A small sample TBM main bearing rolling body fault state recognition method and system
By using a cascaded feature extraction module group and a lightweight fault identification neural network, the problems of accuracy and real-time performance in TBM main bearing rolling element fault identification were solved, achieving efficient and accurate fault status identification under extreme working conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA RAILWEY ENG SERVICE CO LTD
- Filing Date
- 2025-07-10
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies struggle to effectively identify rolling element faults in TBM main bearings under extreme operating conditions, especially with small sample sizes, long time series, and weak abnormal signals. Traditional methods suffer from low accuracy and high computational complexity, failing to meet the requirements for real-time fault identification.
A cascaded feature extraction module group, including a sparse attention mechanism and multi-scale convolution kernels, is adopted. By dividing the fault recognition neural network into time windows and training it, the signal-to-noise ratio of weak signal components is dynamically enhanced. Multiple feature extractions are performed through stacked modules, and the computational complexity is reduced by combining lightweight design.
High-precision fault condition identification was achieved under small sample conditions, improving the accuracy of micro-damage identification, reducing the computational complexity and response time of hardware devices, and meeting the real-time monitoring requirements of TBM main bearings.
Smart Images
Figure CN120670960B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of fault diagnosis, and in particular to a method and system for identifying the fault status of rolling elements in small-sample TBM main bearings. Background Technology
[0002] Full-face tunnel boring machines (TBMs) are core equipment in modern underground engineering. The TBM main bearing plays a crucial role in decoupling the cutterhead rotation drive from the front and rear structural mechanics, and its operational status directly affects project safety and progress. However, current mainstream methods still face significant challenges in main bearing fault identification due to difficulties in acquiring monitoring data under extreme conditions, scarcity of actual damage samples, and weak early fault signal energy. 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 identification performance under strong noise disturbances and signal non-stationarity. On the other hand, while emerging deep learning methods (such as deep convolutional neural networks, long short-term memory neural networks, or Transformer architectures) have shown good performance on standard bearing datasets, they still struggle to effectively generalize or extract key features when faced with small samples, long time series, and weak anomalous signals under harsh TBM service conditions, leading to low fault identification accuracy. While recent research has made some progress in fault diagnosis under small sample conditions, including methods such as meta-learning, metric learning, and transfer learning, these methods are mostly aimed at small-sized, ultra-high-speed main bearings in standard industrial scenarios. These small-sample learning methods suffer from insufficient adaptability to fault identification of ultra-large-sized, ultra-low-speed TBM main bearings. Firstly, the extremely harsh operating environment of TBM main bearings results in extremely weak effective fault features in the vibration signal, exhibiting a "strong noise drowning out weak features" characteristic, with extremely low fault signal energy, severely weakening the generalization basis of the model in meta-learning. Secondly, the temporal sparsity caused by the ultra-low speed in the small-sample scenario of TBM main bearings makes it difficult for transfer learning to effectively capture key dynamic patterns. Thirdly, the computational complexity is high: current small-sample fault identification methods often involve complex model structures and a large number of model parameters, especially methods based on standard attention mechanisms, which incur extremely high computational costs and long model response times when processing long-term sequential state monitoring data of TBM main bearings, failing to meet the high-response speed requirements of real-time fault identification for TBM main bearings. Summary of the Invention
[0003] The purpose of this application is to provide a method and system for identifying the fault status of rolling elements in TBM main bearings with a small sample size. This method and system can efficiently and accurately identify the fault status of rolling elements in TBM main bearings using a cascaded feature extraction module group with a small number of samples. When applied to hardware devices such as computers, it can reduce the computational complexity of the hardware devices, reduce the response time of the hardware devices, and improve the computational efficiency of the hardware devices.
[0004] To achieve the above objectives, this application provides the following solution:
[0005] Firstly, this application provides a method for identifying the rolling element fault state of a small-sample TBM main bearing, including:
[0006] The vibration signal to be identified of the rolling elements of the TBM main bearing is obtained, and the vibration signal to be identified is divided into multiple time window sequences using time windows.
[0007] Any time window sequence is input into the trained fault identification neural network to obtain the fault state identification result of the TBM main bearing rolling element; the fault identification neural network is trained using a small sample set; the small sample set is obtained by dividing the historical vibration signal of the TBM main bearing rolling element into different states by time windows; the different states are normal state, scratch damage state or scratch damage state; wherein, one time window is one sample;
[0008] The fault identification 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 based on the variance of the attention score; the local feature enhancement module is used to perform feature enhancement using multi-scale convolution kernels.
[0009] Secondly, this application provides a small-sample TBM main bearing rolling element fault condition identification system, including:
[0010] The time window sequence acquisition module is used to acquire the vibration signal to be identified of the rolling elements of the TBM main bearing, and divide the vibration signal to be identified into multiple time window sequences by using time windows.
[0011] The fault status identification module is used to input the time window sequence into the trained fault identification neural network for any given time window sequence to obtain the fault status identification result of the rolling elements of the TBM main bearing.
[0012] According to the specific embodiments provided in this application, this application has the following technical effects:
[0013] This application provides a method and system for identifying the fault state of rolling elements in TBM main bearings using a small sample. It employs a time window to divide a small set of historical vibration signals from rolling elements of TBM main bearings under different conditions, using these signals as a small sample set to train a fault identification neural network. This enables the network to accurately identify the fault state of the rolling elements in TBM main bearings even with weak signals in a small sample environment. Furthermore, the feature extraction module of the fault identification neural network introduces a variance modulation mechanism within a sparse attention mechanism, dynamically enhancing the signal-to-noise ratio of weak signal components and improving the ability to capture weak signals. The local feature enhancement module of the fault identification neural network constructs multi-scale convolutional kernels to achieve differentiated extraction of features of different sizes, addressing the insufficient adaptability of single-scale feature extraction to diverse damage. The stacking module of the fault identification neural network uses cascaded feature extraction modules to repeatedly extract features, significantly improving the accuracy of fault state identification and reducing computational complexity. Attached Figure Description
[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0015] 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 this application;
[0016] Figure 2 A flowchart illustrating a method for identifying the rolling element fault status of a small-sample TBM main bearing, provided in an embodiment of this application;
[0017] Figure 3 This application provides a structural diagram of the fault identification neural network in a small-sample TBM main bearing rolling element fault state identification method according to an embodiment of the present application.
[0018] Figure 4 This is a schematic diagram of the functional modules of a small-sample TBM main bearing rolling element fault status identification system provided in an embodiment of this application. Detailed Implementation
[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0020] To make the objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0021] The small-sample TBM main bearing rolling element fault state identification method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, the main bearing rolling element vibration signal acquisition system communicates with the intelligent industrial control computer via an industrial bus. The data storage system can store the vibration signals that the intelligent industrial control computer needs to process. The data storage system can be integrated into the intelligent industrial control computer or deployed independently on a local storage device or a cloud server. The main bearing rolling element vibration signal acquisition system acquires the vibration signals of the TBM main bearing rolling elements in real time and transmits the vibration signals to the intelligent industrial control computer. After receiving the vibration signals, the intelligent industrial control computer divides the vibration signals into fixed-length time windows to obtain the current time window data; it then inputs the time window sequence into a 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 a status monitoring system. If the recognition result indicates a fault, a corresponding early warning mechanism is triggered.
[0022] The vibration signal acquisition system includes, but is not limited to: high-precision accelerometers, signal conditioning modules (interference suppression, amplification, filtering), and analog-to-digital converters (24-bit ADCs). The intelligent industrial control computer features, but are not limited to: industrial-grade processors (such as Intel Core i7-1185G7), embedded AI acceleration modules (such as NVIDIA Jetson AGX Xavier), and real-time operating systems (such as Linux RT-Preempt). The condition monitoring system can achieve real-time waveform display (time domain / frequency domain), visualization of fault types and confidence levels, and multi-level early warning of different abnormal damage to the rolling elements of the TBM main bearing.
[0023] In one exemplary embodiment, such as Figure 2 As shown, a method for identifying the rolling element fault state of a small-sample TBM main bearing is provided, including the following steps 201 to 202. Wherein:
[0024] Step 201: Obtain the vibration signal to be identified from the rolling elements of the TBM main bearing, and divide the vibration signal to be identified into multiple time window sequences using time windows. The time window is a multivariable long-sequence time window.
[0025] Step 202: Input any time window sequence into the trained fault identification neural network to obtain the fault state identification result of the TBM main bearing rolling element. The fault identification 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 into different states by time windows; the different states are normal state, scuff damage state, or scratch damage state; wherein, one time window is one sample.
[0026] The fault identification 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 based on the variance of the attention score; the local feature enhancement module is used to perform feature enhancement using multi-scale convolution kernels.
[0027] During the training of the fault identification neural network, historical vibration signals of the rolling elements of the TBM main bearing were analyzed. An overlapping sliding window segmentation strategy is adopted to convert continuous historical vibration signals into time window sets. , where each time window Treating it as a single sample yields a small sample set. For time steps, Number of sensor channels For window length, The matrix is defined as follows. The historical vibration signal is a high-frequency vibration signal. Overlapping sampling within a time window helps preserve the transient characteristics of the impact fault and capture the fault evolution process. In this application, the window length is set to 1024 points, and the sliding step size is 512 points (50% overlap rate). Overlapping sampling preserves the continuity of the fault impact characteristics.
[0028] This application constructs a small sample set by selecting 100 time windows from each of the rolling elements of the TBM main bearing under nine different conditions, including the normal state and eight simulated damage states. The simulated damage is divided into two categories: one is the scratch damage state with a depth of 2mm, 4mm, and 6mm, specifically the arc-shaped scratch damage state; the other is the scratch damage state with a depth of 3mm (width of 3mm, 5mm, and 7mm) and a width of 3mm (depth of 3mm, 5mm, and 7mm).
[0029] 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 damage. In addition, it can greatly improve the accuracy of fault state identification and the complexity of calculation.
[0030] In another exemplary embodiment of this application, such as Figure 3 As shown, step 202 above is replaced by steps 301 to 304:
[0031] Step 301: Input the time window sequence into the data embedding module to obtain the embedding vector.
[0032] Step 302: Input the embedding vector into the stacking module to obtain the final feature vector.
[0033] Step 303: Input the final feature vector into the fault identification module to obtain the fault probability vector.
[0034] Step 304: Based on the fault probability vector, obtain the fault status identification result of the TBM main bearing rolling element.
[0035] In another exemplary embodiment of this application, step 301 is replaced by the following steps 401 to 403:
[0036] Step 401, using the formula The time window sequence is numerically encoded to obtain the numerical encoding result; wherein, The result is a numerical encoding. For one-dimensional convolution, This is a time window sequence.
[0037] Specifically, to alleviate the problem of insufficient local feature extraction in traditional Transformer in industrial signal processing, for each time window sequence... Encoding is performed using a one-dimensional convolution with a relatively wide kernel, and the kernel size is set to 3, with an output dimension of 128.
[0038] Step 402: Considering the coexistence of periodicity and impact in the vibration signal of the TBM main bearing rolling elements, a sinusoidal position code is added to adapt to the characteristics of different frequency ranges. .
[0039] Specifically, the following formula is used to encode each position in the time window sequence to obtain the position encoding result. :
[0040]
[0041]
[0042] in, The position code value in dimension 2g for the l-th position in the time window sequence. The position code value in dimension 2g+1 for the l-th position in the time window sequence. This is the position index within the time window sequence. For dimensional indexing, For output dimensions, For time, L=1024,
[0043] Step 403, using the formula Feature fusion is performed on the numerical encoding results and the positional encoding results to obtain the embedding vector: where, Layer normalization aims to address the issue of inconsistent dimensions across multiple sensors. To improve the adaptability to small samples, a pruning strategy is employed. This is an embedding vector. It achieves a joint representation of the signal's physical properties and spatial location.
[0044] In another exemplary embodiment of this application, in order to meet the multi-level feature requirements of the rolling elements of the TBM main bearing, a stacked module is designed including a cascaded feature extraction module group, namely: a stacked architecture of feature extraction module and local feature enhancement module, which progressively performs feature abstraction.
[0045] In this application, the stacked 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.
[0046] The embedding vector is input into the stacking module to obtain the final feature vector, specifically including:
[0047] The embedding vector is input into the first feature extraction module group to obtain the first feature vector.
[0048] The first feature vector is fed into the second feature extraction module group to obtain the second feature vector.
[0049] The second feature vector is fed into the third feature extraction module group to obtain the third feature vector.
[0050] The third feature vector is fed into the fourth feature extraction module group to obtain the final feature vector.
[0051] The following formula is used to calculate the first... The first feature extraction module group Feature vector:
[0052] }
[0053] in, For the first Feature vector For the first -1 eigenvector, For feature extraction module, This is a local feature enhancement module. At 1 o'clock, For embedding vectors, This results in the final feature vector. Progressive abstraction of features is achieved through a four-layer stacking process: the bottom layer captures local impact features, while the top layer models system-level fault evolution patterns.
[0054] In another exemplary embodiment of this application, the embedding vector is input to the first feature extraction module group to obtain the first feature vector, specifically including:
[0055] The embedding vector is input to the feature extraction module of the first feature extraction module group to obtain the output vector of the feature extraction module.
[0056] The output vector of the feature extraction module is input to the local feature enhancement module of the first feature extraction module group to obtain the first feature vector.
[0057] In another exemplary embodiment of this 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 concatenated with the output of the first normalization layer in the feature extraction module; the input of the feedforward layer is residually concatenated with the output of the second normalization layer in the feature extraction module. The normalization layer uses a layer normalization method for standard processing.
[0058] The embedded vector is input to the feature extraction module of the first feature extraction module group to obtain the output vector of the feature extraction module, specifically including:
[0059] The embedding vector is input into the sparse attention layer to obtain the sparse attention vector.
[0060] The sparse attention vector is input into the first normalization layer in the feature extraction module to obtain the first standard vector.
[0061] The residual vector is obtained by adding the first standard vector and the embedded vector together.
[0062] The residual vector is input into the feedforward layer to obtain the extracted vector.
[0063] The extracted vector is input into the second normalization layer in the feature extraction module to obtain the second standard vector.
[0064] The second standard vector is added to the residual vector to obtain the output vector of the feature extraction module.
[0065] In another exemplary embodiment of this application, the embedding vector is input to a sparse attention layer to obtain a sparse attention vector, specifically including:
[0066] The query vector, key vector, and value vector of the embedded vector are calculated using the following formulas:
[0067] .
[0068] in, For embedding vectors, For query vector, For key vectors, For value vectors, , and This is a trainable weight matrix.
[0069] Based on the query vector, the set quantity is determined using the following formula:
[0070] .
[0071] in, To set the quantity, As a sparsity factor, The length of the query vector.
[0072] Based on the query vector Q With key vector K The maximum absolute value of the dot product, selected from the previous ones. 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:
[0073] .
[0074] in, To find the maximum absolute value of the dot product of the vector and the key vector, Number the positions of the key vectors (j=1,2,…, (Iterate through all keys and calculate the maximum value). (⋅) represents the first u indices of the maximum value, where u is the set number. The set of indices for the first u query vectors. The index in the index corresponds to the location of the local feature in the vibration signal that is most relevant to the global bond.
[0075] Based on the index set of the first u query vectors, the following formula is used to determine the index of the th element in the index set. Attention score for each query vector ( ):
[0076] .
[0077] in, The index of the query vector in the index set. For the first in the index set There are query vectors, where d is the dimension of the key vector. For the first Attention score for each query vector.
[0078] Based on the first The attention score of the query vector and the key vector are used to determine the first query vector using the following formula. Variance of attention scores for each query vector:
[0079] .
[0080] in, For the first query vectors With the Attention score For the first query vectors Attention score for all keys.
[0081] Based on the first The variance of the attention score for the i-th query vector is determined using the following formula. Weights of the variance of the attention score for each query vector:
[0082]
[0083] in, The modulation coefficient, For the first The weights of the attention score for each query vector.
[0084] Based on the first The weights of the variance of the attention score for each query vector are calculated using the formula... Determine the first Modulation weights of query vectors .
[0085] Based on the first The modulation weights of each query vector and the value vector are calculated using the formula... Determine the first The sparse attention vector corresponding to each query vector .
[0086] The sparse attention layer of this application employs an improved sparse attention mechanism to obtain a sparse attention vector. This improved sparse attention mechanism introduces a variance modulation mechanism to determine the weights of attention scores based on the variance of the attention scores. This improved sparse attention mechanism is a single-head sparse attention mechanism.
[0087] In another exemplary embodiment of this application, to address the multi-scale characteristics of vibration signals from the rolling elements of a TBM main bearing (such as the coexistence of local pitting and distributed wear), an adaptive weighted multi-scale convolutional layer is introduced in the local feature enhancement module. This module includes a convolutional layer, a normalization layer, an activation layer, and a max-pooling layer connected in sequence. The convolutional layer comprises multiple convolutional kernels of different scales. The sum of the output vectors of the convolutional layer is input into a pre-designed distillation path to eliminate feature disturbances caused by operating condition fluctuations. The distillation path includes a normalization layer, an activation layer, and a max-pooling layer. This application retains over 95% of the feature energy while compressing the feature dimension by 50%, significantly improving computational efficiency.
[0088] Based on the output vector of the feature extraction module, the formula is used. This yields the sum of the output vectors of the convolutional layer.
[0089] Where 'a' is the size index of the convolution kernel. Let k be the size of the convolution kernel of size a (k=3,7,11). The learnable adaptive weights aim to achieve dynamic feature fusion. The sum of the output vectors of the convolutional layer. This is the output vector of the feature extraction module.
[0090] Based on the sum of the output vectors of the convolutional layers, the formula is used. Determine the first eigenvector.
[0091] BatchNorm is the batch normalization function in the normalization layer, designed to address the signal amplitude differences at different tunneling stages. This is maximum mean pooling within the max pooling layer, designed to retain key features while reducing computational cost. The first eigenvector, is the activation function in the activation layer, designed to preserve negative features and solve the problem of traditional ReLU losing negative features.
[0092] In another exemplary embodiment of this application, after inputting the embedding vector into the stacking module to obtain the final feature vector, the method further includes:
[0093] use The final feature vector is flattened to obtain the flattened feature vector. The flattened feature vectors, For the final feature vector, To flatten out the features.
[0094] In another exemplary embodiment of this application, the fault identification module includes a linear layer, a ReLU activation layer, another linear layer, and a Softmax activation layer connected in sequence. The fault probability vector is determined using the following formula:
[0095] .
[0096] in, This is the failure probability vector. This is the fault status category number. This represents the probability of belonging to a normal state. This belongs to the fault state category. The probability of.
[0097] Based on the fault probability vector, the fault state identification result of the TBM main bearing rolling elements is obtained. Specifically, it manifests as follows:
[0098] .
[0099] This embodiment verifies that, under a small sample size of only 100 samples per fault state time window, the method achieves an accuracy of 90.5% in identifying different degrees of scratches and abrasions on TBM main bearings, demonstrating its effectiveness in practical engineering applications. In particular, the method achieves identification accuracies of 97.1% and 94.8% for minute-sized damage (2mm scratches) and scratches (3mm deep and 3mm wide), respectively, providing reliable technical support for preventative maintenance of TBM main bearings.
[0100] The advantages of this application are: First, it achieves high-precision fault state identification of the rolling elements of TBM main bearings under small sample conditions. Compared with existing technologies, this application achieves over 90% accuracy in identifying arc-shaped scratches of 2mm, 4mm, and 6mm, as well as scratches with depths of 3mm (widths of 3mm, 5mm, and 7mm) and widths of 3mm (depths of 3mm, 5mm, and 7mm) under small sample conditions with only 100 time window samples per fault state (window length 1024 points, sampling rate 2kHz). Through comparative experiments (comparing with traditional DCNN, LSTM, and Transformer methods on the same dataset), the recognition accuracy of this application is improved by 10-30%, verifying the sparse attention mechanism (sparse factor). =5) and the effectiveness of multi-scale feature distillation fusion (k=3,7,11) for small sample feature extraction.
[0101] Second, it achieves multi-scale damage and fault state identification, especially the identification of early micro-damage states of the rolling elements of the main bearing. Tests under simulated noise environment (SNR=5dB) show that the proposed method improves the detection sensitivity of 2mm scratch signals and 3mm×3mm scratch signals to 97.1% and 94.8% respectively, which is about 30% higher than the traditional method, effectively solving the problem of insufficient response of existing technology to early micro-damage (<3mm) of TBM main bearings.
[0102] Third, this application achieves real-time response performance for a lightweight model. Through innovative designs using sparse attention layers, feature distillation techniques, and an alternating stacking architecture, it achieves significant improvements in lightweight design and real-time performance while maintaining high accuracy. Specifically, the dynamic query selection mechanism effectively reduces redundant computation by 70%, and combined with the feature distillation path of batch normalization → ELU → max pooling, it compresses the model parameter count 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, fully meeting the real-time requirements of online monitoring of TBM main bearings. This lightweight design, while maintaining 90.5% recognition accuracy, provides key technical support for embedded deployment of equipment.
[0103] Based on the same inventive concept, this application also provides a small sample TBM main bearing rolling element fault state identification system for implementing the small sample TBM main bearing rolling element fault state identification method mentioned above. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more small sample TBM main bearing rolling element fault state identification system embodiments provided below can be found in the limitations of the small sample TBM main bearing rolling element fault state identification method mentioned above, and will not be repeated here.
[0104] In one exemplary embodiment, such as Figure 4 As shown, a small-sample TBM main bearing rolling element fault condition identification system is provided, including:
[0105] 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 multiple time window sequences by using time windows.
[0106] The fault status identification module 502 is used to input the time window sequence into the trained fault identification neural network for any given time window sequence to obtain the fault status identification result of the rolling elements of the TBM main bearing.
[0107] 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, data stored, data displayed, 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 the relevant data must comply with relevant regulations.
[0108] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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.
[0109] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for identifying the rolling element fault status of a small-sample TBM main bearing, characterized in that, The method includes: The vibration signal to be identified of the rolling elements of the TBM main bearing is obtained, and the vibration signal to be identified is divided into multiple time window sequences using time windows. Any time window sequence is input into the trained fault identification neural network to obtain the fault state identification result of the TBM main bearing rolling element; the fault identification neural network is trained using a small sample set; the small sample set is obtained by dividing the historical vibration signal of the TBM main bearing rolling element into different states by time windows; the different states are normal state, scratch damage state or scratch damage state; wherein, one time window is one sample; The fault identification 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 based on the variance of the attention score; the local feature enhancement module is used to perform feature enhancement using multi-scale convolutional kernels; The time window sequence is input into the trained fault identification neural network to obtain the fault state identification result of the TBM main bearing rolling element, specifically including: The time window sequence is input into the data embedding module to obtain the embedding vector; The embedding vector is input into the stacking module to obtain the final feature vector; The final feature vector is input into the fault identification module to obtain the fault probability vector; Based on the fault probability vector, the fault status identification result of the TBM main bearing rolling element is obtained; The stacked 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, specifically including: The embedding vector is input into the first feature extraction module group to obtain the first feature vector; The first feature vector is fed into the second feature extraction module group to obtain the second feature vector; The second feature vector is fed into the third feature extraction module group to obtain the third feature vector; The third feature vector is fed into the fourth feature extraction module group to obtain the final feature vector; The embedding vector is input into the first feature extraction module group to obtain the first feature vector, specifically including: The embedding vector is input to the feature extraction module of the first feature extraction module group to obtain the output vector of the feature extraction module. The output vector of the feature extraction module is input to the local feature enhancement module of the first feature extraction module group to obtain the first feature vector; 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 concatenated with the output of the first normalization layer in the feature extraction module; the input of the feedforward layer is residually concatenated with the output of the second normalization layer in the feature extraction module. The embedded vector is input to the feature extraction module of the first feature extraction module group to obtain the output vector of the feature extraction module, specifically including: The embedding vector is input into the sparse attention layer to obtain the sparse attention vector; The sparse attention vector is input into the first normalization layer in the feature extraction module to obtain the first standard vector. The residual vector is obtained by adding the first standard vector and the embedded vector together. The residual vector is input into the feedforward layer to obtain the extracted vector; The extracted vector is input into the second normalization layer in the feature extraction module to obtain the second standard vector; The second standard vector is added to the residual vector to obtain the output vector of the feature extraction module.
2. The method for identifying the rolling element fault status of a small-sample TBM main bearing according to claim 1, characterized in that, The time window sequence is input into the data embedding module to obtain the embedding vector, specifically including: Using formula The time window sequence is numerically encoded to obtain the numerical encoding result; wherein, The result is a numerical encoding. For one-dimensional convolution, For time window sequences; The following formula is used to encode each position in the time window sequence to obtain the position encoding result. : in, The position code value in dimension 2g for the l-th position in the time window sequence. This is the position code value of the l-th position in the time window sequence, located in dimension 2g+1. This refers to the position index within the time window sequence. For dimensional indexing, L represents the output dimension, and L represents the window length. Using formula Feature fusion is performed on the numerical encoding results and the positional encoding results to obtain the embedding vector: where, For layer normalization, As a pruning strategy, This is the embedding vector.
3. The method for identifying the rolling element fault status of a small-sample TBM main bearing according to claim 1, characterized in that, The embedding vector is input into the sparse attention layer to obtain the sparse attention vector, specifically including: The query vector, key vector, and value vector of the embedded vector are calculated using the following formulas: ; in, For embedding vectors, For query vector, For key vectors, For value vectors, , and The weight matrix is trainable. Based on the query vector, the set quantity is determined using the following formula: ； in, To set the quantity, As a sparsity factor, The length of the query vector; Based on the query vector, the key vector, and the set quantity, the index set of the first u query vectors is determined using the following formula: ； in, To find the maximum absolute value of the dot product of the vector and the key vector, Number the positions of the key vectors. ( The expression () selects the first u indices of the maximum value, where u is the set number. The set of indices for the first u query vectors; Based on the index set of the first u query vectors, the following formula is used to determine the index of the th element in the index set. Attention score for each query vector: ； in, The index of the query vector in the index set. For the first in the index set There are query vectors, where d is the dimension of the key vector. For the first Attention score for each query vector; Based on the first The attention score of the query vector and the key vector are used to determine the first query vector using the following formula. Variance of attention scores for each query vector: ； in, For the first query vectors With the Key vectors Attention score For the first query vectors Attention score for all keys; Based on the first The variance of the attention score for the i-th query vector is determined using the following formula. Weights of the variance of the attention score for each query vector: in, The modulation coefficient, For the first The weights of the attention score for each query vector; Based on the first The weights of the variance of the attention score for each query vector are calculated using the formula... Determine the first Modulation weights of query vectors : Based on the first The modulation weights of each query vector and the value vector are calculated using the formula... Determine the first The sparse attention vector corresponding to each query vector .
4. The method for identifying the rolling element fault status of a small-sample TBM main bearing according to claim 1, characterized in that, The local feature enhancement module includes a convolutional layer, a normalization layer, an activation layer, and a max pooling layer connected in sequence; the convolutional layer includes multiple convolutional kernels of different scales; The output vector of the feature extraction module is input to the local feature enhancement module of the first feature extraction module group to obtain the first feature vector, specifically including: Based on the output vector of the feature extraction module, the formula is used. This yields the sum of the output vectors of the convolutional layer; where 'a' is the size index of the convolutional kernel. Let be the size of the convolution kernel of size a. For learnable adaptive weights, The sum of the output vectors of the convolutional layer. This is the output vector of the feature extraction module; Based on the sum of the output vectors of the convolutional layers, the formula is used. The first feature vector is determined; where BatchNorm is the batch normalization in the normalization layer. This is the maximum mean pooling in the maximum pooling layer. The first eigenvector, This refers to the activation function in the activation layer.
5. The method for identifying the 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, another linear layer, and a Softmax activation layer connected in sequence.
6. A small-sample TBM main bearing rolling element fault state identification system, employing the small-sample TBM main bearing rolling element fault state identification method according to any one of claims 1-5, characterized in that, The system includes: The time window sequence acquisition module is used to acquire the vibration signal to be identified of the rolling elements of the TBM main bearing, and divide the vibration signal to be identified into multiple time window sequences by using time windows. The fault status identification module is used to input the time window sequence into the trained fault identification neural network for any given time window sequence to obtain the fault status identification result of the rolling elements of the TBM main bearing.