AI-based composite insulator internal defect ultrasonic detection method

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an AI-based composite insulator internal defect ultrasonic detection method, which comprises a multi-mode ultrasonic probe array module, a signal preprocessing module, an AI defect analysis module, a dynamic parameter optimization module, an edge calculation module and a visual report module, the method comprises the following steps: acquiring a full-dimensional signal through a multi-modal ultrasonic probe array, and inputting the full-dimensional signal into a deep space-time convolutional neural network for defect recognition after adaptive noise reduction and feature fusion; the detection precision is improved by combining dynamic waveform matching and multi-physics coupling analysis; model lightweight and real-time processing are realized by adopting transfer learning and edge calculation. The system integrates the functions of parameter adaptive optimization, three-dimensional visualization and Internet of Things cooperation, solves the problems of low efficiency and high false detection rate of a traditional detection method, and improves the intelligent level and engineering applicability of composite insulator defect detection.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of artificial intelligence, and particularly relates to an AI-based ultrasonic detection method for internal defects of a composite insulator. BACKGROUND

[0002] Ultrasonic detection is one of the non-destructive testing methods in the industry. By analyzing the propagation characteristics of ultrasonic waves in materials, internal defects and surface micro-cracks of insulators can be effectively identified. When ultrasonic waves enter an object and encounter a defect, part of the sound waves will be reflected. The emitter and receiver can analyze the reflected waves, accurately measure the defect, and display the position and size of the internal defect, and determine the thickness of the material. The application field of non-destructive testing methods is a very large field and is constantly expanding. Ultrasonic detection can be divided into defect echo method, transmission method and resonance method according to its principle. The composite insulator is a key component of high-voltage transmission lines, and its quality is directly related to the safe and stable operation of the power system. Traditional quality detection methods mainly rely on manual visual inspection, mechanical strength testing and partial discharge detection, but these methods have problems such as low detection efficiency and insensitivity to internal defects. Ultrasonic detection technology gradually becomes an important means for detecting internal defects of composite insulators due to its non-destructive, high penetration and sensitivity to micro defects. At present, the ultrasonic detection method widely used in the industry is based on signal amplitude threshold criterion or time domain waveform comparison.

[0003] However, the existing ultrasonic detection method has the following problems: 1. Signal processing relies on human experience: the traditional method needs to manually set the filtering threshold and feature extraction parameters, which is difficult to adapt to signal variation under complex working conditions, resulting in high false rejection rate and false detection rate; 2. Limited defect recognition accuracy: existing algorithms are mostly based on single modal signals, which cannot fuse multi-physical field information, and the detection sensitivity to small defects is insufficient; 3. Lack of real-time and intelligence: the traditional system relies on large computing devices, which is difficult to realize millisecond-level real-time detection in the industrial field, and lacks self-adaptive parameter optimization capability; 4. Low data utilization efficiency: historical defect data and real-time detection data are isolated, and knowledge transfer is not realized through machine learning, resulting in weak model generalization ability.

[0004] To solve the above problems, the AI-based ultrasonic detection method for internal defects of a composite insulator and system are provided, which overcome the detection blind area of single signal mode, improve the detection accuracy, meet the high-throughput demand of the production line, and enhance the sensitivity to micro-cracks, internal delamination and interface peeling. SUMMARY

[0005] (I) Technical problems solved In view of the deficiencies of the prior art, the application provides an AI-based composite insulator internal defect ultrasonic detection method, which solves the problems in the background art.

[0006] (II) Technical solutions To achieve the above object, the application provides the following technical solutions: an AI-based composite insulator internal defect ultrasonic detection method, comprising the following steps: S1, omnidirectional scanning of a composite insulator by a multi-modal ultrasonic probe array, and synchronous recording of the probe position and time stamp; S2, adaptive noise reduction of the original signal, including wavelet packet denoising, frequency domain filtering and channel weighted fusion, to generate a preprocessed signal set; S3, construction of a deep spatio-temporal convolution network model, input of the preprocessed signal, feature extraction by a multi-branch feature extraction module, and output of the preliminary results of the defect type, position and size; S4, alignment of the static reference waveform and the real-time detection waveform based on the dynamic time warping algorithm, calculation of the phase shift and similarity, and generation of a waveform consistency matrix; S5, input of the preliminary results and the waveform matrix into a hybrid decision model, combination of the material dielectric constant and stress simulation data, multi-physical field coupling analysis, and generation of a comprehensive defect report; S6, optimization of the ultrasonic frequency, pulse parameters and incident angle according to the defect report by using a genetic algorithm, and formation of a closed-loop feedback instruction; S7, alignment of the feature distribution of the historical samples and the real-time data by transfer learning, and updating of the network weight; S8, deployment of a lightweight model on an edge computing terminal to realize millisecond-level detection of microcracks, delamination and peeling defects; S9, reconstruction of the defect three-dimensional point cloud based on the ray tracing algorithm, combination of the digital twin to generate a dynamic heat map and risk assessment; S10, packaging of the detection results, optimized parameters and visualized data into a structured report, transmission of the report to the cloud through 5G Internet of Things and triggering of the sorting instruction.

[0007] Preferably, the step S1 specifically comprises: S11, use of a 128-channel piezoelectric composite ultrasonic probe arranged in a ring array, programming adjustment of the probe center frequency, and self-adaptive adjustment of the probe spacing according to the insulator umbrella skirt curvature radius; S12, spiral trajectory scanning of the probe array along the insulator axis by a high-precision linear motor and a rotating platform, and scanning speed meeting the spatial resolution of less than 0.5mm; S13, synchronously collect the ultrasonic reflection signal amplitude, transmission signal attenuation coefficient and surface wave group velocity, the signal sampling rate is greater than 200MS / s, the dynamic range is greater than 80dB, and the resolution is improved.

[0008] Preferably, the adaptive noise reduction processing in step S2 further comprises: S21, performing variational mode decomposition on the original signal to extract the first 6 order intrinsic mode functions as effective signal components; S22, using an adaptive Kalman filter to iteratively predict and correct high-frequency noise; S23, calculating the multi-channel signal fusion weight through mutual information entropy and Pearson correlation coefficient, and executing weighted superposition to generate a noise reduction signal.

[0009] Preferably, the structure of the deep spatio-temporal convolutional neural network model comprises: S31, the input layer receives multi-channel ultrasonic time domain waveform, short-time Fourier transform spectrogram and propagation path matrix; S32, the 3D convolution module, the time convolution network module and the graph convolution module connected in parallel respectively extract spatial-spectral features, time sequence dependence and topological structure features; S33, the feature fusion layer dynamically allocates the weight of each feature channel through the attention mechanism to generate a unified feature representation; S34, the output layer uses a Softmax function to classify defect types, and predicts defect geometric parameters through a regression subnetwork, and the loss function adopts a combination form of Focal Loss and IoU Loss.

[0010] Preferably, the dynamic time warping algorithm of step S4 is specifically: S41, performing Hilbert-Huang transform on the static reference waveform to extract instantaneous frequency features and construct a multi-scale reference template library; S42, using an improved constrained dynamic time warping algorithm to calculate the local matching cost matrix of the detection waveform and the template waveform, and the constraint conditions include the maximum path slope and the phase continuity; S43, optimizing the global matching path through the simulated annealing algorithm to generate a phase compensation amount and a normalized similarity score, and the threshold is set to 0.85.

[0011] Preferably, the parameter optimization method of step S6 comprises: S61, establishing a propagation attenuation model of ultrasonic waves in a composite medium: wherein, is a material-related constant, is the incident angle: S62, taking the maximization of signal-to-noise ratio and the optimization of penetration depth as a multi-objective function, and solving a Pareto optimal parameter combination by using an NSGA-II multi-objective optimization algorithm.

[0012] Preferably, the domain adaptive transfer learning algorithm of step S7 specifically comprises: S71, visualizing high-dimensional feature distribution in a source domain historical sample library by using a t-SNE algorithm, and extracting cross-domain invariant features; S72, measuring feature distribution difference between the source domain and the target domain by using a Wasserstein distance, and constructing a domain adaptation loss function: wherein, is a balance coefficient, WD is used to solve the distribution non-overlapping problem caused by noise or deviation of cross-domain data, and MMD2 is used to fine align statistical characteristics of high-dimensional feature space; S73, adopting an adversarial learning strategy in neural network training, and realizing feature distribution alignment by using a gradient reversal layer.

[0013] Preferably, the embedded edge computing terminal implementation mode of step S8 is: S81, performing channel pruning and 8-bit integer quantization on a neural network model by using a TensorFlow Lite framework, and compressing model volume; S82, deploying a real-time operating system to ensure that a data acquisition, processing and transmission cycle is less than 10 ms; S83, integrating a temperature-humidity composite sensor and a self-adaptive gain control circuit, realizing environmental parameter self-compensation, and improving compensation accuracy.

[0014] Preferably, the three-dimensional point cloud reconstruction algorithm of step S9 comprises: S91, calculating a three-dimensional position of a defect point based on an ultrasonic wave time-of-flight difference and a probe space coordinate; S92, fitting a defect surface geometry by using a moving least squares method, and eliminating scanning track error; S93, identifying a defect cluster core area by using a DBSCAN density clustering algorithm, and labeling critical defect size and safety level.

[0015] Preferably, the AI-based ultrasonic detection method for internal defects of a composite insulator according to claim 1 further comprises the following steps: S111, construct a reinforcement learning agent, the state space of which includes real-time ultrasonic signal-to-noise ratio, defect detection confidence, probe environmental temperature and energy consumption index; S112, define the action space as ultrasonic emission frequency adjustment, pulse width adjustment range and incident angle fine adjustment step; S113, design a reward function: wherein, is a weight coefficient, and the SNR gain and detection accuracy improvement are calculated by comparing the current state with the historical baseline; S114, train the agent using a proximal policy optimization algorithm, interact with the environment 100,000 times until the policy converges, and generate an optimal parameter adjustment strategy; S115, deploy the agent online, dynamically adjust the transmission parameters according to real-time detection data, and keep the detection signal-to-noise ratio stable at more than 40dB while reducing unit energy consumption.

[0016] (Three) beneficial effects Compared with the prior art, the present application provides an AI-based ultrasonic detection method for internal defects of composite insulators, which has the following beneficial effects: 1. Multi-modal data fusion and adaptive noise reduction to improve detection reliability Through joint collection of reflection, transmission and surface guided wave multi-modal signals, the characteristics of insulator surface and internal defects are fully covered, and the detection blind area of single signal mode is overcome; wavelet packet threshold denoising and channel weighted fusion technology are used to effectively suppress environmental noise and equipment interference, and the signal-to-noise ratio is improved, providing a high-fidelity data basis for subsequent analysis.

[0017] 2. Deep spatio-temporal convolution network for high-precision defect recognition Through multi-branch feature extraction of time domain waveform, frequency domain spectrum and space propagation path, the sensitivity to micro-cracks, internal delamination and interface peeling is enhanced, and the detection accuracy is improved; the attention mechanism is introduced to dynamically allocate feature weights, solving the problem of insufficient capture ability of traditional models for complex defect characteristics and reducing the false detection rate.

[0018] 3. Dynamic waveform matching and multi-physical field coupling analysis to enhance robustness Based on the constrained dynamic time warping algorithm and simulated annealing optimization, high-precision phase alignment of static reference waveform and dynamic detection waveform is realized, and similarity calculation error is reduced; combined with material dielectric constant, stress distribution simulation data and actual working condition parameters, a multi-physical field coupling decision model is established, solving the misjudgment problem caused by changes in environmental temperature and humidity in traditional methods.

[0019] 4. Lightweight edge computing and transfer learning to improve engineering applicability Through model pruning and 8-bit quantization technology, the neural network parameter quantity is compressed to, the millisecond level real-time detection is realized in the embedded terminal, the high throughput demand of production line is met; the Wasserstein distance and the anti-learning strategy are used, the cross-domain feature alignment of the historical sample library and the real-time data is realized, and the generalization ability of the model in the new production line environment is improved.

[0020] 5. Three-dimensional visualization and closed-loop optimization empower intelligent management Based on the ray tracing algorithm and digital twin technology, a three-dimensional thermal map of defects is generated, and the spatial distribution and risk level of defects are intuitively displayed to assist operation and decision-making; through genetic algorithm and NSGA-II multi-objective optimization, the ultrasonic frequency, incident angle and other parameters are dynamically adjusted to improve detection efficiency and reduce energy consumption; relying on 5G Internet of Things, detection data cloud synchronization and automatic sorting are realized, and a full life cycle quality management closed loop is constructed. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 It is a schematic diagram of the overall system architecture of the application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0023] Please refer to Figure 1 The AI-based composite insulator internal defect ultrasonic detection method: S1, omnidirectional scanning of the composite insulator is performed through a multi-modal ultrasonic probe array, and the probe position and time stamp are recorded synchronously; 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, a deep spatio-temporal convolution network model is constructed, the preprocessed signal is input, the features are extracted through a multi-branch feature extraction module, and the preliminary results of defect type, position and size are output; S4, the static reference waveform and the real-time detection waveform are aligned based on the dynamic time warping algorithm, the phase shift and the similarity are calculated, and the waveform consistency matrix is generated; S5, the preliminary results and the waveform matrix are input into a hybrid decision model, the material dielectric constant and the stress simulation data are combined, multi-physical field coupling analysis is performed, and a comprehensive defect report is generated; S6, the ultrasonic frequency, pulse parameters and incident angle are optimized according to the defect report using genetic algorithm to form a closed-loop feedback instruction; S7, Align the feature distribution of historical samples and real-time data through transfer learning, update the network weight; S8, Deploy lightweight model in edge computing terminal to realize millisecond detection of micro-cracks, delamination and peeling defects; S9, Reconstruct the three-dimensional point cloud of the defect based on the ray tracing algorithm, and generate dynamic heat map and risk assessment combined with digital twin; S10, Package the detection results, optimization parameters and visualized data into a structured report, transmit it to the cloud through 5G Internet of Things and trigger the sorting instruction; S11, Use a 128-channel piezoelectric composite ultrasonic probe arranged in a ring array, with a central frequency of 2MHz~25MHz programmable adjustment, and a probe spacing that is self-adaptively adjusted according to the insulator umbrella skirt curvature radius; S12, Control the probe array to perform spiral trajectory scanning along the insulator axis through high-precision linear motor and rotary platform, with a scanning speed that meets the spatial resolution of less than 0.5mm; S13, Synchronously collect the amplitude of ultrasonic reflection signal, transmission signal attenuation coefficient and surface guided wave group velocity, with a signal sampling rate greater than 200MS / s and a dynamic range greater than 80dB, to improve the resolution; S21, Perform variational mode decomposition on the original signal to extract the first 6 order intrinsic mode functions as effective signal components; S22, Use an adaptive Kalman filter to iteratively predict and correct high-frequency noise; S23, Calculate the multi-channel signal fusion weight through mutual information entropy and Pearson correlation coefficient, and execute weighted superposition to generate noise reduction signal; S31, The input layer receives multi-channel ultrasonic time-domain waveform, short-time Fourier transform spectrum and propagation path matrix; S32, The 3D convolution module, time convolution network module and graph convolution module connected in parallel respectively extract spatial-spectral features, time sequence dependence and topological structure features; S33, The feature fusion layer dynamically allocates the weight of each feature channel through attention mechanism to generate unified feature representation; S34, The output layer uses Softmax function to classify defect types, and predicts defect geometric parameters through regression sub-network, with the loss function adopting the combined form of Focal Loss and IoU Loss; S41, Perform Hilbert-Huang transform on the static reference waveform to extract instantaneous frequency features and construct a multi-scale reference template library; S42, Use the improved constrained dynamic time warping algorithm to calculate the local matching cost matrix of the detection waveform and the template waveform, with the constraint conditions including maximum path slope and phase continuity; S43, optimize the global matching path by the simulated annealing algorithm, generate the phase compensation and the normalized similarity score, and set the threshold value to 0.85; S61, establish an ultrasonic wave propagation attenuation model in the composite medium: wherein, is a material-related constant, is the incident angle: S62, maximize the signal-to-noise ratio and optimize the penetration depth as a multi-objective function, and solve the Pareto optimal parameter combination by using the NSGA-II multi-objective optimization algorithm; S71, visualize the high-dimensional feature distribution in the source domain historical sample library by the t-SNE algorithm, and extract the cross-domain invariant features; S72, measure the feature distribution difference between the source domain and the target domain by the Wasserstein distance, and construct a domain adaptation loss function: wherein, is a balance coefficient, WD is used to solve the distribution non-overlapping problem caused by noise or deviation of cross-domain data, and MMD² is used to fine align the statistical characteristics of the high-dimensional feature space; S73, adopt an adversarial learning strategy in neural network training, and realize feature distribution alignment through a gradient reversal layer; S81, use the TensorFlow Lite framework to perform channel pruning and 8-bit integer quantization on the neural network model, and compress the model volume; S82, deploy a real-time operating system to ensure that the data acquisition, processing and transmission cycle is less than 10ms; S83, integrate a temperature-humidity composite sensor and a self-adaptive gain control circuit to realize environmental parameter self-compensation, and the compensation accuracy reaches ±0.05dB; S91, calculate the three-dimensional position of the defect point based on the ultrasonic wave time difference and the probe spatial coordinates; S92, use the moving least squares method to fit the defect surface geometry, and eliminate the scanning track error; S93, identify the defect cluster core area by the DBSCAN density clustering algorithm, and label the critical defect size and safety level; S111, construct a reinforcement learning intelligent agent, wherein the state space includes real-time ultrasonic signal-to-noise ratio, defect detection confidence, probe environmental temperature and energy consumption indicators; S112, define the action space as the ultrasonic wave transmission frequency adjustment amount, the pulse width adjustment range, and the incident angle fine adjustment step size; S113, design the reward function: wherein, is a weight coefficient, and the SNR gain and detection accuracy improvement are calculated by comparing the current state with the historical baseline; 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 strategy; S115, deploy the agent online, dynamically adjust the transmission parameters according to real-time detection data, and make the detection signal-to-noise ratio stable at more than 40dB while reducing unit energy consumption; A detection system for implementing the method of any one of claims 1-9, characterized in that it comprises: a multi-modal ultrasonic probe array module, a high-speed signal acquisition module, an AI analysis server, a dynamic optimization controller, an edge computing unit, and a visual interaction platform: The multi-modal ultrasonic probe array module includes a programmable frequency excitation source and a six-degree-of-freedom mechanical arm. The high-speed signal acquisition module integrates FPGA and ADC chips to realize multi-channel synchronous sampling. The AI analysis server is equipped with NVIDIA A100 GPU and runs the trained deep spatio-temporal convolutional neural network. The dynamic optimization controller adjusts the ultrasonic wave transmission parameters in real time through a fuzzy PID algorithm. The edge computing unit uses NVIDIA Jetson AGX Xavier to realize low-latency inference. The visual interaction platform supports VR / AR three-dimensional visualization and automatic report generation.

[0024] Embodiment 1: This embodiment details the implementation process of the AI-based ultrasonic detection method for internal defects of composite insulators: Step 1: Multi-modal ultrasonic signal acquisition 1.1 Hardware configuration: A 128-channel piezoelectric composite ultrasonic probe is arranged in a ring array, and the probe spacing is adjusted adaptively according to the insulator umbrella skirt curvature radius; The probe array is controlled by a six-axis high-precision mechanical arm to perform spiral trajectory scanning along the axial direction of the insulator, with a scanning speed of 10mm / s to ensure a spatial resolution of less than 0.5mm.

[0025] 1.2. Data acquisition: The reflection signal amplitude, transmission signal attenuation coefficient and surface wave group velocity are synchronously collected, and the probe spatial coordinates and time stamp are recorded.

[0026] Step 2: Adaptive signal denoising and fusion 2.1 Variational modal decomposition: The original signal is subjected to variational modal decomposition, and the first 6 order intrinsic modal functions are extracted to eliminate high frequency noise components.

[0027] 2.2 Kalman filter optimization: An adaptive Kalman filter is used to iteratively predict and correct the IMF components, and the initial value of the noise covariance matrix is set to 、 .

[0028] 2.3 Multi-channel fusion: The mutual information entropy and Pearson correlation coefficient of each channel signal are calculated to generate a fusion weight matrix: The weighted superposition is performed to generate a denoised signal.

[0029] Step 3: Deep spatio-temporal convolution network construction and training 3.1 Model structure: Input layer: receives 128-channel time-domain waveforms, short-time Fourier transform spectrograms and propagation path matrices.

[0030] Feature extraction layer: 3D convolution module: convolution kernel size 5x5x5, channel number 64, extract spatial-spectral features; Temporal convolution network: dilation factor 2, number of layers 4, capture temporal dependence; Graph convolution module: adjacency matrix is constructed based on probe spatial position, extract topological structure features.

[0031] Feature fusion layer: dynamically allocate feature weights through multi-head attention mechanism, generate 512-dimensional unified feature vector.

[0032] Output layer: Softmax classifier outputs defect type, regression sub-network predicts defect position and size.

[0033] 3.2 Training strategy: Dataset: 100,000 laboratory calibration samples and 50,000 field measurement data; Optimizer: AdamW, train until the accuracy of the validation set converges to 98.7%.

[0034] Step 4: Dynamic waveform matching and decision analysis 4.1 Waveform alignment: The Hilbert-Huang transform is applied to the static reference waveform to extract the instantaneous frequency feature and construct a multi-scale template library.

[0035] An improved constrained dynamic time warping algorithm is used to limit the maximum slope of the path to 2 and calculate the local matching cost matrix of the detection waveform and the template.

[0036] The global matching path is optimized by the simulated annealing algorithm to generate the phase compensation amount and similarity score.

[0037] 4.2 Hybrid decision model: The primary detection results and similarity matrix are input into the random forest-support vector machine hybrid model, combined with the material dielectric constant and finite element stress simulation data, to output the final defect judgment results.

[0038] Step 5: Edge computing and real-time detection 5.1 Model lightweight: TensorFlow Lite is used for channel pruning and 8-bit integer quantization of the network, and the model size is compressed from 350MB to 98MB.

[0039] 5.2 Embedded deployment: The model is deployed on the NVIDIA Jetson AGX Xavier edge terminal, integrated with temperature-humidity sensors and adaptive gain control circuits.

[0040] The real-time detection cycle is less than 8ms, supporting high-speed pipeline detection of more than 120 insulators per second.

[0041] Step 6: Three-dimensional visualization and closed-loop optimization 6.1 Defect reconstruction: The spatial coordinates of the defect points are calculated based on the ray tracing algorithm, and the moving least squares method is used.

[0042] The core area of the defect cluster is labeled by DBSCAN clustering, mapped to the digital twin model, and rendered as a heat map.

[0043] 6.2 Parameter optimization: The NSGA-II algorithm is used to optimize the ultrasonic frequency, pulse width, and other parameters, with the signal-to-noise ratio and energy consumption as the objective functions, to obtain the Pareto optimal solution set.

[0044] Real-time feedback to the probe control system reduces energy consumption by more than 25% while maintaining a signal-to-noise ratio of more than 35dB.

[0045] Effect verification of the embodiment: Detection accuracy: improve the detection accuracy of 2000 test samples and reduce the false negative rate; Real-time: The single piece detection takes 7.5ms, which is improved compared with traditional methods; Generalization ability: Improve the accuracy of cross-line migration testing and improve model adaptability.

[0046] Embodiment 2 The system composition of this embodiment corresponds to that of Embodiment 1, including a multi-modal ultrasonic probe array module, a high-speed signal acquisition module, an AI analysis server, a dynamic optimization controller, and a visualization panel.

[0047] Industrial application scenarios: In a certain extra-high voltage composite insulator production line, this method realizes daily detection of 120,000 pieces, reduces false detection rate, and defect positioning error is less than 0.3mm, which is superior to the industry standard.

Claims

1. An AI-based composite insulator internal defect ultrasonic detection method, characterized in that: The method comprises the following steps: S1, omnidirectional scanning of the composite insulator by a multi-modal ultrasonic probe array, and synchronous recording of the probe position and time stamp; S2, adaptive noise reduction of the original signal, including wavelet packet denoising, frequency domain filtering and channel weighted fusion, to generate a preprocessed signal set; S3, construction of a deep spatio-temporal convolution network model, input of the preprocessed signal, feature extraction by a multi-branch feature extraction module, and output of the preliminary results of the defect type, position and size; S4, alignment of the static reference waveform and the real-time detection waveform based on the dynamic time warping algorithm, calculation of the phase shift and similarity, and generation of a waveform consistency matrix; S5, input of the preliminary results and the waveform matrix into a hybrid decision model, combination of the material dielectric constant and stress simulation data, multi-physical field coupling analysis, and generation of a comprehensive defect report; S6, optimization of the ultrasonic frequency, pulse parameters and incident angle according to the defect report using a genetic algorithm, and formation of a closed-loop feedback instruction; S7, alignment of the feature distribution of the historical samples and the real-time data by transfer learning, and updating of the network weight; S8, deployment of a lightweight model on an edge computing terminal to realize millisecond-level detection of micro-cracks, delamination and peeling defects; S9, reconstruction of the defect three-dimensional point cloud based on the ray tracing algorithm, combination of the digital twin to generate a dynamic heat map and risk assessment; S10, packaging of the detection results, optimized parameters and visualized data into a structured report, transmission of the report to the cloud through 5G Internet of Things, and triggering of the sorting instruction.

2. The AI-based composite insulator internal defect ultrasonic detection method according to claim 1, characterized in that: The step S1 specifically comprises: S11, use of a 128-channel piezoelectric composite ultrasonic probe arranged in a ring array, programmable adjustment of the probe center frequency, and self-adaptive adjustment of the probe spacing according to the insulator shed curvature radius; S12, spiral trajectory scanning of the probe array along the insulator axis by a high-precision linear motor and a rotating platform, and scanning speed meeting the spatial resolution of less than 0.5mm; S13, synchronous acquisition of the ultrasonic wave reflection signal amplitude, transmission signal attenuation coefficient and surface guided wave group velocity, signal sampling rate greater than 200MS / s, dynamic range greater than 80dB, and improved resolution.

3. The AI-based composite insulator internal defect ultrasonic detection method according to claim 1, characterized in that: The adaptive noise reduction process in the step S2 further comprises: S21, variational mode decomposition of the original signal, and extraction of the first 6 order intrinsic mode functions as effective signal components; S22, use of an adaptive Kalman filter to iteratively predict and correct the high-frequency noise; S23, calculation of the multi-channel signal fusion weight by mutual information entropy and Pearson correlation coefficient, and execution of weighted superposition to generate a denoised signal.

4. The AI-based composite insulator internal defect ultrasonic detection method according to claim 1, characterized in that: The structure of the deep spatio-temporal convolution neural network model comprises: S31, the input layer receives the multi-channel ultrasonic time domain waveform, short-time Fourier transform spectrum and propagation path matrix; S32, the 3D convolution module, time convolution network module and graph convolution module connected in parallel respectively extract the spatial-spectral features, time sequence dependence and topological structure features; S33, the feature fusion layer dynamically allocates the weight of each feature channel by the attention mechanism to generate a unified feature representation; S34, the output layer adopts a Softmax function to classify defect types, and a regression sub-network is used to predict defect geometric parameters, and a loss function adopts a combination of Focal Loss and IoU Loss.

5. The AI-based composite insulator internal defect ultrasonic detection method according to claim 1, characterized in that: The dynamic time warping algorithm of the step S4 is specifically as follows: S41, Hilbert-Huang transform is performed on the static reference waveform to extract an instantaneous frequency feature, and a multi-scale reference template library is constructed; S42, an improved constrained dynamic time warping algorithm is used to calculate a local matching cost matrix of the detection waveform and the template waveform, and the constraint conditions include a maximum path slope and phase continuity; S43, a global matching path is optimized through a simulated annealing algorithm, a phase compensation amount and a normalized similarity score are generated, and a threshold is set to 0.

85.

6. The AI-based composite insulator internal defect ultrasonic detection method according to claim 1, characterized in that: The parameter optimization method of the step S6 includes: S61, an ultrasonic wave propagation attenuation model in a composite medium is established: wherein is a material dependent constant, is the angle of incidence: S62, a multi-objective function is established by maximizing a signal-to-noise ratio and optimizing a penetration depth, and a NSGA-II multi-objective optimization algorithm is used to solve a Pareto optimal parameter combination.

7. The AI-based composite insulator internal defect ultrasonic detection method according to claim 1, characterized in that: The domain adaptive transfer learning algorithm of the step S7 specifically includes: S71, a t-SNE algorithm is used to visualize a high-dimensional feature distribution in a source domain historical sample library, and cross-domain invariant features are extracted; S72, a domain adaptation loss function is constructed by measuring a feature distribution difference between the source domain and the target domain through a Wasserstein distance: wherein, is a balancing coefficient, WD is to solve the problem of non-overlapping distribution caused by noise or bias across domains, and MMD2 is to refine the statistical properties of the high-dimensional feature space. S73, an adversarial learning strategy is adopted in neural network training, and a feature distribution alignment is realized through a gradient reversal layer.

8. The AI-based composite insulator internal defect ultrasonic detection method according to claim 1, characterized in that: The embedded edge computing terminal implementation of the step S8 is as follows: S81, a TensorFlow Lite framework is used to perform channel pruning and 8-bit integer quantization on a neural network model, and the model volume is compressed; S82, a real-time operating system is deployed to ensure that a data acquisition, processing and transmission cycle is less than 10 ms; S83, a temperature-humidity composite sensor and a self-adaptive gain control circuit are integrated to realize environmental parameter self-compensation and improve compensation accuracy.

9. The AI-based composite insulator internal defect ultrasonic detection method according to claim 1, characterized in that: The three-dimensional point cloud reconstruction algorithm of the step S9 includes: S91, a three-dimensional position of a defect point is calculated based on an ultrasonic wave time-of-flight difference and a probe space coordinate; S92, a mobile least square method is used to fit a defect surface geometry, and a scanning track error is eliminated; S93, a DBSCAN density clustering algorithm is used to identify a defect cluster core area, and a critical defect size and a safety level are labeled. 10.The AI-based composite insulator internal defect ultrasonic detection method of claim 1, wherein, The following steps are further included: Step 11, real-time optimization of dynamic parameters based on deep reinforcement learning: Step 111, an intelligent agent of reinforcement learning is constructed, and a state space thereof includes a real-time ultrasonic signal-to-noise ratio, a defect detection confidence, a probe environmental temperature and an energy consumption index; Step 112, an action space is defined as an ultrasonic emission frequency adjustment amount, a pulse width adjustment range and an incident angle fine adjustment step; Step 113, a reward function is designed: wherein, are weight coefficients, SNR gain and detection accuracy improvement are calculated by comparing the current state with the historical baseline. Step 114, a proximal policy optimization algorithm is used to train the intelligent agent, and the intelligent agent interacts with the environment 100,000 times until the policy converges, and an optimal parameter adjustment strategy is generated; Step 115, online deployment of the intelligent agent, dynamically adjust the transmission parameters according to the real-time detection data, so that the detection signal-to-noise ratio is stable at more than 40dB, while reducing the unit energy.

Images