Ultrasonic live detection method for electrical equipment

Abstract

The invention relates to the technical field of electrical equipment detection, in particular to an ultrasonic live detection method for electrical equipment. According to the technical scheme, the ultrasonic live detection method for the electrical equipment comprises the following steps: a signal acquisition and synchronization step: acquiring an ultrasonic signal generated by the electrical equipment by using an ultrasonic sensor, and acquiring a background noise signal of a current environment by using an independently arranged noise reference sensor, the two signals are ensured to have synchronous timestamps; an edge intelligent preprocessing step: inputting the background noise signal as a reference into a preset adaptive noise cancellation algorithm in a field detection terminal, and performing preliminary noise reduction processing on the original ultrasonic signal to obtain a first-stage purification signal; the method improves the reliability and accuracy of detection, achieves the suppression of interference from the source through the synchronous noise collection and adaptive filtering of the edge side, and remarkably improves the signal-to-noise ratio of a signal.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment testing technology, and in particular to an ultrasonic live-line testing method for electrical equipment. Background Technology

[0002] The safe and stable operation of electrical equipment is crucial to the reliability of power systems. Ultrasonic live-line testing, as an effective condition monitoring method, is widely used to detect early insulation defects and potential mechanical failures in electrical equipment, such as partial discharge and loose components.

[0003] Existing ultrasonic live-line testing methods typically rely on maintenance personnel using handheld testing equipment to collect data on-site. The data is then brought back or processed briefly before being uploaded for preliminary analysis by expert systems or manual intervention. This approach has several significant limitations: First, the testing process lacks intelligence. Signals collected on-site are susceptible to interference from complex electromagnetic environments and background noise, and traditional filtering methods struggle to adaptively eliminate this interference, resulting in poor signal quality and reliability. Second, the testing and analysis stages are disconnected, failing to fully leverage the value of the data. Existing analysis methods often rely on isolated judgments of single test data, lacking trend analysis of the equipment's long-term operating status and thus failing to effectively predict potential risks. Third, the various stages of the testing process are relatively independent. From data acquisition, transmission, and analysis to decision-making and early warning, an effective closed loop is not formed, limiting the accuracy of test results and the timeliness of maintenance decisions, hindering the true shift from "preventive maintenance" to "predictive condition-based maintenance."

[0004] Existing ultrasonic charging detection methods have the following limitations: Poor on-site anti-interference capability and low detection reliability: Existing methods struggle to effectively distinguish between equipment defect signals and environmental background noise in complex on-site environments. Traditional filtering methods are fixed and singular, unable to adapt to varying noise levels, resulting in low signal-to-noise ratios and unstable signal quality in the acquired signals. This directly affects the accuracy of subsequent analysis, leading to misjudgments or missed diagnoses.

[0005] Data analysis is isolated and superficial, lacking predictive capabilities: Existing analytical methods are mostly limited to isolated judgments based on single test data, lacking correlation and trend analysis with the equipment's historical status. They cannot identify the slow degradation process of equipment, let alone provide early warnings and severity assessments of potential failure risks, resulting in reactive and delayed maintenance decisions.

[0006] The testing process is fragmented and fails to form a closed-loop management system: From data collection, transmission, and analysis to operation and maintenance decisions, the various links in the existing technology are loosely connected, failing to form an efficient automated system. In particular, there is a lack of a closed-loop mechanism to guide on-site maintenance based on diagnostic results and to use maintenance results to optimize the diagnostic model, resulting in low overall operation and maintenance efficiency and excessive reliance on human experience. Summary of the Invention

[0007] This invention proposes an ultrasonic live-line testing method for electrical equipment, which solves the problems of weak anti-interference ability, reliance on manual experience, disconnect between detection and analysis, lack of closed-loop optimization mechanism, resulting in poor detection reliability, low level of intelligence, insufficient status early warning capability and low operation and maintenance efficiency in the existing ultrasonic live-line testing methods for electrical equipment.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: An ultrasonic live-line testing method for electrical equipment includes the following steps: Signal acquisition and synchronization steps: Use an ultrasonic sensor to acquire ultrasonic signals generated by electrical equipment, and use a separately set noise reference sensor to acquire background noise signals of the current environment, and ensure that the two signals have synchronized timestamps; Edge intelligent preprocessing steps: In the detection terminal located on site, the background noise signal is used as a reference and input into a preset adaptive noise cancellation algorithm to perform preliminary noise reduction processing on the original ultrasonic signal to obtain the first-level purified signal; Feature extraction and uploading steps: On the detection terminal, a set of basic time-frequency features are extracted from the first-level purification signal, and the basic time-frequency features, the compressed data packet of the first-level purification signal, and the synchronous background noise signal are packaged together and uploaded to the cloud platform.

[0009] Furthermore, it also includes the following steps: Cloud-based deep analysis and diagnostic steps: On the cloud platform, the received first-level purified signal is first subjected to secondary fine noise reduction using a deep learning-based signal purification model to obtain a finely purified signal. Then, deep nonlinear features are extracted from the finely purified signal. The basic time-frequency features and deep nonlinear features are fused and input into the defect identification and classification model to obtain the diagnostic results of defect type and severity. Situation assessment and early warning steps: The cloud platform queries the historical data of the tested equipment, calculates the trend change of the current diagnostic results relative to the historical state, and generates graded early warning information through a risk assessment matrix by combining the importance level of the tested equipment in the power grid.

[0010] Furthermore, the adaptive noise cancellation algorithm in the edge intelligent preprocessing step is a normalized least mean square algorithm. Its execution process includes using the background noise signal as a reference input and the original ultrasonic signal as the main input. A tunable filter is used to filter the reference input to generate an estimated noise value. This estimated value is then subtracted from the main input to generate an error signal, which is the first-level purified signal. Simultaneously, this error signal is fed back to dynamically adjust the coefficients of the tunable filter to minimize the power of the error signal, thereby achieving adaptive cancellation of environmental noise. Further, the basic time-frequency features extracted in the feature extraction and uploading steps include the effective value of the signal within a specific frequency band, the total energy of the signal within a predetermined frequency band, the zero-crossing rate of the signal waveform per unit time, and the standard deviation of the signal amplitude calculated using a fixed time window. The compressed data packet is a data packet obtained by downsampling and encoding the first-level purified signal using a lossy compression algorithm. Furthermore, the signal purification model based on deep learning in the cloud-based deep analysis and diagnosis step is a one-dimensional convolutional neural network, whose network structure includes multiple alternating convolutional layers, activation function layers, and downsampling layers; the model uses signals collected in a noisy environment and their corresponding clean signals as training samples to learn the end-to-end mapping relationship from noisy signals to clean signals. Furthermore, the extraction of deep nonlinear features specifically involves: forward propagating the refined signal after secondary fine noise reduction through the one-dimensional convolutional neural network, and extracting the feature maps output by the deep convolutional layers of the network as the deep nonlinear feature representation of the signal. These feature maps can characterize abstract high-order patterns related to defects in the signal. Furthermore, the defect identification and classification model is a support vector machine model, with a radial basis function as its kernel function. Before classification, the fused feature vectors are first standardized to eliminate the influence of different feature dimensions. Then, the model is used to construct the optimal classification hyperplane in the high-dimensional feature space to achieve classification of various defect types and their severity. Furthermore, the risk assessment matrix in the situation assessment and early warning step is a two-dimensional lookup table. One dimension represents the severity of the defect, and the other dimension represents the degradation rate calculated based on historical data trends. Each cell in the matrix corresponds to a specific early warning level, and the final early warning level will be adjusted according to the preset importance level of the device being tested. Furthermore, the method also includes a feedback optimization step, which is performed after the situation assessment and early warning steps: when the early warning information triggers on-site maintenance, the actual results confirmed by the maintenance are compared with the diagnostic results, and the detection data and confirmation results are used as new training samples to incrementally learn the defect identification and classification model, so as to achieve continuous optimization of the model's diagnostic capabilities. Furthermore, before the signal acquisition and synchronization step, a detection path planning step is also included: the cloud platform calculates an optimal detection point access sequence using an optimization algorithm based on the detection cycle of the device, the importance weight of the device, and the current location of the detection terminal, with the optimization goal of minimizing the total path length and prioritizing the detection of key devices, and sends the sequence to the detection terminal to guide the detection execution. The method also includes a cross-domain knowledge transfer step for initializing the defect identification and classification model: first, the model is pre-trained using desensitized defect datasets from multiple different data sources, and then the model is fine-tuned using a private dataset from the current target region to improve the initial performance and generalization ability of the model when the target region has limited data.

[0011] The positive effects of this invention are: This enhances the reliability and accuracy of detection. By employing synchronous noise acquisition and adaptive filtering at the edge, interference is suppressed at its source, significantly improving the signal-to-noise ratio. Combined with deep signal cleansing and intelligent feature fusion in the cloud, features relevant to the essence of defects can be extracted more accurately, laying a reliable data foundation for subsequent diagnosis and effectively reducing false alarms and false negatives.

[0012] It achieves deep perception and intelligent early warning of equipment status. By correlating and analyzing real-time data with historical trends and equipment family trees, it has made a leap from single-point judgment to continuous situation assessment. The system can not only identify current defects, but also see the trend of status deterioration, and generate graded early warnings based on the importance of equipment, providing a forward-looking scientific basis for operation and maintenance strategies and realizing truly intelligent status management.

[0013] A self-optimizing closed-loop operation and maintenance ecosystem has been built. By introducing a feedback optimization mechanism, the system can continuously correct and improve the accuracy of the diagnostic model using on-site inspection results, forming a virtuous cycle of "practice-verification-learning-improvement". This enables the entire detection system to continuously evolve, becoming smarter with use and offering high long-term maintenance value.

[0014] The overall efficiency of inspection and maintenance has been improved. The cloud-edge collaborative architecture rationally allocates the computing load, edge processing ensures real-time performance and reduces bandwidth requirements, while cloud-based deep analysis leverages the advantages of centralized computing. Combined with intelligent path planning, the inspection workflow has been optimized, thereby improving the management and execution efficiency of large-scale equipment inspection while ensuring inspection depth. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0016] Example 1 An ultrasonic live-line testing method for electrical equipment includes the following steps: Signal acquisition and synchronization steps: Use an ultrasonic sensor to acquire ultrasonic signals generated by electrical equipment, and use a separately set noise reference sensor to acquire background noise signals of the current environment, and ensure that the two signals have synchronized timestamps; This step is performed on-site at the electrical equipment and requires a dedicated testing hardware unit. This unit contains two core sensors: an ultrasonic sensor, whose acoustic probe must be tightly coupled or oriented to a predetermined detection point on the surface of the equipment under test, to capture ultrasonic acoustic signals generated by partial discharge or mechanical vibration; and a noise reference sensor, whose installation location must be carefully selected to effectively sense ambient background noise rather than the equipment's own signals, for example, it can be placed on the housing of the testing unit or on the adjacent equipment support structure.

[0017] The key to achieving signal synchronization lies in employing a unified time base system. Specifically, the detection unit integrates a high-precision clock source, assigning identical timestamps to the data streams from both sensors each time an acquisition event is triggered. Alternatively, the detection unit can support GPS or network time synchronization, ensuring that data collected from different terminals and batches are aligned under a unified absolute time reference. This rigorous synchronization mechanism provides a crucial timing correlation foundation for distinguishing between the device's actual signals and random environmental noise in subsequent steps.

[0018] Edge intelligent preprocessing steps: In the detection terminal located on site, the background noise signal is used as a reference and input into a preset adaptive noise cancellation algorithm to perform preliminary noise reduction processing on the original ultrasonic signal to obtain the first-level purified signal; This step is completed within the embedded system of the on-site detection terminal. Its core task is to run a pre-configured adaptive noise cancellation algorithm. The execution flow of this algorithm is as follows: the system uses the signal acquired by the noise reference sensor as the reference input channel and the signal acquired by the ultrasonic sensor as the main input channel. The algorithm internally maintains a digital filter with dynamically adjustable coefficients (e.g., a finite-length unit impulse response filter).

[0019] After initialization, the algorithm begins iterative processing: First, it uses the current coefficients of the filter to calculate the reference input signal, generating an estimate of the noise components in the main input channel. Then, this estimate is subtracted from the main input signal in real time to obtain the difference signal, i.e., the error signal. This error signal is considered the "first-stage purified signal." Simultaneously, the value of this error signal is fed back to the filter coefficient update module, which dynamically adjusts the filter coefficients according to a specific adaptive law (such as the least mean square criterion), aiming to continuously minimize the power of the error signal. Through this closed-loop feedback adjustment, the system can automatically track and cancel changing environmental noise, achieving initial enhancement of the original ultrasonic signal.

[0020] Feature extraction and uploading steps: On the detection terminal, a set of basic time-frequency features are extracted from the first-level purification signal, and the basic time-frequency features, the compressed data packet of the first-level purification signal, and the synchronous background noise signal are packaged together and uploaded to the cloud platform.

[0021] After generating the first-level purified signal, the detection terminal does not directly upload the massive amount of raw waveform data. Instead, it performs intelligent feature extraction and data compression. The feature extraction module calculates a set of fundamental features that characterize the key time-domain and frequency-domain attributes of the signal. These features typically include, but are not limited to: the integral energy of the signal within a certain frequency band, the zero-crossing rate reflecting the complexity of the signal waveform, and the statistical fluctuation characteristics of the signal amplitude. Calculating these features aims to summarize the main characteristics of the signal with a small number of numerical values.

[0022] Meanwhile, to preserve the complete information of the original signal for in-depth cloud analysis, the system compresses the first-stage purified signal. The compression process employs a lossy compression algorithm, significantly reducing the data volume while ensuring no loss of key signal components. Finally, the calculated basic time-frequency characteristic values, the compressed signal data packet, and the synchronized background noise signal sample are encapsulated into a standard format data packet and uploaded to the cloud platform via a wireless network (such as 4G / 5G or Wi-Fi). This strategy effectively balances the conflict between data transmission efficiency and the integrity of information required for cloud analysis.

[0023] It also includes the following steps: Cloud-based deep analysis and diagnostic steps: On the cloud platform, the received first-level purified signal is first subjected to secondary fine noise reduction using a deep learning-based signal purification model to obtain a finely purified signal. Then, deep nonlinear features are extracted from the finely purified signal. The basic time-frequency features and deep nonlinear features are fused and input into the defect identification and classification model to obtain the diagnostic results of defect type and severity. After receiving the data packets uploaded by the edge terminal, the cloud platform initiates a deep analysis process. First, a secondary fine-tuning denoising is performed: the platform calls a pre-trained deep learning signal purification model (such as a one-dimensional convolutional neural network). This model runs on powerful cloud computing resources, capable of performing more complex calculations than those at the edge. The model receives compressed signal data (decompressed) uploaded from the edge terminal as input, undergoes multiple layers of nonlinear transformations within the network, and outputs the predicted "clean" defect signal, i.e., the "fine-purified signal." This step aims to eliminate any residual noise or distortion that may have been introduced during edge processing.

[0024] Subsequently, the feature extraction module mines deeper features from the refined signal, namely "deep nonlinear features." This is typically achieved by feeding the signal into an intermediate layer of a trained network and extracting its activation values. These features represent abstract, higher-order patterns in the signal that are related to the physical mechanisms of the defects.

[0025] Finally, feature fusion and diagnosis are performed: the basic time-frequency features from the edge are concatenated or weighted with the deep nonlinear features extracted from the cloud to form a comprehensive feature vector. This vector is then input into another specialized defect identification and classification model (such as a support vector machine or a deep learning classifier). This model analyzes the comprehensive feature vector and outputs probabilistic diagnostic results regarding the defect type (such as corona discharge, surface discharge) and its severity (such as mild, moderate, severe).

[0026] Situation assessment and early warning steps: The cloud platform queries the historical data of the tested equipment, calculates the trend change of the current diagnostic results relative to the historical state, and generates graded early warning information through a risk assessment matrix by combining the importance level of the tested equipment in the power grid.

[0027] This step represents a leap from single-time diagnosis to continuous condition management. The cloud platform's data management system queries the historical database of the device under test, retrieving previous diagnostic results or relevant characteristic values ​​for the same measurement point. By comparing current data with historical sequences, it calculates the changing trends of key indicators (such as signal strength and defect probability), determining whether they are steadily increasing, accelerating deterioration, or remaining stable.

[0028] Simultaneously, the system maintains a device knowledge base, defining the importance level of each device in the power grid (e.g., critical, important, general). The platform's built-in risk assessment matrix is ​​a pre-defined two-dimensional decision table. One dimension represents the severity of the currently diagnosed defect, and the other is the calculated degradation trend. Based on these two dimensions, the matrix maps an initial warning level (e.g., attention, abnormal, severe). Finally, the system adjusts this warning level based on the device's importance level (e.g., increasing the warning level for critical equipment), thereby generating the final graded warning information, which is then pushed to relevant maintenance personnel or the production management system via an interface.

[0029] A basic architecture of "edge perception - preprocessing - feature uploading" was constructed. Its benefits lie in improving signal quality from the source through synchronous signal acquisition and intelligent edge preprocessing, and significantly reducing data transmission bandwidth requirements, making large-scale, frequent live-line testing feasible. Based on this, the introduction of cloud-based deep analysis and situation assessment steps forms an efficient cloud-edge collaboration. Its benefits lie in transforming single-point detection data into time-dimensional equipment health situation awareness, and making intelligent decisions based on comprehensive diagnostic results, historical trends, and equipment importance. Ultimately, this achieves a leap from simple "anomaly detection" to "precise diagnosis and risk assessment," providing direct and reliable decision-making basis for condition-based maintenance.

[0030] The adaptive noise cancellation algorithm in the edge intelligent preprocessing step is a normalized least mean square algorithm. Its execution process includes taking the background noise signal as the reference input and the original ultrasonic signal as the main input, filtering the reference input through an adjustable filter to generate an estimated value of the noise, and then subtracting the estimated value from the main input to generate an error signal. This error signal is the first-level purification signal. At the same time, the error signal is fed back to dynamically adjust the coefficients of the adjustable filter to minimize the power of the error signal, thereby achieving adaptive cancellation of environmental noise.

[0031] The algorithm involves continuous data stream processing. In each iteration, the algorithm performs the following core operation: the reference input signal (i.e., background noise) is passed through a digital filter with variable coefficients to produce an output signal that is the current best estimate of the noise component in the main input signal (i.e., the original ultrasonic signal). This estimate is subtracted from the main input signal to obtain the instantaneous error signal.

[0032] The error signal serves a dual purpose: firstly, it is itself the noise-reduced output signal (the first-stage clean signal); secondly, it is used as a control signal to update the filter coefficients. The coefficient update is proportional to the product of the error signal and the reference input signal, but to ensure the stability of the algorithm under different input signal powers, the update step size is normalized according to the power of the reference input signal. This means that during periods of high noise, the adjustment step size will automatically decrease to prevent system instability; during periods of low noise, the step size will relatively increase to accelerate convergence.

[0033] Through this real-time, instantaneous error-based anti-caking mechanism, the filter coefficients can dynamically track the changes in the correlation between ambient noise and noise components in the main input channel, thereby achieving efficient and robust adaptive noise cancellation.

[0034] By employing a normalized least mean square algorithm for adaptive noise cancellation, the system can dynamically track and cancel complex and varied environmental noise, significantly improving the signal-to-noise ratio. This adaptive characteristic makes the detection method more robust to different environmental conditions, effectively reducing false alarms and missed alarms caused by environmental interference, and ensuring the reliability and accuracy of the primary signal preprocessing stage.

[0035] The basic time-frequency features extracted in the feature extraction and uploading steps include the effective value of the signal in a specific frequency band, the total energy of the signal in a predetermined frequency band, the zero-crossing rate of the signal waveform per unit time, and the standard deviation of the signal amplitude calculated with a fixed time window; the compressed data packet is a data packet obtained by downsampling and encoding the first-level purified signal using a lossy compression algorithm.

[0036] At the detection terminal, the feature extraction module performs a series of predefined mathematical operations on the first-level purification signal: Bandwidth energy: The integral value of the signal's energy within a specific band of interest (e.g., a frequency range corresponding to typical discharge characteristics) is calculated using a digital filter bank or Fast Fourier Transform. This feature helps to focus on specific frequency band activities related to defects.

[0037] Total Energy: Calculates the total energy of the signal within the preset analysis frequency band, reflecting the overall strength of the signal.

[0038] Zero-crossing rate: The number of times a signal waveform crosses the zero level per unit time. This characteristic is related to the frequency components of the signal; signals rich in high-frequency components generally have a higher zero-crossing rate.

[0039] Amplitude standard deviation: The standard deviation of signal sample values ​​is calculated within a sliding time window of fixed length. This feature quantifies the degree of fluctuation in signal amplitude and can be used to identify impulsive or fluctuating discharges.

[0040] Data compression implementation details: The compression process aims to reduce the amount of data for transmission. The lossy compression algorithms used are typically based on transform coding principles (such as wavelet transform): First, the time-domain signal is transformed to another domain (such as the wavelet domain), where the signal's energy is concentrated on a few coefficients. Then, the transformed coefficients are quantized and entropy-coded, discarding subtle information that has little impact on signal reconstruction quality (i.e., introducing distortion), thereby significantly reducing the data rate. Although the decompressed signal is not a perfect reconstruction of the original signal, it retains its main form and key information, sufficient to meet the needs of deep processing in the cloud.

[0041] By extracting a set of carefully designed fundamental time-frequency features and combining them with data compression transmission, an optimal balance between data volume and information integrity is achieved. This fully preserves key information reflecting the device status while greatly reducing the burden on the communication link, ensuring the efficiency and real-time performance of the detection system. It is particularly suitable for application scenarios with limited network conditions or a large number of devices.

[0042] In the cloud-based deep analysis and diagnostic steps, the signal purification model based on deep learning is a one-dimensional convolutional neural network. Its network structure includes multiple alternating convolutional layers, activation function layers, and downsampling layers. The model uses signals collected in a noisy environment and their corresponding clean signals as training samples to learn the end-to-end mapping relationship from noisy signals to clean signals. A one-dimensional convolutional neural network model is constructed and operates as follows: its input is a sequence of one-dimensional ultrasonic signals that has undergone edge preprocessing and compression. The network structure consists of multiple functional layers stacked alternately. Convolutional layers: These layers use one-dimensional convolutional kernels that slide across the input signal to extract local features. Each kernel is responsible for detecting a specific local pattern (such as a pulse of a specific shape).

[0043] Activation function layer: Non-linear activation functions such as linear rectified functions are typically used to introduce non-linear transformation capabilities into the network, enabling it to learn complex mapping relationships.

[0044] Downsampling layers (such as pooling layers): reduce computational cost and enhance the model's robustness to small signal displacements by reducing data dimensionality.

[0045] During the training phase, the network learns using a large number of sample pairs, each containing a "noisy signal" (input) and a corresponding "ideal clean signal" (target output). The training process adjusts all network weight parameters using optimization algorithms (such as gradient descent) with the goal of making the network's output signal as close as possible to the "ideal clean signal." Once training is complete, the network possesses the ability to estimate a clean signal from a noisy input, achieving end-to-end signal cleansing.

[0046] By employing a one-dimensional convolutional neural network for secondary fine-tuning in the cloud, the powerful feature learning capabilities of deep learning models can be leveraged to remove complex noise and distortions from signals that are difficult to completely eliminate through edge processing, resulting in cleaner defect signals. This lays a more reliable data foundation for subsequent accurate feature extraction and defect identification.

[0047] The extraction of deep nonlinear features specifically involves: forward propagating the refined cleaned signal after secondary fine noise reduction through the one-dimensional convolutional neural network, and extracting the feature map output of the deep convolutional layer in the network as the deep nonlinear feature representation of the signal. This feature map can characterize the abstract high-order patterns related to defects in the signal.

[0048] This process occurs in the cloud, using a pre-trained signal to refine the CNN model. Specifically, the refined signal (i.e., the output or intermediate result of the CNN model) is taken as input and propagated forward through the CNN again. However, we do not use the network's final output, but rather truncate it to a deeper layer of the network (e.g., after the last convolutional layer and before the fully connected layer).

[0049] The activation values ​​(i.e., output values) of all neurons (or convolutional kernels) in this layer are treated as a single output. This output is a multi-dimensional feature tensor (or feature map), where each dimension represents some abstract, high-order characteristic of the input signal. These characteristics are automatically acquired by the network through layered learning and may correspond to more complex firing patterns or fault symptoms, far exceeding the capabilities of manually designed simple time-frequency features. This feature tensor is the "deep nonlinear feature" used for subsequent classification.

[0050] By extracting nonlinear features from deep learning models, it is possible to capture abstract, high-order patterns in signals that are difficult to describe using traditional methods. These patterns are more closely related to the physical mechanisms of equipment defects, thus greatly enriching the information content of fault features and providing crucial support for subsequent more refined and accurate defect classification. The defect identification and classification model is a support vector machine model, and its kernel function is a radial basis function. Before classification, the fused feature vectors are first standardized to eliminate the influence of different feature dimensions. Then, the model is used to construct the optimal classification hyperplane in the high-dimensional feature space to classify various defect types and their severity.

[0051] The implementation process is divided into two phases: training and application. Training Phase: Collect a large number of ultrasound signal samples with known defect types and severity, and extract their fused feature vectors. Use these labeled samples to train the SVM model. For multi-class classification problems, a one-to-one or one-to-many strategy is often used to construct multiple binary classifiers. The role of the radial basis function kernel is to map the original feature space to a higher-dimensional space, so that samples of different classes can be linearly separated by a hyperplane in this higher-dimensional space. The training process is the process of finding this optimal hyperplane.

[0052] Application Phase: Before classifying a new signal, its fused feature vector must be standardized. This is typically achieved by subtracting the mean of the training set features and dividing by the standard deviation. The goal is to ensure all feature dimensions have the same scale, preventing certain features with large values ​​from dominating the classification decision. Then, the standardized feature vector is input into a trained SVM model. The model then provides the classification result based on the relationship between the vector's position in the feature space and the optimal hyperplane.

[0053] By employing a support vector machine model and standardizing features, an optimal classification boundary can be constructed in a high-dimensional feature space, thereby achieving high-precision and high-stability identification of various defect types and their severity. This method maintains good generalization ability even with small sample sizes, effectively improving the practicality and reliability of the diagnostic model.

[0054] The risk assessment matrix in the situation assessment and early warning steps is a two-dimensional lookup table. One dimension represents the severity of the defect, and the other dimension represents the degradation rate calculated based on historical data trends. Each cell in the matrix corresponds to a specific early warning level, and the final early warning level will be adjusted according to the preset importance level of the tested equipment.

[0055] The risk assessment matrix is ​​a lookup table pre-defined in the cloud platform software. Its two dimensions are defined as follows: Defect severity: Based on the results of the diagnostic model output (such as probability value or confidence level), it is divided into several levels (such as low, medium, high).

[0056] Degradation rate: Determined based on time series analysis of the device's historical test data (such as the slope of linear regression fitting), and also divided into several levels (such as slow, medium, and fast).

[0057] Each cell in the matrix (the intersection of rows and columns) has a pre-set suggested alert level (e.g., low severity + slow degradation = "Caution"; high severity + rapid degradation = "Urgent").

[0058] During actual assessment, the system automatically queries this matrix based on the current diagnostic results (determining the severity level) and the calculated trend (determining the rate of degradation level) to obtain a basic warning level. Subsequently, the system reads the preset importance weights for the equipment (e.g., critical equipment has a higher weight) and corrects the basic warning level according to predetermined rules. For example, the rule might stipulate that for critical equipment, the warning level is automatically increased by one level. Finally, a corrected, graded warning message incorporating the equipment's importance is generated.

[0059] By constructing a two-dimensional risk assessment matrix that comprehensively considers the severity of defects and the rate of degradation, and adjusting it according to the importance of equipment, intelligent and refined decision-making on early warning levels is achieved. It avoids the limitations of the single threshold method, can identify risks that accelerate degradation earlier, and can allocate maintenance resources differently based on the criticality of equipment, making early warning information more practically instructive.

[0060] Example 2 Based on Example 1, the method further includes a feedback optimization step, which is performed after the situation assessment and early warning steps: when the early warning information triggers on-site maintenance, the actual results of the maintenance confirmation are compared with the diagnostic results, and the detection data and confirmation results are used as new training samples to incrementally learn the defect identification and classification model in order to achieve continuous optimization of the model's diagnostic capabilities. A self-optimizing feedback loop has been added, with the feedback optimization steps triggered after maintenance activities. When a system-issued warning leads to on-site repairs, maintenance personnel will manually enter or transmit the actual fault information (i.e., the real label) back to the cloud platform via system interface.

[0061] The platform binds the original data and diagnostic results corresponding to this detection with the real labels to form a new, validated training sample. The platform will periodically (or when a certain number of new samples have been accumulated) initiate the incremental learning (or online learning) process of the model. In this process, the parameters of the defect identification and classification model (such as SVM or deep learning model) are updated and retrained (or fine-tuned) using the original training set plus the newly collected feedback samples.

[0062] In this way, the model can learn from new practical experience, correct possible misjudgment patterns, and gradually adapt to new defect types, thereby achieving continuous evolution of diagnostic capabilities and making the entire system increasingly intelligent and reliable.

[0063] By introducing an incremental learning mechanism based on on-site maintenance feedback, the entire detection system acquires the ability to self-optimize and continuously evolve. The model can continuously learn new defect patterns and correct errors from practice, thereby constantly improving its diagnostic accuracy and adaptability over time, forming a virtuous cycle of becoming smarter with use.

[0064] Before the signal acquisition and synchronization step, there is also a detection path planning step: the cloud platform calculates an optimal detection point access sequence based on the detection cycle of the device, the importance weight of the device, and the current location of the detection terminal, with the goal of minimizing the total path length and prioritizing the detection of key devices, and sends the sequence to the detection terminal to guide the detection execution. The method also includes a cross-domain knowledge transfer step for initializing the defect identification and classification model: first, the model is pre-trained using desensitized defect datasets from multiple different data sources, and then the model is fine-tuned using a private dataset from the current target region to improve the initial performance and generalization ability of the model when the target region has limited data.

[0065] Implementation details of the inspection path planning step: Before the planned inspection task begins, the cloud platform's path planning module starts working. The inputs to this module include: a list of devices to be inspected, the geographical coordinates of each device, the importance weight of each device (from the asset management system), the preset inspection cycle requirements, and the current location of the inspection terminal (or inspection robot).

[0066] Planning algorithms (such as genetic algorithms, simulated annealing algorithms, or heuristic rules) target all detection points, with the optimization objectives of "shortest total path travel distance" and "prioritizing coverage of high-weight devices," to calculate an optimal sequence of detection point visits. This sequence ensures the efficient execution of the detection task. The final generated detection path plan is then distributed to the on-site detection terminals or autonomous inspection robots, guiding them to execute the detection tasks sequentially.

[0067] Implementation details of the cross-domain knowledge transfer step: This step is used to optimize the initialization of the defect identification and classification model, especially when local data is insufficient. Implementation is divided into two phases: Pre-training: Under centralized, privacy-preserving conditions (e.g., through a federated learning framework), a general defect identification model is jointly trained using ultrasonic defect datasets from multiple different sources (e.g., different regions, different power grid companies) that have undergone anonymization (removal of sensitive information). This model learns the universality and fundamental knowledge about defects in various types of electrical equipment.

[0068] Fine-tuning: Using the pre-trained general model as the initial model, and then using a small number of private datasets held by the local region (target domain), perform a few additional rounds of training (fine-tuning) on ​​all or some of the parameters of the model.

[0069] This strategy enables the model to transfer general defect knowledge to local specific tasks and quickly adapt to local device characteristics and environmental features, significantly improving the model's performance and generalization ability in the initial stage and effectively overcoming the challenge of data scarcity.

[0070] By adding a detection path planning step, the execution of detection tasks is made more intelligent and efficient, significantly improving inspection efficiency. The introduction of a cross-domain knowledge transfer strategy has the advantage of rapidly improving the model's initial performance and generalization ability in specific scenarios by utilizing external knowledge, effectively overcoming the difficulties in model training caused by insufficient local data, and accelerating the deployment and application of the system.

[0071] The above-described embodiments are detailed and specific, illustrating preferred embodiments of the present invention. They are only used to illustrate the technical ideas and features of the present invention, with the aim of enabling those skilled in the art to understand the content of the present invention and implement it accordingly. However, they are not limited to the present invention, and the patent scope of the present invention cannot be limited by this embodiment alone. That is, any equivalent changes or modifications made to the spirit disclosed in the present invention, without departing from the structure of the present invention, such as local improvements within the system and modifications or transformations between subsystems, are still within the patent scope of the present invention.

Claims

1. A method for ultrasonic live-line testing of electrical equipment, characterized in that, Includes the following steps: Signal acquisition and synchronization steps: Use an ultrasonic sensor to acquire ultrasonic signals generated by electrical equipment, and use a separately set noise reference sensor to acquire background noise signals of the current environment, and ensure that the two signals have synchronized timestamps; Edge intelligent preprocessing steps: In the detection terminal located on site, the background noise signal is used as a reference and input into a preset adaptive noise cancellation algorithm to perform preliminary noise reduction processing on the original ultrasonic signal to obtain the first-level purified signal; Feature extraction and uploading steps: On the detection terminal, a set of basic time-frequency features are extracted from the first-level purification signal, and the basic time-frequency features, the compressed data packet of the first-level purification signal, and the synchronous background noise signal are packaged together and uploaded to the cloud platform.

2. The ultrasonic live-line detection method for electrical equipment according to claim 1, characterized in that, It also includes the following steps: Cloud-based deep analysis and diagnostic steps: On the cloud platform, the received first-level purified signal is first subjected to secondary fine noise reduction using a deep learning-based signal purification model to obtain a finely purified signal. Then, deep nonlinear features are extracted from the finely purified signal. The basic time-frequency features and deep nonlinear features are fused and input into the defect identification and classification model to obtain the diagnostic results of defect type and severity. Situation assessment and early warning steps: The cloud platform queries the historical data of the tested equipment, calculates the trend change of the current diagnostic results relative to the historical state, and generates graded early warning information through a risk assessment matrix by combining the importance level of the tested equipment in the power grid.

3. The ultrasonic live-line detection method for electrical equipment according to claim 2, characterized in that, The adaptive noise cancellation algorithm in the edge intelligent preprocessing step is a normalized least mean square algorithm. Its execution process includes taking the background noise signal as a reference input and the original ultrasonic signal as the main input, filtering the reference input through an adjustable filter to generate an estimated value of the noise, and then subtracting the estimated value from the main input to generate an error signal. This error signal is the first-level purification signal. At the same time, the error signal is fed back to dynamically adjust the coefficients of the adjustable filter to minimize the power of the error signal, thereby achieving adaptive cancellation of environmental noise.

4. The ultrasonic live-line detection method for electrical equipment according to claim 2, characterized in that, The basic time-frequency features extracted in the feature extraction and uploading steps include the effective value of the signal in a specific frequency band, the total energy of the signal in a predetermined frequency band, the zero-crossing rate of the signal waveform per unit time, and the standard deviation of the signal amplitude calculated with a fixed time window; the compressed data packet is a data packet obtained by downsampling and encoding the first-level purified signal using a lossy compression algorithm.

5. The ultrasonic live-line detection method for electrical equipment according to claim 2, characterized in that, The signal purification model based on deep learning in the cloud-based deep analysis and diagnosis step is a one-dimensional convolutional neural network. Its network structure includes multiple alternating convolutional layers, activation function layers, and downsampling layers. The model uses signals collected in a noisy environment and their corresponding clean signals as training samples to learn the end-to-end mapping relationship from noisy signals to clean signals.

6. The ultrasonic live-line detection method for electrical equipment according to claim 5, characterized in that, The extraction of the deep nonlinear features specifically involves: forward propagating the refined signal after secondary fine noise reduction through the one-dimensional convolutional neural network, and extracting the feature maps output by the deep convolutional layers of the network as the deep nonlinear feature representation of the signal. This feature map can characterize abstract high-order patterns in the signal that are related to defects.

7. The ultrasonic live-line detection method for electrical equipment according to claim 2, characterized in that, The defect identification and classification model is a support vector machine model, and its kernel function is a radial basis function. Before classification, the fused feature vectors are first standardized to eliminate the influence of different feature dimensions. Then, the model is used to construct the optimal classification hyperplane in the high-dimensional feature space to achieve the classification of multiple defect types and their severity.

8. The ultrasonic live-line detection method for electrical equipment according to claim 2, characterized in that, The risk assessment matrix in the situation assessment and early warning steps is a two-dimensional lookup table. One dimension represents the severity of the defect, and the other dimension represents the degradation rate calculated based on historical data trends. Each cell in the matrix corresponds to a specific early warning level, and the final early warning level will be adjusted according to the preset importance level of the device being tested.

9. The ultrasonic live-line detection method for electrical equipment according to claim 2, characterized in that, The method further includes a feedback optimization step, which is performed after the situation assessment and early warning steps: when the early warning information triggers on-site maintenance, the actual results confirmed by the maintenance are compared with the diagnostic results, and the detection data and confirmation results are used as new training samples to incrementally learn the defect identification and classification model in order to achieve continuous optimization of the model's diagnostic capabilities.

10. The ultrasonic live-line detection method for electrical equipment according to claim 2, characterized in that, Before the signal acquisition and synchronization step, there is also a detection path planning step: the cloud platform calculates an optimal detection point access sequence based on the detection cycle of the device, the importance weight of the device, and the current location of the detection terminal, with the goal of minimizing the total path length and prioritizing the detection of key devices, and sends the sequence to the detection terminal to guide the detection execution. The method also includes a cross-domain knowledge transfer step for initializing the defect identification and classification model: first, the model is pre-trained using desensitized defect datasets from multiple different data sources, and then the model is fine-tuned using a private dataset from the current target region to improve the initial performance and generalization ability of the model when the target region has limited data.