A method for correcting the wear coefficient of a shield cutter

Abstract

The application provides a wear coefficient correction method of a shield cutter, and relates to the technical field of wear monitoring.The method collects time sequence magnetic field signals obtained by a magnetic sensor array arranged on a non-rotating part of a shield machine, and synchronously collects cutter disc angle encoding signals; performs time-frequency transformation and filtering on the magnetic field signals to obtain noise reduction time sequence signals; determines the geometric position sequence of each shield cutter in combination with the angle encoding signals; constructs a theoretical magnetic field mapping between the cutters and the sensors based on the magnetic sensor pose, cutter position and magnetic dipole model, generates a system response matrix; reconstructs the independent magnetic feature vectors of each cutter to obtain an optimal feature set; generates a vector difference by comparing with a reference vector, inputs the wear identification model, obtains the wear amount and type identification of each cutter, and corrects the wear coefficient; the application improves the wear identification precision and reliability, and is suitable for tunnel construction under complex geological conditions.

Description

Technical Field

[0001] This application relates to the field of wear monitoring technology, and more specifically, to a method for correcting the wear coefficient of tunnel boring machine cutters. Background Technology

[0002] As the core equipment in tunnel construction, the tunnel boring machine (TBM) has multiple cutters mounted on its cutterhead that come into direct contact with the geological strata during excavation, enduring high-intensity friction and impact. This makes cutter wear a critical factor affecting construction efficiency and safety. Current technologies primarily rely on manual inspection during machine shutdowns or indirect methods based on vibration and acoustic signals. These methods suffer from poor real-time performance, high invasiveness, and low accuracy. In practical applications, the rotation of the cutterhead causes multiple cutter wear signals to overlap and mix, and these signals are further affected by noise and geological disturbances, making it difficult to separate the independent wear characteristics of each cutter from the composite signal. This results in insufficient accuracy in wear assessment and an inability to reliably correct the wear coefficient to optimize tunneling parameters. Therefore, accurately extracting the independent wear information of each cutter from the multiple cutters of the TBM to improve the accuracy and reliability of wear assessment and effectively correct the wear coefficient has become an urgent technical problem to be solved. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this application provides a method for correcting the wear coefficient of tunnel boring machine (TBM) cutters, comprising:

[0004] Acquire the time-series magnetic field signals collected by the magnetic sensor array deployed on the non-rotating parts of the tunnel boring machine within a preset sampling period; and simultaneously acquire the angle encoding signal associated with the rotation angle of the tunnel boring machine cutterhead within the sampling period;

[0005] The time-series magnetic field signal is subjected to time-frequency transformation to generate a time-frequency distribution matrix; the time-frequency distribution matrix is ​​filtered to generate a noise-reduced time-series signal; based on the angle-encoded signal, the geometric position sequence of each shield cutter configured on the cutterhead is determined as a function of time within the sampling period;

[0006] Based on the pose parameters of the magnetic sensor array, the geometric position sequence of each shield cutter, and the magnetic dipole field model, a theoretical magnetic field mapping relationship between the cutter sensors is established, and a system response matrix is ​​generated.

[0007] The noise reduction time series signal, the geometric position sequence, and the system response matrix are input into the regularized inverse problem solver to separate and reconstruct independent magnetic feature vectors for each shield cutter on the cutterhead, so as to form an optimal set of magnetic feature vectors.

[0008] Each independent magnetic feature vector in the optimal magnetic feature vector set is compared with its corresponding reference magnetic feature vector to obtain a first vector difference;

[0009] Based on the first vector difference, the current wear amount and wear type identifier of each tunnel boring machine cutter are generated using a wear identification model; based on the current wear amount and wear type identifier, the wear coefficient of each tunnel boring machine cutter is adjusted according to a preset update rule.

[0010] Optionally, filtering the time-frequency distribution matrix to generate a denoised time-series signal includes:

[0011] In the time-frequency distribution matrix, identify the noise frequency band corresponding to the preset motor power frequency and its harmonic frequencies;

[0012] A filtering operation is performed on the time-frequency distribution matrix to suppress the noise frequency band, and an inverse time-frequency transformation operation is performed to reconstruct and generate a denoised timing signal.

[0013] Optionally, the reconstructing and generating the denoised timing signal includes:

[0014] Based on the angle-encoded signal, the rotation frequency of the cutter head is determined; and the target harmonic frequency point, which is an integer multiple of the rotation frequency as the fundamental frequency, is determined in the time-frequency distribution matrix.

[0015] The target harmonic frequency points are divided into contaminated harmonic frequency points located within the noise frequency band and clean harmonic frequency points located outside the noise frequency band.

[0016] Extract the first spectral amplitude corresponding to all the clean harmonic frequency points in the time-frequency distribution matrix;

[0017] Based on the first spectral amplitude and its corresponding harmonic order, a harmonic attenuation model is constructed and fitted.

[0018] Input the harmonic order of each polluted harmonic frequency point into the harmonic attenuation model to generate the corresponding predicted spectral amplitude.

[0019] The predicted spectral amplitude is used to replace the original spectral data at the polluted harmonic frequency point in the time-frequency distribution matrix to generate a harmonic reconstruction time-frequency distribution matrix.

[0020] The inverse time-frequency transformation operation is performed on the harmonic reconstruction time-frequency distribution matrix to generate the noise-reduced timing signal.

[0021] Optionally, determining the cutter head rotation frequency based on the angle-encoded signal includes:

[0022] Perform a time-series resampling operation on the angle-encoded signal to generate a resampled angle sequence with equal time intervals;

[0023] Based on the resampled angle sequence, calculate the angle increment sequence per unit time.

[0024] Based on the aforementioned angle increment sequence, the number of rotations per unit time is determined;

[0025] The rotation frequency of the cutter head is determined based on the number of rotations.

[0026] Optionally, the time-frequency distribution matrix consists of multiple frequency units divided in the frequency dimension and multiple time frames divided in the time dimension. Each matrix element stores the original spectral amplitude at the corresponding time frame and frequency unit, wherein the original spectral data is the matrix element of the time-frequency distribution matrix at the frequency unit corresponding to the polluted harmonic frequency point.

[0027] Optionally, the generated system response matrix includes:

[0028] Read the pose parameters of the magnetic sensor array and the spatial coordinates of each shield cutter in the geometric position sequence;

[0029] Based on the magnetic dipole field model, the theoretical magnetic field contribution of each shield cutter to each magnetic sensor is calculated in each time frame within the sampling period.

[0030] Based on the correspondence between the shield cutter number and the magnetic sensor number, the theoretical magnetic field contribution is arranged as matrix elements to generate the system response matrix.

[0031] Optionally, forming the optimal set of magnetic eigenvectors includes:

[0032] Based on the system response matrix, a dynamic system response model is established by introducing a medium disturbance parameter to characterize the influence of unknown geological permeability.

[0033] An alternating iterative optimization process is performed to jointly estimate the independent magnetic eigenvectors and the medium perturbation parameters. The alternating iterative optimization process iteratively executes the following steps:

[0034] Under the condition of fixing the medium perturbation parameters, solve the first regularized optimization problem to update the independent magnetic eigenvectors;

[0035] Under the condition of fixing the updated independent magnetic eigenvectors, solve the second optimization problem to update the medium perturbation parameters;

[0036] The process is iterated until the estimation results of the independent magnetic eigenvectors and the medium perturbation parameters converge to a preset stability threshold.

[0037] The converged independent magnetic eigenvectors are determined as the final independent magnetic eigenvectors to form the optimal set of magnetic eigenvectors.

[0038] Optionally, generating the current wear amount and wear type identifier for each tunnel boring machine cutter includes:

[0039] The first vector difference is normalized to obtain the normalized vector difference;

[0040] The normalized vector difference is input into a multi-task deep neural network with a shared coding layer and a dual-branch decoder head, wherein the output of the first decoding branch of the dual-branch decoder head represents the continuous regression value of the current wear amount, and the output of the second decoding branch represents the discrete category probability of the wear type identifier.

[0041] The current wear amount and wear type identifier are determined based on the continuous regression value and the discrete category probability, respectively.

[0042] Optionally, determining the current wear amount and wear type identifier based on the continuous regression value and the discrete category probability respectively includes:

[0043] Calculate the maximum probability value among the discrete category probabilities, and take the corresponding index as the first candidate wear type identifier;

[0044] In response to the maximum probability value being greater than or equal to a preset threshold, the first candidate wear type identifier is output;

[0045] In response to the maximum probability value being less than the preset confidence threshold, based on the landing point of the continuous regression value in the preset wear level scale, the mapping rule is invoked to generate a second candidate wear type identifier, and the second candidate wear type identifier is output.

[0046] The continuous regression value is mapped to the wear level scale to obtain the corrected current wear amount, wherein when the second candidate wear type identifier is output, the corrected current wear amount is taken as the wear level value corresponding to the second candidate wear type identifier.

[0047] Optionally, adjusting the wear coefficient of each tunnel boring machine cutter based on the current wear amount and wear type identifier according to a preset update rule includes:

[0048] Read the wear coefficient and corresponding cumulative cutting distance recorded in the previous update cycle;

[0049] Using the wear coefficient and the cumulative cutting distance, the predicted wear amount for this update cycle is calculated according to the empirical wear model;

[0050] The wear error is obtained by comparing the difference between the predicted wear amount and the current wear amount.

[0051] Based on the wear type identifier, an applicable learning rate is selected from the learning rate mapping table, and the wear coefficient adjustment amount is generated by the learning rate, the wear error, and the cumulative cutting distance;

[0052] The wear coefficient is adjusted by increasing or decreasing the wear coefficient according to the wear coefficient adjustment amount to obtain the updated wear coefficient.

[0053] Apply upper and lower limit constraints to the updated wear coefficient to generate a corrected wear coefficient;

[0054] The corrected wear coefficient is written into the memory, replacing the wear coefficient recorded in the previous update cycle.

[0055] Compared with existing technologies, this invention can combine time-series magnetic field signals collected by a magnetic sensor array with angle-coded signals to achieve time-frequency transformation and filtering of the mixed signals, effectively suppressing noise interference and reconstructing the denoised signal. Furthermore, based on the magnetic dipole field model and the regularized inverse problem, the independent magnetic feature vector of each shield cutter is separated, thereby accurately quantifying the amount and type of wear, realizing dynamic adjustment of the wear coefficient, improving the real-time and non-contact nature of shield cutter wear monitoring, significantly enhancing the reliability of tunneling parameter optimization, reducing the risk of construction interruption, and extending the service life of the cutters, ultimately improving the overall efficiency and safety of tunnel engineering. Attached Figure Description

[0056] Figure 1 A flowchart illustrating a method for correcting the wear coefficient of a tunnel boring machine cutter, provided in an embodiment of this application;

[0057] Figure 2 A flowchart illustrating a method for generating denoised timing signals provided in this application embodiment;

[0058] Figure 3 This is a flowchart illustrating a method for reconstructing and generating denoised timing signals, as provided in an embodiment of this application. Detailed Implementation

[0059] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0060] See Figure 1 The diagram shows a flowchart of a method for correcting the wear coefficient of a tunnel boring machine cutter according to an embodiment of this application. The method includes steps S101 to S105, wherein:

[0061] S101: Acquire the time-series magnetic field signal collected by the magnetic sensor array deployed on the non-rotating parts of the tunnel boring machine within a preset sampling period; and simultaneously acquire the angle encoding signal associated with the rotation angle of the tunnel boring machine cutterhead within the sampling period;

[0062] S102: Perform time-frequency transformation on the time-series magnetic field signal to generate a time-frequency distribution matrix; filter the time-frequency distribution matrix to generate a noise-reduced time-series signal; based on the angle-encoded signal, determine the geometric position sequence of each shield cutter configured on the cutterhead as it changes with time within the sampling period;

[0063] S103: Based on the pose parameters of the magnetic sensor array, the geometric position sequence of each shield cutter, and the magnetic dipole field model, establish the theoretical magnetic field mapping relationship between the cutter sensors and generate the system response matrix;

[0064] S104: Input the noise-reduced timing signal, the geometric position sequence, and the system response matrix into the regularized inverse problem solver to separate and reconstruct independent magnetic feature vectors for each shield cutter on the cutterhead, so as to form an optimal magnetic feature vector set; compare each independent magnetic feature vector in the optimal magnetic feature vector set with the corresponding reference magnetic feature vector to obtain the first vector difference;

[0065] S105: Based on the first vector difference, use the wear identification model to generate the current wear amount and wear type identifier of each shield cutter; based on the current wear amount and wear type identifier, adjust the wear coefficient of each shield cutter according to the preset update rules.

[0066] Regarding the above S101:

[0067] In practical implementation, the magnetic sensor array is first fixedly mounted on non-rotating components of the tunnel boring machine (TBM), such as the main frame or the fixed support of the propulsion system. These components remain stationary during cutterhead rotation, ensuring that the sensors can stably capture the magnetic field changes caused by the cutterhead's rotation. The magnetic sensor array can consist of multiple magnetic sensors, such as a triaxial magnetometer as the basic unit. These sensors are evenly distributed in the array to cover the magnetic field influence area along the cutterhead's rotation path. For example, four magnetic sensors can be arranged in a rectangular array, with the array center aligned with the cutterhead axis, to improve the comprehensiveness of signal acquisition.

[0068] Next, a preset sampling period is set, for example, 1 second. Within this period, magnetic field signals are acquired at a fixed sampling rate, such as 100 sample points per second, thereby generating a magnetic field intensity sequence containing multiple time points as a time-series magnetic field signal. The specific acquisition process is implemented through a data acquisition module connected to the magnetic sensor array. This module can be a microcontroller-based embedded system that periodically triggers each magnetic sensor to read the current magnetic field intensity value, including components along the x, y, and z axes, and stores these values ​​as a vector sequence in chronological order.

[0069] For example, within a sampling period, if the sampling rate is 100 Hz, the time-series magnetic field signal will contain 100 magnetic field vectors corresponding to timestamps, each vector representing the magnetic field strength at that moment.

[0070] Simultaneously, to ensure synchronization with the magnetic field signal, the same clock source or synchronous triggering mechanism is used to acquire the angle-encoded signal. This signal comes from an angle encoder mounted on the cutter head shaft, such as an incremental photoelectric encoder, which outputs pulse signals or digital values ​​related to the cutter head rotation angle in real time. Within the same preset sampling period, the data acquisition module synchronously reads the output of the angle encoder. For example, whenever a magnetic field sample is acquired, the corresponding angle value is recorded simultaneously, thus forming an angle-encoded sequence aligned with the time-series magnetic field signal.

[0071] For example, if there are 100 samples of the magnetic field signal, the angle-encoded signal will also have 100 angle values, ensuring that each magnetic field measurement point corresponds to a precise cutter head rotation position. This synchronization can be achieved through hardware interrupts or software polling to avoid signal mismatch caused by time offset.

[0072] Regarding S102 above:

[0073] In practice, the first step is to perform a time-frequency transformation on the time-series magnetic field signal to generate a time-frequency distribution matrix. Specifically, a suitable time-frequency analysis method, such as the short-time Fourier transform, is selected to process the time-series magnetic field signal. This transformation decomposes the signal into representations in both time and frequency dimensions by applying a sliding window to the signal, thereby capturing the frequency components of the signal at different time points.

[0074] For example, if a time-series magnetic field signal contains 100 sample points, a Hamming window with a window length of 32 sample points can be used. The window can be slidable with a 50% overlap rate, and a Fourier transform can be performed on the signal segment within each window to calculate the amplitude spectrum, thus generating a time-frequency distribution matrix. The rows of this matrix correspond to frequency units, for example, from 0 Hz to half the sampling rate, divided into multiple frequency intervals; the columns correspond to time frames, for example, multiple time points generated based on the window sliding; each matrix element represents the signal amplitude value at a specific time frame and frequency unit. In this way, the time-frequency distribution matrix provides a dynamic spectral view of the magnetic field signal, facilitating subsequent noise identification and processing.

[0075] Secondly, the time-frequency distribution matrix is ​​filtered to generate a denoised timing signal. Specifically, noise components, such as frequency bands related to the motor's power frequency or other interference sources, are identified in the time-frequency distribution matrix. Then, filtering operations are applied to suppress this noise, such as using band-stop filters or adaptive filtering methods to attenuate the amplitude of the noise frequency band while preserving the spectral portion of the useful signal.

[0076] After filtering, an inverse time-frequency transform, such as an inverse short-time Fourier transform, is performed on the processed time-frequency distribution matrix. By converting the frequency domain information back to the time domain, the denoised time-series signal is reconstructed. This signal has the same length and structure as the original time-series magnetic field signal, but the noise level is significantly reduced.

[0077] For example, if the original matrix has 50 rows of frequency units and 60 columns of time frames, the inverse transform after filtering will generate a time-series vector sequence containing the corresponding number of sample points, ensuring the continuity and fidelity of the signal.

[0078] Finally, based on the angle-encoded signal, the geometric position sequence of each shield cutter configured on the cutterhead is determined as it changes over time within the sampling period.

[0079] In practice, the initial installation positions of each shield cutterhead on the cutterhead are first read, such as the radial distance and angular offset of each cutterhead relative to the center of the cutterhead. This positional information can be obtained from the design parameters or calibration data of the tunnel boring machine. Then, using the cutterhead rotation angle sequence provided by the angle encoding signal, the cumulative rotation of the cutterhead at each time point is calculated.

[0080] For example, if the angle encoding signal shows an angle increment of 90 degrees from the start time to a certain sample point, this increment is added to the initial angle of each tool to obtain the instantaneous angle position at that time point. Furthermore, combined with the geometric model of the tool head, for example assuming the tool head is a circular plane, the angle position is converted into spatial coordinates in the Cartesian coordinate system, such as x, y, and z coordinates, where the z coordinate can be considered fixed.

[0081] By repeating this calculation at all time points within the sampling period, a geometric position sequence is generated for each cutterhead, containing position vectors corresponding to multiple timestamps. For example, for a cutterhead with four cutters, if the sampling period has 100 time points, the geometric position sequence for each cutterhead will include 100 position vectors, ensuring that position changes are synchronized with the cutterhead rotation, thus providing a basis for subsequent magnetic field mapping.

[0082] Regarding the above S103:

[0083] In practice, the first step is to read the pose parameters of the magnetic sensor array. These parameters include the spatial position coordinates and orientation of each magnetic sensor, such as the x, y, and z coordinates and rotation angle in the tunnel boring machine coordinate system. This information can be obtained during the installation and calibration process or pre-stored in the system.

[0084] For example, if the magnetic sensor array has four sensors, the pose parameters will include four position vectors, each describing the sensor's three-dimensional coordinates and orientation, ensuring that the sensor's relative position to the tool turret is taken into account in subsequent calculations.

[0085] At the same time, the previously generated geometric position sequence of each shield cutter is invoked. This sequence provides a spatial coordinate sequence for each cutter at each time point within the sampling period. For example, for a cutter, the sequence may contain 100 position vectors, corresponding to 100 time frames of the sampling period.

[0086] Next, based on the magnetic dipole field model, a theoretical magnetic field mapping relationship between the cutting tool and the sensor is established. Specifically, the magnetic dipole field model treats each shield cutter as a magnetic dipole source, and the magnetic field generated by this source decays with distance and direction in space.

[0087] In the calculation, for each time frame within the sampling period, for each shield cutter and each magnetic sensor, the theoretical magnetic field contribution of the cutter to the sensor at its current geometric position is calculated. This can be achieved by considering the cutter position, sensor pose, and magnetic field propagation characteristics. For example, the relative distance and direction vector between the cutter and the sensor are first calculated, and then the magnetic field strength components, including contributions along the x, y, and z axes, are estimated according to the attenuation rules of the magnetic dipole model, without introducing the actual magnetic moment value as a variable, thus obtaining a theoretical magnetic field vector.

[0088] Repeat this calculation process, iterating through all time frames, all tunnel boring machine cutters, and all magnetic sensors, to generate a multidimensional set of theoretical contributions.

[0089] For example, if there are 4 cutting tools, 4 sensors, and 100 time frames, a set containing a large number of elements will be generated, each element representing the magnetic field contribution of a specific cutting tool to a specific sensor at a specific time frame.

[0090] Finally, these theoretical magnetic field contributions are organized into a system response matrix according to a preset arrangement rule. Specifically, the contribution values ​​are filled into matrix elements according to the shield cutter number as the row dimension and the magnetic sensor number combined with the time frame as the column dimension, or a similar structure, thus forming a matrix describing the magnetic field response of the entire system. This matrix captures the theoretical influence of the cutter motion on the sensor signals, providing a basis for subsequent signal separation.

[0091] For example, the generated matrix can be a matrix with the number of rows equal to the number of tools multiplied by some feature dimensions, and the number of columns corresponding to the number of sensors multiplied by the number of time frames, ensuring that the matrix structure facilitates linear operations.

[0092] Regarding S104 above:

[0093] In practice, the denoised time-series signal, geometric position sequence, and system response matrix are first fed as input to the regularized inverse problem solver. This solver is a computational module used to inverse the source parameters from the observation data, while suppressing noise and instability through regularization techniques.

[0094] Specifically, the denoised time-series signal is first organized into an observation vector. For example, the time-series values ​​of all magnetic sensors are concatenated into a long vector. If there are 4 sensors and 100 time frames, with each sensor providing three-axis components, the observation vector length could be 1200 elements. Simultaneously, the system response matrix serves as the transition matrix, describing the theoretical magnetic field mapping, while the geometric position sequence can be used to dynamically adjust matrix elements or verify position consistency during the solution process. For example, the correspondence between the position sequence and the matrix can be confirmed during iteration.

[0095] Next, within the solver, an optimization framework is established that aims to minimize the difference between the observed vector and the predicted vector after the system response matrix acts on the estimated magnetic eigenvector. At the same time, a regularization term is introduced to constrain the rationality of the estimated vector. For example, a penalty term is added to promote the smoothness of the vector or limit its magnitude to be too large, thereby avoiding overfitting or unstable solutions.

[0096] The specific process is implemented through iterative optimization algorithms, such as gradient descent or conjugate gradient methods, to progressively update the magnetic eigenvectors starting from the initial estimate. In each iteration step, the predicted magnetic field signal under the current estimate is calculated, compared with the observed signal, the estimate is adjusted to reduce the difference, and regularization is applied to balance accuracy and stability. Iteration continues until the residual is below a preset threshold or the maximum number of iterations is reached, for example, setting the threshold to one-hundredth of the norm of the observed signal and limiting the number of iterations to 100.

[0097] This process separates and reconstructs an independent magnetic eigenvector for each shield cutter on the cutterhead. This vector represents the magnetization characteristics of the cutter, for example, a vector containing multiple components that captures the direction and intensity of its magnetic moment without mixing with the signals from other cutters. For example, for four cutters, the solver will output four independent magnetic eigenvectors, each of which may have several dimensions depending on the model complexity. All these vectors are combined to form the optimal set of magnetic eigenvectors, which represents the optimized magnetic characteristic representation of the cutters.

[0098] Subsequently, each independent magnetic eigenvector in the optimal set of magnetic eigenvectors is compared with its corresponding reference magnetic eigenvector to obtain the first vector difference.

[0099] Specifically, the reference magnetic eigenvector is a pre-stored reference value, such as a magnetic eigenvector obtained during tool installation or initial calibration, and has the same structure as the independent magnetic eigenvector. The comparison process is achieved by element-wise subtraction. For example, for a vector, if the independent magnetic eigenvector is [0.5, 0.3, 0.2] and the reference is [0.4, 0.3, 0.1], then the first vector difference is [0.1, 0.0, 0.1]. This operation is repeated for all vectors in the set, generating a corresponding set of difference vectors for subsequent wear analysis.

[0100] Regarding the above S105:

[0101] In practice, firstly, based on the first vector difference, the wear identification model is used to generate the current wear amount and wear type identifier for each tunnel boring machine cutter.

[0102] Specifically, the wear identification model is a pre-trained computational model, such as a neural network-based classification and regression model, that infers the wear state by learning the correspondence between historical magnetic feature changes and wear data. First, the first vector difference is preprocessed, for example, by normalization, scaling its element values ​​to a uniform range, such as between 0 and 1, to match the model input requirements.

[0103] Furthermore, the processed vector difference is fed into the model, which includes a feature extraction layer to extract wear-related patterns from the vectors, for example, by processing the vector components through convolutional or fully connected layers. The model outputs two parts: a continuous numerical value representing the current wear amount, such as a floating-point number ranging from 0 to 100 mm, quantifying the physical wear degree of the tool; and a discrete category label, such as an integer label, like 0 for normal, 1 for slight wear, and 2 for severe wear, representing the wear type. This process is performed independently for the vector difference of each shield cutter; for example, for four cutters, four sets of outputs will be generated, ensuring that the wear information of each cutter is evaluated independently.

[0104] For example, if a vector difference is [0.1, 0.0, 0.1], after model processing, it might output a wear measurement of 5 mm and a type identifier of 1, indicating slight wear. This model can be trained on historical datasets, for example, using supervised learning methods. During the training phase, known wear vector differences and corresponding labels are input, and the accuracy is improved by iteratively optimizing the weights.

[0105] Secondly, based on the current wear amount and wear type identifier, the wear coefficient of each shield cutter is adjusted according to the preset update rules.

[0106] Specifically, the wear coefficient and corresponding cumulative cutting distance stored in the previous update cycle are first read. These values ​​can be obtained from system memory or a database; for example, the previous coefficient was 0.05 and the cumulative distance was 1000 meters. Then, the predicted wear amount is calculated using an empirical wear model. For example, the expected wear value is estimated based on the product of the coefficient and the distance, such as a predicted value of 50 millimeters. This predicted value is compared with the current wear amount, and the difference is calculated as the wear error. For example, if the current wear amount is 55 millimeters, the error is 5 millimeters.

[0107] Furthermore, adjustment parameters are selected from a preset mapping table based on the wear type identifier. For example, for type 1, a moderate learning rate such as 0.01 is chosen. Then, this parameter is used in conjunction with the error and cumulative distance to generate an adjustment amount, for example, by calculating the increase or decrease through a weighted product, such as the adjustment amount being the error multiplied by the learning rate and then by the distance factor. Finally, the original wear coefficient is corrected by applying this adjustment amount. For example, if the adjustment amount is positive, the coefficient is increased; otherwise, it is decreased. Upper and lower limits are imposed, such as the coefficient range being between 0.01 and 0.1, to avoid outliers. The updated coefficient is written to memory, replacing the old value, and the current cumulative distance is recorded in preparation for the next update.

[0108] For example, for a cutting tool, if the error is 5 mm, the learning rate is 0.01, and the distance factor is 0.001, the adjustment amount might be 0.00005, adjusting the original coefficient 0.05 to 0.05005, while ensuring it remains within the limit. This rule ensures that the coefficient dynamically adapts to actual wear, supporting the optimization of tunneling parameters.

[0109] Optional, see Figure 2 The flowchart below illustrates a method for generating a denoised timing signal according to an embodiment of this application, including steps S201 to S202, wherein:

[0110] S201: In the time-frequency distribution matrix, identify the noise frequency band corresponding to the preset motor power frequency and its harmonic frequencies;

[0111] S202: Perform a filtering operation on the time-frequency distribution matrix to suppress the noise frequency band, and perform an inverse time-frequency transformation operation to reconstruct and generate a noise-reduced timing signal.

[0112] In one optional embodiment, to address the technical problem that the periodic noise interference from the motor's power frequency and its harmonic frequencies during the rotation of the tunnel boring machine's cutterhead causes the useful magnetic field signal to be submerged, thereby reducing the noise reduction effect and the accuracy of subsequent wear assessment, this embodiment performs filtering by identifying and suppressing specific noise frequency bands in the time-frequency distribution matrix. This effectively removes motor-related noise, improves the signal-to-noise ratio and purity of the signal, and achieves more accurate noise reduction timing signal reconstruction, which is beneficial to improving the reliability and real-time performance of the overall wear coefficient correction.

[0113] Specifically, firstly, the noise frequency bands corresponding to the preset motor power frequency and its harmonic frequencies are identified in the time-frequency distribution matrix. The specific process includes reading the preset motor power frequency value, for example, the standard industrial motor power frequency is 50 Hz, and then calculating its harmonic frequencies, such as the first harmonic being 100 Hz, the second harmonic being 150 Hz, etc. Usually, the first few harmonics, such as the first three, are considered to cover the main interference sources.

[0114] Understandably, these frequency values ​​can be obtained from system configuration parameters or predefined according to the tunnel boring machine motor specifications. In the time-frequency distribution matrix, rows along the frequency dimension are traversed to locate frequency cells that match or are close to these frequencies. For example, if the matrix frequency resolution is 1 Hz, the rows corresponding to 50 Hz, 100 Hz, and 150 Hz are identified as noise bands. At the same time, a small bandwidth such as ±2 Hz is considered to tolerate slight offsets, thus forming a list of noise bands to ensure that the identification process covers the entire time frame dimension.

[0115] For example, if the time-frequency distribution matrix has 50 rows of frequency cells, with a frequency range from 0 Hz to half the sampling rate (50 Hz), assuming a sampling rate of 100 Hz, then the Nyquist frequency is 50 Hz. The noise band may include the 50th row (corresponding to 50 Hz, but due to Nyquist limitations, it may need to be adjusted to a value close to it) and its multiples. However, in practice, for the 50 Hz power frequency, it is necessary to identify its precise position in the matrix.

[0116] Secondly, filtering operations are performed on the time-frequency distribution matrix to suppress noise bands. Specifically, digital filtering techniques, such as band-stop filters, are used to apply attenuation operations to the matrix elements in the identified noise bands. This can be done by multiplying the amplitude values ​​of these elements by an attenuation factor less than 1, such as 0.1, to partially suppress noise, or by setting them to zero to completely remove it, depending on the noise intensity assessment. Simultaneously, the original values ​​of other frequency units are preserved to avoid over-filtering that could lead to the loss of useful signals.

[0117] This operation can be achieved by traversing the matrix row by row and column by column. For example, the software can process the rows corresponding to the noise frequency band in a loop and apply the filtering rules uniformly to the elements of all time frames to generate a filtered time-frequency distribution matrix.

[0118] After filtering, an inverse time-frequency transform is performed to reconstruct and generate a denoised time-series signal. Specifically, an inverse transform method corresponding to the time-frequency transform, such as the inverse short-time Fourier transform, is applied to the filtered time-frequency distribution matrix: first, the amplitude spectrum is converted to recapitulation form; then, the time-domain signal is reconstructed through overlapping addition, for example, using the same Hamming window and 50% overlap as the original window. The inverse Fourier transform results of each time frame are superimposed to generate a continuous time-series signal sequence. This signal has the same length as the original time-series magnetic field signal, for example, 100 sample points, ensuring consistent time resolution.

[0119] For example, after filtering a matrix containing 50 rows of frequency units and 60 columns of time frames, the inverse transform will output a vector sequence of approximately 100 sample points (depending on the window parameters). This sequence represents the magnetic field signal after removing motor noise, which facilitates the integration of subsequent geometric position sequences and the construction of the system response matrix.

[0120] Optional, see Figure 3 The flowchart below illustrates a method for reconstructing and generating a denoised timing signal according to an embodiment of this application, including steps S301 to S305, wherein:

[0121] S301: Determine the tool head rotation frequency based on the angle-encoded signal; determine the target harmonic frequency point that is an integer multiple of the rotation frequency as the fundamental frequency in the time-frequency distribution matrix;

[0122] S302: Divide the target harmonic frequency point into polluted harmonic frequency points located within the noise frequency band and clean harmonic frequency points located outside the noise frequency band;

[0123] S303: Extract the first spectral amplitude corresponding to all the clean harmonic frequency points in the time-frequency distribution matrix; construct and fit a harmonic attenuation model based on the first spectral amplitude and its corresponding harmonic order; input the harmonic order of each polluted harmonic frequency point into the harmonic attenuation model to generate the corresponding predicted spectral amplitude.

[0124] S304: In the time-frequency distribution matrix, the predicted spectral amplitude is used to replace the original spectral data at the polluted harmonic frequency point to generate a harmonic reconstruction time-frequency distribution matrix;

[0125] S305: Perform the inverse time-frequency transformation operation on the harmonic reconstruction time-frequency distribution matrix to generate the noise-reduced timing signal.

[0126] In one optional implementation, to address the technical problem that useful harmonic signals generated by the rotation of the cutter head may overlap with the motor noise frequency band during the time-frequency distribution matrix filtering process, leading to excessive suppression of these harmonics and loss of key magnetic field information, thereby affecting the reconstruction accuracy of the noise reduction timing signal and the reliability of subsequent wear assessment, this implementation performs reconstruction processing by identifying target harmonics based on the cutter head rotation frequency, dividing clean and contaminated frequency points, fitting an attenuation model, and predicting replacement of contaminated harmonics. This restores the spectral amplitude of useful harmonics within the noise frequency band, improves signal integrity and fidelity, achieves a more refined noise reduction effect, and enhances the robustness and accuracy of the overall wear coefficient correction method.

[0127] Specifically, firstly, the rotation frequency of the cutter head is determined based on the angle-encoded signal. This process involves analyzing the sequence values ​​of the angle-encoded signal, calculating the average rotational speed per unit time, for example, by dividing the total increment of angle change within a statistical sampling period by the period duration to obtain the rotational angle rate, and then converting it into a frequency value, such as the number of rotations per second. For instance, if the sampling period is 1 second and the total angle increment is 360 degrees, the rotational frequency is 1 Hz. This frequency serves as the fundamental frequency for subsequent harmonic identification, ensuring synchronization with the actual cutter head movement.

[0128] Secondly, target harmonic frequencies that are integer multiples of the fundamental frequency are determined in the time-frequency distribution matrix. Specifically, integer multiples of the fundamental frequency are calculated; for example, the first harmonic is the fundamental frequency itself, the second harmonic is twice the fundamental frequency, and so on. Typically, the first few orders, such as the first ten orders, are considered to cover the main signal components. Then, these frequencies are located along the frequency dimension of the matrix. For example, if the fundamental frequency is 1 Hz and the matrix frequency resolution is 0.5 Hz, the target harmonic frequencies are the corresponding frequency cell rows for 1 Hz, 2 Hz, 3 Hz, etc., forming a list of target frequencies.

[0129] For example, for a matrix with a fundamental frequency of 1 Hz, the target harmonic frequency may include elements corresponding to row number 2 (1 Hz), row number 4 (2 Hz), and row number 6 (3 Hz), assuming that the row number starts from 0 and the resolution is 0.5 Hz.

[0130] Next, the target harmonic frequencies are divided into contaminated harmonic frequencies located within the noise band and clean harmonic frequencies located outside the noise band. Specifically, each target harmonic frequency is compared with a previously identified list of noise bands. If the frequency falls within the noise band (e.g., 50 Hz ± 2 Hz), it is classified as contaminated; otherwise, it is classified as clean. For example, a conditional check is used to iterate through the list, assigning overlapping frequencies to the contaminated group and non-overlapping frequencies to the clean group, thus generating two sub-lists.

[0131] For example, if the target harmonics include 1 Hz (clean), 50 Hz (contaminated), and 100 Hz (contaminated), then the clean group is 1 Hz, and the contaminated groups are 50 Hz and 100 Hz.

[0132] Then, the first spectral amplitude corresponding to all clean harmonic frequency points in the time-frequency distribution matrix is ​​extracted. Specifically, for each clean frequency point, the time frame dimension of the matrix is ​​traversed, and all element values ​​of the corresponding frequency cell row are read. These values ​​represent the amplitude spectrum of that frequency point. Then, the average amplitude is calculated or the sequence is directly collected as the first spectral amplitude set.

[0133] For example, for a clean frequency point, extract a vector containing the amplitude values ​​of 60 time frames. If simplification is needed, the average of these values ​​can be taken as a single amplitude value.

[0134] For example, in a matrix with 60 columns of time frames, for a 1 Hz clean frequency point, extract the 60 amplitude values ​​of the corresponding elements in the row to form an amplitude sequence.

[0135] Based on the first spectral amplitude and its corresponding harmonic order, a harmonic attenuation model is constructed and fitted. Specifically, the order of the clean harmonics, such as 1, 3, and 5, is used as the independent variable, and the first spectral amplitude is used as the dependent variable. A suitable model form is selected, such as an exponential attenuation or power-law attenuation model. Then, the model parameters are fitted using the least squares method or similar optimization methods, for example, by iteratively adjusting the parameters to make the model's predicted values ​​close to the actual amplitudes. This model describes the law of amplitude attenuation with order; for example, higher-order harmonics have smaller amplitudes.

[0136] For example, for magnitudes of 0.8, 0.4, and 0.2 for orders 1, 3, and 5, a model in which the magnitude decreases exponentially with the order can be fitted, and the resulting parameters can be used for prediction.

[0137] The harmonic order of each polluted harmonic frequency is input into the harmonic attenuation model to generate the corresponding predicted spectral amplitude. Specifically, for each polluted frequency, its order (e.g., 2 or 4) is input into the fitting model, and the output amplitude is calculated. If the first spectral amplitude is a sequence, the prediction is also a corresponding sequence; if it is an average, a single value is generated. For example, for an order of 2, the model predicts an amplitude of 0.6.

[0138] For example, if the model is in exponential form, inputting order 4 will result in a predicted magnitude of 0.3.

[0139] In the time-frequency distribution matrix, the predicted spectral amplitude is used to replace the original spectral data at the polluted harmonic frequency points, generating a harmonic reconstruction time-frequency distribution matrix. Specifically, the list of polluted frequency points is traversed, and for each matrix row corresponding to a frequency point, the original element value is replaced with the predicted amplitude. For example, if the prediction is a single value, all time frames are replaced uniformly; if it is a sequence, each frame is replaced. After the replacement, the matrix retains other parts unchanged, forming the reconstruction matrix.

[0140] For example, for a 50 Hz contamination row, replace all 60 elements with a predicted amplitude of 0.6, or replace the sequence values ​​one by one.

[0141] Finally, an inverse time-frequency transform is performed on the harmonic reconstruction time-frequency distribution matrix to generate a denoised time-series signal. Specifically, methods such as the inverse short-time Fourier transform are used to transform the reconstruction matrix back to the time domain. For example, the signal sequence is reconstructed by performing an inverse Fourier transform on each time frame and overlapping windows, ensuring that the length is consistent with the original, such as 100 sample points. The useful harmonics have been recovered from the signal, and the noise has been suppressed. For example, after the inverse transform, the reconstruction matrix outputs a 100-point time-series vector, representing the optimized denoised signal.

[0142] Optionally, determining the cutter head rotation frequency based on the angle-encoded signal includes:

[0143] Perform a time-series resampling operation on the angle-encoded signal to generate a resampled angle sequence with equal time intervals;

[0144] Based on the resampled angle sequence, calculate the angle increment sequence per unit time.

[0145] Based on the aforementioned angle increment sequence, the number of rotations per unit time is determined;

[0146] The rotation frequency of the cutter head is determined based on the number of rotations.

[0147] In one optional embodiment, to address the technical problem that the angle-encoded signal may have unequal time intervals due to uneven encoder pulses or sampling clock jitter, resulting in deviations in the calculation of the cutter head rotation frequency and thus affecting the accurate identification of the target harmonic frequency and the reliability of noise reduction timing signal reconstruction, this embodiment accurately determines the cutter head rotation frequency by performing time-series resampling of the angle-encoded signal, calculating the angle increment, and determining the number of rotations. This achieves stable tracking of rotational dynamics, improves the accuracy and robustness of frequency estimation, and enhances the accuracy of harmonic reconstruction and the adaptability of the overall wear coefficient correction method.

[0148] Specifically, firstly, a time-series resampling operation is performed on the angle-encoded signal to generate a resampled angle sequence with equal time intervals. The process involves checking the timestamps of the original angle-encoded signal. If the intervals are uneven (e.g., some points are 0.009 seconds apart while others are 0.011 seconds), a fixed interval, such as 0.01 seconds, is selected as the target. Linear interpolation or nearest-neighbor interpolation methods are used to insert or adjust points in the original sequence to generate a sequence with uniform intervals. For example, by iterating through the timestamps of the original sequence, new time points at the target interval are calculated. Then, interpolation is performed based on the angle values ​​of adjacent original points to estimate the new angle values, ensuring that the resampled sequence covers the entire sampling period and maintains the continuity and monotonically increasing characteristics of the angles. This operation can be implemented by a signal processing software module, such as using an interpolation function library to process the sequence data.

[0149] For example, in an original sequence, if the timestamps of 0 seconds, 0.009 seconds, and 0.02 seconds correspond to angles of 0 degrees, 3 degrees, and 7 degrees, then when resampled to 0 seconds, 0.01 seconds, and 0.02 seconds, the new sequence may be 0 degrees, approximately 3.33 degrees (linear interpolation), and 7 degrees, forming an equally spaced sequence.

[0150] Secondly, based on the resampled angle sequence, an angle increment sequence per unit time is calculated. Specifically, the angle difference between adjacent points in the sequence is used as the increment. For example, for each consecutive pair of points, the angle value of the previous point is subtracted to obtain a positive increment. If there is a jump, such as exceeding 360 degrees, it is expanded using modulo operation or phase expansion to avoid negative values ​​or discontinuities. Simultaneously, considering unit time, such as per second or per interval, the increment is divided by the corresponding time difference. However, since the intervals are equal, the increment sequence can be directly used to represent the standardized change. This sequence reflects the fluctuations in the rotational speed, facilitating subsequent averaging processing.

[0151] Based on the angle increment sequence, the number of rotations per unit time is determined. Specifically, the total angle change is obtained by summing the increments of the entire sequence, and then divided by 360 degrees to calculate the number of rotations. For example, if the total increment is 720 degrees, the number of rotations is 2. Simultaneously, to handle noise, a smoothing filter such as a moving average can be applied to preprocess the increment sequence to ensure that the number of rotations reflects the average rotation behavior. This value represents the total amount of rotation within the sampling period.

[0152] For example, if the sum of the increment sequence is 1080 degrees, then the number of rotations is 3, indicating that it has rotated three times within the period.

[0153] The rotation frequency of the cutter head is determined based on the number of rotations. Specifically, the frequency value is obtained by dividing the number of rotations by the sampling period. For example, if the period is 1 second and the number of rotations is 2, the frequency is 2 Hz. This frequency serves as the fundamental frequency for subsequent harmonic correlation processing to ensure matching with the actual mechanical motion. If the period is long, it can be further averaged to improve stability.

[0154] For example, for a 1-second period and 3 rotations, the frequency is determined to be 3 Hz, providing accurate rotational dynamics metrics.

[0155] In one optional implementation, the time-frequency distribution matrix consists of multiple frequency units divided in the frequency dimension and multiple time frames divided in the time dimension. Each matrix element stores the original spectral amplitude at the corresponding time frame and frequency unit, wherein the original spectral data is the matrix element of the time-frequency distribution matrix at the frequency unit corresponding to the polluted harmonic frequency point.

[0156] The specific process involves setting the frequency dimension as rows when generating the matrix after time-frequency transformation. For example, each row corresponds to a different frequency unit from low frequency to high frequency, and each unit represents a frequency interval such as 1 Hz. The time dimension is set as columns, for example, each column corresponds to a time frame, which consists of time points generated by a sliding window. The values ​​of the matrix elements are calculated through Fourier transform, and the original amplitude or power spectrum values ​​are stored to represent the signal strength of the frequency unit within that time frame.

[0157] For example, in a matrix, if the frequency dimension has 50 units (rows) corresponding to 0 Hz to 49 Hz, and the time dimension has 60 frames (columns), then the matrix is ​​a two-dimensional array of 50 rows and 60 columns. The value of each element, such as the value of the 10th row and 20th column, represents the original spectral amplitude of the 20th time frame at 9 Hz (assuming a resolution of 1 Hz).

[0158] For contaminated harmonic frequencies, such as the frequency cell row corresponding to 50 Hz, all column elements are the original spectrum data. During subsequent replacements, the elements in that row are modified directly to ensure that the operation is targeted and does not affect neighboring cells.

[0159] Optionally, the generated system response matrix includes:

[0160] Read the pose parameters of the magnetic sensor array and the spatial coordinates of each shield cutter in the geometric position sequence;

[0161] Based on the magnetic dipole field model, the theoretical magnetic field contribution of each shield cutter to each magnetic sensor is calculated in each time frame within the sampling period.

[0162] Based on the correspondence between the shield cutter number and the magnetic sensor number, the theoretical magnetic field contribution is arranged as matrix elements to generate the system response matrix.

[0163] In one alternative implementation, to address the technical problem that in generating the system response matrix, the lack of explicit reading of pose parameters and spatial coordinates, as well as the lack of refinement in the calculation of magnetic field contribution and matrix arrangement, leads to inaccurate mapping relationships, thus affecting the stability of the inverse problem solution and the accuracy of independent magnetic eigenvector separation, this implementation generates the matrix by reading pose and coordinates, calculating theoretical contributions, and arranging the matrix according to numbers. This establishes an accurate theoretical magnetic field mapping, improves the representativeness and reliability of the matrix, and helps to improve the signal separation effect and the accuracy of the overall wear coefficient correction.

[0164] Specifically, first, the pose parameters of the magnetic sensor array and the spatial coordinates of each tunnel boring machine (TBM) cutter in the geometric position sequence are read. This process includes loading the pose parameters of the magnetic sensor array from system storage or a configuration file, such as the three-dimensional coordinates and orientation angle of each sensor, ensuring these parameters are defined in the TBM coordinate system. Simultaneously, the spatial coordinates of each TBM cutter are extracted from the geometric position sequence generated in the previous steps. This sequence is in time-series form, for example, each cutter has a list of coordinate vectors corresponding to a sampling period time frame.

[0165] By iterating through the tool numbers, the spatial coordinates of all tools across various time frames are collected, forming a structured dataset for subsequent calculations. Specifically:

[0166] For a magnetic sensor array, its pose parameters include the three-dimensional position coordinates (x, y, z) of each sensor in the tunnel boring machine coordinate system and the orientation angle theta, which are used to describe the sensor's installation position and orientation. For example, if there are 4 sensors in the array, their pose parameters can be represented as 4 quadruplet vectors, such as sensor 1: (x1, y1, z1, theta1).

[0167] For tunnel boring machine (TBM) cutters, their spatial coordinates are used to describe the three-dimensional position of each cutter within a specific time frame. Assuming there are four cutters and the sampling period is divided into 100 time frames, a 4×100 coordinate sequence table can be formed, where each element is a three-dimensional coordinate vector (x, y, z), representing the spatial position of a particular cutter within a given time frame. In other words, the rows of this table correspond to the cutter numbers, the columns correspond to the time frame numbers, and the cell content is the three-dimensional coordinate value of that cutter within that time frame.

[0168] Therefore, the pose parameters of the sensor array and the time-series coordinates of the tool are defined in numerical form, ensuring that they can be directly used for subsequent magnetic field calculations and matrix construction.

[0169] Secondly, based on the magnetic dipole field model, the theoretical magnetic field contribution of each shield cutter to each magnetic sensor is calculated in each time frame within the sampling period. Specifically, for each time frame, each shield cutter and each magnetic sensor is iterated. First, the relative position between the current coordinates of the cutter and the pose of the sensor is calculated, including the distance and direction vectors. Then, the magnetic field vector is estimated based on the field strength attenuation characteristics of the magnetic dipole model, for example, considering the effects of the inverse cube ratio of distance and the cosine of direction, to generate a three-dimensional magnetic field contribution vector without involving the actual magnetic moment. This calculation is repeated to cover all combinations, forming a contribution set.

[0170] For example, in a time frame, for tool 1 and sensor 1, if the relative distance is 1 meter and the direction is along the x-axis, the contribution may be biased towards the x-component; traverse 100 frames to generate a complete set.

[0171] Based on the correspondence between the tunnel boring machine cutter numbers and the magnetic sensor numbers, the theoretical magnetic field contributions are arranged as matrix elements to generate the system response matrix. Specifically, a matrix structure is selected, for example, rows correspond to the tunnel boring machine cutter numbers, or it is expanded to cutter multiplied by the feature dimension, and columns correspond to magnetic sensor numbers multiplied by the time frame. Then, the contribution values ​​are filled in sequentially; for example, the contribution of cutter 1 to all frames of all sensors is placed in the first row. Filling is completed through array operations or loop assignment to ensure that the matrix dimensions match the observed data for subsequent inverse problem solving.

[0172] For example, with 4 tools, 4 sensors, and 100 frames, the matrix could be 4 rows and 400 columns, with each element being a scalar contribution, or expanded to multiple dimensions to accommodate vector components, ensuring a linear representation of the system response.

[0173] The magnetic dipole field model is a commonly used physical model to describe the magnetic field distribution generated in space by a point magnetic source (called a magnetic dipole). This model assumes that the magnetic source has a fixed magnetic moment vector, representing the intensity and direction of the magnetic source.

[0174] In practical applications, such as in the scenario of monitoring tunnel boring machine cutters, each cutter can be approximated as a magnetic dipole. Its magnetic field decays with distance at the observation point far from the source and is affected by the position and orientation of the observation point relative to the source.

[0175] Specifically, this model considers the vector form of the magnetic field in space, including radial and tangential components, where the magnetic field strength is stronger near the source and decreases rapidly with increasing distance. When calculating the theoretical contribution of the tool to the magnetic sensor, this model establishes a mapping relationship by inputting the tool position, sensor pose, and magnetic moment direction, and outputting the corresponding magnetic field vector value. Based on the fundamental laws of electromagnetism, this model is suitable for non-contact magnetic field simulation under far-field approximation conditions, ensuring computational efficiency and compliance with physical laws, and facilitating integration into the generation of the system response matrix.

[0176] Optionally, forming the optimal set of magnetic eigenvectors includes:

[0177] Based on the system response matrix, a dynamic system response model is established by introducing a medium disturbance parameter to characterize the influence of unknown geological permeability.

[0178] An alternating iterative optimization process is performed to jointly estimate the independent magnetic eigenvectors and the medium perturbation parameters. The alternating iterative optimization process iteratively executes the following steps:

[0179] Under the condition of fixing the medium perturbation parameters, solve the first regularized optimization problem to update the independent magnetic eigenvectors;

[0180] Under the condition of fixing the updated independent magnetic eigenvectors, solve the second optimization problem to update the medium perturbation parameters;

[0181] The process is iterated until the estimation results of the independent magnetic eigenvectors and the medium perturbation parameters converge to a preset stability threshold.

[0182] The converged independent magnetic eigenvectors are determined as the final independent magnetic eigenvectors to form the optimal set of magnetic eigenvectors.

[0183] In one alternative implementation, to address the technical problem that the disturbance of the magnetic field signal by the unknown geological permeability under complex geological conditions leads to inaccurate system response matrix, thereby affecting the stability and accuracy of independent magnetic eigenvector separation, this implementation introduces a medium disturbance parameter to establish a dynamic model and performs alternating iterative optimization of the joint estimation vector and parameters to form an optimal set of magnetic eigenvectors. This compensates for geological influences, improves the robustness and accuracy of the estimation, and helps to enhance the reliability of wear identification and the adaptability of the overall method.

[0184] Specifically, firstly, based on the system response matrix, a dynamic system response model is established by introducing a medium perturbation parameter to characterize the influence of unknown geological permeability. The process involves extending the original system response matrix, for example, by adding a medium perturbation parameter. This parameter can be a scalar value or a vector, representing the overall influence of the geological layer on magnetic field propagation, such as adjusting matrix elements with a multiplicative factor to simulate permeability variations. This model represents the observed signal as the result of the dynamic response acting on the magnetic eigenvector. Adding the perturbation term forms a more comprehensive equation framework, facilitating the handling of environmental uncertainties.

[0185] For example, if the original matrix describes the magnetic field contribution under ideal conditions, the dynamic model can simulate the geologically enhanced magnetic field by multiplying it by a perturbation parameter such as 1.05, ensuring that the model captures deviations from the actual scenario.

[0186] Secondly, an alternating iterative optimization process is performed to jointly estimate the independent magnetic eigenvectors and the medium perturbation parameters. This alternating iterative process cyclically executes the following steps: First, under the condition of fixed medium perturbation parameters, a first regularized optimization problem is solved to update the independent magnetic eigenvectors. Specifically, the dynamic response model is adjusted using the current perturbation parameters, and then the difference between the observed vector and the predicted vector is minimized, while adding regularization terms such as vector magnitude constraints. The magnetic eigenvectors of each tool are then progressively updated using an iterative algorithm such as gradient descent until this step converges.

[0187] For example, with the perturbation parameter fixed at an initial value of 1.0, the updated vector reduces the residual, and after several iterations, a temporary vector is obtained.

[0188] Then, with the updated independent magnetic eigenvectors fixed, a second optimization problem is solved to update the medium perturbation parameters. Specifically, the updated vectors are substituted into the model to minimize the overall residuals, and the perturbation parameters are adjusted through simple optimization methods such as least squares to make the model better fit the observed data.

[0189] For example, the prediction is calculated using the new vector, and the parameter is adjusted to 1.02 to further reduce the error.

[0190] The iteration continues until the estimated results of the independent magnetic eigenvectors and the medium perturbation parameters converge to a preset stability threshold. Specifically, the magnitude of the change in vectors and parameters after each iteration is checked. If the change is less than a preset threshold, such as a relative change of 0.01, the iteration stops. This typically requires 10 to 50 iterations, depending on the initial values ​​and data complexity.

[0191] For example, if the change in the previous loop vector is less than 0.01 and the change in the parameter is less than 0.005, it is considered to have converged.

[0192] The converged independent magnetic eigenvectors are determined as the final independent magnetic eigenvectors to form the optimal magnetic eigenvector set. Specifically, the vector set from the last iteration is extracted, with each vector corresponding to a tunnel boring machine cutter, for subsequent comparison and wear analysis, ensuring that the optimized set represents magnetic features unaffected by disturbances. For example, for four cutters, a set containing four vectors is formed, each vector having compensated for geological disturbances.

[0193] In one optional implementation, to address the technical problem that inconsistent vector difference scales lead to model input sensitivity and prediction bias when generating wear amount and type identifiers based on the first vector difference, thereby affecting the accuracy and stability of wear assessment, this implementation normalizes the vector difference, inputs it into a multi-task deep neural network, and determines the wear amount and type based on the output. This standardizes the input data, simultaneously processes regression and classification tasks, improves the robustness of prediction and multi-dimensional analysis capabilities, and helps to enhance the accuracy and applicability of wear coefficient correction.

[0194] Specifically, firstly, the first vector difference is normalized to obtain a normalized vector difference. This process involves calculating the statistical range of each element in the vector difference, for example, finding the maximum and minimum values ​​of all elements. Then, a min-max normalization method is applied element-wise, mapping each value to the interval between zero and one. This process ensures that the scales of different vector differences are consistent, facilitating model learning without altering the relative relationships.

[0195] For example, if a first vector difference contains several elements, and the range is from a smaller value to a larger value, then the normalized element values ​​are distributed in the standardized interval, forming an input suitable for the model.

[0196] Secondly, the normalized vector difference is input into a multi-task deep neural network with a shared encoding layer and a dual-branch decoding head. The first decoding branch outputs a continuous regression value representing the current wear amount, while the second decoding branch outputs a discrete category probability representing the wear type identifier. Specifically, the network includes a shared encoding layer, such as composed of multiple fully connected or convolutional layers, to extract common features from the input vector, such as magnetic feature variation patterns. It then branches into two decoding branches: the first branch uses a regression head, such as a linear layer, to output a continuous floating-point number as the wear amount estimate; the second branch uses a classification head, such as a layer with softmax activation, to output a probability vector corresponding to the probabilities of predefined categories such as normal, minor, and severe wear. The network is pre-trained on a historical dataset, and the learning of the two branches is balanced using a multi-task loss function.

[0197] For example, after inputting a normalized vector, the shared coding layer extracts feature vectors, the first branch outputs a continuous value representing the degree of physical wear, and the second branch outputs the probability distribution corresponding to different categories.

[0198] The training method for this multi-task deep neural network includes the following steps: First, prepare a historical dataset, which is collected from previous shield tunneling cutter monitoring experiments or simulations. This dataset includes multiple samples, each containing a normalized vector difference as input, the corresponding wear amount as a continuous label (e.g., a physical measurement value), and the wear type as a discrete label (e.g., a category index). The dataset is divided into training, validation, and test sets, for example, proportionally allocated to avoid overfitting.

[0199] Secondly, a multi-task loss function is defined, which weights and combines the losses from the regression branch and the classification branch to construct the overall optimization objective. The regression branch predicts continuous wear and tear, and its loss function can be the mean squared error (MSE) to measure the deviation between the predicted and actual values. The classification branch determines the wear type, and its loss function can be cross-entropy to evaluate the degree of matching between the predicted category distribution and the true labels. The weights of each branch loss can be set according to the importance of the task; for example, the weight of the regression branch can be increased to enhance the focus on the accuracy of wear and tear prediction.

[0200] Then, the supervised learning training process is performed: Network weights are initialized, for example, using random initialization or pre-trained weights; an optimizer, such as the Adam optimizer, is selected to gradually adjust the weights; the batch size is set, for example, processing a certain number of samples per batch; within each training epoch, the training set is traversed, the input is fed into a shared encoding layer to extract features, and then the output is calculated through two branches respectively, compared with the labels to calculate the loss, and the weights are updated through backpropagation. This process is repeated for multiple training epochs, for example, tens to hundreds of epochs, until the total loss stabilizes or reaches a minimum on the validation set.

[0201] Finally, the model performance is evaluated using a validation set, such as calculating the mean absolute error of regression and the accuracy of classification, and hyperparameters such as the learning rate or the number of layers are adjusted based on the results. The generalization ability is then confirmed on the test set to ensure the model can handle new vector differences. This training method ensures that the network simultaneously learns both quantified and categorical features of wear, improving its adaptability to complex magnetic field changes.

[0202] The current wear level and wear type identifier are determined based on continuous regression values ​​and discrete category probabilities, respectively. Specifically, the continuous regression value of the first branch is directly used as the current wear level, for example, its numerical value represents the degree of physical wear. For the second branch, the category index corresponding to the highest probability in the probability vector is used as the wear type identifier; for example, if a certain probability is significantly higher than others, it is identified as the corresponding wear type. This determination process is applied independently to the input of each tunnel boring machine cutter, ensuring that the output is directly usable.

[0203] For example, if the regression value is a moderate value, and a certain category's probability is dominant in the probability distribution, then the wear amount is that continuous value, and the type is identified by that dominant category, ensuring a comprehensive and efficient assessment.

[0204] In one optional implementation, to address the technical problem that when multi-task deep neural networks output discrete category probabilities, the uncertainty of the probability distribution may lead to low confidence in the type identifier, resulting in misjudgment of wear type and inconsistency between the current wear amount and the probability distribution, this implementation determines the type identifier by calculating the maximum probability and applying a confidence threshold, generating candidate identifiers using backup mapping rules, and correcting the wear amount. This improves the reliability of the type identifier and the consistency of the wear amount, enhances the robustness and accuracy of the overall assessment, and facilitates the optimization of the adaptability of the wear coefficient adjustment.

[0205] Specifically, first, the maximum probability value in the discrete category probability vector is calculated, and the corresponding index is taken as the first candidate wear type identifier. The process involves iterating through the probability vector elements, finding the element with the largest value as the maximum probability, and then recording its position index in the vector as the candidate identifier. For example, if the vector corresponds to multiple predefined categories, this index represents a specific wear type such as normal or severe. This step ensures that the classification result that the network is most confident in is given priority.

[0206] For example, for a probability vector, if the value of a certain element is significantly higher than the others, its index is selected as the first candidate, representing the initial assessment of the wear type.

[0207] In response to the maximum probability value being greater than or equal to a preset confidence threshold, the first candidate wear type identifier is output. Specifically, the maximum probability is compared with a preset threshold, such as a value at a moderate level. If the condition is met, the first candidate wear type identifier is directly adopted as the final type identifier, ensuring rapid decision-making under high confidence conditions.

[0208] For example, if the maximum probability is higher than the threshold, it indicates that the classification is reliable, and the type identifier of the corresponding index is output.

[0209] In response to a maximum probability value being less than a preset confidence threshold, the system uses the location of the continuous regression value on a preset wear level scale to generate a second candidate wear type identifier based on mapping rules, and outputs this second candidate wear type identifier. Specifically, if the threshold is not met, indicating classification uncertainty, a preset scale, such as a segmented wear range, is referenced to locate the continuous regression value to the corresponding segment. Then, mapping rules, such as interval-corresponding categories, are applied to determine the segment's type as the second candidate, which is then used as the final output. This rule provides a backup mechanism to avoid errors in low-confidence classification.

[0210] For example, if the maximum probability is below a threshold, the regression value falls into a lower wear segment, is mapped to the normal type as a second candidate, and is output.

[0211] The continuous regression values ​​are mapped to a wear level scale to obtain the corrected current wear amount. When a second candidate wear type identifier is output, the corrected current wear amount is taken as the wear level value corresponding to the second candidate wear type identifier. Specifically, in either case, the regression values ​​are projected onto the scale, for example, by finding the interval to which they belong and taking the representative value of the interval, such as the median, as the correction amount. Especially when using a second candidate, to maintain consistency, a fixed level value associated with that type is directly used as the correction amount to ensure that the amount and type are consistent.

[0212] For example, if the first candidate is output directly, the correction amount is the projection of the regression value onto the scale; if the second candidate is output, the correction amount is the preset typical level value for that type, representing the adjusted wear level.

[0213] In one alternative implementation, to address the technical problem that in adjusting the wear coefficient, the lack of integration of historical records and wear type-specific parameters leads to inaccurate predictions and excessive fluctuations in corrections, thereby affecting the stability and safety of tunnel boring machine (TBM) excavation parameter optimization, this implementation adjusts the coefficient by reading data from the previous cycle, predicting wear using an empirical wear model, calculating errors, generating adjustment amounts based on the type and selecting a learning rate, performing corrections and constraints, and storing updates. This achieves adaptive and robust coefficient optimization, improves the accuracy and consistency of corrections, and helps reduce construction risks and extend tool life.

[0214] Specifically, first, the wear coefficient and corresponding cumulative cutting distance recorded in the previous update cycle are read. This process involves retrieving the output data of the previous correction operation from the system memory or database, for example, locating the record by keyword or timestamp, ensuring that the read value corresponds to a specific tunnel boring machine cutter. The coefficient represents an estimate of the wear rate in the previous cycle, and the cumulative distance represents the total tunneling length from initial installation to the current cycle, used to track long-term wear accumulation.

[0215] For example, for a tunnel boring machine cutter, if the previous cycle record shows a medium level coefficient and a relatively long cumulative distance, these values ​​are used as basic inputs to ensure continuous updates.

[0216] Using the wear coefficient and the cumulative cutting distance, the predicted wear amount for the current update cycle is calculated according to an empirical wear model. Specifically, the empirical wear model is a simplified relationship based on historical observations. For example, a linear model multiplies the wear coefficient by the cumulative distance or its increment to estimate the expected wear level. This model is derived from engineering practice data and is used for rapid prediction without complex simulations. Alternatively, an exponential model uses the coefficient as a decay rate, combined with the distance to calculate the progressive wear value, to account for nonlinear cumulative effects. Those skilled in the art can choose an appropriate form based on the geological conditions of the tunnel boring machine. By inputting the read values, the predicted amount is calculated, representing the expected wear if there are no new changes.

[0217] For example, if the coefficients represent wear rate per unit distance, the model can obtain a moderate prediction value by product, reflecting historical trends; if an exponential form is used, a gradually varying prediction value that takes into account cumulative distance can be generated, representing long-term wear saturation.

[0218] The wear error is obtained by comparing the difference between the predicted wear amount and the current wear amount.

[0219] Specifically, the predicted amount is directly subtracted from the current amount (obtained from previous steps), and the absolute value or signed difference is taken as the error. This error quantifies the deviation between the actual and expected values ​​and is used to guide the direction and magnitude of adjustments.

[0220] For example, if the current quantity is higher than the prediction, the error is positive, indicating accelerated wear and tear, and a corresponding correction is needed.

[0221] Based on the wear type identifier, an applicable learning rate is selected from the learning rate mapping table, and the wear coefficient adjustment amount is generated by the learning rate, the wear error, and the cumulative cutting distance.

[0222] Specifically, the process begins by querying a pre-defined mapping table that associates type identifiers with learning rates. For example, a low rate corresponds to a normal type, a medium rate to a slight type, and a high rate to a severe type, to accommodate different levels of wear severity. Then, the selected learning rate is combined with the error and distance, for example, by calculating the adjustment amount through a product or weighted method. A common weighted method is to multiply the learning rate by the wear error and then by a distance factor, such as the normalized value obtained by dividing the cumulative distance by the standard reference distance. This ensures that the adjustment amount reflects the magnitude of the error and the impact of historical distances.

[0223] For example, for mild cases, a moderate learning rate is chosen, and a moderate positive adjustment is generated with the positive error and long distance, indicating that the coefficients need to be increased slightly.

[0224] The wear coefficient is adjusted by the specified adjustment amount to obtain an updated wear coefficient. Specifically, the adjustment amount is added to or subtracted from the original coefficient, depending on its sign. For example, a positive adjustment increases the coefficient to reflect faster wear, while a negative adjustment decreases it, ensuring that the update reflects the latest error.

[0225] For example, if the adjustment amount is a medium value and the original coefficient is medium, the updated coefficient will be slightly higher, indicating the corrected rate.

[0226] The updated wear coefficient is subjected to upper and lower limit constraints to generate a corrected wear coefficient. Specifically, the upper and lower limit ranges are preset. For example, the lower limit is based on engineering experience to avoid negative values, and the upper limit is to prevent excessive amplification. Then, the updated coefficient is compared. If it exceeds the limit, the nearest limit value is truncated to ensure that the coefficient value is within the physically permissible range and to maintain the stability of the system calculation.

[0227] For example, if the updated coefficient exceeds the upper limit, it is corrected to the upper limit value to prevent unrealistic adjustments.

[0228] The corrected wear coefficient is written to memory, replacing the wear coefficient recorded in the previous update cycle. Specifically, a write operation is used to overwrite the storage location, while the associated cumulative distance is updated (adding the current cycle increment), ensuring that the next read uses the latest value, and possibly recording a timestamp to track history.

[0229] For example, after a write operation, the memory holds a moderate correction factor and update distance, providing a basis for continuous optimization.

[0230] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for correcting the wear coefficient of tunnel boring machine cutters, characterized in that, include: Acquire time-series magnetic field signals collected within a preset sampling period by a magnetic sensor array deployed on a non-rotating component of a tunnel boring machine; Simultaneously acquire the angle-encoded signal associated with the rotation angle of the tunnel boring machine cutterhead within the sampling period; The time-series magnetic field signal is subjected to time-frequency transformation to generate a time-frequency distribution matrix; the time-frequency distribution matrix is ​​filtered to generate a noise-reduced time-series signal; based on the angle-encoded signal, the geometric position sequence of each shield cutter configured on the cutterhead is determined as a function of time within the sampling period; Based on the pose parameters of the magnetic sensor array, the geometric position sequence of each shield cutter, and the magnetic dipole field model, a theoretical magnetic field mapping relationship between the cutter sensors is established, and a system response matrix is ​​generated. The noise reduction time series signal, the geometric position sequence, and the system response matrix are input into the regularized inverse problem solver to separate and reconstruct independent magnetic feature vectors for each shield cutter on the cutterhead, so as to form an optimal set of magnetic feature vectors. Each independent magnetic feature vector in the optimal magnetic feature vector set is compared with its corresponding reference magnetic feature vector to obtain a first vector difference; Based on the first vector difference, the current wear amount and wear type identifier of each tunnel boring machine cutter are generated using a wear identification model; based on the current wear amount and wear type identifier, the wear coefficient of each tunnel boring machine cutter is adjusted according to a preset update rule.

2. The method for correcting the wear coefficient of a tunnel boring machine cutter according to claim 1, characterized in that, The step of filtering the time-frequency distribution matrix to generate a denoised time-series signal includes: In the time-frequency distribution matrix, identify the noise frequency band corresponding to the preset motor power frequency and its harmonic frequencies; A filtering operation is performed on the time-frequency distribution matrix to suppress the noise frequency band, and an inverse time-frequency transformation operation is performed to reconstruct and generate a denoised timing signal.

3. The method for correcting the wear coefficient of a tunnel boring machine cutter according to claim 2, characterized in that, The reconstructing and generating of the denoised timing signal includes: Based on the angle-encoded signal, the rotation frequency of the cutter head is determined; and the target harmonic frequency point, which is an integer multiple of the rotation frequency as the fundamental frequency, is determined in the time-frequency distribution matrix. The target harmonic frequency points are divided into contaminated harmonic frequency points located within the noise frequency band and clean harmonic frequency points located outside the noise frequency band. Extract the first spectral amplitude corresponding to all the clean harmonic frequency points in the time-frequency distribution matrix; Based on the first spectral amplitude and its corresponding harmonic order, a harmonic attenuation model is constructed and fitted. Input the harmonic order of each polluted harmonic frequency point into the harmonic attenuation model to generate the corresponding predicted spectral amplitude. The predicted spectral amplitude is used to replace the original spectral data at the polluted harmonic frequency point in the time-frequency distribution matrix to generate a harmonic reconstruction time-frequency distribution matrix. The inverse time-frequency transformation operation is performed on the harmonic reconstruction time-frequency distribution matrix to generate the noise-reduced timing signal.

4. The method for correcting the wear coefficient of a tunnel boring machine cutter according to claim 3, characterized in that, Determining the cutter head rotation frequency based on the angle-encoded signal includes: Perform a time-series resampling operation on the angle-encoded signal to generate a resampled angle sequence with equal time intervals; Based on the resampled angle sequence, calculate the angle increment sequence per unit time. Based on the aforementioned angle increment sequence, the number of rotations per unit time is determined; The rotation frequency of the cutter head is determined based on the number of rotations.

5. The method for correcting the wear coefficient of a tunnel boring machine cutter according to claim 3, characterized in that, The time-frequency distribution matrix consists of multiple frequency units divided in the frequency dimension and multiple time frames divided in the time dimension. Each matrix element stores the original spectral amplitude at the corresponding time frame and frequency unit. The original spectral data is the matrix element of the time-frequency distribution matrix at the frequency unit corresponding to the polluted harmonic frequency point.

6. The method for correcting the wear coefficient of a tunnel boring machine cutter according to claim 1, characterized in that, The generated system response matrix includes: Read the pose parameters of the magnetic sensor array and the spatial coordinates of each shield cutter in the geometric position sequence; Based on the magnetic dipole field model, the theoretical magnetic field contribution of each shield cutter to each magnetic sensor is calculated in each time frame within the sampling period. Based on the correspondence between the shield cutter number and the magnetic sensor number, the theoretical magnetic field contribution is arranged as matrix elements to generate the system response matrix.

7. The method for correcting the wear coefficient of a tunnel boring machine cutter according to claim 1, characterized in that, The formation of the optimal magnetic eigenvector set includes: Based on the system response matrix, a dynamic system response model is established by introducing a medium disturbance parameter to characterize the influence of unknown geological permeability. An alternating iterative optimization process is performed to jointly estimate the independent magnetic eigenvectors and the medium perturbation parameters. The alternating iterative optimization process iteratively executes the following steps: Under the condition of fixing the medium perturbation parameters, solve the first regularized optimization problem to update the independent magnetic eigenvectors; Under the condition of fixing the updated independent magnetic eigenvectors, solve the second optimization problem to update the medium perturbation parameters; The process is iterated until the estimation results of the independent magnetic eigenvectors and the medium perturbation parameters converge to a preset stability threshold. The converged independent magnetic eigenvectors are determined as the final independent magnetic eigenvectors to form the optimal set of magnetic eigenvectors.

8. The method for correcting the wear coefficient of a tunnel boring machine cutter according to claim 1, characterized in that, The generation of the current wear amount and wear type identifier for each tunnel boring machine cutter includes: The first vector difference is normalized to obtain the normalized vector difference; The normalized vector difference is input into a multi-task deep neural network with a shared coding layer and a dual-branch decoder head, wherein the output of the first decoding branch of the dual-branch decoder head represents the continuous regression value of the current wear amount, and the output of the second decoding branch represents the discrete category probability of the wear type identifier. The current wear amount and wear type identifier are determined based on the continuous regression value and the discrete category probability, respectively.

9. The method for correcting the wear coefficient of a tunnel boring machine cutter according to claim 8, characterized in that, The step of determining the current wear amount and wear type identifier based on the continuous regression value and the discrete category probability includes: Calculate the maximum probability value among the discrete category probabilities, and take the corresponding index as the first candidate wear type identifier; In response to the maximum probability value being greater than or equal to a preset threshold, the first candidate wear type identifier is output; In response to the maximum probability value being less than the preset confidence threshold, based on the landing point of the continuous regression value in the preset wear level scale, the mapping rule is invoked to generate a second candidate wear type identifier, and the second candidate wear type identifier is output. The continuous regression value is mapped to the wear level scale to obtain the corrected current wear amount, wherein when the second candidate wear type identifier is output, the corrected current wear amount is taken as the wear level value corresponding to the second candidate wear type identifier.

10. The method for correcting the wear coefficient of a tunnel boring machine cutter according to claim 9, characterized in that, The adjustment of the wear coefficient of each tunnel boring machine cutter based on the current wear amount and wear type identifier, according to a preset update rule, includes: Read the wear coefficient and corresponding cumulative cutting distance recorded in the previous update cycle; Using the wear coefficient and the cumulative cutting distance, the predicted wear amount for this update cycle is calculated according to the empirical wear model; The wear error is obtained by comparing the difference between the predicted wear amount and the current wear amount; Based on the wear type identifier, an applicable learning rate is selected from the learning rate mapping table, and the wear coefficient adjustment amount is generated by the learning rate, the wear error, and the cumulative cutting distance; The wear coefficient is adjusted by increasing or decreasing the wear coefficient according to the wear coefficient adjustment amount to obtain the updated wear coefficient. Apply upper and lower limit constraints to the updated wear coefficient to generate a corrected wear coefficient; The corrected wear coefficient is written into the memory, replacing the wear coefficient recorded in the previous update cycle.

Images