An AI-based composite insulator internal defect ultrasonic detection method

Abstract

The application relates to the technical field of artificial intelligence, and discloses an AI-based composite insulator internal defect ultrasonic detection method, which comprises a multi-modal ultrasonic probe array module, a signal preprocessing module, an AI defect analysis module, a dynamic parameter optimization module, an edge computing module and a visual report module. Full-dimensional signals are collected through the multi-modal ultrasonic probe array, input into a deep space-time convolutional neural network for defect identification after adaptive noise reduction and feature fusion; the detection accuracy is improved through dynamic waveform matching and multi-physical field coupling analysis; and model lightening and real-time processing are realized through transfer learning and edge computing. The system integrates parameter adaptive optimization, three-dimensional visualization and Internet of Things collaborative functions, solves the problems of low efficiency and high false detection rate of traditional detection methods, and improves the intelligent level and engineering applicability of composite insulator defect detection.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an AI-based ultrasonic detection method for internal defects in composite insulators. Background Technology

[0002] Ultrasonic testing is one of the non-destructive testing methods in industry. By analyzing the propagation characteristics of ultrasonic waves in materials, it can effectively identify internal defects and surface microcracks in insulators. When ultrasonic waves enter an object and encounter a defect, some of the sound waves are reflected. The transmitter and receiver can analyze the reflected waves to accurately measure the defect, displaying its location and size, and determining material thickness. The application field of non-destructive testing methods is vast and continuously expanding. Ultrasonic testing can be categorized according to its principles into defect echo method, transmission method, and resonance method. Composite insulators are key components of high-voltage transmission lines, and their quality directly affects the safe and stable operation of the power system. Traditional quality inspection methods mainly rely on manual visual inspection, mechanical strength testing, and partial discharge detection, but these methods suffer from low detection efficiency and insensitivity to internal defects. Ultrasonic testing technology, due to its non-destructive nature, high penetration, and sensitivity to microscopic defects, has gradually become an important means of detecting internal defects in composite insulators. Currently, the ultrasonic testing methods commonly used in industry are based on signal amplitude threshold criteria or time-domain waveform comparison.

[0003] However, existing ultrasonic testing methods have shortcomings:

[0004] 1. Signal processing relies on human experience: Traditional methods require manual setting of filtering thresholds and feature extraction parameters, which is difficult to adapt to signal variations under complex working conditions, resulting in high false negative and false positive rates;

[0005] 2. Limited accuracy of defect identification: Existing algorithms are mostly based on single-mode signals and cannot integrate multi-physical field information, resulting in insufficient sensitivity to detect minute defects;

[0006] 3. Insufficient real-time performance and intelligence: Traditional systems rely on large computing devices, making it difficult to achieve millisecond-level real-time detection in industrial settings, and they lack adaptive parameter optimization capabilities;

[0007] 4. Low data utilization efficiency: Historical defect data and real-time detection data are isolated, and knowledge transfer is not achieved through machine learning, resulting in weak model generalization ability.

[0008] To address the aforementioned issues, this invention provides an AI-based ultrasonic testing method and system for internal defects in composite insulators. This method overcomes the detection blind spots of single signal modes, improves detection accuracy, meets the high-throughput requirements of production lines, and enhances sensitivity to microcracks, internal delamination, and interface peeling. Summary of the Invention

[0009] (a) Technical problems to be solved

[0010] To address the shortcomings of existing technologies, this invention provides an AI-based ultrasonic detection method for internal defects in composite insulators, solving the problems mentioned in the background section.

[0011] (II) Technical Solution

[0012] To achieve the above objectives, the present invention provides the following technical solution: an AI-based ultrasonic detection method for internal defects in composite insulators, the method comprising the following steps:

[0013] S1. Perform omnidirectional scanning of the composite insulator using a multimodal ultrasonic probe array, and simultaneously record the probe position and timestamp;

[0014] S2. Adaptive noise reduction is performed on the original signal, including wavelet packet denoising, frequency domain filtering and channel weighted fusion, to generate a preprocessed signal set;

[0015] S3. Construct a deep spatiotemporal convolutional network model, input the preprocessed signal, extract features through the multi-branch feature extraction module, and output preliminary results of defect type, location and size;

[0016] S4. Align the static reference waveform and the real-time detected waveform based on the dynamic time warping algorithm, calculate the phase offset and similarity, and generate a waveform consistency matrix.

[0017] S5. Input the initial results and waveform matrix into the hybrid decision model, combine the material dielectric constant and stress simulation data, perform multi-physics coupling analysis, and generate a comprehensive defect report.

[0018] S6. Based on the defect report, use a genetic algorithm to optimize the ultrasonic frequency, pulse parameters and incident angle to form a closed-loop feedback command.

[0019] S7. Align the feature distributions of historical samples with real-time data through transfer learning and update the network weights;

[0020] S8. Deploy lightweight models on edge computing terminals to achieve millisecond-level detection of microcracks, delamination, and peeling defects;

[0021] S9. Reconstruct the 3D point cloud of the defect based on the ray tracing algorithm, and generate a dynamic heat map and risk assessment by combining digital twin;

[0022] S10, packaging test results, optimization parameters and visualization data are compiled into a structured report, which is transmitted to the cloud via 5G IoT and triggers sorting instructions.

[0023] Preferably, step S1 specifically includes:

[0024] S11. A 128-channel piezoelectric composite ultrasonic probe with a ring array arrangement is used. The probe center frequency is programmable and the probe spacing is adaptively adjusted according to the curvature radius of the insulator skirt.

[0025] S12. A high-precision linear motor and a rotating platform control the probe array to perform a spiral trajectory scan along the insulator axis, with the scanning speed meeting the requirement of a spatial resolution of less than 0.5 mm.

[0026] S13. Synchronously acquire the amplitude of ultrasonic reflected signals, the attenuation coefficient of transmitted signals, and the velocity of surface guided waves. The signal sampling rate is greater than 200MS / s, the dynamic range is greater than 80dB, and the resolution is improved.

[0027] Preferably, the adaptive noise reduction process in step S2 further includes:

[0028] S21. Perform variational mode decomposition on the original signal and extract the first 6 eigenmode functions as effective signal components.

[0029] S22. An adaptive Kalman filter is used to iteratively predict and correct high-frequency noise.

[0030] S23. Calculate the fusion weight of the multi-channel signals using mutual information entropy and Pearson correlation coefficient, and perform weighted superposition to generate a noise-reduced signal.

[0031] Preferably, the structure of the deep spatiotemporal convolutional neural network model includes:

[0032] S31. The input layer receives multi-channel ultrasonic time-domain waveforms, short-time Fourier transform spectra, and propagation path matrices.

[0033] S32, the parallel-connected 3D convolutional module, temporal convolutional network module and graph convolutional module extract spatial-spectral features, temporal dependencies and topological features respectively;

[0034] S33. The feature fusion layer dynamically allocates the weights of each feature channel through an attention mechanism to generate a unified feature representation.

[0035] S34. The output layer uses the Softmax function to classify defect types and predicts defect geometric parameters through a regression sub-network. The loss function is a combination of Focal Loss and IoU Loss.

[0036] Preferably, the dynamic time warping algorithm in step S4 is as follows:

[0037] S41. Extract instantaneous frequency features from the static reference waveform using Hilbert-Huang transform and construct a multi-scale reference template library;

[0038] S42. The improved constrained dynamic time warping algorithm is used to calculate the local matching cost matrix between the detected waveform and the template waveform. The constraints include the maximum path slope and phase continuity.

[0039] S43. Optimize the global matching path using the simulated annealing algorithm to generate phase compensation amount and normalized similarity score, with the threshold set at 0.85.

[0040] Preferably, the parameter optimization method in step S6 includes:

[0041] S61. Establish a propagation attenuation model for ultrasonic waves in composite media:

[0042]

[0043] in, These are material-related constants. Angle of incidence:

[0044]

[0045]

[0046]

[0047] S62. Using the maximization of signal-to-noise ratio and the optimization of penetration depth as multi-objective functions, the NSGA-II multi-objective optimization algorithm is used to solve for the Pareto optimal parameter combination.

[0048] Preferably, the domain adaptive transfer learning algorithm in step S7 specifically includes:

[0049] S71. Visualize the high-dimensional feature distribution in the source domain historical sample database using the t-SNE algorithm and extract cross-domain invariant features;

[0050] S72. Construct a domain adaptation loss function by measuring the difference in feature distributions between the source and target domains using the Wasserstein distance:

[0051]

[0052] in, WD is the balance coefficient, which solves the problem of non-overlapping distribution of cross-domain data due to noise or offset, and MMD² is the statistical property for finely aligning high-dimensional feature spaces.

[0053] S73. In neural network training, an adversarial learning strategy is adopted to achieve feature distribution alignment through a gradient inversion layer.

[0054] Preferably, the embedded edge computing terminal in step S8 is implemented as follows:

[0055] S81. The TensorFlow Lite framework is used to perform channel pruning and 8-bit integer quantization on the neural network model to compress the model size.

[0056] S82, Deploy a real-time operating system to ensure that the data acquisition, processing and transmission cycle is less than 10ms;

[0057] The S83 integrates a temperature-humidity composite sensor and an adaptive gain control circuit to achieve self-compensation of environmental parameters and improve compensation accuracy.

[0058] Preferably, the 3D point cloud reconstruction algorithm in step S9 includes:

[0059] S91. Calculate the three-dimensional position of the defect point based on the ultrasonic time-of-flight difference and probe spatial coordinates;

[0060] S92. The moving least squares method is used to fit the geometry of the defect surface to eliminate scanning trajectory error;

[0061] S93. Identify the core region of the defect cluster using the DBSCAN density clustering algorithm, and label the critical defect size and safety level.

[0062] Preferably, the following steps are also included:

[0063] S111. Construct a reinforcement learning agent whose state space includes real-time ultrasonic signal-to-noise ratio, defect detection confidence, probe ambient temperature and energy consumption indicators.

[0064] S112. Define the motion space as the ultrasonic emission frequency adjustment amount, pulse width adjustment range, and incident angle fine-tuning step size.

[0065] S113. Design the reward function:

[0066]

[0067] in, The SNR gain and detection accuracy improvement are calculated by comparing the current state with the historical baseline, where the weighting coefficients are used.

[0068] S114. Train the agent using the proximal policy optimization algorithm, interact with the environment 100,000 times until the policy converges, and generate the optimal parameter adjustment policy.

[0069] S115. Online deployment of intelligent agents dynamically adjusts transmission parameters based on real-time detection data, ensuring that the detection signal-to-noise ratio is stable at greater than 40dB, while reducing unit energy consumption.

[0070] (III) Beneficial Effects

[0071] Compared with existing technologies, this invention provides an AI-based ultrasonic detection method for internal defects in composite insulators, which has the following advantages:

[0072] 1. Multimodal data fusion and adaptive noise reduction improve detection reliability

[0073] By jointly acquiring multi-mode signals from reflection, transmission, and surface guided waves, the system comprehensively covers the surface and internal defect characteristics of insulators, overcoming the detection blind spots of single signal modes. The use of wavelet packet threshold denoising and channel weighted fusion technology effectively suppresses environmental noise and equipment interference, improving the signal-to-noise ratio and providing a high-fidelity data foundation for subsequent analysis.

[0074] 2. High-precision defect recognition achieved through deep spatiotemporal convolutional networks.

[0075] By integrating time-domain waveforms, frequency-domain spectra, and spatial propagation paths for multi-branch feature extraction, the sensitivity to microcracks, internal delamination, and interface peeling is enhanced, improving detection accuracy. An attention mechanism is introduced to dynamically allocate feature weights, addressing the problem of insufficient ability of traditional models to capture complex defect features and reducing false detection rate.

[0076] 3. Enhanced robustness through dynamic waveform matching and multiphysics coupling analysis.

[0077] Based on the constrained dynamic time warping algorithm and simulated annealing optimization, high-precision phase alignment between static reference waveforms and dynamic detection waveforms is achieved, reducing similarity calculation errors. By combining material dielectric constant, stress distribution simulation data and actual working condition parameters, a multi-physics coupled decision model is established to solve the misjudgment problem caused by changes in environmental temperature and humidity in traditional methods.

[0078] 4. Lightweight edge computing and transfer learning enhance engineering applicability

[0079] By using model pruning and 8-bit quantization techniques, the number of neural network parameters is compressed to achieve millisecond-level real-time detection on embedded terminals, meeting the high-throughput requirements of production lines. By employing Wasserstein distance and adversarial learning strategies, cross-domain feature alignment between historical sample databases and real-time data is achieved, improving the model's generalization ability in the new production line environment.

[0080] 5. 3D visualization and closed-loop optimization empower intelligent management

[0081] Based on ray tracing algorithms and digital twin technology, a 3D heat map of defects is generated, which intuitively displays the spatial distribution and risk level of defects, and assists in operation and maintenance decisions. Through genetic algorithms and NSGA-II multi-objective optimization, parameters such as ultrasonic frequency and incident angle are dynamically adjusted to improve detection efficiency and reduce energy consumption. Relying on 5G Internet of Things, detection data is synchronized in the cloud and automatically sorted, thus building a closed loop for full life cycle quality management. Attached Figure Description

[0082] Figure 1 This is a schematic diagram of the overall system architecture of the present invention. Detailed Implementation

[0083] 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.

[0084] Please see Figure 1 This is an AI-based ultrasonic detection method for internal defects in composite insulators.

[0085] S1. Perform omnidirectional scanning of the composite insulator using a multimodal ultrasonic probe array, and simultaneously record the probe position and timestamp;

[0086] S2. Adaptive noise reduction is performed on the original signal, including wavelet packet denoising, frequency domain filtering and channel weighted fusion, to generate a preprocessed signal set;

[0087] S3. Construct a deep spatiotemporal convolutional network model, input the preprocessed signal, extract features through the multi-branch feature extraction module, and output preliminary results of defect type, location and size;

[0088] S4. Align the static reference waveform and the real-time detected waveform based on the dynamic time warping algorithm, calculate the phase offset and similarity, and generate a waveform consistency matrix.

[0089] S5. Input the initial results and waveform matrix into the hybrid decision model, combine the material dielectric constant and stress simulation data, perform multi-physics coupling analysis, and generate a comprehensive defect report.

[0090] S6. Based on the defect report, use a genetic algorithm to optimize the ultrasonic frequency, pulse parameters and incident angle to form a closed-loop feedback command.

[0091] S7. Align the feature distributions of historical samples with real-time data through transfer learning and update the network weights;

[0092] S8. Deploy lightweight models on edge computing terminals to achieve millisecond-level detection of microcracks, delamination, and peeling defects;

[0093] S9. Reconstruct the 3D point cloud of the defect based on the ray tracing algorithm, and generate a dynamic heat map and risk assessment by combining digital twin;

[0094] S10, packaging test results, optimization parameters and visualization data are compiled into a structured report, which is transmitted to the cloud via 5G IoT and triggers sorting instructions;

[0095] S11. A 128-channel piezoelectric composite ultrasonic probe with a ring array arrangement is used. The probe center frequency is programmable and adjustable from 2MHz to 25MHz. The probe spacing is adaptively adjusted according to the curvature radius of the insulator skirt.

[0096] S12. A high-precision linear motor and a rotating platform control the probe array to perform a spiral trajectory scan along the insulator axis, with the scanning speed meeting the requirement of a spatial resolution of less than 0.5 mm.

[0097] S13. Synchronously acquire ultrasonic reflected signal amplitude, transmitted signal attenuation coefficient, and surface guided wave group velocity; signal sampling rate greater than 200MS / s; dynamic range greater than 80dB; improve resolution.

[0098] S21. Perform variational mode decomposition on the original signal and extract the first 6 eigenmode functions as effective signal components.

[0099] S22. An adaptive Kalman filter is used to iteratively predict and correct high-frequency noise.

[0100] S23. Calculate the fusion weights of the multi-channel signals using mutual information entropy and Pearson correlation coefficient, and perform weighted superposition to generate a denoised signal;

[0101] S31. The input layer receives multi-channel ultrasonic time-domain waveforms, short-time Fourier transform spectra, and propagation path matrices.

[0102] S32, the parallel-connected 3D convolutional module, temporal convolutional network module and graph convolutional module extract spatial-spectral features, temporal dependencies and topological features respectively;

[0103] S33. The feature fusion layer dynamically allocates the weights of each feature channel through an attention mechanism to generate a unified feature representation.

[0104] S34. The output layer uses the Softmax function to classify defect types and predicts defect geometric parameters through a regression subnetwork. The loss function is a combination of Focal Loss and IoU Loss.

[0105] S41. Extract instantaneous frequency features from the static reference waveform using Hilbert-Huang transform and construct a multi-scale reference template library;

[0106] S42. The improved constrained dynamic time warping algorithm is used to calculate the local matching cost matrix between the detected waveform and the template waveform. The constraints include the maximum path slope and phase continuity.

[0107] S43. Optimize the global matching path using simulated annealing algorithm to generate phase compensation amount and normalized similarity score, with the threshold set at 0.85;

[0108] S61. Establish a propagation attenuation model for ultrasonic waves in composite media:

[0109]

[0110] in, These are material-related constants. Angle of incidence:

[0111]

[0112]

[0113]

[0114] S62. Using the maximization of signal-to-noise ratio and optimization of penetration depth as multi-objective functions, the NSGA-II multi-objective optimization algorithm is employed to solve for the optimal Pareto parameter combination.

[0115] S71. Visualize the high-dimensional feature distribution in the source domain historical sample database using the t-SNE algorithm and extract cross-domain invariant features;

[0116] S72. Construct a domain adaptation loss function by measuring the difference in feature distributions between the source and target domains using the Wasserstein distance:

[0117]

[0118] in, WD is the balance coefficient, which solves the problem of non-overlapping distribution of cross-domain data due to noise or offset, and MMD² is the statistical property for finely aligning high-dimensional feature spaces.

[0119] S73. Adversarial learning strategies are employed in neural network training, and feature distribution alignment is achieved through gradient inversion layers;

[0120] S81. The TensorFlow Lite framework is used to perform channel pruning and 8-bit integer quantization on the neural network model to compress the model size.

[0121] S82, Deploy a real-time operating system to ensure that the data acquisition, processing and transmission cycle is less than 10ms;

[0122] The S83 integrates a temperature-humidity composite sensor and an adaptive gain control circuit, enabling self-compensation of environmental parameters with a compensation accuracy of ±0.05dB.

[0123] S91. Calculate the three-dimensional position of the defect point based on the ultrasonic time-of-flight difference and probe spatial coordinates;

[0124] S92. The moving least squares method is used to fit the geometry of the defect surface to eliminate scanning trajectory error;

[0125] S93. Identify the core region of defect clusters using the DBSCAN density clustering algorithm, and label the critical defect size and safety level;

[0126] S111. Construct a reinforcement learning agent whose state space includes real-time ultrasonic signal-to-noise ratio, defect detection confidence, probe ambient temperature and energy consumption indicators.

[0127] S112. Define the motion space as the ultrasonic emission frequency adjustment amount, pulse width adjustment range, and incident angle fine-tuning step size.

[0128] S113. Design the reward function:

[0129]

[0130] in, The SNR gain and detection accuracy improvement are calculated by comparing the current state with the historical baseline, where the weighting coefficients are used.

[0131] S114. Train the agent using the proximal policy optimization algorithm, interact with the environment 100,000 times until the policy converges, and generate the optimal parameter adjustment policy.

[0132] S115. Online deployment of intelligent agents dynamically adjusts transmission parameters based on real-time detection data, stabilizing the detection signal-to-noise ratio at greater than 40dB while reducing unit energy consumption;

[0133] A detection system for implementing the method includes:

[0134] Multimodal ultrasound probe array module, high-speed signal acquisition module, AI analysis server, dynamic optimization controller, edge computing unit, and visualization interaction platform:

[0135] The multimodal ultrasound probe array module includes a programmable frequency excitation source and a six-degree-of-freedom robotic arm;

[0136] The high-speed signal acquisition module integrates an FPGA and an ADC chip to achieve multi-channel synchronous sampling;

[0137] The AI ​​analysis server is equipped with an NVIDIA A100 GPU and runs a pre-trained deep spatiotemporal convolutional neural network.

[0138] The dynamic optimization controller adjusts the ultrasonic emission parameters in real time using a fuzzy PID algorithm.

[0139] The edge computing unit uses NVIDIA Jetson AGX Xavier for low-latency inference;

[0140] The visualization and interactive platform supports VR / AR 3D visualization and automated report generation. Specific Implementation

[0142] Example 1: This example details the implementation process of an AI-based ultrasonic testing method for internal defects in composite insulators.

[0143] Step 1: Multimodal ultrasound signal acquisition

[0144] 1.1 Hardware Configuration:

[0145] A 128-channel piezoelectric composite ultrasonic probe with a ring array arrangement is used, and the probe spacing is adaptively adjusted according to the curvature radius of the insulator skirt.

[0146] A six-axis high-precision robotic arm controls the probe array to perform a spiral trajectory scan along the insulator axis, with the scanning speed set at 10 mm / s to ensure a spatial resolution of less than 0.5 mm.

[0147] 1.2. Data Acquisition:

[0148] The amplitude of the reflected signal, the attenuation coefficient of the transmitted signal, and the velocity of the surface guided wave group are acquired simultaneously, and the spatial coordinates and timestamps of the probe are recorded.

[0149] Step 2: Adaptive Signal Denoising and Fusion

[0150] 2.1 Variational Mode Decomposition:

[0151] Variational mode decomposition is performed on the original signal to extract the first 6 intrinsic mode functions and remove high-frequency noise components.

[0152] 2.2 Kalman Filter Optimization:

[0153] An adaptive Kalman filter is used for iterative prediction and correction of the IMF components, with the initial value of the noise covariance matrix set to... , .

[0154] 2.3 Multi-channel fusion:

[0155] Calculate the mutual information entropy and Pearson correlation coefficient of each channel signal to generate the fusion weight matrix:

[0156]

[0157] The weighted superposition is performed to generate a noise-reduced signal.

[0158] Step 3: Construction and Training of Deep Spatiotemporal Convolutional Networks

[0159] 3.1 Model Structure:

[0160] Input layer: Receives 128 channels of time-domain waveforms, short-time Fourier transform spectra, and propagation path matrices.

[0161] Feature extraction layer:

[0162] 3D Convolution Module: Convolution kernel size 5×5×5, number of channels 64, extracts spatial-spectral features;

[0163] Temporal convolutional network: dilation factor 2, number of layers 4, capturing temporal dependencies;

[0164] Graph convolution module: The adjacency matrix is ​​constructed based on the spatial location of the probe to extract topological features.

[0165] Feature fusion layer: Dynamically allocates feature weights through a multi-head attention mechanism to generate a 512-dimensional unified feature vector.

[0166] Output layer: The Softmax classifier outputs the defect type, and the regression subnetwork predicts the defect location and size.

[0167] 3.2 Training Strategy:

[0168] Data set: 100,000 sets of laboratory-calibrated samples and 50,000 sets of field-measured data;

[0169] Optimizer: AdamW, training to the validation set accuracy converged to 98.7%.

[0170] Step 4: Dynamic Waveform Matching and Decision Analysis

[0171] 4.1 Waveform Alignment:

[0172] The static reference waveform is subjected to Hilbert-Huang transform to extract instantaneous frequency features and a multi-scale template library is constructed.

[0173] An improved constrained dynamic time warping algorithm is adopted, which limits the maximum slope of the path to 2, and calculates the local matching cost matrix between the detected waveform and the template.

[0174] The global matching path is optimized by using simulated annealing algorithm to generate phase compensation amount and similarity score.

[0175] 4.2 Hybrid Decision Model:

[0176] Input the initial detection results and similarity matrix into the random forest-support vector machine hybrid model, combine the material dielectric constant and finite element stress simulation data, and output the final defect judgment result.

[0177] Step 5: Edge Computing and Real-Time Detection

[0178] 5.1 Model Lightweighting:

[0179] Using TensorFlow Lite to perform channel pruning and 8-bit integer quantization on the network, the model size was compressed from 350MB to 98MB.

[0180] 5.2 Embedded Deployment:

[0181] The model is deployed on the NVIDIA Jetson AGX Xavier edge terminal, integrating a temperature-humidity sensor and adaptive gain control circuitry.

[0182] It achieves a real-time detection cycle of less than 8ms and supports high-speed pipeline detection that can process more than 120 insulators per second.

[0183] Step 6: 3D Visualization and Closed-Loop Optimization

[0184] 6.1 Defect Refactoring:

[0185] The spatial coordinates of the defect point are calculated based on the ray tracing algorithm, and the moving least squares method is used.

[0186] The core regions of defect clusters are labeled using DBSCAN clustering, mapped to a digital twin model, and a heatmap is rendered.

[0187] 6.2 Parameter Optimization:

[0188] The NSGA-II algorithm is used to optimize parameters such as ultrasonic frequency and pulse width, with the objective function being signal-to-noise ratio and energy consumption, to obtain the Pareto optimal solution set.

[0189] Real-time feedback is sent to the probe control system, reducing energy consumption by more than 25% while maintaining a signal-to-noise ratio of more than 35dB.

[0190] Example effect verification:

[0191] Detection accuracy: Improves the detection accuracy of 2000 test samples and reduces the false negative rate;

[0192] Real-time performance: Single-item inspection takes 7.5ms, which is an improvement over traditional methods;

[0193] Generalization capability: Improves the accuracy of cross-production line migration testing and enhances model adaptability.

[0194] Example 2:

[0195] The system composition of this embodiment corresponds to that of Embodiment 1, including a multimodal ultrasound probe array module, a high-speed signal acquisition module, an AI analysis server, a dynamic optimization controller, and a visualization platform.

[0196] Industrial application scenarios:

[0197] In a certain ultra-high voltage composite insulator production line, this method achieves an average daily inspection volume of 120,000 pieces, reduces the false detection rate, and the defect location error is less than 0.3mm, which is better than the industry standard.

Claims

1. An AI-based ultrasonic testing method for internal defects in composite insulators, characterized in that: The method includes the following steps: S1. Perform omnidirectional scanning of the composite insulator using a multimodal ultrasonic probe array, and simultaneously record the probe position and timestamp; S2. Adaptive noise reduction is performed on the original signal, including wavelet packet denoising, frequency domain filtering and channel weighted fusion, to generate a preprocessed signal set; S3. Construct a deep spatiotemporal convolutional neural network model, input the preprocessed signal, extract features through the multi-branch feature extraction module, and output preliminary results of defect type, location, and size; S4. Align the static reference waveform and the real-time detected waveform based on the dynamic time warping algorithm, calculate the phase offset and similarity, and generate a waveform consistency matrix. S5. Input the initial results and waveform matrix into the hybrid decision model, combine the material dielectric constant and stress simulation data, perform multi-physics coupling analysis, and generate a comprehensive defect report. S6. Based on the defect report, use a genetic algorithm to optimize the ultrasonic frequency, pulse parameters and incident angle to form a closed-loop feedback command. S7. Align the feature distributions of historical samples with real-time data through transfer learning and update the network weights; S8. Deploy lightweight models on edge computing terminals to achieve millisecond-level detection of microcracks, delamination, and peeling defects; S9. Reconstruct the 3D point cloud of the defect based on the ray tracing algorithm, and generate a dynamic heat map and risk assessment by combining digital twin; S10, packaging test results, optimization parameters and visualization data are compiled into a structured report, which is transmitted to the cloud via 5G IoT and triggers sorting instructions.

2. The AI-based ultrasonic testing method for internal defects in composite insulators according to claim 1, characterized in that: Step S1 specifically includes: S11. A 128-channel piezoelectric composite ultrasonic probe with a ring array arrangement is used. The probe center frequency is programmable and adjustable. The probe spacing is adaptively adjusted according to the curvature radius of the insulator skirt. S12. A high-precision linear motor and a rotating platform control the probe array to perform a spiral trajectory scan along the insulator axis, with the scanning speed meeting the requirement of a spatial resolution of less than 0.5 mm. S13. Synchronously acquire the amplitude of ultrasonic reflected signals, the attenuation coefficient of transmitted signals, and the velocity of surface guided waves. The signal sampling rate is greater than 200MS / s, the dynamic range is greater than 80dB, and the resolution is improved.

3. The ultrasonic detection method for internal defects of composite insulators based on AI according to claim 1, characterized in that: The adaptive noise reduction process in step S2 further includes: S21. Perform variational mode decomposition on the original signal and extract the first 6 eigenmode functions as effective signal components. S22. An adaptive Kalman filter is used to iteratively predict and correct high-frequency noise. S23. Calculate the fusion weight of the multi-channel signals using mutual information entropy and Pearson correlation coefficient, and perform weighted superposition to generate a noise-reduced signal.

4. The ultrasonic detection method for internal defects of composite insulators based on AI according to claim 1, characterized in that: The structure of the deep spatiotemporal convolutional neural network model includes: S31. The input layer receives multi-channel ultrasonic time-domain waveforms, short-time Fourier transform spectra, and propagation path matrices. S32, the parallel-connected 3D convolutional module, temporal convolutional network module and graph convolutional module extract spatial-spectral features, temporal dependencies and topological features respectively; S33. The feature fusion layer dynamically allocates the weights of each feature channel through an attention mechanism to generate a unified feature representation. S34. The output layer uses the Softmax function to classify defect types and predicts defect geometric parameters through a regression sub-network. The loss function is a combination of Focal Loss and IoU Loss.

5. The ultrasonic detection method for internal defects of composite insulators based on AI according to claim 1, characterized in that: The dynamic time warping algorithm in step S4 is as follows: S41. Extract instantaneous frequency features from the static reference waveform using Hilbert-Huang transform and construct a multi-scale reference template library; S42. The improved constrained dynamic time warping algorithm is used to calculate the local matching cost matrix between the detected waveform and the template waveform. The constraints include the maximum path slope and phase continuity. S43. Optimize the global matching path using the simulated annealing algorithm to generate phase compensation amount and normalized similarity score, with the threshold set at 0.

85.

6. The ultrasonic detection method for internal defects of composite insulators based on AI according to claim 1, characterized in that: The parameter optimization method in step S6 includes: S61. Establish a propagation attenuation model for ultrasonic waves in composite media: in, These are material-related constants. Angle of incidence: S62. Using the maximization of signal-to-noise ratio and the optimization of penetration depth as multi-objective functions, the NSGA-II multi-objective optimization algorithm is used to solve for the Pareto optimal parameter combination.

7. The ultrasonic testing method for internal defects in composite insulators based on AI according to claim 1, characterized in that: The transfer learning in step S7 is a domain-adaptive transfer learning algorithm, specifically including: S71. Visualize the high-dimensional feature distribution in the source domain historical sample database using the t-SNE algorithm and extract cross-domain invariant features; S72. Construct a domain adaptation loss function by measuring the difference in feature distributions between the source and target domains using the Wasserstein distance: in, WD is the balance coefficient, which solves the problem of non-overlapping distribution of cross-domain data due to noise or offset, and MMD² is the statistical property for finely aligning high-dimensional feature spaces. S73. In neural network training, an adversarial learning strategy is adopted to achieve feature distribution alignment through a gradient inversion layer.

8. The ultrasonic testing method for internal defects in composite insulators based on AI according to claim 1, characterized in that: The embedded edge computing terminal implementation method of step S8 is as follows: S81. The TensorFlow Lite framework is used to perform channel pruning and 8-bit integer quantization on the neural network model to compress the model size. S82, Deploy a real-time operating system to ensure that the data acquisition, processing and transmission cycle is less than 10ms; The S83 integrates a temperature-humidity composite sensor and an adaptive gain control circuit to achieve self-compensation of environmental parameters and improve compensation accuracy.

9. The AI-based ultrasonic testing method for internal defects in composite insulators according to claim 1, characterized in that, It also includes the following steps: Step 11: Real-time optimization of dynamic parameters based on deep reinforcement learning: Step 111: Construct a reinforcement learning agent whose state space includes real-time ultrasonic signal-to-noise ratio, defect detection confidence, probe ambient temperature, and energy consumption indicators. Step 112: Define the motion space as the ultrasonic emission frequency adjustment amount, pulse width adjustment range, and incident angle fine-tuning step size; Step 113: Design the reward function: in, The SNR gain and detection accuracy improvement are calculated by comparing the current state with the historical baseline, where the weighting coefficients are used. Step 114: Train the agent using the proximal policy optimization algorithm, interact with the environment 100,000 times until the policy converges, and generate the optimal parameter adjustment policy. Step 115: Deploy the intelligent agent online and dynamically adjust the transmission parameters based on real-time detection data to stabilize the detection signal-to-noise ratio at greater than 40dB while reducing unit energy consumption.

Images