TBM cutterhead fault identification method and system based on relational network model

Abstract

The invention discloses a TBM cutterhead fault identification method and system based on a relational network model, belongs to the technical field of fault identification, and solves the problem that the fault type of an existing TBM cutterhead cannot be identified when the existing TBM cutterhead breaks down. Based on a relational network model, aiming at a TBM cutterhead fault waveform signal received in a signal sensor, a query data set and a support data set are mapped to a low-dimensional space from a high-dimensional feature through an embedded module, feature information corresponding to images in the two data sets is cascaded through a related module, and the fault waveform signal of the TBM cutterhead is obtained. The similarity score between the two images is calculated, the similarity score with the highest image similarity degree in the two data sets is output, and the corresponding fault category is output according to the similarity score, so that engineering technicians can maintain the fault in time conveniently, the time cost is reduced, the maintenance efficiency is improved, and the economic value is good.

Description

Technical Field

[0001] The invention belongs to the technical field of fault identification, and relates to a TBM cutter head fault identification method and system based on a relational network model. Background Art

[0002] During the construction of a full-section tunnel boring machine (TBM), the existence of unfavorable geological bodies such as faults, broken zones, and changes in soft and hard rocks, as well as rock hardness, have become one of the reasons for the failure of the cutterhead of a full-section tunnel boring machine. Therefore, how to effectively monitor the state of the shield machine's excavation cutterhead and determine whether the cutterhead is faulty is of great significance to the TBM cutterhead. For example, the invention patent with application publication number CN117145493A discloses a method for early warning of damage to the side scraper of a full-section tunnel boring machine. According to the historical data and real-time excavation data of the shield construction line, a real-time early warning of damage to the side scraper of an earth pressure balance shield and a slurry balance shield is achieved. By independently extracting the historical construction data and real-time excavation data of the line, the changes in the characteristic data are obtained, and real-time and rapid judgment of damage to the side scrapers of multiple line shields is achieved. However, this technology only provides an early warning method for identifying damage to the side scraper, and at the same time relies on a set fixed judgment standard. When other types of cutterhead failures occur, this method is difficult to respond.

[0003] According to statistics, more than 50% of the overall risk accidents of full-face tunnel boring machines are cutterhead risk accidents, and the main manifestations of the accidents include wear of the cutterhead panel, wear of the tool, and the cutterhead losing sufficient thrust and torque, etc. Cutterhead failure accidents lead to a significant decrease in the excavation efficiency and service life of the full-face tunnel boring machine. At the same time, since the full-face tunnel boring machine is located in a complex underground space and is huge and difficult to move, once a cutterhead failure accident occurs, it is difficult to diagnose and eliminate it in time, which not only delays the project schedule and increases the project time cost, but also causes the maintenance time cost to rise sharply. Therefore, accurate and effective identification of TBM cutterhead failures has great economic value for coal mine TBM tunneling projects. Summary of the invention

[0004] The technical solution of the present invention is used to solve the problem that the type of cutter disc fault cannot be identified when a cutter disc fault occurs in the existing TBM.

[0005] The present invention solves the above technical problems through the following technical solutions:

[0006] A TBM cutterhead fault identification method based on a relational network model comprises the following steps:

[0007] S1. Collect the waveform signal data set X = {x1, x2, x3, ..., x n}, where n is a positive integer;

[0008] S2. Perform time-frequency conversion on the waveform signal dataset X to obtain the time-frequency spectrum dataset Y of TBM cutterhead damage = {y1, y2, y3, ..., y m}, where m is a positive integer;

[0009] S3, extracting image features of the time-frequency spectrum data set Y, and annotating the image features according to the fault category of TBM cutterhead damage;

[0010] S4. Construct a relationship network model, wherein the relationship network model includes an embedding module and relationship modules

[0011] S5. Supervise the similarity score based on the objective function and optimize the similarity score;

[0012] S6. Perform fault identification on the waveform signal received by the signal sensor and output the fault category.

[0013] Furthermore, the data annotation includes a top event and a first-level intermediate event, the top event includes a cutter disc failure and a normal cutter disc, the first-level intermediate event includes a panel failure and a tool failure, and the panel failure and the tool failure are sub-events of the cutter disc failure.

[0014] Furthermore, the image features of the spectrogram dataset Y are extracted using the following logic in S3:

[0015]

[0016] Where W(a,b) represents the local features of the input data at different scales and translation positions, x(t) represents the input original signal, Ψ*(·) represents the complex conjugate of the wavelet function, t represents the time variable of the signal, a represents the scale factor, and b represents the translation parameter.

[0017] Further, the S4 comprises the following steps:

[0018] S41. Using embedded modules The network is trained on the annotated time-frequency spectrum dataset in S3 to construct the query dataset Q = {c1, c2, c3, ..., c i}, where i is a positive integer;

[0019] S42, perform real-time time-frequency conversion on the waveform signal received from the signal sensor, and construct a supporting data set S = {d1, d2, d3, ..., d j}, where j is a positive integer;

[0020] S43, based on the query dataset Q and the support dataset S, using the embedding module Map high-dimensional features to low-dimensional space:

[0021] S44. Utilize relevant modules Cascade query dataset Q and support dataset S for image feature information;

[0022] S45. Construct an objective function and calculate a similarity score.

[0023] Further, the following logic is used in S43 to express the use of embedded modules: Map high-dimensional features to low-dimensional space:

[0024]

[0025] Where, d j represents a sample in the query dataset Q, c i represents a sample in the support dataset S, Indicates that c in the support dataset S i The high-dimensional features of Indicates the query data set Q in d j The high-dimensional features of E ci Indicates that c in the query dataset Q i The low-dimensional features, E dj Indicates the query data set Q in d j low-dimensional features.

[0026] Furthermore, the objective function is constructed and the similarity score is calculated using the following logic in S45:

[0027]

[0028] In the formula, r i,j Indicates c i With d j The similarity score between them, C[·] represents the connection function, Represents the correlation calculation function.

[0029] Furthermore, the following logic is used in S5 to optimize the similarity score:

[0030]

[0031] In the formula, φ represents the target loss function value, Y i Represents the i-th label output by the model, Y j represents the jth label output by the model, 1(Y i = = Y j ) means when Y i With Y j If they are the same, the matching similarity is 1, otherwise it is 0.

[0032] The present invention also provides a TBM cutter head fault identification system based on a relational network model, comprising a training set acquisition module, a training set processing module, a fault category labeling module, a relational network model module, a similarity optimization module and a fault detection module;

[0033] The training set acquisition module is used to collect the waveform signal data set X={x1,x2,x3,...,x n}, where n is a positive integer;

[0034] The training set processing module is used to perform time-frequency conversion on the waveform signal data set X to obtain a time-frequency spectrum data set Y of TBM cutterhead damage = {y1, y2, y3, ..., y m}, where m is a positive integer;

[0035] The fault category labeling module is used to extract image features of the time-frequency spectrum data set Y and label the image features according to the fault category of the TBM cutterhead damage;

[0036] The relationship network model module is used to construct a relationship network model, and the relationship network model includes an embedding module and relationship modules

[0037] The similarity optimization module is used to supervise the similarity score based on the objective function and optimize the similarity score;

[0038] The fault detection module is used to perform fault identification on the waveform signal received by the signal sensor and output the fault category.

[0039] Further, the relationship network model module includes a query unit, a support unit, a mapping unit, a cascade unit and a similarity unit;

[0040] The query unit is used to use the embedded module The network training is performed on the time-frequency spectrum data set after data annotation in the fault category annotation module, and the query data set Q = {c1, c2, c3, ..., c i}, where i is a positive integer;

[0041] The supporting unit is used to perform real-time time-frequency conversion on the waveform signal received by the signal sensor, and construct a supporting data set S = {d1, d2, d3, ..., d j}, where j is a positive integer;

[0042] The mapping unit is used to use the embedding module based on the query data set Q and the support data set S. Map high-dimensional features to low-dimensional space;

[0043] The cascade unit is used to utilize related modules Cascade query dataset Q and support dataset S for image feature information;

[0044] The similarity unit is used to construct an objective function and calculate a similarity score.

[0045] The advantages of the present invention are:

[0046] (1) The present invention targets the continuous vibration characteristics of the TBM cutterhead during operation, and uses the time-frequency domain to identify faults by using the continuous time-frequency spectrum when the cutterhead is working and the discontinuous time-frequency spectrum when a fault occurs, thereby greatly improving the accuracy of fault identification. In addition, based on the relational network model, the present invention maps the query data set and the support data set from high-dimensional features to low-dimensional space through an embedding module for the TBM cutterhead fault waveform signal received in the signal sensor, cascades the feature information corresponding to the images in the two data sets through the correlation module, calculates the similarity score between the two images, outputs the similarity score of the image with the highest degree of similarity in the two data sets, and outputs the corresponding fault category according to the similarity score, which facilitates timely maintenance by engineering and technical personnel, reduces time costs, improves maintenance efficiency, and has good economic value.

[0047] (2) The present invention can repeatedly perform data training on cutter disc fault categories based on the relational network model to expand the number of cutter disc fault database sets, and can accurately identify and predict various types of cutter disc faults, with good scalability.

[0048] (3) The embedded module based on the relational network model has strong real-time processing capabilities and can complete the fault identification task within the specified time. At the same time, the embedded module is easy to integrate. The related modules cascade the image feature information of the query data set and the support data set and adopt a standardized interface design to facilitate the connection and communication between different modules, thereby improving the flexibility of model design and the convenience of maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of a TBM cutter head fault identification method based on a relational network model according to the first embodiment of the present invention;

[0050] Figure 2 is a time-frequency diagram of a TBM cutterhead wear fault according to the first embodiment of the present invention;

[0051] Figure 3 is a schematic diagram of the overall structure of a fault tree according to the first embodiment of the present invention;

[0052] Figure 4 is a schematic diagram of a relationship network model of the first embodiment of the present invention;

[0053] Figure 5 It is a time-frequency diagram of the TBM cutter head fault identification result of the first embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments:

[0056] Embodiment 1

[0057] like Figure 1 Specifically, a TBM cutter head fault identification method based on a relational network model is disclosed, comprising:

[0058] S1. Collect the waveform signal data set X = {x1, x2, x3, ..., x n}, where n is a positive integer;

[0059] S2. Perform time-frequency conversion on the waveform signal dataset X to obtain the time-frequency spectrum dataset Y of TBM cutterhead damage = {y1, y2, y3, ..., y m}, where m is a positive integer;

[0060] In this embodiment, the waveform signal data set X is converted into a time-frequency data set using a wavelet function to obtain a cutter head damage time-frequency spectrum data set Y = {y1, y2, y3, ..., y m}, where m is a positive integer. Figure 2 The figure shows the time-frequency diagram of the waveform signal containing TBM cutterhead damage after time-frequency conversion.

[0061] S3, extracting image features of the time-frequency spectrum data set Y, and annotating the image features according to the fault category of the TBM cutterhead damage, wherein the data annotation includes a top event and a first-level intermediate event, wherein the top event includes a cutterhead failure and a normal cutterhead, and the first-level intermediate event includes a panel failure and a tool failure, wherein the panel failure and the tool failure are sub-events of the cutterhead failure;

[0062] In this embodiment, the continuous wavelet transform method is used to extract the image features of the time-frequency spectrum data set Y, and then the image features are data labeled according to the fault categories of TBM cutterhead damage. This embodiment is based on the perspective of the relationship network and hierarchically labels the fault categories.

[0063] This embodiment uses the following logic to extract the image features of the spectrogram dataset Y:

[0064]

[0065] Wherein, W(a,b) represents the local features of the input data at different scales and translation positions, x(t) represents the input original signal, and in this embodiment, the time-frequency spectrum data set Y is used as the input signal, Ψ*(·) represents the complex conjugate of the wavelet function, t represents the time variable of the signal, a represents the scale factor, and b represents the translation parameter.

[0066] like Figure 3 The figure shows the overall structure of the fault tree of this embodiment. The normal and faulty cutter discs are taken as top events. By classifying the image features, it is first determined whether the cutter disc is "faulty" or "normal" as the starting point of the relationship network. Since "normal cutter disc" itself is a stable state, it does not need to be refined into further sub-events, while "cutter disc fault" needs to be further divided into specific fault types. The fault categories are further refined. This embodiment takes "panel fault" and "tool fault" as examples, and takes "panel fault" and "tool fault" as sub-events of "cutter disc fault", and marks the image features with the above fault categories.

[0067] S4. Construct a relationship network model, wherein the relationship network model includes an embedding module and relationship modules

[0068] like Figure 4 As shown, the S4 comprises the following steps:

[0069] S41. Using embedded modules The network is trained on the annotated time-frequency spectrum dataset in S3 to construct the query dataset Q = {c1, c2, c3, ..., c i}, where i is a positive integer;

[0070] In this embodiment, the embedded module It includes an input layer, an embedding layer and an output layer; the input layer is used to take the time-frequency spectrum after data annotation as input.

[0071] The embedding layer includes a BN convolution layer and a ReLU convolution layer, wherein the BN convolution layer is used to process the feature distribution, stabilize the training process by normalizing the intermediate features, and accelerate convergence, and the ReLU convolution layer is used to provide the ability for nonlinear transformation and extract complex features. In this embodiment, the embedding layer maps the high-dimensional features of the input layer to a low-dimensional space by using a convolutional neural network (CNN), while retaining the similarity information between the features.

[0072] The output layer is used to use the output of the embedding layer as the subsequent relationship module The input is used to calculate the similarity between different sample data.

[0073] S42, perform real-time time-frequency conversion on the waveform signal received from the signal sensor, and construct a supporting data set S = {d1, d2, d3, ..., d j}, where j is a positive integer;

[0074] In this embodiment, the waveform signal received by the signal sensor is used as the key data set for training the relational network model, and real-time time-frequency conversion is performed to obtain the real-time time-frequency spectrum of TBM cutterhead damage, and the supporting data set S = {d1, d2, d3, ..., d j}, where j is a positive integer.

[0075] S43, based on the query dataset Q and the support dataset S, using the embedding module Map high-dimensional features to low-dimensional space;

[0076] In this embodiment, the following logic is used to express the use of embedded modules Map high-dimensional features to low-dimensional space:

[0077]

[0078] Where, d j represents a sample in the query dataset Q, c i represents a sample in the support data set S. Since there is only one sample vector for each category in the support data set S, then c i Represents a classification vector information, Indicates that c in the support dataset S i The high-dimensional features of Indicates the query data set Q in d j The high-dimensional features of E ci Indicates that c in the query dataset Q i The low-dimensional features, E dj Indicates the query data set Q in d j low-dimensional features.

[0079] In this embodiment, using the embedded module Map the high-dimensional features in the query dataset Q and the support dataset S to a low-dimensional space to better measure similarity, using the embedding module Perform feature extraction on the input data to extract feature vectors that reflect the characteristics of the data, which are then used to calculate the relationship scores between samples. It has strong real-time processing capabilities and can complete fault identification tasks within the specified time. At the same time, the embedded system can be easily integrated with other devices or modules to form a complete solution.

[0080] S44. Utilize relevant modules Cascade query dataset Q and support dataset S for image feature information;

[0081] In this embodiment, the related modules It includes an input layer, an embedding layer and an output layer; the input layer is used to receive The extracted feature vector is used to input the relevant modules using the input layer preprocessing the data.

[0082] The embedding layer includes a BN fully connected layer and a ReLU fully connected layer. The BN fully connected layer is used to input related modules. The data is normalized to solve the problem of internal covariate shift in the neural network. The introduction of the BN fully connected layer enables the relational network model to use the information of other samples in the mini-batch when predicting independent samples, thereby making the loss surface smoother and easier to find the optimal solution. The ReLU fully connected layer is used to introduce nonlinear factors so that the network can learn more complex and abstract feature combinations.

[0083] The output layer determines the relationship or similarity between the query sample and the support samples based on the features extracted by the embedding layer.

[0084] In this embodiment, the relevant modules Receives data from embedded modules The feature information corresponding to the image in the query dataset Q and the feature information corresponding to the image in the support dataset S are concatenated by pairwise connection. Specifically, if there are N images in the query dataset Q and M images in the support dataset S, N×M concatenated feature vectors will be generated. Each concatenated feature vector contains the combined information between the query set image and the support set image, which is used for correlation calculation.

[0085] S45, constructing an objective function and calculating a similarity score;

[0086] In this embodiment, the objective function is constructed and the similarity score is calculated using the following logic:

[0087]

[0088] In the formula, r i,j Indicates c i With d j The similarity score between them, C[·] represents the connection function, Represents the correlation calculation function.

[0089] In this embodiment, through the relevant modules To calculate the similarity score between the two images, and finally output a one-shot vector, which represents the category with the highest similarity between the image in the query dataset Q and the image in the support dataset S.

[0090] S5. Supervise the similarity score based on the objective function and optimize the similarity score;

[0091] In this embodiment, the similarity score is supervised by using the improved minimization mean square error function as the objective function, and the similarity score is optimized using the following logic:

[0092]

[0093] In the formula, φ represents the target loss function value, Y i Represents the i-th label output by the model, Y j represents the jth label output by the model, 1(Y i = = Y j ) means when Y i With Y j If they are the same, the matching similarity is 1, otherwise it is 0. This condition is used to determine whether two samples belong to the same category, so that the relationship network model has a higher score for correct classification and a lower score for wrong classification.

[0094] S6, performing fault identification on the waveform signal received by the signal sensor and outputting the fault category;

[0095] like Figure 4-5 As shown, for the waveform signal received in the signal sensor, this embodiment maps the query data set and the support data set from high-dimensional features to low-dimensional space through an embedding module, cascades the feature information corresponding to the images in the two data sets through a correlation module, calculates the similarity score between the two images, outputs the similarity score of the image with the highest degree of similarity in the two data sets, and outputs the corresponding fault category according to the similarity score.

[0096] This embodiment is based on capturing the waveform signal existing in the signal sensor, identifying and judging the fault category of the tool through the relational network model, and at the same time, based on the relational network model, it can repeatedly perform data training on the cutter disc fault category to expand the number of cutter disc fault database sets, and can accurately identify and predict various types of cutter disc faults, with good scalability.

[0097] Embodiment 2

[0098] The present invention also discloses a TBM cutter head fault identification system based on a relational network model, comprising a training set acquisition module, a training set processing module, a fault category labeling module, a relational network model module, a similarity optimization module and a fault detection module;

[0099] The training set acquisition module is used to collect the waveform signal data set X={x1,x2,x3,...,x n}, where n is a positive integer;

[0100] The training set processing module is used to perform time-frequency conversion on the waveform signal data set X to obtain a time-frequency spectrum data set Y of TBM cutterhead damage = {y1, y2, y3, ..., y m}, where m is a positive integer;

[0101] In this embodiment, the waveform signal data set X is converted into a time-frequency data set using a wavelet function to obtain a cutter head damage time-frequency spectrum data set Y = {y1, y2, y3, ..., y m}, where m is a positive integer.

[0102] The fault category labeling module is used to extract image features of the time-frequency spectrum data set Y, and to label the image features according to the fault category of the TBM cutter head damage, wherein the data labeling includes a top event and a first-level intermediate event, wherein the top event includes a cutter head failure and a normal cutter head, and the first-level intermediate event includes a panel failure and a tool failure, wherein the panel failure and the tool failure are sub-events of the cutter head failure;

[0103] In this embodiment, the continuous wavelet transform method is used to extract the image features of the time-frequency spectrum data set Y, and then the image features are data labeled according to the fault categories of TBM cutterhead damage. This embodiment is based on the perspective of the relationship network and hierarchically labels the fault categories.

[0104] This embodiment uses the following logic to extract the image features of the spectrogram dataset Y:

[0105]

[0106] Wherein, W(a,b) represents the local features of the input data at different scales and translation positions, x(t) represents the input original signal, and in this embodiment, the time-frequency spectrum data set Y is used as the input signal, Ψ*(·) represents the complex conjugate of the wavelet function, t represents the time variable of the signal, a represents the scale factor, and b represents the translation parameter.

[0107] like Figure 3 The figure shows the overall structure of the fault tree of this embodiment. The normal and faulty cutter discs are taken as top events. By classifying the image features, it is first determined whether the cutter disc is "faulty" or "normal" as the starting point of the relationship network. Since "normal cutter disc" itself is a stable state, it does not need to be refined into further sub-events, while "cutter disc fault" needs to be further divided into specific fault types. The fault categories are further refined. This embodiment takes "panel fault" and "tool fault" as examples, and takes "panel fault" and "tool fault" as sub-events of "cutter disc fault", and marks the image features with the above fault categories.

[0108] The relationship network model module is used to construct a relationship network model, and the relationship network model includes an embedding module and relationship modules

[0109] The relationship network model module includes a query unit, a support unit, a mapping unit, a cascade unit and a similarity unit;

[0110] The query unit is used to use the embedded module The network training is performed on the time-frequency spectrum data set after data annotation in the fault category annotation module, and the query data set Q = {c1, c2, c3, ..., c i}, where i is a positive integer;

[0111] In this embodiment, the embedded module It includes an input layer, an embedding layer and an output layer; the input layer is used to take the time-frequency spectrum after data annotation as input.

[0112] The embedding layer includes a BN convolution layer and a ReLU convolution layer, wherein the BN convolution layer is used to process the feature distribution, stabilize the training process by normalizing the intermediate features, and accelerate convergence, and the ReLU convolution layer is used to provide the ability for nonlinear transformation and extract complex features. In this embodiment, the embedding layer maps the high-dimensional features of the input layer to a low-dimensional space by using a convolutional neural network (CNN), while retaining the similarity information between the features.

[0113] The output layer is used to use the output of the embedding layer as the subsequent relationship module The input is used to calculate the similarity between different sample data.

[0114] The supporting unit is used to perform real-time time-frequency conversion on the waveform signal received by the signal sensor, and construct a supporting data set S = {d1, d2, d3, ..., d j}, where j is a positive integer;

[0115] In this embodiment, the waveform signal received by the signal sensor is used as the key data set for training the relational network model, and real-time time-frequency conversion is performed to obtain the real-time time-frequency spectrum of TBM cutterhead damage, and the supporting data set S = {d1, d2, d3, ..., d j}, where j is a positive integer.

[0116] The mapping unit is used to use the embedding module based on the query data set Q and the support data set S. Map high-dimensional features to low-dimensional space;

[0117] In this embodiment, the following logic is used to express the use of embedded modules Map high-dimensional features to low-dimensional space:

[0118]

[0119] Where, d j represents a sample in the query dataset Q, c i represents a sample in the support data set S. Since there is only one sample vector for each category in the support data set S, then c i Represents a classification vector information, Indicates that c in the support dataset S i The high-dimensional features of Indicates the query data set Q in d j The high-dimensional features of E ci Indicates that c in the query dataset Q i The low-dimensional features, E dj Indicates the query data set Q in d j low-dimensional features.

[0120] In this embodiment, using the embedded module Map the high-dimensional features in the query dataset Q and the support dataset S to a low-dimensional space to better measure similarity, using the embedding module Perform feature extraction on the input data to extract feature vectors that reflect the characteristics of the data, which are then used to calculate the relationship scores between samples. It has strong real-time processing capabilities and can complete fault identification tasks within the specified time. At the same time, the embedded system can be easily integrated with other devices or modules to form a complete solution.

[0121] The cascade unit is used to utilize related modules Cascade query dataset Q and support dataset S for image feature information;

[0122] In this embodiment, the related modules It includes an input layer, an embedding layer and an output layer; the input layer is used to receive The extracted feature vector is used to input the relevant modules using the input layer preprocessing the data.

[0123] The embedding layer includes a BN fully connected layer and a ReLU fully connected layer. The BN fully connected layer is used to input related modules. The data is normalized to solve the problem of internal covariate shift in the neural network. The introduction of the BN fully connected layer enables the relational network model to use the information of other samples in the mini-batch when predicting independent samples, thereby making the loss surface smoother and easier to find the optimal solution. The ReLU fully connected layer is used to introduce nonlinear factors so that the network can learn more complex and abstract feature combinations.

[0124] The output layer determines the relationship or similarity between the query sample and the support samples based on the features extracted by the embedding layer.

[0125] In this embodiment, the relevant modules Receives data from embedded modules The feature information corresponding to the image in the query dataset Q and the feature information corresponding to the image in the support dataset S are concatenated by pairwise connection. Specifically, if there are N images in the query dataset Q and M images in the support dataset S, N×M concatenated feature vectors will be generated. Each concatenated feature vector contains the combined information between the query set image and the support set image, which is used for correlation calculation.

[0126] The similarity unit is used to construct an objective function and calculate a similarity score;

[0127] In this embodiment, the objective function is constructed and the similarity score is calculated using the following logic:

[0128]

[0129] In the formula, r i,j Indicates c i With d j The similarity score between them, C[·] represents the connection function, Represents the correlation calculation function.

[0130] In this embodiment, through the relevant modules To calculate the similarity score between the two images, and finally output a one-shot vector, which represents the category with the highest similarity between the image in the query dataset Q and the image in the support dataset S.

[0131] The similarity optimization module is used to supervise the similarity score based on the objective function and optimize the similarity score;

[0132] In this embodiment, the similarity score is supervised by using the improved minimization mean square error function as the objective function, and the similarity score is optimized using the following logic:

[0133]

[0134] In the formula, φ represents the target loss function value, Y i Represents the i-th label output by the model, Y j represents the jth label output by the model, 1(Y i = = Y j ) means when Y i With Y j If they are the same, the matching similarity is 1, otherwise it is 0. This condition is used to determine whether two samples belong to the same category, so that the relationship network model has a higher score for correct classification and a lower score for wrong classification.

[0135] The fault detection module is used to perform fault identification on the waveform signal received by the signal sensor and output the fault category;

[0136] For the waveform signal received in the signal sensor, this embodiment maps the query data set and the support data set from high-dimensional features to low-dimensional space through an embedding module, cascades the feature information corresponding to the images in the two data sets through a correlation module, calculates the similarity score between the two images, outputs the similarity score with the highest degree of similarity between the two data sets, and outputs the corresponding fault category based on the similarity score.

[0137] This embodiment is based on capturing the waveform signal existing in the signal sensor, identifying and judging the fault category of the tool through the relational network model, and at the same time, based on the relational network model, it can repeatedly perform data training on the cutter disc fault category to expand the number of cutter disc fault database sets, and can accurately identify and predict various types of cutter disc faults, with good scalability.

[0138] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A TBM cutterhead fault identification method based on a relational network model, characterized in that: The following steps are involved: S1. Collect the waveform signal data set X = {x1, x2, x3, ..., x n }, where n is a positive integer; S2. Perform time-frequency conversion on the waveform signal dataset X to obtain the time-frequency spectrum dataset Y of TBM cutterhead damage = {y1, y2, y3, ..., y m }, where m is a positive integer; S3, extracting image features of the time-frequency spectrum data set Y, and annotating the image features according to the fault category of TBM cutterhead damage; S4. Construct a relationship network model, wherein the relationship network model includes an embedding module and relationship modules S5. Supervise the similarity score based on the objective function and optimize the similarity score; S6. Perform fault identification on the waveform signal received by the signal sensor and output the fault category.

2. The TBM cutterhead fault identification method based on the relational network model according to claim 1 is characterized in that: The data annotation in S3 includes top events and first-level intermediate events. The top events include cutter head failure and cutter head normal. The first-level intermediate events include panel failure and tool failure. The panel failure and tool failure are sub-events of cutter head failure.

3. The TBM cutterhead fault identification method based on the relational network model according to claim 1 is characterized in that: In S3, the following logic is used to extract the image features of the spectrogram dataset Y: Where W(a,b) represents the local features of the input data at different scales and translation positions, x(t) represents the input original signal, Ψ*(·) represents the complex conjugate of the wavelet function, t represents the time variable of the signal, a represents the scale factor, and b represents the translation parameter.

4. The TBM cutterhead fault identification method based on a relational network model according to claim 1 is characterized in that: The S4 comprises the following steps: S41. Using embedded modules The network is trained on the annotated time-frequency spectrum dataset in S3 to construct the query dataset Q = {c1, c2, c3, ..., c i }, where i is a positive integer; S42, perform real-time time-frequency conversion on the waveform signal received from the signal sensor, and construct a supporting data set S = {d1, d2, d3, ..., d j }, where j is a positive integer; S43, based on the query dataset Q and the support dataset S, using the embedding module Map high-dimensional features to low-dimensional space: S44. Utilize relevant modules Cascade query dataset Q and support dataset S for image feature information; S45. Construct an objective function and calculate a similarity score.

5. The TBM cutterhead fault identification method based on the relational network model according to claim 4 is characterized in that: The following logic is used in S43 to express the use of embedded modules Map high-dimensional features to low-dimensional space: Where, d j represents a sample in the query dataset Q, c i represents a sample in the support dataset S, Indicates that c in the support dataset S i The high-dimensional features of Indicates the query data set Q in d j The high-dimensional features of E ci Indicates c in the query dataset Q i The low-dimensional features, E dj Indicates the query data set Q in d j low-dimensional features.

6. The TBM cutterhead fault identification method based on the relational network model according to claim 5 is characterized in that: In S45, the following logic is used to construct the objective function and calculate the similarity score: In the formula, r i,j Indicates c i With d j The similarity score between them, C[·] represents the connection function, Represents the correlation calculation function.

7. The TBM cutterhead fault identification method based on the relational network model according to claim 6 is characterized in that: The following logic is used in S5 to optimize the similarity score: In the formula, φ represents the target loss function value, Y i represents the i-th label output by the model, Y j represents the jth label output by the model, 1(Y i = = Y j ) means when Y i With Y j If they are the same, the matching similarity is 1, otherwise it is 0.

8. A TBM cutterhead fault identification system based on a relational network model, characterized in that: It includes a training set acquisition module, a training set processing module, a fault category labeling module, a relationship network model module, a similarity optimization module and a fault detection module; The training set acquisition module is used to collect the waveform signal data set X={x1,x2,x3,...,x n }, where n is a positive integer; The training set processing module is used to perform time-frequency conversion on the waveform signal data set X to obtain a time-frequency spectrum data set Y of TBM cutterhead damage = {y1, y2, y3, ..., y m }, where m is a positive integer; The fault category labeling module is used to extract image features of the time-frequency spectrum data set Y and label the image features according to the fault category of the TBM cutterhead damage; The relationship network model module is used to construct a relationship network model, and the relationship network model includes an embedding module and relationship modules The similarity optimization module is used to supervise the similarity score based on the objective function and optimize the similarity score; The fault detection module is used to perform fault identification on the waveform signal received by the signal sensor and output the fault category.

9. The TBM cutterhead fault identification system based on the relational network model according to claim 8 is characterized in that: The relationship network model module includes a query unit, a support unit, a mapping unit, a cascade unit and a similarity unit; The query unit is used to use the embedded module The network training is performed on the time-frequency spectrum data set after data annotation in the fault category annotation module, and the query data set Q = {c1, c2, c3, ..., c i }, where i is a positive integer; The supporting unit is used to perform real-time time-frequency conversion on the waveform signal received by the signal sensor, and construct a supporting data set S = {d1, d2, d3, ..., d j }, where j is a positive integer; The mapping unit is used to use the embedding module based on the query data set Q and the support data set S. Map high-dimensional features to low-dimensional space; The cascade unit is used to utilize related modules Cascade query dataset Q and support dataset S for image feature information; The similarity unit is used to construct an objective function and calculate a similarity score.

Images