CN-122017038-A - Pipeline defect acoustic detection method and detection system based on characteristic pattern decomposition and transducer neural network

CN122017038ACN 122017038 ACN122017038 ACN 122017038ACN-122017038-A

Abstract

The invention discloses a pipeline defect acoustic detection method and a detection device thereof based on characteristic mode decomposition and a transducer neural network, wherein the detection method comprises the following steps of S1, sound wave excitation signal emission and echo acquisition; the method comprises the steps of S2, preprocessing echo signals, designing a Butterworth band-pass filter, filtering low-frequency environmental noise and high-frequency electronic noise, removing low-frequency drift components in the signals by adopting a polynomial fitting or high-pass filtering method, normalizing signal amplitude to the range of < -1,1 >, eliminating the influence of different sensor gain differences, calculating effective time length of the signals according to pipeline length and sound velocity, intercepting effective data segments, S3, decomposing characteristic modes based on relevant kurtosis, S4, extracting and identifying deep learning characteristics based on a transducer, and S5, outputting and visualizing detection results.

Inventors

ZHU XUEFENG
Wang Senle
HUANG GUOYONG
Guo Xuanwen
Hei Shengrui

Assignees

昆明理工大学

Dates

Publication Date: 20260512
Application Date: 20260325

Claims (10)

1. A pipeline defect acoustic detection method based on characteristic pattern decomposition and a transducer neural network comprises the following steps: Step S1, transmitting acoustic wave excitation signals and collecting echoes Transmitting an acoustic wave excitation signal in a preset frequency range into a pipeline to be tested through an acoustic wave transmitting module, wherein the acoustic wave excitation signal is a linear frequency modulation signal or a multi-frequency sine superposition signal; The system comprises a synchronous starting sound wave receiving module, a pipeline monitoring module and a pipeline monitoring module, wherein the synchronous starting sound wave receiving module collects reflected echo signals which are transmitted in the pipeline and are subjected to defect modulation, the sound wave receiving module comprises a linear array formed by a plurality of microphone sensors, the sampling frequency is not lower than 4 times of the highest frequency of an excitation signal, and the collection duration is the duration of the excitation signal plus the pipeline round trip transmission time; Step S2, preprocessing echo signals, namely performing the following preprocessing operation on the acquired original echo signals: Firstly, band-pass filtering, namely designing a Butterworth band-pass filter, wherein the passband range is 300Hz to 6kHz, filtering low-frequency environmental noise and high-frequency electronic noise, secondly, trending treatment, namely removing low-frequency drift components in signals by adopting a polynomial fitting or high-pass filtering method, thirdly, normalizing the amplitude of the signals to the range of < -1 >, eliminating the influence of gain difference of different sensors, and fourthly, segmenting data, namely calculating the effective duration of the signals according to the length of a pipeline and the sound velocity, and intercepting the effective data segment; Step S3, characteristic mode decomposition based on the correlation kurtosis Inputting the preprocessed echo signals into a characteristic mode decomposition model to carry out self-adaptive decomposition, wherein the method specifically comprises the following steps of: constructing an adaptive finite impulse response filter bank, wherein the filter bank comprises K FIR filters connected in parallel, the order of each filter is M, the value range of K is 4 to 8, and the value range of M is 64 to 256; Defining the relevant kurtosis CK as an optimization objective function, wherein the calculation formula is as follows: ; Wherein, the For the filter output signal, T is the period delay parameter, The correlation kurtosis measures the impact and periodicity of the signal at the same time to effectively enhance the periodic impact characteristic; Iterative optimization of filter coefficients by a gradient descent or particle swarm optimization algorithm is performed, and the relevant kurtosis value output by each filter is maximized; Obtaining K eigenmode components, wherein each eigenmode component IMF corresponds to defect characteristics of different frequency bands and different physical mechanisms, and enhancing and separating the defect characteristics are realized; S4, deep learning feature extraction and recognition based on transformers The K eigen mode components obtained in the step S3 are used as multichannel input and sent into a pre-trained transducer deep learning model to perform feature extraction and defect recognition, and the method specifically comprises the following steps: The data embedding comprises the steps of segmenting each intrinsic mode component IMF in a sliding window mode, wherein the window length is W, the step length is S, S=W/2, mapping each window into d-dimensional feature vectors through a one-dimensional convolution layer or a full connection layer, and d is the embedding dimension; Position coding, namely adding sine-cosine position coding to the embedded feature vector, and reserving time sequence information, wherein the position coding formula is as follows: ; ; Wherein, the For the position index to be used, Is a dimension index; Multiple head self-attention encoding, namely globally modeling embedded features through N layers of converters encoders, wherein each layer of encoder comprises multiple head self-attention sublayers and feedforward neural network sublayers; the calculation formula of the multi-head self-attention mechanism is as follows: ; ; Wherein, the , In order to pay attention to the number of heads, The self-attention mechanism can adaptively pay attention to the time region and the channel which are most relevant to the defect characteristics, and capture the interaction between different eigenmode components and the long-distance dependence between different time steps; Feature aggregation, namely carrying out global average pooling on all time step features output by an encoder or using a set [ CLS ] token to obtain global feature representation with fixed dimension; double-task output, namely realizing defect classification and positioning through two independent output heads: the classification head is used for outputting probability distribution of C type defect types by a full connection layer+Softmax activation function, wherein C is the defect type number, and an output formula is as follows: ; Wherein, the As a global feature vector of the object, Classifying the header parameters; regression head, full connection layer + linear activation function, output the distance value of defect from pipeline mouth, output formula is: ; Wherein the method comprises the steps of Is a regression header parameter; step S5, outputting and visualizing the detection result And displaying the detection result on a man-machine interaction interface in real time, wherein the detection result comprises defect type, defect position, confidence level, original echo waveform, IMF component waveform and attention thermodynamic diagram.
2. The method for detecting pipeline defects based on characteristic pattern decomposition and a transducer neural network according to claim 1, wherein the method is characterized by comprising the steps of adaptively detecting, when the confidence of the type of the defects output in the step S5 is lower than a preset threshold, judging that the signal quality of the current position is poor or the signal is too far away from the defects, controlling a mobile platform to advance into a pipeline for a preset distance, repeating the steps S1 to S5, and circularly executing the processes until a high confidence detection result is obtained or the end of the pipeline is reached.
3. The method for detecting pipeline defect acoustics based on characteristic pattern decomposition and transform neural network of claim 1, wherein in step S3, the period delay parameter T is adaptively determined according to the pipeline length L and the sound velocity c, and the calculation formula is as follows: ; Wherein the method comprises the steps of For the sampling frequency to be the same, The expression rounds down and the formula ensures that the period delay corresponds to the time that the sound wave travels back and forth in the pipe once, thus effectively extracting the periodic reflection characteristics.
4. The method for acoustic detection of pipeline defects based on eigenmode decomposition and transform neural network of claim 1, wherein the detection results are stored in a database to generate detection reports supporting historical data query and trend analysis.
5. The method for acoustic detection of pipeline defects based on eigenmode decomposition and a transducer neural network according to claim 1, wherein in step S4, the transducer model is pre-trained by: The method comprises the steps of data preparation, namely collecting acoustic echo data of pipelines of different types, defect types and defect positions, and establishing a labeling data set, wherein the data set comprises at least 1000 normal samples and more than 500 samples of each type of defect; Data enhancement, namely expanding training data by adopting methods of time stretching, frequency shifting, gaussian noise addition, time shifting and the like, so as to improve the generalization capability of the model; the loss function is to adopt a weighted combination of classification loss and regression loss, and the formula is as follows: ; Wherein, the In order to cross-entropy categorize the loss, In order to be a mean square error regression loss, Is a weight coefficient; Adopting AdamW optimizers, wherein the initial learning rate is 0.0001, using cosine annealing learning rate scheduling, training 100-200 epochs, and the batch size is 32-64; Regularization-using Dropout and label smoothing to prevent overfitting.
6. A detection system for realizing the acoustic detection method for the pipeline defects based on the characteristic pattern decomposition and the transducer neural network according to any one of claims 1-5, which is characterized by comprising a mobile platform capable of moving in a pipeline autonomously or remotely, an acoustic wave transmitting module, an acoustic wave receiving module, a control and data processing module, a man-machine interaction module and an auxiliary function module comprising a plurality of optional functions; the robot platform comprises a vehicle body, a driving mechanism, a power supply device, a communication device for realizing real-time data transmission with a ground control station and a sensor platform are arranged on the vehicle body, an acoustic wave transmitting module and an acoustic wave receiving module are arranged on the sensor platform, The driving mechanism is a four-wheel independent driving mechanism or a crawler-type driving mechanism, and each driving wheel of the four-wheel independent driving mechanism is provided with an independent motor and an encoder.
7. A detection system for implementing a method for acoustic detection of pipeline defects based on eigenmode decomposition and transform neural network as set forth in claim 6, wherein said vehicle body is provided with an attitude adjustment mechanism for performing attitude adjustment of the sensor platform, the attitude adjustment mechanism comprising pitch adjustment and roll adjustment means for ensuring that the horizontal attitude of the sensor platform is maintained in a non-horizontal pipeline.
8. The detection system for implementing the method for acoustic detection of pipeline defects based on eigenmode decomposition and transform neural network as defined in claim 7, wherein said acoustic wave transmitting module further comprises a power monitoring unit for monitoring the transmitting power and the impedance of the transducer in real time, and automatically adjusting the transmitting power to prevent equipment damage and signal saturation when the transducer is abnormal or the environment is reflected too strongly.
9. The detection system for realizing the acoustic detection method for the pipeline defects based on the characteristic pattern decomposition and the transducer neural network according to claim 8, wherein the acoustic wave receiving module is arranged at the front end of the sensor platform and is used for receiving reflected echo signals in the pipeline, the sensor arrays of the acoustic wave receiving module are linearly arranged at equal intervals, and the interval d is as follows: Where c is the speed of sound, each sensor is connected to a separate low noise preamplifier for analog low pass filtering prior to ADC sampling to prevent aliasing.
10. A detection system for realizing the pipeline defect acoustic detection method based on the characteristic pattern decomposition and the transducer neural network is characterized in that a sensor array of an acoustic wave receiving module adopts differential signal transmission and is provided with a shielding cable to effectively inhibit electromagnetic interference and common mode noise, an auxiliary functional module comprises a video monitoring unit, an environment sensing unit, a GPS (global positioning system) and inertial navigation unit and a data synchronization functional unit, a waterproof camera and an LED illuminating lamp are arranged in the video monitoring unit, videos inside a pipeline are transmitted in real time to assist in confirming defects, the environment sensing unit comprises a temperature sensor, a humidity sensor and a gas sensor to monitor environmental parameters in the pipeline to evaluate safety, and the GPS and inertial navigation unit is used for accurately recording detection positions.

Description

Pipeline defect acoustic detection method and detection system based on characteristic pattern decomposition and transducer neural network Technical Field The invention belongs to the technical field of pipeline nondestructive testing, and particularly relates to a pipeline defect acoustic detection method and system based on characteristic pattern decomposition and deep learning, which are particularly suitable for intelligent defect identification and accurate positioning of complex pipeline systems such as urban drainage pipelines. Background Urban drainage pipelines are used as important underground infrastructures for guaranteeing sustainable development of modern cities, bear the burden of treating sewage and rainwater, and are easy to block due to complex internal environments of the pipelines, so that a high-efficiency low-cost method for detecting the blocking condition of the drainage pipelines is urgently needed at present. The existing pipeline detection technology has the following problems: the method comprises the steps of firstly judging according to manual experience, judging the subjectivity of a detection result by the traditional acoustic detection method, and realizing standardization and automation difficultly, secondly, under a complex noise environment in a pipeline, effectively extracting weak defect characteristics, ensuring high signal flooding ratio and low detection sensitivity, thirdly, solving the problems of mode aliasing, end-point effect and the like of the traditional signal processing method (such as wavelet transformation and empirical mode decomposition), ensuring insufficient characteristic extraction accuracy, fourthly, ensuring low defect positioning accuracy, and being difficult to meet the requirements of accurate maintenance and preventive maintenance, and thirdly, limiting the application of the traditional video detection method in a water-filled pipeline due to illumination and turbidity limitation of the water body and ensuring high equipment cost. Disclosure of Invention The technical problem to be solved by the invention is to provide an automatic all-weather pipeline defect detection technical scheme, so that intelligent identification and accurate positioning of internal defects (including blockage, cracks, corrosion, breakage and the like) of a pipeline are realized, the detection efficiency is improved, and the detection cost is reduced. In order to solve the technical problems, the technical scheme adopted by the invention is that the pipeline defect acoustic detection method based on the characteristic mode decomposition and the transducer neural network comprises the following steps: Transmitting an acoustic wave excitation signal with a preset frequency range into a pipeline to be detected through an acoustic wave transmitting module, wherein the acoustic wave excitation signal is preferably a linear frequency modulation signal (Chirp) or a multi-frequency sine superposition signal, the frequency range is 500Hz to 5kHz, and the excitation duration is 1 to 3 seconds; The system comprises a synchronous starting sound wave receiving module, a pipeline monitoring module and a pipeline monitoring module, wherein the synchronous starting sound wave receiving module collects reflected echo signals which are transmitted in the pipeline and are subjected to defect modulation, the sound wave receiving module comprises a linear array formed by a plurality of microphone sensors, the sampling frequency is not lower than 4 times of the highest frequency of an excitation signal, and the collection duration is the duration of the excitation signal plus the pipeline round trip transmission time; Step S2, preprocessing echo signals, namely performing the following preprocessing operation on the acquired original echo signals: Firstly, band-pass filtering, namely designing a Butterworth band-pass filter, wherein the passband range is 300Hz to 6kHz, filtering low-frequency environmental noise and high-frequency electronic noise, secondly, trending treatment, namely removing low-frequency drift components in signals by adopting a polynomial fitting or high-pass filtering method, thirdly, normalizing the amplitude of the signals to the range of < -1 >, eliminating the influence of gain difference of different sensors, and fourthly, segmenting data, namely calculating the effective duration of the signals according to the length of a pipeline and the sound velocity, and intercepting the effective data segment; Step S3, characteristic mode decomposition based on the correlation kurtosis Inputting the preprocessed echo signals into a characteristic mode decomposition model to carry out self-adaptive decomposition, wherein the method specifically comprises the following steps of: Constructing an adaptive finite impulse response filter bank, wherein the filter bank comprises K FIR filters connected in parallel, the order of each filter is M, the value range of K is 4 to 8, and