A shield machine main drive bearing fault diagnosis method and system based on VMD-FFT and variable working condition adaptive deep learning
Patent Information
- Application Number
- CN202611019770.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-09
- Publication Date
- 2026-10-20
AI Technical Summary
若仍采用固定故障特征频带进行 FFT 特征提取，则容易造成故障频率漂移条件下的漏检或误检；若仅依赖振动信号进行深度学习诊断，也难以区分由工况突变引起的振动响应变化与由轴承故障引起的振动响应变化
1、本发明通过建立基于实时转速的故障特征频率动态跟踪机制，解决了变转速工况下固定频带特征提取失效的问题。根据同步采集的主驱动转速及轴承结构参数，实时计算外圈、内圈、滚动体和保持架的时变故障特征频率，并据此构建动态特征频带，替代现有固定频带提取方式；同时，工况特征分支将转速、扭矩、推进力、贯入度、温度等工况参数映射为工况嵌入特征，通过交叉注意力机制对时序特征和时频特征进行工况约束下的自适应融合，实现工况扰动与真实故障特征的有效分离。经实验验证，相较于固定频带特征提取结合支持向量机的方法，本发明方法准确率显著提升，误报率和漏报率明显降低，在变转速、扭矩波动和冲击扰动等复杂工况下召回率均保持在较高水平，表明本发明具有显著的变工况诊断稳定性。
Smart Images
Figure FT_1 
Figure FT_2 
Figure SMS_1
Abstract
Description
Technical Field
[0001] This invention relates to the field of soil testing technology, specifically to a method and system for diagnosing faults in the main drive bearing of a tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions. Background Technology
[0002] As a core load-bearing and transmission component, the health of the main drive bearing of a tunnel boring machine directly determines the overall tunneling efficiency and construction safety. Operating under harsh conditions of heavy loads, strong impacts, and high dust levels for extended periods, the main drive bearing is highly susceptible to pitting, spalling, and cracking. Failure to detect early-stage faults can lead to bearing failure or even the complete shutdown of the main drive system, resulting in significant economic losses and potential engineering safety hazards.
[0003] Currently, vibration signal analysis is the mainstream method for bearing fault diagnosis. However, fault diagnosis of the main drive bearing of a tunnel boring machine (TBM) faces many challenges: First, the early fault characteristic signals are extremely weak and easily drowned out by strong background noise; second, during actual tunneling, the load and rotational speed of the TBM constantly change, causing the vibration signals to exhibit significant non-stationary and nonlinear characteristics, making feature extraction difficult; third, traditional fault diagnosis methods rely heavily on the prior knowledge of experts, have limited automation and intelligence levels, and are insufficiently adaptable to complex and variable working conditions.
[0004] In recent years, deep learning technology has provided new solutions for intelligent fault diagnosis of mechanical equipment. However, a single deep learning model often struggles to simultaneously and effectively capture the spatial and temporal features of fault signals, and its diagnostic accuracy and robustness still need improvement when dealing with weak, complex faults in a noisy environment.
[0005] Furthermore, during the actual tunneling process of a tunnel boring machine, the main drive bearing is often under low-speed, heavy-load, variable-speed, and variable-load conditions. Changes in cutterhead torque, propulsion force, penetration depth, and ground resistance will cause changes in the amplitude and spectral distribution of vibration signals. If a fixed fault characteristic frequency band is still used for FFT feature extraction, it is easy to cause missed or false detections under fault frequency drift conditions. If deep learning diagnosis is performed solely based on vibration signals, it is also difficult to distinguish between vibration response changes caused by sudden changes in operating conditions and vibration response changes caused by bearing failure.
[0006] Therefore, a fault diagnosis method and system for the main drive bearing of a tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions is needed to solve the above problems. Summary of the Invention
[0007] The purpose of this invention is to provide a method and system for diagnosing faults in the main drive bearing of a tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions, so as to solve the problems existing in the prior art mentioned in the background.
[0008] To achieve the above objectives, the present invention provides the following technical solution: A fault diagnosis method for the main drive bearing of a tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions includes the following steps: S1: Signal Acquisition and Preprocessing: The original vibration signal during the tunneling process is acquired by an acceleration sensor arranged in the main drive bearing seat of the tunnel boring machine, and the main drive speed and operating parameters are acquired simultaneously; the original vibration signal is denoised and normalized, and the operating parameters are aligned with the original vibration signal according to the sampling timestamp. S2: VMD Adaptive Decomposition: The preprocessed vibration signal is decomposed using Variational Mode Decomposition (VMD) to adaptively decompose the signal into K quasi-orthogonal eigenmode function (IMF) components. S3: FFT Frequency Domain Feature Extraction: Perform Fast Fourier Transform (FFT) on each Intrinsic Mode Function (IMF) component, calculate the time-varying fault characteristic frequency based on the synchronously acquired main drive speed and main bearing structural parameters, set a dynamic characteristic frequency band according to the time-varying fault characteristic frequency, extract frequency domain features from the dynamic characteristic frequency band, and construct a dynamic frequency domain feature vector; the dynamic characteristic frequency band is updated as the main drive speed changes; S4: Time-frequency feature sample construction: Construct a one-dimensional time-series feature sample set based on the intrinsic mode function (IMF) components; generate a two-dimensional time-frequency feature sample set based on the combination of the Fast Fourier Transform (FFT) spectrum and the time-domain statistical features of the IMF components; S5: Deep feature extraction and fusion: The time-series feature sample set, time-frequency feature sample set, and operating condition parameter features are respectively input into a deep learning model that integrates a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), operating condition feature branches, and a cross-attention mechanism for diagnosis; S6: Fault Diagnosis and Output: Input the fused joint feature vector into the classifier and output the probability distribution of different fault states of the main bearing to achieve accurate fault classification and early warning.
[0009] Preferably, in step S2, the variational mode decomposition (VMD) is achieved by solving the following constrained variational problem:
[0010]
[0011] in, and These represent the K IMF components and their center frequencies, respectively; f is the original signal; and δ(t) is the Dirac function. The specific process of step S2 includes: S21: Initialize the number of modes K, penalty parameter α, convergence tolerance ε, and set the initial center frequency of each IMF component. ; S22: Iteratively update each IMF component using the Alternating Direction Multiplier Method (ADMM). and its center frequency ; S22: Determine if the iteration satisfies the convergence condition:
[0012] If the condition is met, output K IMF components; otherwise, return to step S22 to continue iterating.
[0013] Preferably, in step S3, for the k-th IMF component Perform an FFT transform to obtain the frequency domain representation. :
[0014] and from the spectrum The amplitude near the preset fault characteristic frequency is extracted as the frequency domain feature.
[0015] Preferably, in step S3, the time-varying fault characteristic frequency is calculated based on the synchronously acquired main drive speed n(t) and main bearing structural parameters. The main bearing structural parameters include the number of rolling elements z, the rolling element diameter d, the pitch circle diameter D, and the contact angle θ. The time-varying fault characteristic frequencies include the outer ring fault frequency BPFO(t), the inner ring fault frequency BPFI(t), the rolling element fault frequency BSF(t), and the cage fault frequency FTF(t), and the calculation formulas are as follows:
[0016]
[0017]
[0018]
[0019] Based on the characteristic frequency of time-varying faults, a dynamic characteristic frequency band is set. ,in The frequency is the time-varying fault characteristic frequency corresponding to the fault type, and r is the bandwidth coefficient.
[0020] Preferably, the steps for constructing the deep learning model include: S51: The CNN branch processes the time-frequency feature sample set and extracts spatial features through convolutional layers. Its convolutional operation is represented as follows:
[0021] in, To input a time-frequency graph, and Here, σ represents the kernel weights and biases, and σ is the activation function. This represents the convolution operation; S52: The BiLSTM branch processes the temporal feature sample set and extracts temporal features through bidirectional LSTM units; for a time step t, the forward LSTM outputs... With backward LSTM output Through splicing and fusion:
[0022] The hidden states at all time steps constitute the temporal feature matrix. ; S53: Cross-attention module with temporal features As a query Based on spatial characteristics As a key Sum The calculation process is as follows: Attention weight calculation:
[0023] Context vector generation:
[0024] Feature fusion:
[0025] in, Let be the dimension of the key vector. The fused joint feature vector and For learnable fusion weights and biases.
[0026] Preferably, in step S5, the cross-attention mechanism further includes a condition constraint fusion step: Features embedded in working conditions As auxiliary query or gating variables, the correlation weights between operating condition characteristics and time series characteristics, and between operating condition characteristics and time-frequency characteristics, are calculated respectively, and adjustments are made based on the correlation weights. and The proportion in the final fusion features.
[0027] Preferably, in step S6, a Softmax classifier is used to output the probability distribution of the fault type:
[0028] Where C represents the total number of fault categories. and These are the weights and biases corresponding to category c.
[0029] Preferably, in step S1, the operating parameters include at least one of the following: main drive speed, cutterhead torque, propulsion force, penetration depth, and main drive temperature; when the sampling frequency of the operating parameters is lower than the sampling frequency of the vibration signal, the operating parameters are mapped to the vibration signal sample window using linear interpolation, nearest neighbor hold, or piecewise averaging.
[0030] A fault diagnosis system for the main drive bearing of a tunnel boring machine includes: The sensing and acquisition layer includes a vibration sensor array and a working condition parameter acquisition module. The vibration sensor array is used to acquire bearing vibration signals, and the working condition parameter acquisition module is used to synchronously acquire at least one working condition parameter among the main drive speed, cutter head torque, propulsion force, penetration depth, and main drive temperature. The edge processing unit includes a signal preprocessing module, a VMD-FFT feature calculation engine, a dynamic fault frequency tracking module, and a multimodal feature construction module. The signal preprocessing module is used to preprocess the acquired signals. The VMD-FFT feature calculation engine is used to perform VMD decomposition and FFT transformation. The dynamic fault frequency tracking module is used to calculate the time-varying fault characteristic frequency based on the main drive speed and main bearing structural parameters. The multimodal feature construction module is used to generate time-series features, time-frequency features, and operating condition features. The intelligent diagnostic platform is equipped with a deep learning model that integrates convolutional neural networks, bidirectional long short-term memory networks, operating condition feature branches, and cross-attention mechanisms. It is used to receive the time-series features, time-frequency features, and operating condition features, perform fault diagnosis, and output diagnostic results. The application interactive terminal includes a status monitoring interface and an early warning module for result display and alarm.
[0031] Preferably, the vibration sensor array consists of multiple accelerometers mounted in an array at key radial positions on the main drive bearing housing; the intelligent diagnostic platform is deployed on a server, equipped with a deep learning model that integrates CNN, BiLSTM, and cross-attention mechanisms, and provides an API interface to receive feature samples and return diagnostic results.
[0032] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention solves the problem of fixed-band feature extraction failure under variable speed conditions by establishing a dynamic tracking mechanism for fault feature frequencies based on real-time rotational speed. Based on synchronously acquired main drive speed and bearing structural parameters, the time-varying fault feature frequencies of the outer ring, inner ring, rolling elements, and cage are calculated in real time, and a dynamic feature band is constructed accordingly, replacing the existing fixed-band extraction method. Simultaneously, the operating condition feature branch maps operating condition parameters such as speed, torque, thrust, penetration, and temperature into operating condition embedded features. Through a cross-attention mechanism, time-series and time-frequency features are adaptively fused under operating condition constraints, achieving effective separation of operating condition disturbances from actual fault features. Experimental verification shows that compared to the fixed-band feature extraction combined with support vector machines, the accuracy of this invention is significantly improved, and the false alarm rate and false negative rate are significantly reduced. The recall rate remains at a high level under complex operating conditions such as variable speed, torque fluctuations, and impact disturbances, indicating that this invention has significant stability in variable operating condition diagnosis.
[0033] 2. This invention overcomes the limitations of single models in simultaneously capturing spatial local features, long-term temporal dependencies, and operational condition coupling information by constructing a three-modal deep learning fusion architecture encompassing time series, time frequency, and operational condition. A convolutional neural network branch extracts spatial fault imprints from a two-dimensional time-frequency graph, while a bidirectional long short-term memory network branch extracts bidirectional long-term temporal dependencies from a one-dimensional time series signal. A cross-attention module enables adaptive alignment and refined fusion of heterogeneous modal features. Operational condition embedded features participate in the fusion process with gated weights, dynamically adjusting the contribution ratio of each modality feature, ensuring the model maintains high recognition sensitivity under conditions of strong background noise and weak early faults. Experimental results demonstrate that, compared to methods combining convolutional neural networks and bidirectional long short-term memory networks without cross-attention mechanisms, this invention effectively improves both accuracy and F1 score, and the single-window inference time meets the timeliness requirements for online engineering deployment. Attached Figure Description
[0034] Figure 1 This is a flowchart of the method of the present invention.
[0035] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation
[0036] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0037] Please see Figure 1-2 The present invention provides the following technical solutions: A fault diagnosis method for the main drive bearing of a tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions includes the following steps: S1: Signal Acquisition and Preprocessing: Raw vibration signals during the tunneling process are acquired using accelerometers mounted on the main drive bearing housing of the tunnel boring machine (TBM), along with the main drive speed and operating parameters. The raw vibration signals are then denoised and normalized, and the operating parameters are aligned with the raw vibration signals according to the sampling timestamps. The specific implementation process includes: S11: An ICP accelerometer with model PCB 608A11 is used to collect the radial vibration acceleration signal of the main drive bearing at a sampling frequency of 12.8 kHz. S12: Signal preprocessing: The original signal is initially denoised using a bandpass filter from 20 Hz to 2000 Hz to remove high-frequency noise and low-frequency drift. The filtered signal is normalized using the Z-Score normalization method to make its mean 0 and standard deviation 1. S13: While acquiring vibration acceleration signals, synchronously read operating parameters such as main drive speed, cutterhead torque, thrust, penetration depth, and main drive temperature from the tunnel boring machine PLC or main control system. Each operating parameter is aligned with the vibration signal according to the acquisition timestamp. When the sampling frequency of the operating parameters is lower than the sampling frequency of the vibration signal, linear interpolation, nearest neighbor hold, or piecewise averaging can be used to map the operating parameters to the vibration signal sample window.
[0038] S2: VMD Adaptive Decomposition: The preprocessed vibration signal is decomposed using Variational Mode Decomposition (VMD), adaptively decomposing the signal into K quasi-orthogonal Intrinsic Mode Function (IMF) components. The specific implementation process includes: S21: Initialize VMD parameters: Set the number of modes K=5, the penalty parameter α=2000, and the convergence tolerance ε=1e-7; S22: Solve the constrained variational problem iteratively using the Alternating Direction Multiplier Method (ADMM) to update each IMF component. and its center frequency ; S23: Determine if the iteration satisfies the convergence condition: ; S24: Iteration terminates, output 5 IMF components (IMF1 to IMF5).
[0039] S3: FFT Frequency Domain Feature Extraction: A Fast Fourier Transform (FFT) is performed on each Intrinsic Mode Function (IMF) component. Based on the synchronously acquired main drive speed and main bearing structural parameters, time-varying fault characteristic frequencies are calculated. Dynamic feature frequency bands are set according to these time-varying fault characteristic frequencies. Frequency domain features are extracted from these dynamic feature frequency bands to construct a dynamic frequency domain feature vector. The dynamic feature frequency bands are updated as the main drive speed changes. The specific implementation process includes: For the k-th IMF component Perform FFT transformation , , to obtain the frequency domain representation From the spectrum In this study, the amplitude values within the characteristic frequency bands corresponding to the failure frequencies of the bearing outer ring, inner ring, and rolling elements are extracted to form a multi-dimensional frequency domain feature vector.
[0040] As a further implementation, considering that the main drive bearing of the tunnel boring machine is in a variable speed condition, the outer ring failure frequency is calculated based on the synchronously acquired main drive speed n(t) and main bearing structural parameters. Inner ring failure frequency Rolling element failure frequency and cage failure frequency The structural parameters of the main bearing include the number of rolling elements z, the rolling element diameter d, the pitch circle diameter D, and the contact angle. ;in The main driving speed.
[0041]
[0042]
[0043]
[0044]
[0045] Based on the aforementioned time-varying fault characteristic frequencies, a dynamic characteristic frequency band is set. , in The frequency is the time-varying fault characteristic frequency corresponding to the fault type, and r is the bandwidth coefficient. Features such as amplitude, energy, spectral kurtosis, and sideband energy are extracted from the dynamic characteristic frequency band to form a dynamic frequency domain feature vector. This implementation does not change the original FFT feature extraction framework; it only replaces the original fixed fault frequency band with a dynamic fault frequency band that updates with the main drive speed.
[0046] S4: Time-frequency feature sample construction: A one-dimensional time-series feature sample set is constructed based on the intrinsic mode function (IMF) components; a two-dimensional time-frequency feature sample set is generated by combining the Fast Fourier Transform (FFT) spectrum with the time-domain statistical features of the IMF components; the specific implementation process includes: S41: Construction of one-dimensional time series feature sample set: Select time domain signals of IMF2 to IMF4 components that are sensitive to faults; take 1024 data points as a sample, use the sliding window method to extract data, and the window overlap rate is 50%.
[0047] S42: Construction of a two-dimensional time-frequency feature sample set: Combine the FFT spectrum amplitude (range of 0-2000 Hz) of each IMF component with the statistical features of the corresponding time domain signal, such as kurtosis, peak value, and root mean square value, and reshape the combined features into a 64-pixel × 64-pixel grayscale image to construct a time-frequency image sample set.
[0048] S43: Normalize the working condition parameters corresponding to each sliding window in S13 to construct a working condition feature vector. The working condition feature vector includes at least one of the following: main drive speed, cutterhead torque, propulsion force, penetration depth, and main drive temperature.
[0049] S5: Deep Feature Extraction and Fusion: The time-series feature sample set, time-frequency feature sample set, and operating condition parameter features are respectively input into a deep learning model that integrates a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), an operating condition feature branch, and a cross-attention mechanism for diagnosis; the specific implementation process includes: S51: The CNN branch processes the time-frequency feature sample set: the input is the reshaped 64×64 time-frequency map. The network structure includes: First convolutional layer: 32 3×3 convolutional kernels, ReLU activation function, output connected to a 2×2 max pooling layer; Second convolutional layer: 64 3×3 convolutional kernels, ReLU activation function, output connected to a 2×2 max pooling layer; flattening the feature map, outputting a spatial feature vector. The convolution operation is represented as:
[0050] in, To input a time-frequency graph, and Here, σ represents the kernel weights and biases, and σ is the activation function. This represents the convolution operation.
[0051] S52: BiLSTM branch processing of time-series feature sample sets: The input is a one-dimensional time-series signal with 1024 points; the network structure contains a bidirectional LSTM layer, with 128 hidden units per LSTM layer. For one time step t, the forward LSTM output... With backward LSTM output By splicing and fusing, the hidden states of all time steps constitute a temporal feature matrix. .
[0052] S53: Cross attention module with temporal features As a query Based on spatial characteristics As a key Sum The calculation process is as follows: Attention weight calculation:
[0053] Context vector generation:
[0054] Feature fusion:
[0055] in, Let be the dimension of the key vector. The fused joint feature vector and For learnable fusion weights and biases.
[0056] S54: Add a working condition feature branch to the existing CNN and BiLSTM branches; input the working condition feature vector into a fully connected network or a one-dimensional convolutional network to obtain the working condition embedded features. ;Will and , After linearly mapping to the same dimension, gating weights or attention weights are generated using the embedded features of the working conditions. The temporal and time-frequency features are then adaptively weighted and fused to obtain the joint feature vector under the working condition constraints. .
[0057] S55: The working condition constraint fusion adopts the following method: embedding features based on working conditions. As auxiliary query or gating variables, the correlation weights between operating condition characteristics and time series characteristics, and between operating condition characteristics and time-frequency characteristics, are calculated respectively, and adjustments are made based on the correlation weights. and The proportion in the final fusion features; when the sudden change in cutterhead torque or propulsion force causes non-fault changes in the vibration signal, the model can reduce the impact of such disturbances on the fault classification results based on the embedded features of the working condition.
[0058] S6: Fault Diagnosis and Output: Input the fused joint feature vector into the classifier and output the probability distribution of different fault states of the main bearing to achieve accurate fault classification and early warning.
[0059] Use the Softmax classifier to output the probability distribution of fault types:
[0060] Where C represents the total number of fault categories. and These correspond to the weights and biases for category c. When the predicted probability of a certain fault type exceeds a preset threshold of 90%, an early warning signal is issued via the application's interactive terminal.
[0061] like Figure 2 As shown, the present invention also provides a fault diagnosis system for the main drive bearing of a tunnel boring machine, comprising: The sensing and acquisition layer includes a vibration sensor array for collecting bearing vibration signals. In this embodiment, multiple high-precision accelerometers are installed in an array at key radial positions (e.g., 0°, 90°, 180°, 270°) on the main drive bearing housing, and the signals are transmitted to the edge processing unit via shielded cables. As a further implementation, the sensing and acquisition layer also includes a working condition parameter acquisition module for synchronously acquiring working condition parameters such as main drive speed, cutterhead torque, thrust, penetration depth, and main drive temperature from the tunnel boring machine's main control system.
[0062] The edge processing unit includes a signal preprocessing module and a VMD-FFT feature calculation engine for signal processing and feature extraction. In this embodiment, an industrial-grade edge computing gateway is used, which has built-in signal conditioning circuitry and a highly efficient digital signal processing program to perform signal preprocessing, VMD decomposition, FFT transformation, and feature sample construction tasks (S1 to S4) in real time. As a further implementation, the edge processing unit also includes a dynamic fault frequency tracking module and a multimodal feature construction module, used to calculate time-varying fault feature frequencies and construct time-series features, time-frequency features, and operating condition features, respectively.
[0063] Intelligent diagnostic platform: This includes a deep diagnostic model framework and model inference for fault diagnosis. In this embodiment, the platform is deployed on a data center server at the project site, built on the Python and TensorFlow frameworks, and equipped with the hybrid deep learning model described in Embodiment 1. It provides a RESTful API interface to receive feature samples and return diagnostic results. As a further implementation, the intelligent diagnostic platform receives time-series features, time-frequency features, and operating condition features, and performs multimodal feature fusion under operating condition constraints based on a cross-attention mechanism.
[0064] Application Interaction Terminal: Includes a status monitoring interface and an early warning module for result display and alarm. In this embodiment, the terminal is an industrial monitoring computer deployed in the main control room of the tunnel boring machine. It has a graphical visualization interface for displaying the health status curve and fault probability distribution of the main drive bearing in real time, and issuing an early warning in the form of sound and light when a fault is diagnosed.
[0065] The system in this invention, through the coordinated operation of the above modules, realizes online, automatic, and intelligent diagnosis and early warning of the status of the main drive bearing of the tunnel boring machine, effectively solving the industry problem of difficulty in extracting and diagnosing weak faults under strong noise and variable working conditions.
[0066] Experimental verification Experimental verification is used to illustrate the diagnostic effectiveness of this invention under low-speed, heavy-load, variable-speed, and high-impact conditions. Since it is difficult to obtain actual early-stage fault samples of the main drive bearing of a tunnel boring machine (TBM) through artificial construction on-site, this invention uses equivalent low-speed, heavy-load main bearing test bench data, and simultaneously constructs or collects operating condition parameter data corresponding to the TBM construction state for verification. This equivalent data is used to illustrate the implementation process and technical effects of the method of this invention, and does not limit the invention to implementation only under these data conditions; in actual engineering deployment, the equivalent test bench data in this invention can be replaced with historical monitoring data or online monitoring data of the TBM main drive system.
[0067] The experimental verification data comes from two sources: the first part is the main bearing vibration data, which is collected by an accelerometer installed in the radial position of the bearing housing, and is used to characterize the vibration response under normal, outer ring failure, inner ring failure, rolling element failure and cage failure conditions; the second part is the operating condition parameter data, including the main drive speed, simulated cutterhead torque, axial load, equivalent propulsion load and temperature data, which are used to characterize the impact of low speed heavy load, variable speed and load fluctuation on the vibration signal during shield tunneling.
[0068] Table 1 Experimental Data and Operating Conditions
[0069] During sample construction, the vibration signal is subjected to bandpass filtering, normalization, VMD decomposition, and FFT transformation according to the method of this invention, and samples are constructed using a sliding window approach. Each sample includes three types of inputs: one-dimensional time-series features, two-dimensional time-frequency features, and operating condition parameter features. Fault labels are marked according to the fault type set on the test bench or engineering maintenance records and anomaly records.
[0070] Table 2 Fault Categories and Sample Division
[0071] To verify the improvement effect of the present invention compared with the traditional solution, the following comparison methods were set up: Method A is a fixed frequency band FFT plus SVM classifier; Method B is VMD-FFT plus SVM classifier; Method C is VMD-FFT plus CNN classifier; Method D is VMD-FFT plus CNN-BiLSTM classifier; Method E is the method of the present invention, namely the VMD-FFT multimodal fusion diagnostic method that combines dynamic fault frequency tracking, operating condition feature branching and cross-attention mechanism.
[0072] Evaluation metrics included overall accuracy, recall, F1 score, false alarm rate, missing alarm rate, and single-window inference time. The model was trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, and 100 training epochs. An early stopping strategy was implemented to prevent overfitting.
[0073] Table 3. Comparison of diagnostic performance of different methods
[0074] As shown in Table 3, under the same training, validation, and test set partitioning conditions, the method of this invention shows a certain improvement compared to fixed-band FFT+SVM, VMD-FFT+SVM, and deep learning methods without introducing operating condition constraint fusion, but the improvement remains within a reasonable range. These results indicate that VMD adaptive decomposition can weaken some of the influence of strong noise, dynamic fault frequency tracking can alleviate the fault frequency drift problem caused by variable speed, and operating condition feature branching can reduce the interference of sudden load changes on fault judgment to a certain extent.
[0075] Table 4. Stability verification results of the method of the present invention under different operating conditions.
[0076] As shown in Table 4, the method of the present invention can maintain a basically stable fault identification effect under different working conditions such as stable low speed, continuous variable speed, torque fluctuation, and impact disturbance. However, the accuracy decreases to some extent under torque fluctuation and impact disturbance conditions, which is consistent with the actual characteristics of fault diagnosis under strong disturbance environment of the main drive bearing of the tunnel boring machine. Especially in the working conditions of continuous variable speed and impact disturbance, the dynamic fault frequency tracking module can update the characteristic frequency band in real time according to the rotation speed, and the working condition feature branch can introduce load disturbance information into the model fusion process, thereby reducing the misjudgment caused by fixed frequency band extraction and single vibration signal diagnosis.
[0077] In summary, the above experiments demonstrate the feasibility and technical effectiveness of this invention from the aspects of data construction, comparison methods, evaluation indicators, comparative tests, and working condition stability verification. Experimental verification results show that this invention can improve the stability of main drive bearing fault diagnosis in tunnel boring machines (TBMs) under low-speed, heavy-load, variable-speed, and strong-impact conditions, reduce some false alarms and missed alarms, and meet the basic timeliness requirements of online inference for edge processing units and intelligent diagnostic platforms. The above experimental data are exemplary verification data; in specific engineering deployments, retraining and calibration can be performed based on the TBM model, main bearing structural parameters, sensor arrangement, and construction geological conditions.
[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for fault diagnosis of the main drive bearing of a tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions, characterized in that, Includes the following steps: S1: Signal Acquisition and Preprocessing: The original vibration signal during the tunneling process is acquired by an acceleration sensor arranged in the main drive bearing seat of the tunnel boring machine, and the main drive speed and operating parameters are acquired simultaneously; the original vibration signal is denoised and normalized, and the operating parameters are aligned with the original vibration signal according to the sampling timestamp. S2: VMD Adaptive Decomposition: The preprocessed vibration signal is decomposed using Variational Mode Decomposition (VMD) to adaptively decompose the signal into K quasi-orthogonal eigenmode function (IMF) components. S3: FFT Frequency Domain Feature Extraction: Perform Fast Fourier Transform (FFT) on each Intrinsic Mode Function (IMF) component, calculate the time-varying fault characteristic frequency based on the synchronously acquired main drive speed and main bearing structural parameters, set a dynamic characteristic frequency band according to the time-varying fault characteristic frequency, extract frequency domain features from the dynamic characteristic frequency band, and construct a dynamic frequency domain feature vector; the dynamic characteristic frequency band is updated as the main drive speed changes; S4: Time-frequency feature sample construction: Construct a one-dimensional time-series feature sample set based on the intrinsic mode function (IMF) components; generate a two-dimensional time-frequency feature sample set based on the combination of the Fast Fourier Transform (FFT) spectrum and the time-domain statistical features of the IMF components; S5: Deep feature extraction and fusion: The time-series feature sample set, time-frequency feature sample set, and operating condition parameter features are respectively input into a deep learning model that integrates a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), operating condition feature branches, and a cross-attention mechanism for diagnosis; S6: Fault Diagnosis and Output: Input the fused joint feature vector into the classifier and output the probability distribution of different fault states of the main bearing to achieve accurate fault classification and early warning.
2. The method for fault diagnosis of main drive bearing of tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions as described in claim 1, characterized in that, In step S2, the variational mode decomposition (VMD) is achieved by solving the following constrained variational problem: in, and These represent the K IMF components and their center frequencies, respectively; f is the original signal; and δ(t) is the Dirac function. The specific process of step S2 includes: S21: Initialize the number of modes K, penalty parameter α, convergence tolerance ε, and set the initial center frequency of each IMF component. ; S22: Iteratively update each IMF component using the Alternating Direction Multiplier Method (ADMM). and its center frequency ; S22: Determine if the iteration satisfies the convergence condition: If the condition is met, output K IMF components; otherwise, return to step S22 to continue iterating.
3. The method for diagnosing faults in the main drive bearing of a tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions, as described in claim 2, is characterized in that... In step S3, for the k-th IMF component Perform an FFT transform to obtain the frequency domain representation. : and from the spectrum The amplitude near the preset fault characteristic frequency is extracted as the frequency domain feature.
4. The method for diagnosing the main drive bearing fault of a tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions, as described in claim 1, is characterized in that... In step S3, the time-varying fault characteristic frequency is calculated based on the synchronously acquired main drive speed n(t) and main bearing structural parameters. The main bearing structural parameters include the number of rolling elements z, the rolling element diameter d, the pitch circle diameter D, and the contact angle θ. The time-varying fault characteristic frequencies include the outer ring fault frequency BPFO(t), the inner ring fault frequency BPFI(t), the rolling element fault frequency BSF(t), and the cage fault frequency FTF(t), and the calculation formulas are as follows: Based on the characteristic frequency of time-varying faults, a dynamic characteristic frequency band is set. , in The frequency is the time-varying fault characteristic frequency corresponding to the fault type, and r is the bandwidth coefficient.
5. The method for diagnosing faults in the main drive bearing of a tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions, as described in claim 1, is characterized in that... In step S5, the construction steps of the deep learning model include: S51: The CNN branch processes the time-frequency feature sample set and extracts spatial features through convolutional layers. Its convolutional operation is represented as follows: in, To input a time-frequency graph, and Here, σ represents the kernel weights and biases, and σ is the activation function. This represents the convolution operation; S52: The BiLSTM branch processes the temporal feature sample set and extracts temporal features through bidirectional LSTM units; for a time step t, the forward LSTM outputs... With backward LSTM output Through splicing and fusion: The hidden states at all time steps constitute the temporal feature matrix. ; S53: Cross-attention module with temporal features As a query Based on spatial characteristics As a key Sum The calculation process is as follows: Attention weight calculation: Context vector generation: Feature fusion: in, Let be the dimension of the key vector. The fused joint feature vector and For learnable fusion weights and biases.
6. The method for diagnosing the main drive bearing fault of a tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions, as described in claim 5, is characterized in that... In step S5, the cross-attention mechanism further includes a condition constraint fusion step: Features embedded in working conditions As auxiliary query or gating variables, the correlation weights between operating condition characteristics and time series characteristics, and between operating condition characteristics and time-frequency characteristics, are calculated respectively, and adjustments are made based on the correlation weights. and The proportion in the final fusion features.
7. The method for fault diagnosis of the main drive bearing of a tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions, as described in claim 5, is characterized in that... In step S6, the Softmax classifier is used to output the probability distribution of the fault type: Where C represents the total number of fault categories. and These are the weights and biases corresponding to category c.
8. The method for diagnosing the main drive bearing fault of a tunnel boring machine based on VMD-FFT and adaptive deep learning under varying working conditions, as described in claim 1, is characterized in that... In step S1, the operating parameters include at least one of the following: main drive speed, cutterhead torque, propulsion force, penetration depth, and main drive temperature; when the sampling frequency of the operating parameters is lower than the sampling frequency of the vibration signal, the operating parameters are mapped to the vibration signal sample window using linear interpolation, nearest neighbor hold, or piecewise averaging.
9. A fault diagnosis system for the main drive bearing of a tunnel boring machine for implementing the method according to any one of claims 1-8, characterized in that, include: The sensing and acquisition layer includes a vibration sensor array and a working condition parameter acquisition module. The vibration sensor array is used to acquire bearing vibration signals, and the working condition parameter acquisition module is used to synchronously acquire at least one working condition parameter among the main drive speed, cutter head torque, propulsion force, penetration depth, and main drive temperature. The edge processing unit includes a signal preprocessing module, a VMD-FFT feature calculation engine, a dynamic fault frequency tracking module, and a multimodal feature construction module; The signal preprocessing module is used to preprocess the acquired signals; the VMD-FFT feature calculation engine is used to perform VMD decomposition and FFT transformation; the dynamic fault frequency tracking module is used to calculate the time-varying fault characteristic frequency based on the main drive speed and main bearing structural parameters; and the multimodal feature construction module is used to generate time-series features, time-frequency features, and operating condition features. The intelligent diagnostic platform is equipped with a deep learning model that integrates convolutional neural networks, bidirectional long short-term memory networks, operating condition feature branches, and cross-attention mechanisms. It is used to receive the time-series features, time-frequency features, and operating condition features, perform fault diagnosis, and output diagnostic results. The application interactive terminal includes a status monitoring interface and an early warning module for result display and alarm.
10. The tunnel boring machine main drive bearing fault diagnosis system according to claim 9, characterized in that, The vibration sensor array consists of multiple accelerometers mounted in an array at key radial positions on the main drive bearing housing; the intelligent diagnostic platform is deployed on a server, equipped with a deep learning model that integrates CNN, BiLSTM and cross-attention mechanisms, and provides an API interface to receive feature samples and return diagnostic results.