A semi-supervised anomaly detection method for tunnel voids and loose defects

Abstract

The present invention belongs to the technical field of tunnel defect detection and relates to a semi-supervised anomaly detection method for tunnel voids and loose defects. The method integrates GPS technology with ground penetrating radar equipment, collects and segments radar data blocks carrying corresponding position area information of a target tunnel lining structure, extracts signal features of each radar data block through dual channels and fuses them by referring to a mutual attention mechanism, significantly enhancing the integrity, comprehensiveness and accuracy of data extraction while helping to identify and emphasize the most critical feature parts for void and loose defect detection. Then, data is reconstructed on the fused signal features of each radar data block to analyze the data reconstruction error, and each radar data block with void and loose defects is identified with high precision and its specific position is located, which not only achieves a significant improvement in the accuracy and efficiency of tunnel lining structure defect detection, but also provides tunnel maintenance personnel with easy-to-understand defect indications.

Description

Technical Field

[0001] The invention belongs to the technical field of tunnel defect detection and relates to a semi-supervised anomaly detection method for tunnel voids and loose defects. Background Art

[0002] During tunnel construction, due to complex geological conditions, many interference factors in the on-site construction environment, and the influence of construction technology and construction equipment, the lining structure may have loose holes, which will not only reduce the bearing capacity and stability of the tunnel, but also accelerate the aging of the structure, increase the risk of water leakage, and even cause safety accidents such as lining collapse in severe cases, posing a serious threat to driving safety. Therefore, the development of efficient and accurate detection technology for loose holes in tunnel lining structures has become an urgent need in the field of tunnel engineering.

[0003] As a key means of tunnel engineering quality inspection, the geological radar detection method realizes effective detection of underground media through the reflection and refraction mechanism of high-frequency electromagnetic waves in the medium. In view of the significant electrical difference between the air in the cavity and the surrounding concrete, the cavity in the tunnel exhibits the following significant characteristics in the radar wave phase characteristics: the lining interface reflection signal is strong, the three-phase characteristics are obvious, and there is still a strong reflection interface signal below it, and the time difference between the two sets of signals is large. In the process of interpreting and visualizing the data of the cavity, the difference between the cavity structure and the normal structure can be easily distinguished based on the above significant characteristics, which is suitable for non-destructive detection of quality defects of tunnel lining structure and helps to effectively detect the cavity target of tunnel lining structure.

[0004] Existing tunnel void defect detection methods gradually apply artificial intelligence technology to ground penetrating radar detection, but the application of existing technologies is still relatively rudimentary. On the one hand, most methods simply apply target detection and image segmentation technology directly to radar data processing, but fail to be specially designed for the unique properties of radar data. Radar data not only contains rich spatial domain information, but also contains important time domain information. When converting radar data into image data, it is often difficult to fully retain it, resulting in the loss of valuable information, which in turn reduces the accuracy of subsequent recognition and segmentation.

[0005] On the other hand, the situation of hollow and loose areas is relatively rare in tunnel detection, so the relevant data sets are relatively scarce. For object detection and image segmentation algorithms that rely on a large amount of data training, it is difficult to learn the typical features of hollow and loose areas due to the lack of sufficient training samples, which makes it difficult to accurately identify and segment abnormal areas in practical applications. Summary of the invention

[0006] In view of this, in order to solve the problems raised in the above background technology, a semi-supervised anomaly detection method for tunnel voids and loose defects is proposed.

[0007] The purpose of the present invention can be achieved through the following technical solutions: The present invention provides a semi-supervised anomaly detection method for tunnel voids and loose defects, including: S1. Radar data cutting: The target tunnel lining structure is detected by a ground penetrating radar device compatible with GPS, and its corresponding B-Scan radar data is collected. The B-Scan radar data is cut into radar data blocks according to preset specification requirements and continuous geographical location requirements, and each radar data block carries its corresponding location area information.

[0008] S2. Signal feature fusion: Data preprocessing is performed on each radar data block, and the signal features of each radar data block are extracted through dual channels. The dual channels include a time series graph attention channel and a spatiotemporal convolutional network channel. The mutual attention mechanism is introduced to fuse the signal features extracted from the dual channels of each radar data block to obtain the fused signal features of each radar data block.

[0009] S3. Identification of void and loose defects: Input the fused signal features of each radar data block into the reconstruction network for data reconstruction, calculate the data reconstruction error of each radar data block, identify the radar data blocks with void and loose defects and mark them as candidate data blocks.

[0010] S4. Positioning of void and loose defect: Each candidate data block is input into the segmentation network, and the segmentation network outputs the position of the void and loose defect in the corresponding position area of ​​each candidate data block, and the position of each void and loose defect in the target tunnel lining structure is obtained.

[0011] S5. Feedback on void and loose defects: Visually display and feedback on the locations of void and loose defects in the target tunnel lining structure.

[0012] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention integrates GPS technology with ground penetrating radar equipment to ensure that the collected radar data of the target tunnel lining structure are automatically accompanied by precise geographic location information, providing accurate geographic reference for subsequent analysis, thereby improving the practicality of the data and the accuracy of analysis.

[0013] (2) The present invention extracts radar signal features through dual channels. One channel combines a one-dimensional convolutional network and a graph attention mechanism to capture the spatiotemporal features of each radar data block, and the other channel uses a temporal convolutional network to directly extract the spatiotemporal features of the B-scan data. This increases the adaptability of the semi-supervised anomaly detection method for tunnel voids and loose defects to different feature types and changes, making it more flexible and robust when facing the diversity of tunnel structures and complex defect patterns, reducing single limitations.

[0014] (3) The present invention introduces a mutual attention mechanism to fuse the signal features extracted from the dual channels of each radar data block. The promotion of dual-channel signal interaction makes it possible to automatically identify and emphasize the most critical feature parts for the detection of void and loose defects, while reducing the interference of irrelevant or noise information, effectively highlighting the key details in the radar data block, improving the pertinence and accuracy of detection, and further enhancing the versatility and robustness of the detection method.

[0015] (4) The present invention reconstructs the fused signal features of each radar data block and analyzes the data reconstruction error, so as to accurately identify each radar data block with void and loose defects and locate their specific positions. This not only significantly improves the accuracy and efficiency of tunnel lining structure defect detection, but also enhances the intelligence and interpretability of the detection process, providing strong technical support for tunnel safety monitoring and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0017] Figure 1 The present invention is a flowchart of the steps for implementing the method.

[0018] Figure 2 This is a flow chart of the tunnel void loose semi-supervised anomaly detection algorithm of the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] See also Figure 1 and Figure 2 As shown, the present invention provides a semi-supervised anomaly detection method for tunnel voids and loose defects, including: S1. Radar data cutting: The target tunnel lining structure is detected by a ground penetrating radar device compatible with GPS, and its corresponding B-Scan radar data is collected, and the B-Scan radar data is cut into radar data blocks according to preset specification requirements and continuous geographical location requirements, and each radar data block carries its corresponding location area information.

[0021] The embodiment of the present invention integrates GPS technology with ground penetrating radar equipment to ensure that the collected radar data of the target tunnel lining structure is automatically accompanied by accurate geographic location information, providing accurate geographic reference for subsequent analysis, thereby improving the practicality of the data and the analysis accuracy.

[0022] S2. Signal feature fusion: Data preprocessing is performed on each radar data block, and the signal features of each radar data block are extracted through dual channels. The dual channels include a time series graph attention channel and a spatiotemporal convolutional network channel. The mutual attention mechanism is introduced to fuse the signal features extracted from the dual channels of each radar data block to obtain the fused signal features of each radar data block.

[0023] Specifically, the data preprocessing of each radar data block includes: B-Scan radar data is processed by The channel timing code length is A-Scan composition, according to the preset time series length and preset number of features The B-Scan radar data is cut, and the B-Scan radar data can be expressed as , where each radar data block can be expressed as , is the number of each radar data block, , , that is, the unit radar data block has The time length of the number of features is sequence.

[0024] The maximum and minimum values ​​of the radar data training set stored in the WEB cloud are extracted and subtracted to obtain the reference training value. The data in each radar data block is subtracted from the maximum value of the radar data training set and then further compared with the reference training value to realize the maximum and minimum normalization method of the data in each radar data block.

[0025] The Z-score detection method is used to detect abnormal timestamps in each time series in each radar data block, and the abnormal timestamps are replaced with normal timestamps to complete the data preprocessing of each radar data block.

[0026] It should be noted that the specific process of the above-mentioned Z-score detection method for detecting abnormal timestamps in each time series in each radar data block is as follows: obtaining the data point corresponding to each unit timestamp in each time series in each radar data block, calculating the average value and standard deviation of the data points in each time series in each radar data block, subtracting the value of each data point in each time series in each radar data block from the average value of the time series in which it is located, and taking the ratio of the difference value to the standard deviation of the time series in which it is located as the Z-score value of each data point in each time series in each radar data block, and comparing them with the preset Z-score values ​​respectively, and screening out the timestamps corresponding to each data point in each time series in each radar data block whose Z-score value is greater than the preset Z-score value and marking them as abnormal timestamps.

[0027] Specifically, the signal features of each radar data block are extracted through the timing diagram attention channel, including: randomly converting a radar data block Input a one-dimensional convolutional neural network to extract its time series features. The calculation process is: ,in Represents the output timing characteristics of the one-dimensional convolutional neural network for the radar data block, represents a one-dimensional convolution operation, Respectively represent the preset nonlinear activation function and convolution kernel group, Represents the preset number of convolution kernels, which determines the number of feature channels of the output features. Indicates the symbol for the matrix matching operation. Represents the convolution bias, Relu represents the convolution layer activation function, and then the output timing characteristics of the one-dimensional convolutional neural network for each radar data block are obtained.

[0028] The output time series features of each radar data block are further input into the feature dimension graph attention layer to generate a reference schematic diagram of each radar data block. A time subsequence corresponding to each feature is used as each node of the reference schematic diagram. The reference schematic diagram is set to have nodes, and the input vector of each node can be expressed as ,in is the number of each node, , obtain the attention weight of each node in the reference diagram for any node and the total number of neighbor nodes of each node, then the input vector of each node corresponds to the output vector of the attention layer of the feature dimension graph The calculation formula is: ,in and The dimension is the same, used to replace the original vector , as the feature dimension graph attention layer output, represents the sigmoid activation function, Indicates the preset For the first The attention weight of each node, Indicates the preset The total number of neighbor nodes of a node is calculated, and then the output spatial features of each radar data block are obtained.

[0029] The residual structure is introduced to concatenate the output temporal features and output spatial features of each radar data block to obtain the signal features of each radar data block.

[0030] Specifically, the The calculation formula is: ,in represents the concatenation of two vectors, Indicates node input vector, LeakyReLU represents the nonlinear activation function, They represent the preset learnable weight parameters and the preset timestamp number respectively.

[0031] Specifically, the signal features of each radar data block are extracted through the spatiotemporal convolutional network channel, including: the spatiotemporal convolutional network channel is a framework structure composed of causal convolution, hole convolution and residual network, which converts B-Scan radar data into Directly input the spatiotemporal convolutional network channel and extract the temporal characteristics of B-Scan radar data through causal convolution. The analysis formula is: ,in represents the filter bank, Indicates the number of filters in the preset filter group, Indicates the number of radar data blocks of the preset B-Scan radar data.

[0032] Setting the dilation factor via dilated convolution Increase the receptive field of the convolution kernel and extract the spatial features of the B-Scan radar data. The analysis formula is: , Represents the preset one-dimensional convolution kernel length, It means that after a hole convolution calculation, the new sequence is The settlement results on the position, Indicates the past direction. Indicates A filter.

[0033] A residual network is composed of two layers of dilated convolution, ReLU function, and a weight normalization kernel dropout layer. The temporal and spatial features of B-Scan radar data are fused through the residual network to obtain the signal characteristics of B-Scan radar data.

[0034] The signal characteristics of the B-Scan radar data are segmented according to the preset specifications to obtain the signal characteristics of each radar data block.

[0035] The embodiment of the present invention extracts radar signal features through dual channels. One channel combines a one-dimensional convolutional network and a graph attention mechanism to capture the spatiotemporal features of each radar data block, and the other channel uses a temporal convolutional network to directly extract the spatiotemporal features of B-scan data, thereby increasing the adaptability of the semi-supervised anomaly detection method for tunnel voids and loose defects to different feature types and changes, making it more flexible and robust when facing the diversity of tunnel structures and complex defect patterns, and reducing single limitations.

[0036] Specifically, the mutual attention mechanism is introduced to fuse the signal features extracted from the dual channels of each radar data block, including: exchanging the signal features extracted from the dual channels of each radar data block through the mutual attention mechanism feature fusion module for collaborative learning, and recording the input of the mutual attention mechanism feature fusion module corresponding to each radar data block as , through the formula The features of all modes are spliced ​​together to output the signal feature fusion extracted by dual channels of each radar data block ,in Indicates The feature matrix extracted by the one-dimensional convolutional network of radar data blocks, Both indicate The feature matrix extracted by the spatiotemporal convolutional network of radar data blocks, , They represent the learnable matrices of the attention corresponding matrix in the preset spatiotemporal convolutional network feature extraction, Indicates the preset zoom factor, represents the normalized exponential function.

[0037] The embodiment of the present invention introduces a mutual attention mechanism to fuse the signal features extracted by dual channels of each radar data block. The promotion of dual-channel signal interaction makes it possible to automatically identify and emphasize the most critical feature parts for void and loose defect detection, while reducing the interference of irrelevant or noise information, achieving effective highlighting of key details in the radar data block, improving the pertinence and accuracy of detection, and further enhancing the versatility and robustness of the detection method.

[0038] S3. Identification of void and loose defects: Input the fused signal features of each radar data block into the reconstruction network for data reconstruction, calculate the data reconstruction error of each radar data block, identify the radar data blocks with void and loose defects and mark them as candidate data blocks.

[0039] Specifically, the fusion signal features of each radar data block are input into the reconstruction network for data reconstruction, including: the reconstruction network has a gated loop unit, including a reset gate and an update gate, and the fusion signal features of each radar data block are input into the reconstruction network for data reconstruction, including: The gated recurrent unit is input in the form of a sequence, with the time step corresponding to the features in the sequence one by one. The hidden state of the unit time step is updated through the reset gate calculation, update gate calculation and candidate hidden state calculation of the gated recurrent unit.

[0040] The reset gate calculation formula is: ,in Indicates The hidden state of each radar data block corresponds to a time step, They represent the preset learnable weight parameters respectively. When the output of the reset gate is close to 1, it means that the reconstructed network needs to obtain more information from the previous state. When the output of the reset gate is close to 0, it means that the reconstructed network needs to rely more on the current input information.

[0041] The update gate calculation formula is: ,in They represent the preset learnable weight parameters respectively. When the output of the update gate is close to 1, it means that the reconstructed network needs to completely remember the previous state information. When the output of the update gate is close to 0, it means that the reconstructed network completely ignores the previous state information and only depends on the current input information.

[0042] The candidate hidden state calculation formula is: ,in They represent the preset learnable weight parameters respectively, and ⊙ represents element-by-element multiplication.

[0043] The output formula of the gated recurrent unit is obtained as follows: ,in Respectively represent Each radar data block corresponds to a hidden state, an old state, and a new candidate state.

[0044] The hidden states of a preset number of terminal time steps output by the gated recurrent unit generate the reconstructed data of each radar data block.

[0045] Specifically, the calculation of the data reconstruction error of each radar data block includes: using an RNN decoder to reconstruct the output of the gated recurrent unit of the reconstruction network to obtain the original vector and the reconstruction vector , using the loss function The data reconstruction error is obtained, and then the data reconstruction error of each radar data block is obtained.

[0046] Specifically, the identification of each radar data block with a void and non-compact defect includes: extracting a preset data reconstruction error threshold corresponding to the void and non-compact defect stored in the WEB cloud, and comparing it with the data reconstruction error of each radar data block; if the data reconstruction error of a radar data block is greater than or equal to the preset data reconstruction error threshold corresponding to the void and non-compact defect, then the radar data block is identified as having a void and non-compact defect.

[0047] S4. Positioning of void and loose defect: Each candidate data block is input into the segmentation network, and the segmentation network outputs the position of the void and loose defect in the corresponding position area of ​​each candidate data block, and the position of each void and loose defect in the target tunnel lining structure is obtained.

[0048] Specifically, the segmentation network outputs the position of the void and loose defect in the corresponding position area of ​​each candidate data block, including: converting each candidate data block into a preset deep learning processing format through preprocessing and data enhancement, and extracting the pixel-level prediction and spatial resolution of the corresponding position area of ​​each candidate data block through the built-in architecture of the segmentation network.

[0049] The segmentation network extracts a training data set with the same spatial resolution as the candidate data block and applies it to pixel-level prediction of the candidate data block, identifying whether each pixel in the corresponding position area of ​​the candidate data block is a hollow and loose pixel, and then outlining the hollow and loose defect area corresponding to the candidate data block, thereby outputting the hollow and loose defect position in the corresponding position area of ​​each candidate data block.

[0050] The embodiment of the present invention reconstructs the fused signal features of each radar data block and analyzes the data reconstruction error, so as to accurately identify each radar data block with void and loose defects and locate their specific positions. This not only significantly improves the accuracy and efficiency of tunnel lining structure defect detection, but also enhances the intelligence and interpretability of the detection process, providing strong technical support for tunnel safety monitoring and maintenance.

[0051] S5. Void and loose defect feedback: Visually display and feedback the locations of each void and loose defect in the target tunnel lining structure.

[0052] The above contents are merely examples and explanations of the concept of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, they shall all fall within the protection scope of the present invention.

Claims

1. A semi-supervised anomaly detection method for tunnel voids and loose defects, characterized in that: The method includes: S1. Radar data cutting: Use GPS-compatible ground-penetrating radar equipment to detect the target tunnel lining structure, collect its corresponding B-Scan radar data, and cut the B-Scan radar data into radar data blocks according to preset specification requirements and continuous geographical location requirements. Each radar data block carries its corresponding location area information; S2. Signal feature fusion: perform data preprocessing on each radar data block, extract the signal features of each radar data block through dual channels, the dual channels include a time series graph attention channel and a spatiotemporal convolutional network channel, introduce a mutual attention mechanism to fuse the signal features extracted by the dual channels of each radar data block, and obtain the fused signal features of each radar data block; S3. Identification of void and loose defects: Input the fusion signal features of each radar data block into the reconstruction network for data reconstruction, calculate the data reconstruction error of each radar data block, identify each radar data block with void and loose defects and mark them as candidate data blocks; S4. Positioning of hollow and loose defects: inputting each candidate data block into the segmentation network, and the segmentation network outputs the hollow and loose defect positions in the corresponding position area of ​​each candidate data block, and sorting out the hollow and loose defect positions of the target tunnel lining structure; S5. Void and loose defect feedback: Visually display and feedback the location of each void and loose defect in the target tunnel lining structure; The method of extracting the signal features of each radar data block through the timing diagram attention channel includes: randomly converting a radar data block Input a one-dimensional convolutional neural network to extract its time series features. The calculation process is: ,in Represents the output timing characteristics of the one-dimensional convolutional neural network for the radar data block, represents a one-dimensional convolution operation, Respectively represent the preset nonlinear activation function and convolution kernel group, Represents the preset number of convolution kernels, which determines the number of feature channels of the output features. Indicates the symbol for the matrix matching operation. Relu represents the convolution bias, and Relu represents the activation function of the convolution layer, thereby obtaining the output timing characteristics of the one-dimensional convolutional neural network for each radar data block; The output time series features of each radar data block are further input into the feature dimension graph attention layer to generate a reference schematic diagram of each radar data block. A time subsequence corresponding to each feature is used as each node of the reference schematic diagram. The reference schematic diagram is set to have nodes, and the input vector of each node can be expressed as ,in is the number of each node, , obtain the attention weight of each node in the reference diagram for any node and the total number of neighbor nodes of each node, then the input vector of each node corresponds to the output vector of the attention layer of the feature dimension graph The calculation formula is: ,in and The dimension is the same, used to replace the original vector , as the feature dimension graph attention layer output, represents the sigmoid activation function, Indicates the preset For the first node The attention weight of each node, Indicates the preset The total number of neighbor nodes of each node is calculated, and then the output spatial features of each radar data block are obtained; The residual structure is introduced to splice the output time series features and output spatial features of each radar data block to obtain the signal features of each radar data block; Said The calculation formula is: ,in represents the concatenation of two vectors, Indicates node input vector, LeakyReLU represents the nonlinear activation function, They represent the preset learnable weight parameters and the preset timestamp number respectively.

2. A tunnel void and loose defect semi-supervised anomaly detection method according to claim 1, characterized in that: The data preprocessing of each radar data block includes: B-Scan radar data is processed by The channel timing code length is A-Scan composition, according to the preset time series length and preset number of features The B-Scan radar data is cut, and the B-Scan radar data can be expressed as , where each radar data block can be expressed as , is the number of each radar data block, , , that is, the unit radar data block has The time length of the number of features is sequence; The maximum and minimum values ​​of the radar data training set stored in the WEB cloud are extracted and subtracted to obtain the reference training value. The data in each radar data block is subtracted from the maximum value of the radar data training set and then further compared with the reference training value to realize the maximum and minimum normalization method processing of the data in each radar data block; The Z-score detection method is used to detect abnormal timestamps in each time series in each radar data block, and the abnormal timestamps are replaced with normal timestamps to complete the data preprocessing of each radar data block.

3. The semi-supervised anomaly detection method for tunnel voids and loose defects according to claim 1 is characterized by: The signal features of each radar data block are extracted through the spatiotemporal convolutional network channel, including: the spatiotemporal convolutional network channel is a framework structure composed of causal convolution, hole convolution and residual network, which converts B-Scan radar data into Directly input the spatiotemporal convolutional network channel and extract the temporal characteristics of B-Scan radar data through causal convolution. The analysis formula is: ,in represents the filter bank, Indicates the number of filters in the preset filter group, Indicates the number of radar data blocks of the preset B-Scan radar data; Setting the dilation factor via dilated convolution Increase the receptive field of the convolution kernel and extract the spatial features of the B-Scan radar data. The analysis formula is: , Represents the preset one-dimensional convolution kernel length, It means that after a hole convolution calculation, the new sequence is The settlement results on the position, Indicates the past direction. Indicates filters; A residual network is composed of two layers of dilated convolution, ReLU function, and a weight normalization kernel dropout layer. The temporal and spatial features of B-Scan radar data are fused through the residual network to obtain the signal features of B-Scan radar data. The signal characteristics of the B-Scan radar data are segmented according to the preset specifications to obtain the signal characteristics of each radar data block.

4. A tunnel void and loose defect semi-supervised anomaly detection method according to claim 3, characterized in that: The mutual attention mechanism is introduced to fuse the signal features extracted from the dual channels of each radar data block, including: exchanging the signal features extracted from the dual channels of each radar data block through the mutual attention mechanism feature fusion module for collaborative learning, and recording the input of the mutual attention mechanism feature fusion module corresponding to each radar data block as , through the formula The features of all modes are spliced ​​together to output the signal feature fusion extracted by dual channels of each radar data block ,in Indicates The feature matrix extracted by the one-dimensional convolutional network of radar data blocks, Both indicate The feature matrix extracted by the spatiotemporal convolutional network of radar data blocks, , They represent the learnable matrices of the attention corresponding matrix in the preset spatiotemporal convolutional network feature extraction, Indicates the preset zoom factor, represents the normalized exponential function.

5. A tunnel void and loose defect semi-supervised anomaly detection method according to claim 4, characterized in that: The fusion signal features of each radar data block are input into the reconstruction network for data reconstruction, including: the reconstruction network has a gated loop unit, including a reset gate and an update gate, and the fusion signal features of each radar data block are input into the reconstruction network for data reconstruction, including: The gated recurrent unit is input in the form of a sequence, with the time step corresponding to the features in the sequence one by one. The hidden state of the unit time step is updated through the reset gate calculation, update gate calculation and candidate hidden state calculation of the gated recurrent unit; The reset gate calculation formula is: ,in Indicates The hidden state of each radar data block corresponds to a time step, They represent preset learnable weight parameters. When the output of the reset gate is close to 1, it means that the reconstructed network needs to obtain more information from the previous state. When the output of the reset gate is close to 0, it means that the reconstructed network needs to rely more on the current input information. The update gate calculation formula is: ,in They represent preset learnable weight parameters respectively. When the output of the update gate is close to 1, it means that the reconstructed network needs to completely remember the previous state information. When the output of the update gate is close to 0, it means that the reconstructed network completely ignores the previous state information and only depends on the current input information. The candidate hidden state calculation formula is: ,in They represent the preset learnable weight parameters respectively, and ⊙ represents element-by-element multiplication; The output formula of the gated recurrent unit is obtained as follows: ,in Respectively represent Each radar data block corresponds to a hidden state, an old state, and a new candidate state; The hidden states of a preset number of terminal time steps output by the gated recurrent unit generate the reconstructed data of each radar data block.

6. A semi-supervised anomaly detection method for tunnel voids and loose defects according to claim 5, characterized in that: The calculation of the data reconstruction error of each radar data block includes: using the RNN decoder to reconstruct the output of the gated recurrent unit of the reconstruction network to obtain the original vector and the reconstruction vector , using the loss function The data reconstruction error is obtained, and then the data reconstruction error of each radar data block is obtained.

7. A semi-supervised anomaly detection method for tunnel voids and loose defects according to claim 6, characterized in that: The identification of each radar data block with a void and non-density defect includes: extracting a preset data reconstruction error threshold corresponding to the void and non-density defect stored in the WEB cloud, and comparing it with the data reconstruction error of each radar data block; if the data reconstruction error of a radar data block is greater than or equal to the preset data reconstruction error threshold corresponding to the void and non-density defect, then the radar data block is identified as having a void and non-density defect.

8. The semi-supervised anomaly detection method for tunnel voids and loose defects according to claim 1 is characterized by: The segmentation network outputs the position of the hollow and loose defect in the corresponding position area of ​​each candidate data block, including: converting each candidate data block into a preset deep learning processing format through preprocessing and data enhancement, and extracting the pixel-level prediction and spatial resolution of the corresponding position area of ​​each candidate data block through the built-in architecture of the segmentation network; The segmentation network extracts a training data set with the same spatial resolution as the candidate data block and applies it to pixel-level prediction of the candidate data block, identifying whether each pixel in the corresponding position area of ​​the candidate data block is a hollow and loose pixel, and then outlining the hollow and loose defect area corresponding to the candidate data block, thereby outputting the hollow and loose defect position in the corresponding position area of ​​each candidate data block.

Images