A method and system for fault diagnosis and life prediction of tunnel boring machine components.

Abstract

This invention discloses a method and system for fault diagnosis and life prediction of tunnel boring machine (TBM) components, relating to the field of intelligent operation and maintenance technology. The method includes receiving and storing multi-source time-series data, equipment files, and historical maintenance records through a cloud platform big data center, and issuing model threshold update instructions, sensor configuration instructions, and diagnostic task instructions. The multi-source time-series data includes preprocessed feature vectors, alarm logs, and sensor health codes. A digital twin is constructed in the cloud platform big data center, and virtual-real mapping and dynamic calibration are performed on the TBM components. The digital twin includes a parametric geometric model, a multibody dynamics model, a fault injection model, and a virtual-real calibration submodule. This invention ensures the reliability of the model mechanism and improves the sensitivity and classification accuracy for early and complex faults through real-time calibration of the digital twin and the physical entity.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology, and in particular to a method and system for fault diagnosis and life prediction of tunnel boring machine components. Background Technology

[0002] As the core equipment in tunnel engineering, tunnel boring machines (TBMs) are characterized by complex structures, harsh working conditions, high unit costs, and high maintenance costs. Failures in key components such as the main drive bearing, gearbox, and hydraulic cylinders can easily lead to unplanned downtime, with direct economic losses from a single unplanned downtime typically exceeding 500,000 yuan.

[0003] Current tunnel boring machine (TBM) monitoring solutions in the industry still have significant limitations: edge computing power is limited and can only support real-time alarms, unable to achieve in-depth diagnosis and lifespan prediction; after long-term operation of cloud-based digital twins, the deviation between simulation and actual measurement increases, with some deviations exceeding 10%, leading to state distortion; existing deep learning models mostly rely on pure data-driven approaches without incorporating mechanistic information, resulting in an accuracy rate of less than 85% in identifying common multi-fault coupling scenarios; at the same time, there is a lack of effective collaboration mechanisms between the edge and the cloud, which prevents the formation of data interaction and command loops, thus restricting maintenance response efficiency. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for fault diagnosis and life prediction of tunnel boring machine components to solve the problems existing in fault diagnosis and life prediction of tunnel boring machines, such as the virtual-real deviation of digital twins, low fault identification accuracy, difficulty in identifying multiple coupled faults, lack of edge collaboration, and inaccurate prediction of remaining life.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for fault diagnosis and life prediction of tunnel boring machine (TBM) components, comprising: receiving and storing multi-source time-series data, equipment files, and historical maintenance records through a cloud platform big data center, and issuing model threshold update instructions, sensor configuration instructions, and diagnostic task instructions; wherein the multi-source time-series data includes preprocessed feature vectors, alarm logs, and sensor health codes; constructing a digital twin in the cloud platform big data center, and performing virtual-real mapping and dynamic calibration on the TBM components; wherein the digital twin includes a parametric geometric model, a multibody dynamics model, a fault injection model, and a virtual-real calibration submodule; Deep time-series diagnostics and predictions are performed in the cloud platform big data center. The deep time-series diagnostics and predictions are based on an encoder-decoder network. The encoder of the encoder-decoder network includes a bidirectional long short-term memory network, and independent weights are assigned to the residual features of the digital twin through an attention layer. A remote interactive terminal is connected to the cloud platform big data center for communication, and the status of the digital twin, diagnostic reports, and RUL prediction curves are visualized. The remote interactive terminal integrates a visualization knowledge graph module.

[0007] As a preferred embodiment of the method for fault diagnosis and life prediction of tunnel boring machine components according to the present invention, the model threshold update instruction is set to be issued regularly every month, and when the false alarm rate of a certain type of fault exceeds the preset false alarm rate threshold, it is issued immediately. The sensor configuration command includes sampling frequency and filtering parameters, and is sent out immediately when the sensor data signal-to-noise ratio is lower than the signal-to-noise ratio threshold. The diagnostic task instruction includes key monitoring components and diagnostic frequency, and is issued when the edge system triggers either a level-two early warning or an abnormal rate of decline in remaining service life.

[0008] As a preferred embodiment of the method for fault diagnosis and life prediction of tunnel boring machine components according to the present invention, the parametric geometric model is constructed based on ANSYS Mechanical; The multibody dynamics model is driven by real-time operating conditions. The simulation step size is set according to the sensor sampling frequency requirements, and the simulation time per cycle is set according to the real-time requirements of the project. The fault injection model is used to simulate single-fault scenarios and multi-fault coupled scenarios. The single-fault scenario includes increased clearance of the main drive bearing, pitting of gear teeth, and internal leakage of the hydraulic cylinder. The multi-fault coupling scenario refers to the combination of increased clearance of the main drive bearing and pitting of gear teeth. The virtual-real calibration submodule extracts the measured feature vector and the simulated feature vector according to a preset calibration cycle, and calculates the residual and the mean residual. When the residual mean exceeds the residual safety threshold, historical normal operation data with a similarity to the current operating condition not lower than the operating condition similarity threshold are matched from the historical database. The operating condition similarity is calculated by cosine similarity. The friction coefficient and stiffness coefficient of the multibody dynamics model are corrected based on historical normal operation data, and the residual calculation and coefficient correction are repeated until the mean residual value meets the preset stability condition. The friction coefficient, stiffness coefficient and residual change curves at this time are recorded and stored in the cloud platform big data center.

[0009] As a preferred embodiment of the method for fault diagnosis and life prediction of tunnel boring machine components according to the present invention, the encoder of the encoder-decoder network adopts a bidirectional long short-term memory network and performs bidirectional encoding on the input temporal feature sequence to extract deep spatiotemporal features. The attention layer assigns independent weights to the residual features of the digital twin, and the sum of the weights of other features is: And satisfy the normalization constraint; The normalization constraint refers to the independent weights of the residual features in the attention layer. The sum of them is one; When the residual mean exceeds the residual safety threshold, the value of the independent weight of the residual feature in the attention layer is dynamically increased. The decoder output of the encoder-decoder network includes fault classification results and remaining useful life prediction sequences, and the accuracy and prediction error are calculated based on the fault classification results and remaining useful life prediction sequences.

[0010] In a preferred embodiment of the method for fault diagnosis and life prediction of tunnel boring machine components according to the present invention, the input feature vector for depth-time series diagnosis and prediction is generated through a multi-source feature fusion strategy, and the expression is: ; ; in, This represents the input feature vector; This represents the preprocessed multidimensional features uploaded by the edge system, including the frequency domain features of the vibration signal and the temperature statistical features; This indicates the real-time operating parameters of the tunnel boring machine, including tunneling speed, cutterhead torque, and total thrust. Represents the characteristics of the simulation-measured residuals; This represents the residual between different measured eigenvectors and simulated eigenvectors.

[0011] As a preferred embodiment of the method for fault diagnosis and life prediction of tunnel boring machine components according to the present invention, the encoder-decoder network includes an encoding stage, an attention-weighted stage, and a decoding stage.

[0012] In a preferred embodiment of the method for fault diagnosis and life prediction of tunnel boring machine components according to the present invention, the coding stage has the following workflow: The input feature vectors are arranged according to a fixed time step to form an input time series feature sequence; The encoder encodes the input temporal feature sequence bidirectionally through a bidirectional long short-term memory network, and outputs the hidden state at each fixed time step to form a hidden state sequence. The workflow for the attention-weighted phase is as follows: The attention layer connects the encoder and decoder, and computes the decoder's first... The temporal attention weights for each prediction time step are generated by independent weights of the residual features of the attention layer, and the weights corresponding to the residual features of the digital twin and the attention weights corresponding to each time step satisfy the normalization constraint. The decoding stage has the following workflow. The decoder takes the context vector and its previous output as input and calculates the prediction result at the current time step through the LSTM unit of the Long Short-Term Memory network. The "output of itself at the previous moment" refers to the decoder's output at the previous consecutive time step, denoted as... At what time, the prediction result is output after the calculation is completed; After iterating step by step over time, the probability distribution of health status and the remaining lifespan value are output for a future period of time.

[0013] As a preferred embodiment of the method for fault diagnosis and life prediction of tunnel boring machine components according to the present invention, the encoder-decoder network is trained through a phased strategy, including a first phase and a second phase. The first stage follows the process as follows: Using a large amount of historical data and corresponding remaining useful life labels, a remaining useful life prediction branch is pre-trained with mean squared error as the loss function. The expression for the mean squared error loss function is as follows: ; in, Indicates mean square error; Indicates the total number of training samples; This represents the index number of the training sample, with a value ranging from 1 to... ; Indicates the first The model predicts the remaining useful life value for each training sample; Indicates the first The actual remaining lifespan value corresponding to each training sample; The second stage follows the procedure below. Using samples labeled with fault types, including samples with multiple coupled faults, the fault classification branch is fine-tuned using cross-entropy as the loss function. The expression for the cross-entropy loss function is as follows: ; in, This represents the cross-entropy loss value; Represented as the number of fault categories; Indicates the first The true labels of each training sample; For the first Each training sample belongs to the fault category. The predicted probability; The index number representing the fault category, with a value ranging from 1 to... This is used to iterate through all fault categories.

[0014] As a preferred embodiment of the method for fault diagnosis and life prediction of tunnel boring machine components according to the present invention, a potential anomaly early warning mechanism is added to the virtual-real calibration submodule; The aforementioned potential anomaly early warning mechanism refers to the mechanism that when the residual characteristics of the digital twin continuously exceed the residual safety threshold, but the measured data uploaded by the edge system does not trigger an alarm, i.e. the fault is in the early latent stage, the cloud platform big data center generates a potential anomaly prompt and pushes it through the remote interactive terminal, and issues sensor configuration instructions to the edge system to increase the sampling frequency of the corresponding component sensors.

[0015] Secondly, this invention provides a system for fault diagnosis and life prediction of tunnel boring machine (TBM) components, comprising a data receiving module, a digital twin construction module, a deep diagnosis and prediction module, and an interactive display module. The data receiving module receives and stores multi-source time-series data, equipment files, and historical maintenance records through a cloud platform big data center, and issues model threshold update instructions, sensor configuration instructions, and diagnostic task instructions. The multi-source time-series data includes preprocessed feature vectors, alarm logs, and sensor health codes. The digital twin construction module constructs a digital twin in the cloud platform big data center and performs virtual-real mapping and dynamic calibration on the TBM components. The digital twin... The system includes a parametric geometric model, a multibody dynamics model, a fault injection model, and a virtual-real calibration submodule; a deep diagnostic prediction module for performing deep temporal diagnostics and predictions in the cloud platform big data center, wherein the deep temporal diagnostics and predictions are based on an encoder-decoder network, the encoder of which includes a bidirectional long short-term memory network, and an attention layer assigns independent weights to the residual features of the digital twin; and an interactive display module for communicating between a remote interactive terminal and the cloud platform big data center, and visually displaying the status of the digital twin, diagnostic reports, and RUL prediction curves, wherein the remote interactive terminal integrates a visualization knowledge graph module.

[0016] The beneficial effects of this invention are as follows: Real-time calibration of the digital twin and the physical entity ensures the reliability of the model mechanism; the fusion of attention mechanism and physical mechanism improves the sensitivity and classification accuracy for early and complex faults, effectively covering complex working conditions; the constructed life prediction model combines mechanistic robustness and data adaptability, with small prediction errors, providing accurate basis for maintenance decisions; relying on an efficient edge-cloud collaborative closed loop, it realizes dynamic optimization of monitoring strategies and global sharing of diagnostic knowledge, solving the problem of information gap between the field and remote areas; the overall system responds quickly, has reliable indicators, and is directly adaptable to existing tunnel boring machine equipment based on engineering field verification, possessing good field applicability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for fault diagnosis and life prediction of tunnel boring machine components.

[0019] Figure 2 A schematic diagram of a system for fault diagnosis and life prediction of tunnel boring machine components.

[0020] Figure 3 This is a diagram showing the overall system architecture and its relationship with edge systems.

[0021] Figure 4 This is a flowchart illustrating the composition and virtual-to-real calibration process of a digital twin.

[0022] Figure 5 A diagram of the encoder-decoder network architecture enhanced with attention mechanism. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0026] Reference Figures 1-5 As one embodiment of the present invention, this embodiment provides a method for fault diagnosis and life prediction of tunnel boring machine components, including the following steps: S1. Receives and stores multi-source time-series data, equipment files and historical maintenance records through the cloud platform big data center, and issues model threshold update instructions, sensor configuration instructions and diagnostic task instructions. The multi-source time-series data includes preprocessed feature vectors, alarm logs and sensor health codes.

[0027] It should be noted that this embodiment of the invention is based on the main drive bearing of a 6.3-meter diameter earth pressure balance shield tunneling machine. The shield tunneling machine is used in a subway tunnel project in a certain city. The tunneling diameter is 6.3m, the current tunneling distance is 1200m, the geological conditions are clay layer, the real-time operating parameters are 30mm / min, the total thrust is 18000kN, and the cutterhead torque is 3800kN·m. This embodiment aims to achieve fault diagnosis and remaining service life (RUL) prediction of the main drive bearing through the system and method of this invention, so as to avoid unplanned downtime.

[0028] The edge system to which this invention pertains, namely the shield tunneling machine edge computing system, is a local computing node deployed at the shield tunneling machine's operating site, mainly composed of a sensor network, an edge computing gateway, and a local execution terminal.

[0029] The sensor network collects multi-source time-series data such as vibration, temperature, and pressure from key components of the tunnel boring machine; the edge computing gateway performs real-time preprocessing on the collected data, extracts frequency domain features and statistical features, and uploads them to the cloud platform's big data center at a set frequency, while also receiving various instructions issued by the cloud platform; the local execution terminal realizes functions such as local data storage, real-time alarm triggering, and instruction execution feedback. It is a key hub connecting physical devices and the cloud platform, undertaking the responsibilities of "real-time monitoring + data preprocessing + instruction execution", and forming a collaborative closed loop with the cloud platform of "real-time edge response + deep cloud analysis".

[0030] The edge system uploads multi-source timing data of the main drive bearing at a frequency of 5 minutes / time: vibration signal peak frequency 320Hz, root mean square frequency 150Hz, average temperature 78℃, variance 2.3℃, sensor health code is "normal", and there are no alarm logs.

[0031] The tunnel boring machine's PLC uploads real-time operating parameters Creal=[30mm / min,3800kN·m,18000kN], which are synchronously stored in the equipment archive by the cloud platform's big data center. This is linked to the tunnel boring machine's historical maintenance records, and there are no records of main drive bearing failures in the past 6 months.

[0032] It receives multi-source time-series data uploaded by the edge system, including vibration signal frequency domain characteristics: peak frequency and root mean square frequency; temperature statistical characteristics: mean and variance; alarm logs; and sensor health codes. Under normal operating conditions, it uploads data once every 5 minutes, and pushes data in real time under early warning conditions. It also receives real-time operating parameters uploaded by the tunnel boring machine PLC.

[0033] The model threshold update command is set to be issued monthly, and immediately when the false alarm rate for a certain type of fault exceeds 1%; the monthly issuance is scheduled for the 1st of each month, and the expression for calculating the false alarm rate is as follows. ; in, Indicates the false alarm rate; This indicates the number of false alarms, specifically the number of times a normal state was misdiagnosed as a fault. This indicates the total number of diagnoses, which refers to the total number of fault judgments within the statistical period.

[0034] The sensor configuration command includes sampling frequency and filtering parameters. It is sent immediately when the sensor data signal-to-noise ratio (SNR) falls below the SNR threshold. The initial sampling frequency is specified in the command. for Low-pass filter cutoff frequency for A signal-to-noise ratio (SNR) threshold of 20 dB is set. When the SNR falls below this threshold, the sampling frequency is immediately adjusted to 32 kHz. The expression for calculating the SNR is: ; in, Indicates the signal-to-noise ratio; Indicates the effective signal power; This indicates the background noise power.

[0035] The diagnostic task command includes the key monitoring components and the diagnostic frequency. It is issued when the edge system triggers either a level 2 warning or an abnormal rate of remaining useful life (RUL) decline. The default diagnostic frequency is specified in the command. The frequency is set to 1 time / hour, with the main drive bearing being the key monitoring component. When the edge system triggers a level 2 warning or the RUL decrease rate is abnormal, the diagnostic frequency is adjusted to 2 times / hour. The expression for calculating the RUL decrease rate is: ; in, Indicates the rate of RUL descent; , This indicates that the abnormal rate of decrease of RUL between two adjacent prediction times exceeds the historical normal fluctuation range; express The predicted remaining useful life at that moment; express The predicted remaining useful life at that time.

[0036] The historical normal fluctuation range refers to the normal distribution range of the RUL descent rate obtained by statistically analyzing the full life cycle historical data of key components of the tunnel boring machine (main drive bearing in this scheme) under similar working conditions. Specifically, it is determined as follows: 50 sets of full life cycle historical data of the same model and working conditions of the main drive bearing are selected, and the RUL descent rate at adjacent prediction times in each set of data is calculated. Historical values, using After removing outlier historical values, the remaining values ​​are selected. The minimum and maximum historical values ​​constitute the historical normal fluctuation range. , When calculated in real time At that time, it was determined that the RUL descent rate was abnormal.

[0037] This invention sets up a three-level early warning mechanism to distinguish the severity of the fault and the priority of handling: Level 1 early warning (emergency early warning) indicates that the component has suffered a serious fault or RUL ≤ 48 hours, and the machine needs to be stopped for maintenance immediately. Level 2 warning (potential fault warning) refers to the residual characteristics of the digital twin exceeding the safety threshold and the operating parameters of the component deviating from the normal range but not reaching the level of an emergency fault, or the RUL being in the range of 48 to 100 hours, requiring increased monitoring frequency and the development of a maintenance plan; Level 3 warning (indicative warning) refers to the slight fluctuations in the operating parameters of the component and the residual characteristics approaching the safety threshold, requiring continuous monitoring but no adjustment to the diagnostic strategy.

[0038] The triggering conditions for Level II early warning include: residual of a single core feature of the digital twin >4%, sensor data exceeding the normal fluctuation range for three consecutive times, and RUL prediction value entering the range within 100 hours for the first time. Specific triggering thresholds can be optimized through training with historical fault data.

[0039] If the edge system detects that the peak vibration frequency of the main drive bearing exceeds the normal range of 300~350Hz three times consecutively, a level two warning is triggered, or the RUL decrease rate is detected. When the fluctuations exceed the historical normal range, the cloud platform will adjust the diagnostic frequency from once per hour to twice per hour to achieve accurate tracking of the faulty parts.

[0040] S2. Construct a digital twin in the cloud platform big data center and perform virtual-real mapping and dynamic calibration on the tunnel boring machine components; the digital twin includes a parametric geometric model, a multibody dynamics model, a fault injection model, and a virtual-real calibration submodule.

[0041] It should be noted that the parametric geometric model is based on ANSYS Mechanical to construct the parametric geometric model of the main drive bearing. The number of meshes is set to 1.2 million, and the mesh type is a hexahedral structured mesh to ensure a balance between simulation accuracy and computational efficiency.

[0042] The multibody dynamics model is driven by real-time operating conditions. The simulation step size is set according to the sensor sampling frequency requirements, and the simulation time per cycle is set according to the real-time requirements of the project. The simulation step size is set to 0.01 seconds to match the sensor sampling frequency. Input real-time operating condition parameters. The driving model runs, and the time for a single simulation is controlled within 7.5 minutes, which meets the real-time requirements of engineering. The fault injection model is used to simulate single-fault scenarios and multi-fault coupled scenarios. The single-fault scenario includes increased clearance of the main drive bearing, pitting of the gear tooth surface, and internal leakage of the hydraulic cylinder. The multi-fault coupled scenario refers to the combination of increased clearance of the main drive bearing and pitting of the gear tooth surface. Specifically, the simulated single-fault scenario is a clearance increase of 0.3 mm, which is in the typical wear range of 0.1 mm to 0.5 mm. The multi-fault coupled scenario is a clearance increase of 0.3 mm and a gear tooth surface pitting area accounting for 15%. The model outputs virtual sensor vibration and temperature response curves.

[0043] The virtual-real calibration submodule extracts the measured feature vector and the simulated feature vector according to a preset calibration cycle, and calculates the residuals and the mean of the residuals; the preset calibration cycle is once per hour, and the expression for calculating the residuals is as follows: ; in, Represents the residual of the j-th core feature; Indicates the first Measured values ​​of the core features; Indicates the first Simulated values ​​of the core features; Indicates the index number of the core feature.

[0044] The expression for calculating the mean of the residuals is: ; in, This represents the mean of the residuals; This represents the total number of core features.

[0045] When the mean residual exceeds the residual safety threshold (5%), historical normal operation data with a similarity to the current operating condition not lower than the operating condition similarity threshold (90%) is matched from the historical database. To eliminate the difference in the dimensions of parameters under different operating conditions, each dimension of the operating condition feature vector is normalized (e.g., min-max or z-score) to obtain a normalized vector. The operating condition similarity is calculated using cosine similarity, expressed as follows: ; in, This represents the similarity between the current working condition and the historical working condition, with a value range of [0,1]. The closer the value is to 1, the higher the similarity. This represents the feature vector of the current working condition, which includes working condition parameters such as the current advance speed, tunneling pressure, geological hardness, and rotation speed. This represents the historical operating condition feature vector, derived from operating condition parameter records under normal operating conditions in the historical database. Its dimension is... Consistent.

[0046] The friction coefficient and stiffness coefficient of the multibody dynamics model are corrected based on historical normal operation data. The expressions for the corrected friction coefficient and stiffness coefficient are as follows: ; ; in, This represents the corrected friction coefficient, used to update the friction parameters in the multibody dynamics model. This represents the friction coefficient before correction, i.e., the initial value of the friction coefficient currently used in the multibody dynamics model or the value after the last calibration; This represents the friction coefficient correction factor. This represents the stiffness coefficient correction factor, with a value ranging from 0.05 to 0.2. This represents the corrected stiffness coefficient, used to update the stiffness parameters in the multibody dynamics model; This represents the stiffness coefficient before correction, i.e., the initial value of the stiffness coefficient currently used in the multibody dynamics model or the value after the last calibration.

[0047] Repeat the residual calculation and coefficient correction until the mean residual value is stable within 3%. Record the friction coefficient, stiffness coefficient and residual change curve at this time and store them in the cloud platform big data center.

[0048] The residual variation curve is a continuous curve plotted with the number of iterations on the x-axis and the mean residual value corresponding to the number of iterations on the y-axis. It is used to visually reflect the decay trend of the mean residual value during the iterative calibration process and to verify the effectiveness and convergence of the calibration process.

[0049] Digital twins with As the driving condition, run a multibody dynamics simulation and output simulation vectors of 15 core features. The simulated peak vibration frequency is 315Hz, and the simulated average temperature is 75℃.

[0050] S3. Perform deep time series diagnosis and prediction in the cloud platform big data center. The deep time series diagnosis and prediction is based on an encoder-decoder network. The encoder of the encoder-decoder network includes a bidirectional long short-term memory network and assigns independent weights to the residual features of the digital twin through an attention layer.

[0051] It should be noted that the encoder uses a bidirectional long short-term memory network with 128 hidden layer units and a time step T of 60 (corresponding to 5 hours of time-series data, with one data point every 5 minutes); the residual features in the attention layer have independent weights. for The sum of the weights of other features is and satisfy and The sum of the values ​​is 1, and the residual mean dynamically increases to 0.4 when it exceeds 5%. The decoder uses a single hidden layer LSTM with an output dimension of 2, which is the fault classification probability and RUL prediction value. The encoder performs bidirectional encoding on the input temporal feature sequence to extract deep spatiotemporal features.

[0052] The decoder output includes fault classification results and RUL prediction sequences. Based on these results, it calculates accuracy and prediction error. The expression for calculating accuracy is: ; in, Indicates the accuracy of fault classification; This represents the number of correctly classified samples, that is, the total number of samples whose fault classification results match the actual fault category of the samples. This represents the total number of samples participating in the performance evaluation, which is consistent with the total number of labeled samples used to train the cross-entropy loss function mentioned above.

[0053] RUL prediction series refers to the RUL prediction series for the next 300 hours, and it meets the classification requirements. When the RUL prediction series has a RUL > 200h, the prediction error is... If the error is ≤±3.75%, the expression for calculating the prediction error is as follows: ; in, Indicates prediction error; This represents the predicted remaining useful life at a given prediction point in time. It represents the actual remaining service life under the actual operating conditions at a given predicted time. Actual remaining useful life at a given predicted time Determined from equipment files and historical maintenance / replacement records: For samples that experienced malfunctions or underwent replacement / repair, the time of malfunction confirmation or the time of completion of replacement / repair is recorded as follows: (Unit: h), for any observation time The actual remaining lifespan is marked as follows: ; Samples that have not yet failed to malfunction and have not been replaced or repaired are considered truncated samples, with the data acquisition termination time used as the cutoff point. This is a right-hand truncation point, used only during training / evaluation. By eliminating sample sets that require precise labeling, the constraints are mitigated, thereby ensuring that the true source of RUL values ​​is clear and traceable.

[0054] When RUL≤100h, the error If the error is ≤±2.1%, the expression for calculating the prediction error is: ; The input feature vector for deep temporal diagnostics and prediction is generated through a multi-source feature fusion strategy, expressed as follows: ; ; in, This represents the input feature vector, with dimensions of 26×60; The preprocessed multidimensional features uploaded by the edge system include vibration signal frequency domain features and temperature statistical features, with a dimension of 8×60; This indicates the real-time operating parameters of the tunnel boring machine, including tunneling speed, cutterhead torque, and total thrust, with a dimension of 3×60. This represents the simulation-measured residual characteristics, with a dimension of 15×60; This represents the residual between different measured eigenvectors and simulated eigenvectors.

[0055] The workflow of an encoder-decoder network includes an encoding phase, an attention-weighted phase, and a decoding phase.

[0056] The workflow during the coding phase is as follows: Input feature vector Arranged according to a fixed time step, forming the input time series feature sequence. , .

[0057] The encoder uses a bidirectional long short-term memory network to bidirectionally encode the input temporal feature sequence, outputting the hidden state at each fixed time step, expressed as follows: ; in, Indicates a hidden state; This represents a bidirectional long short-term memory network; The input feature vector at time t is composed of the fusion of the key component operation features, operating condition parameter features, and residual features output by the digital twin after preprocessing of the edge system. express The hidden state vector of the Bi-LSTM at time step 1 is the core feature representation result of the previous time step, which is used to convey temporal dependency information and realize the temporal coherence of features. The time index represents the time step position of the current feature input and hidden state update.

[0058] The hidden states are arranged to form a hidden state sequence. .

[0059] The workflow of the attention-weighted phase is as follows: The attention layer connects the encoder and decoder, and computes the decoder's first... Time series focus weights for each prediction time step The expression for weight calculation is, ; ; in, Indicates attention weight; Indicates the decoder's first Step into hidden state; The variable representing the traversal index of the encoder's hidden state sequence; This represents the total time length of the encoder input time sequence, i.e., the length of the hidden state sequence; This represents the hidden state of the encoder at step k; This represents the dot product function, used to quantify the correlation between temporal features and the prediction target; This indicates the decoding step index of the decoder.

[0060] To ensure dimensionality consistency in context vector computation, the simulation-measured residual features are aligned into a residual sequence by time steps. The encoder outputs the hidden state at each time step. , The hidden state dimension is 128 in this invention. Since the simulation-measured residual characteristics differ from the hidden state dimension, a linear mapping is first performed on the residual vectors in the residual sequence to obtain the hidden state dimension. ; in, Let represent the residual vector after the mapping at step k; This represents the original residual vector at step k; Represents the residual mapping weight matrix; This represents the bias vector.

[0061] Further aggregation of the mapped residual sequences yields the residual representation vector. ; in, This represents the residual representation vector.

[0062] Assign independent weights to the residual features of the digital twin Generate a context vector; the expression for generating the context vector is: ; in, Represents the context vector; Indicates the independent weights of the residual features in the attention layer; This represents the attention weight at step k.

[0063] The weights corresponding to the residual features of the digital twin and the attention weights corresponding to each time step satisfy the normalization constraint, i.e. ,and , The normalization constraint means that the sum of the feature weights of the digital twin parameters and the attention weights corresponding to each time step is one.

[0064] The workflow of the decoding stage is as follows: The decoder uses context vectors and its output in the previous moment As input, the prediction result at the current time step is calculated using the LSTM unit of the bidirectional long short-term memory network, expressed as follows: ; in, Indicates the decoder's first The output vector of the step; Indicates the first The output vector of the step.

[0065] After iterating step by step over time, output the probability distribution of health status over a future period. With remaining useful life (RUL) value.

[0066] Health status probability distribution The decoder extracts temporal features step-by-step through bidirectional LSTM units for a preset future time window. The extracted features are then normalized by a Softmax layer to output the probability values ​​of key components in a healthy state, a potential fault state, and a fault state at each time step, forming a continuous probability distribution curve of future health states. In the probability calculation process at each time step, the "output probability of the previous time step and the context vector of the current time step" are used as joint inputs to ensure the continuity and traceability of the health state evolution sequence.

[0067] The Remaining Usage Limit (RUL) value, with a preset fault determination threshold of ≥0.5, is used to locate the first time point that meets the fault determination threshold by traversing the probability distribution curve of future health states. Based on the current prediction time ,calculate and The difference is used to obtain the initial RUL value. To improve prediction accuracy, the simulation results of the digital twin fault injection model are introduced to weight and correct the initial RUL value. The deep learning model prediction result has a weight of 0.7, and the digital twin simulation result has a weight of 0.3. After correction, the accurate RUL value is output. The remaining useful life RUL value does not refer to the previous RUL prediction sequence. The RUL prediction sequence is "an ordered set of RUL prediction values ​​at each time step in the future" (a set of time series data). The final output RUL value is the result of processing the RUL prediction sequence (which can be a single value or an interval value). The RUL prediction sequence is the core data support for the output RUL value.

[0068] The encoder-decoder network is trained using a phased strategy, which includes a first phase and a second phase. The first phase process is as follows. Using a large amount of historical data and corresponding RUL labels, the RUL prediction branch is pre-trained with mean squared error as the loss function. The expression for the mean squared error loss function is as follows: ; in, Indicates mean square error; Indicates the total number of training samples; This represents the index number of the training sample, with a value ranging from 1 to... ; Indicates the first The model predicts the remaining useful life value for each training sample; Indicates the first The actual remaining lifespan value corresponding to each training sample.

[0069] The first stage of pre-training the RUL prediction branch, with a certain number of training samples. =5000, the mean squared error (MSE) loss function convergence threshold is 10. 4 .

[0070] The second phase process is as follows: Using samples labeled with fault types, including samples with multiple coupled faults, the fault classification branch is fine-tuned using cross-entropy as the loss function. The expression for the cross-entropy loss function is as follows: ; in, This represents the cross-entropy loss value; Represented as the number of fault categories; Indicates the first The true labels of each training sample; For the first Each training sample belongs to the fault category. The predicted probability; The index number representing the fault category, with a value ranging from 1 to... This is used to iterate through all fault categories.

[0071] The second stage involves fine-tuning the fault classification branches and the number of fault categories. =4, representing normal, increased clearance, pitting, and coupling faults, cross-entropy. The convergence threshold of the loss function is .

[0072] S4. Connect the remote interactive terminal to the cloud platform big data center and visualize the status, diagnostic reports, and RUL prediction curves of the digital twin. The remote interactive terminal integrates a visualization knowledge graph module.

[0073] It should be noted that the remote interactive terminal integrates a visual knowledge graph module, which associates the four types of failure modes, root causes, affected components, and recommended maintenance measures of the main drive bearing; it supports displaying the real-time status of the digital twin, fault diagnosis reports, RUL prediction curves, and edge-cloud collaborative logs.

[0074] A potential anomaly early warning mechanism is added to the virtual-real calibration submodule. The potential anomaly early warning mechanism means that when the residual characteristics of the digital twin continuously exceed the residual safety threshold, but the measured data uploaded by the edge system does not trigger an alarm, that is, when the fault is in the early latent stage, the cloud platform big data center generates a potential anomaly prompt and pushes it through the remote interactive terminal, and issues sensor configuration instructions to the edge system to increase the sampling frequency of the corresponding component sensors.

[0075] This embodiment also provides a system for fault diagnosis and life prediction of tunnel boring machine components, including: a data receiving module, a digital twin construction module, a deep diagnosis and prediction module, and an interactive display module; The data receiving module is used to receive and store multi-source time-series data, equipment files and historical maintenance records through the cloud platform big data center, and issue model threshold update instructions, sensor configuration instructions and diagnostic task instructions. The multi-source time-series data includes preprocessed feature vectors, alarm logs and sensor health codes. The digital twin construction module is used to build a digital twin in the cloud platform big data center and perform virtual-real mapping and dynamic calibration on the components of the tunnel boring machine; the digital twin includes a parametric geometric model, a multibody dynamics model, a fault injection model and a virtual-real calibration submodule; The deep diagnostic prediction module is used to perform deep temporal diagnostics and predictions in the cloud platform big data center. The deep temporal diagnostics and predictions are based on an encoder-decoder network. The encoder of the encoder-decoder network includes a bidirectional long short-term memory network and assigns independent weights to the residual features of the digital twin through an attention layer. The interactive display module is used to connect the remote interactive terminal with the cloud platform big data center and visualize the status, diagnostic reports and RUL prediction curves of the digital twin. The remote interactive terminal integrates a visualization knowledge graph module.

[0076] In summary, this invention achieves the following: Real-time calibration of the digital twin and the physical entity ensures the reliability of the model mechanism; the fusion of attention mechanisms and physical mechanisms enhances the sensitivity and classification accuracy for early and complex faults, effectively covering complex operating conditions; the constructed life prediction model possesses both mechanistic robustness and data adaptability, with small prediction errors, providing accurate basis for maintenance decisions; relying on an efficient edge-cloud collaborative closed loop, it realizes dynamic optimization of monitoring strategies and global sharing of diagnostic knowledge, solving the information gap between the field and remote locations; the overall system responds rapidly, has reliable indicators, and is directly adaptable to existing tunnel boring machine equipment based on engineering field verification, demonstrating good field applicability.

[0077] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for fault diagnosis and life prediction of tunnel boring machine components, characterized in that: include, The cloud platform big data center receives and stores multi-source time-series data, equipment files and historical maintenance records, and issues model threshold update instructions, sensor configuration instructions and diagnostic task instructions. The multi-source time-series data includes preprocessed feature vectors, alarm logs and sensor health codes. A digital twin is constructed in the cloud platform big data center, and virtual-real mapping and dynamic calibration are performed on the components of the tunnel boring machine; the digital twin includes a parametric geometric model, a multibody dynamics model, a fault injection model, and a virtual-real calibration submodule; Deep time-series diagnostics and predictions are performed in the cloud platform big data center. The deep time-series diagnostics and predictions are based on an encoder-decoder network. The encoder of the encoder-decoder network includes a bidirectional long short-term memory network and assigns independent weights to the residual features of the digital twin through an attention layer. The remote interactive terminal communicates with the cloud platform's big data center and visualizes the status, diagnostic reports, and RUL prediction curves of the digital twin. The remote interactive terminal integrates a visualization knowledge graph module.

2. The method for fault diagnosis and life prediction of tunnel boring machine components as described in claim 1, characterized in that: The model threshold update instruction is set to be issued regularly every month, and when the false alarm rate of a certain type of fault exceeds the preset false alarm rate threshold, it will be issued immediately. The sensor configuration command includes sampling frequency and filtering parameters, and is sent out immediately when the sensor data signal-to-noise ratio is lower than the signal-to-noise ratio threshold. The diagnostic task instruction includes key monitoring components and diagnostic frequency, and is issued when the edge system triggers either a level-two early warning or an abnormal rate of decline in remaining service life.

3. The method for fault diagnosis and life prediction of tunnel boring machine components as described in claim 1, characterized in that: The parametric geometric model is constructed based on ANSYS Mechanical; The multibody dynamics model is driven by real-time operating conditions. The simulation step size is set according to the sensor sampling frequency requirements, and the simulation time per cycle is set according to the real-time requirements of the project. The fault injection model is used to simulate single-fault scenarios and multi-fault coupled scenarios. The single-fault scenario includes increased clearance of the main drive bearing, pitting of gear teeth, and internal leakage of the hydraulic cylinder. The multi-fault coupling scenario refers to the combination of increased clearance of the main drive bearing and pitting of gear teeth. The virtual-real calibration submodule extracts the measured feature vector and the simulated feature vector according to a preset calibration cycle, and calculates the residual and the mean residual. When the residual mean exceeds the residual safety threshold, historical normal operation data with a similarity to the current operating condition not lower than the operating condition similarity threshold are matched from the historical database. The operating condition similarity is calculated by cosine similarity. The friction coefficient and stiffness coefficient of the multibody dynamics model are corrected based on historical normal operation data, and the residual calculation and coefficient correction are repeated until the mean residual value meets the preset stability condition. The friction coefficient, stiffness coefficient and residual change curves at this time are recorded and stored in the cloud platform big data center.

4. The method for fault diagnosis and life prediction of tunnel boring machine components as described in claim 1, characterized in that: The encoder-decoder network employs a bidirectional long short-term memory network for its encoder, and performs bidirectional encoding on the input temporal feature sequence to extract deep spatiotemporal features. The attention layer assigns independent weights to the residual features of the digital twin, and the sum of the weights of other features is: And satisfy the normalization constraint; The normalization constraint refers to the independent weights of the residual features in the attention layer. The sum of them is one; When the residual mean exceeds the residual safety threshold, the value of the independent weight of the residual feature in the attention layer is dynamically increased. The decoder output of the encoder-decoder network includes fault classification results and remaining useful life prediction sequences, and the accuracy and prediction error are calculated based on the fault classification results and remaining useful life prediction sequences.

5. The method for fault diagnosis and life prediction of tunnel boring machine components as described in claim 1, characterized in that: The input feature vector for deep temporal diagnosis and prediction is generated through a multi-source feature fusion strategy, expressed as follows: ； ； in, This represents the input feature vector; This represents the preprocessed multidimensional features uploaded by the edge system, including the frequency domain features of the vibration signal and the temperature statistical features; This indicates the real-time operating parameters of the tunnel boring machine, including tunneling speed, cutterhead torque, and total thrust. Represents the characteristics of the simulation-measured residuals; This represents the residual between different measured eigenvectors and simulated eigenvectors.

6. The method for fault diagnosis and life prediction of tunnel boring machine components as described in claim 1, characterized in that: The encoder-decoder network's workflow includes an encoding phase, an attention-weighted phase, and a decoding phase.

7. The method for fault diagnosis and life prediction of tunnel boring machine components as described in claim 6, characterized in that: The coding phase follows this workflow. The input feature vectors are arranged according to a fixed time step to form an input time series feature sequence; The encoder encodes the input temporal feature sequence bidirectionally through a bidirectional long short-term memory network, and outputs the hidden state at each fixed time step to form a hidden state sequence. The workflow for the attention-weighted phase is as follows: The attention layer connects the encoder and decoder, and computes the decoder's first... The temporal attention weights for each prediction time step are generated by independent weights of the residual features of the attention layer, and the weights corresponding to the residual features of the digital twin and the attention weights corresponding to each time step satisfy the normalization constraint. The decoding stage has the following workflow. The decoder takes the context vector and its previous output as input and calculates the prediction result at the current time step through the LSTM unit of the Long Short-Term Memory network. The "output of itself at the previous moment" refers to the decoder's output at the previous consecutive time step, denoted as... At what time, the prediction result is output after the calculation is completed; After iterating step by step over time, the probability distribution of health status and the remaining lifespan value are output for a future period of time.

8. The method for fault diagnosis and life prediction of tunnel boring machine components as described in claim 1, characterized in that: The encoder-decoder network is trained using a phased strategy, including a first phase and a second phase. The first stage follows the process as follows: Using a large amount of historical data and corresponding remaining useful life labels, a remaining useful life prediction branch is pre-trained with mean squared error as the loss function. The expression for the mean squared error loss function is as follows: ； in, Indicates mean square error; Indicates the total number of training samples; This represents the index number of the training sample, with a value ranging from 1 to... ; Indicates the first The model predicts the remaining useful life value for each training sample; Indicates the first The actual remaining lifespan value corresponding to each training sample; The second stage follows the procedure below. Using samples labeled with fault types, including samples with multiple coupled faults, the fault classification branch is fine-tuned using cross-entropy as the loss function. The expression for the cross-entropy loss function is as follows: ； in, This represents the cross-entropy loss value; Represented as the number of fault categories; Indicates the first The true labels of each training sample; For the first Each training sample belongs to the fault category. The predicted probability; The index number representing the fault category, with a value ranging from 1 to... This is used to iterate through all fault categories.

9. The method for fault diagnosis and life prediction of tunnel boring machine components as described in claim 3, characterized in that: A potential anomaly warning mechanism is added to the virtual-real calibration submodule; The aforementioned potential anomaly early warning mechanism refers to the mechanism that when the residual characteristics of the digital twin continuously exceed the residual safety threshold, but the measured data uploaded by the edge system does not trigger an alarm, i.e. the fault is in the early latent stage, the cloud platform big data center generates a potential anomaly prompt and pushes it through the remote interactive terminal, and issues sensor configuration instructions to the edge system to increase the sampling frequency of the corresponding component sensors.

10. A system for fault diagnosis and life prediction of tunnel boring machine components, based on the method for fault diagnosis and life prediction of tunnel boring machine components according to any one of claims 1 to 9, characterized in that: It includes a data receiving module, a digital twin construction module, a deep diagnostic prediction module, and an interactive display module; The data receiving module is used to receive and store multi-source time-series data, equipment files and historical maintenance records through the cloud platform big data center, and issue model threshold update instructions, sensor configuration instructions and diagnostic task instructions. The multi-source time-series data includes preprocessed feature vectors, alarm logs and sensor health codes. The digital twin construction module is used to build a digital twin in the cloud platform big data center and perform virtual-real mapping and dynamic calibration on the components of the tunnel boring machine; the digital twin includes a parametric geometric model, a multibody dynamics model, a fault injection model and a virtual-real calibration submodule; The deep diagnostic prediction module is used to perform deep temporal diagnostics and predictions in the cloud platform big data center. The deep temporal diagnostics and predictions are based on an encoder-decoder network. The encoder of the encoder-decoder network includes a bidirectional long short-term memory network and assigns independent weights to the residual features of the digital twin through an attention layer. The interactive display module is used to connect the remote interactive terminal with the cloud platform big data center and visualize the status, diagnostic reports and RUL prediction curves of the digital twin. The remote interactive terminal integrates a visualization knowledge graph module.

Images