TBM cutter wear monitoring method based on multi-source data fusion and abnormal self-adaptation

Abstract

This invention relates to a TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptation. The steps are as follows: Using spatiotemporal reference nodes deployed within a distributed sensor cluster, the raw multi-source data collected in real-time by each sensor during the TBM tool's operation are subjected to spatiotemporal deviation compensation and timestamp alignment, followed by standardized preprocessing. Based on the signal dimensions of the standardized preprocessed multi-source data and the characteristic parameters of the TBM tool wear stage, the correlation weights of the two are calculated. Anomaly types are identified and differentiated using a multi-factor three-dimensional adaptive threshold model, resulting in purified multi-source data. Through hierarchical fusion processing, a multi-dimensional fused wear index is generated, and wear classification diagnosis is performed based on the multi-dimensional fused wear index and its changing trends. This invention, through multi-source data fusion and anomaly adaptive processing, effectively filters out interference such as sudden changes in operating conditions and equipment vibration, significantly reducing the monitoring misjudgment rate.

Description

Technical Field

[0001] This invention relates to a TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptation, belonging to the field of TBM tool condition monitoring and intelligent operation and maintenance technology. Background Technology

[0002] As tunnel engineering construction develops towards deeper and larger-scale operations, TBMs (Tunnel Boring Machines) have become core equipment in tunnel construction due to their advantages of high efficiency, safety, and environmental friendliness. As key working components that directly contact the rock, the wear condition of TBM cutters directly determines tunneling efficiency, construction safety, and project costs. During tunneling under complex geological conditions, cutters are prone to wear, chipping, and other malfunctions. If these issues are not monitored accurately and in a timely manner, they can lead to cutterhead jamming, a sharp drop in tunneling efficiency, and even equipment damage and construction accidents. Therefore, achieving real-time and accurate monitoring of TBM cutter wear is indispensable for ensuring smooth construction progress and optimizing operation and maintenance strategies.

[0003] Real-time TBM tool wear monitoring refers to a technology that uses deployed sensors to capture various physical signals during tool operation, combined with data processing and analysis techniques, to assess tool wear levels and diagnose fault conditions in real time. This technology provides maintenance personnel with accurate data support on tool status, guiding tool replacement timing and optimizing tunneling parameters, thereby reducing downtime for maintenance, minimizing ineffective maintenance costs, and improving the economy and safety of engineering construction.

[0004] Currently, some related technologies have been explored for TBM tool wear monitoring: 1) The TBM cutter wear monitoring device and method disclosed in announcement number CN112184630A, based on machine vision, involves a controller driving a mobile device and an industrial camera to move along an arc-shaped track, capturing images of the cutter, and then matching and fusing these images with images of unworn cutters. The wear value is then calculated using the minimum bounding rectangle algorithm. This method relies on visual image acquisition and processing technology, which is susceptible to interference from factors such as rock debris, dust, and changes in lighting at the construction site, leading to reduced image quality and affecting the accuracy of wear calculation. Furthermore, the need for a mechanical structure to mount the camera for mobile monitoring results in a slow response time, making it unsuitable for the real-time monitoring requirements of continuous tunneling. In addition, this method struggles to distinguish between surface contaminants on the cutter ring and actual wear, posing a risk of misjudgment.

[0005] 2) The TBM cutter wear real-time monitoring device and system based on induced current disclosed in announcement number CN118857192A generates a magnetic field through an excitation coil, collects current change data using an inductive sensor and temperature data using an infrared temperature sensor, and then converts the data into cutter wear amount using an empirical formula. This technology relies only on two data sources, current and temperature, resulting in a single data dimension. It cannot distinguish the impact of external interference such as geological changes and equipment vibration on the signal, nor can it differentiate between the impact of cutter wear on the signal. Furthermore, it uses a fixed empirical formula for wear amount conversion, lacking an adaptive adjustment mechanism for different geological conditions and tunneling speeds. When the working conditions change, the monitoring accuracy decreases significantly. In addition, no abnormal data processing scheme is designed, and the data reliability cannot be guaranteed when the sensor is affected by electromagnetic interference or temperature drift.

[0006] In summary, most existing TBM tool wear monitoring technologies rely on single data dimensions or visual monitoring methods, which generally suffer from weak anti-interference capabilities and insufficient real-time performance. Furthermore, they lack working condition adaptive mechanisms and abnormal data processing solutions, making it difficult to meet the actual application requirements of complex tunneling conditions in terms of monitoring accuracy and stability. Therefore, there is an urgent need for a TBM tool wear monitoring method based on multi-source data fusion and abnormal adaptive operation to solve the above problems and promote the development of intelligent operation and maintenance technology for TBMs. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptive processing. Through multi-source data fusion and anomaly adaptive processing, interference such as sudden changes in working conditions and equipment vibration can be effectively filtered out, and the monitoring misjudgment rate can be greatly reduced.

[0008] To address the aforementioned technical problems, the technical solution proposed in this invention is: a TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptation, comprising the following steps: utilizing spatiotemporal reference nodes deployed within a distributed sensor cluster, spatiotemporal deviation compensation and timestamp alignment are performed on the raw multi-source data collected in real time by each sensor during the TBM tool's operation. The multi-source data after spatiotemporal deviation compensation and timestamp alignment undergoes standardized preprocessing, while the sampling frequency of the distributed sensor cluster is dynamically adjusted. Based on the signal dimensions of the standardized preprocessed multi-source data and the characteristic parameters of the TBM tool wear stage, the correlation weight between the two is calculated. An adaptive threshold is calculated using a multi-factor three-dimensional adaptive threshold model. Anomaly types are identified and differentiated based on the adaptive threshold, resulting in purified multi-source data. A cross-modal feature alignment enhancement fusion model is constructed, and hierarchical fusion processing is performed on the purified multi-source data to generate a multi-dimensional fused wear index. Wear classification diagnosis is performed based on the multi-dimensional fused wear index and its changing trends.

[0009] Preferably, the process of spatiotemporal deviation compensation and timestamp alignment is as follows: the fixed non-rotating end of the TBM cutterhead spindle is used as the spatial reference origin of the spatiotemporal reference node, and the industrial-grade synchronous clock of the TBM main control system is used as the time reference source of the spatiotemporal reference node. Based on the spatial reference origin and the time reference source, a unified reference for time synchronization and spatial positioning of all sensors is constructed. The timestamp deviation data and spatial position deviation data of each sensor are obtained through the spatiotemporal reference node to construct a multi-dimensional spatiotemporal deviation compensation matrix. The original multi-source data collected by each sensor is subjected to spatiotemporal deviation compensation through linear transformation, so that the spatial position deviation of the multi-source data after spatiotemporal deviation compensation is aligned with the timestamp deviation.

[0010] Preferably, the standardization preprocessing process is as follows: based on the multidimensional spatiotemporal deviation compensation matrix, signal amplitude calibration is performed on the multi-source data after spatiotemporal deviation compensation to eliminate signal amplitude distortion caused by sensor hardware nonlinearity error, installation spatial position deviation and TBM tool excavation environment interference; using the linear normalization method, the mechanical parameters and working condition parameters with inconsistent dimensions and ranges of the multi-source data after signal amplitude calibration are mapped to a unified range to eliminate the interference of dimension differences on the correlation weight solution and multi-source feature fusion process.

[0011] Preferably, the method for dynamically adjusting the sampling frequency of the distributed sensor cluster is as follows: Based on the original multi-source data, determine the corresponding geological grade and historical tool wear rate; construct a coupled correlation model based on the variation law of the geological grade and historical tool wear rate; preset the maximum value of the historical tool wear rate as... The minimum preset historical tool wear rate is The local geological grade is hard rock, and the historical tool wear rate is... At that time, the sampling frequency was adjusted to high frequency; when the geological grade was hard rock and the historical tool wear rate was... At that time, the sampling frequency was adjusted to high frequency; when the geological grade was soft rock and the historical tool wear rate was... At that time, the sampling frequency was adjusted to low frequency; when the geological grade was soft rock and the historical tool wear rate was... At that time, adjust to a lower frequency sampling.

[0012] Preferably, the characteristic parameters of the TBM tool wear stage include: the Pearson correlation coefficient of the wear amount, the sensitivity coefficient of the wear stage, and the prediction confidence of the current wear stage; and the correlation weights of each signal dimension of the standardized preprocessed multi-source data and the characteristic parameters of the TBM tool wear stage. The calculation is as follows: ,in, For the first Pearson correlation coefficients after dimension normalization of multi-source data signals; For the first The wear stage sensitivity coefficient after dimensional normalization of multi-source data signals; The confidence level for the current wear stage prediction; The total number of each signal dimension in the standardized preprocessed multi-source data; the prediction confidence of the current wear stage. The calculation formula is as follows: ,in, The number of sampling points; For the first Predicted tool wear value for each sample; For the first The actual value of tool wear for each sample.

[0013] Preferred anomaly types include: sensor failure, transmission interference, and sudden change in operating conditions; correspondingly, the identification process is as follows: first, an adaptive threshold is calculated using a multi-factor three-dimensional adaptive threshold model. The adaptive threshold The calculation formula is: ,in, , , For preset weighting coefficients and ; Geological grade; This refers to the tunneling speed; This is the wear and tear stage; This represents the historical anomaly incidence rate. This is the initial baseline value; Based on an adaptive threshold, suspected outliers are identified, and the following steps are performed on consecutive suspected outliers: Extract its time-domain features, frequency-domain energy distribution features, and spatial correlation features, and integrate the coupling correlation between geological parameters and TBM operating parameters to complete the anomaly type classification; if the signal amplitude of the output signal of a single sensor in the distributed sensor cluster exceeds the adaptive threshold for no less than Q consecutive sampling points. Or a single sensor experiences a complete signal interruption, and If the signal amplitudes of multiple sensors output from a distributed sensor cluster synchronously generate high-frequency random fluctuations, then it is determined to be a sensor fault; and ,at the same time If the signal amplitude of all sensors related to the TBM cutting tool's tunneling conditions synchronously exhibits a step change, then it is determined to be transmission interference; and ,at the same time If the changing trend of geological parameters is consistent with the predicted trend of geological parameters, it is determined to be a sudden change in working conditions. in, The standard deviation of the signal amplitude; This indicates the preset standard deviation threshold; Indicates the spatial correlation coefficient of sensors of the same type; This indicates the preset spatial correlation coefficient threshold; Represents the time-domain characteristic fluctuation coefficient; This represents the preset threshold for the time-domain characteristic fluctuation coefficient; The proportion of high-frequency energy representing the frequency domain energy distribution characteristics; This indicates the preset threshold for the proportion of energy in the high-frequency band; This indicates the consistency of signal trends among adjacent sensors of the same type. This indicates the preset signal trend consistency threshold; Indicates the rate of change of signal amplitude; This indicates the preset threshold for the rate of change of signal amplitude; Indicates duration; Indicates the preset duration threshold; This represents the product of the rate of change of the geological parameters of the TBM and the anomaly of the TBM signal; This indicates the preset product threshold.

[0014] Preferably, the differential processing of the anomaly type is as follows: using the spatial correlation matrix constructed by multiple adjacent sensors of the same type as spatial constraints, and combining it with the sliding window time series trend model to complete the missing data of sensor faults; for noise data caused by transmission interference, a denoising method combining wavelet packet band energy weighted decomposition and the introduction of masking signal to suppress mode aliasing is adopted; the interference of sudden changes in working conditions is determined by geological-equipment signal coupling verification, and at the same time, new anomaly patterns are learned in real time through a lightweight incremental learning unit to update the local parameters associated with the current wear stage and anomaly type.

[0015] Preferably, the hierarchical fusion processing steps are as follows: using the mechanical parameter features and operating condition parameter features in the purified multi-source data as training samples, mapping the mechanical parameter features and operating condition parameter features to the same space to achieve cross-modal feature alignment; performing consistency verification on the mechanical parameter features and operating condition parameter features, and combining the attention-weighted fusion alignment of the cross-modal features through deep feature extraction.

[0016] Preferably, the generation process of the multi-dimensional fusion wear index is as follows: based on the measured data of different working condition parameters and TBM tool wear stage feature parameters in the purified multi-source data, a three-dimensional mapping table of working condition-wear stage-feature weight is trained and generated, which serves as the generation benchmark for the basic weights of working condition parameters and the corrected weights of wear stages in the two-factor weight model, thereby generating dynamic weights; the dynamic weights The calculation formula is: ,in, For the first The basic weights of each feature based on its operating conditions; For the first Correction weights for each wear stage; For the first The cross-modal alignment coefficients of each feature are used; combined with normalized feature values ​​and fusion residual compensation, a multi-dimensional fused wear index is generated; the multi-dimensional fused wear index... The calculation formula is as follows: ,in, These are supplementary coefficients for the normalized fusion residuals; The total number of features; For the first The normalized eigenvalues.

[0017] Preferably, based on a multi-dimensional fusion wear index Preset minimum multi-dimensional fusion wear index And the preset maximum multi-dimensional fusion wear index And the rising rate of the multi-dimensional integrated wear index change trend. Preset minimum ascent speed and preset maximum rate of ascent Wear grading diagnosis is performed; the wear grading diagnosis process is as follows: ,and If it is in a stable state, it indicates slight wear; ,or This indicates moderate wear. ,or If so, it indicates severe wear.

[0018] The beneficial effects of this invention are as follows: Through multi-source data fusion and anomaly adaptive processing, interference factors such as sudden changes in working conditions and equipment vibration can be effectively filtered out, significantly reducing the monitoring misjudgment rate. Simultaneously, by combining dynamic acquisition frequency adjustment and lightweight optimization of the deep network, it adapts to complex and changing geological conditions, meeting the real-time monitoring requirements of continuous TBM tunneling and effectively compensating for the insufficient early warning accuracy of existing technologies. Furthermore, through multi-indicator precise hierarchical diagnosis and intelligent early warning, it can provide a scientific basis for TBM tool replacement decisions, reducing the frequency of unplanned downtime and further improving the safety and economic efficiency of tunnel excavation construction. Attached Figure Description

[0019] The present invention will be further described below with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the method flow of Example 1; Figure 2 This is a flowchart of the distributed sensor cluster in Example 1; Figure 3 This is the flowchart of the anomaly adaptive handling in Example 1; Figure 4 This is a flowchart of multi-source information fusion and wear classification diagnosis in Example 1. Detailed Implementation

[0020] Example 1 This embodiment presents a TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptive processing. Addressing the common problems of weak anti-interference capabilities and insufficient real-time performance in existing TBM tool wear monitoring technologies that largely rely on single data dimensions or visual monitoring methods, this method effectively filters out interference from sudden changes in operating conditions and equipment vibrations through multi-source data fusion and anomaly adaptive processing, significantly reducing the monitoring misjudgment rate. Figure 1 As shown, the steps are as follows: By utilizing spatiotemporal reference nodes deployed within a distributed sensor cluster, spatiotemporal deviation compensation and timestamp alignment are performed on the raw multi-source data collected in real-time by each sensor during the TBM tool's operation. The multi-source data after spatiotemporal deviation compensation and timestamp alignment undergoes standardized preprocessing, while the sampling frequency of the distributed sensor cluster is dynamically adjusted. In this embodiment, the distributed sensor cluster includes piezoelectric load sensors, triaxial acceleration and vibration sensors, acoustic emission sensors, three-dimensional force sensors, strain gauge torque sensors, and photoelectric speed sensors. The multi-source data includes mechanical parameters such as tool load, vibration, acoustic emission, penetration force, and torque, as well as working condition parameters such as geological conditions and tunneling speed.

[0021] Spatiotemporal reference nodes provide a unified time and space reference, completing spatiotemporal deviation compensation and timestamp alignment for raw multi-source data. The distributed sensor cluster and spatiotemporal reference nodes are deployed according to the TBM (tunnel boring machine) cutter and key preset positions. Specifically, two piezoelectric load sensors are deployed at each cutter shaft, with a range of... precision Eight triaxial acceleration vibration sensors are installed at the cutter head support, with a measurement range of [missing information]. Frequency response Sixteen acoustic emission sensors are attached to the side of the blade ring using a bonding method, covering a frequency range of... One three-dimensional force sensor is installed at the central axis of the cutter head, with a range of... Two strain gauge torque sensors are installed at the output shaft of the main drive motor, with a range of [missing information]. One photoelectric speed sensor is installed at the cutter head drive gear, with a measurement range of... precision In this embodiment, a total of 96 sensors are deployed. All sensors are connected to the data transmission line via M12 waterproof connectors to resist vibration and dust interference during the tunneling process.

[0022] Based on the signal dimensions of the standardized preprocessed multi-source data and the characteristic parameters of the TBM tool wear stage, the correlation weights of the two are calculated. An adaptive threshold is calculated using a multi-factor three-dimensional adaptive threshold model. Anomalies are identified and differentiated based on the adaptive thresholds, thus obtaining purified multi-source data. In this embodiment, the correlation weights are calculated using an anomaly adaptive processing algorithm with lightweight incremental learning and adaptive attention mechanism. The characteristic parameters of the TBM tool wear stage include the Pearson correlation coefficient of wear amount, the wear stage sensitivity coefficient, and the prediction confidence of the current wear stage. By determining the correlation weights of the signal dimensions of the standardized preprocessed multi-source data and the characteristic parameters of the TBM tool wear stage, the characteristics of the wear stage sensitivity dimension are strengthened.

[0023] Anomaly types include sensor malfunctions, transmission interference, and sudden changes in operating conditions. Anomaly types are identified and differentiated using adaptive thresholds, resulting in purified multi-source data. When the purified data deviates from subsequent normal data... At the same time, the abnormal feature library and threshold model parameters are automatically updated and fed back to the dynamic attention weight calculation stage to form a closed-loop optimization.

[0024] A cross-modal feature alignment enhancement fusion model is constructed to perform hierarchical fusion processing on purified multi-source data, generating a multi-dimensional fused wear index. Wear grading diagnosis is then performed based on the multi-dimensional fused wear index and its changing trends. In this embodiment, a cross-modal feature alignment enhancement fusion model is constructed by combining a deep feature extraction network with an attention mechanism. This model aligns features from different modalities in a unified semantic or geometric space, thereby achieving more effective fusion and improving the performance of wear stage identification and prediction. The multi-dimensional fused wear index provides a scientific basis for TBM tool replacement decisions through accurate grading diagnosis and early warning using multiple indicators, reducing unplanned downtime and improving construction safety and economy.

[0025] In this embodiment, a 12.18m open-face TBM installed in a deep-buried hard rock tunnel project will undergo a 300-hour full-condition trial run, covering tunneling scenarios with rock types of 3-5, and recording TBM operating parameters and multi-dimensional fusion wear index in real time. The calculation results are compared with the actual tool wear measurement data from disassembly. Based on the trial operation data, the parameter configuration was optimized, the geological grade weight coefficient was adjusted to 0.35, the normalization boundary of the correlation weight calculation was optimized, the wear stage correction weight in the two-factor weight mapping model under soft rock conditions was corrected, and the multi-dimensional fused wear index was adjusted. The grading and diagnostic rule boundaries will include moderate wear. The lower limit was adjusted from 0.3 to 0.32 to avoid misjudgments caused by slight fluctuations; the anomaly feature library was updated to supplement new transmission interference feature parameters that appeared during tunneling, and the system timestamp error was stabilized through iterative optimization. Wear prediction error Inference delay Once all indicators meet the preset requirements, it will be put into formal use.

[0026] Furthermore, such as Figure 2 As shown, the process of spatiotemporal deviation compensation and timestamp alignment is as follows: The fixed, non-rotating end of the TBM cutterhead spindle is used as the spatial reference origin of the spatiotemporal reference node, and the industrial-grade synchronous clock of the TBM main control system is used as the time reference source of the spatiotemporal reference node. Based on the spatial reference origin and the time reference source, a unified reference is established for time synchronization and spatial positioning of all sensors. In this embodiment, three spatiotemporal reference nodes are deployed in a structurally stable area of ​​the TBM's supporting system, free from strong vibration and impact disturbances, arranged in a triangular pattern. The fixed, non-rotating end of the TBM cutterhead spindle is used as the spatial reference origin, and the industrial-grade synchronous clock of the TBM main control system is used as the time reference source. Each node integrates a high-precision time synchronization module (time resolution 1μs) and a laser ranging and positioning module (range measurement accuracy...). ( ), serving as a unified benchmark for time synchronization and spatial positioning of all sensors.

[0027] By acquiring timestamp deviation data and spatial location deviation data of each sensor through a spatiotemporal reference node, a multidimensional spatiotemporal deviation compensation matrix is ​​constructed; in this embodiment, using Taking the spatiotemporal deviation compensation matrix as an example, the spatiotemporal reference node collects the timestamp deviation and spatial position deviation of each sensor to construct... 3D spatiotemporal bias compensation matrix, The matrix represents the number of sensors, i.e., 96 sensors. The three rows correspond to timestamp deviation, radial spatial position deviation, and axial spatial position deviation, respectively. The 97 columns correspond to one reference node and 96 sensor nodes. The matrix elements are the measured deviation values ​​of the corresponding sensors relative to the spatiotemporal reference node. The matrix representation is as follows: , in, It is a time variable; This refers to the change in position in the horizontal direction; This represents the change in position in the vertical direction.

[0028] Spatiotemporal bias compensation is performed on the raw multi-source data collected by each sensor through linear transformation, aligning the spatial position deviation of the multi-source data with the timestamp deviation after spatiotemporal bias compensation; in this embodiment, for the first... Raw data sequences from each sensor Perform spatiotemporal deviation compensation, after spatiotemporal deviation compensation The expression is: , in, For time; For the first The timestamp deviation of each sensor relative to the spatiotemporal reference node; For time dimension correction; For the first Radial spatial position deviation of each sensor relative to a spatiotemporal reference node; For the first Axial spatial position deviation of each sensor relative to a spatiotemporal reference node; It is a spatial correction function; Spatial dimension correction; ultimately achieving multi-source data timestamp alignment error. Spatial position deviation compensation accuracy .

[0029] Furthermore, the standardization preprocessing process is as follows: Based on a multi-dimensional spatiotemporal deviation compensation matrix, amplitude calibration is performed on the multi-source data after spatiotemporal deviation compensation to eliminate signal amplitude distortion caused by sensor hardware nonlinearity errors, installation spatial position deviations, and interference from the TBM cutting tool's tunneling environment. In this embodiment, a three-level calibration correction is performed on the multi-source data after spatiotemporal deviation compensation based on a 3×97-dimensional spatiotemporal deviation compensation matrix. Standard force sources, standard vibration tables, and standard torque sources, verified by legal metrology institutions, are used to perform full-range static calibration on all sensors. The linear calibration coefficients and zero-point offsets of each sensor are obtained through least-squares fitting, eliminating the nonlinearity errors and zero-point drift inherent in the sensor hardware. During TBM tunneling, real-time ambient temperature data is collected and used to calculate dynamic temperature compensation based on the temperature drift coefficients of each sensor as specified by the manufacturer. This process performs secondary correction on the statically calibrated data to eliminate dynamic amplitude drift caused by temperature changes and strong vibrations during tunneling. Based on the radial and axial spatial position deviations of each sensor recorded in the 3×97-dimensional spatiotemporal deviation compensation matrix, a spatial position-amplitude calibration function is constructed using the real-time rotational angular velocity of the cutterhead and a rock cutting mechanics model. This function compensates for the amplitude of load, vibration, and acoustic emission sensor signals at different installation positions, eliminating signal amplitude deviations caused by differences in installation spatial position. After amplitude calibration, the dynamic measurement errors of all sensors are corrected. .

[0030] A linear normalization method is adopted to map the mechanical parameters and operating parameters with inconsistent dimensions and ranges of the multi-source data after amplitude calibration to a unified interval, so as to eliminate the interference of dimensional differences on the correlation weight solution and multi-source feature fusion process. In this embodiment, the linear normalization method takes Min-Max as an example. The Min-Max linear normalization method is used to perform dimensional unification processing on the multi-source data after amplitude calibration. For each type of signal dimension, the upper and lower limits of normalization are determined based on the theoretical maximum and minimum values ​​of the corresponding sensor across the full range. The real-time sampled values ​​are mapped to the unified interval [0, 1] through linear transformation. Amplitude limiting processing is performed on sampled values ​​that exceed the theoretical range. When the sampled value is greater than the theoretical maximum value, it is fixed at 1, and when it is less than the theoretical minimum value, it is fixed at 0, to avoid distortion of normalization results caused by abnormal extreme values. After processing, the mechanical parameters and operating parameters with different dimensions and ranges are all in the same numerical range, completely eliminating the interference of dimensional differences on subsequent data processing.

[0031] Furthermore, the method for dynamically adjusting the sampling frequency of the distributed sensor cluster is as follows: Based on the original multi-source data, determine the corresponding geological grade and historical tool wear rate; construct a coupled correlation model based on the variation law of the geological grade and historical tool wear rate; preset the maximum value of the historical tool wear rate as... The minimum preset historical tool wear rate is The geological grade is hard rock, and the historical wear rate of the cutting tool is... At that time, the sampling frequency was adjusted to high frequency; the geological grade was hard rock, and the historical wear rate of the cutting tool was also considered. At that time, the sampling frequency was adjusted to high frequency; the geological grade was soft rock, and the historical wear rate of the cutting tool was [data missing]. At that time, the sampling frequency was adjusted to low frequency; the geological grade was soft rock, and the historical wear rate of the cutting tool was [data missing]. At that time, adjust to a lower frequency sampling.

[0032] In this embodiment, a coupled correlation model is established based on the variation patterns of both geological grade and historical tool wear rate to dynamically adjust the sampling frequency of each sensor. Geological grades 4-5 represent hard rock, and grades 1-3 represent soft rock. The highest sampling frequency is... High frequency Low frequency is small frequency is When the geological grade is 4-5 (hard rock) and the historical wear rate of the cutting tool is greater than or equal to... At that time, adjust the sampling frequency to a higher frequency. Sampling; when the geological grade is 4-5 (hard rock) and the historical wear rate of the cutting tool is less than... At that time, adjust the sampling frequency to a high frequency. Sampling; when the geological grade is 1-3 (soft rock) and the historical wear rate of the cutting tool is greater than or equal to... At that time, adjust the sampling frequency to a low frequency. Sampling; when the geological grade is 1-3 (soft rock) and the historical wear rate of the cutting tool is less than... At that time, adjust the sampling frequency to a lower frequency. sampling.

[0033] Furthermore, the characteristic parameters of the TBM tool wear stage include: the Pearson correlation coefficient of the wear amount, the sensitivity coefficient of the wear stage, and the prediction confidence of the current wear stage; and the correlation weights of each signal dimension of the standardized preprocessed multi-source data and the characteristic parameters of the TBM tool wear stage. The calculation is as follows: , in, For the first Pearson correlation coefficients after dimension normalization of multi-source data signals; For the first The wear stage sensitivity coefficient after dimensional normalization of multi-source data signals; The confidence level for the current wear stage prediction; The total number of dimensions of the multi-source data signal; The prediction confidence level of the current wear stage The calculation formula is as follows: , in, The number of sampling points; For the first Predicted tool wear value for each sample; For the first The actual value of tool wear for each sample. In this embodiment, the prediction confidence level is set to 0.8-1.0, and is calculated by the matching degree between the predicted and actual wear values ​​of the last 5 sampling points.

[0034] In this embodiment, the signal characteristics of the standardized preprocessed multi-source data, including input load, vibration, and acoustic emission, are used. First, the Pearson correlation coefficient between each signal characteristic and the wear amount is calculated. Taking the absolute value of the correlation coefficient for the load signal characteristic as 0.87, the absolute value of the correlation coefficient for the vibration signal characteristic as 0.79, and the absolute value of the correlation coefficient for the acoustic emission signal characteristic as 0.82 as an example, the wear stage sensitivity coefficients for the load signal characteristic, vibration signal characteristic, and acoustic emission signal characteristic are then calculated. The wear stage sensitivity coefficient is the ratio of the coefficient of variation of each signal characteristic at the current wear stage to the average coefficient of variation of the entire wear stage. At the moderate wear stage, the wear stage sensitivity coefficient for the load signal characteristic is 1.2, the wear stage sensitivity coefficient for the vibration signal characteristic is 1.1, and the wear stage sensitivity coefficient for the acoustic emission signal characteristic is 1.15. Simultaneously, the prediction confidence of the current wear stage is calculated. The prediction confidence is obtained by matching the predicted and actual values ​​of the wear amount at the last 5 sampling points. The actual values ​​of the wear amount at the 5 sampling points are as follows: , , , , The predicted values ​​are respectively , , , , The prediction confidence level is calculated. It falls within the value range of 0.8-1.0.

[0035] Based on the above parameter coupling calculation, the correlation weights between each signal feature and the feature parameters of the TBM tool wear stage are calculated, and the total number of core signal features is set to The correlation weights of the load signal features were calculated to be 0.37, the vibration signal features to be 0.30, and the acoustic emission signal features to be 0.33. The sum of the correlation weights of each signal feature was 1, thereby enhancing the characteristics of the wear sensitivity coefficient.

[0036] Furthermore, the anomaly types include: sensor failure, transmission interference, and sudden change in operating conditions; correspondingly, the identification process is as follows: First, calculate the adaptive threshold using a multi-factor three-dimensional adaptive threshold model. The adaptive threshold The calculation formula is: ,in, , , Preset weighting coefficients; Geological grade, ranging from 1 to 5; The tunneling speed is expressed in m / h. This is the wear stage, with values ​​ranging from 1 to 3. For historical anomaly incidence rates, the historical anomaly incidence rate over the past hour is preferred; As the initial baseline, the statistical data of normal data in the past 10 minutes is preferred, and the mean is taken as ± 3 times the standard deviation.

[0037] Multi-factor 3D adaptive threshold building block inputs a statistical benchmark of normal multi-source data from the past 10 minutes (such as mean load). Standard deviation ), geological grade tunneling speed Wear and tear stage Historical anomaly incidence rate Adaptive thresholds are calculated using a multi-factor coupling formula. Based on adaptive thresholds, point-by-point threshold determination is performed on the standardized preprocessed multi-source data. When the signal amplitude exceeds the upper and lower limits of the adaptive threshold, it is marked as a suspected anomaly.

[0038] Based on an adaptive threshold, suspected outliers are identified, and the following steps are performed on consecutive suspected outliers: The time-domain features, frequency-domain energy distribution features, and spatial correlation features are extracted, and the coupling relationship between geological parameters and TBM operating parameters is integrated to complete the anomaly type classification. In this embodiment, the time-domain amplitude features, frequency-domain energy distribution features, and spatial correlation features are extracted for continuously occurring suspected anomaly points. The coupling relationship between real-time acquired geological parameters and TBM equipment operating parameters is combined to complete the accurate classification of three types of anomalies: sensor failure, transmission interference, and sudden change in operating conditions.

[0039] If the signal amplitude of a single sensor output from a distributed sensor cluster exceeds the adaptive threshold for at least Q consecutive sampling points... Or a single sensor experiences a complete signal interruption, and If the error is not detected, it is determined to be a sensor malfunction; in this embodiment, The standard deviation of the signal amplitude. This indicates the preset standard deviation threshold. This represents the spatial correlation coefficient of sensors of the same type. This indicates a preset spatial correlation coefficient threshold. When the signal amplitude of a single-channel piezoelectric load sensor exceeds the adaptive threshold for five or more consecutive sampling points... And the standard deviation of the signal amplitude The signal exhibits an approximately constant value or is completely interrupted, while the spatial correlation coefficient between this sensor and three adjacent sensors of the same type is... If this occurs, it is determined to be a sensor malfunction.

[0040] If the signal amplitudes of multiple sensors output from a distributed sensor cluster synchronously generate high-frequency random fluctuations... and ,at the same time If so, it is determined to be transmission interference; in this embodiment, Represents the time-domain characteristic fluctuation coefficient. This represents the preset threshold for the time-domain characteristic fluctuation coefficient. The proportion of high-frequency energy representing the characteristics of frequency domain energy distribution. This indicates the preset threshold for the proportion of energy in the high-frequency band. This indicates the consistency of signal trends between adjacent sensors of the same type. This indicates the preset signal trend consistency threshold. Multi-channel triaxial accelerometers and acoustic emission sensors are multi-channel sensors. When both multi-channel triaxial accelerometers and acoustic emission sensors simultaneously exhibit high-frequency random fluctuations, the time-domain characteristic fluctuation coefficient of the signal... Furthermore, analysis of the frequency domain energy distribution characteristics shows that the energy proportion in the high-frequency band above 1000Hz is... At the same time, the signal trends of adjacent sensors of the same type are consistent. When this occurs, it is determined to be an abnormal transmission interference.

[0041] If all sensors related to the TBM cutting tool's tunneling conditions simultaneously experience a step change in output signal amplitude... and ,at the same time If the changing trend of geological parameters matches the predicted trend of geological parameters, it is determined to be a sudden change in working conditions; in this embodiment, Indicates the rate of change of signal amplitude. This indicates the preset threshold for the rate of change of signal amplitude. Indicates duration, Indicates the preset duration threshold. This represents the product of the rate of change of the geological parameters of the TBM and the anomaly of the TBM signal. This represents the preset product threshold. When all load, vibration, torque, and speed sensor signals related to the tunneling condition undergo synchronous step changes, the rate of change of the signal amplitude... And duration Simultaneously, the product of the rate of change of geological parameters and the degree of anomaly in equipment signals. When the trend of change is completely consistent with the geological forecast data, it is determined to be an abnormal disturbance of the working condition.

[0042] Furthermore, such as Figure 3 As shown, the differential processing procedure for the aforementioned anomaly type is as follows: Using the spatial correlation matrix constructed from multiple adjacent sensors of the same type as spatial constraints, and combining it with a sliding window time-series trend model, missing data for sensor faults is completed. In this embodiment, a time-series-spatial joint optimization algorithm is used for differentiated processing. Sensor fault data is completed using the spatial correlation matrix of three adjacent sensors of the same type as spatial constraints, combined with a sliding window time-series trend model, and interpolation errors are considered. .

[0043] To address noise data caused by transmission interference, a denoising method combining wavelet packet band energy weighting decomposition with an improved EMD (Empirical Mode Decomposition) that introduces a masking signal to suppress mode aliasing is employed. In this embodiment, wavelet packet band energy weighting is first used for initial screening, band segmentation, and coarse denoising. Then, improved EMD (Empirical Mode Decomposition) with an introduced masking signal is used for fine decomposition, mode aliasing suppression, fine denoising, and feature separation. This coarse band segmentation followed by fine screening is specifically adapted to complex noise data caused by transmission interference. The decomposition weights are obtained by normalizing the Pearson correlation coefficients between each band and the wear amount, and only the correlation with the wear amount's Pearson correlation coefficient is retained. The intrinsic mode function, the number of decomposition layers satisfies the signal-to-noise ratio. .

[0044] The system uses geological-equipment signal coupling verification to determine sudden changes in operating conditions. Simultaneously, a lightweight incremental learning unit learns new anomaly patterns in real time, updating local parameters associated with the current wear stage and anomaly type. In this embodiment, the sudden change in operating conditions is determined through geological-equipment signal coupling verification, where the product of the geological parameter change rate and the equipment signal anomaly degree is used. The time markers are used as disturbances, and new anomaly patterns are learned in real time through lightweight incremental learning units. Only local parameters associated with the current wear stage and anomaly type are updated, reducing the computational load by 75%.

[0045] Furthermore, such as Figure 4 As shown, the steps of the layered fusion process are as follows: Using the mechanical and operational parameters from the purified multi-source data as training samples, the mechanical and operational parameters are mapped to the same space to achieve cross-modal feature alignment. In this embodiment, a training sample set is constructed by collecting measured data from multiple operational parameters and wear stages, covering three geological operational parameters (Level 3 soft rock, Level 4 hard rock, and Level 5 hard rock) and three tool states (slight wear, moderate wear, and severe wear), totaling 1200 valid samples. 70% of these samples are used for training, 20% for validation, and 10% for testing. A cross-modal aligner is constructed using cross-modal feature alignment, and the aligner is trained based on the training sample set. The cosine similarity of the aligned features is ≥0.92.

[0046] Consistency verification is performed on mechanical parameter features and operating condition parameter features through deep feature extraction, combined with attention-weighted fusion and aligned cross-modal features. In this embodiment, the deep feature extraction consists of a 3-layer convolutional-pooling structure (3×3 kernel size, 2×2 pooling window) and a 2-layer bidirectional LSTM, with consistency coefficient... For example, local time-domain-frequency domain features and long-term evolution features are extracted, and cross-modal features after attention-weighted fusion alignment are combined to handle the uncertainty of cross-modal features based on DS evidence theory.

[0047] Furthermore, the generation process of the multi-dimensional fusion wear index is as follows: Based on measured data of different working condition parameters and wear stages in the characteristic parameters of TBM tool wear stages from the purified multi-source data, a three-dimensional mapping table of working condition-wear stage-feature weights is trained and generated. This table serves as the basis for generating the basic weights of working condition parameters and the corrected weights of wear stages in the two-factor weight model, thus generating dynamic weights. The calculation formula is: , in, For the first The basic weights of each feature based on its operating conditions; For the first Correction weights for each wear stage; For the first The cross-modal alignment coefficients of each feature; in this embodiment, the basic weights of the load features under level 5 hard rock conditions. The correction weight is 0.25 for the moderate wear stage. The cross-modal alignment coefficient of this feature is 1.2. The value is 0.95, and the dynamic weight of this feature is calculated. It is 0.285.

[0048] By combining normalized eigenvalues ​​and fusion residual compensation, a multi-dimensional fusion wear index is generated; the multi-dimensional fusion wear index The calculation formula is as follows: , in, These are supplementary coefficients for the normalized fusion residuals; The total number of features; For the first Normalized eigenvalues; dynamic weights After normalization, it satisfies The final generated The value range is [0, 1]. In this embodiment, using... Taking a value of 0.98 as an example, the current time is calculated. According to the corresponding grading and diagnostic rules, it is determined to be moderate wear.

[0049] Furthermore, based on a multi-dimensional fusion wear index Preset minimum multi-dimensional fusion wear index And the preset maximum multi-dimensional fusion wear index And the rising rate of the multi-dimensional integrated wear index change trend. Preset minimum ascent speed and preset maximum rate of ascent Wear grading diagnosis is performed; the wear grading diagnosis process is as follows: ,and If it is in a stable state, it indicates slight wear; ,or This indicates moderate wear. ,or If it is, then it is considered severe wear; in this embodiment, Furthermore, the trend of change is stable, indicating only slight wear; Or the rate of increase in the trend of change Moderate wear; Or the rate of increase in the trend of change It is severely worn.

[0050] It outputs real-time tool wear classification results, wear prediction values ​​after residual correction, and anomaly warning information. The anomaly warning information includes the anomaly type, occurrence time, impact range, and handling suggestions based on the current wear stage. It connects with the TBM tunneling control system and operation and maintenance management system through the linkage interface, providing accurate data support for the selection of TBM tool replacement timing and the adjustment of tunneling parameters, thereby reducing downtime and operation and maintenance risks.

[0051] This invention is not limited to the specific technical solutions described in the above embodiments. Besides the above embodiments, this invention may have other implementation methods. For those skilled in the art, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the protection scope of this invention.

Claims

1. A TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptive method, characterized in that, The steps are as follows: Using the spatiotemporal reference nodes deployed within the distributed sensor cluster, spatiotemporal deviation compensation and timestamp alignment are performed on the raw multi-source data collected in real time by each sensor during the TBM tool's working process. The multi-source data after spatiotemporal deviation compensation and timestamp alignment is then standardized and preprocessed, while the sampling frequency of the distributed sensor cluster is dynamically adjusted. Based on the signal dimensions of the multi-source data after standardized preprocessing and the characteristic parameters of the TBM tool wear stage, the correlation weight between the two is calculated. An adaptive threshold is calculated through a multi-factor three-dimensional adaptive threshold model. Based on the adaptive threshold, the abnormal type is identified and differentiated, thus obtaining the purified multi-source data. A cross-modal feature alignment enhancement fusion model is constructed, and hierarchical fusion processing is performed on the purified multi-source data to generate a multi-dimensional fusion wear index. Wear classification diagnosis is performed based on the multi-dimensional fusion wear index and its changing trend.

2. The TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptation according to claim 1, characterized in that, The process of spatiotemporal deviation compensation and timestamp alignment is as follows: The fixed non-rotating end of the TBM cutter head spindle is used as the spatial reference origin of the spatiotemporal reference node, and the industrial-grade synchronous clock of the TBM main control system is used as the time reference source of the spatiotemporal reference node. Based on the spatial reference origin and the time reference source, a unified reference for time synchronization and spatial positioning of all sensors is constructed. By acquiring the timestamp deviation data and spatial position deviation data of each sensor through the spatiotemporal reference node, a multidimensional spatiotemporal deviation compensation matrix is ​​constructed. By performing spatiotemporal deviation compensation on the raw multi-source data collected by each sensor through linear transformation, the spatial position deviation of the multi-source data after spatiotemporal deviation compensation is aligned with the timestamp deviation.

3. The TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptation according to claim 2, characterized in that, The standardization preprocessing process is as follows: Based on the multidimensional spatiotemporal deviation compensation matrix, signal amplitude calibration is performed on the multi-source data after spatiotemporal deviation compensation is completed, eliminating signal amplitude distortion caused by sensor hardware nonlinearity error, installation spatial position deviation and interference from the TBM tool tunneling environment. A linear normalization method is used to map the mechanical parameters and operating parameters with inconsistent dimensions and ranges of the multi-source data after signal amplitude calibration to a unified range, so as to eliminate the interference of dimensional differences on the correlation weight solution and multi-source feature fusion process.

4. The TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptation according to claim 1, characterized in that, The method for dynamically adjusting the sampling frequency of the distributed sensor cluster is as follows: Based on the original multi-source data, the corresponding geological grade and historical tool wear rate are determined. A coupled correlation model is constructed based on the variation patterns of the geological grade and historical tool wear rate, with the maximum value of the historical tool wear rate preset to be [value missing]. The minimum preset historical tool wear rate is , When the geological grade is hard rock and the historical tool wear rate At that time, adjust to high-frequency sampling; When the geological grade is hard rock and the historical tool wear rate At that time, adjust to high-frequency sampling; When the geological grade is soft rock and the historical tool wear rate When necessary, adjust to low-frequency sampling; When the geological grade is soft rock and the historical tool wear rate At that time, adjust to a lower frequency sampling.

5. The TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptation according to claim 1, characterized in that, The characteristic parameters of the TBM tool wear stage include: Pearson correlation coefficient of wear amount, sensitivity coefficient of wear stage, and prediction confidence of current wear stage; The correlation weights of each signal dimension of the standardized preprocessed multi-source data and the characteristic parameters of the TBM tool wear stage. The calculation is as follows: ， in, For the first Pearson correlation coefficients after dimension normalization of multi-source data signals; For the first The wear stage sensitivity coefficient after dimensional normalization of multi-source data signals; The confidence level for the current wear stage prediction; The total number of each signal dimension in the standardized preprocessed multi-source data; The prediction confidence level of the current wear stage The calculation formula is as follows: ， in, The number of sampling points; For the first Predicted tool wear value for each sample; For the first The actual value of tool wear for each sample.

6. The TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptation according to claim 1, characterized in that, The anomaly types include: sensor failure, transmission interference, and sudden change in operating conditions; the corresponding identification process is as follows: First, calculate the adaptive threshold using a multi-factor three-dimensional adaptive threshold model. The adaptive threshold The calculation formula is: ， in, , , For preset weighting coefficients and ; Geological grade; This refers to the tunneling speed; This is the wear and tear stage; This represents the historical anomaly incidence rate. This is the initial baseline value; Based on an adaptive threshold, suspected outliers are identified, and the following steps are performed on consecutive suspected outliers: Extract its time-domain features, frequency-domain energy distribution features, and spatial correlation features, and integrate the coupling relationship between geological parameters and TBM operating parameters to complete the anomaly type classification; If the signal amplitude of a single sensor output from a distributed sensor cluster exceeds the adaptive threshold for at least Q consecutive sampling points... Or a single sensor experiences a complete signal interruption, and If so, it is determined to be a sensor malfunction; If the signal amplitudes of multiple sensors output from a distributed sensor cluster synchronously generate high-frequency random fluctuations... and ,at the same time If so, it is determined to be transmission interference; If all sensors related to the TBM cutting tool's tunneling conditions simultaneously experience a step change in output signal amplitude... and ,at the same time If the changing trend of geological parameters is consistent with the predicted trend of geological parameters, it is determined to be a sudden change in working conditions. in, The standard deviation of the signal amplitude; This indicates the preset standard deviation threshold; Indicates the spatial correlation coefficient of sensors of the same type; This indicates the preset spatial correlation coefficient threshold; Represents the time-domain characteristic fluctuation coefficient; This represents the preset threshold for the time-domain characteristic fluctuation coefficient; The proportion of high-frequency energy representing the frequency domain energy distribution characteristics; This indicates the preset threshold for the proportion of energy in the high-frequency band; This indicates the consistency of signal trends among adjacent sensors of the same type. This indicates the preset signal trend consistency threshold; Indicates the rate of change of signal amplitude; This indicates the preset threshold for the rate of change of signal amplitude; Indicates duration; Indicates the preset duration threshold; This represents the product of the rate of change of the geological parameters of the TBM and the anomaly of the TBM signal. This indicates the preset product threshold.

7. The TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptation according to claim 6, characterized in that, The differential processing procedure for the aforementioned anomaly types is as follows: Using the spatial correlation matrix constructed from multiple adjacent sensors of the same type as spatial constraints, and combining it with a sliding window time series trend model, missing data for sensor faults is completed. To address noise data caused by transmission interference, a noise reduction method combining wavelet packet band energy weighting decomposition and improved EMD that introduces masking signals to suppress mode aliasing is adopted. The system uses geological-equipment signal coupling verification to determine the interference caused by sudden changes in operating conditions. At the same time, it uses a lightweight incremental learning unit to learn new abnormal patterns in real time and update local parameters associated with the current wear stage and abnormal type.

8. The TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptation according to claim 1, characterized in that, The steps of the layered fusion process are as follows: Using the mechanical parameter features and operating condition parameter features in the purified multi-source data as training samples, the mechanical parameter features and operating condition parameter features are mapped to the same space to achieve cross-modal feature alignment; Consistency verification is performed on mechanical parameter features and working condition parameter features. Through deep feature extraction, cross-modal features are fused and aligned using attention-weighted fusion.

9. The TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptation according to claim 1, characterized in that, The process of generating the multi-dimensional fusion wear index is as follows: Based on measured data of different working condition parameters and wear stages in the characteristic parameters of TBM tool wear stages from the purified multi-source data, a three-dimensional mapping table of working condition-wear stage-feature weights is trained and generated. This table serves as the basis for generating the basic weights of working condition parameters and the corrected weights of wear stages in the two-factor weight model, thus generating dynamic weights. The calculation formula is: ， in, For the first The basic weights of each feature based on its operating conditions; For the first Correction weights for each wear stage; For the first Cross-modal alignment coefficients for each feature; By combining normalized eigenvalues ​​and fusion residual compensation, a multi-dimensional fusion wear index is generated; the multi-dimensional fusion wear index The calculation formula is as follows: ， in, These are supplementary coefficients for the normalized fusion residuals; The total number of features; For the first The normalized eigenvalues.

10. The TBM tool wear monitoring method based on multi-source data fusion and anomaly adaptation according to claim 9, characterized in that, Based on multi-dimensional fusion wear index Preset minimum multi-dimensional fusion wear index And the preset maximum multi-dimensional fusion wear index And the rising rate of the multi-dimensional integrated wear index change trend. Preset minimum ascent speed and preset maximum rate of ascent Wear grading diagnosis is performed; the wear grading diagnosis process is as follows: ,and If it is in a stable state, it indicates slight wear; ,or This indicates moderate wear. ,or If so, it indicates severe wear.

Images