Ultrasonic detection system for transformer fault

Abstract

The invention discloses a transformer fault ultrasonic detection system, which comprises an ultrasonic sensor module, a signal conditioning module, a data acquisition and processing module, a data fusion module, a learning module, a diagnosis decision module and a human-computer interaction interface, according to the ultrasonic sensor module, 6-8 ultrasonic sensors are arranged in a transformer shell in an array mode, piezoelectric ceramic sensors with the resonant frequency of 40 kHZ are adopted as the ultrasonic sensors, the sensitivity of the ultrasonic sensors is-65 dBV / uBar, and effective capture is weak. According to the ultrasonic detection system for the transformer fault, single ultrasonic detection is easily influenced by accidental interference, the system is synchronously connected into equipment such as an infrared thermal imager and a vibration accelerometer, and a multi-physical-quantity correlation analysis model is established. When high-frequency ultrasonic pulses are detected, the system automatically calls temperature data of the corresponding position, and if the temperature gradient exceeds 2 DEG C / cm, it is confirmed that discharging heating is conducted; and if the temperature rise does not exist and the low-frequency vibration of 10-200Hz is accompanied, the mechanical looseness is judged.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer detection, and particularly to an ultrasonic detection system for transformer faults. Background Art

[0002] A transformer is an important device used in the power system, mainly for changing the voltage of alternating current. An ultrasonic detection system for transformer faults is a device that uses ultrasonic technology to monitor the internal and external states of a transformer, which can help detect potential faults early, thereby improving the safety and reliability of the transformer. When ultrasonic waves encounter bubbles, cracks, insulation material defects or mechanical structure problems inside the transformer, the propagation of the waves will be affected, resulting in changes in the reflected waves.

[0003] In the prior art, a Chinese patent with the application number CN202111298479.8 discloses a transformer detection system based on ultrasonic detection, including a processing module for performing comparative processing of data; an ultrasonic detection module connected to the processing module, where the ultrasonic detection module is distributed at each signal acquisition point on the transformer for collecting internal ultrasonic signals of the transformer; a storage module connected to the processing module for storing reference values; and a communication module connected to the processing module for performing remote transmission of data. By using an ultrasonic receiving device to receive the ultrasonic waves emitted during the operation of the transformer and analyzing the ultrasonic waves to judge the working state of the transformer, it can perform real-time detection of transformer discharge faults.

[0004] Again, in the prior art, a Chinese patent with the application number CN202210892568.3 discloses a transformer partial discharge detection system and method, including an ultrasonic sensing module, a moving component, and a positioning device. The moving component is configured to adjust the relative position between the ultrasonic sensing module and the target transformer to achieve multi-point detection of the ultrasonic signals generated by the partial discharge of the target transformer. The positioning device is used to locate the actual position of the partial discharge of the target transformer according to the ultrasonic signals generated by the partial discharge of the target transformer obtained from multi-point detection. By using the ultrasonic signals of multiple detection points to locate the actual position of the partial discharge of the transformer, it can overcome the problem of inaccurate positioning that may occur when using the ultrasonic signals of a single detection point to locate the transformer.

[0005] For another example, in the prior art, a Chinese patent with the application number CN201610302660.4 discloses a detection system for local discharge location of a transformer, which includes a Rogowski coil radio frequency sensor, a filter amplifier, an ultrasonic phased array sensor, a charge amplifier, a synchronous collector, and a computer. The Rogowski coil radio frequency sensor sends the detected radio frequency current to the filter amplifier for filtering and amplification, and then sends it to the synchronous collector. Each element of the ultrasonic phased array sensor sends the detected ultrasonic signal to the charge amplifier for amplification, and then sends it to the synchronous collector. The synchronous collector sends the received radio frequency current signal after filtering and amplification and the amplified ultrasonic signal to the computer, and the computer locates the local discharge source of the transformer.

[0006] Combined with the above materials, it can be seen that the ultrasonic detection systems in the prior art generally detect local discharge to achieve the purpose of fault detection. However, in the actual use process, there are many fault situations in the transformer, and a single detection direction cannot achieve the purpose of comprehensive detection. Summary of the Invention

[0007] The purpose of the present invention is to provide an ultrasonic detection system for transformer faults to solve the problem of incomplete detection range proposed in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solution: An ultrasonic detection system for transformer faults, including an ultrasonic sensor module, a signal conditioning module, a data acquisition and processing module, a data fusion module, a learning module, a diagnostic decision module, and a human-machine interface. The ultrasonic sensor module consists of 6-8 ultrasonic sensors arranged in an array inside the transformer housing. The ultrasonic sensors use piezoelectric ceramic sensors with a resonant frequency of 40 kHz, and the sensitivity of the ultrasonic sensors is -65 dBV / uBar, effectively capturing weak signals. The signal conditioning module filters, amplifies the weak signals output by the ultrasonic sensors, and converts them into standardized signals suitable for digital acquisition. The hardware platform of the data acquisition and processing module is Xilinx Zynq-7020 SoC, integrating FPGA and ARM Cortex-A9 to achieve high-speed signal processing and flexible control. The data fusion module receives ultrasonic features, infrared temperature, and vibration acceleration from the sensors. The ultrasonic features include amplitude, frequency, and phase. The accuracy of the infrared temperature is ±1°C, and the resolution is 0.1°C. The frequency of the vibration acceleration is 0.1-10 kHz, and the dynamic range is ±50 g. The diagnostic decision module includes threshold judgment and severity assessment. The threshold judgment is adjusted according to the dynamic baseline, and the background noise level is updated every 24 hours. The severity assessment is divided into four levels: normal, attention, abnormal, and critical based on the pulse density and energy growth rate. The model structure of the learning module includes a CNN branch, an LSTM branch, and a fusion layer. The CNN branch is used to process time-frequency images, the LSTM branch is used to process time-series features, and the fusion layer is used to splice the outputs of the CNN branch and the LSTM branch and then classify them through a fully connected layer.

[0009] Preferably, the ultrasonic sensors are distributed in an equiangular ring array, and the sensors at the top and bottom are tilted 45° to achieve three-dimensional sound field coverage.

[0010] Preferably, the distance between adjacent ultrasonic sensors is ≥30 cm to avoid acoustic interference, and the ultrasonic sensors are provided with a μ-Metal alloy housing with a permeability >50,000 to suppress strong electromagnetic field interference of the transformer.

[0011] Preferably, the circuit composition of the signal conditioning module includes high-pass filtering, pre-amplification, and band-pass filtering. High-pass filtering: Remove low-frequency vibration noise of the transformer, use a second-order Butterworth filter with a roll-off slope of -40 dB / dec. Pre-amplification: Use a low-noise instrumentation amplifier with an input noise density of 1.1 nV / √Hz and a gain error <0.01%. Band-pass filtering: Use a switched-capacitor filter with a Q value of 5 to suppress high-frequency switching noise and out-of-band interference.

[0012] Preferably, the processing flow of the data acquisition and processing module is as follows: Adaptive noise cancellation: The LMS algorithm is used to cancel environmental noise in real time; Time-frequency analysis: Short-time Fourier transform, Hanning window, window length 512 points, frequency resolution 97.6 Hz; Feature extraction: The extracted features are peak frequency, pulse repetition rate, signal energy entropy, and phase difference positioning. Among them, the peak frequency is the frequency component with the largest amplitude detected in the spectrum, and the pulse repetition rate is calculated as the reciprocal of the pulse interval time through zero-crossing detection.

[0013] Preferably, the fusion method of the data fusion module is based on Dempster-Shafer evidence theory to solve the problem of uncertain information fusion, including: Basic probability assignment: Assign confidence levels to the fault hypotheses for each sensor data; Decision output: Select the fault type with the highest confidence level.

[0014] Preferably, the training data of the learning module comes from more than 2000 sets of on-site data augmentation, and the accuracy of the test set reaches 96.2%. Among them, the data augmentation includes adding -10dB noise, time shift, and frequency shift.

[0015] Preferably, a fault feature library is set between the data acquisition and processing module and the machine learning module. The construction method of the fault feature library is based on historical fault cases and laboratory simulation data. The fault types in the fault feature library include: Partial discharge: Main frequency, PRF, pulse rise time Core looseness: Low-frequency harmonic components, vibration correlation Winding deformation: High-frequency burst pulses, energy entropy change rate Air bubbles in oil: Wideband energy distribution, no significant periodicity.

[0016] Preferably, the training strategy of the learning module is as follows: Input data: Time-frequency diagram, generated by STFT, window function: Kaiser β = 6; Data augmentation: Add -20dB Gaussian noise, time shift jitter, frequency band masking; Transfer learning: Pre-train on the CWRU bearing dataset, and freeze the first 3 layers during fine-tuning.

[0017] Preferably, the functional modules of the human-computer interaction interface include: Real-time monitoring: The three-dimensional sound pressure cloud map shows the sound source position; Historical analysis: Support the playback of time-frequency diagrams and feature trend curves; Alarm management: Hierarchical alarm, push SMS and email notifications; Report generation: Automatically output a PDF diagnostic report, including failure probability and maintenance suggestions.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: The ultrasonic detection system for transformer faults adopts a novel structural design, and the specific content is as follows: 1. Multi-source data fusion: Single ultrasonic detection is vulnerable to accidental interference. This system synchronously accesses devices such as infrared thermal imagers and vibration accelerometers to establish a multi-physical quantity correlation analysis model. When high-frequency ultrasonic pulses are detected, the system automatically retrieves the temperature data at the corresponding position. If the temperature gradient exceeds 2°C / cm, it is confirmed as discharge heating; if there is no temperature rise but accompanied by low-frequency vibration of 10 - 200 Hz, it is determined as mechanical looseness.

[0019] 2. Intelligent diagnostic algorithm: The traditional threshold alarm method is vulnerable to operating condition fluctuations. This system innovatively integrates convolutional neural network (CNN) and long short-term memory network (LSTM) to construct a hybrid diagnostic model. The CNN branch focuses on analyzing the spatial features in the ultrasonic time-frequency image, such as the pulse cluster morphology of partial discharge; the LSTM branch tracks the change trends of time series parameters such as pulse repetition rate and energy entropy. After the dual-branch features are fused at the decision layer, the recognition accuracy of the model for complex faults such as winding deformation and core looseness is increased to 96.7%, which is 12% higher than that of a single model.

[0020] 3. Strong anti-interference ability: The biggest challenge faced by ultrasonic detection is the strong electromagnetic interference and complex background noise in the transformer operating environment. This system ensures stable operation in high-interference scenarios such as high-voltage substations through the dual guarantee of differential signal transmission architecture and adaptive noise reduction algorithm. The differential transmission adopts a twisted pair shielding structure to effectively suppress common-mode interference. The measured signal distortion degree is less than 0.1% in a strong electric field of 30 V / m. Combined with wavelet noise reduction technology, the system can automatically identify and filter out background interferences such as oil pump vibration and fan noise, and the signal-to-noise ratio is increased by more than 16 dB.

[0021] 4. Economic advantages: Compared with traditional oil chromatography analysis, the single detection cost of this system is reduced by 83%, and there is no need to take oil samples regularly, reducing the frequency of manual inspections by more than 50%. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is the overall process system block diagram of the present invention; Figure 2 is the system diagram of the sensor array topology optimization of the present invention; Figure 3 is the system diagram of the analog signal link design of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0023] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0024] Embodiment 1: Please refer to Figures 1 - 3 , the present invention provides a technical solution: an ultrasonic detection system for transformer faults, including an ultrasonic sensor module, a signal conditioning module, a data acquisition and processing module, a data fusion module, a learning module, a diagnostic decision module, and a human-machine interface; the ultrasonic sensor module consists of 6-8 ultrasonic sensors installed in an array inside the transformer shell, and the ultrasonic sensors use piezoelectric ceramic sensors with a resonant frequency of 40 kHz. The sensitivity of the ultrasonic sensors is -65 dBV / uBar, effectively capturing weak signals; the signal conditioning module filters, amplifies the weak signals output by the ultrasonic sensors, and converts them into standardized signals suitable for digital acquisition; the hardware platform of the data acquisition and processing module is Xilinx Zynq-7020 SoC, integrating FPGA and ARM Cortex-A9 to achieve high-speed signal processing and flexible control; the data fusion module receives ultrasonic characteristics, infrared temperature, and vibration acceleration from the sensors. Among them, the ultrasonic characteristics include amplitude, frequency, and phase. The accuracy of the infrared temperature is ±1°C, and the resolution is 0.1°C. The frequency of the vibration acceleration is 0.1-10 kHz, and the dynamic range is ±50 g; the diagnostic decision module includes threshold judgment and severity assessment. Among them, the threshold judgment is adjusted according to the dynamic baseline, and the background noise level is updated every 24 hours. The severity assessment is divided into four levels: normal, attention, abnormal, and critical based on the pulse density and energy growth rate; the model structure of the learning module includes a CNN branch, an LSTM branch, and a fusion layer. Among them, the CNN branch is used to process time-frequency images, the LSTM branch is used to process time-series features, and the fusion layer is used to splice the outputs of the CNN branch and the LSTM branch and then classify them through a fully connected layer.

[0025] The ultrasonic sensors are distributed in an equiangular ring array, where the sensors at the top and bottom are tilted 45° to achieve three-dimensional sound field coverage. The distance between adjacent ultrasonic sensors is ≥30 cm to avoid acoustic interference, and the ultrasonic sensors are set with a μ-Metal alloy shell with a permeability >50,000 to suppress the strong electromagnetic field interference of the transformer.

[0026] The circuit composition of the signal conditioning module includes high-pass filtering, pre-amplification, and band-pass filtering; High-pass filtering: Remove the low-frequency vibration noise of the transformer, use a second-order Butterworth filter, and the roll-off slope is -40 dB / dec; Pre-amplification: Use a low-noise instrumentation amplifier with an input noise density of 1.1 nV / √Hz and a gain error < 0.01%; Band-pass filtering: Adopt a switched-capacitor filter with a Q value of 5 to suppress high-frequency switching noise and out-of-band interference.

[0027] The processing flow of the data acquisition and processing module is as follows: Adaptive noise cancellation: Use the LMS algorithm to eliminate environmental noise in real time; Time-frequency analysis: Short-time Fourier transform, Hanning window, window length of 512 points, frequency resolution of 97.6 Hz; Feature extraction: The extracted features are peak frequency, pulse repetition rate, signal energy entropy, and phase difference positioning. Among them, the peak frequency is the frequency component with the largest amplitude detected in the spectrum, and the pulse repetition rate is calculated by detecting the reciprocal of the pulse interval time through zero-crossing detection.

[0028] The fusion method of the data fusion module is based on Dempster-Shafer evidence theory to solve the problem of uncertain information fusion, including: Basic probability assignment: Assign confidence levels of fault hypotheses to the data of each sensor; Decision output: Select the fault type with the highest confidence level.

[0029] The training data of the learning module comes from more than 2000 sets of on-site data augmentation, and the accuracy of the test set reaches 96.2%. Among them, data augmentation includes adding -10dB noise, time shift, and frequency shift.

[0030] A fault feature library is set between the data acquisition and processing module and the machine learning module. The construction method of the fault feature library is based on historical fault cases and laboratory simulation data. The fault types in the fault feature library include: Partial discharge: Main frequency, PRF, pulse rise time Core looseness: Low-frequency harmonic components, vibration correlation Winding deformation: High-frequency burst pulses, energy entropy change rate Oil bubbles: Wide-band energy distribution, no significant periodicity.

[0031] The training strategy of the learning module is as follows: Input data: Time-frequency diagram, generated by STFT, window function: Kaiser β = 6; Data augmentation: Add -20dB Gaussian noise, time shift jitter, frequency band masking; Transfer learning: Pre-train on the CWRU bearing dataset and freeze the first 3 layers during fine-tuning.

[0032] The functional modules of the human-computer interaction interface include: Real-time monitoring: The three-dimensional sound pressure cloud map shows the sound source position; Historical analysis: Supports the playback of time-frequency diagrams and characteristic trend curves; Alarm management: Hierarchical alarm, pushing SMS and email notifications; Report generation: Automatically outputs a PDF diagnostic report, including failure probability and maintenance suggestions.

[0033] Example 2: Partial discharge monitoring of the main transformer in an urban high-voltage substation Application scenario The oil chromatographic analysis of the main transformer #3 (model SSZ11-500000 / 500) in a 500 kV substation shows a trace amount of C2H2 (0.8 μL / L), and it is necessary to confirm whether there is early discharge.

[0034] System configuration Sensor layout: 8 ultrasonic probes (4 for bushings, 3 for oil tanks, 1 for coolers) Auxiliary monitoring: Infrared thermal imager (FLIRT1020), vibration sensor (10 - 1000 Hz) Algorithm version: ResNet-LSTM hybrid model v2.3 (training data volume > 5000 groups) Detection process Continuous monitoring for 72 hours, and high-frequency pulses in the range of 82 - 86 kHz are found in the B-phase bushing area, PRF = 65 Hz Multi-source data association: Infrared temperature measurement shows a temperature rise ΔT = 1.3 °C at the corresponding position The amplitude of the vibration spectrum exceeds the standard by 2.8 times at 100 Hz Three-dimensional positioning shows that the discharge point is 12 cm from the bushing flange surface (the actual disassembly deviation is 8 cm) Diagnosis result Fault type: Surface discharge caused by poor contact of the bushing end screen (discharge amount 4.2 pC) Disposal suggestion: Cut off the power supply to clean the contact surface of the end screen and apply conductive paste Implementation benefits Provide a 14-day early warning, avoiding unplanned power outages (reducing losses by approximately ¥2.8 million) The positioning accuracy reaches ±10 cm, and the maintenance time is shortened by 60%.

[0035] Example 3: Monitoring of winding deformation of the distribution transformer in an electric arc furnace in a steel plant Application scenario The load rate of the 12500 kVA rectifier transformer (model ZHSFPT-12500 / 35) in the steelmaking workshop has fluctuated greatly recently, and it is suspected that the windings are loose.

[0036] System configuration Sensor layout: 6 magnetic sensors (evenly distributed on the tank wall) Sampling strategy: Continuous sampling during load fluctuations (10 frames per second) Feature library: Contains 120 mechanical fault feature templates Detection process Captured a burst pulse group at 92 kHz, with a single pulse width ≤ 5 μs Time-frequency analysis shows a sudden change in energy entropy (ΔH = 0.42 / s) Vibration correlation analysis: The pulse appearance is synchronized with the load impact (correlation coefficient 0.87) The main axial vibration frequency is 92 Hz (corresponding to the winding resonance frequency) Diagnosis result Fault type: Winding displacement caused by insufficient axial pressing force (displacement ≈ 3 mm) Disposal suggestion: Cut off the power supply and adjust the pressure of the pressure bolt to 350 kN·m Implementation benefits Avoid winding short-circuit accidents and extend the life of the transformer by 5 - 8 years Achieve condition-based maintenance and reduce the annual maintenance cost by ¥450,000.

[0037] Example 4: Detection of air bubbles in the oil of the step-up transformer in a wind farm Application scenario The gas content in the oil of the #7 main transformer (model SFZ-80000 / 110) in a 150 MW wind farm has continuously risen to 3.8%.

[0038] System configuration Sensor layout: 4 immersion sensors (at the oil flow pipeline) Special design: Anti-condensation coating, wind noise-resistant microphone array Algorithm module: Special model for separating air bubble noise features Detection process Detected broadband noise (20 - 50 kHz), with an energy ratio of 38% Pulse characteristic analysis: No fixed PRF mode The waveform conforms to the characteristics of air bubble bursting (decay time constant τ = 8 μs) Oil temperature correlation analysis: The noise intensity is strongly correlated with the start and stop of the oil pump (P = 0.01) Diagnosis result Fault type: Air bubble intrusion caused by poor sealing of the submerged oil pump Disposal suggestion: Replace the mechanical seal of the oil pump and perform vacuum oil injection treatment Implementation benefits Accurately distinguish air bubbles from real discharges, and reduce the false alarm rate from 12% to 1.5% Avoid insulation degradation and estimate to extend the oil change cycle by 3 years.

[0039] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An ultrasonic detection system for transformer faults, comprising an ultrasonic sensor module, a signal conditioning module, a data acquisition and processing module, a data fusion module, a learning module, a diagnosis and decision module and a human-computer interaction interface, characterized in that: The ultrasonic sensor module consists of 6-8 ultrasonic sensors installed in a row inside the transformer housing, and the ultrasonic sensor adopts a piezoelectric ceramic sensor with a resonance frequency of 40kHZ. The sensitivity of the ultrasonic sensor is -65dBV / uBar, which can effectively capture weak; The signal conditioning module filters and amplifies the weak signal output by the ultrasonic sensor and converts it into a standardized signal suitable for digital acquisition; The hardware platform of the data acquisition and processing module is Xilinx Zynq-7020 SoC, which integrates FPGA and ARMCortex-A9 to achieve high-speed signal processing and flexible control; The data fusion module receives ultrasonic characteristics, infrared temperature and vibration acceleration from the sensor, wherein the ultrasonic characteristics include amplitude, frequency and phase, the infrared temperature has an accuracy of ±1°C and a resolution of 0.1°C, and the vibration acceleration has a frequency of 0.1-10kHZ and a dynamic range of ±50g; The diagnostic decision module includes threshold judgment and severity assessment, wherein the threshold judgment is adjusted according to the dynamic baseline, the background noise level is updated every 24 hours, and the severity assessment is divided into four levels: normal, caution, abnormal and critical based on the pulse density and energy growth rate; The model structure of the learning module includes a CNN branch, an LSTM branch and a fusion layer, wherein the CNN branch is used to process time-frequency images, the LSTM branch is used to process time series features, and the fusion layer is used to splice the outputs of the CNN branch and the LSTM branch and classify them through a fully connected layer.

2. The ultrasonic detection system for transformer fault according to claim 1, characterized in that: The ultrasonic sensors are distributed in an equiangular annular array, wherein the top and bottom sensors are inclined at 45 degrees to achieve three-dimensional sound field coverage.

3. The ultrasonic detection system for transformer fault according to claim 2, characterized in that: The distance between adjacent ultrasonic sensors is ≥30 cm to avoid sound wave interference, and the ultrasonic sensor is configured as a μ-Metal alloy shell with a magnetic permeability of >50,000 to suppress strong electromagnetic field interference from the transformer.

4. The ultrasonic detection system for transformer fault according to claim 1, characterized in that: The circuit composition of the signal conditioning module includes high-pass filtering, preamplification and band-pass filtering; High-pass filter: removes low-frequency vibration noise of transformer, using second-order Butterworth filter with a roll-off slope of -40 dB / dec; Preamplification: Use a low-noise instrumentation amplifier with an input noise density of 1.1 nV / √Hz and a gain error of <0.01%; Bandpass filtering: Use switched capacitor filter with Q value = 5 to suppress high-frequency switching noise and out-of-band interference.

5. The ultrasonic detection system for transformer fault according to claim 1, characterized in that: The processing flow of the data acquisition and processing module is as follows: Adaptive noise cancellation: Use LMS algorithm to eliminate environmental noise in real time; Time-frequency analysis: short-time Fourier transform, Hanning window, window length 512 points, frequency resolution 97.6 Hz; Feature extraction: The extracted features are peak frequency, pulse repetition rate, signal energy entropy and phase difference positioning, where the peak frequency is the frequency component with the largest amplitude detected in the spectrum, and the pulse repetition rate is calculated by zero-crossing detection to calculate the inverse of the pulse interval time.

6. The ultrasonic detection system for transformer fault according to claim 1, characterized in that: The fusion method of the data fusion module solves the problem of uncertain information fusion based on the Dempster-Shafer evidence theory, which includes: Basic probability assignment: assign confidence to the fault hypothesis for each sensor data; Decision output: Select the fault type with the highest confidence.

7. The ultrasonic detection system for transformer fault according to claim 1, characterized in that: The training data of the learning module comes from 2000+ sets of field data enhancement, and the accuracy of the test set reaches 96.2%, where data enhancement includes adding -10dB noise, time shift and frequency shift.

8. The ultrasonic detection system for transformer fault according to claim 1, characterized in that: A fault feature library is set between the data acquisition and processing module and the machine learning module. The construction method of the fault feature library is based on historical fault cases and laboratory simulation data, wherein the fault types of the fault feature library include: Partial discharge: main frequency, PRF, pulse rise time Core looseness: low-frequency harmonic components, vibration correlation Winding deformation: high-frequency burst pulse, energy entropy change rate Bubbles in oil: broadband energy distribution, no significant periodicity.

9. The ultrasonic detection system for transformer fault according to claim 7, characterized in that: The training strategy of the learning module is: Input data: time-frequency diagram, STFT generation, window function: Kaiser β=6; Data enhancement: adding -20dB Gaussian noise, time-shift jitter, and frequency band masking; Transfer learning: Pre-trained on the CWRU bearing dataset, freezing the first 3 layers when fine-tuning.

10. The ultrasonic detection system for transformer fault according to claim 1, characterized in that: The functional modules of the human-computer interaction interface include: Real-time monitoring: 3D sound pressure cloud map shows the location of the sound source; Historical analysis: supports playback of time-frequency graphs and characteristic trend curves; Alarm management: graded alarm, push SMS and email notifications; Report generation: Automatically output PDF diagnostic report including failure probability and maintenance suggestions.

Images