Acoustic emission nonmetal pipeline damage detection method based on intelligent sensor

Abstract

The invention discloses an acoustic emission nonmetal pipeline damage detection method based on an intelligent sensor, and particularly relates to the field of nondestructive testing. Intelligent acoustic emission sensor nodes with multiple built-in modules are arranged on the outer wall of a non-metal pipeline, a distributed ad hoc network monitoring system is constructed, and microsecond-level time synchronization is achieved; after each node collects an acoustic emission signal, an effective waveform is extracted through pre-amplification, adaptive filtering and background noise suppression, time domain-frequency domain characteristic parameters are extracted through an edge calculation unit, and primary identification of damage types is completed by using a lightweight convolutional neural network; during multi-node cooperative response, the edge gateway positions a damage source based on a frequency dispersion characteristic correction time difference of arrival method, and rejects an abnormal point in combination with a signal attenuation trend and confidence; and finally, the cloud platform fuses the detection data, performs damage evolution trend analysis and residual life prediction, and triggers multi-level early warning.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, and more specifically, to a method for detecting damage to nonmetallic pipes based on acoustic emission using intelligent sensors. Background Technology

[0002] Non-metallic pipelines are widely used in municipal water supply, gas transmission, chemical media transportation, and marine engineering due to their advantages such as corrosion resistance, light weight, low cost, and long service life. However, non-metallic materials are prone to failure modes such as creep deformation, impact damage, crack propagation, and aging degradation during service. Once a leak or rupture occurs, it will cause serious economic losses and environmental hazards. Therefore, real-time, online damage detection and condition assessment of non-metallic pipelines are of significant engineering importance.

[0003] Acoustic emission (AE) testing technology has become an important technical means for health monitoring of non-metallic pipelines due to its high sensitivity to dynamic damage inside materials, passive monitoring, and strong real-time performance. However, existing AE-based non-metallic pipeline testing technologies still have the following problems in practical applications: Firstly, non-metallic pipe materials (such as polyethylene and fiberglass) severely attenuate the high-frequency components of acoustic emission signals. Traditional piezoelectric sensors have a low signal-to-noise ratio and are easily affected by complex background noise such as pump and valve vibration, fluid turbulence, and electromagnetic interference, resulting in a high rate of missed detections and false alarms for effective events.

[0004] Secondly, existing acoustic emission detection systems mostly adopt a centralized data acquisition and processing architecture. The sensors are connected to the central acquisition host through long cables, resulting in complex system wiring, poor scalability, and a large amount of raw data that needs to be transmitted to the central processor. The data transmission volume is large and the real-time performance is limited, making it difficult to achieve long-distance and large-scale distributed monitoring.

[0005] Third, damage location depends on the propagation speed of acoustic emission signals in the pipeline medium. Traditional methods treat the wave speed as a constant, but non-metallic materials exhibit significant dispersion characteristics—signals of different frequency components have different propagation speeds, resulting in large errors in location methods based on fixed wave speeds and making it difficult to meet engineering requirements in terms of positioning accuracy.

[0006] Therefore, there is an urgent need to develop a non-metallic pipeline damage detection method that can overcome the above-mentioned defects and achieve high-precision positioning, intelligent identification and distributed collaborative monitoring. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for detecting damage to non-metallic pipes based on acoustic emission using intelligent sensors, thereby addressing the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A1: Multiple intelligent acoustic emission sensor nodes are equally spaced along the axial direction on the outer wall of a non-metallic pipe. Each node has a built-in piezoelectric thin film sensor, temperature compensation module, edge computing unit and wireless communication module. The nodes achieve microsecond-level synchronization through a time synchronization protocol to form a distributed self-organizing network monitoring system. A2: Each intelligent sensor node collects acoustic emission signals in real time. After being amplified by a preamplifier, the edge computing unit performs adaptive filtering and background noise suppression to extract the effective acoustic emission waveform. A3: The edge computing unit extracts the time-domain and frequency-domain feature parameters of the acoustic emission signal, constructs a feature classification model based on a lightweight convolutional neural network, and realizes the preliminary identification of damage type at the node end; A4: When multiple adjacent nodes detect a valid acoustic emission event simultaneously, each node uploads the event timestamp, feature vector, and preliminary classification results to the edge gateway. The edge gateway calculates the location of the damage source based on the time difference of arrival method and the modified wave velocity model combined with the dispersion characteristics of the acoustic emission signal in the non-metallic pipe. At the same time, based on the signal attenuation trend and the confidence of the node response, abnormal location points are eliminated, and the final damage location is output. A5: The edge gateway uploads the damage type, location information and intensity level to the cloud management platform. Combined with historical data and pipeline operation parameters, it performs damage evolution trend analysis and remaining life prediction, and triggers multi-level early warnings according to preset thresholds.

[0009] Preferably, in step A1, multiple intelligent acoustic emission sensor nodes are arranged at equal intervals along the axial direction on the outer wall of the non-metallic pipe, with a spacing of 10 meters, to ensure effective overlap of the sensing range of adjacent nodes and avoid monitoring blind spots. Each intelligent sensor node adopts an integrated structure, which includes a piezoelectric thin film sensor, a temperature compensation module, an edge computing unit, and a wireless communication module. The piezoelectric thin film sensor uses a wide-band response polyvinylidene fluoride (PVDF) piezoelectric thin film material. The temperature compensation module is composed of a high-precision digital temperature sensor, which collects the pipe surface temperature in real time for subsequent temperature drift correction of signal propagation speed and sensor sensitivity. The edge computing unit adopts a low-power embedded microprocessor (such as ARM Cortex-M series or RISC-V architecture), with built-in flash memory and static random access memory, for performing signal preprocessing, feature extraction, and lightweight neural network inference tasks. The wireless communication module supports low-power wide area network protocols (such as LoRa, NB-IoT) or short-range wireless protocols, which can be selected according to the site conditions to achieve reliable data transmission between the node and the edge gateway.

[0010] Preferably, in A2, each smart sensor node acquires the acoustic emission signals generated by the non-metallic pipeline in service state in real time through a piezoelectric thin film sensor. The original signal picked up by the sensor is a weak voltage signal with an amplitude typically in the range of microvolts to millivolts. It is first amplified with low noise by a preamplifier integrated inside the node. The amplification gain is dynamically set according to the background noise level of the pipeline, with a typical gain range of 40dB to 60dB, to ensure that the signal amplitude meets the input range requirements of the analog-to-digital converter. The amplified analog signal is converted into a digital signal by the analog-to-digital converter at a preset sampling rate and sent to the edge computing unit for subsequent processing. The edge computing unit performs adaptive filtering and background noise suppression on the original digital signal to extract the effective acoustic emission waveform; an improved wavelet threshold denoising algorithm is used to implement adaptive filtering; specifically, the acquired discrete signal... ( Perform discrete wavelet transform, decomposing the material into 10 layers. , to obtain the detail coefficients of each layer With approximation coefficient ,in, Represented as the decomposition layer number, This serves as the coefficient index; for different decomposition levels, a dynamic threshold function is used for coefficient processing, and the threshold calculation formula is as follows: in, Represented as a threshold, Represented as the first Noise standard deviation estimation of layer detail coefficients, Represented as signal length, It is an adaptive adjustment factor with a value ranging from 0.6 to 1.4, which is dynamically adjusted according to the degree of matching between this frequency component and the spectrum of a typical acoustic emission signal; The robust median absolute bias estimation method is used to calculate: When the frequency of this layer is within the main sensitive range of acoustic emission (50kHz~300kHz), Take the smaller value to retain more effective signal; when in a period dominated by background noise, A larger value is chosen to enhance noise suppression capability.

[0011] Preferably, in step A3, after completing the adaptive filtering and effective waveform extraction of the acoustic emission signal, the edge computing unit performs multi-dimensional feature parameter extraction on the extracted effective acoustic emission waveform to construct feature vectors representing different damage types; the feature extraction is divided into two parallel branches: time-domain feature parameter extraction and frequency-domain feature parameter extraction. The time-domain feature parameters include: amplitude. (Absolute peak value of waveform, in dB), rise time (Time interval from the first crossing of the trigger threshold to the peak of the waveform, in μs), Duration (The time interval from the first time the trigger threshold is exceeded to the last time the threshold is returned to below the threshold, in μs), ringing count (Number of times the waveform crosses the preset threshold level), energy (Area under the waveform envelope, relative unit); The above time-domain parameters reflect the intensity, duration, and burst characteristics of the acoustic emission event; Frequency domain feature parameters are obtained by performing a Fast Fourier Transform on the truncated effective waveform segments. The extracted parameters include: the dominant frequency. (Frequency corresponding to the peak of the power spectrum, in kHz), centroid frequency (in For the first One frequency component, (corresponding power spectral amplitude), peak frequency bandwidth (Frequency width between the half-power points on both sides of the peak frequency, in kHz). The time-domain and frequency-domain feature parameters together constitute the original feature vector. This vector serves as the input to a lightweight convolutional neural network model; For typical damage types of non-metallic pipes, a lightweight convolutional neural network model is constructed and deployed within the edge computing unit to perform preliminary node-level identification of damage types. This lightweight convolutional neural network uses a depthwise separable convolutional structure instead of standard convolution to reduce the number of parameters and computational complexity. The specific network structure is as follows: the input layer receives the original feature vector... First, it is reshaped by dimensional transformation. The tensor form facilitates convolution operations; subsequently, it passes through three depthwise separable convolutional blocks, each containing a depthwise convolutional layer (with a kernel size of 1). (with a stride of 1 and padding mode of SAME) and a pointwise convolutional layer (with a kernel size of 1). Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The depthwise convolutional layer convolves each channel of the input tensor independently, while the pointwise convolutional layer fuses the channels of the depthwise convolution output. Combining these two approaches reduces the computational cost to a fraction of that of standard convolution. ,in Number of output channels The kernel size is [size]. The first depthwise separable convolutional block has 16 output channels, the second has 32, and the third has 64. A global average pooling layer is then applied after the three convolutional blocks to compress the feature map size. Subsequently, the damage is passed through a flattening layer and a fully connected layer, and finally a Softmax classifier outputs the probability distribution of various damage types. ,in The total number of damage categories, satisfying Finally, the category corresponding to the highest probability is taken as the preliminary identification result of the node-end damage type.

[0012] Preferably, in A4, when multiple adjacent intelligent acoustic emission sensor nodes simultaneously detect valid acoustic emission events within a preset time window, the system determines it as a multi-node response event triggered by the same damage source. At this time, each detection node encapsulates its locally stored event timestamp, feature vector, and preliminary classification result of node-end damage type into a data packet, and uploads it to the edge gateway via the wireless communication module. The uploaded data packet contains at least: node number. Event trigger timestamp (Precision in microseconds), feature vector Node-end classification results and node-end classification confidence After receiving data reported by at least three nodes, the edge gateway initiates the damage source location calculation process.

[0013] The location calculation is based on the time difference of arrival (TDOA) method. Its basic principle is that there is a definite geometric relationship between the time difference in the propagation of the acoustic emission signal from the damage source to different sensor nodes and the distance difference between the nodes; let the coordinates of the damage source be... (For straight pipe sections, this can be simplified to a one-dimensional coordinate along the pipe axis.) (For bends or complex pipe sections, two-dimensional coordinates are used to describe the nodes.) The known position coordinates are The signal travels from the damage source to the node. The propagation time is ,in The moment of earthquake origination of the damage source; any two nodes Time difference of arrival between a and 0 The corresponding distance difference ,in It is expressed as the propagation speed of the acoustic emission signal in the pipe medium; for non-metallic pipes, the traditional method will... While the wave velocity is considered constant, the positioning accuracy is severely reduced because non-metallic materials exhibit dispersion characteristics—the acoustic emission signals of different frequency components have different propagation speeds. To address this problem, a dispersion characteristic-corrected wave velocity model is proposed.

[0014] Preferably, in A5, after completing the damage location calculation and confidence verification, the edge gateway encapsulates the final determined damage detection result into a standard data format and uploads it to the cloud management platform via a cellular network or industrial Ethernet; the uploaded data packet includes the following information: damage type. Damage location coordinates Location reliability (Calculated based on the consistency between geometric constraint verification and secondary review, with a value range of 0-1), Strength Grade (Based on the maximum signal amplitude among participating nodes) and event energy Based on comprehensive assessment, the events were categorized into three levels: Level 1 (Minor), Level 2 (Moderate), and Level 3 (Severe). (Event timestamp) (The weighted median of the timestamps of each node) and the list of nodes participating in the localization and their respective response characteristic parameters. After receiving the data, the cloud management platform first stores the current damage event in the pipeline health monitoring database and correlates it with the historical monitoring data of the pipeline to construct an event sequence with time as the axis; Damage evolution trend analysis is based on historical acoustic emission event data throughout the pipeline's entire lifecycle. Time series analysis is used to uncover the evolution patterns of damage at the same location or of the same type. Specifically, the cloud platform divides the pipeline into spatial grids (1-meter axial intervals, with additional circumferential divisions for complex pipe sections), establishes a spatial indexing mechanism, and associates each damage event with its corresponding spatial grid cell. For each grid cell, a damage intensity time series is constructed. Time series of damage events ,in The event occurred at time. A sliding window autoregressive model was used to fit the trend of the sequence. The specific calculation method is as follows: in, For constant terms, These are the autoregressive coefficients. This represents the white noise error term. The optimal order is determined using the Akaike information criterion. The short-term predicted value of damage intensity is calculated; at the same time, the cumulative damage index is introduced. ,in According to the damage intensity level The assigned weights (first-level weight 1, second-level weight 3, third-level weight 9) The time decay factor (values ​​range from 0.01 to 0.05, unit: ...) The cumulative damage index reflects the decay of the impact of a damage event on the pipeline condition over time. When the value exceeds a multiple of the historical baseline, it is determined to be in a stage of accelerated damage evolution, triggering a key monitoring and early warning system.

[0015] The remaining life prediction is comprehensively evaluated by integrating damage evolution trends, pipeline operating parameters, and material aging models. Pipeline operating parameters include operating pressure, temperature, and medium flow rate, which are synchronized in real-time to a cloud platform via a Supervisory Control and Data Acquisition (SCADA) system. For non-metallic pipeline materials, a remaining life prediction model based on the coupling of fracture mechanics and creep damage is established. The crack propagation rate follows the Paris formula. in, The crack depth. The number of load cycles, This represents the stress intensity factor amplitude. Where u is a material constant. Crack propagation events obtained from acoustic emission monitoring are converted into equivalent crack increments. Based on the pipeline operating pressure fluctuation spectrum, the current crack size is calculated. To the critical crack size (Remaining number of cycles required for pipeline burst failure) Converted into remaining service time ,in The average frequency of pressure fluctuations; for fiber-reinforced composite pipes (FRP), a cumulative damage mechanics model is used to establish a correlation between the energy accumulation of acoustic emission events and material stiffness degradation, through the residual stiffness. Estimate the remaining life of the structure when its load-bearing capacity degrades to the design threshold, where The initial elastic modulus, These are the fitting coefficients. To accumulate acoustic emission energy.

[0016] The technical effects and advantages of this invention are as follows: 1. This invention constructs a lightweight convolutional neural network based on depthwise separable convolution at the node end to achieve intelligent identification of damage types. By combining laboratory samples and field samples through transfer learning strategy, the model is trained, which not only ensures the lightweight model is compatible with embedded processors, but also improves the classification accuracy and generalization ability under actual complex working conditions. 2. This invention proposes a dispersion characteristic-corrected wave velocity model. Through laboratory testing, it establishes the mapping relationship between different frequency components of the acoustic emission signal and the propagation speed. In time difference of arrival (TDOA) positioning, it performs weighted correction on the centroid frequency of the received signal at each node, effectively eliminating the wave velocity uncertainty error caused by the dispersion characteristics of non-metallic pipes and significantly improving the damage positioning accuracy. At the same time, it combines signal attenuation trend verification and node response confidence verification to eliminate abnormal positioning points, thereby enhancing the robustness of the positioning results. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0018] Figure 2 This is a schematic diagram of the damage evolution analysis and life prediction process of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 As shown, this invention provides a method for detecting damage to non-metallic pipes based on acoustic emission using intelligent sensors, comprising: A1: Multiple intelligent acoustic emission sensor nodes are equally spaced along the axial direction on the outer wall of a non-metallic pipe. Each node has a built-in piezoelectric thin film sensor, temperature compensation module, edge computing unit and wireless communication module. The nodes achieve microsecond-level synchronization through a time synchronization protocol to form a distributed self-organizing network monitoring system. In A1, multiple intelligent acoustic emission sensor nodes are arranged at equal intervals along the axial direction on the outer wall of the non-metallic pipe, with a spacing of 10 meters to ensure effective overlap of the sensing range of adjacent nodes and avoid monitoring blind spots. Each intelligent sensor node adopts an integrated structure, which includes a piezoelectric thin film sensor, a temperature compensation module, an edge computing unit, and a wireless communication module. The piezoelectric thin film sensor uses wide-band response polyvinylidene fluoride (PVDF) piezoelectric thin film material. The temperature compensation module is composed of a high-precision digital temperature sensor, which collects the pipe surface temperature in real time for subsequent temperature drift correction of signal propagation speed and sensor sensitivity. The edge computing unit adopts a low-power embedded microprocessor (such as ARM Cortex-M series or RISC-V architecture), with built-in flash memory and static random access memory, for performing signal preprocessing, feature extraction, and lightweight neural network inference tasks. The wireless communication module supports low-power wide area network protocols (such as LoRa, NB-IoT) or short-range wireless protocols, which can be selected according to the site conditions to achieve reliable data transmission between nodes and edge gateways.

[0021] To achieve distributed collaborative monitoring, all intelligent sensor nodes achieve microsecond-level time synchronization through a high-precision time synchronization protocol (such as the IEEE 1588 Precision Time Protocol or an improved wireless synchronization protocol). This ensures that each node has a unified time reference when collecting acoustic emission events, providing accurate timestamps for subsequent damage localization based on the time difference of arrival method. The nodes physically form a self-organizing network topology, supporting dynamic joining and leaving without relying on centralized cabling, offering excellent scalability and deployment flexibility. During system operation, each node independently completes the acquisition and local processing of acoustic emission signals. Only when a damage event occurs and meets the conditions for multi-node collaborative judgment is the key feature data and timestamp uploaded to the edge gateway, significantly reducing data transmission volume and system power consumption, thus forming an efficient and reliable distributed self-organizing network monitoring system.

[0022] A2: Each intelligent sensor node collects acoustic emission signals in real time. After being amplified by a preamplifier, the edge computing unit performs adaptive filtering and background noise suppression to extract the effective acoustic emission waveform. In A2, each smart sensor node acquires acoustic emission signals generated by the non-metallic pipeline in service in real time through a piezoelectric thin-film sensor. The raw signal picked up by the sensor is a weak voltage signal with an amplitude typically in the range of microvolts to millivolts. It is first amplified with low noise by a preamplifier integrated inside the node. The amplification gain is dynamically set according to the background noise level of the pipeline, with a typical gain range of 40dB to 60dB, to ensure that the signal amplitude meets the input range requirements of the analog-to-digital converter. The amplified analog signal is then converted into a digital signal by the analog-to-digital converter at a preset sampling rate and sent to the edge computing unit for further processing. The edge computing unit performs adaptive filtering and background noise suppression on the original digital signal to extract the effective acoustic emission waveform; an improved wavelet threshold denoising algorithm is used to implement adaptive filtering; specifically, the acquired discrete signal... ( Perform discrete wavelet transform, decomposition level is . , to obtain the detail coefficients of each layer With approximation coefficient ,in, Represented as the decomposition layer number, This serves as the coefficient index; for different decomposition levels, a dynamic threshold function is used for coefficient processing, and the threshold calculation formula is as follows: in, Represented as a threshold, Represented as the first Noise standard deviation estimation of layer detail coefficients, Represented as signal length, It is an adaptive adjustment factor with a value ranging from 0.6 to 1.4, which is dynamically adjusted according to the degree of matching between this frequency component and the spectrum of a typical acoustic emission signal; The robust median absolute bias estimation method is used to calculate: When the frequency of this layer is within the main sensitive range of acoustic emission (50kHz~300kHz), Take the smaller value to retain more effective signal; when in a period dominated by background noise, To enhance noise suppression capability, a larger value is selected. The threshold processing employs an improved soft threshold function, calculated as follows: in, The attenuation factor has a range of values. to This is used to control the degree of coefficient shrinkage and avoid the constant deviation caused by traditional soft thresholds; The dead zone factor has a range of values. to This is used to introduce a transition dead zone below the threshold to further suppress low-amplitude noise fluctuations; for values ​​between... and The coefficients between them are processed using a linear transition to ensure the continuity of the signal after denoising; The following linear transition function is used for processing: ,in Transition coefficient and ; The processed wavelet coefficients are reconstructed using inverse wavelet transform to obtain the denoised transmitted signal. .

[0023] After wavelet denoising, the edge computing unit further extracts the effective acoustic emission waveform using a combination of short-time energy and adaptive thresholding; the short-time energy calculation method within the sliding window is as follows: in, The starting position of the window. The window length; when the short-time energy of multiple consecutive windows exceeds the adaptive background noise threshold. And the peak amplitude exceeds the preset emission threshold (set as the root mean square value of the background noise). to When the energy decays to below a threshold for a duration less than the preset maximum event duration, it is considered a valid acoustic emission event. The corresponding waveform segment is captured, and the event trigger time is recorded as a timestamp for subsequent damage localization. For event identification in continuous background noise environments, a time delay confirmation mechanism is introduced. This involves continuously monitoring subsequent windows after the initial trigger. If the energy decays to below a threshold for a duration less than the preset maximum event duration, the event is considered valid. to If the duration is too long or the energy fluctuates repeatedly, it is determined to be continuous noise interference and is eliminated.

[0024] A3: The edge computing unit extracts the time-domain and frequency-domain feature parameters of the acoustic emission signal, constructs a feature classification model based on a lightweight convolutional neural network, and realizes the preliminary identification of damage type at the node end; In A3, after completing the adaptive filtering and effective waveform extraction of the acoustic emission signal, the edge computing unit performs multi-dimensional feature parameter extraction on the extracted effective acoustic emission waveform to construct feature vectors representing different damage types. Feature extraction is divided into two parallel branches: time-domain feature parameter extraction and frequency-domain feature parameter extraction. Time-domain feature parameters include: amplitude. (Absolute peak value of waveform, in dB), rise time (Time interval from the first crossing of the trigger threshold to the peak of the waveform, in μs), Duration (The time interval from the first time the trigger threshold is exceeded to the last time the threshold is returned to below the threshold, in μs), ringing count (Number of times the waveform crosses the preset threshold level), energy (Area under the waveform envelope, relative unit); The above time-domain parameters reflect the intensity, duration, and burst characteristics of the acoustic emission event; Frequency domain feature parameters are obtained by performing a Fast Fourier Transform on the truncated effective waveform segments. The extracted parameters include: the dominant frequency. (Frequency corresponding to the peak of the power spectrum, in kHz), centroid frequency (in For the first One frequency component, (corresponding power spectral amplitude), peak frequency bandwidth (Frequency width between the half-power points on both sides of the peak frequency, in kHz). The time-domain and frequency-domain feature parameters together constitute the original feature vector. This vector serves as the input to a lightweight convolutional neural network model; For typical damage types of non-metallic pipes, a lightweight convolutional neural network model is constructed and deployed within the edge computing unit to perform preliminary node-level identification of damage types. This lightweight convolutional neural network uses a depthwise separable convolutional structure instead of standard convolution to reduce the number of parameters and computational complexity. The specific network structure is as follows: the input layer receives the original feature vector... First, it is reshaped by dimensional transformation. The tensor form facilitates convolution operations; subsequently, it passes through three depthwise separable convolutional blocks, each containing a depthwise convolutional layer (with a kernel size of 1). (with a stride of 1 and padding mode of SAME) and a pointwise convolutional layer (with a kernel size of 1). Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The depthwise convolutional layer convolves each channel of the input tensor independently, while the pointwise convolutional layer fuses the channels of the depthwise convolution output. Combining these two approaches reduces the computational cost to a fraction of that of standard convolution. ,in Number of output channels The kernel size is [size]. The first depthwise separable convolutional block has 16 output channels, the second has 32, and the third has 64. A global average pooling layer is then applied after the three convolutional blocks to compress the feature map size. Subsequently, the damage is passed through a flattening layer and a fully connected layer, and finally a Softmax classifier outputs the probability distribution of various damage types. ,in The total number of damage categories, satisfying Finally, the category corresponding to the highest probability is taken as the preliminary identification result of the node-end damage type.

[0025] In the model training phase, a transfer learning strategy is adopted. First, the network is pre-trained using labeled samples collected in a laboratory environment (including various damage simulation signals and background noise samples) to enable the model to learn general acoustic emission feature representations. After pre-training, the model is deployed to field nodes and a small number (typically 100-500 sets) of acoustic emission event samples under actual field conditions are collected. After being labeled by operators, the weights of the last fully connected layer and some high-level convolutional blocks of the model are updated through fine-tuning.

[0026] A4: When multiple adjacent nodes detect a valid acoustic emission event simultaneously, each node uploads the event timestamp, feature vector, and preliminary classification results to the edge gateway. The edge gateway calculates the location of the damage source based on the time difference of arrival method and the modified wave velocity model combined with the dispersion characteristics of the acoustic emission signal in the non-metallic pipe. At the same time, based on the signal attenuation trend and the confidence of the node response, abnormal location points are eliminated, and the final damage location is output. In A4, when multiple adjacent smart acoustic emission sensor nodes simultaneously detect valid acoustic emission events within a preset time window, the system determines it as a multi-node response event triggered by the same damage source. At this time, each detection node encapsulates its locally stored event timestamp, feature vector, and preliminary classification result of node-end damage type into a data packet, and uploads it to the edge gateway via the wireless communication module. The uploaded data packet contains at least: node number. Event trigger timestamp (Precision in microseconds), feature vector Node-end classification results and node-end classification confidence After receiving data reported by at least three nodes, the edge gateway initiates the damage source location calculation process.

[0027] The location calculation is based on the time difference of arrival (TDOA) method. Its basic principle is that there is a definite geometric relationship between the time difference in the propagation of the acoustic emission signal from the damage source to different sensor nodes and the distance difference between the nodes; let the coordinates of the damage source be... (For straight pipe sections, this can be simplified to a one-dimensional coordinate along the pipe axis.) (For bends or complex pipe sections, two-dimensional coordinates are used to describe the nodes.) The known position coordinates are The signal travels from the damage source to the node. The propagation time is ,in The moment of earthquake origination of the damage source; any two nodes Time difference of arrival between a and 0 The corresponding distance difference ,in It is expressed as the propagation speed of the acoustic emission signal in the pipe medium; for non-metallic pipes, the traditional method will... While the wave velocity is considered constant, the non-metallic material exhibits dispersion characteristics—the acoustic emission signals of different frequency components have different propagation speeds, which leads to a serious decrease in positioning accuracy. To address this problem, a dispersion characteristic-corrected wave velocity model is proposed. In the dispersion-corrected wave velocity model, wave velocity curves of acoustic emission signals at different frequency components are pre-established through laboratory testing. The specific testing method involves simulating a quasi-acoustic emission source on a non-metallic pipe specimen using a lead-breaking method, placing broadband sensors at different distances from the source, recording the propagation delay, and using narrowband filtering to decompose different center frequencies. Calculate the propagation speed of each frequency component of the signal. A functional relationship between wave speed and frequency was established through polynomial fitting: in, For low-frequency asymptotic velocity, These are the fitting coefficients. In actual positioning calculations, the edge gateway performs a Fast Fourier Transform on each valid acoustic emission waveform segment reported by each node to calculate its power spectral density. This allows for the determination of the centroid frequency of the signal received by the node. Based on the above wave velocity model, the equivalent propagation velocity corresponding to each node is corrected to... For node pairs The weighted average correction rate is used, and the specific calculation method is as follows: This leads to the corrected distance difference equation: When three or more nodes are involved in localization, an overdetermined system of equations is formed, and the location of the damage source is solved using the least squares method. Specifically, define the objective function: in, The total number of nodes that reported the event ( The Levenberg-Marquardt optimization algorithm is used to iteratively solve for the position parameters that minimize Q. The initial values ​​for iteration are obtained using the weighted method of nodal response amplitude.

[0028] After obtaining the initial location results, the edge gateway further executes an abnormal location point elimination mechanism, performing dual verification based on signal attenuation trend and node response confidence. The signal attenuation trend verification is based on the propagation attenuation law of acoustic emission signals in the pipeline medium: the signal amplitude decreases exponentially with the propagation distance, i.e. ,in, For nodes The detected signal amplitude, The source point amplitude, The attenuation coefficient (calibrated by actual measurement). Nodes inferred from the positioning results Distance from the damage source; calculate the theoretical amplitude for each participating node. Compared with the actual detected amplitude relative error between If the relative error of a certain node If the contribution weight of a node exceeds a preset threshold (0.4), its contribution weight will be reduced or removed in subsequent optimizations; the node response confidence verification is based on the consistency of the classification results at the node end: if more than half of the reporting nodes provide the initial classification results... Inconsistencies (e.g., some nodes are identified as "crack propagation" while others are identified as "leakage impact"), or the classification confidence of individual nodes. If the data is below a threshold (e.g., 0.6), the node's data is affected by strong local noise interference and is marked as a low-confidence node. Combining the above two verification methods, a geometric constraint model is constructed: when the positioning result deviates from the physical range (exceeds the coverage area of ​​adjacent nodes by more than 20%) or contradicts the node response time sequence (i.e., the timestamp of the node farther from the damage source is earlier than that of the closer node, and the difference exceeds the allowable error of propagation delay), the positioning event is automatically marked as a low-confidence event. For low-confidence events, the edge gateway triggers a secondary verification mechanism for neighboring nodes: instructs up to 3 nodes near the damage source to re-collect a signal of a fixed duration at a higher sampling rate, and repeats the above positioning process. The statistical mean and variance of the positioning results are calculated through multiple sampling or the sliding window method, and finally the damage location after confidence-weighted fusion is output, completing the elimination of abnormal positioning points and robust positioning output.

[0029] A5: The edge gateway uploads the damage type, location information and intensity level to the cloud management platform. Combined with historical data and pipeline operation parameters, it performs damage evolution trend analysis and remaining life prediction, and triggers multi-level early warnings according to preset thresholds.

[0030] In A5, after completing the damage location calculation and confidence verification, the edge gateway encapsulates the final damage detection result into a standard data format and uploads it to the cloud management platform via cellular network or industrial Ethernet. The uploaded data packet contains the following information: damage type. Damage location coordinates Location reliability (Calculated based on the consistency between geometric constraint verification and secondary review, with a value range of 0-1), Strength Grade (Based on the maximum signal amplitude among participating nodes) and event energy Based on comprehensive assessment, the events were categorized into three levels: Level 1 (Minor), Level 2 (Moderate), and Level 3 (Severe). (Event timestamp) (The weighted median of the timestamps of each node) and the list of nodes participating in the localization and their respective response characteristic parameters. After receiving the data, the cloud management platform first stores the current damage event in the pipeline health monitoring database and correlates it with the historical monitoring data of the pipeline to construct an event sequence with time as the axis; Damage evolution trend analysis is based on historical acoustic emission event data throughout the pipeline's entire lifecycle. Time series analysis is used to uncover the evolution patterns of damage at the same location or of the same type. Specifically, the cloud platform divides the pipeline into spatial grids (1-meter axial intervals, with additional circumferential divisions for complex pipe sections), establishes a spatial indexing mechanism, and associates each damage event with its corresponding spatial grid cell. For each grid cell, a damage intensity time series is constructed. Time series of damage events ,in The event occurred at time. A sliding window autoregressive model was used to fit the trend of the sequence. The specific calculation method is as follows: in, For constant terms, These are the autoregressive coefficients. This represents the white noise error term. The optimal order is determined using the Akaike information criterion. The short-term predicted value of damage intensity is calculated; at the same time, the cumulative damage index is introduced. ,in According to the damage intensity level The assigned weights (first-level weight 1, second-level weight 3, third-level weight 9) The time decay factor (values ​​range from 0.01 to 0.05, unit: ...) The cumulative damage index reflects the decay of the impact of a damage event on the pipeline condition over time. When the value exceeds a multiple of the historical baseline, it is determined to be in a stage of accelerated damage evolution, triggering a key monitoring and early warning system.

[0031] The remaining life prediction is comprehensively evaluated by integrating damage evolution trends, pipeline operating parameters, and material aging models. Pipeline operating parameters include operating pressure, temperature, and medium flow rate, which are synchronized in real-time to a cloud platform via a Supervisory Control and Data Acquisition (SCADA) system. For non-metallic pipeline materials, a remaining life prediction model based on the coupling of fracture mechanics and creep damage is established. The crack propagation rate follows the Paris formula. in, The crack depth. The number of load cycles, This represents the stress intensity factor amplitude. Where u is a material constant. Crack propagation events obtained from acoustic emission monitoring are converted into equivalent crack increments. Based on the pipeline operating pressure fluctuation spectrum, the current crack size is calculated. To the critical crack size (Remaining number of cycles required for pipeline burst failure) Converted into remaining service time ,in The average frequency of pressure fluctuations; for fiber-reinforced composite pipes (FRP), a cumulative damage mechanics model is used to establish a correlation between the energy accumulation of acoustic emission events and material stiffness degradation, through the residual stiffness. Estimate the remaining life of the structure when its load-bearing capacity degrades to the design threshold, where The initial elastic modulus, These are the fitting coefficients. To accumulate acoustic emission energy.

[0032] The early warning mechanism is divided into a multi-level response system based on preset thresholds, including four levels: blue, yellow, orange, and red. A blue warning corresponds to a single minor injury event (level 1 intensity) and a cumulative injury index lower than twice the baseline value. The triggering condition is the event location reliability. Furthermore, the damage type is non-background noise, and the handling measure is to record the event and include it in the routine inspection plan; a yellow warning corresponds to a single moderate damage event (level 2 intensity), or more than 3 level 1 events occurring cumulatively at the same location within 30 days, or the cumulative damage index exceeding twice the baseline value but less than three times the baseline value. After triggering, the system automatically generates a monitoring report and pushes it to the mobile terminal of the maintenance personnel. It is recommended to increase the inspection frequency or arrange professional testing; an orange warning corresponds to a single severe damage event (level 3 intensity), or the damage evolution trend analysis shows that the short-term predicted value of the damage intensity time series reaches the moderate intensity threshold within 7 days, or the remaining life prediction result is less than 6. Once triggered, the system automatically sends an emergency notification to the operations and maintenance department supervisor, simultaneously activating encrypted monitoring mode for nearby nodes (sampling rate increased to 5MHz, reporting time window shortened to 1 minute), and recommends immediate on-site investigation. A red alert corresponds to two or more consecutive severe damage events with an expanding damage location, or a remaining life prediction of less than one month, or a cumulative damage index exceeding five standard deviations of the baseline value. Upon triggering, the system automatically initiates the emergency response process, simultaneously pushing alert information to the pipeline operation dispatch center, safety management department, and emergency repair team, including precise damage location, estimated failure probability, and recommended handling measures. All alert information is stored on the cloud platform, forming a complete event traceability chain, supporting subsequent accident analysis, operations and maintenance strategy optimization, and pipeline integrity management.

[0033] Because non-metallic pipes (such as PE / PVC materials) have severe intrinsic attenuation characteristics for high-frequency sound waves, and there is a lot of low-frequency mechanical vibration interference in the field environment, the sensitive frequency band of the piezoelectric film sensor in this embodiment is specifically selected as 50kHz-300kHz, and is matched with a preamplifier gain of 40dB-60dB to obtain the best signal-to-noise ratio boundary in this material medium. This method employs depthwise separable convolution in edge computing units, reducing its computational cost to that of standard convolution. Its fundamental purpose is to adapt to the hardware constraints of battery power supply and limited microprocessor computing power (such as the use of ARM Cortex-M series chips) in wireless ad hoc network nodes, so as to meet the real-time requirements of microsecond-level event response.

[0034] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting damage to non-metallic pipes based on acoustic emission using intelligent sensors, characterized in that, include: A1: Multiple intelligent acoustic emission sensor nodes are equally spaced along the axial direction on the outer wall of a non-metallic pipe. Each node has a built-in piezoelectric thin film sensor, temperature compensation module, edge computing unit and wireless communication module. The nodes achieve microsecond-level synchronization through a time synchronization protocol to form a distributed self-organizing network monitoring system. A2: Each intelligent sensor node collects acoustic emission signals in real time. After being amplified by a preamplifier, the edge computing unit performs adaptive filtering and background noise suppression to extract the effective acoustic emission waveform. A3: The edge computing unit extracts the time-domain and frequency-domain feature parameters of the acoustic emission signal, constructs a feature classification model based on a lightweight convolutional neural network, and realizes the preliminary identification of damage type at the node end; A4: When multiple neighboring nodes detect a valid acoustic emission event simultaneously, each node uploads the event timestamp, feature vector, and preliminary classification results to the edge gateway. The edge gateway calculates the location of the damage source based on the time difference of arrival method and the modified wave velocity model by combining the dispersion characteristics of acoustic emission signals in non-metallic pipes. At the same time, based on the signal attenuation trend and node response confidence, abnormal location points are eliminated and the final damage location is output. A5: The edge gateway uploads the damage type, location information and intensity level to the cloud management platform. Combined with historical data and pipeline operation parameters, it performs damage evolution trend analysis and remaining life prediction, and triggers multi-level early warnings according to preset thresholds.

2. The method for detecting damage to non-metallic pipes based on acoustic emission according to claim 1, characterized in that: In the intelligent acoustic emission sensor node, the temperature compensation module consists of a high-precision digital temperature sensor, which is used to collect the surface temperature of the pipe in real time and correct the nonlinear effect of temperature on the propagation speed of acoustic emission signals and the sensitivity of the sensor.

3. The method for detecting damage to non-metallic pipes based on acoustic emission according to claim 1, characterized in that: In A2, the adaptive filtering employs an improved wavelet threshold denoising algorithm, combined with typical pipeline noise spectrum characteristics, to dynamically... Adjust the threshold function; in the improved wavelet thresholding denoising algorithm, the threshold processing adopts an improved soft threshold function.

4. The method for detecting damage to non-metallic pipes based on acoustic emission according to claim 1, characterized in that: In the A3, the lightweight convolutional neural network adopts a depthwise separable convolutional structure and performs real-time inference on the embedded processor at the node. The network structure includes three depthwise separable convolutional blocks. Each convolutional block contains a depthwise convolutional layer and a pointwise convolutional layer, followed by a batch normalization layer and a ReLU activation function. Finally, it outputs the probability distribution of damage type through a global average pooling layer and a Softmax classifier.

5. The method for detecting damage to non-metallic pipes based on acoustic emission according to claim 4, characterized in that: The lightweight convolutional neural network adopts a transfer learning strategy during the model training phase. First, it is pre-trained using labeled samples collected in a laboratory environment. Then, it is fine-tuned using labeled samples under actual working conditions to adapt the model to specific pipe materials and environmental characteristics.

6. The method for detecting damage to non-metallic pipes based on acoustic emission according to claim 1, characterized in that: In A4, the dispersion characteristic modified wave velocity model is established in advance through laboratory testing to establish the mapping relationship between different frequency components of the acoustic emission signal and the propagation speed. In the time difference of arrival calculation, the centroid frequency of the received signal at each node is weighted and corrected, and the distance difference equation is constructed by using a weighted average correction velocity.

7. The method for detecting damage to non-metallic pipes based on acoustic emission according to claim 6, characterized in that: Abnormal location points are eliminated based on signal attenuation trends and node response confidence. This includes: calculating the theoretical amplitude based on the exponential attenuation relationship between signal amplitude and propagation distance, comparing it with the actual detected amplitude to obtain the relative error, and eliminating nodes whose errors exceed a preset threshold; at the same time, based on the consistency of node classification results and classification confidence, low-confidence nodes are marked, and a geometric constraint model is constructed to eliminate abnormal location points.

8. The method for detecting damage to non-metallic pipes based on acoustic emission according to claim 6, characterized in that: The method of combining confidence level to remove outliers specifically includes a two-stage verification mechanism: When the confidence level of the damage source location initially calculated by the edge gateway is lower than the preset threshold, the sensor nodes in the target area are scheduled to re-collect data at a sampling rate higher than the initial set value. The edge gateway uses the sliding window method to calculate the positioning results multiple times and extracts the statistical mean and variance of the multiple positioning coordinates; The mean coordinates with variance less than or equal to the limit value are used as the location of the damage source in the final output.

9. The method for detecting damage to non-metallic pipes based on acoustic emission according to claim 6, characterized in that: When the location result is marked as a low-confidence event, the edge gateway triggers a secondary verification mechanism for neighboring nodes, instructing three nodes near the damage source to re-collect signals at a higher sampling rate. Through multiple rounds of location calculation, the statistical mean and variance are calculated, and the damage location is output after confidence-weighted fusion.

10. The method for detecting damage to non-metallic pipes based on acoustic emission according to claim 1, characterized in that: In A5, the damage evolution trend analysis uses a sliding window autoregressive model to fit the damage intensity time series and introduces a cumulative damage index to determine the damage evolution stage. The remaining service life prediction uses a crack propagation model based on the Paris formula for polyethylene pipes to calculate the remaining number of cycles and the remaining service time; for fiber-reinforced composite pipes, a cumulative damage mechanics model is used to estimate the remaining service life through the remaining stiffness; the multi-level early warning includes four levels: blue, yellow, orange and red. Each level of early warning is determined based on the damage intensity level, cumulative damage index, remaining service life prediction results and event location confidence, and corresponds to different triggering conditions and handling measures.

Images