a cascaded feature extraction module group; the feature extraction module group includes a feature extraction module and a local feature enhancement module connected in sequence; the feature extraction module is used to introduce a variance modulation mechanism in the sparse attention mechanism to determine the weight of the attention score according to the variance of the attention score; the local feature enhancement module is used to use a multi-scale convolution kernel for feature enhancement.

2. The method for identifying rolling element fault status of a small sample TBM main bearing according to claim 1 is characterized in that: The time window sequence is input into the trained fault recognition neural network to obtain the fault state recognition result of the TBM main bearing rolling element, which specifically includes: Inputting the time window sequence into the data embedding module to obtain an embedding vector; Input the embedding vector into the stacking module to obtain the final feature vector; Inputting the final feature vector into the fault identification module to obtain a fault probability vector; According to the fault probability vector, a fault state identification result of the TBM main bearing rolling element is obtained.

3. The method for identifying rolling element fault status of a small sample TBM main bearing according to claim 2 is characterized in that: Inputting the time window sequence into the data embedding module to obtain an embedding vector specifically includes: Using the formula The time window sequence is numerically encoded to obtain a numerical encoding result; wherein, is the numerical encoding result, is a one-dimensional convolution, is a time window sequence; The following formula is used to encode each position in the time window sequence to obtain the position encoding result: : in, is the position code value of the lth position in the time window sequence in dimension 2g, is the position code value of the lth position in the time window sequence in dimension 2g+1, is the position index in the time window sequence, is the dimension index, is the output dimension, For time; Using the formula The feature fusion of the numerical encoding result and the position encoding result is obtained to obtain the embedding vector: is layer normalization, is the pruning strategy, is the embedding vector.

4. The method for identifying rolling element fault status of a small sample TBM main bearing according to claim 2 is characterized in that: The stacking module includes a first feature extraction module group, a second feature extraction module group, a third feature extraction module group and a fourth feature extraction module group connected in sequence; The embedding vector is input into the stacking module to obtain the final feature vector, which specifically includes: Inputting the embedding vector into a first feature extraction module group to obtain a first feature vector; Applying the first feature vector to a second feature extraction module group to obtain a second feature vector; Applying the second feature vector to a third feature extraction module group to obtain a third feature vector; The third feature vector to the fourth feature extraction module group are combined to obtain a final feature vector.

5. The method for identifying rolling element fault status of a small sample TBM main bearing according to claim 4 is characterized in that: Inputting the embedding vector into the first feature extraction module group to obtain a first feature vector specifically includes: Inputting the embedding vector into a feature extraction module of a first feature extraction module group to obtain an output vector of the feature extraction module; The output vector of the feature extraction module is input into the local feature enhancement module of the first feature extraction module group to obtain a first feature vector.

6. The method for identifying rolling element fault status of a small sample TBM main bearing according to claim 5 is characterized in that: The feature extraction module includes a sparse attention layer, a normalization layer, a feedforward layer and a normalization layer connected in sequence; the output of the data embedding module is residually connected to the output of the first normalization layer in the feature extraction module; the input of the feedforward layer is residually connected to the output of the second normalization layer in the feature extraction module; Inputting the embedding vector into the feature extraction module of the first feature extraction module group to obtain an output vector of the feature extraction module specifically includes: Inputting the embedding vector into the sparse attention layer to obtain a sparse attention vector; Inputting the sparse attention vector into a first normalization layer in the feature extraction module to obtain a first normalized vector; Perform residual addition on the first standard vector and the embedded vector to obtain a residual vector; Inputting the residual vector into the feedforward layer to obtain an extracted vector; Inputting the extracted vector into a second normalization layer in the feature extraction module to obtain a second normalized vector; The second standard vector is added to the residual vector residual to obtain an output vector of the feature extraction module.

7. The method for identifying rolling element fault status of a small sample TBM main bearing according to claim 6 is characterized in that: The embedding vector is input into the sparse attention layer to obtain a sparse attention vector, which specifically includes: The query vector, key vector, and value vector of the embedding vector are calculated using the following formulas: ; in, is the embedding vector, is the query vector, is the key vector, is a value vector, 、 and is the trainable weight matrix; Based on the query vector, the set quantity is determined using the following formula: ； in, To set the quantity, is the sparse factor, is the length of the query vector; Based on the query vector, the key vector, and the set number, the index set of the first u query vectors is determined using the following formula: ； in, is the maximum absolute value of the dot product of the query vector and the key vector, is the position number of the key vector, (⋅) is the first u indexes of the maximum value, u is the set number, is the index set of the first u query vectors; Based on the index set of the first u query vectors, the following formula is used to determine the first Attention scores for query vectors: ； in, is the index of the query vector in the index collection, The index in the collection query vectors, d is the dimension of the key vector, For the The attention score of the query vector; Based on the The attention score of the query vector and the key vector is determined by the following formula The variance of the attention scores of the query vectors: ； in, For the query vector With the Key vector The attention score, For the query vector Attention scores for all keys; Based on the The variance of the attention score of the query vector is determined by the following formula The weight of the variance of the attention scores of the query vectors: in, is the modulation coefficient, For the The weight of the attention score of the query vector; Based on the The weight of the variance of the attention score of the query vector is calculated using the formula , determine the The modulation weight of the query vector : Based on the The modulation weights of the query vector and the value vector are calculated using the formula , determine the The sparse attention vector corresponding to the query vector .

8. The method for identifying rolling element fault status of a small sample TBM main bearing according to claim 5 is characterized in that: The local feature enhancement module includes a convolution layer, a normalization layer, an activation layer and a maximum pooling layer connected in sequence; the convolution layer includes a plurality of convolution kernels of different scales; Inputting the output vector of the feature extraction module into the local feature enhancement module of the first feature extraction module group to obtain a first feature vector specifically includes: Based on the output vector of the feature extraction module, the formula , and the sum of the output vectors of the convolution layer is obtained; where a is the size sequence of the convolution kernel, is the size of the convolution kernel of size a, is a learnable adaptive weight, is the sum of the output vectors of the convolutional layer, is the output vector of the feature extraction module; Based on the sum of the output vectors of the convolutional layer, the formula , determine the first eigenvector; where BatchNorm is the batch normalization in the normalization layer, is the maximum mean pooling in the maximum pooling layer, is the first eigenvector, is the activation function in the activation layer.

9. The method for identifying rolling element fault status of a small sample TBM main bearing according to claim 1, characterized in that: The fault identification module includes a linear layer, a ReLU activation layer, a linear layer and a Softmax activation layer connected in sequence.

10. A small sample TBM main bearing rolling element fault state identification system, using the small sample TBM main bearing rolling element fault state identification method according to any one of claims 1 to 9, characterized in that: The system comprises: A time window sequence acquisition module is used to obtain the vibration signal to be identified of the rolling element of the TBM main bearing, and divide the vibration signal to be identified into multiple time window sequences by using time windows; The fault state identification module is used to input any time window sequence into a trained fault identification neural network to obtain a fault state identification result of the TBM main bearing rolling element.

Images