A method, device and medium for predicting the wear amount of a shield machine cutter

Abstract

The embodiment of the present application discloses a shield machine tool wear prediction method, device and medium. It belongs to the technical field of shield machines and solves the problem of poor prediction accuracy and robustness of shield machine tool wear. The historical working data of the shield machine is used to generate sample data; the wavelet transform image of the main shaft vibration signal and the geological cross-section map in the sample data are subjected to feature extraction through a dual-channel fusion shield machine lightweight feature extraction network to obtain a feature vector; based on the feature vector corresponding to the sample data, a multi-perspective information fusion feature vector is constructed; based on the multi-perspective information fusion feature vector, a time series feature vector is constructed to construct an improved VIT shield machine tool wear prediction model through the time series feature vector; the improved VIT shield machine tool wear prediction model is subjected to parameter optimization, and when the parameter optimization result meets the preset error condition, the shield machine tool wear is predicted.

Description

Technical Field

[0001] The present application relates to the technical field of shield machines, and in particular to a method, device and medium for predicting tool wear of shield machines. Background Art

[0002] Predicting shield machine tool wear is crucial for ensuring the safety, economy, and efficiency of shield construction. Accurate wear prediction allows for the rational planning of tool replacements, avoiding unplanned downtime and construction delays caused by tool damage, thereby reducing project risks and costs.

[0003] Shield machine tool wear is affected by a variety of factors, including excavation parameters, geological structure, and tool parameters. These various influencing factors have numerous parameters and are coupled to each other. Therefore, traditional tool wear prediction methods based on prior knowledge struggle to establish a definitive mathematical mapping relationship through mechanism analysis and mathematical modeling to describe the wear changes of shield machine tools in complex environments. Traditional machine learning-based methods still rely on artificial features during the feature extraction stage, and the constructed shield machine tool wear prediction features have defects such as low information density and incomplete features. Furthermore, traditional machine learning models are generally not good at capturing complex nonlinear relationships in data, so the accuracy of tool wear prediction needs to be further improved.

[0004] In recent years, with the development of artificial intelligence (AI) technology, deep learning has been widely used to predict tool wear. However, the information used to predict tool wear in shield machines is mostly heterogeneous and heterogeneous, with inconsistent data modalities. Current deep learning-based methods struggle to fully utilize this information, resulting in poor accuracy and robustness in predicting tool wear. Summary of the Invention

[0005] The embodiments of the present application provide a method, device and medium for predicting the wear of shield machine tools, which are used to solve the following technical problems: the information used by shield machines to characterize the prediction of tool wear is mostly heterogeneous and heterogeneous, and the data modalities are not unified. At present, deep learning-based methods are difficult to fully utilize the prediction information, resulting in poor prediction accuracy and robustness of shield machine tool wear.

[0006] The embodiments of this application adopt the following technical solutions:

[0007] The present application provides a shield machine tool wear prediction method, which includes sorting and segmenting acquired shield machine historical working data to generate sample data of different categories; extracting features from the wavelet transform image of the main shaft vibration signal and the geological cross-section map in the sample data using a dual-channel fusion shield machine lightweight feature extraction network to obtain a feature vector; constructing a multi-perspective information fusion feature vector based on the feature vector corresponding to the dual-channel fusion shield machine lightweight feature extraction network and the feature vectors corresponding to other data in the sample data; constructing a time series feature vector based on the multi-perspective information fusion feature vector, and constructing an improved VIT shield machine tool wear prediction model using the time series feature vector; and optimizing the parameters of the improved VIT shield machine tool wear prediction model. When the parameter optimization results meet preset error conditions, the shield machine tool wear is predicted.

[0008] The embodiment of the present application constructs multi-perspective complete tool prediction information, including spindle vibration signals, geological parameters, tunneling parameters and shield machine tool parameters. In order to extract richer spindle vibration features, the embodiment of the present application converts the time domain signal of the spindle vibration into an image representation based on wavelet continuous transform. Then, a dual-channel fusion shield machine feature extraction network is constructed, and the feature extraction of the spindle continuous wavelet transform image and the geological section image is integrated into one network. On the one hand, it significantly improves the depth of image feature fusion, and on the other hand, it effectively reduces the complexity of the feature extraction network. Secondly, principal component analysis and feature normalization are used to construct a multi-information perspective feature-level fusion vector to achieve cross-modal fusion of heterogeneous and heterogeneous perception data. Finally, the encoding layer and prediction head of the Vision Transformer are improved, and with the help of the attention mechanism and Transformer encoding, the tool wear prediction problem with multi-factor coupling is converted into a time series reasoning problem, effectively improving the prediction accuracy of the model.

[0009] In one implementation of the present application, the sample data also includes one of the historical maintenance information of the shield machine excavation parameters, the shield machine tool parameters and the shield machine tool wear; the acquired historical working data of the shield machine is sorted and divided to generate sample data of different categories, specifically including: sorting and dividing the historical working data of the shield machine in chronological order; dividing the sample data into sample input data and sample output data according to the detection data and maintenance data of the shield machine; filtering out invalid information in the non-working state of the shield machine and information irrelevant to the tool wear prediction, and using the interpolation method to fill the missing values of the filtered data; encoding the non-numerical information in the sample data by one-hot vector encoding; wherein the non-numerical information includes one of the soil type information, the tool position information and the segment selection information.

[0010] In one implementation of the present application, a dual-channel fusion shield machine lightweight feature extraction network is used to perform feature extraction on the wavelet transform image and geological cross-section of the spindle vibration signal in the sample data, specifically including: performing wavelet transformation on the spindle vibration signal to convert the spindle vibration signal into a wavelet transformation image; establishing a dual-channel fusion shield machine feature extraction network; wherein the dual-channel fusion shield machine feature extraction network includes a wavelet transform image feature extraction channel and a geological cross-section feature extraction channel; through different channels in the dual-channel fusion shield machine feature extraction network, feature extraction is performed on the wavelet transform image and geological cross-section of the spindle vibration signal, respectively.

[0011] In one implementation of the present application, the wavelet transform image feature extraction channel is a convolutional network structure including five layers, and the filling mode of each layer is set to Same; the size of the first convolution layer is 300×300×32, the number of convolution kernels is 32, the convolution step is [1 1], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the maximum pooling layer size is [5 5], and the step is [1 1]; the size of the second convolution layer is 150×150×48, the number of convolution kernels is 48, the convolution step is

[22] , the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the maximum pooling layer size is [5 5], and the step is [1 1]; the size of the third convolution layer is 75×75×64, the number of convolution kernels is 64, the convolution step is [2 2], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the maximum pooling layer size is [5 5], with a step size of [1 1]; the size of the fourth convolution layer is 38×38×128, the number of convolution kernels is 128, the convolution step size is [2 2], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the size of the maximum pooling layer is [5 5], with a step size of [1 1]; the size of the fifth convolution layer is 19×19×128, the number of convolution kernels is 128, the convolution step size is [2 2], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the size of the maximum pooling layer is [5 5], with a step size of [1 1]; a dropout layer is set between the batch normalization layer and the excitation layer, and the dropout probability is set to 0.5; the last layer of the wavelet transform image feature extraction channel is a global normalization layer, and the size of the global normalization layer is 1×1×256.

[0012] In one implementation of the present application, after feature extraction is performed on the wavelet transform image and geological cross-section of the main shaft vibration signal, respectively, the method further includes: combining and cascading the maximum pooling layers corresponding to the wavelet transform image feature extraction channel and the geological cross-section feature extraction channel through the feature cascade layer of the dual-channel fusion shield machine feature extraction network to obtain a cascaded feature vector; performing dimensionality reduction processing on the cascaded feature vector through the principal component analysis method, and transmitting the dimensionality reduction feature vector to the fully connected layer and the prediction regression layer.

[0013] In one implementation of the present application, a multi-perspective information fusion feature vector is constructed based on the feature vector corresponding to the dual-channel fusion shield machine lightweight feature extraction network and the feature vectors corresponding to other data in the sample data, specifically including: determining the 8-dimensional statistical characteristics of the main shaft vibration signal based on the main shaft vibration signal; and matching the current excavation position with the geological section map to obtain the current geological section information, and determining the 2-dimensional excavation geological information characteristics based on the current geological section information; and determining the 6-dimensional shield machine excavation parameter characteristics; and determining the 6-dimensional shield machine tool parameter characteristics; and determining the 36-dimensional extraction features corresponding to the dual-channel fusion shield machine lightweight feature extraction network; and fusing various features to obtain a 58-dimensional multi-perspective information fusion feature vector.

[0014] In one implementation of the present application, a time series feature vector is constructed based on a multi-perspective information fusion feature vector to construct an improved VIT shield machine tool wear prediction model through the time series feature vector, specifically including: constructing a time series feature vector on a time scale; wherein each time series contains multiple sampling moments, each sampling moment contains a 58-dimensional multi-perspective information fusion feature vector, and the input of each time series is a 10×58-dimensional feature vector; through feature normalization processing, the difference in data distribution range is eliminated, and the sequence position of the feature is marked through time coding, and the time series features are encoded through multiple superimposed Transformer encoding modules to improve the VIT shield machine tool wear prediction model; wherein each encoding module includes a normalization layer, a multi-head attention layer and a multi-layer perceptron; wherein the input of the improved VIT shield machine tool wear prediction model is the time series feature, and the output is the tool wear; the improved VIT shield machine tool wear prediction model is iteratively trained through supervised learning.

[0015] In one implementation of the present application, the parameters of the improved VIT shield machine tool wear prediction model are optimized, specifically including: using a test data set to test the improved VIT shield machine tool wear prediction model, and when the test prediction error is less than a preset error condition, determining that the test requirements are met; otherwise, using a genetic algorithm to optimize the parameters of the improved VIT shield machine tool wear prediction model, and retraining the model until the test accuracy of the model meets the preset error condition.

[0016] An embodiment of the present application provides a shield machine tool wear prediction device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can: sort and segment the acquired historical working data of the shield machine to generate sample data of different categories; perform feature extraction on the wavelet transform image and geological cross-section of the main shaft vibration signal in the sample data through a dual-channel fusion shield machine lightweight feature extraction network to obtain a feature vector; construct a multi-perspective information fusion feature vector based on the feature vector corresponding to the dual-channel fusion shield machine lightweight feature extraction network and the feature vector corresponding to other data in the sample data; construct a time series feature vector based on the multi-perspective information fusion feature vector, and construct an improved VIT shield machine tool wear prediction model through the time series feature vector; perform parameter optimization on the improved VIT shield machine tool wear prediction model, and predict the shield machine tool wear when the parameter optimization result meets the preset error condition.

[0017] A non-volatile computer storage medium provided by an embodiment of the present application stores computer-executable instructions, which are configured to: sort and segment the acquired historical working data of a shield machine to generate sample data of different categories; perform feature extraction on the wavelet transform image and geological cross-section of the main shaft vibration signal in the sample data through a dual-channel fusion shield machine lightweight feature extraction network to obtain a feature vector; construct a multi-perspective information fusion feature vector based on the feature vector corresponding to the dual-channel fusion shield machine lightweight feature extraction network and the feature vectors corresponding to other data in the sample data; construct a time series feature vector based on the multi-perspective information fusion feature vector, and construct an improved VIT shield machine tool wear prediction model through the time series feature vector; perform parameter optimization on the improved VIT shield machine tool wear prediction model, and predict the shield machine tool wear when the parameter optimization result meets the preset error condition.

[0018] At least one of the above technical solutions adopted in the embodiments of the present application can achieve the following beneficial effects: The embodiments of the present application construct multi-perspective complete tool prediction information, including spindle vibration signals, geological parameters, tunneling parameters, and shield machine tool parameters. In order to extract richer spindle vibration features, the embodiments of the present application convert the time domain signal of the spindle vibration into an image representation based on wavelet continuous transform. Then, a dual-channel fusion shield machine feature extraction network DCFSMNet is constructed to fuse the feature extraction of the spindle continuous wavelet transform image and the geological cross-section image in one network, which significantly improves the depth of image feature fusion on the one hand and effectively reduces the complexity of the feature extraction network on the other hand. Secondly, principal component analysis and feature normalization are used to construct a multi-information perspective feature-level fusion vector to achieve cross-modal fusion of heterogeneous and heterogeneous perception data. Finally, the encoding layer and prediction head of the Vision Transformer are improved, and with the help of the attention mechanism and Transformer encoding, the tool wear prediction problem with multi-factor coupling is converted into a time series reasoning problem, effectively improving the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments described in the present application. For those skilled in the art, other drawings can be obtained based on these drawings without inventive work. In the drawings:

[0020] Figure 1 A flow chart of a shield machine tool wear prediction method provided in an embodiment of the present application;

[0021] Figure 2 A DCFSMNet network structure diagram provided in an embodiment of the present application;

[0022] Figure 3 A continuous wavelet transform image of a spindle vibration signal provided in an embodiment of the present application;

[0023] Figure 4 A continuous wavelet transform image feature extraction visualization image provided in the embodiment of the present application;

[0024] Figure 5 A geological cross-section feature extraction visualization image provided in an embodiment of the present application;

[0025] Figure 6 A schematic diagram of an improved VTI tool wear prediction model provided in an embodiment of the present application;

[0026] Figure 7A flowchart of a shield machine tool wear prediction process provided in an embodiment of the present application;

[0027] Figure 8 A schematic structural diagram of a shield machine tool wear prediction device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The embodiments of the present application provide a shield machine tool wear prediction method, equipment and medium.

[0029] In order to enable those skilled in the art to better understand the technical solutions in this application, the following will clearly and completely describe the technical solutions in the embodiments of this application in conjunction with the drawings in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0030] The technical solutions proposed in the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0031] Figure 1 A flow chart of a shield machine tool wear prediction method provided in an embodiment of the present application is as follows: Figure 1 As shown in Figure 2, the shield machine tool wear prediction method includes the following steps:

[0032] S101. Sort and segment the acquired historical working data of the shield machine to generate sample data of different categories.

[0033] In one embodiment of the present application, historical working data of a shield machine is sorted and segmented in chronological order; sample data is divided into sample input data and sample output data based on the detection data and maintenance data of the shield machine; invalid information in a non-working state of the shield machine and information irrelevant to tool wear prediction are screened out, and missing values are filled in the screened out data using interpolation; non-numerical information in the sample data is encoded using one-hot vector encoding; wherein the non-numerical information includes one of soil type information, tool position information, and segment selection information.

[0034] Specifically, the embodiment of the present application collects historical working data of shield machines under different working environments, sorts and segments the historical working data in chronological order, and forms trainable sample data through data cleaning and encoding. Specifically, the following steps are included:

[0035] (1) Based on the inspection and maintenance data of the shield machine, the historical data is divided into sample input data and sample output data. The sample input data includes four types of information: spindle vibration signal, geological parameters, shield machine excavation parameters, and shield machine tool parameters. The sample output data is the historical maintenance information of the shield machine tool wear.

[0036] (2) Invalid information when the shield machine is not in working state and information irrelevant to tool wear prediction are eliminated, and the missing values of the data are filled using the interpolation method.

[0037] (3) One-hot vector encoding is used to encode non-numerical information in the data, such as soil type, tool position, and segment selection.

[0038] S102. By using a dual-channel fusion shield machine lightweight feature extraction network, feature extraction is performed on the wavelet transform image of the main shaft vibration signal and the geological cross-section diagram in the sample data to obtain a feature vector.

[0039] In one embodiment of the present application, shield machine tool wear prediction information feature extraction specifically includes the following steps:

[0040] (1) Extraction of spindle vibration information features. The spindle vibration information is sliced into time series, with a signal length of 3 seconds and a moving step of 1 second for each sampling. Statistical methods are used to extract seven types of statistical features from the spindle vibration signal, including mean, maximum, minimum, variance, median, kurtosis, and skewness.

[0041] (2) Extraction of geological information features during excavation. The current excavation position is matched with the geological section map drawn before construction to obtain the current geological section information. The current soil type and quartz content are determined based on the geological section information.

[0042] (3) Extraction of shield machine excavation parameter features. According to the construction specifications and current system settings, the shield machine excavation parameter information is obtained, including: total excavation thrust, cutter head torque, excavation speed, soil bin pressure, grouting pressure, cutter speed, screw speed, and modifier content, etc.

[0043] (4) Shield machine tool parameter feature extraction. Shield machine tool parameter information is obtained based on the current shield machine model and type, including: segment selection, cutterhead diameter, cumulative working time, segment ring width, stone ring thickness, tool position, inner diameter size, outer diameter size, and other tool features.

[0044] In one embodiment of the present application, a wavelet transform is performed on the spindle vibration signal to convert it into a wavelet-transformed image. A dual-channel fusion shield machine feature extraction network is established; the dual-channel fusion shield machine feature extraction network includes a wavelet transform image feature extraction channel and a geological cross-section feature extraction channel. Feature extraction is performed on the wavelet transform image of the spindle vibration signal and the geological cross-section image, respectively, using different channels in the dual-channel fusion shield machine feature extraction network.

[0045] Specifically, the embodiment of the present application proposes a dual-channel fusion shield machine lightweight feature extraction network DCFSMNet (Double-channel Fusion of Shield Machine for Feature Extraction Network), Figure 2 A DCFSMNet network structure diagram provided in an embodiment of the present application is as follows: Figure 2 As shown in the figure, the DCFSMNet network structure is used to extract features from the wavelet transform image of the main shaft vibration signal and the geological section map, and convert the high-dimensional image information into a low-dimensional trainable feature vector.

[0046] Furthermore, the spindle vibration signal is first converted into an image representation using continuous wavelet transform. Figure 3 A continuous wavelet transform image of a main shaft vibration signal provided in an embodiment of the present application is shown as follows: Figure 3 As shown, the wavelet transform uses the Morse (3, 60) wavelet to analyze the main axis signal, where 3 represents symmetry, 60 represents the time-bandwidth product, and the wavelet transform uses 10 parts per octave. The wavelet transformed image is normalized to a standard size of 300×300×3.

[0047] Furthermore, a dual-channel fusion shield machine feature extraction network DCFSMNet is established, such as Figure 2 As shown in Figure 3, the model contains 47 layers and has a total of only 827K parameters.

[0048] Furthermore, DCFSMNet contains two feature extraction channels. The backbone networks of the two feature extraction channels are symmetrically distributed, one channel is used for wavelet transform image feature extraction, and the other channel is used for geological section map feature extraction.

[0049] In one embodiment of the present application, the wavelet transform image feature extraction channel is a convolutional network structure including five layers, and the padding mode of each layer is set to Same. The size of the first convolution layer is 300×300×32, the number of convolution kernels is 32, the convolution step is [1 1], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the size of the maximum pooling layer is [5 5], and the step is [1 1]. The size of the second convolution layer is 150×150×48, the number of convolution kernels is 48, the convolution step is

[22] , the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the size of the maximum pooling layer is [5 5], and the step is [1 1]. The size of the third convolution layer is 75×75×64, the number of convolution kernels is 64, the convolution stride is [2 2], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the size of the maximum pooling layer is [5 5], and the stride is [1 1]. The size of the fourth convolution layer is 38×38×128, the number of convolution kernels is 128, the convolution stride is [2 2], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the size of the maximum pooling layer is [5 5], and the stride is [1 1]. The size of the fifth convolution layer is 19×19×128, the number of convolution kernels is 128, the convolution stride is [2 2], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the size of the maximum pooling layer is [5 5], and the stride is [1 1]. A dropout layer is set between the batch normalization layer and the excitation layer, and the dropout probability is set to 0.5. The last layer of the wavelet transform image feature extraction channel is a global normalization layer with a size of 1×1×256.

[0050] Specifically, in the embodiment of the present application, for the channel of wavelet transform image feature extraction, the network input is a 300×300×3 image, and the size of the first convolution layer is 300×300×32. Among them, 32 represents the number of convolution kernels in the convolution layer, the convolution step is [1 1], and the padding mode is "Same". Figure 4 The embodiment of the present application provides a continuous wavelet transform image feature extraction visualization image, the convolution layer 1 extracts some features of the visualization as shown in FIG. Figure 4 As shown in Figure 2, the batch normalization layer’s mean decay MeanDecay and decay noise variance VarianceDecay are both set to 0.1, the maximum pooling layer size is [5 5], and the stride is [1 1].

[0051] Furthermore, the size of the second convolution layer is 150×150×48, where 48 represents the number of convolution kernels of the convolution layer, the convolution step is [2 2], and the padding is “Same”. The settings of other layers are consistent with the first convolution layer, and the embodiments of this application will not be repeated. Furthermore, the size of the third convolution layer is 75×75×64, where 64 represents the number of convolution kernels of the convolution layer, the convolution step is [2 2], and the padding is “Same”. The settings of other layers are consistent with the first convolution layer, and the embodiments of this application will not be repeated. Furthermore, the size of the fourth convolution layer is 38×38×128, where 128 represents the number of convolution kernels of the convolution layer, the convolution step is [2 2], and the padding is “Same”. The settings of other layers are consistent with the first convolution layer, and the embodiments of this application will not be repeated. Furthermore, the size of the fifth convolutional layer is 19×19×128, where 128 represents the number of convolution kernels in the convolutional layer, the convolution stride is [2 2], and the padding is “Same”. A dropout layer is added between the batch normalization layer and the excitation layer, with a dropout probability set to 0.5. The last layer is a global normalization layer with a size of 1×1×256.

[0052] In one embodiment of the present application, a dual-channel fusion shield machine feature extraction network uses a feature concatenation layer to combine and concatenate the maximum pooling layers corresponding to the wavelet transform image feature extraction channel and the geological cross-section feature extraction channel, respectively, to obtain a concatenated feature vector. This concatenated feature vector is then subjected to dimensionality reduction using principal component analysis and fed to the fully connected layer and the prediction and regression layer.

[0053] Specifically, Figure 5 A geological cross-section feature extraction visualization image is provided in the embodiment of the present application, such as Figure 5 As shown in Figure 3, the network structure of geological cross-section feature extraction is consistent with the feature extraction network of wavelet transform images. The backbone network of each feature extraction channel contains 22 layers, and the 45th layer is a feature cascade layer. The maximum pooling layers (1×1×256) of the two feature extraction channels are combined and cascaded to obtain a 1×1×512 feature vector, followed by a fully connected layer and prediction regression.

[0054] Furthermore, the parameters of the feature concatenation layer are used as the feature output of DCFSMNet. This layer contains 1×512 features. To ensure the balance of feature fusion, the embodiment of the present application uses principal component analysis to reduce the dimensionality of the features, retaining 98% of the contribution rate of the original features and compressing the 1×512 feature vector into a 1×36 representation.

[0055] S103. Construct a multi-perspective information fusion feature vector based on the feature vector corresponding to the dual-channel fusion shield machine lightweight feature extraction network and the feature vectors corresponding to other data in the sample data.

[0056] In one embodiment of the present application, based on the spindle vibration signal, an 8-dimensional statistical feature of the spindle vibration signal is determined. Furthermore, the current excavation position is matched with a geological cross-section map to obtain current geological cross-section information, and based on this information, a 2-dimensional feature of the excavation geological information is determined. Furthermore, a 6-dimensional feature of the shield machine excavation parameter is determined. Furthermore, a 6-dimensional feature of the shield machine tool parameter is determined. Furthermore, a 36-dimensional feature extraction corresponding to a dual-channel fusion shield machine lightweight feature extraction network is determined. These various features are fused to obtain a 58-dimensional multi-view information fusion feature vector.

[0057] Specifically, the embodiment of the present application constructs a feature vector of multi-perspective information fusion, in which the statistical features of the spindle vibration signal are 8-dimensional, the excavation geological information features are 2-dimensional, the shield machine excavation parameter features are 6-dimensional, the shield machine tool parameter features are 6-dimensional, and the DCFSMNet network extraction features are 36-dimensional. Various features are fused to obtain a 58-dimensional feature vector of multi-perspective information fusion.

[0058] S104. Construct a time series feature vector based on the multi-view information fusion feature vector, and construct an improved VIT shield machine tool wear prediction model through the time series feature vector.

[0059] In one embodiment of the present application, a time series feature vector is constructed on a time scale; wherein each time series includes multiple sampling moments, each sampling moment includes a 58-dimensional multi-view information fusion feature vector, and the input of each time series is a 10×58-dimensional feature vector. Through feature normalization processing, the data distribution range is eliminated, and the sequence position of the feature is marked through time coding. The time series features are encoded through multiple superimposed Transformer encoding modules to improve the VIT shield machine tool wear prediction model; wherein each encoding module includes a normalization layer, multi-head attention and a multi-layer perceptron; wherein the input of the improved VIT shield machine tool wear prediction model is the time series feature, and the output is the tool wear. The improved VIT shield machine tool wear prediction model is iteratively trained through supervised learning.

[0060] Specifically, a time series shield machine tool wear prediction feature is constructed. A time series feature vector is constructed on the time scale. Each time series contains 10 sampling moments, and each moment contains a 58-dimensional feature vector of multi-view information fusion. The input of each time series is a 10×58-dimensional feature vector.

[0061] Furthermore, the VIT (Vision Transformer) model is improved. Figure 6 A schematic diagram of an improved VTI tool wear prediction model provided in an embodiment of the present application is shown in FIG. Figure 6 As shown. The input of the model is the time series feature, and the output is the wear of the tool. First, the feature normalization is performed to eliminate the difference in data distribution range, and the sequence position of the feature is marked by time coding. Then, multiple superimposed Transformer encoding modules are used to encode the time series features. Each module contains a normalization layer, multi-head attention and multilayer perceptron MLP (Multilayer Perceptron). The multilayer perceptron predicts the wear of the tool through nonlinear fitting.

[0062] Furthermore, model training and parameter setting are performed. The main parameter settings of the improved VIT model include the Adam solver, 10 attention heads, 4 MLP ratios, 1e-4 learning rate, and 16 batch sizes. The model is trained using supervised learning, with epochs set to 200, and model training stops after 200 iterations.

[0063] S105. Parameter optimization is performed on the improved VIT shield machine tool wear prediction model. When the parameter optimization result meets the preset error condition, the shield machine tool wear is predicted.

[0064] In one embodiment of the present application, a test dataset is used to test the improved VIT shield machine tool wear prediction model. If the test prediction error is less than a preset error condition, the test requirements are determined to be met. Otherwise, a genetic algorithm is used to optimize the parameters of the improved VIT shield machine tool wear prediction model and the model is retrained until the model's test accuracy meets the preset error condition.

[0065] Specifically, the embodiment of the present application uses a test set to test the model, optimizes the model that does not meet the test accuracy, and deploys the model that meets the test accuracy. The specific steps are as follows:

[0066] (1) The model is tested using a test data set. If the test prediction error is less than 6%, the test requirements are met. If the test accuracy is not met, the model parameters are optimized using a genetic algorithm and the model is retrained until the model's test accuracy meets the requirements. It should be noted that the embodiment of the present application preferably sets the error threshold to 6%. It can be adjusted according to actual needs in the application, and the embodiment of the present application does not limit this.

[0067] (2) Use TensorRT to perform inference deployment on the model that meets the test accuracy, deploy the optimized and trimmed model to the NVIDIA edge computing device, and define the data interface.

[0068] Figure 7 A flowchart of a shield machine tool wear prediction process provided in an embodiment of the present application is shown as follows: Figure 7 As shown, the shield machine tool wear prediction information feature extraction is first performed, including: (1) spindle vibration information feature extraction. The spindle vibration information is sliced into time series, the signal length is 3 seconds, and the moving step length of each signal sampling is 1 second. Statistical methods are used to extract 7 types of statistical features from the spindle vibration signal, including: mean, maximum, minimum, variance, median, kurtosis and skewness; (2) excavation geological information feature extraction. The current excavation position is matched with the geological section map drawn before construction to obtain the current geological section information. The current soil type and quartz content are determined based on the address section information. (3) Shield machine excavation parameter feature extraction. According to the construction specifications and current system settings, the shield excavation parameter information is obtained, including: total excavation thrust, cutter head torque, excavation speed, soil bin pressure, grouting pressure, tool speed, screw speed, and improver content and other excavation features. (4) Shield machine tool parameter feature extraction: According to the current shield machine model and type, shield machine tool parameter information is obtained, including: segment selection, cutterhead diameter, cumulative working time, segment ring width, stone ring thickness, tool position, inner diameter size, outer diameter size and other tool features. A dual-channel fusion shield machine feature extraction network DCFSMNet is established. The backbone network of the two feature extraction channels is symmetrically distributed, one channel is used for wavelet transform image feature extraction, and the other channel is used for geological cross-section feature extraction. Secondly, based on the acquired features, a feature vector of multi-view information fusion is constructed to construct a time series shield machine tool wear prediction feature. The time series feature vector is constructed on the time scale. Finally, the encoding layer and prediction head of the Vision Transformer are improved. With the help of the attention mechanism and Transformer encoding, the tool wear prediction problem with multi-factor coupling is transformed into a time series reasoning problem, effectively improving the prediction accuracy of the model.

[0069] Figure 8 This is a schematic diagram of the structure of a shield machine tool wear prediction device provided in an embodiment of the present application. Figure 8As shown, a shield machine tool wear prediction device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can: sort and segment the acquired historical working data of the shield machine to generate sample data of different categories; extract features of the wavelet transform image and geological cross-section of the main shaft vibration signal in the sample data through a dual-channel fusion shield machine lightweight feature extraction network to obtain a feature vector; construct a multi-view information fusion feature vector based on the feature vector corresponding to the dual-channel fusion shield machine lightweight feature extraction network and the feature vector corresponding to other data in the sample data; construct a time series feature vector based on the multi-view information fusion feature vector, and construct an improved VIT shield machine tool wear prediction model through the time series feature vector; optimize the parameters of the improved VIT shield machine tool wear prediction model, and predict the shield machine tool wear when the parameter optimization result meets the preset error condition.

[0070] A non-volatile computer storage medium provided by an embodiment of the present application stores computer-executable instructions, which are configured to: sort and segment the acquired historical working data of a shield machine to generate sample data of different categories; perform feature extraction on the wavelet transform image and geological cross-section of the main shaft vibration signal in the sample data through a dual-channel fusion shield machine lightweight feature extraction network to obtain a feature vector; construct a multi-perspective information fusion feature vector based on the feature vector corresponding to the dual-channel fusion shield machine lightweight feature extraction network and the feature vectors corresponding to other data in the sample data; construct a time series feature vector based on the multi-perspective information fusion feature vector, and construct an improved VIT shield machine tool wear prediction model through the time series feature vector; perform parameter optimization on the improved VIT shield machine tool wear prediction model, and predict the shield machine tool wear when the parameter optimization result meets the preset error condition.

[0071] The various embodiments in this application are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from the other embodiments. In particular, the device, apparatus, and non-volatile computer storage medium embodiments are generally similar to the method embodiments, so their descriptions are relatively simple. For relevant portions, refer to the descriptions of the method embodiments.

[0072] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. It will be apparent to those skilled in the art that various modifications and variations may be made to the embodiments of the present application. However, such modifications or substitutions do not deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A shield machine tool wear prediction method, characterized in that: The method comprises: The acquired historical working data of the shield machine are sorted and divided to generate sample data of different categories; the main shaft vibration in the sample data is extracted by a dual-channel fusion shield machine lightweight feature extraction network. The wavelet transform image of the motion signal and the geological cross-section are used to extract features to obtain feature vectors; Based on the feature vector corresponding to the dual-channel fusion shield machine lightweight feature extraction network and the feature vectors corresponding to other data in the sample data, a multi-view information fusion feature vector is constructed, specifically including: Based on the spindle vibration signal, multi-dimensional statistical characteristics of the spindle vibration signal are determined; and the current excavation position is matched with the geological section map to obtain current geological section information, so as to determine multi-dimensional excavation geological information characteristics based on the current geological section information; and multi-dimensional shield machine excavation parameter characteristics are determined; and multi-dimensional shield machine tool parameter characteristics are determined; and multi-dimensional extraction features corresponding to the dual-channel fusion shield machine lightweight feature extraction network are determined; and various features are fused to obtain a feature vector of multi-dimensional and multi-view information fusion; A time series feature vector is constructed based on the multi-view information fusion feature vector, so as to construct an improved VIT (Vision Transformer) shield machine tool wear prediction model through the time series feature vector, specifically comprising: constructing a time series feature vector on a time scale; wherein each time series includes multiple sampling moments, each sampling moment includes a multi-dimensional multi-view information fusion feature vector, and the input of each time series is a 10× multi-dimensional feature vector; wherein the multi-dimensional feature vector is obtained by multi-dimensional multi-view information fusion; through feature normalization processing, the difference in data distribution range is eliminated, and the sequence position of the feature is marked through time coding, and the time series features are encoded through multiple superimposed Transformer encoding modules to improve the VIT shield machine tool wear prediction model; wherein each encoding module includes a normalization layer, a multi-head attention layer, and a multi-layer perceptron; wherein the input of the improved VIT shield machine tool wear prediction model is the time series feature, and the output is the tool wear; and the improved VIT shield machine tool wear prediction model is iteratively trained through supervised learning; Parameters of the improved VIT shield machine tool wear prediction model are optimized, and when the parameter optimization results meet the preset error conditions, the shield machine tool wear is predicted.

2. A shield machine tool wear prediction method according to claim 1, characterized in that: The sample data also includes one of the following items: shield machine excavation parameters, shield machine tool parameters, and shield machine tool wear history maintenance information; The obtained shield machine historical working data is sorted and divided to generate sample data of different categories, specifically including: Sorting and dividing the historical working data of the shield machine in chronological order; According to the detection data and maintenance data of the shield machine, the sample data is divided into sample input data and sample output data; Invalid information when the shield machine is not in operation and information irrelevant to tool wear prediction are screened out, and the missing values of the screened data are filled using interpolation method. The non-numerical information in the sample data is encoded by one-hot vector encoding; wherein the non-numerical information includes one of soil type information, tool position information and segment selection information.

3. A shield machine tool wear prediction method according to claim 1, characterized in that: The dual-channel fusion shield machine lightweight feature extraction network is used to extract the main shaft vibration in the sample data. The wavelet transform image of the motion signal and the geological cross-section are used for feature extraction, including: performing a wavelet transform on the main shaft vibration signal to convert the main shaft vibration signal into a wavelet transform image; Establishing a dual-channel fusion shield machine feature extraction network; wherein the dual-channel fusion shield machine feature extraction network includes a wavelet transform image feature extraction channel and a geological section map feature extraction channel; By using different channels in the dual-channel fusion shield machine feature extraction network, feature extraction is performed on the wavelet transform image of the main shaft vibration signal and the geological cross-section diagram respectively.

4. A shield machine tool wear prediction method according to claim 1, characterized in that: The wavelet transform image feature extraction channel is a convolutional network structure including five layers, and the filling mode of each layer is set to Same; The size of the first convolution layer is 300×300×32, the number of convolution kernels is 32, the convolution step is [1 1], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the size of the maximum pooling layer is [5 5], and the step is [1 1]; The size of the second convolution layer is 150×150×48, the number of convolution kernels is 48, the convolution step is [2 2], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the size of the maximum pooling layer is [5 5], and the step is [1 1]; The size of the third convolution layer is 75×75×64, the number of convolution kernels is 64, the convolution step is [2 2], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the size of the maximum pooling layer is [5 5], and the step is [1 1]; The size of the fourth convolution layer is 38×38×128, the number of convolution kernels is 128, the convolution step is [2 2], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the size of the maximum pooling layer is [5 5], and the step is [1 1]; The size of the fifth convolution layer is 19×19×128, the number of convolution kernels is 128, the convolution step is [2 2], the average attenuation and attenuation noise variance of the batch normalization layer are both set to 0.1, the size of the maximum pooling layer is [5 5], and the step is [1 1]; Among them, a dropout layer is set between the batch normalization layer and the incentive layer, and the dropout probability is set to 0.5; The last layer of the wavelet transform image feature extraction channel is a global normalization layer, and the size of the global normalization layer is 1×1×256.

5. The shield machine tool wear prediction method according to claim 1, characterized in that: After extracting features from the wavelet transform image of the main shaft vibration signal and the geological cross-section, the method further includes: Through the feature cascade layer of the dual-channel fusion shield machine feature extraction network, the maximum pooling layers corresponding to the wavelet transform image feature extraction channel and the geological section map feature extraction channel are combined and cascaded to obtain a cascaded feature vector; The cascaded feature vector is subjected to dimensionality reduction processing by principal component analysis, and the feature vector after dimensionality reduction processing is transmitted to the fully connected layer and the prediction regression layer.

6. A shield machine tool wear prediction method according to claim 1, characterized in that: The parameter optimization of the improved VIT shield machine tool wear prediction model specifically includes: The improved VIT shield machine tool wear prediction model is tested using a test data set, and when the test prediction error is less than a preset error condition, it is determined that the test requirements are met; Otherwise, a genetic algorithm is used to optimize the parameters of the improved VIT shield machine tool wear prediction model, and the model is retrained until the test accuracy of the model meets the preset error condition.

7. A shield machine tool wear prediction device, characterized in that: The device comprises a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to execute the method according to any one of claims 1 to 6.

8. A non-volatile computer storage medium storing computer-executable instructions, characterized in that: The computer executable instructions can execute the method according to any one of claims 1 to 6.

Images