Multi-defect identification method for shield main bearing

Abstract

The invention discloses a multi-defect identification method for a shield main bearing. The method comprises the steps of multi-source signal acquisition, feature extraction, feature fusion and normalization, mode identification and threshold analysis. After a vibration signal and a current signal are collected, time domain and frequency domain feature extraction is carried out respectively, a feature vector set is constructed through fusion and normalization processing is carried out; training and classifying the normalized feature vector set by adopting a convolutional neural network mode recognition algorithm, recognizing whether the main bearing has the defects of cracks and spalling or not, and accurately distinguishing the defects; and meanwhile, a threshold analysis rule is set to carry out verification and auxiliary judgment on a pattern recognition result. According to the method, by fusing multi-source signal information, constructing a complete identification process and combining mode identification and threshold analysis, the accuracy and reliability of multi-defect identification of the shield main bearing can be effectively improved, and powerful support is provided for equipment maintenance and safe operation.

Description

Technical Field

[0001] This invention relates to the field of tunnel boring machine technology, and in particular to a method for identifying multiple defects in the main bearing of a tunnel boring machine. Background Technology

[0002] Tunnel boring machines (TBMs) play a crucial role in tunnel excavation, and the main bearing, as a core component of the TBM, directly affects the stability and construction efficiency of the entire equipment. During long-term operation, main bearings are prone to various typical defects such as cracks and spalling. If these defects are not identified and addressed promptly and accurately, they can lead to equipment failure or even serious accidents.

[0003] However, existing methods for identifying defects in main bearings often have limitations. They struggle to comprehensively analyze multiple source signals, such as vibration and current, simultaneously, and their ability to accurately distinguish different defect types still needs improvement. Therefore, there is an urgent need for a method that can effectively fuse multi-source signal features and utilize advanced pattern recognition algorithms to achieve accurate identification of various defects. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for identifying multiple defects in the main bearing of a tunnel boring machine.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying multiple defects in the main bearing of a tunnel boring machine includes the following specific steps: S1: Collects multi-source signals during the operation of the shield tunnel main bearing, including vibration signals and current signals; S2: Extract time-domain features from the collected vibration signal and calculate its mean, variance, root mean square value, and kurtosis; extract frequency-domain features from the vibration signal, obtain the spectrum using fast Fourier transform, and determine the prominent frequency points and their amplitudes in the spectrum; extract time-domain features from the current signal and calculate its mean, variance, root mean square value, and peak value. S3: The extracted vibration signal time-domain features, frequency-domain features, and current signal time-domain features are fused to construct a feature vector set; S4: Normalize the constructed feature vector set so that the feature data are located in the interval [0,1]. S5: The pattern recognition algorithm is used to train and classify the normalized feature vector set to identify and accurately distinguish whether the main bearing has various typical defects such as cracks and spalling. S6: Set threshold analysis rules to verify and assist in the judgment of pattern recognition results. When the confidence level of the pattern recognition result is lower than the threshold, combine the threshold analysis to further confirm the defect type or judge it as an abnormal state.

[0006] As a further technical solution of the present invention, in S1, the vibration signal is acquired using a high-precision accelerometer with a sampling frequency of not less than 10kHz; the current signal is acquired using a current transformer with a sampling frequency of not less than 1kHz.

[0007] As a further technical solution of the present invention, in step S2, when extracting frequency domain features from the vibration signal, the spectrum is segmented to determine the characteristic frequency points and their amplitudes in different frequency bands, so as to more comprehensively reflect the vibration characteristics of the main bearing.

[0008] As a further technical solution of the present invention, step S2, specifically includes extracting frequency domain features from the vibration signal, including: a: Fast Fourier Transform: Let the acquired vibration signal sequence be... The sampling frequency is The number of sampling points is The spectrum is then obtained through Fast Fourier Transform (FFT). The calculation formula is: ; b: Calculate frequency and amplitude: the calculated spectrum The corresponding frequency is: The spectral amplitude is taken from the amplitude spectrum, and the calculation formula is: ; c: Determine the characteristic frequency points and their amplitudes: First, calculate the amplitude of all frequency points, then sort them from largest to smallest amplitude, and take the first... The frequency points corresponding to each amplitude value are used as feature frequency points, and together with their corresponding amplitude values, they form a frequency domain feature vector; where The number of preset characteristic frequency points is determined based on actual experience or needs.

[0009] As a further technical solution of the present invention, S3 specifically includes the following steps: S31: Definition of Feature Dimension: Let the time-domain feature vector of the vibration signal be... ,in The average value of the vibration signal. This represents the standard deviation (square root of variance) of the vibration signal. The root mean square value of the vibration signal. The kurtosis value of the vibration signal; the frequency domain feature vector of the vibration signal is ,in For characteristic frequency points, The amplitude at the corresponding characteristic frequency point is; the time-domain feature vector of the current signal is... ,in The average value of the current signal. This represents the standard deviation (square root of variance) of the current signal. This is the root mean square value of the current signal. The peak value of the current signal; S32: Feature Fusion to Construct a Feature Vector Set: Combining the Time-Domain Feature Vectors of Vibration Signals and the time-domain eigenvector of the current signal Concatenate them sequentially to form a complete feature vector. ,Right now: .

[0010] As a further technical solution of the present invention, S4 specifically includes the following steps: S41: Define feature dimension and sample size: Let the constructed feature vector set , where each feature vector Include Each feature dimension The number of samples; S42: Calculate the maximum and minimum values ​​for each feature dimension: For each feature dimension Calculate its maximum value among all samples. and minimum value ; ; ; S43: Perform a linear transformation on each feature dimension: for each feature vector Each feature dimension The normalized set of eigenvectors is obtained by performing a linear transformation according to the following formula. : ;in, Indicates the first The th eigenvector of the th feature vector The original values ​​of each feature dimension. Indicates the first The th eigenvector of the th feature vector The normalized values ​​of each feature dimension.

[0011] As a further technical solution of the present invention, in step S5, the pattern recognition algorithm adopts a convolutional neural network based on deep learning, the structure of which includes an input layer, multiple convolutional layers, a pooling layer, a fully connected layer, and an output layer; the input layer receives a normalized feature vector, the dimension of which is... (Consistent with the feature dimension of the feature vector); a convolutional layer consists of multiple convolutional kernels, let the first kernel be... Each convolutional layer has There are n convolutional kernels, each with a size of n. Step size is The padding method is "valid" or "same"; the pooling layer uses max pooling or average pooling, and the pooling window size is [size missing]. Step size is The fully connected layer maps the features extracted by the convolutional and pooling layers to the category space. Let the first layer be the first layer. There are 1 fully connected layer The number of neurons in the output layer equals the number of defect categories. (e.g., cracks, peeling, etc.), the softmax activation function is used to output the probability of each category.

[0012] As a further technical solution of the present invention, S5 specifically includes the following steps: S51: Training process: Define the loss function as the cross-entropy loss function: ; in, It is the true label of the sample (one-hot encoded form). The sample predicted by the network belongs to the first... The probability of the class; The network weights are updated using stochastic gradient descent (SGD) or Adam optimization algorithms, with a learning rate of... The number of iterations is epoch; S52: Classification process: Input the normalized feature vector into the trained CNN model, output the probability distribution of each sample belonging to different defect categories, and take the category corresponding to the maximum probability as the classification result.

[0013] As a further technical solution of the present invention, S6 specifically includes the following steps: S61: Define the confidence threshold and feature threshold: Set the confidence threshold for the pattern recognition result as follows. (Typically, the value range is 0-1, such as 0.7, 0.8, etc.), used to judge the reliability of the pattern recognition results; for each typical defect of the main bearing (such as cracks, spalling, etc.), corresponding feature thresholds are set according to historical data and expert experience; taking crack defects as an example, the threshold for the root mean square value of the vibration signal is set to... The threshold for vibration signal kurtosis is The threshold value of the current signal peak is Taking peeling defects as an example, the corresponding feature threshold is set as follows: And so on; S62: Confidence Determination of Pattern Recognition Results: For each sample output by the pattern recognition algorithm, the probability distribution of belonging to various types of defects is used as the confidence level of the pattern recognition result. ,Right now: ,in Indicates that the sample belongs to the first The probability of class defects, Total number of defect categories; S63: Threshold Analysis Logic: If confidence level If the confidence level is low, the defect category corresponding to the pattern recognition result is directly adopted as the final judgment result; if the confidence level is low, the defect category corresponding to the pattern recognition result is adopted as the final judgment result. If so, the threshold analysis auxiliary judgment process will be triggered.

[0014] As a further technical solution of the present invention, the trigger threshold analysis auxiliary determination process in S63 specifically includes: S631: For each type of typical defect, check whether the corresponding feature value of the current sample exceeds the feature threshold of that type of defect; for example, check the root mean square value of the vibration signal. Is it greater than or vibration signal kurtosis value Is it greater than or Peak current signal Is it greater than or ; S632: If the feature value of a sample meets the feature threshold condition of a certain type of defect, it is determined to be a defect of that type; if it meets the feature threshold conditions of multiple defect categories, it is determined by combining expert experience or further diagnostic rules; if it does not meet the feature threshold conditions of any type of defect, it is determined to be an abnormal state, indicating that further inspection and analysis of the main bearing's operating status is required.

[0015] The beneficial effects of this invention are as follows: 1. This invention, by integrating multi-source signals such as vibration and current during the operation of the main bearing, can more comprehensively obtain information about the main bearing under different working conditions, which improves the accuracy and reliability of defect identification compared with single signal analysis methods; 2. This invention constructs a complete process of feature extraction and pattern recognition. By utilizing time-domain and frequency-domain feature extraction, it fully mines the key feature information in the main bearing signal. Combined with advanced pattern recognition algorithms, it can effectively learn and identify the feature patterns of various typical defects such as cracks and spalling in the main bearing, achieving accurate differentiation and improving the level of intelligent defect identification. 3. This invention integrates threshold analysis and pattern recognition algorithms, and performs threshold verification and auxiliary judgment based on the pattern recognition results, which further enhances the accuracy and stability of defect identification, reduces the false judgment rate, and provides a more reliable basis for fault diagnosis and maintenance of the main bearing of the tunnel boring machine. It helps to discover and deal with the defects of the main bearing in a timely manner, and ensures the normal operation of the tunnel boring machine and construction safety. Attached Figure Description

[0016] Figure 1This is a flowchart of a method for identifying multiple defects in the main bearing of a tunnel boring machine (TBM) proposed in this invention. Detailed Implementation

[0017] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0018] Please see the appendix Figure 1 A method for identifying multiple defects in the main bearing of a tunnel boring machine includes the following specific steps: S1: Collects multi-source signals during the operation of the shield tunnel main bearing, including vibration signals and current signals; S2: Extract time-domain features from the collected vibration signal and calculate its mean, variance, root mean square value, and kurtosis; extract frequency-domain features from the vibration signal, obtain the spectrum using fast Fourier transform, and determine the prominent frequency points and their amplitudes in the spectrum; extract time-domain features from the current signal and calculate its mean, variance, root mean square value, and peak value. The extraction of time-domain features from vibration signals specifically includes: 1. Mean Calculation: Let the collected vibration signal sequence be... ,in Number of sampling points, mean The calculation formula is: ; 2. Variance Calculation: Variance The formula reflecting the degree of dispersion of a signal is as follows: ; 3. Root Mean Square (RMS) Value Calculation: Root Mean Square (RMS) The effective amplitude of a signal is represented by the following formula: ; 4. Kurtosis Calculation: Kurtosis Value The formula used to measure the flatness of a signal waveform is: .

[0019] The extraction of time-domain features from current signals specifically includes: 1. Calculation of the mean value of the current signal: Let the acquired current signal sequence be... ,in The number of sampling points for the current signal, and the mean value. The calculation formula is: ; 2. Calculation of current signal variance: Variance The formula reflecting the dispersion of the current signal is as follows: ; 3. Calculation of the root mean square (RMS) value of the current signal: RMS value The effective amplitude of a current signal is characterized by the following formula: ; 4. Calculation of peak value of current signal: The peak value of the current signal. It is the maximum amplitude of the signal, calculated using the following formula: .

[0020] S3: The extracted vibration signal time-domain features, frequency-domain features, and current signal time-domain features are fused to construct a feature vector set; S4: Normalize the constructed feature vector set so that the feature data are located in the interval [0,1]. S5: The pattern recognition algorithm is used to train and classify the normalized feature vector set to identify and accurately distinguish whether the main bearing has various typical defects such as cracks and spalling. S6: Set threshold analysis rules to verify and assist in the judgment of pattern recognition results. When the confidence level of the pattern recognition result is lower than the threshold, combine the threshold analysis to further confirm the defect type or judge it as an abnormal state.

[0021] In a preferred embodiment, in S1, the vibration signal is acquired using a high-precision accelerometer with a sampling frequency of not less than 10kHz; the current signal is acquired using a current transformer with a sampling frequency of not less than 1kHz.

[0022] In a preferred embodiment, in S2, when extracting frequency domain features from the vibration signal, the spectrum is segmented to determine the characteristic frequency points and their amplitudes in different frequency bands, so as to more comprehensively reflect the vibration characteristics of the main bearing.

[0023] In a preferred embodiment, step S2, extracting frequency domain features from the vibration signal, specifically includes: a: Fast Fourier Transform: Let the acquired vibration signal sequence be... The sampling frequency is The number of sampling points is The spectrum is then obtained through Fast Fourier Transform (FFT). The calculation formula is: ; b: Calculate frequency and amplitude: the calculated spectrum The corresponding frequency is: The spectral amplitude is taken from the amplitude spectrum, and the calculation formula is: ; c: Determine the characteristic frequency points and their amplitudes: First, calculate the amplitude of all frequency points, then sort them from largest to smallest amplitude, and take the first... The frequency points corresponding to each amplitude value are used as feature frequency points, and together with their corresponding amplitude values, they form a frequency domain feature vector; where The number of preset characteristic frequency points is determined based on actual experience or needs.

[0024] In a preferred embodiment, S3 specifically includes the following steps: S31: Definition of Feature Dimension: Let the time-domain feature vector of the vibration signal be... ,in The average value of the vibration signal. This represents the standard deviation (square root of variance) of the vibration signal. The root mean square value of the vibration signal. The kurtosis value of the vibration signal; the frequency domain feature vector of the vibration signal is ,in For characteristic frequency points, The amplitude at the corresponding characteristic frequency point is; the time-domain feature vector of the current signal is... ,in The average value of the current signal. This represents the standard deviation (square root of variance) of the current signal. This is the root mean square value of the current signal. The peak value of the current signal; S32: Feature Fusion to Construct a Feature Vector Set: Combining the Time-Domain Feature Vectors of Vibration Signals and the time-domain eigenvector of the current signal Concatenate them sequentially to form a complete feature vector. ,Right now: .

[0025] In a preferred embodiment, S4 specifically includes the following steps: S41: Define feature dimension and sample size: Let the constructed feature vector set , where each feature vector Include Each feature dimension The number of samples; S42: Calculate the maximum and minimum values ​​for each feature dimension: For each feature dimension Calculate its maximum value among all samples. and minimum value ; ; ; S43: Perform a linear transformation on each feature dimension: for each feature vector Each feature dimension The normalized set of eigenvectors is obtained by performing a linear transformation according to the following formula. : ;in, Indicates the first The th eigenvector of the th feature vector The original values ​​of each feature dimension. Indicates the first The th eigenvector of the th feature vector The normalized values ​​of each feature dimension.

[0026] In a preferred embodiment, in S5, the pattern recognition algorithm employs a deep learning-based convolutional neural network, the structure of which includes an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer; the input layer receives a normalized feature vector with a dimension of [missing information]. (Consistent with the feature dimension of the feature vector); a convolutional layer consists of multiple convolutional kernels, let the first kernel be... Each convolutional layer has There are n convolutional kernels, each with a size of n. Step size is The padding method is "valid" or "same"; the pooling layer uses max pooling or average pooling, and the pooling window size is [size missing]. Step size is The fully connected layer maps the features extracted by the convolutional and pooling layers to the category space. Let the first layer be the first layer. There are 1 fully connected layer The number of neurons in the output layer equals the number of defect categories. (e.g., cracks, peeling, etc.), the softmax activation function is used to output the probability of each category.

[0027] In a preferred embodiment, S5 specifically includes the following steps: S51: Training process: Define the loss function as the cross-entropy loss function: ; in, It is the true label of the sample (one-hot encoded form). The sample predicted by the network belongs to the first... The probability of the class; The network weights are updated using stochastic gradient descent (SGD) or Adam optimization algorithms, with a learning rate of... The number of iterations is epoch; S52: Classification process: Input the normalized feature vector into the trained CNN model, output the probability distribution of each sample belonging to different defect categories, and take the category corresponding to the maximum probability as the classification result.

[0028] In a preferred embodiment, S6 specifically includes the following steps: S61: Define the confidence threshold and feature threshold: Set the confidence threshold for the pattern recognition result as follows. (Typically, the value range is 0-1, such as 0.7, 0.8, etc.), used to judge the reliability of the pattern recognition results; for each typical defect of the main bearing (such as cracks, spalling, etc.), corresponding feature thresholds are set according to historical data and expert experience; taking crack defects as an example, the threshold for the root mean square value of the vibration signal is set to... The threshold for the kurtosis value of the vibration signal is The threshold value of the current signal peak is Taking peeling defects as an example, the corresponding feature threshold is set as follows: And so on; S62: Confidence Determination of Pattern Recognition Results: For each sample output by the pattern recognition algorithm, the probability distribution of belonging to various types of defects is used as the confidence level of the pattern recognition result. ,Right now: ,in Indicates that the sample belongs to the first The probability of class defects, Total number of defect categories; S63: Threshold Analysis Logic: If confidence level If the confidence level is low, the defect category corresponding to the pattern recognition result is directly adopted as the final judgment result; if the confidence level is low, the defect category corresponding to the pattern recognition result is adopted as the final judgment result. If so, the threshold analysis auxiliary judgment process will be triggered.

[0029] In a preferred embodiment, S63 specifically includes the following steps in triggering the threshold analysis-assisted determination process: S631: For each type of typical defect, check whether the corresponding feature value of the current sample exceeds the feature threshold of that type of defect; for example, check the root mean square value of the vibration signal. Is it greater than or vibration signal kurtosis value Is it greater than or Circuit signal peak value Is it greater than or ; S632: If the feature value of a sample meets the feature threshold condition of a certain type of defect, it is determined to be a defect of that type; if it meets the feature threshold conditions of multiple defect categories, it is determined by combining expert experience or further diagnostic rules; if it does not meet the feature threshold conditions of any type of defect, it is determined to be an abnormal state, indicating that further inspection and analysis of the main bearing's operating status is required.

[0030] As can be seen from the above description, the above embodiments of the present invention achieve the following technical effects: by integrating multi-source signals such as vibration and current during the operation of the main bearing, information about the main bearing under different working conditions can be obtained more comprehensively, which improves the accuracy and reliability of defect identification compared with single signal analysis methods.

[0031] A complete process of feature extraction and pattern recognition was constructed. By utilizing time-domain and frequency-domain feature extraction, key feature information in the main bearing signal was fully explored. Combined with advanced pattern recognition algorithms, it can effectively learn and identify the feature patterns of various typical defects such as cracks and spalling in the main bearing, achieve accurate differentiation, and improve the level of intelligent defect identification.

[0032] By integrating threshold analysis and pattern recognition algorithms, threshold verification and auxiliary judgment are performed based on the pattern recognition results, which further enhances the accuracy and stability of defect identification, reduces the false judgment rate, and provides a more reliable basis for fault diagnosis and maintenance of the main bearing of the tunnel boring machine. This helps to detect and deal with the defects of the main bearing in a timely manner, and ensures the normal operation of the tunnel boring machine and construction safety.

[0033] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in the details for the sake of brevity.

[0034] This invention is intended to cover all such substitutions, modifications, and variations falling within the broad scope of the claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for identifying multiple defects in the main bearing of a tunnel boring machine, characterized in that, The specific steps include the following: S1: Collects multi-source signals during the operation of the shield tunnel main bearing, including vibration signals and current signals; S2: Extract time-domain features from the collected vibration signal and calculate its mean, variance, root mean square value, and kurtosis value; extract frequency-domain features from the vibration signal, obtain the spectrum using fast Fourier transform, and determine the prominent characteristic frequency points and their amplitudes in the spectrum. Time-domain features of the current signal are extracted, and its mean, variance, root mean square value, and peak value are calculated. S3: The extracted vibration signal time-domain features, frequency-domain features, and current signal time-domain features are fused to construct a feature vector set; S4: Normalize the constructed feature vector set so that the feature data are located in the interval [0,1]. S5: Use pattern recognition algorithms to train and classify the normalized feature vector set; S6: Set threshold analysis rules to verify and assist in the judgment of pattern recognition results. When the confidence level of the pattern recognition result is lower than the threshold, combine the threshold analysis to further confirm the defect type or judge it as an abnormal state.

2. The method for identifying multiple defects in a shield tunnel main bearing according to claim 1, characterized in that, In S1, the vibration signal is acquired using a high-precision accelerometer with a sampling frequency of not less than 10kHz; the current signal is acquired using a current transformer with a sampling frequency of not less than 1kHz.

3. The method for identifying multiple defects in a shield tunnel main bearing according to claim 1, characterized in that, In step S2, when extracting frequency domain features from the vibration signal, the characteristic frequency points and their amplitudes in different frequency bands are determined by segmenting the spectrum.

4. The method for identifying multiple defects in a shield tunnel main bearing according to claim 3, characterized in that, In step S2, the frequency domain feature extraction of the vibration signal specifically includes: a: Fast Fourier Transform: Let the acquired vibration signal sequence be... The sampling frequency is The number of sampling points is The spectrum is then obtained through Fast Fourier Transform. The calculation formula is: ; b: Calculate frequency and amplitude: the calculated spectrum The corresponding frequency is: The spectral amplitude is taken from the amplitude spectrum, and the calculation formula is: ; c: Determine the characteristic frequency points and their amplitudes: First, calculate the amplitude of all frequency points, then sort them from largest to smallest amplitude, and take the first... The frequency points corresponding to each amplitude are used as feature frequency points, and together with the corresponding amplitude, they form a frequency domain feature vector.

5. The method for identifying multiple defects in a shield tunnel main bearing according to claim 1, characterized in that, S3 specifically includes the following steps: S31: Definition of Feature Dimension: Let the time-domain feature vector of the vibration signal be... ,in The average value of the vibration signal. The standard deviation of the vibration signal. The root mean square value of the vibration signal. The kurtosis value of the vibration signal; the frequency domain feature vector of the vibration signal is ,in For characteristic frequency points, The amplitude at the corresponding characteristic frequency point is; the time-domain feature vector of the current signal is... ,in The average value of the current signal. The standard deviation of the current signal. This is the root mean square value of the current signal. The peak value of the current signal; S32: Feature Fusion to Construct a Feature Vector Set: Combining the Time-Domain Feature Vectors of Vibration Signals and the time-domain eigenvector of the current signal Concatenate them sequentially to form a complete feature vector. ,Right now: .

6. The method for identifying multiple defects in a shield tunnel main bearing according to claim 1, characterized in that, S4 specifically includes the following steps: S41: Define feature dimension and sample size: Let the constructed feature vector set , where each feature vector Include Each feature dimension The number of samples; S42: Calculate the maximum and minimum values ​​for each feature dimension: For each feature dimension Calculate its maximum value among all samples. and minimum value ; S43: Perform a linear transformation on each feature dimension: for each feature vector Each feature dimension The normalized set of eigenvectors is obtained by performing a linear transformation according to the following formula. : ;in, Indicates the first The th eigenvector of the th feature vector The original values ​​of each feature dimension. Indicates the first The th eigenvector of the th feature vector The normalized values ​​of each feature dimension.

7. The method for identifying multiple defects in a shield tunnel main bearing according to claim 1, characterized in that, In S5, the pattern recognition algorithm uses a deep learning-based convolutional neural network, whose structure includes an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer.

8. The method for identifying multiple defects in a shield tunnel main bearing according to claim 7, characterized in that, S5 specifically includes the following steps: S51: Training process: Define the loss function as the cross-entropy loss function: ； in, These are the true labels of the samples. The sample predicted by the network belongs to the first... The probability of a class; Update network weights using stochastic gradient descent or Adam optimization algorithm, with a learning rate of [missing value]. The number of iterations is epoch; S52: Classification process: Input the normalized feature vector into the trained CNN model, output the probability distribution of each sample belonging to different defect categories, and take the category corresponding to the maximum probability as the classification result.

9. The method for identifying multiple defects in a shield tunnel main bearing according to claim 1, characterized in that, S6 specifically includes the following steps: S61: Define the confidence threshold and feature threshold: Set the confidence threshold for the pattern recognition result as follows. For each type of typical defect in the main bearing, corresponding characteristic thresholds are set based on historical data and expert experience. S62: Confidence Determination of Pattern Recognition Results: For each sample output by the pattern recognition algorithm, the probability distribution of belonging to various types of defects is used as the confidence level of the pattern recognition result. ,Right now: ,in Indicates that the sample belongs to the first The probability of class defects, Total number of defect categories; S63: Threshold Analysis Logic: If confidence level If the confidence level is low, the defect category corresponding to the pattern recognition result is directly adopted as the final judgment result; if the confidence level is low, the defect category corresponding to the pattern recognition result is adopted as the final judgment result. If so, the threshold analysis auxiliary judgment process will be triggered.

10. The method for identifying multiple defects in a shield tunnel main bearing according to claim 9, characterized in that, In S63, the trigger threshold analysis-assisted determination process specifically includes: S631: For each type of typical defect, check whether the corresponding feature value of the current sample exceeds the feature threshold of that type of defect; S632: If the feature value of the sample meets the feature threshold condition of a certain type of defect, it is determined to be that type of defect; if the feature threshold conditions of multiple defect categories are met, a comprehensive judgment is made by combining expert experience or further diagnostic rules; if the feature threshold conditions of any type of defect are not met, it is determined to be an abnormal state, indicating that further inspection and analysis of the main bearing's operating status is required.

Images