Abnormality detection method and system before shutdown of shield tunneling machine

Abstract

The invention belongs to the field of tunnel engineering construction, and provides an anomaly detection method and system before shutdown of a shield tunneling machine, and the method comprises the steps: collecting and preprocessing multi-system sensor data of the shield tunneling machine; performing grouping, feature extraction and dimension reduction processing on the data by adopting a hierarchical hybrid dimension reduction strategy; a multi-scale space-time attention residual error neural network model is constructed and trained; and real-time monitoring and shutdown early warning are carried out on the operation state of the shield tunneling machine. Experiments prove that the accuracy rates of the method on four test sets all exceed 95%, the F1 score is 0.90 or above, and the abnormal state of the shield tunneling machine before shutdown can be accurately predicted. According to the method, engineers can take measures in advance, construction period delay and economic loss caused by sudden shutdown are avoided, the safety and efficiency of shield construction are improved, and the method has important significance in promoting shield tunneling machine construction intelligentization.

Description

Technical Field

[0001] The present invention belongs to the field of tunnel engineering construction, and particularly relates to a method and system for abnormal detection before a shield machine stops, which is a method and system for abnormal detection before a shield machine stops based on hierarchical hybrid dimensionality reduction and multi-scale spatio-temporal attention, and is used to realize the prediction and early warning of unplanned shutdowns of the shield machine. Background Art

[0002] With the acceleration of the global urbanization process and the increasing scarcity of ground space, the demand for the utilization of underground space has increased exponentially. The shield method has become the main construction method for underground infrastructure projects due to its environmental sustainability, high operating efficiency, and little interference to the ground surface. The tunnel boring machine (TBM) is the core equipment in these projects. However, when it operates in a complex geological environment, it faces multiple challenges and requires more advanced monitoring and control strategies.

[0003] During the tunnel boring process, the occurrence of shield machine shutdown events is a key operational issue that seriously affects the construction progress of the project. These interruptions stem from various factors, including mechanical component failures, complex geological conditions, and operating parameter deviations. Statistical analysis of recent large urban construction projects shows that unplanned TBM shutdowns can account for about 30% of the total construction period, which will not only cause huge economic losses and delay the project progress, but these accidental interruptions will also trigger a series of secondary effects, including ground settlement, deformation of adjacent structures, and potential damage to existing underground facilities.

[0004] The traditional TBM maintenance mode mainly relies on regular inspections and pre-determined maintenance plans, and has limitations in dealing with problems of accidental operation interruptions. In addition, the complex interaction mechanism between the TBM and the surrounding geological medium and the shutdown events pose great challenges to fault mode identification and prediction. The dynamic nature of the tunnel environment and the multi-faceted relationship between operating parameters will further exacerbate this complexity.

[0005] In the prior art, the performance prediction of shield machines mainly focuses on cutterhead blockage, attitude prediction, propulsion rate, geological condition assessment, etc., and relatively few studies have been conducted on the prediction of machine shutdowns. Secondly, unplanned shutdowns frequently occur during the tunneling process of the shield, resulting in problems such as construction period delays, economic losses, and increased safety risks. The prior art has not provided an effective solution.

[0006] Therefore, there is an urgent need for a method that can effectively predict unplanned shutdowns of shield machines to reduce the adverse effects caused by shutdowns and improve the tunneling efficiency of shields. Summary of the Invention

[0007] The purpose of the present invention is to provide a shield machine pre-shutdown anomaly detection method and system based on hierarchical hybrid dimensionality reduction and multi-scale spatiotemporal attention, so as to solve the technical problem of low accuracy in the prediction of unplanned shutdown of existing shield machines.

[0008] To achieve the above objectives, the present invention provides the following technical solutions:

[0009] A method for detecting anomalies before a shield machine stops, comprising the following steps:

[0010] Step 1: Collect and pre-process shield machine operation data;

[0011] 1.1) Collecting shield machine multi-system sensor data, including cutterhead speed, cutterhead torque, cutterhead power, cutterhead angle, penetration, chamber soil pressure, screw machine speed, screw machine torque, screw machine pressure, motor frequency, motor current, motor temperature, air flow, foam flow, etc., a total of 131 dimensions of shield machine sensor data;

[0012] 1.2) Valid data from the TBM during tunneling operations was screened and a systematic preprocessing process was implemented. First, the unstable data generated by the TBM restart after a prolonged power outage was identified and removed from the first two tunneling cycles. Furthermore, based on the principle of data authenticity, outliers in the original data collection were retained to ensure data integrity and a true reflection of the operating conditions. Subsequently, multidimensional time series parameters were standardized to effectively eliminate dimensional differences between sensor measurements, ensuring a balanced weight distribution across all features during subsequent model training. This mitigates feature dominance caused by inconsistent numerical scales and improves model generalization and prediction accuracy.

[0013] Step 2: Process the data according to the hierarchical hybrid dimensionality reduction strategy;

[0014] 2.1) Divide the pre-processed sensor data into multiple functional subsystems according to their functions, namely, the tunneling system, attitude system, earth pressure control system, grouting system, power and transmission system, and auxiliary system;

[0015] 2.2) Extract features from each subsystem and construct physical features as follows:

[0016] 2.3) Apply principal component analysis (PCA) to each subsystem for intra-group dimensionality reduction;

[0017] 2.4) Hierarchical hybrid dimensionality reduction through inter-group feature fusion;

[0018] Step 3: Establish a Multi-Scale Spatio-Temporal Attention Residual Network (MSTAR) model;

[0019] Connect them in sequence according to the order of the input layer, multi-scale time encoding layer, enhanced time attention layer, residual connection feature fusion layer, and output layer to construct the MSTAR neural network; the specific details of each layer are as follows:

[0020] Input layer design: Construct an input projection layer, receive the dimensionality-reduced feature vector as input, and perform a linear transformation on the input data;

[0021] Multi-scale time encoding implementation: Design a multi-scale strategy to divide the input sequence according to different time window lengths; for each scale i, construct an encoder that processes the input sequence with different window sizes, that is, a multi-scale time encoder; deploy bidirectional gated recurrent units (BiGRUs) at each scale to capture the front and back data information of the past and future at the same time; process the features extracted from each scale in parallel to obtain dependencies with different time granularities;

[0022] Enhanced time attention layer: Introduce an enhanced time attention module after the multi-scale time encoding layer to implement the multi-head self-attention mechanism and calculate the correlation strength between different time steps; the attention mechanism enhances the model's ability to recognize abnormal patterns before shutdown by dynamically adjusting the attention weights for different time steps;

[0023] Deployment of the residual connection structure: Introduce a residual connection structure after the enhanced time attention layer and inside the feature fusion layer; each residual block consists of two fully connected layers, a batch normalization layer, and a ReLU activation function; by directly adding the input to the output of the network layer, a "shortcut connection" is formed to build an information highway; the residual structure ensures that the gradient can flow smoothly during backpropagation and effectively alleviates the problem of gradient disappearance in deep networks;

[0024] Feature fusion and output layer design: Concatenate and fuse the multi-scale features and the features after residual processing; design the output layer to use the sigmoid activation function to output the probability value of the occurrence of a shutdown anomaly; when the output probability value exceeds the preset threshold, it is determined as an abnormal state where a shutdown is about to occur.

[0025] Through the above design, the MSTAR model can effectively capture the change patterns of the shield machine's operating parameters at different time scales. The arrangement of the residual structure ensures the training stability of the deep network, and the spatio-temporal attention mechanism enhances the ability to recognize abnormal patterns, thereby achieving accurate prediction of the abnormal state before the shield machine shuts down.

[0026] Step 4: Train the model and perform anomaly detection;

[0027] 4.1) Mark the data 15 minutes before the shutdown point as abnormal, and perform the processing of steps 1 and 2 to obtain the dimensionality-reduced data;

[0028] 4.2) Train the model constructed in step 3 using the weighted binary cross - entropy loss function;

[0029] 4.3) Use the trained model to perform anomaly detection on the dimension - reduced data, thereby determining the normal and abnormal state data under the shield tunneling condition.

[0030] The physical feature extraction in step 2 includes:

[0031] Tunneling system features: Extract the cutterhead efficiency (the relationship between cutterhead rotation speed and penetration), energy consumption ratio (the ratio of actual power to theoretical power), working state (angle and torque changes), and time - cumulative features. At the same time, calculate the propulsion damping, cylinder balance, and propulsion attitude angle to comprehensively characterize the tunneling performance.

[0032] Attitude control system features: Analyze the attitude stability from the vertical, horizontal, and rolling dimensions, calculate the trajectory cumulative deviation and change rate, and evaluate the response characteristics of the articulation system to reflect the position control ability of the shield machine.

[0033] Earth pressure control system features: Evaluate the earth pressure state of the tunneling face through pressure distribution, gradient, and fluctuation characteristics, calculate the equilibrium indicators in the horizontal and vertical directions, monitor the operation of the screw conveyor and the parameters of the hydraulic system to prevent the risk of earth pressure out - of - control.

[0034] Grouting system features: Analyze the statistical characteristics of chamber pressure, axial and radial pressure gradients, and pressure fluctuation laws to evaluate the grouting efficiency and uniformity and ensure the performance of the sealing system.

[0035] Power and transmission system features: Monitor the motor synchronism, load balance, efficiency index, and temperature distribution, analyze the phase current balance and power quality, and evaluate the energy utilization status of the system.

[0036] Auxiliary system features: Include the pressure balance and consumption rate of the foam system, the liquid level change characteristics of the bentonite system, the industrial water circulation parameters, and the mixing liquid transportation uniformity index to ensure the reliable operation of the auxiliary system.

[0037] Through the construction of these physical features, the efficient conversion of the original sensor data is realized, laying a foundation for subsequent dimension reduction and anomaly detection. At the same time, the interpretability and prediction accuracy of the model are improved.

[0038] The principal component analysis (PCA) dimension - reduction process in step 2 includes: calculating the covariance matrix of the extracted features; performing eigenvalue decomposition on the covariance matrix; automatically selecting the number of principal components according to the cumulative variance contribution rate, and selecting the minimum number of principal components such that the cumulative variance contribution rate ≥ 95%.

[0039] The multi-scale time encoder in step 3 processes the input sequence at different time scales. The i-th scale focuses on the most recent part of the sequence with a length of L / (2^i), where L is the length of the original sequence.

[0040] The enhanced time attention module in step 3 is implemented by the following formula:

[0041] Attention mechanism:

[0042]

[0043] Processing after multi-head attention:

[0044] X′ = LayerNorm(X + MultiHead(X, X, X))

[0045] Processing after the feed-forward network (FFN):

[0046] X″ = LayerNorm(X′ + FFN(X′))

[0047] where Q, K, and V represent the query, key, and value matrices derived from the input, d k is the dimension of the key vector, FFN represents the feed-forward network, X represents the input of the current layer, X′ represents the result after multi-head attention, residual connection, and layer normalization and is used as the input for the next step, X″ represents the result after FFN, residual connection, and layer normalization, Attention represents the attention mechanism for calculating the relationships between parts of the input, softmax represents the normalization function that converts dot product scores into probability distributions, LayerNorm represents layer normalization, and MultiHead represents the multi-head attention mechanism which is a core component of the Transformer.

[0048] The residual structure in step 3 is implemented by the following formula:

[0049] F(x) = σ(BN(W2σ(BN(W1x)) + H(x))

[0050] where W1 and W2 are learnable weight matrices, BN represents batch normalization, σ is the ReLU activation function, H(x) represents the shortcut connection, F(x) represents the result of feature extraction for the input x, and x represents the output of the enhanced time attention.

[0051] A shield machine abnormal detection system before shutdown adopts the above detection method and specifically includes:

[0052] Data acquisition and preprocessing module: used to acquire shield machine multi-system sensor data and perform preprocessing;

[0053] Hierarchical Hybrid Dimensionality Reduction Module: used for grouping data, feature extraction, and dimensionality reduction processing;

[0054] Multi-scale Spatiotemporal Attention Residual Neural Network Module: used for establishing a prediction model and training;

[0055] Anomaly Detection and Early Warning Module: used for real-time monitoring of the operation status of the shield machine and generating a shutdown early warning.

[0056] The Hierarchical Hybrid Dimensionality Reduction Module includes a feature grouping unit, a feature extraction unit, a PCA dimensionality reduction unit, and a feature fusion unit.

[0057] The Multi-scale Spatiotemporal Attention Residual Neural Network Module includes an input projection layer, a multi-scale time encoder, an enhanced time attention module, a feature fusion component with a residual connection, and an output layer.

[0058] The Anomaly Detection and Early Warning Module can generate a warning signal according to the predicted anomaly probability when it exceeds a preset threshold.

[0059] Compared with the existing technologies, the present invention has the following beneficial effects:

[0060] (1) Through the hierarchical hybrid dimensionality reduction strategy, the feature extraction is targeted and the key features of each subsystem are retained, improving the interpretability of the model;

[0061] (2) The multi-scale spatiotemporal attention mechanism can capture the abnormal behavior patterns of the shield machine at different time scales, enhancing the prediction ability of the model;

[0062] (3) The residual structure effectively alleviates the gradient vanishing problem of the deep network, improving the training stability and generalization ability of the model;

[0063] (4) Experiments show that the accuracy of this method is between 95.01% and 97.25% on four test sets, and the F1 score is between 0.9029 and 0.9194, with high prediction reliability.

[0064] (5) The present invention helps engineering personnel take measures in advance to avoid the construction period delay and economic losses caused by sudden shutdowns, improve the safety and efficiency of shield construction, and is of great significance for promoting the intelligentization of shield machine construction. Brief Description of the Drawings

[0065] Figure 1 is the framework diagram of the hierarchical hybrid dimensionality reduction model of the present invention;

[0066] Figure 2 is the framework diagram of the MSTAR model of the present invention;

[0067] Figure 3 is the structural diagram of the residual neural network of the present invention;

[0068] Figure 4 Structural diagram of the enhanced time attention module of the present invention;

[0069] Figure 5 Confusion matrix result diagram of the present invention on the experimental dataset;

[0070] Figure 6 ROC curve diagram of the present invention on the experimental dataset. Specific implementation manner

[0071] The following further describes the implementation manner of the present invention through an example in conjunction with the accompanying drawings.

[0072] I. System architecture

[0073] The abnormal detection method for the shield machine before shutdown provided by the present invention mainly includes a data acquisition and preprocessing module, a hierarchical hybrid dimensionality reduction module, a multi-scale spatio-temporal attention residual neural network module, and an abnormal detection and warning module.

[0074] 1. Data acquisition and preprocessing module

[0075] This module is responsible for collecting multi-system sensor data of the shield machine, including but not limited to the following system parameters:

[0076] (1) Boring system parameters: including cutter head rotation speed, cutter head torque, cutter head power, cutter head angle, cutter head ring cumulative working time, cutter head total cumulative working time, penetration rate, propulsion speed, total thrust, propulsion pressure, and propulsion cylinder displacement and pressure values;

[0077] (2) Attitude control system parameters: including VMT guide vertical front middle, VMT guide horizontal middle, VMT guide vertical trend, cut horizontal deviation, VMT guide roll angle, and articulated system related pressure values and guide wheel front inclination angle;

[0078] (3) Earth pressure control system parameters: including upper left middle earth pressure, upper right middle earth pressure, screw conveyor rotation speed, torque, pressure, oil temperature, make-up oil pressure, oil degree, and screw conveyor upper and lower gate strokes;

[0079] (4) Grouting system parameters: including shield tail seal front, middle and rear cavity pressures, EP2 outer seal pressure, grouting times, and cumulative amounts of various grouting liquids;

[0080] (5) Power and transmission system parameters: including motor frequency, power, current, torque, temperature, phase current, power factor, and fuel tank temperature;

[0081] (6) Auxiliary system parameters: including cumulative amount of foam liquid, cumulative amount of bentonite, bentonite tank liquid level, air flow, foam pressure, mixed liquid flow, and industrial water system parameters.

[0082] This module preprocesses the collected data, including screening data under the tunneling working state, removing data under the power-off and shutdown state, and performing standardization processing to eliminate the scale differences between the measurement values of different sensors.

[0083] 2. Hierarchical Hybrid Dimensionality Reduction Module

[0084] As Figure 1 shown, the hierarchical hybrid dimensionality reduction module uses a hierarchical structure to reduce the dimensionality of the high-dimensional data of the shield machine:

[0085] (1) Feature grouping guided by professional knowledge: Divide the sensor data into subsystem groups such as the tunneling system, attitude control system, earth pressure control system, grouting system, power and transmission system, and auxiliary system;

[0086] (2) Feature engineering guided by physical information: For each subsystem, design specific feature extraction algorithms to transform the original sensor data into high-level features with physical meanings;

[0087] a. Tunneling system feature extraction: Focus on cutterhead efficiency, cutterhead energy consumption, cutterhead working state, and time accumulation features;

[0088] b. Attitude system feature extraction: Focus on attitude stability analysis, trajectory deviation analysis, and articulated response analysis;

[0089] c. Earth pressure control system feature extraction: Focus on earth pressure distribution features, earth pressure dynamic features, earth pressure balance features, and screw conveyor system features;

[0090] d. Grouting system feature extraction: Focus on chamber pressure statistical features, pressure gradient features, pressure fluctuation features, and grouting features;

[0091] e. Power and transmission system feature extraction: Focus on motor features and electrical system features;

[0092] f. Auxiliary system feature extraction: Focus on foam injection features, bentonite supply features, industrial water circulation features, and mixed liquid transportation features;

[0093] (3) Adaptive PCA dimensionality reduction: Apply principal component analysis to each subsystem for dimensionality reduction, and automatically select the optimal number of principal components through the cumulative variance contribution rate;

[0094] (4) Feature fusion: Fuse the features after dimensionality reduction of each subsystem to form the final feature set.

[0095] 3. Multi-scale Spatiotemporal Attention Residual Neural Network Module

[0096] As Figure 2As shown, the structure of the multi-scale spatio-temporal attention residual neural network (MSTAR) module includes:

[0097] (1) Input projection layer: performs initial processing on the fused features;

[0098] (2) Multi-scale time encoder: processes the input sequence using BiGRU encoders with different time scales. The i-th scale focuses on the most recent part of the sequence (length L / (2^i));

[0099] (3) Enhanced time attention layer: As Figure 4 shown, adopts the multi-head self-attention mechanism, combined with residual connection and layer normalization to stabilize training;

[0100] (4) Feature fusion component with residual connection: As Figure 3 shown, avoids signal attenuation during multi-layer transmission through residual connection, ensuring the smooth transmission of gradients during backpropagation;

[0101] (5) Output layer: outputs the probability of anomaly occurrence through the sigmoid activation function.

[0102] The mathematical expression of this module is:

[0103]

[0104] In the formula, represents the input sequence with scale i and length L / (2 i ), E i represents the bidirectional GRU encoder, A i is the enhanced time attention module, and R i represents the residual feature processing block.

[0105] The multi-scale features are fused with the original input through the residual block:

[0106]

[0107] In the formula, represents the predicted output, [·] represents concatenation, F i represents the feature extracted at scale i, x T is the input at the last time step, F fusion is the sequence of the residual block and the linear layer, and σ is the sigmoid activation function that outputs the probability of anomaly occurrence.

[0108] 4. Anomaly detection and warning module

[0109] Based on the trained MSTAR model, this module monitors the operation status of the shield machine in real time and issues a warning signal when an abnormal pattern is detected. The specific steps include:

[0110] (1) Real-time data acquisition and preprocessing: Continuously acquire the operation data of the shield machine and perform preprocessing;

[0111] (2) Feature extraction and dimensionality reduction: Extract features and reduce the dimensionality of the preprocessed data;

[0112] (3) Anomaly detection: Use the trained MSTAR model to detect anomalies in the data after dimensionality reduction;

[0113] (4) Early warning signal generation: When the detected anomaly probability exceeds the preset threshold, generate an early warning signal; The threshold setting needs to adopt a differential threshold strategy according to the characteristics of different shield machine models, considering environmental factors such as the current geological conditions and tunneling rate, and dynamically adjust the threshold.

[0114] II. Method implementation steps

[0115] The specific implementation steps of the method of the present invention are as follows:

[0116] 1. Dataset construction and preprocessing

[0117] Select the historical tunneling data of the shield machine in the tunnel engineering project for model verification analysis, and screen the data in the tunneling working state. Data preprocessing includes:

[0118] (1) Remove the data in the power-off and shutdown state;

[0119] (2) Remove unstable data (such as the data of the first two rings after initial startup);

[0120] (3) Standardization processing, standardize the feature values into a distribution with a mean of 0 and a standard deviation of 1:

[0121]

[0122] In the formula, x is the original feature, and μ and σ are the mean and standard deviation respectively.

[0123] 2. Hierarchical hybrid dimensionality reduction processing

[0124] (1) Feature grouping: Divide the sensor data into subsystem groups such as the tunneling system, attitude control system, earth pressure control system, grouting system, power and transmission system, and auxiliary system according to functions;

[0125] (2) Feature extraction: Apply the corresponding feature extraction algorithm to each subsystem to construct physical features;

[0126] (3) PCA dimensionality reduction: Apply principal component analysis to each subsystem for dimensionality reduction, and determine the number of principal components by calculating the cumulative variance contribution rate:

[0127]

[0128] Among them, λ i is the i-th eigenvalue, γ i is the variance contribution rate of the i-th principal component, and Γk is the cumulative variance contribution rate of the first k principal components. Select the smallest k value that satisfies Γ k ≥ 0.95 as the final number of principal components.

[0129] (4) Feature fusion: Fuse the features after dimensionality reduction of each subsystem to form the final feature set.

[0130] 3. MSTAR Model Construction and Training

[0131] (1) Model construction: Construct the MSTAR model according to the Figure 2 shown architecture;

[0132] (2) Data labeling: Label the data 15 minutes before the stop point as abnormal data (category 0), and the remaining data as normal data (category 1);

[0133] (3) Model training: Train the model using the weighted binary cross-entropy loss function:

[0134]

[0135] Among them, N is the sample size, y i is the ground truth label, is the predicted probability.

[0136] The training strategy includes:

[0137] Use the AdamW optimizer with cosine annealing scheduling; gradient clipping (maximum norm 1.0) to prevent gradient explosion; ReduceLROnPlateau to adaptively adjust the learning rate; early stopping to avoid overfitting; regularization measures such as weight decay and Dropout.

[0138] 4. Model Evaluation and Validation

[0139] Evaluate the model performance using accuracy, precision, recall, and F1 score:

[0140]

[0141] Among them, TP, TN, FP, and FN represent the number of true positive, true negative, false positive, and false negative samples respectively.

[0142] In addition, draw the confusion matrix and ROC curve, and calculate the AUC value to further evaluate the model performance.

[0143] III. Embodiment

[0144] The application effects of the present invention are illustrated by specific embodiments as follows:

[0145] Embodiment 1: Experimental dataset and parameter settings

[0146] Select the historical tunneling data of the shield machine in the Dayun - Pingshan section of the Shenzhen - Dapeng Intercity Railway for verification and analysis, specifically including:

[0147] Data range: Ring data from 1121 to 1193;

[0148] Data acquisition frequency: 5 s / time;

[0149] Number of features: 131 original features, expanded to 147 features after feature extraction, and reduced to 16 features after dimensionality reduction;

[0150] Number of outliers: 52 abnormal shutdown points;

[0151] Data division: According to the complete ring numbers, the data is evenly divided into 4 test sets.

[0152] Embodiment 2: Experiments with different abnormal marking step lengths before shutdown

[0153] For dataset 1, mark the data 10 minutes, 15 minutes, and 20 minutes before the shutdown point as abnormal data respectively, and compare the prediction effects:

[0154] 10 - minute step length: Accuracy 96.33%, Precision 0.8333, Recall 0.8291, F1 - score 0.8312;

[0155] 15 - minute step length: Accuracy 97.25%, Precision 0.9286, Recall 0.8996, F1 - score 0.9139;

[0156] 20 - minute step length: Accuracy 93.88%, Precision 0.8726, Recall 0.8821, F1 - score 0.8773.

[0157] The results show that the prediction effect marked with a 15 - minute step length is the best, and it can capture key features more effectively.

[0158] Embodiment 3: Verification of prediction performance on four test sets

[0159] On four independent test sets, the performance of the proposed model is as follows:

[0160] Dataset 1: Accuracy 97.25%, Precision 0.9286, Recall 0.8996, F1 - score 0.9139;

[0161] Dataset 2: Accuracy 96.33%, Precision 0.9057, Recall 0.9335, F1 - score 0.9194;

[0162] Dataset 3: Accuracy 95.01%, Precision 0.9299, Recall 0.8774, F1-score 0.9029;

[0163] Dataset 4: Accuracy 96.50%, Precision 0.9140, Recall 0.9116, F1-score 0.9129.

[0164] The confusion matrix results are as Figure 5 shown. On the four test sets, the number of true negatives (TN) is significantly higher than that of false positives (FP), and the number of true positives (TP) also remains at a relatively high level compared to false negatives (FN), proving that the model can accurately identify the normal operating state and pre-shutdown abnormal state of the shield machine.

[0165] The area under the ROC curve (AUC) of all test sets is between 0.98 - 0.99, as specifically Figure 6 shown, indicating that the model has stable classification performance.

Claims

1. An abnormal detection method before the shutdown of a shield machine, characterized in that The following steps are involved: Step 1: Collect and pre-process shield machine operation data; 1.1) Collecting shield machine multi-system sensor data, including cutterhead speed, cutterhead torque, cutterhead power, cutterhead angle, penetration, chamber soil pressure, screw machine speed, screw machine torque, screw machine pressure, motor frequency, motor current, motor temperature, air flow, and foam flow; 1.2) Valid data from the TBM during tunneling operations was screened and a systematic preprocessing process was implemented. First, the unstable data generated by the TBM restart after a long power outage was identified and removed from the first two tunneling cycles. Furthermore, outliers in the original data collection were retained to ensure data authenticity. Subsequently, multidimensional time series parameters were standardized. Step 2: Process the data according to the hierarchical hybrid dimensionality reduction strategy; 2.1) Divide the pre-processed sensor data into multiple functional subsystems according to their functions, namely, the tunneling system, attitude system, earth pressure control system, grouting system, power and transmission system, and auxiliary system; 2.2) Extract features from each subsystem and construct physical features as follows: 2.3) Apply principal component analysis (PCA) to each subsystem for intra-group dimensionality reduction; 2.4) Hierarchical hybrid dimensionality reduction through inter-group feature fusion; Step 3: Establish a multi-scale spatiotemporal attention residual neural network model MSTAR; The MSTAR neural network is constructed by sequentially connecting the input layer, multi-scale temporal encoding layer, enhanced temporal attention layer, residual connection feature fusion layer, and output layer. The details of each layer are as follows: Input layer design: Construct an input projection layer, receive the reduced-dimensional feature vector as input, and perform a linear transformation on the input data; Multi-scale temporal coding implementation: A multi-scale strategy is designed to partition the input sequence into time windows of varying lengths. For each scale i, an encoder is constructed to process input sequences of varying window lengths, known as a multi-scale temporal encoder. A bidirectional gated recurrent unit (BiGRU) is deployed at each scale to simultaneously capture both past and future context. Features extracted at each scale are processed in parallel to capture dependencies at different temporal granularities. Enhanced Temporal Attention Layer: An enhanced temporal attention module is introduced after the multi-scale temporal encoding layer to implement a multi-head self-attention mechanism and calculate the correlation strength between different time steps. The attention mechanism dynamically adjusts the attention weights for different time steps, enhancing the model's ability to identify abnormal patterns before downtime. Residual connection structure deployment: The residual connection structure is introduced after the enhanced temporal attention layer and within the feature fusion layer. Each residual block consists of two fully connected layers, a batch normalization layer, and a ReLU activation function. Feature fusion and output layer design: concatenate and fuse multi-scale features with features after residual processing; The output layer is designed to use a sigmoid activation function to output the probability value of a shutdown anomaly. When the output probability value exceeds the preset threshold, it is determined to be an abnormal state of an impending shutdown. Step 4: Train the model and perform anomaly detection; 4.1) Mark the data of the 15 minutes before the stop point as abnormal, and perform the processing of steps 1 and 2 to obtain the data after dimensionality reduction; 4.2) Use the weighted binary cross-entropy loss function to train the model constructed in step 3; 4.3) Use the trained model to perform anomaly detection on the data after dimensionality reduction, so as to determine the normal and abnormal state data under the condition of shield tunneling.

2. The abnormal detection method before the shutdown of a shield machine according to claim 1, wherein, The physical feature extraction in step 2 includes: Tunneling system features: Extract the cutter head efficiency, energy consumption ratio, working status and time accumulation features, and at the same time calculate the propulsion damping, cylinder balance and propulsion attitude angle to comprehensively characterize the tunneling performance; Attitude control system features: Analyze the attitude stability from the vertical, horizontal and rolling dimensions, calculate the trajectory cumulative deviation and change rate, and evaluate the response characteristics of the articulation system to reflect the position control ability of the shield machine; Earth pressure control system features: Evaluate the earth pressure state of the tunneling face through pressure distribution, gradient and fluctuation features, calculate the equilibrium indicators in the horizontal and vertical directions, monitor the operation of the screw conveyor and the parameters of the hydraulic system, and prevent the risk of earth pressure out of control; Grouting system features: Analyze the statistical characteristics of the chamber pressure, axial and radial pressure gradients, and pressure fluctuation laws, evaluate the grouting efficiency and uniformity, and ensure the performance of the sealing system; Power and transmission system features: Monitor the motor synchronization, load balance, efficiency index and temperature distribution, analyze the phase current balance and power quality, and evaluate the energy utilization status of the system; Auxiliary system features: Include the pressure balance and consumption rate of the foam system, the liquid level change characteristics of the bentonite system, the industrial water circulation parameters and the mixing liquid delivery uniformity index to ensure the reliable operation of the auxiliary system.

3. The abnormal detection method before the shutdown of a shield machine according to claim 1, characterized in that The principal component analysis PCA dimensionality reduction process in step 2 includes: calculating the covariance matrix of the extracted features; performing eigenvalue decomposition on the covariance matrix; automatically selecting the number of principal components according to the cumulative variance contribution rate, and selecting the smallest number of principal components that makes the cumulative variance contribution rate ≥ 95%.

4. The abnormal detection method before the shutdown of a shield machine according to claim 1, characterized in that The multi-scale time encoder in step 3 processes the input sequence at different time scales. The i-th scale focuses on the nearest part of the sequence, with a length of L / (2^i), where L is the length of the original sequence.

5. The abnormal detection method before the shutdown of a shield machine according to claim 1, characterized in that The enhanced time attention module in step 3 is implemented by the following formula: Attention mechanism: Processing after multi-head attention: X ′ = LayerNorm(X + MultiHead(X, X, X)) Processing after the feed-forward network FFN: X ″ = LayerNorm(X ′ + FFN(X ′ )) where Q, K, and V represent query, key, and value matrices derived from the input, and d k is the dimension of the key vectors, FFN represents the feed-forward network, X represents the input to the current layer, and X ′ represents the result after multi-head attention, residual connection, and layer normalization, which is used as the input for the next step. X ″ represents the result after FFN, residual connection, and layer normalization. Attention represents the attention mechanism for calculating the relationships between different parts of the input. Softmax represents the normalization function that converts dot-product scores into a probability distribution. LayerNorm represents layer normalization, and MultiHead represents the multi-head attention mechanism.

6. The abnormal detection method before the shutdown of a shield machine according to claim 1, wherein, The residual structure in step 3 is implemented by the following formula: F(x) = σ(BN(W2σ(BN(W1x)) + H(x)) where W1 and W2 are learnable weight matrices, BN represents batch normalization, σ is the ReLU activation function, H(x) represents the shortcut connection, F(x) represents the result of feature extraction for the input x, and x represents the output of the enhanced time attention.

7. An abnormal detection system before the shutdown of a shield machine, adopting the detection method described in any one of claims 1-6, characterized in that, Specifically include: Data acquisition and preprocessing module: Used to collect the data of the multi-system sensors of the shield machine and perform preprocessing; Hierarchical hybrid dimensionality reduction module: Used to group, extract features and perform dimensionality reduction processing on the data; Multi-scale spatio-temporal attention residual neural network module: Used to establish a prediction model and perform training; Anomaly detection and early warning module: Used to monitor the operation status of the shield machine in real time and perform shutdown early warning.

8. The abnormal detection system before the shutdown of a shield machine according to claim 7, characterized in that, The hierarchical hybrid dimensionality reduction module includes a feature grouping unit, a feature extraction unit, a PCA dimensionality reduction unit, and a feature fusion unit.

9. The abnormal detection system before the shutdown of a shield machine according to claim 7, characterized in that, The multi-scale spatio-temporal attention residual neural network module includes an input projection layer, a multi-scale time encoder, an enhanced time attention module, a feature fusion component with residual connections, and an output layer.

10. The abnormal detection system before the shutdown of a shield machine according to claim 7, characterized in that, The anomaly detection and warning module can generate a warning signal according to the predicted anomaly probability when it exceeds a preset threshold.

Images