A method for monitoring abnormal wear of shield machine disc cutters

Abstract

The present invention provides a method for monitoring abnormal wear of disc cutters of shield machines. According to the demand for monitoring abnormal wear of disc cutters during the tunneling process of the shield machine, the tunneling parameters closely related to the wear status of the disc cutters are screened and the data missing due to acquisition problems are filled. The long short-term memory autoencoder model based on the abnormal attention mechanism uses a learnable Gaussian kernel to obtain the prior association in the original sequence, and improves the distinguishability between normal and abnormal by obtaining the correlation difference of the sequence. Combined with the extreme value theory, the abnormal wear of the disc cutters of the shield machine is accurately and reliably monitored. The method of the present invention more accurately identifies the tunneling parameters closely related to the operating status of the disc cutters, and improves the model training accuracy and the accuracy of the abnormal wear monitoring of the cutters. The present invention is of great significance to the application of abnormal wear monitoring of disc cutters of shield machines.

Description

Technical field

[0001] The present invention relates to the field of tool wear monitoring, and in particular to a method for monitoring abnormal wear of a shield machine disc cutter. [Background Technology]

[0002] The shield method refers to tunnel excavation using a shield machine. Due to its advantages such as high automation, rapid construction speed, and high safety, it is widely used in tunneling. As a key component of the shield machine, the shield cutter performs the primary rock-breaking work in shield tunneling projects. Abnormal cutter wear increases the load on other cutters on the cutterhead and, in severe cases, can even render the entire cutterhead useless. Researching intelligent monitoring methods for abnormal shield cutter wear has important practical implications for guiding cutter replacement and shortening construction schedules.

[0003] Disc cutters are mounted on the cutterhead of a shield machine and perform the primary earthmoving task during shield tunneling. During tunneling, the cutterhead presses against the excavation surface and slowly rotates along the normal to the excavation surface. This causes the disc cutters to rotate about their own axes and orbit around the cutterhead's central bearing, breaking up and removing the soil from the excavation surface. Disc cutter wear can be categorized as normal wear or abnormal wear. To improve construction efficiency, real-time cutter condition monitoring technology has emerged. Typically, such methods use sensors such as magnetoresistance, ultrasonic monitoring, eddy current monitoring, and acoustic emission monitoring to directly monitor cutter wear, helping operators determine the optimal time to change cutters. However, sensors mounted on the cutterhead are affected by harsh operating environments such as high temperature, high humidity, and strong vibration, resulting in low reliability for direct sensor monitoring methods.

[0004] In recent years, data-driven methods for monitoring disc cutter wear have received widespread attention. In particular, deep learning unsupervised anomaly monitoring models driven by normal data have been widely used. For example, the combination of an attention mechanism and a long short-term memory neural network (LSTM) algorithm has a certain effect on abnormal wear monitoring of disc cutters. However, in practical engineering problems, due to the scarcity of anomalies in time series, it is extremely difficult to associate anomalies with the entire sequence. Existing methods for monitoring abnormal wear of shield machine disc cutters can improve the performance of anomaly monitoring to a certain extent. However, due to the scarcity of disc cutter anomalies, it is extremely difficult to associate anomalies with the entire sequence. As a result, the existing attention mechanism has not effectively improved the distinguishability of abnormal states.

[0005] The above content is only used to assist in understanding the technical solution of the present invention and does not constitute an admission that the above content is prior art. [Summary of the invention]

[0006] To solve the above problems, the present invention proposes a method for monitoring abnormal wear of disc cutters of shield machines. The tunneling parameters of the shield machine are screened and missing values ​​are filled in during the tunneling process of the shield machine. The optimized tunneling parameter data is used for model training, and a long short-term memory autoencoder network model based on the abnormal attention mechanism is constructed. The extreme value theory is combined to perform abnormal monitoring. The method of the present invention improves the accuracy and reliability of abnormal wear monitoring of disc cutters of shield machines.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A method for monitoring abnormal wear of a shield machine disc cutter, characterized by comprising:

[0009] Step S1: Collect various tunneling parameter data during shield machine excavation, and determine whether it is overall missing data or overall unchanged data based on its data length and data variance, and delete the parameters of overall missing data and / or overall unchanged data; in this way, the tunneling parameter data can be optimized, the model training accuracy and the accuracy of abnormal cutter wear monitoring can be improved.

[0010] Step S2: using a screening strategy to screen tunneling parameters that are closely related to the operating status of the disc cutter;

[0011] Preferably, the screening strategy is a strategy combining Pearson and Spearman correlation coefficients for screening, and the screening strategy is as follows:

[0012]

[0013]

[0014]

[0015] In the formula, f represents the number of model training and case test sets used, are Pearson correlation coefficient and Spearman correlation coefficient, respectively. Represents the standard deviation of two random variables a and b, cov(a,b) is the covariance matrix of variables a and b, Y i It represents the difference in the rank value of the i-th data pair, N is the total number of observation samples of this type of tunneling parameters, ε, η are weight coefficients, 0≤ε, η≤1, the sum of ε, η is equal to 1, and correlation is the total correlation coefficient.

[0016] Through the screening strategy combining the two correlation coefficients, the trend characteristics of the data are retained, and the weight coefficient is adjusted according to the actual working conditions of the shield machine disc cutter, which can more accurately identify the excavation parameters closely related to the operating status of the disc cutter.

[0017] Step S3: Normalize the filtered tunneling parameter data and fill in the missing rows with the mean of the data in the rows before and after the missing rows. This can fill the data gaps and improve the data quality. The original sequence sample x of size [U×C] is intercepted through the sliding window. By filling in the missing rows, the integrity and continuity of the data are maintained, and the accurate training and prediction effects of the network model are improved.

[0018] Fill in the empty data by filling in the data. The data filling method is as follows:

[0019]

[0020] Among them, x Nan is the missing value of the data, The data before the missing row is not empty in the first five rows. They are the 5 rows of data that are not null values ​​after the row with missing data, and avg means finding the average value of the data.

[0021] The filled data is sampled through a sliding window and divided into model training set and case test set.

[0022] Step S4: Constructing a long short-term memory autoencoder network model with improved abnormal attention mechanism Set the hyperparameters of the abnormal attention mechanism-modified long short-term memory autoencoder network, the abnormal attention mechanism layer E, and the long short-term memory autoencoder network layer D;

[0023] Long short-term memory autoencoder network layer with improved abnormal attention mechanism The hyperparameters include the number of network layers, the number of hidden layer units, and the number of neurons in the fully connected layer.

[0024] Step S5: Randomly initialize the network parameters, input the sample x in step S3 into the abnormal attention mechanism layer, output the abnormal attention mechanism hidden layer feature h, and calculate the abnormal attention mechanism associated difference loss L association , input the abnormal attention mechanism hidden layer feature h into the long short-term memory autoencoder network layer and output the reconstructed sample with the same size as x Calculate x and The root mean square error reconstruction loss L RMSE ;

[0025] The abnormal attention mechanism layer in step S5 includes the prior association of each time point with adjacent time periods and the serial association with the entire time series. The prior association uses a learnable Gaussian kernel to represent the inductive bias of the neighboring set of each time point, while the serial association corresponds to the self-attention weights learned from the original sequence, and is calculated as follows:

[0026] Q,K,V,σ=x l-1 w l Q ,x l-1 w l K ,x l-1 w l V ,x l-1 w l σ

[0027]

[0028]

[0029] h=S l V

[0030] Where Q, K, V∈R U×r are the query, key, and value vectors of attention, σ∈R U×g is the learnable Gaussian kernel scale parameter, w l Q ,w l K ,w l V ∈R r×r , w l σ ∈R r×1 are the parameter matrices of Q, K, V, and σ at the lth layer, respectively, and x l-1 is the sample data of the input l-1 layer, P l is the prior association, i and j represent the i-th time and j-th time step respectively, σ i is the scale corresponding to the learned i-th time step, and Rescale(·) is the discretization operation by removing rows. l ,S l ∈R U×U is the serial correlation, K T is the transpose of the attention key-value vector, r is the number of model channels, l is the number of model layers, Softmax(·) is the attention map normalization operation, h,h∈R U×r is the hidden layer representation of the abnormal attention mechanism. The general parameters can be set according to actual needs.

[0031] The abnormal attention mechanism associated difference loss is as follows:

[0032]

[0033]

[0034] Where KL(·) is the KL divergence operation between distributions. is the rounding operation, max(.) is the maximum value operation, U is the sliding window interception size window length, and g is the number of multi-head attention.

[0035] The root mean square error loss formula is as follows:

[0036]

[0037] Where, is the reconstructed sample. ||·||2 is the 2-norm;

[0038] Step S6: Define the loss function L total , through L association With L RMSE Calculate L total , back propagation calculates the gradient of each network layer and uses the optimizer to update the network parameters;

[0039] Optionally, the optimizer is an Adam optimizer.

[0040] The loss function L in step S6 total The calculation formula is as follows:

[0041] L total =L RMSE -λL association

[0042] Where, L RMSE is the root mean square error reconstruction loss, L association is the abnormal attention mechanism associated difference loss, and λ is the penalty term weight.

[0043] Step S7: Repeat steps S5 to S6. When the training reaches a preset number of rounds M, save the trained abnormal attention mechanism improved long short-term memory autoencoder model network parameters;

[0044] Step S8: Input the sample x into the trained The reconstruction error is obtained and the anomaly score is calculated; the formula for obtaining the sample reconstruction error and calculating the anomaly score in step S8 is as follows:

[0045] e rec =L RMSE

[0046] Score=Softmax(L association )⊙e rec

[0047] Where e rec is the sample reconstruction error, Score is the anomaly score, and ⊙ is the dot product operation.

[0048] Step S9: Use extreme value theory to define an abnormal threshold for each case test sample and output the abnormal wear monitoring result of the shield machine disc cutter.

[0049] In step S9, the extreme value theory is used to calculate the abnormal threshold for each case test sample as follows:

[0050]

[0051] Where thr F is the final threshold, thr is the initial threshold of the model error, q is the probability of abnormal data in the entire sequence, N is the number of observations, N thr is the number of times the abnormal score Score>thr, is the parameter in the generalized Pareto distribution.

[0052] The beneficial effects of the present invention are as follows: the shield machine disc cutter abnormal wear monitoring method of the present invention performs data screening and missing value filling on the excavation parameters during the shield machine excavation process, optimizes the excavation parameter data, and screens the excavation parameters closely related to the disc cutter operation status through a strategy combining the Pearson and Spearman correlation coefficients, retaining the trend characteristics of the data. The long short-term memory autoencoder based on the abnormal attention mechanism uses a learnable Gaussian kernel to obtain the prior association in the original sequence, obtains the sequence association from the attention weight learned from the original sequence, improves the distinguishability between normal and abnormal by obtaining the correlation difference of the sequence, and combines the extreme value theory to accurately and reliably monitor the abnormal wear of the shield machine disc cutter. The method of the present invention more accurately identifies the excavation parameters closely related to the disc cutter operation status, and improves the model training accuracy and the accuracy of the abnormal wear monitoring of the cutter.

Brief Description of the Drawings

[0053] Figure 1 Schematic diagram of the method for monitoring abnormal wear of disc cutters of a shield machine in the present invention;

[0054] Figure 2 2. It is a Pearson and Spearman correlation coefficient screening effect diagram in an embodiment of the present invention;

[0055] Figure 3 This is a diagram showing the effect of filling in part of the cutter head torque data in Case 4 according to an embodiment of the present invention;

[0056] Figure 4 2 is a diagram showing the effect of monitoring abnormal wear of a hob in an embodiment of the present invention.

[0057] Figure 5 This is a comparison chart of the monitoring effects of abnormal hob wear using the method of the present invention and the conventional method. [Specific implementation method]

[0058] To make the objects, technical solutions, and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with the examples and drawings. It should be understood that the exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined in any manner as long as they do not conflict with each other.

[0059] See also Figure 1-5 The present invention provides a method for monitoring abnormal wear of a shield machine disc cutter, comprising:

[0060] Step S1: Collect tunneling parameter data during shield machine excavation, and determine whether it is overall missing data or overall unchanged data based on its data length and data variance, and delete the parameters of overall missing data and overall unchanged data; the tunneling parameter data can be optimized to improve the accuracy of model training.

[0061] Preferably, if the data length does not reach 60% of other normal data lengths and / or the data variance is less than 0.01, it is determined to be entirely missing data or entirely unchanged data, and the entirely missing data or entirely unchanged data is deleted.

[0062] Optionally, the wear state of the disc hob can be divided into normal wear and abnormal wear, wherein normal wear refers to a small amount of uniform wear generated by the disc hob along the circumference of the cutter ring, and abnormal wear mainly includes excessive wear, uneven wear of the cutter ring, etc.

[0063] In one embodiment, through the data preprocessing in step S1, the optimized categories of various tunneling parameters of the shield machine during tunneling are obtained as shown in Table 1.

[0064] Table 1 Optimized data parameters after deleting missing data and unchanged data

[0065]

[0066]

[0067] Step S2: using a screening strategy to screen tunneling parameters that are closely related to the operating status of the disc cutter;

[0068] Preferably, the screening strategy is a strategy combining Pearson and Spearman correlation coefficients for screening, and the screening strategy is as follows:

[0069]

[0070]

[0071]

[0072] Where, are the Pearson correlation coefficient and the Spearman correlation coefficient respectively, and f represents the number of model training and case test sets used. Represents the standard deviation of two random variables a and b, cov(a,b) is the covariance matrix of variables a and b, Y i It represents the difference in the rank value of the i-th data pair, N is the total number of observation samples of this type of tunneling parameters, ε, η are weight coefficients, and correlation is the total correlation coefficient.

[0073] Optionally, ε and η are 0.4 and 0.6 respectively.

[0074] The ε,η are dynamically adjusted according to the actual working requirements of the shield machine or the excavation distance of the shield machine. For example, the ε,η change linearly with the excavation distance of the shield machine, and the sum of ε,η is kept equal to 1. When the total correlation coefficient is greater than the set threshold, the variable is a excavation parameter closely related to the operating state of the disc cutter. Optionally, the threshold is set to 0.4.

[0075] Through the strategy of combining the two correlation coefficients, the tunneling parameters closely related to the operating status of the disc cutter are screened, the trend characteristics of the data are retained, and the weight coefficient can be adjusted according to the actual working conditions of the shield machine disc cutter, so as to more accurately identify the tunneling parameters closely related to the operating status of the disc cutter.

[0076] Step S3: Normalize the filtered tunneling parameter data, and use the mean of the data of the previous and next rows to fill the missing data rows, and intercept a sample x of size [U×C] through a sliding window; by filling the missing data rows, the data quality, accurate training of the network model and prediction effect are improved.

[0077] Fill the empty value data by data filling, and the data filling method is as follows:

[0078]

[0079] Among them, x Nan is the missing value of the data, The data before the missing row is not empty in the first five rows. The 5 rows after the missing row are not null values, and avg represents the average value of the data. Of course, those skilled in the art know that the number of non-null values ​​before and after the missing value can be selected to fill the null value data according to the actual situation.

[0080] The padded data is intercepted through a sliding window to extract samples and divided into model training samples and case test samples. The sliding window interception size is U, the window length, and C, the number of variables in the input model or the number of variables screened out, preferably 8. In one embodiment, the window length is set to 10, the sliding step size is set to 3, and the model training samples and case test samples are set as shown in Table 2.

[0081] Table 2 Model training sample and case test sample collection settings

[0082]

[0083] In the table, "ring" is an engineering definition. Based on the actual engineering situation, the tunnel is divided into a "ring" of a certain length for construction and management. The number of collected samples can be adjusted according to the specific model training.

[0084] Step S4: Constructing a long short-term memory autoencoder network model with improved abnormal attention mechanism Including an improved long short-term memory autoencoder network with an unusual attention mechanism Hyperparameters of

[0085] The long short-term memory autoencoder network layer improved by the abnormal attention mechanism in step S4 The hyperparameters include the number of network layers, the number of hidden layer units, and the number of neurons in the fully connected layer.

[0086] Step S5: Randomly initialize the network parameters, input the sample x in step S3 into the abnormal attention mechanism layer according to the batch size, output the abnormal attention mechanism hidden layer feature h of size [U×r], and calculate the abnormal attention mechanism associated difference loss L association , the abnormal attention mechanism hidden layer feature h is input into the long short-term memory autoencoder network layer Output the reconstructed sample with the same size as x x and The dimensions are [U×C], calculate x and The root mean square error reconstruction loss L RMSE ;

[0087] The abnormal attention mechanism layer in step S5 includes modeling the prior association of each time point with its adjacent time periods and the serial association with the entire time series. The prior association uses a learnable Gaussian kernel to represent the inductive bias of the neighboring set of each time point, while the serial association corresponds to the self-attention weights learned from the original sequence. The calculation formula of the hidden layer feature h of the abnormal attention mechanism is as follows:

[0088] Q,K,V,σ=x l-1 w l Q ,xl-1 w l K ,x l-1 w l V ,x l-1 w l σ

[0089]

[0090]

[0091] h=S l V

[0092] Where Q, K, V∈R U×r are the query, key, and value vectors of attention, σ∈R U×g is the learnable Gaussian kernel scale parameter, w l Q ,w l K ,w l V ∈R r×r , w l σ ∈R r×1 are the parameter matrices of Q, K, V, and σ at the lth layer, respectively. l is the prior association, i and j represent the i-th time and j-th time step respectively, σ i is the scale corresponding to the learned i-th time step, Rescale(·) is the discretization operation by removing rows, S l ,S l ∈R U×U is the serial correlation, K T is the transpose of the attention key-value vector, r is the number of model channels, l is the number of model layers, Softmax(·) is the attention map normalization operation, h,h∈R U×r is the hidden layer representation of the abnormal attention mechanism.

[0093] The abnormal attention mechanism associated difference loss is as follows:

[0094]

[0095]

[0096] Where KL(·) is the KL divergence operation between distributions. is a rounding operation, max(.) is a maximum value operation, U is the window length, g is the number of multi-head attention, and g is preferably 8.

[0097] The root mean square error loss formula is as follows:

[0098]

[0099] Where, is the reconstructed sample. 2 is the 2-norm;

[0100] Step S6: Define the loss function L total , through L association With L RMSE Calculate L total , back propagation calculates the gradient of each network layer and uses the optimizer to update the network parameters;

[0101] Optionally, the optimizer is an Adam optimizer.

[0102] The loss function L in step S6 total The calculation formula is as follows:

[0103] L total =L RMSE -λL association

[0104] Where, L RMSE is the root mean square error reconstruction loss, L association is the abnormal attention mechanism associated difference loss, λ is the penalty term weight, which is set according to the model training requirements, and λ is preferably 0.003;

[0105] Step S7: Repeat steps S5 to S6. When the training reaches the preset number of rounds M, save the trained abnormal attention mechanism improved long short-term memory autoencoder model. Network parameters;

[0106] Optionally, the preset number of rounds M=200.

[0107] Step S8: Input the case test sample into the trained The reconstruction error of the case test sample is obtained and the anomaly score is calculated;

[0108] The formula for calculating the abnormality score in step S8 is as follows:

[0109] Score=Softmax(L association )⊙L RMSE

[0110] Where L RMSE is the root mean square error loss, Score is the anomaly score, Softmax is the activation function, and ⊙ is the dot product operation.

[0111] Step S9: Use extreme value theory to define an abnormal threshold for each case test sample and output the abnormal wear monitoring result of the shield machine disc cutter.

[0112] In step S9, the extreme value theory is used to calculate the abnormal threshold for each case test sample as follows:

[0113]

[0114] Where thr F is the abnormal threshold, thr is the initial threshold of the model error, q is the probability of abnormal data in the entire sequence, N is the number of observations, N thr is the number of times Score>thr, is the parameter in the generalized Pareto distribution.

[0115] When the abnormality score is greater than the threshold, the disc cutter is determined to be abnormally worn.

[0116] Optionally, in order to ensure reliable and effective collection of excavation parameter data, the present invention may be provided with a step of manually or automatically detecting tool wear after each excavation interval is completed, so as to ensure that the excavation process will not affect the excavation quality due to tool wear.

[0117] Optionally, during the step of automatically detecting tool wear, a sensor for detecting tool wear is provided. The sensor can be any conventional sensor, such as a visual sensor. If tool wear exceeding a set wear threshold is detected, a manual or automatic tool replacement procedure is executed. Furthermore, the tool wear detected by the sensor can be compared and analyzed with the results of the abnormal wear monitoring method of the present invention, and model training can be performed to further improve the accuracy and reliability of the present method.

[0118] The present invention provides a model construction embodiment.

[0119] The model is built based on the Pytorch deep learning framework. The experimental equipment configuration is: Windows 10 64-bit operating system, CPU is Intel Xeon Silver 4210R @ 2.4GHz, GPU is NVIDIA GeForce RTX3090. The specific parameters of the network structure are shown in Table 3.

[0120] Table 3 Specific parameter settings of the network structure of the method according to the embodiment of the present invention

[0121]

[0122] Among them, after the mask layer of the abnormal attention mechanism, the LeakyRelu activation function is sequentially connected (the negative activation slope is set to 0.2); the Sigmoid activation function is sequentially connected to the fully connected layer of the long short-term memory autoencoder.

[0123] The network model uses the Adam optimizer for parameter optimization, with a learning rate of 0.001, a batch size of 50, and a total of 200 rounds of training. In this embodiment of the present invention, a sequence encoder anomaly monitoring model (Tranad) based on a self-attention mechanism, an anomaly monitoring model with a memory query module (MemAE), a time series feature extraction anomaly monitoring model (LSTM-AE), and an inference generative anomaly monitoring model (VAE) are selected as comparison methods for comparison. The parameters of the long short-term memory autoencoder model based on the anomaly attention mechanism are loaded, and the case test samples in the data sample are input into the trained model to obtain the anomaly score. The threshold is delineated in combination with the extreme value theory to output the anomaly monitoring result.

[0124] Attachment Figure 2 The tunneling parameters closely related to the disc cutter operating status were screened out by combining the Pearson and Spearman correlation coefficients. Among them, the total correlation coefficient greater than 0.4 indicated that it had a strong correlation with the disc cutter operating status.

[0125] Attachment Figure 3 This is a data filling diagram of the shield machine cutter head torque (N·m). The data is filled by the five rows of data before and after the missing data row, which effectively fills the data gaps and retains the trend characteristics of the data.

[0126] The monitoring results of the test samples in each case in Table 2 are shown in the attached Figure 4 After each excavation interval is completed, since manual or sensor detection may detect that the tool wear exceeds the abnormal threshold, the tool replacement procedure is executed manually or programmatically. Therefore, there is a situation where the number of samples exceeding the abnormal threshold in cases with longer excavation distances is less than the number of samples exceeding the abnormal threshold in cases with shorter excavation distances.

[0127] The comparison between the method of the present invention and conventional monitoring methods (TranAD, Memae, LSTM-AE, VAE) is shown in the attached figure. Figure 5 As shown in the figure, because all models in the test sample of Case 1 had false reporting, the test sample of Case 1 was not included in the early warning quantity comparison. Among them, "*" and "#" indicate the presence of false reporting or missed reporting. The abnormal wear monitoring method of the shield machine disc cutter of the present invention did not have false reporting or missed reporting, and it can achieve effective early warning.

[0128] The present invention provides a method for monitoring abnormal wear of shield machine disc cutters. The method filters and fills missing values ​​in the excavation parameters during the shield machine excavation process, optimizes the excavation parameter data, and uses a strategy combining Pearson and Spearman correlation coefficients to filter excavation parameters closely related to the operating status of the disc cutters, thereby retaining the trend characteristics of the data. The long short-term memory autoencoder based on the abnormal attention mechanism uses a learnable Gaussian kernel to obtain prior associations in the original sequence, obtains sequence associations from the attention weights learned from the original sequence, and improves the distinguishability between normal and abnormal conditions by obtaining the correlation differences of the sequences. Combined with extreme value theory, the method can accurately and reliably monitor abnormal wear of shield machine disc cutters. The method of the present invention more accurately identifies excavation parameters closely related to the operating status of the disc cutters, and improves the accuracy of model training and the accuracy of abnormal cutter wear monitoring.

[0129] The above embodiments are intended to illustrate the present invention, not to limit the present invention. Any person skilled in the art can easily conceive of various equivalent modifications or replacements within the technical scope disclosed in the present invention, and these modifications or replacements should all be included in the scope of protection of the present invention.

Claims

1. A method for monitoring abnormal wear of a shield machine disc cutter, characterized in that: include: Step S1: Collect tunneling parameter data during shield machine tunneling, determine whether it is overall missing data or overall unchanged data based on data length and data variance, and delete the overall missing data and overall unchanged data; Step S2: using a screening strategy to screen tunneling parameters closely related to the operating status of the disc cutter, wherein the screening strategy is a combination of Pearson and Spearman correlation coefficients; Step S3: normalize the selected excavation parameter data, fill the missing rows with the mean of the previous and next rows, and intercept the sample x through the sliding window; Step S4: Constructing a long short-term memory autoencoder network model with improved abnormal attention mechanism set up Hyperparameters of Step S5: Randomly initialize the network parameters, input the sample x in step S3 into the abnormal attention mechanism layer, output the abnormal attention mechanism hidden layer feature h, and calculate the abnormal attention mechanism associated difference loss L association , input the hidden layer feature h of the abnormal attention mechanism into the long short-term memory autoencoder network layer, and output a reconstructed sample with the same size as the sample x Calculate x and The root mean square error reconstruction loss L RMSE ; Step S6: Define the loss function L total , through L association With L RMSE Calculate L total , back propagation calculates the gradient of each network layer and uses the optimizer to update Network parameters; Loss function L total The calculation formula is as follows: THE total =L RMSE -λL association Where, L RMSE is the root mean square error reconstruction loss, L association is the abnormal attention mechanism associated difference loss, λ is the penalty term weight; Step S7: Repeat steps S5 to S6 until the training reaches the preset number of rounds M, and save the trained abnormal attention mechanism improved long short-term memory autoencoder model Network parameters; Step S8: Input the case test sample into the trained The reconstruction error of the case test sample is obtained and the anomaly score is calculated; Step S9: Use extreme value theory to define an abnormal threshold for each case test sample and output the abnormal wear monitoring result of the shield machine disc cutter.

2. The method for monitoring abnormal wear of a shield machine disc cutter according to claim 1, wherein the strategy for combining the Pearson and Spearman correlation coefficients is: Where, ρ a,b , are the Pearson correlation coefficient and the Spearman correlation coefficient respectively, and f represents the number of model training and case test sets used. Represents the standard deviation of two random variables a and b, cov(a,b) is the covariance matrix of variables a and b, Y i It represents the difference in the rank value of the i-th data pair, N is the total number of observation samples of this type of tunneling parameters, ε, η are weight coefficients, and correlation is the total correlation coefficient.

3. The method for monitoring abnormal wear of a shield machine disc cutter according to claim 2, wherein the data filling method in step S3 is: in, x Nan is the missing value of the data, The data before the missing row is not empty in the first five rows. They are the 5 rows of data that are not null values ​​after the row with missing data, and avg means finding the average value of the data.

4. A shield machine disc cutter abnormal wear monitoring method as described in claim 1, wherein the long short-term memory autoencoder network layer hyperparameters of the improved abnormal attention mechanism in step S4 include the number of network layers, the number of hidden layer units, and the number of fully connected layer neurons.

5. A shield machine disc cutter abnormal wear monitoring method according to claim 3, wherein the abnormal attention mechanism layer in step S5 includes modeling the prior association of adjacent time periods at each time point and the serial association with the entire time series, the prior association using a learnable Gaussian kernel to represent the adjacency set inductive bias of each time point, the serial association corresponding to the self-attention weight learned from the original sequence, and the calculation formula of the hidden layer feature h of the abnormal attention mechanism is as follows: Q,K,V,σ=x l-1 w l Q ,x l-1 w l K ,x l-1 w l V ,x l-1 w l σ h=S l V Where Q, K, V∈R U×r are the query, key, and value vectors of attention, σ∈R U×g is the learnable Gaussian kernel scale parameter, w l Q ,w l K ,w l V ∈R r×r , w l σ ∈R r×1 are the parameter matrices of Q, K, V, and σ at the lth layer, respectively. l is the prior association, i and j represent the i-th time and j-th time step respectively, σ i is the scale corresponding to the learned i-th time step, Rescale(·) is the discretization operation by removing rows, S l ,S l ∈R U×U is the serial correlation, K T is the transpose of the attention key-value vector, r is the number of model channels, l is the number of model layers, Softmax(·) is the attention map normalization operation, h,h∈R U×r is the hidden layer representation of the abnormal attention mechanism.

6. According to the method for monitoring abnormal wear of a shield machine disc cutter according to claim 5, the abnormal attention mechanism associated difference loss is as follows: Where KL(·) is the KL divergence operation between distributions. is the rounding operation, max(.) is the maximum value operation, U is the window length, and g is the number of multi-head attention.

7. A method for monitoring abnormal wear of a shield machine disc cutter according to claim 6, The root mean square error loss formula is as follows: Where, is the reconstructed sample, ||·||2 is the 2-norm; 8. A shield machine disc cutter abnormal wear monitoring method according to claim 7, characterized in that: The formula for calculating the abnormality score in step S8 is as follows: Score=Softmax(L association )⊙L RMSE Where L RMSE is the root mean square error loss, Score is the anomaly score, Softmax is the activation function, and ⊙ is the dot product operation.

9. A shield machine disc cutter abnormal wear monitoring method according to claim 8, characterized in that: In step S9, the extreme value theory is used to calculate the abnormal threshold for each case test sample as follows: Where thr F is the final threshold, thr is the initial threshold of the model error, q is the probability of abnormal data in the entire sequence, N is the number of observations, N thr is the number of times the abnormal score Score>thr, is the parameter in the generalized Pareto distribution.

Images