A PCCP wire breakage monitoring and identification analysis method based on deep learning fused features
By using distributed fiber optic sensors and deep learning-based feature fusion technology, broken wire signals in PCCP pipelines can be identified and analyzed, solving the identification problem in existing technologies and ensuring the safe operation and maintenance of pipelines.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NORTHWEST ENGINEERING CORPORATION LIMITED
- Filing Date
- 2025-08-01
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies are insufficient to effectively identify and analyze broken wire signals in prestressed concrete pipes (PCCPs), leading to potential pipe cracks and bursts, which affect the safety and stability of water conveyance projects.
Distributed optical fiber sensors are used to monitor the acoustic signals of PCCP. A two-dimensional image dataset of broken wire signals is constructed by deep learning feature fusion technology. Multi-scale features are extracted using Inception-v3, ResNet and Inception-Resnet-v2 models to achieve accurate identification and classification of signals.
It enables accurate identification of wire breakage signals during PCCP pipe operation, improves the safety and reliability of pipeline operation, and provides an effective means of monitoring structural performance.
Smart Images
Figure CN120971585B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of monitoring signal recognition for prestressed concrete duct (PCCP), specifically involving a PCCP wire breakage monitoring and identification analysis method based on deep learning fusion features. Background Technology
[0002] PCCPs (Prestressed Concrete Pipeline Cylinders) are widely used in long-distance water conveyance projects due to their pressure-bearing capacity and durability. The prestressed steel wires in PCCPs are the main structure bearing water pressure; once a wire breaks, the increased internal and external pressure leads to tensile failure of the prestressed steel wire. Prestressed steel wire breakage is a major factor causing PCCP failure. If the breakage continues, it may lead to cracks in the pipe body or even pipe bursts. Pipe bursts not only interrupt the operation of water conveyance projects but can also cause property damage and loss of life. Therefore, corresponding measures should be taken to monitor PCCP wire breakage to provide effective information for PCCP safety assessment. However, the reconstruction and analysis of PCCP wire breakage signal characteristics remains a challenge in engineering. Monitoring systems based on distributed optical fiber technology automatically filter data through primary sorting to find noise with similar prestressed steel wire breakage characteristics. While capturing and recording sound events with similar characteristics still require further analysis and identification to definitively determine the wire breakage event, deep learning has significant advantages in feature extraction. Deep learning methods can be used to process the acoustic signals monitored during PCCP operation, which is of great significance for the monitoring and identification of wire breakage signals during PCCP operation. Summary of the Invention
[0003] The purpose of this invention is to provide a PCCP wire breakage monitoring and identification analysis method based on deep learning fusion features. By establishing PCCP wire breakage signal fusion features, the invention achieves accurate identification of wire breakage signals during PCCP pipe operation and conducts model comparison and evaluation. This provides an advanced and reliable wire breakage signal monitoring method for PCCP pipeline operation, ensuring the safe operation of the pipeline.
[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0005] A PCCP (Polycystic Cyclops Fiber) breakage detection and identification analysis method based on deep learning fusion features includes the following steps:
[0006] S1, PCCP prototype test was conducted using distributed fiber optic sensors;
[0007] S2, based on the PCCP prototype experiment, collect various types of acoustic wave signals generated during the experiment and construct a one-dimensional PCCP acoustic wave signal dataset;
[0008] S3. Process the one-dimensional acoustic signal dataset, convert it into a two-dimensional spectrogram, construct a PCCP-standardized two-dimensional image dataset of broken wire signals, divide the dataset, and select a test set and a validation set.
[0009] S4. Input the two-dimensional image dataset of the monitoring signal into the deep learning model to extract signal features and construct PCCP broken wire signal fusion features;
[0010] S5 trains and tests the model to evaluate the accuracy of signal recognition under single-model feature extraction and feature fusion.
[0011] Preferably, the specific steps of S1 are as follows:
[0012] S101, distributed fiber optic sensors are longitudinally arranged in four directions: 0°, 90°, 180° and 270° on the inner wall of the inner concrete core.
[0013] S102, a corrosion wire breakage test was conducted at 0° circumferentially on the pipe body and 2.7m from the bottom of the pipe;
[0014] S103, hydrogen embrittlement and wire breakage tests were conducted at different internal pressures at positions 180° circumferentially and 2-3m longitudinally on the tube body;
[0015] S104, under an internal water pressure of 1.15 MPa, simulated noise interference in the operating environment by striking the outer layer of mortar with a geological hammer. As the area of broken wires increased, the core began to crack. The test ended when a large number of abnormal cracks appeared in the outermost layer of mortar, or when the pipeline leaked and could not stabilize the pressure.
[0016] Preferably, the specific steps of S2 are as follows:
[0017] S201, MS-DAS2000II equipment was used to monitor wire breakage signals, noise signals and core cracking signals during the test;
[0018] S202. The collected acoustic signals are processed and classified into four types: hydrogen embrittlement wire breakage signal, corrosion wire breakage signal, core cracking signal, and environmental noise. A one-dimensional acoustic signal dataset for PCCP is constructed.
[0019] Preferably, the specific steps of S3 are as follows:
[0020] S301, synchronous squeezing wavelet transform is used on the one-dimensional acoustic wave signal data of PCCP to transform the one-dimensional acoustic wave signal into a two-dimensional spectrum, and a two-dimensional image dataset of PCCP standardized wire breakage signal is constructed.
[0021] S302, divide the two-dimensional spectrum of various types of acoustic signals into data, number and sort the data of various types of signals, select 70% of all two-dimensional spectrum of acoustic signals as the test set, and the remaining 30% as the verification set;
[0022] S303. For the two-dimensional spectrum diagrams of hydrogen embrittlement wire breakage signals under different internal pressures, the hydrogen embrittlement wire breakage signal data are numbered and sorted, and 75% of the two-dimensional spectrum diagrams of hydrogen embrittlement wire breakage signals are selected as the test set, and the remaining 25% are selected as the validation set.
[0023] Preferably, the specific steps of S4 are as follows:
[0024] S401, Input the two-dimensional image dataset of the monitoring signal into the deep learning model Inception-v3 to extract signal features;
[0025] S402, input the two-dimensional image dataset of the monitored signal into the deep learning model ResNet to extract signal features;
[0026] S403, input the two-dimensional image dataset of the monitoring signal into the deep learning model Inception-Resnet-v2 to extract signal features;
[0027] S404 constructs PCCP wire breakage signal fusion features by concatenating the signal features extracted by the single model based on multi-scale features of different dimensions obtained from three single models.
[0028] Preferably, the specific steps of S5 are as follows:
[0029] S501, compare the classification results of the single model with the classification results of the fusion model to verify the effectiveness of feature fusion;
[0030] S502 uses accuracy and confusion matrix to evaluate different models.
[0031] Preferably, in step S401, the Inception-v3 model structure takes a two-dimensional image of the acoustic signal as input. The Inception-v3 model structure adopts a multi-branch convolutional structure, which extracts features by connecting convolutional kernels of different sizes and pooling operations in parallel, and merges the outputs of all branches in the channel dimension through a concatenated layer to achieve the fusion of multi-scale information.
[0032] Preferably, in S402, the ResNet model structure adopts a standard residual connection form, each module consists of multiple convolutional layers, and the input is directly added to the output through skip connections to realize cross-layer information transfer; the entire network does not use a multi-branch structure, the modules are sequentially connected, and max pooling is concatenated in the model architecture; the tail of the model contains a fully connected layer for classification, and no Dropout layer is introduced.
[0033] Preferably, in S403, the Inception-Resnet-v2 model structure combines the design of Inception and ResNet. Each module contains multiple convolutional layers of different sizes, average pooling, and max pooling branches. The output is concatenated through a cascaded layer and then residually connected to the input. The middle to the end of the model contains a Dropout layer to prevent overfitting, and the classification result is output with a fully connected layer.
[0034] The beneficial effects of this invention are as follows: Based on a deep learning fusion model, this invention achieves accurate identification of PCCP wire breakage signals during PCCP pipe operation by establishing PCCP wire breakage signal fusion features. Based on the wire breakage signal identification results, the structural performance of PCCP pipe during operation can be effectively grasped, providing new theories and methods for the operation and maintenance of actual pipeline structures. Attached Figure Description
[0035] Figure 1 This is a simplified flowchart of the present invention.
[0036] Figure 2 This is a detailed flowchart of the present invention.
[0037] Figure 3 This is a schematic diagram of the deployment of distributed fiber optic sensors.
[0038] Figure 4 This is a diagram of the corrosion-induced wire breakage test process.
[0039] Figure 5 This is a diagram of the hydrogen embrittlement wire breakage test process.
[0040] Figure 6 This is a diagram showing cracks in the concrete pipe core.
[0041] Figure 7 This is a typical acoustic signal diagram from the PCCP one-dimensional acoustic hydrogen embrittlement wire breakage signal dataset.
[0042] Figure 8 This is a typical acoustic signal diagram from the PCCP one-dimensional acoustic corrosion wire breakage signal dataset.
[0043] Figure 9 This is a typical acoustic signal image from a PCCP one-dimensional acoustic wave core cracking signal dataset.
[0044] Figure 10 This is a typical acoustic signal diagram from the PCCP one-dimensional acoustic environmental noise signal dataset.
[0045] Figure 11 The image shows a two-dimensional image dataset of PCCP-standardized wire breakage signals; where (a) hydrogen embrittlement wire breakage signal, (b) corrosion wire breakage signal, (c) die cracking signal, and (d) environmental noise.
[0046] Figure 12 This is a diagram of a deep learning model architecture and a schematic diagram of signal feature fusion.
[0047] Figure 13 This is a diagram illustrating the training process of the fusion model.
[0048] Figure 14 The image shows the recognition results of different types of signals using the fusion model. Detailed Implementation
[0049] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings.
[0050] As attached Figures 1-2 As shown, this invention proposes a PCCP (Polycystic Cyclops Fiber) breakage monitoring and identification analysis method based on deep learning fusion features, comprising the following steps:
[0051] S1 uses distributed fiber optic sensors to conduct PCCP prototype tests and monitor the acoustic signals generated during PCCP operation.
[0052] S101, distributed fiber optic sensors are longitudinally arranged in four directions: 0°, 90°, 180° and 270° on the inner wall of the inner concrete core.
[0053] Before deploying the sensors, the inner concrete surface of the PCCP was cleaned to ensure a smooth and flat surface. NZS-DSS-C09 fiber optic sensors were selected to monitor fiber breakage signals. Fiber optic cables were longitudinally deployed along the inner wall of the inner concrete core in four directions: 0°, 90°, 180°, and 270°. A schematic diagram of the distributed fiber optic sensor deployment is attached. Figure 3 As shown, the pipe is 5.0m high, with the socket end at the bottom and the spigot end at the top. Longitudinal fiber optic sensors are installed at four positions on the inner concrete pipe wall at 0°, 90°, 180° and 270°. Starting from 0.5m from the bottom of the pipe, each longitudinal fiber is 4.1m long and extends out of the pipe body after turning at 4.6m from the bottom of the pipe, covering the area of broken wires in the pipe.
[0054] S102, a corrosion wire breakage test was conducted at 0° circumferentially on the pipe body and 2.7m from the bottom of the pipe.
[0055] A corrosion wire breakage test was conducted at a circumferential angle of 0° and a distance of 2.7m from the bottom of the pipe. When the water pressure inside the pipe reached 0.8MPa, a 5% NaCl solution was used to electrochemically corrode the prestressed steel wire, with a corrosion current of 10 mA / cm². 2 The corrosion wire breakage test process diagram is attached. Figure 4 As shown.
[0056] S103, hydrogen embrittlement and wire breakage tests were conducted at different internal pressures at positions 180° circumferentially and 2-3m longitudinally on the tube body.
[0057] Hydrogen embrittlement wire breakage tests were conducted at a circumferential angle of 180° and a longitudinal angle of 2-3m along the pipe body under different internal pressures. This was achieved by manually cutting the broken wires under varying internal pressure conditions. During the test, the internal water pressure was increased in increments of 0.1 MPa. After stabilizing the pressure at 0.1 MPa, 0.4 MPa, 0.8 MPa, and 1.15 MPa for 5 minutes, five steel wires were sequentially cut from top to bottom in the breakage area. A diagram of the hydrogen embrittlement wire breakage test process is attached. Figure 5 As shown.
[0058] S104, under an internal water pressure of 1.15 MPa, simulated noise interference in the operating environment by striking the outer layer of mortar with a geological hammer. As the area of broken wires increased, cracks began to appear in the pipe core. The test ended when a large number of abnormal cracks appeared in the outermost layer of mortar, or when the pipe leaked and could not stabilize the pressure. A diagram of the concrete pipe core cracking is attached. Figure 6 As shown.
[0059] S2, based on the PCCP prototype experiment, collects various types of acoustic wave signals generated during the experiment and constructs a PCCP one-dimensional acoustic wave signal dataset.
[0060] S201, during the prototype test, the MS-DAS2000II device (distributed fiber optic acoustic wave sensing system) was used to monitor the wire breakage signal, noise signal and core cracking signal during the test. The sampling frequency was 1kHz and the spatial resolution was 2m.
[0061] S202. The collected acoustic signals are processed and classified into four types: hydrogen embrittlement wire breakage signal, corrosion wire breakage signal, core cracking signal, and environmental noise. A PCCP one-dimensional acoustic signal dataset is constructed, and typical one-dimensional acoustic signal images of each type are attached. Figures 7-10 As shown.
[0062] S3 uses synchronous squeezing wavelet transform to process the one-dimensional acoustic signal dataset, converting it into a two-dimensional spectrogram, and constructs a PCCP-standardized two-dimensional image dataset of wire breakage signals.
[0063] S301 uses synchronous squeezing wavelet transform on PCCP one-dimensional acoustic signal data to convert the one-dimensional acoustic signal into a two-dimensional spectrogram, and constructs a two-dimensional image dataset of PCCP standardized wire breakage signal.
[0064] A typical two-dimensional spectrum diagram of a signal is attached. Figure 11 As shown. The formula for synchronous squeezing wavelet transform is as follows:
[0065]
[0066]
[0067] Among them, W s (m,n) represents the wavelet coefficients, and s(t) represents the time-domain signal. It is the wavelet mother function. It is a wavelet function The conjugate complex number of , m>0, is the scale parameter, and n∈R is the time parameter. and These are the wavelet mother functions. The Fourier transform of the time-domain signal s(t). ω is the frequency, ε is the angular frequency, and ω s (m,n) represents the instantaneous frequency. T s (ω l ,n) represents the synchronous extrusion transformation value, ω l The center frequency.
[0068] S302, divide the two-dimensional spectrum diagrams of various types of acoustic signals into data, number and sort the data of each type of signal, select 70% of all two-dimensional spectrum diagrams of acoustic signals as the test set, and the remaining 30% as the validation set.
[0069] The two-dimensional spectrograms of various types of acoustic signals were divided into data sets, totaling 410 images. The signal data were numbered and sorted as follows: hydrogen embrittlement wire breakage, corrosion wire breakage, core cracking, and environmental noise were numbered SWF-i, CWB-i, CCS-i, and EN-i (i = 1, 2, 3, 4, 5…). 70% of all two-dimensional spectrograms of the acoustic signals were randomly selected as the test set, and the remaining 30% as the validation set.
[0070] S303. For the two-dimensional spectrum diagrams of hydrogen embrittlement wire breakage signals under different internal pressures, the hydrogen embrittlement wire breakage signal data are numbered and sorted, and 75% of the two-dimensional spectrum diagrams of hydrogen embrittlement wire breakage signals are selected as the test set, and the remaining 25% are selected as the validation set.
[0071] For the two-dimensional spectrograms of hydrogen embrittlement wire breakage signals under different internal pressures, the dataset includes 155 images. The hydrogen embrittlement wire breakage signals are numbered and sorted as follows: signals at 0.1 MPa, 0.4 MPa, 0.8 MPa, and 1.15 MPa are numbered S010-i, S040-i, S080-i, and S115-i (i = 1, 2, 3, 4, 5...). Due to the small dataset size, the weight of the test set is increased, and 75% of the spectrograms are randomly selected as the test set, with the remaining 25% used as the validation set.
[0072] S4 involves inputting the two-dimensional image dataset of the monitoring signal into a deep learning model to extract signal features, including single models Inception-v3, ResNet, and Inception-Resnet-v2, to obtain multi-scale features of the broken wire signal in different dimensions. The signal features extracted by the single models are then concatenated to construct a fusion feature set for the PCCP broken wire signal.
[0073] S401, input the two-dimensional image dataset of the monitoring signal into the deep learning model Inception-v3 to extract signal features.
[0074] A schematic diagram of the Inception-v3 model structure is attached. Figure 12 As shown, the Inception-v3 model uses a two-dimensional image of the acoustic signal as input. It employs a multi-branch convolutional structure, combining convolutional kernels of different sizes (e.g., 1×1, 3×3, 5×5) with pooling operations (average pooling, max pooling) in parallel for feature extraction. A concatenated layer merges the outputs of all branches along the channel dimension, achieving multi-scale information fusion. The model consists of multiple stacked Inception modules connected by ordinary convolutional layers. Dropout layers and fully connected layers are added near the output to prevent overfitting and facilitate final classification. The overall structure ensures both network width and depth.
[0075] S402, input the two-dimensional image dataset of the monitored signal into the deep learning model ResNet to extract signal features.
[0076] A schematic diagram of the ResNet structure is attached. Figure 12 As shown, the ResNet model structure employs a standard residual connection approach. Each module in the diagram consists of multiple convolutional layers, and skip connections directly add the input to the output, enabling cross-layer information transfer. The entire network does not use a multi-branch structure; modules are sequentially connected, with max pooling concatenated throughout the model architecture. A fully connected layer is included at the end of the model for classification, and no Dropout layer is introduced. The overall structure has a moderate depth, emphasizing the use of residual structures to improve training stability and feature reuse capabilities.
[0077] S403: Input the two-dimensional image dataset of the monitoring signal into the deep learning model Inception-Resnet-v2 to extract signal features.
[0078] A schematic diagram of the Inception-Resnet-v2 architecture is attached. Figure 12As shown, the Inception-ResNet-v2 model structure combines the designs of Inception and ResNet. Each module in the diagram contains multiple convolutional layers of different sizes, average pooling, and max pooling branches. The outputs are concatenated through cascaded layers and then residually connected to the inputs, achieving a combination of multi-scale information extraction and efficient training. Dropout layers are included from the middle to the end of the model to prevent overfitting, and fully connected layers output the classification results.
[0079] S404 constructs PCCP wire breakage signal fusion features by concatenating the signal features extracted by the single model based on multi-scale features of different dimensions obtained from three single models.
[0080] The feature fusion process is shown in the attached figure. Figure 12 As shown in the feature fusion module, the converted image is input into three classic deep learning models: Inception-v3, ResNet, and Inception-Resnet-v2, which perform multi-scale and multi-level feature extraction respectively. The feature maps output by the three models are fused in the feature concatenation module to improve the classifier's ability to recognize broken wire signals. Finally, the model performance is evaluated using model evaluation metrics, verifying the effectiveness of the feature fusion strategy in improving recognition results.
[0081] S5. The model is trained and tested to evaluate the accuracy of signal recognition under single model feature extraction and feature fusion. The characteristics of feature extraction of each single model are compared and the differences in signal recognition under single model and fused features are analyzed. The effectiveness of signal recognition under feature fusion is proved, and accurate recognition of PCCP wire breakage monitoring signals is achieved.
[0082] S501 compares the classification results of the single model with the classification results of the fusion model to verify the effectiveness of feature fusion.
[0083] The training process diagram of the fusion model is attached. Figure 13 As shown, in the early stages of training, both training accuracy and validation accuracy rapidly increased and remained stable above 90% for a long period, while the cross-entropy loss continuously decreased and tended to converge. The classification and recognition results are attached. Figure 14As shown in the figure, the confusion matrix for the four-class classification task is as follows: the horizontal axis represents the predicted class, and the vertical axis represents the actual class. It includes the following four signal categories: Category 0 – Hydrogen embrittlement wire breakage signal category; Category 1 – Corrosion wire breakage signal category; Category 2 – Die cracking signal category; and Category 3 – Environmental noise. The results show that of the 50 hydrogen embrittlement wire breakage signals, 49 were correctly identified, with only 1 misclassified as environmental noise, demonstrating a high recognition accuracy. All 13 corrosion wire breakage signals were correctly classified, with no misclassifications. Of the 52 die cracking signals, all but 1 were correctly identified. All 10 environmental noise signals were correctly identified. Overall, the model performed excellently in recognizing the four signal categories, with only a very small number of misclassifications in the hydrogen embrittlement wire breakage and die cracking signals.
[0084] S502 uses accuracy and confusion matrix to evaluate different models. Precision, recall, and F1 score, based on the confusion matrix, are also used for model evaluation. Precision refers to the proportion of samples predicted as positive by the classifier that are actually positive; recall refers to the proportion of samples that are actually positive that were predicted as positive by the classifier; the F1 score is the harmonic mean of precision and recall, comprehensively considering the accuracy and completeness of the classifier. Its expression is as follows:
[0085]
[0086] In this system, TP represents a positive sample that is predicted as a positive sample (true positive), FP represents a negative sample that is predicted as a positive sample (false positive), FN represents a positive sample that is predicted as a negative sample (false negative), and TN represents a negative sample that is predicted as a negative sample (true negative). The closer the above evaluation index is to 1, the more accurate the classification prediction.
[0087] Table 1 summarizes the classification results of single and fusion models for various types of acoustic signals. Table 2 summarizes the classification results of single and fusion models for hydrogen embrittlement wire breakage signals under different internal pressures. The model comparison and evaluation results show that the deep learning fusion feature method has the highest accuracy for PCCP monitoring signal identification, and can accurately predict the wire breakage situation during the operation of the PCCP pipe structure, ensuring the safety and reliability of pipeline operation.
[0088] Table 1 Summary of Classification Results of Single and Fusion Models for Various Types of Acoustic Signals
[0089]
[0090] Table 2 Summary of Classification Results of Single and Fusion Models for Hydrogen Embrittlement Filament Breakage Signals under Different Internal Pressures
[0091]
Claims
1. A PCCP broken wire monitoring and identification analysis method based on deep learning fusion features, characterized in that, Includes the following steps: S1, PCCP prototype test was conducted using distributed fiber optic sensors; S2, based on the PCCP prototype experiment, collect various types of acoustic wave signals generated during the experiment and construct a PCCP one-dimensional acoustic wave signal dataset; S3. Process the one-dimensional acoustic signal dataset, convert it into a two-dimensional spectrogram, construct a PCCP-standardized two-dimensional image dataset of broken wire signals, divide the dataset, and select a test set and a validation set. S4. Input the two-dimensional image dataset of the monitoring signal into the deep learning model to extract signal features and construct PCCP broken wire signal fusion features; The specific steps of S4 are as follows: S401, Input the two-dimensional image dataset of the monitoring signal into the deep learning model Inception-v3 to extract signal features; S402, input the two-dimensional image dataset of the monitored signal into the deep learning model ResNet to extract signal features; S403, input the two-dimensional image dataset of the monitoring signal into the deep learning model Inception-Resnet-v2 to extract signal features; S404, based on the multi-scale features of different dimensions obtained from three single models, concatenates the signal features extracted by the single models to construct PCCP wire breakage signal fusion features; S5, train and test the model to evaluate the accuracy of signal recognition under single model feature extraction and feature fusion; The specific steps of S5 are as follows: S501, compare the classification results of the single model with the classification results of the fusion model to verify the effectiveness of feature fusion; S502 uses accuracy and confusion matrix to evaluate different models.
2. The PCCP filament breakage monitoring and identification analysis method based on deep learning fusion features according to claim 1, characterized in that, The specific steps of S1 are as follows: S101, distributed fiber optic sensors are longitudinally arranged in four directions: 0°, 90°, 180° and 270° on the inner wall of the inner concrete core. S102, a corrosion wire breakage test was conducted at 0° circumferentially on the pipe body and 2.7m from the bottom of the pipe; S103, hydrogen embrittlement and wire breakage tests were conducted at different internal pressures at positions 180° circumferentially and 2-3m longitudinally on the tube body; S104, under an internal water pressure of 1.15 MPa, simulated noise interference in the operating environment by striking the outer layer of mortar with a geological hammer. As the area of broken wires increased, the core began to crack. The test ended when a large number of abnormal cracks appeared in the outermost layer of mortar, or when the pipeline leaked and could not stabilize the pressure.
3. The PCCP filament breakage monitoring and identification analysis method based on deep learning fusion features according to claim 1, characterized in that, The specific steps of S2 are as follows: S201, MS-DAS2000II equipment was used to monitor wire breakage signals, noise signals and core cracking signals during the test; S202. The collected acoustic signals are processed and classified into four types: hydrogen embrittlement wire breakage signal, corrosion wire breakage signal, core cracking signal, and environmental noise. A one-dimensional acoustic signal dataset for PCCP is constructed.
4. The PCCP filament breakage monitoring and identification analysis method based on deep learning fusion features according to claim 1, characterized in that, The specific steps of S3 are as follows: S301, synchronous squeezing wavelet transform is used on the one-dimensional acoustic wave signal data of PCCP to transform the one-dimensional acoustic wave signal into a two-dimensional spectrum, and a two-dimensional image dataset of PCCP standardized wire breakage signal is constructed. S302, divide the two-dimensional spectrum of various types of acoustic signals into data, number and sort the data of various types of signals, select 70% of all two-dimensional spectrum of acoustic signals as the test set, and the remaining 30% as the verification set; S303. For the two-dimensional spectrum of hydrogen embrittlement wire breakage signal under different internal pressures, the hydrogen embrittlement wire breakage signal data are numbered and sorted, and 75% of the two-dimensional spectrum of hydrogen embrittlement wire breakage signal is selected as the test set, and the remaining 25% is used as the validation set.
5. The PCCP filament breakage monitoring and identification analysis method based on deep learning fusion features according to claim 1, characterized in that, In S401, the Inception-v3 model structure takes a two-dimensional image of the acoustic signal as input. The Inception-v3 model structure adopts a multi-branch convolutional structure, which extracts features by connecting convolutional kernels of different sizes and pooling operations in parallel, and merges the outputs of all branches in the channel dimension through a concatenated layer to achieve the fusion of multi-scale information.
6. The PCCP filament breakage monitoring and identification analysis method based on deep learning fusion features according to claim 1, characterized in that, In S402, the ResNet model structure adopts a standard residual connection form. Each module consists of multiple convolutional layers, and the input is directly added to the output through skip connections to realize the cross-layer transfer of information. The entire network does not use a multi-branch structure, and the modules are sequentially connected. Max pooling is concatenated in the model architecture. The tail of the model contains a fully connected layer for classification, and no Dropout layer is introduced.
7. The PCCP filament breakage monitoring and identification analysis method based on deep learning fusion features according to claim 1, characterized in that, In S403, the Inception-Resnet-v2 model structure combines the Inception and ResNet designs. Each module contains multiple convolutional layers of different sizes, average pooling, and max pooling branches. The output is concatenated through a cascaded layer and then residually connected to the input. The model includes Dropout layers from the middle to the end to prevent overfitting, and outputs classification results with fully connected layers.