Intelligent Recognition System and Method Based on Time-Frequency Feature Fusion and Adaptive Classification

Abstract

This invention discloses an intelligent identification system and method based on time-frequency feature fusion and adaptive classification, relating to the field of power equipment condition monitoring technology. It includes a UHF signal processing unit, a filtering and amplification unit, an analog-to-digital conversion unit, a digital signal preprocessing unit, a multi-dimensional feature extraction unit, a discharge phase analysis unit, a classification model optimization unit, an interference elimination unit, a defect type identification and diagnosis unit, and a discharge source location unit. The UHF signal processing unit collects and transmits weak electromagnetic signals from partial discharges, while the filtering and amplification unit filters out interference and amplifies the signal amplitude. This invention, through the UHF signal processing unit, filtering and amplification unit, and analog-to-digital conversion unit, accurately acquires weak electromagnetic signals, efficiently filters out interference, amplifies and converts signals, improving the accuracy and effectiveness of the detection process. Through pulse segmentation and normalization, multi-dimensional feature extraction, and classification model establishment and optimization, it reduces the risk of misjudgment.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, and in particular to an intelligent identification system and method based on time-frequency feature fusion and adaptive classification. Background Technology

[0002] In modern power systems, the safe and stable operation of power equipment is of paramount importance. Partial discharge (PD) is a key indicator reflecting the insulation status of power equipment, and its accurate detection and analysis are of great significance for ensuring the reliable operation of the power system. However, the detection and processing of PD signals face many challenges. When power equipment is running, the electromagnetic signals generated by PD are extremely weak and are often submerged in complex electromagnetic environmental noise. For example, in high-voltage substations and other locations, various electrical devices are densely packed, and the electromagnetic interference generated by different devices intertwines, making the effective extraction of partial discharge signals extremely difficult. Although the ultra-high frequency (UHF) band (300MHz-3GHz) can capture partial discharge signals, the signals are weak and require high-sensitivity sensors and low-loss transmission methods; otherwise, the signals are easily distorted or lost, leading to detection failure. In addition to the electromagnetic interference generated by the equipment itself, the power system is also affected by external communication signals and natural environmental noise. Communication signals and partial discharge signals may overlap in the spectrum, and power frequency interference can also seriously affect detection. Traditional detection methods are difficult to accurately identify partial discharge signals in such complex interference environments, and are prone to misjudgment and omission, affecting the accuracy of equipment condition assessment. Partial discharge signals contain rich information about the insulation status of equipment, but it is not easy to accurately extract and analyze these features. The characteristics of the signal in the time and frequency domains are complex and related to factors such as equipment type and operating conditions. For example, different types of power transformers exhibit varying partial discharge (PD) signal characteristics, and the PD signals of the same equipment also differ under different aging conditions. Therefore, multi-dimensional and refined feature extraction and analysis methods are needed to comprehensively and deeply mine signal characteristics to achieve accurate assessment of equipment insulation status. Accurately determining the type of PD signal and equipment defect is crucial for formulating reasonable maintenance strategies. However, due to the diversity and complexity of PD signals, existing classification models struggle to meet high-precision identification requirements. Different defect types may correspond to similar PD signal characteristics, easily leading to classification errors. Furthermore, with the expansion of power system scale and equipment upgrades, higher demands are placed on the adaptability and accuracy of classification models, requiring continuous optimization and improvement. Therefore, an intelligent identification system and method based on time-frequency feature fusion and adaptive classification is needed to address the aforementioned problems. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent recognition system and method based on time-frequency feature fusion and adaptive classification. This addresses the problems of difficulty in detecting weak signals, high requirements for signal feature analysis in complex interference environments, and urgent need for classification and recognition accuracy in the aforementioned technical solutions.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intelligent recognition system based on time-frequency feature fusion and adaptive classification, comprising a UHF signal processing unit, wherein the output of the UHF signal processing unit is connected to the input of a filtering and amplification unit, the output of the filtering and amplification unit is connected to the input of an analog-to-digital conversion unit, the output of the analog-to-digital conversion unit is connected to the input of a digital signal preprocessing unit, the output of the digital signal preprocessing unit is connected to the input of a multi-dimensional feature extraction unit, the output of the multi-dimensional feature extraction unit is connected to the input of a discharge phase analysis unit, the output of the discharge phase analysis unit is connected to the input of a classification model optimization unit, the output of the classification model optimization unit is connected to the input of an interference elimination unit, the output of the interference elimination unit is connected to the input of a defect type identification and diagnosis unit, and the output of the defect type identification and diagnosis unit is connected to the input of a discharge source location unit.

[0005] The ultra-high frequency signal processing unit is used to collect and transmit weak electromagnetic signals from partial discharge. The filtering and amplification unit is used to filter out interference and amplify the signal amplitude. The analog-to-digital conversion unit is used to convert analog electrical signals into digital signals according to the theorem. The digital signal preprocessing unit is used to denoise, smooth, and resample the signal. The multi-dimensional feature extraction unit is used to extract various features such as signal time and frequency. The discharge phase analysis unit is used to draw a spectrum to determine the severity of partial discharge. The classification model optimization unit is used to select the best parameters and fuse them to improve the recognition rate. The interference elimination unit is used to identify and suppress various interference signals. The defect type identification and diagnosis unit is used to identify the defect type from the input data. The discharge power source positioning unit is used to locate the position of the discharge power source according to the time difference.

[0006] Furthermore, the output of the digital signal preprocessing unit is also connected to the input of the anomaly judgment unit, and the output of the anomaly judgment unit is connected to the input of the discharge power source positioning unit. The anomaly judgment unit is used to judge the local anomaly of the digital signal output by the digital signal preprocessing unit and output the abnormal data to the discharge power source positioning unit to locate the position of the discharge power source according to the time difference.

[0007] Furthermore, the multi-dimensional feature extraction unit includes a time-frequency analysis module, a statistical feature calculation module, and a phase feature extraction module. The time-frequency analysis module is used to analyze the change of signal frequency over time, the statistical feature calculation module is used to calculate the mean, variance, and other quantized signals, and the phase feature extraction module is used to lock the phase and extract relevant features.

[0008] Furthermore, the classification model optimization unit includes a parameter tuning module and a model fusion module. The parameter tuning module is used to traverse parameter combinations to find the optimal parameters, and the model fusion module is used to combine multiple models to improve the recognition effect.

[0009] Furthermore, the interference elimination unit includes an interference identification module and an interference suppression module. The interference identification module is used to identify current signal interference using a classifier, and the interference suppression module is used to suppress various types of interference using different methods.

[0010] Furthermore, intelligent recognition methods based on time-frequency feature fusion and adaptive classification include the following specific recognition methods:

[0011] S1. Signal Acquisition: The UHF information processing unit uses a high-sensitivity UHF sensor with a working frequency of 300MHz-3GHz. Its built-in high-gain antenna or high-coupling-efficiency capacitor component accurately captures the weak electromagnetic signals generated by partial discharge in the space around the power equipment. The signal is then transmitted to the filtering and amplification unit through a low-loss coaxial cable for accurate acquisition and transmission, ensuring that the signal is not distorted.

[0012] S2. Filtering and Amplification: The filtering and amplification unit suppresses low-frequency interference below 300MHz and high-frequency noise above 3GHz, while effectively preserving the effective frequency band of the UHF partial discharge signal. Then, a low-noise amplifier is used to increase the signal amplitude. The noise figure of the amplifier is less than 2dB, and the signal amplification factor can be flexibly adjusted between 5 and 50 times. The effective frequency band of the partial discharge signal is output to the analog-to-digital conversion unit, making the filtering and amplification highly efficient and optimizing the overall signal quality.

[0013] S3. Analog-to-digital conversion: By employing a high-precision analog-to-digital converter with 12 bits or more, and setting the sampling frequency to above 5GHz according to the Nyquist sampling theorem, the effective frequency band of the filtered and amplified partial discharge signal is converted into a digital signal stored and transmitted in binary code form, with a quantization error of less than 0.1%. The digital signal is then transmitted to the digital signal preprocessing unit, achieving high-standard analog-to-digital conversion and reducing information loss.

[0014] S4. Pulse Segmentation and Normalization: In the digital signal preprocessing unit stage, a sliding window-based detection method is used, with the window width dynamically adjusted between 10 and 2000 ns. The pulse start point is determined by setting an adaptive threshold, and independent pulse sequences are segmented and extracted. Then, these sequences are normalized to unify the signal amplitude to the range of [-1, 1], ensuring the comparability of different pulse amplitudes. The sequences are then input into the multi-dimensional feature extraction unit for accurate segmentation and normalization, improving the accuracy of the analysis.

[0015] S5. Multi-dimensional Feature Extraction: Within the multi-dimensional feature extraction unit, the time-frequency analysis module stitches and fuses the frequency domain features obtained from multi-scale frequency domain analysis with the features aligned by dynamic time warping in the time domain; the statistical feature calculation module extracts the time domain features; the phase feature extraction module analyzes the relationship between the pulse and the power frequency phase and constructs a phase distribution model; simultaneously, the data extracted by the phase features is transmitted to the discharge phase analysis unit, which, based on the phase features, plots a discharge phase spectrum. The horizontal axis of the spectrum represents the power frequency phase, and the vertical axis represents the number of discharges or the discharge amount. The distribution pattern of discharge points on the spectrum is analyzed to determine the severity of partial discharge, extracting features from multiple dimensions to comprehensively characterize the signal.

[0016] S6. Classification Model Establishment and Optimization: The data after time-frequency analysis, statistical feature calculation and discharge phase analysis are transmitted to the classification model optimization unit. Through decision tree, random forest and neural network algorithm models, the parameters are optimized and adjusted based on a large amount of historical data. At the same time, the classification ability of the model for different types of pulses is evaluated, and the classification effect is adjusted according to the evaluation results to optimize the classification model and improve the classification accuracy.

[0017] S7. Interference signal removal: Data in the classification model optimization unit is transmitted to the interference judgment unit and compared with the partial discharge data under normal conditions. Data with range error within the specified threshold is marked as normal data. Clustering algorithm is used to determine the distribution range of normal data. When new data deviates from the range and exceeds a certain threshold, it is judged as abnormal data. At the same time, normal data and abnormal data are transmitted to the defect type identification and diagnosis unit to effectively remove interference and ensure data purity.

[0018] S8. Classification and Recognition: The defect type recognition and diagnosis unit uses a neural network algorithm model to extract features from the convolutional layer and classify the data in the fully connected layer of abnormal data, and outputs the defect type. For normal data, the forest algorithm model is used to calculate and analyze the partial discharge signal probability, determine the pulse type, accurately classify and identify, and clarify the equipment status.

[0019] S9. Result Output and Alarm: After normal data is analyzed by the defect type identification and diagnosis unit, the pulse type and confidence level are displayed in real time. The confidence level is displayed to two decimal places. A configurable alarm threshold can be set. When the probability of the faulty discharge signal reaches or exceeds the threshold, an alarm is triggered. At the same time, the data that triggers the alarm threshold is input into the discharge power supply location unit for discharge power supply location. Abnormal data is directly processed by the discharge power supply location unit for discharge power supply location. The alarm is output intuitively to ensure system safety.

[0020] Furthermore, the S5 multi-dimensional feature extraction step includes time-domain feature extraction and frequency-domain feature extraction. The time-domain feature extraction calculates the pulse rise time, the number of peaks, and the waveform symmetry to quantitatively describe the characteristics of the pulse in the time domain. The frequency-domain feature extraction uses a wavelet scattering network to perform multi-scale frequency domain analysis to obtain frequency domain features. After dimensionality reduction, these features are spliced ​​and fused with the time-domain features aligned by dynamic time warping (DTW). During the dimensionality reduction process, more than 95% of the key information is retained. The time-frequency fusion feature extraction provides accurate and comprehensive analysis of partial discharge, offering precise data for partial discharge analysis.

[0021] Furthermore, the S7 interference signal removal step includes phase distribution clustering and dynamic amplitude thresholding. The phase distribution clustering constructs a phase distribution model by analyzing the relationship between the pulse and the power frequency phase. If the pulse phase concentration is greater than 90%, it is determined to be communication interference and removed. The dynamic amplitude thresholding removes abnormal pulses with amplitudes exceeding the average value ±3σ by calculating the average value and standard deviation (σ) of the signal amplitude. Interference is removed based on phase amplitude, improving the accuracy and stability of diagnosis. The phase amplitude thresholding effectively removes interference signals.

[0022] In summary, this invention provides an intelligent identification method and system for ultra-high frequency partial discharge pulses based on time-frequency feature fusion and adaptive classification, which has the following beneficial effects:

[0023] 1. Through the ultra-high frequency signal processing unit, filtering and amplification unit and analog-to-digital conversion unit, weak electromagnetic signals are accurately acquired, interference is efficiently filtered out, and the signals are amplified and converted, providing a high-precision, low-error data foundation for subsequent processing and improving the accuracy and effectiveness of the detection process.

[0024] 2. Through the time-frequency analysis module, statistical feature calculation module, and phase feature extraction module in the multi-dimensional feature extraction unit, the characteristics of partial discharge signals are deeply mined from multiple perspectives in the time-frequency domain, a comprehensive feature vector is constructed, the severity of partial discharge is accurately determined, and rich and accurate data support is provided for subsequent analysis.

[0025] 3. By segmenting and normalizing pulses, extracting multi-dimensional features, and establishing and optimizing classification models, pulses are accurately segmented and normalized, signal features are comprehensively mined and fused, and the optimized model improves classification accuracy, providing a reliable basis for equipment status assessment and reducing the risk of misjudgment.

[0026] 4. Through interference signal elimination, classification and identification and discharge source location unit, interference signals are effectively identified and eliminated, defects are accurately classified and pulse types are judged, discharge sources are quickly located, the stability of the detection system is ensured, the risk of equipment failure is reduced, and the safe and stable operation of the power system is contributed. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the architecture of the intelligent recognition system based on time-frequency feature fusion and adaptive classification according to the present invention;

[0028] Figure 2 This is a schematic diagram of the multi-dimensional feature extraction unit architecture of the intelligent recognition system based on time-frequency feature fusion and adaptive classification according to the present invention;

[0029] Figure 3 This is a schematic diagram of the classification model optimization unit architecture of the intelligent recognition system based on time-frequency feature fusion and adaptive classification of the present invention;

[0030] Figure 4 This is a schematic diagram of the interference elimination unit architecture of the intelligent recognition system based on time-frequency feature fusion and adaptive classification according to the present invention;

[0031] Figure 5 This is a schematic diagram of the process architecture of the intelligent recognition method based on time-frequency feature fusion and adaptive classification of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Example:

[0034] Please see Figure 1 - Figure 4As shown, the present invention provides a technical solution: an intelligent recognition system based on time-frequency feature fusion and adaptive classification, comprising a UHF signal processing unit, the output of which is connected to the input of a filtering and amplification unit, the output of which is connected to the input of an analog-to-digital conversion unit, the output of which is connected to the input of a digital signal preprocessing unit, the output of which is connected to the input of a multi-dimensional feature extraction unit, the output of which is connected to the input of a discharge phase analysis unit, the output of which is connected to the input of a classification model optimization unit, the output of which is connected to the input of an interference elimination unit, the output of which is connected to the input of a defect type identification and diagnosis unit, and the output of which is connected to the input of a discharge source location unit.

[0035] The ultra-high frequency signal processing unit is used to collect and transmit weak electromagnetic signals from partial discharge. The filtering and amplification unit is used to filter out interference and amplify the signal amplitude. The analog-to-digital conversion unit is used to convert analog electrical signals into digital signals according to the theorem. The digital signal preprocessing unit is used to denoise, smooth, and resample the signal. The multi-dimensional feature extraction unit is used to extract various features such as signal time and frequency. The discharge phase analysis unit is used to draw a spectrum to determine the severity of partial discharge. The classification model optimization unit is used to select the best parameters and fuse them to improve the recognition rate. The interference elimination unit is used to identify and suppress various interference signals. The defect type identification and diagnosis unit is used to identify the defect type from the input data. The discharge power supply location unit is used to locate the position of the discharge power supply according to the time difference.

[0036] Please see Figure 1 As shown, the output of the digital signal preprocessing unit is also connected to the input of the anomaly judgment unit, and the output of the anomaly judgment unit is connected to the input of the discharge power supply positioning unit. The anomaly judgment unit is used to judge the local anomaly of the digital signal output by the digital signal preprocessing unit and output the anomaly data to the discharge power supply positioning unit to locate the position of the discharge power supply according to the time difference.

[0037] Please see Figure 2 As shown, the multi-dimensional feature extraction unit includes a time-frequency analysis module, a statistical feature calculation module, and a phase feature extraction module. The time-frequency analysis module is used to analyze the change of signal frequency over time, the statistical feature calculation module is used to calculate the mean, variance, and other quantized signals, and the phase feature extraction module is used to lock the phase and extract relevant features.

[0038] Please see Figure 3 As shown, the classification model optimization unit includes a parameter tuning module and a model fusion module. The parameter tuning module is used to traverse parameter combinations to find the best parameters, and the model fusion module is used to combine multiple models to improve the recognition effect.

[0039] Please see Figure 4 As shown, the interference elimination unit includes an interference identification module and an interference suppression module. The interference identification module is used to identify the current signal interference using a classifier, and the interference suppression module is used to suppress various types of interference using different methods.

[0040] Please see Figure 1 - Figure 4 As shown, the intelligent recognition method based on time-frequency feature fusion and adaptive classification includes the following specific steps:

[0041] S1. Signal Acquisition: The UHF information processing unit utilizes a high-sensitivity UHF sensor operating in the 300MHz-3GHz frequency band. Its built-in high-gain antenna or high-coupling-efficiency capacitor accurately captures the weak electromagnetic signals generated by partial discharge in the surrounding space of the power equipment. The signal is then transmitted to the filtering and amplification unit via a low-loss coaxial cable. This ensures that the extremely weak electromagnetic signals generated by partial discharge of the power equipment are accurately captured, with picometer-level capture accuracy that does not miss any minute changes. The low-loss coaxial cable has a signal attenuation rate of less than 5%, ensuring that the signal is not distorted during transmission. This provides a true and reliable original signal for subsequent filtering, amplification, and other processing, greatly improving the accuracy and effectiveness of the entire detection process.

[0042] S2. Filtering and Amplification: The filtering and amplification unit suppresses low-frequency interference below 300MHz and high-frequency noise above 3GHz, effectively preserving the effective frequency band of the UHF partial discharge signal. A low-noise amplifier then boosts the signal amplitude. The amplifier's noise figure is less than 2dB, and the signal amplification factor can be flexibly adjusted between 5 and 50 times. The effective frequency band of the partial discharge signal is output to the analog-to-digital conversion unit. The bandpass filter's strong suppression capability of low-frequency and high-frequency interference (suppression capability can reach over 60dB) effectively removes noise interference. The low-noise amplifier maintains low noise while boosting the signal amplitude, ensuring high signal quality output to the analog-to-digital conversion unit. Simultaneously, the signal amplification factor can be flexibly adjusted to adapt to raw signals of different intensities, laying the foundation for accurate analog-to-digital conversion and ensuring more precise data processing in subsequent steps.

[0043] S3. Analog-to-Digital Conversion: By employing a high-precision analog-to-digital converter (ADC) with 12 bits or more, and setting the sampling frequency to above 5 GHz based on the Nyquist sampling theorem, the effective frequency band of the filtered and amplified partial discharge signal is converted into a digital signal stored and transmitted in binary code form. The quantization error is less than 0.1%, and the digital signal is transmitted to the digital signal preprocessing unit. The high-resolution ADC and high sampling frequency enable the details of the analog signal to be accurately restored with minimal quantization error, minimizing information loss and providing a high-precision data foundation for subsequent digital signal preprocessing. This allows subsequent pulse segmentation, feature extraction, and other operations to be performed based on more accurate data, thereby improving the reliability and detection accuracy of the entire detection system.

[0044] S4. Pulse Segmentation and Normalization: In the digital signal preprocessing unit stage, a sliding window-based detection method is used, with the window width dynamically adjusted between 10 and 2000 ns. The pulse start point is determined by setting an adaptive threshold, and independent pulse sequences are segmented and extracted. These sequences are then normalized to unify the signal amplitude to the [-1, 1] interval, ensuring the comparability of different pulse amplitudes. The sequences are then input into the multi-dimensional feature extraction unit. The dynamically adjusted sliding window and adaptive threshold can accurately segment pulses, effectively avoiding analysis errors caused by pulse overlap. The normalization process unifies pulses of different amplitudes to the same interval, making the features of different pulses comparable during subsequent multi-dimensional feature extraction. This improves the accuracy and stability of feature extraction and provides strong support for accurate analysis of partial discharge signals.

[0045] S5. Multi-dimensional Feature Extraction: Within the multi-dimensional feature extraction unit, the time-frequency analysis module stitches and fuses the frequency domain features obtained from multi-scale frequency domain analysis with the features aligned by dynamic time warping in the time domain; the statistical feature calculation module extracts time domain features; the phase feature extraction module analyzes the relationship between the pulse and the power frequency phase and constructs a phase distribution model; simultaneously, the data extracted by the phase feature is transmitted to the discharge phase analysis unit, where a discharge phase spectrum is drawn based on the phase features. The horizontal axis of the spectrum represents the power frequency phase, and the vertical axis represents the number of discharges or the discharge amount. The distribution pattern of discharge points on the spectrum is analyzed to determine the severity of partial discharge. Multi-dimensional feature extraction comprehensively and deeply mines signal features, and the fused feature vector can more accurately represent the partial discharge signal. The fusion of time domain and frequency domain features, as well as phase feature analysis and discharge phase spectrum drawing, provide detailed information about the partial discharge signal from multiple perspectives, which helps to more accurately determine the severity of partial discharge and provides richer and more accurate data support for subsequent classification model establishment and equipment status assessment.

[0046] S6. Classification Model Establishment and Optimization: The data after time-frequency analysis, statistical feature calculation, and discharge phase analysis are transmitted to the classification model optimization unit. Parameters are optimized and adjusted based on a large amount of historical data using decision tree, random forest, and neural network algorithms. Simultaneously, the model's classification ability for different types of pulses is evaluated, and the classification effect is adjusted according to the evaluation results. Optimizing model parameters using a large amount of historical data enables the decision tree, random forest, and neural network algorithms to better learn the characteristic patterns of different pulses such as partial discharge signals and noise. Through continuous evaluation and adjustment of the classification effect, the model's accuracy in classifying different types of pulses is improved, providing a reliable classification basis for accurate subsequent judgment of equipment status and reducing the risk of misjudgment.

[0047] S7. Interference Signal Removal: Data from the classification model optimization unit is transmitted to the interference judgment unit and compared with partial discharge data under normal conditions. Data with a range error within a specified threshold is marked as normal data, and a clustering algorithm is used to determine the distribution range of normal data. When new data deviates from this range by more than a certain threshold, it is judged as abnormal data. Simultaneously, both normal and abnormal data are transmitted to the defect type identification and diagnosis unit. By comparing with normal data and using a clustering algorithm, the normal range is determined, effectively identifying interference signals. Accurate removal of interference signals significantly improves the purity of the data input to the defect type identification and diagnosis unit, reduces the impact of interference on the diagnostic results, and makes subsequent defect type identification and diagnosis more accurate and reliable, thus improving the stability and reliability of the entire detection system.

[0048] S8. Classification and Recognition: The defect type recognition and diagnosis unit uses a neural network algorithm model to extract features from convolutional layers and classify abnormal data using fully connected layers, outputting the defect type. For normal data, it uses a forest algorithm model to calculate and analyze the partial discharge signal probability and determine the pulse type. Different algorithm models are used to process abnormal and normal data, fully leveraging the advantages of neural network algorithms in feature extraction and classification, and the ability of forest algorithms to determine the probability of partial discharge signals. This accurately outputs the defect type and determines the pulse type, providing equipment maintenance personnel with clear and concise equipment status information, helping them to understand the equipment situation in a timely manner and take appropriate measures.

[0049] S9. Result Output and Alarm: After normal data is analyzed by the defect type identification and diagnosis unit, the pulse type and confidence level are displayed in real time. The confidence level is displayed to two decimal places. Configurable alarm thresholds are set. When the probability of a discharge signal reaches or exceeds the threshold, an alarm is triggered. Simultaneously, the data triggering the alarm threshold is input into the discharge power source location unit for discharge power source location. Abnormal data is directly processed by the discharge power source location unit for discharge power source location. The pulse type and confidence level are displayed in real time, intuitively showing the test results to the user, allowing the user to quickly understand the equipment status. The configurable alarm thresholds can be flexibly set according to actual needs, triggering alarms in a timely manner, helping maintenance personnel to quickly respond to potential faults. Discharge power source location based on alarm-triggered data helps maintenance personnel quickly locate the fault point, take targeted measures, reduce equipment failure risks, and ensure the safe and stable operation of the power system.

[0050] The S5 multi-dimensional feature extraction process includes time-domain feature extraction and frequency-domain feature extraction. Time-domain feature extraction calculates the pulse rise time, number of peaks, and waveform symmetry to quantitatively describe the pulse's characteristics in the time domain. Frequency-domain feature extraction uses wavelet scattering networks for multi-scale frequency domain analysis to obtain frequency domain features. After dimensionality reduction, these features are then stitched together with the time-domain features aligned using Dynamic Time Warping (DTW). During dimensionality reduction, over 95% of the key information is retained, simplifying the data while preserving important features. The remaining features are then stitched together with the time-domain features after DTW alignment, presenting the partial discharge signal comprehensively from multiple perspectives. This provides a comprehensive and accurate data foundation for subsequent accurate assessment of partial discharge severity, establishment of classification models, and equipment status evaluation, greatly improving the reliability and accuracy of the analysis.

[0051] The S7 interference signal removal step includes phase distribution clustering and dynamic amplitude thresholding. Phase distribution clustering analyzes the relationship between pulses and power frequency phases to construct a phase distribution model. If the pulse phase concentration is greater than 90%, it is identified as communication interference and removed. Dynamic amplitude thresholding calculates the average signal amplitude and standard deviation (σ) to remove abnormal pulses with amplitudes exceeding the average ±3σ. This significantly improves the purity of the data input to the defect type identification and diagnosis unit, reduces interference on the diagnostic results, and makes subsequent defect type identification and diagnosis more accurate and reliable. It ensures the stable operation of the detection system and provides strong support for equipment fault diagnosis.

[0052] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any person skilled in the art may utilize the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes for application in other fields. However, any simple modification, equivalent change, and modification of the above embodiments made in accordance with the technical essence of the present invention without departing from the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. An intelligent recognition system based on time-frequency feature fusion and adaptive classification, comprising a UHF signal processing unit, characterized in that: The output of the ultra-high frequency signal processing unit is connected to the input of the filtering and amplification unit. The output of the filtering and amplification unit is connected to the input of the analog-to-digital conversion unit. The output of the analog-to-digital conversion unit is connected to the input of the digital signal preprocessing unit. The output of the digital signal preprocessing unit is connected to the input of the multi-dimensional feature extraction unit. The output of the multi-dimensional feature extraction unit is connected to the input of the discharge phase analysis unit. The output of the discharge phase analysis unit is connected to the input of the classification model optimization unit. The output of the classification model optimization unit is connected to the input of the interference elimination unit. The output of the interference elimination unit is connected to the input of the defect type identification and diagnosis unit. The output of the defect type identification and diagnosis unit is connected to the input of the discharge source positioning unit. The ultra-high frequency signal processing unit is used to collect and transmit weak electromagnetic signals from partial discharge. The filtering and amplification unit is used to filter out interference and amplify the signal amplitude. The analog-to-digital conversion unit is used to convert analog electrical signals into digital signals according to the theorem. The digital signal preprocessing unit is used to denoise, smooth, and resample the signal. The multi-dimensional feature extraction unit is used to extract multiple time-frequency features of the signal. The discharge phase analysis unit is used to draw a spectrum to determine the severity of partial discharge. The classification model optimization unit is used to select the best parameters and fuse them to improve the recognition rate. The interference elimination unit is used to identify and suppress various interference signals. The defect type identification and diagnosis unit is used to identify the defect type from the input data. The discharge power source positioning unit is used to locate the position of the discharge power source according to the time difference.

2. The intelligent recognition system based on time-frequency feature fusion and adaptive classification according to claim 1, characterized in that: The output of the digital signal preprocessing unit is also connected to the input of the anomaly judgment unit. The output of the anomaly judgment unit is connected to the input of the discharge power source positioning unit. The anomaly judgment unit is used to judge the local anomaly of the digital signal output by the digital signal preprocessing unit and output the abnormal data to the discharge power source positioning unit to locate the position of the discharge power source according to the time difference.

3. The intelligent recognition system based on time-frequency feature fusion and adaptive classification according to claim 1, characterized in that: The multi-dimensional feature extraction unit includes a time-frequency analysis module, a statistical feature calculation module, and a phase feature extraction module. The time-frequency analysis module is used to analyze the change of signal frequency over time. The statistical feature calculation module is used to calculate the mean-variance quantized signal. The phase feature extraction module is used to lock the phase and extract relevant features.

4. The intelligent recognition system based on time-frequency feature fusion and adaptive classification according to claim 1, characterized in that: The classification model optimization unit includes a parameter tuning module and a model fusion module. The parameter tuning module is used to traverse parameter combinations to find the optimal parameters, and the model fusion module is used to combine multiple models to improve the recognition effect.

5. The intelligent recognition system based on time-frequency feature fusion and adaptive classification according to claim 1, characterized in that: The interference elimination unit includes an interference identification module and an interference suppression module. The interference identification module is used to identify the current signal interference using a classifier, and the interference suppression module is used to suppress various types of interference using different methods.

6. An intelligent recognition method based on time-frequency feature fusion and adaptive classification, employing the intelligent recognition system based on time-frequency feature fusion and adaptive classification as described in any one of claims 1-5, characterized in that: The specific steps include the following: S1. Signal Acquisition: The UHF information processing unit uses a high-sensitivity UHF sensor with a working frequency of 300MHz-3GHz. Its built-in high-gain antenna or high-coupling-efficiency capacitor component accurately captures the weak electromagnetic signals generated by partial discharge in the space around the power equipment, and transmits the signal to the filtering and amplification unit through a low-loss coaxial cable. S2. Filtering and Amplification: The filtering and amplification unit suppresses low-frequency interference below 300MHz and high-frequency noise above 3GHz, while effectively preserving the effective frequency band of the UHF partial discharge signal. The signal amplitude is then boosted by a low-noise amplifier with a noise figure of less than 2dB and a signal amplification factor that can be flexibly adjusted between 5 and 50 times. The effective frequency band of the partial discharge signal is then output to the analog-to-digital conversion unit. S3. Analog-to-digital conversion: By using a high-precision analog-to-digital converter with 12 bits or more, and setting the sampling frequency to above 5GHz according to the Nyquist sampling theorem, the effective frequency band of the filtered and amplified partial discharge signal is converted into a digital signal stored and transmitted in binary code form, with a quantization error of less than 0.1%, and the digital signal is transmitted to the digital signal preprocessing unit. S4. Pulse Segmentation and Normalization: In the digital signal preprocessing unit stage, a sliding window-based detection method is used, with the window width dynamically adjusted between 10 and 2000 ns. The pulse start point is determined by setting an adaptive threshold, and independent pulse sequences are segmented and extracted. Then, these sequences are normalized to unify the signal amplitude to the range of [-1, 1], ensuring that different pulse amplitudes are comparable. The sequences are then input into the multi-dimensional feature extraction unit. S5. Multi-dimensional feature extraction: Within the multi-dimensional feature extraction unit, the time-frequency analysis module splices and fuses the frequency domain features obtained from multi-scale frequency domain analysis with the features aligned by dynamic time warping in the time domain. The statistical feature calculation module extracts time-domain features; The phase feature extraction module analyzes the relationship between the pulse and the power frequency phase and constructs a phase distribution model. At the same time, the data of the phase feature extraction is transmitted to the discharge phase analysis unit. Based on the phase features, a discharge phase spectrum is drawn. The horizontal axis of the spectrum is the power frequency phase, and the vertical axis is the number of discharges or the discharge amount. The distribution pattern of discharge points on the spectrum is analyzed to determine the severity of partial discharge. S6. Classification Model Establishment and Optimization: The data after time-frequency analysis, statistical feature calculation and discharge phase analysis are transmitted to the classification model optimization unit. The parameters are optimized and adjusted based on a large amount of historical data through decision tree, random forest and neural network algorithm models. At the same time, the classification ability of the model for different types of pulses is evaluated and the classification effect is adjusted according to the evaluation results. S7. Interference signal removal: Data in the classification model optimization unit is transmitted to the interference judgment unit and compared with the partial discharge data under normal conditions. Data with range error within the specified threshold is marked as normal data. Clustering algorithm is used to determine the distribution range of normal data. When new data deviates from the range and exceeds a certain threshold, it is judged as abnormal data. At the same time, normal data and abnormal data are transmitted to the defect type identification and diagnosis unit. S8. Classification and Recognition: The defect type recognition and diagnosis unit uses a neural network algorithm model to extract features from the convolutional layer and classify the abnormal data using the fully connected layer, and outputs the defect type. For normal data, it uses a forest algorithm model to calculate and analyze the partial discharge signal probability and determine the pulse type. S9. Result Output and Alarm: After normal data is analyzed by the defect type identification and diagnosis unit, the pulse type and confidence level are displayed in real time. The confidence level is displayed to two decimal places. A configurable alarm threshold can be set. When the probability of the discharge signal reaches or exceeds the threshold, an alarm is triggered. At the same time, the data that triggers the alarm threshold is input into the discharge power supply positioning unit for discharge power supply positioning. Abnormal data is directly located in the discharge power supply positioning unit.

7. The intelligent recognition method based on time-frequency feature fusion and adaptive classification according to claim 6, characterized in that: The S5 multi-dimensional feature extraction step includes time-domain feature extraction and frequency-domain feature extraction. The time-domain feature extraction calculates the time-domain features of pulse rise time, number of peaks, and waveform symmetry to quantify the characteristics of the pulse in the time domain. The frequency-domain feature extraction uses a wavelet scattering network to perform multi-scale frequency domain analysis to obtain frequency-domain features. After dimensionality reduction, these features are spliced ​​and fused with the time-domain features aligned by dynamic time warping (DTW). During the dimensionality reduction process, more than 95% of the key information is retained.

8. The intelligent recognition method based on time-frequency feature fusion and adaptive classification according to claim 7, characterized in that: The S7 interference signal removal step includes phase distribution clustering and dynamic amplitude thresholding. The phase distribution clustering constructs a phase distribution model by analyzing the relationship between the pulse and the power frequency phase. If the pulse phase concentration is greater than 90%, it is determined to be communication interference and removed. The dynamic amplitude thresholding removes abnormal pulses with amplitudes exceeding the average value ±3σ by calculating the average value and standard deviation (σ) of the signal amplitude.

Images