Partial discharge recognition and positioning method based on multi-modal calibration and blind source separation

Abstract

The application discloses a partial discharge recognition and positioning method based on multi-modal calibration and blind source separation, which comprises the following steps: step 1, arranging a high-frequency current sensor at the line of a transformer core grounding flat iron or a clamp grounding flat iron; step 2, collecting a mixed pulse signal and performing time-frequency conversion to output multi-modal features; step 3, using blind source separation on the mixed pulse signal to output a plurality of independent pulse source signals and signal parameters; step 4, periodically injecting a reference pulse signal into the flat iron, estimating and adaptively updating the signal propagation speed along the flat iron based on the time difference of arrival; and step 5, acquiring the multi-modal features and the signal parameters, inputting the multi-modal features and the signal parameters into a deep learning classifier, outputting classification confidence, comprehensively judging the interference source of the pulse signal, and realizing internal and external discharge recognition and flat iron edge positioning.

Description

Technical Field

[0001] This invention relates to the field of partial discharge signal recognition technology, and in particular to a method for partial discharge recognition and localization based on multimodal calibration and blind source separation. Background Technology

[0002] Partial discharge is an important early indicator of insulation defects in transformers. For transformers of 500kV and above, commonly used online detection methods include high-frequency current sensors (HFCT) and ultrasonic sensors. However, in actual operation, the transformer's core grounding flat iron or clamp grounding flat iron is connected to the power grid grounding grid. Discharge or switching impacts generated by other equipment in the grounding grid can couple to the grounding point, resulting in a large amount of external interference and coupling noise in the detection signal.

[0003] Existing rule-based discrimination methods based on high-frequency current sensors at both ends, such as those based on time difference of arrival, polarity, and waveform similarity, can perform basic internal and external differentiation and location along flat iron under simple operating conditions. However, their performance deteriorates significantly under complex coupling, multi-pulse superposition, low signal-to-noise ratio, and the influence of on-site structural and temperature changes on propagation speed, leading to false alarms or missed alarms. Furthermore, traditional solutions typically rely on fixed signal propagation speed estimation, lack on-site adaptive calibration methods, and are insufficiently adaptable to overlapping pulses or new interference sources. Summary of the Invention

[0004] The purpose of this invention is to provide a partial discharge identification and localization method based on multimodal calibration and blind source separation, which solves the problems of traditional schemes that usually rely on fixed signal propagation speed estimation, lack on-site adaptive calibration means, and are not adaptable to overlapping pulses or new interference sources.

[0005] The present invention solves the technical problem by adopting the following technical solution:

[0006] A partial discharge identification and localization method based on multimodal calibration and blind source separation.

[0007] Step 1: Deploy a high-frequency current sensor, connect the output terminal of the high-frequency current sensor to a high-speed data acquisition device, and set the sampling channel clock synchronization. The high-frequency current sensor is connected along the grounding flat iron of the transformer core or the grounding flat iron of the clamp.

[0008] Step 2: Acquire a mixed pulse signal, which includes a high-frequency sensor signal and a signal from at least one other sensor. Perform time-frequency transformation on the mixed pulse signal and output multimodal features. The multimodal features include: the original time-series waveform, time-spectrum diagram, ultrasound envelope signal, PRPD image, energy ratio, spectral centroid, and instantaneous phase information.

[0009] Step 3: Define a blind source separation convolutional hybrid model, use blind source separation to output several independent pulse source signals for the hybrid pulse signal, and calculate the signal parameters of the independent pulse source signals. The signal parameters include: waveform similarity, pulse polarity, time difference of arrival, energy and spectral characteristics.

[0010] Step 4: Periodically inject a reference pulse signal into the flat iron, and estimate and adaptively update the signal propagation speed along the flat iron based on the arrival time difference of the reference pulse;

[0011] Step 5: Obtain the multimodal features and the signal parameters, input them into the deep learning classifier, output the classification confidence score, and comprehensively determine the source of pulse signal interference.

[0012] Preferably, the spacing between the high-frequency current sensors is set to 2-3 meters, the sampling rate of the high-speed data acquisition device is not less than 1 GS / s, and the clock synchronization error between the sampling channels is less than 1 nanosecond.

[0013] Preferably, in step 5, the comprehensive judgment specifically involves fusing the classification confidence level with evidence based on polarity, time difference, and similarity to determine whether the pulse signal originates from external interference or internal discharge of the transformer.

[0014] If the signal is determined to be an internal discharge, the system will locate the signal along the flat iron vector based on the signal propagation speed and arrival time difference, record the location information, and save the determination result, interpretability output, and security timestamp.

[0015] Preferably, in step 3, blind source separation employs independent component analysis, sparse component analysis, or a decomposition method based on sparse representation, and prioritizes segmenting short-time energy segments before implementing blind source separation.

[0016] Preferably, the deep learning classifier includes: an anomaly detection module, an online incremental learning module, and a pseudo-label self-learning module.

[0017] Preferably, it also includes: an anomaly detection module, which uploads anomaly events detected and identified based on historical events, and performs manual review and training set expansion. The historical events specifically refer to an unsupervised detection model constructed based on historical features and events that are significantly different from historical samples.

[0018] A partial discharge identification and localization system based on multimodal calibration and blind source separation includes:

[0019] A sensor module that acquires and outputs mixed pulse signals, the sensor module comprising: several high-frequency current sensors and at least one other type of sensor;

[0020] A high-speed data acquisition module receives the mixed pulse signal and performs analog-to-digital conversion, with a sampling rate ≥1GS / s and a clock synchronization error ≤1ns.

[0021] The processing and judgment module processes the mixed pulse data to generate classification confidence scores. The processing includes: blind source separation, multimodal feature extraction, adaptive propagation speed estimation, deep learning classification, and evidence fusion judgment.

[0022] The self-test calibration module is used to inject a reference pulse signal into the flat iron, calibrate the channel delay and gain, and output the calibration result.

[0023] The anomaly detection module uploads anomaly events detected and identified based on historical events, and performs manual review and training set expansion.

[0024] The storage module is used to save process data and output results. The process data includes: mixed pulse signals, independent pulse source signals, and multimodal features. The output results include: judgment results, interpretable output, and secure timestamps.

[0025] Preferably, the number of high-frequency current sensors is ≥3, and they are arranged along the flat iron linear array.

[0026] Preferably, the processing and judgment module processes the mixed pulse signal through beamforming to enhance the signal and suppress external interference.

[0027] A computer-readable storage medium storing at least one program, which is loaded and executed by a processor to implement the partial discharge identification and localization method described above.

[0028] This invention enhances pulse signal separation capabilities by deploying multi-channel high-frequency current sensors and other sensors along the flat iron line and employing blind source separation processing for mixed pulse signals; it improves positioning accuracy by adaptively estimating propagation speed based on periodic reference pulse signals; it achieves high-precision automatic discrimination of complex coupled interference by using a deep learning classifier to output classification confidence; and it realizes internal and external discharge identification and positioning along the flat iron by comprehensively judging rule features and data-driven results through evidence fusion methods. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0030] The technical solution of the present invention will be further described below with reference to the embodiments and accompanying drawings.

[0031] Example 1

[0032] A partial discharge identification and localization method based on multimodal calibration and blind source separation.

[0033] Step 1: Deploy a high-frequency current sensor. Connect the output terminal of the high-frequency current sensor to a high-speed data acquisition device and set the sampling channel clock synchronization. Connect the high-frequency current sensor to the grounding flat iron of the transformer core or the grounding flat iron of the clamp.

[0034] Step 2: Acquire a mixed pulse signal, which includes a high-frequency sensor signal and a signal from at least one other sensor. Perform time-frequency transformation on the mixed pulse signal to output multimodal features. The multimodal features include: the original time-series waveform, time-spectrum graph, ultrasonic envelope signal, PRPD image, energy ratio, spectral centroid, and instantaneous phase information. The other sensor can be at least one of a temperature sensor, humidity sensor, ultrasonic sensor, ultra-high frequency sensor, and humidity sensor.

[0035] Step 3: Define a convolutional mixing model for blind source separation. Use blind source separation to output several independent pulse source signals for the mixed pulse signal. Calculate the signal parameters of the independent pulse source signals, including: waveform similarity, pulse polarity, time difference of arrival, energy, and spectral characteristics.

[0036] Step 4: Periodically inject reference pulse signals into the flat iron, estimate and adaptively update the signal propagation speed along the flat iron based on the arrival time difference of the reference pulses;

[0037] Step 5: Obtain multimodal features and signal parameters, input them into a deep learning classifier, output classification confidence, and comprehensively determine the source of pulse signal interference.

[0038] In a further embodiment of this invention, the spacing between high-frequency current sensors is set to 2-3 meters, the sampling rate of the high-speed data acquisition device is not less than 1 GS / s, and the clock synchronization error between sampling channels is less than 1 nanosecond.

[0039] In a further implementation of this embodiment, in step 5, the comprehensive judgment specifically involves fusing the classification confidence level with evidence based on polarity, time difference, and similarity to determine whether the pulse signal originates from external interference or internal discharge of the transformer.

[0040] If the signal is determined to be internal discharge, the system will locate the signal along the flat iron vector based on the signal propagation speed and arrival time difference, record the location information, and save the judgment result, interpretability output, and security timestamp.

[0041] In a further implementation of this embodiment, in step 3, blind source separation is performed using independent component analysis, sparse component analysis, or a decomposition method based on sparse representation, and short-time energy segments are segmented first before blind source separation is implemented.

[0042] In a further implementation of this embodiment, the deep learning classifier includes: an anomaly detection module, an online incremental learning module, and a pseudo-label self-learning module.

[0043] A further implementation of this embodiment includes: an anomaly detection module, which uploads anomaly events detected and identified based on historical events, performs manual review and expands the training set. The historical events specifically refer to an unsupervised detection model constructed based on historical features and events that are significantly different from historical samples.

[0044] In a further embodiment of this example, four high-frequency current sensors are installed every 2.5 meters on the grounding flat iron of the 500kV transformer core, and two ultrasonic sensors are installed on the transformer shell. The output of the high-frequency current sensors is connected to a 4-channel high-speed data acquisition card.

[0045] The high-speed data acquisition card has a sampling rate of 1.5GS / s and a clock synchronization error of ≤0.8ns. It synchronously acquires mixed pulse signals, which include high-frequency current signals, ultrasonic envelope signals, and temperature sensor data.

[0046] The frequency band is set to 80-300MHz. Wavelet transform is performed on the high-frequency current signal to generate a time spectrum. The peak value of the time waveform and PRPD image are extracted. The energy ratio of the ultrasonic envelope signal is calculated. The energy ratio is high frequency / low frequency. Values ​​≥3 are selected and marked as valid pulses.

[0047] When using blind source separation to process mixed pulse signals, sparse component analysis is employed to handle mixed pulse segments whose energy exceeds a threshold (3 times the standard deviation of background noise).

[0048] Constructing a hybrid convolutional model: Where X is the observed signal, s k As the source signal, a k τ is the mixing coefficient. k To mitigate the time delay, the independent pulse source signals are separated by minimizing the demixing matrix using the L1 norm.

[0049] Independent pulse source signals are separated, and the parameters of each source signal need to be calculated. The waveform similarity is quantified by normalizing the cross-correlation to determine the consistency of the morphology of the two pulse signals. Experimental data shows that the range of internal discharge pulses from the same source is 0.88 to 0.99, while that of external interference pulses is 0.30 to 0.65. Therefore, in this embodiment, signals ≥0.85 are determined to be signals from the same source. The arrival time difference to the independent pulse source signal is calculated by extracting 10%-90% of the waveform rising edge, avoiding pulse initiation noise (0-10%) and top oscillation (90-100%). The spectral characteristics are used to determine whether it is an internal discharge. Based on the propagation characteristics of electromagnetic waves in oil-paper medium, the discharge pulse width at the insulation defect in the transformer is narrow, ranging from 1-5 ns, and the main frequency of the spectrum is concentrated in 80-120 MHz. The main frequency of external interference such as switching operations is usually <30 MHz or >300 MHz.

[0050] In step 4, a reference pulse with a width of 5ns is injected into the flat iron through the self-test module at midnight every day, the arrival time difference Δt of the high-frequency current sensors at both ends is measured, and the Kalman filter is used to smooth the measurement values ​​for 7 consecutive days.

[0051] In step 5, multimodal features are input into the pre-trained CNN-Transformer hybrid model. The CNN-Transformer hybrid model includes a CNN module, a Transformer module, and a fusion module. The CNN module extracts image features, and the Transformer module processes time-series waveforms. Pre-training is based on selected real and simulated high-frequency, ultrasonic, and other partial discharge data, which are input into the CNN-Transformer hybrid model, and the model outputs the recognition accuracy. The model parameters are adjusted based on the recognition accuracy, and training is completed after reaching 98%.

[0052] The CNN-Transformer hybrid model outputs three results: internal discharge, external disturbance, and unknown events.

[0053] The evidence fusion specifically involves the following: all three high-frequency current sensors detected pulses with negative polarity, consistent with internal discharge characteristics, and a weight of 0.3 was set; arrival time difference matching was used, with the time difference ΔT between adjacent sensors being used for matching. 12 = 3.2ns, ΔT 23 =3.0ns, substituting into the positioning equation, the error is <5%, so the weight is set to 0.4; for waveform similarity, the cross-correlation coefficient of the separated pulses ρ = 0.91 > 0.85, so the weight is set to 0.3;

[0054] The specific calculation process for the deep learning results is as follows: when an abnormal pulse is detected on the transformer grounding flat iron, the deep learning classifier starts working. First, the raw pulse signal captured by the high-frequency current sensor undergoes wavelet transform to output a time-frequency spectrum. Simultaneously, the phase-amplitude relationship of the raw pulse within the power frequency cycle is extracted to output a PRPD image. The signal acquired by the ultrasonic sensor undergoes Hilbert transform to extract the envelope waveform, which is then interpolated and aligned to the same time axis.

[0055] The pulse's energy ratio and spectral centroid are quantized into numerical values, normalized, and converted into grayscale values. These values ​​are then integrated into a four-channel 256×256 pixel tensor. The first channel is a time-spectrum graph, recording frequency domain characteristics; the second channel is a PRPD image, recording the discharge phase pattern; the third channel is an ultrasonic envelope, capturing the correlation of mechanical vibration; and the fourth channel incorporates quantitative indicators of energy and spectrum to fully describe the physical characteristics of the pulse.

[0056] Next, the pixel tensors are input into a pre-trained CNN-Transformer hybrid model. The CNN module parses the temporal spectrum and spatial patterns in the PRPD image through a three-layer convolutional network. The first convolutional layer uses a small 3×3 kernel to slide across the image, capturing local details. The residual connection structure ensures that the deep network can continuously optimize feature extraction capabilities and avoid loss of details. The Transformer module processes the raw waveform data of high-frequency current, dividing the 1024 sampling points into 16 time segments. The temporal relationship of each segment is marked by position encoding, and then the correlation of key parts such as the pulse rising edge and oscillation decay is analyzed by using a multi-head attention mechanism. The morphological consistency of the waveform in the 10%-90% rising phase is as high as 89%, which is highly correlated with the internal discharge.

[0057] The outputs of the CNN and Transformer modules are fed into a fusion module. The 64×64×128 dimensional spatial features extracted by the CNN module and the 16×64 dimensional temporal features generated by the Transformer module are flattened and concatenated to form a joint feature vector exceeding 520,000 dimensions. This joint feature vector is processed through two fully connected neural network layers. The first layer uses the ReLU activation function to filter salient features, and the second layer outputs three raw score values, corresponding to the probabilities of internal discharge, external interference, and unknown events, respectively. Finally, the Softmax function transforms the scores into a probability distribution. In this embodiment, the score for internal partial discharge is significantly higher than the other two categories, with a calculated quantization result of P_in = 0.88, meaning the model has an 88% confidence level in determining that the current pulse is a true internal partial discharge.

[0058] To verify the reliability of this result, the model provides interpretability through gradient-weighted class activation maps. The 100-105MHz frequency band in the time-spectrum graph exhibits a bright response, typical of internal discharge. The PRPD image shows dense clusters of scattered points in the first quadrant, consistent with the phase characteristics of oil-paper insulation discharge. The attention weights of the Transformer module are concentrated at 72% in the pulse rising edge region, perfectly matching the physical laws of electrical equipment discharge, thus confirming the scientific validity of the probability values.

[0059] The generation of deep learning probability values ​​establishes a transparent mapping from raw signals to decision-making criteria through a triple guarantee of multimodal feature collaboration, hybrid model architecture optimization, and visualization of physical laws. It inherits the adaptability advantages of data-driven methods and meets the stringent requirements for result reliability in the power monitoring field through an interpretable mechanism, providing a highly reliable criterion for transformer condition-based maintenance.

[0060] After obtaining the deep learning result P_in = 0.88, we first determine whether there is a conflict between the deep learning result and the rule feature detection. If there is no conflict, the conflict factor is close to 0. Then, we synthesize the confidence score using the weighted fusion formula: Comprehensive Confidence Score = P_in * (Polarity Weight + Time Difference Weight + Waveform Weight). Substituting the above values, we output the comprehensive confidence score as 0.88 × (0.3 + 0.4 + 0.3) = 0.88 × 1.0 = 0.88.

[0061] If the threshold trigger is met, and the overall confidence level is ≥0.88 (preset threshold 0.85), then the conclusion of internal partial discharge is given.

[0062] Based on the reference pulses from previous injections, the current propagation velocity of the flat iron is updated to 1.65 × 10⁻⁶. 8 m / s, positioning along the flat iron: Substitute the time difference ΔT 12 =3.2ns and sensor spacing d=2.5m, fault location = (d×ΔT) 12 ) / (ΔT 12 +ΔT 23 = (2.5 × 3.2) / (3.2 + 3.0) = 1.29 meters. The positioning accuracy is ±0.2 meters at 1.29 meters from the grounding end of the iron core. The original waveform segment, fusion confidence level 0.88, positioning coordinates and timestamp are encrypted and saved. The system synchronously triggers an alarm SMS to the maintenance personnel.

[0063] Example 2

[0064] A partial discharge identification and localization system based on multimodal calibration and blind source separation includes:

[0065] The sensor module acquires and outputs mixed pulse signals. The sensor module includes: several high-frequency current sensors and at least one other type of sensor.

[0066] High-speed data acquisition module: The high-speed data acquisition module receives mixed pulse signals and performs analog-to-digital conversion, with a sampling rate ≥1GS / s and a clock synchronization error ≤1ns;

[0067] The processing and judgment module processes the mixed pulse data to generate classification confidence scores. The processing includes: blind source separation, multimodal feature extraction, adaptive propagation speed estimation, deep learning classification, and evidence fusion judgment.

[0068] The self-test calibration module is used to inject a reference pulse signal into the flat iron, calibrate the channel delay and gain, and output the calibration result.

[0069] The anomaly detection module uploads anomaly events detected and identified based on historical events, and performs manual review and training set expansion.

[0070] The storage module is used to save process data and output results. The process data includes: mixed pulse signals, independent pulse source signals, and multimodal characteristics. The output results include: judgment results, interpretable output, and secure timestamps.

[0071] In a further embodiment of this example, the number of high-frequency current sensors is ≥3, and they are arranged along the flat iron linear array.

[0072] In a further embodiment of this invention, the processing and judgment module processes the mixed pulse signal through beamforming to enhance the signal and suppress external interference.

[0073] Example 3

[0074] A computer-readable storage medium storing at least one program, which is loaded and executed by a processor to implement the partial discharge identification and localization method described above.

[0075] Embodiment 3 of the present invention provides a server, which includes a processor and a storage medium. The storage medium stores at least one instruction, at least one program, code set or instruction set. The at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement the partial discharge identification and location method provided in Embodiment 1 above.

[0076] Storage media can be used to store software programs and modules. The processor executes various functional applications and data processing by running the software programs and modules stored in the storage media. The storage media may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for the functions, etc.; the data storage area may store data created based on the use of the device, etc. Furthermore, the storage media may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory may also include a memory controller to provide the processor with access to the memory.

[0077] In this embodiment 3, the storage medium can be located in at least one of the multiple network servers in a computer network.

[0078] In this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0079] The order of the above embodiments is for ease of description only and does not represent the superiority or inferiority of the embodiments.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for partial discharge identification and localization based on multimodal calibration and blind source separation, characterized in that, Includes the following steps: Step 1: Deploy a high-frequency current sensor, connect the output terminal of the high-frequency current sensor to a high-speed data acquisition device, and set the sampling channel clock synchronization. The high-frequency current sensor is connected along the grounding flat iron of the transformer core or the grounding flat iron of the clamp. Step 2: Acquire a mixed pulse signal, which includes a high-frequency sensor signal and a signal from at least one other sensor. Perform time-frequency transformation on the mixed pulse signal and output multimodal features. The multimodal features include: the original time-series waveform, time-spectrum diagram, ultrasound envelope signal, PRPD image, energy ratio, spectral centroid, and instantaneous phase information. Step 3: Define a blind source separation convolutional hybrid model, use blind source separation to output several independent pulse source signals for the hybrid pulse signal, and calculate the signal parameters of the independent pulse source signals. The signal parameters include: waveform similarity, pulse polarity, time difference of arrival, energy and spectral characteristics. Step 4: Periodically inject a reference pulse signal into the flat iron, and estimate and adaptively update the signal propagation speed along the flat iron based on the arrival time difference of the reference pulse; Step 5: Obtain the multimodal features and the signal parameters, input them into the deep learning classifier, output the classification confidence score, and comprehensively determine the source of pulse signal interference; The multimodal features are input into a pre-trained CNN-Transformer hybrid model, which includes a CNN module, a Transformer module, and a fusion module. The CNN module extracts image features, and the Transformer module processes time-series waveforms. The pre-training is based on selected real and simulated high-frequency, ultrasonic, and other partial discharge data, which are input into the CNN-Transformer hybrid model and output the recognition accuracy. The model parameters are adjusted based on the recognition accuracy, and training is completed after reaching 98%. The CNN-Transformer hybrid model outputs three results: internal discharge, external disturbance, and unknown events. The specific calculation process of the deep learning results is as follows: when an abnormal pulse is detected on the transformer grounding flat iron, the deep learning classifier starts to work. First, the original pulse signal captured by the high-frequency current sensor is transformed by wavelet transform and outputs a time spectrum. The phase-amplitude relationship of the original pulse in the power frequency cycle is extracted simultaneously and the PRPD image is output. The signal collected by the ultrasonic sensor is processed by Hilbert transform to extract the envelope waveform and interpolated to the same time axis. The pulse's energy ratio and spectral centroid were quantized into numerical values, normalized, and converted into grayscale values. These values ​​were then integrated into a four-channel 256×256 pixel tensor. The first channel is a time-spectrum plot, recording frequency domain characteristics; the second channel is a PRPD image, recording the discharge phase pattern; the third channel is an ultrasonic envelope, capturing the mechanical vibration correlation; and the fourth channel incorporates the quantified energy and spectral parameters to comprehensively describe the pulse's physical characteristics. Next, the pixel tensor is input into the pre-trained CNN-Transformer hybrid model. The CNN module parses the spatial patterns in the temporal spectrogram and PRPD image through a three-layer convolutional network. The first convolutional layer uses a small 3×3 kernel to slide and scan on the image to capture local details. The residual connection structure ensures that the deep network can continuously optimize the feature extraction capability and avoid loss of details. The Transformer module processes the raw waveform data of high-frequency current, divides the 1024 sampling points into 16 time segments, marks the temporal relationship of each segment through position encoding, and then uses a multi-head attention mechanism to analyze the correlation between the pulse rising edge and key parts of oscillation decay. The morphological consistency of the waveform in the 10%-90% rising stage is as high as 89%, which is highly correlated with the internal discharge. The outputs of the CNN and Transformer modules are fed into the fusion module. The 64×64×128-dimensional spatial features extracted by the CNN module and the 16×64-dimensional temporal features generated by the Transformer module are flattened and concatenated to form a joint feature vector of over 520,000 dimensions. The joint feature vector is processed by two layers of fully connected neural networks. The first layer uses the ReLU activation function to filter significant features, and the second layer outputs three raw score values, corresponding to the probability of internal discharge, external interference, and unknown events, respectively. Finally, the Softmax function transforms the scores into a probability distribution.

2. The partial discharge identification and localization method based on multimodal calibration and blind source separation according to claim 1, characterized in that, The spacing between the high-frequency current sensors is set to 2-3 meters, the sampling rate of the high-speed data acquisition device is not less than 1 GS / s, and the clock synchronization error between the sampling channels is less than 1 nanosecond.

3. The partial discharge identification and localization method based on multimodal calibration and blind source separation according to claim 1, characterized in that, In step 5, the comprehensive judgment specifically involves fusing the classification confidence level with evidence based on polarity, time difference, and similarity to determine whether the pulse signal originates from external interference or internal discharge within the transformer. Specifically, the evidence fusion involves the three high-frequency current sensors detecting pulses with negative polarity, consistent with internal discharge characteristics, and setting a weight of 0.

3. Time difference matching is also performed, with the time difference ΔT between adjacent sensors being considered. 12 =3.2ns, ΔT 23 =3.0ns, substituting into the positioning equation, the error is <5%, so the weight is set to 0.4; for waveform similarity, the cross-correlation coefficient of the separated pulses ρ=0.91>0.85, so the weight is set to 0.3; If the signal is determined to be an internal discharge, the system will locate the signal along the flat iron vector based on the signal propagation speed and arrival time difference, record the location information, and save the determination result, interpretability output, and security timestamp.

4. The partial discharge identification and localization method based on multimodal calibration and blind source separation according to claim 1, characterized in that, In step 3, blind source separation employs independent component analysis, sparse component analysis, or a decomposition method based on sparse representation, and prioritizes segmenting short-time energy segments before implementing blind source separation.

5. The partial discharge identification and localization method based on multimodal calibration and blind source separation according to claim 1, characterized in that, The deep learning classifier includes: an anomaly detection module, an online incremental learning module, and a pseudo-label self-learning module.

6. The partial discharge identification and localization method based on multimodal calibration and blind source separation according to claim 5, characterized in that, Also includes: An anomaly detection module uploads anomaly events detected and identified based on historical events, and performs manual review and training set expansion. The historical events specifically refer to an unsupervised detection model constructed based on historical features and events that are significantly different from historical samples.

7. A positioning system based on the partial discharge identification and positioning method based on multimodal calibration and blind source separation as described in any one of claims 1-6, characterized in that, include: A sensor module that acquires and outputs mixed pulse signals, the sensor module comprising: several high-frequency current sensors and at least one other type of sensor; A high-speed data acquisition module receives the mixed pulse signal and performs analog-to-digital conversion, with a sampling rate ≥1GS / s and a clock synchronization error ≤1ns. The processing and judgment module processes the mixed pulse signal to generate classification confidence. The processing includes: blind source separation, multimodal feature extraction, adaptive propagation speed estimation, deep learning classification, and evidence fusion judgment. The self-test calibration module is used to inject a reference pulse signal into the flat iron, calibrate the channel delay and gain, and output the calibration result. The anomaly detection module uploads anomaly events detected and identified based on historical events, and performs manual review and training set expansion. The storage module is used to save process data and output results. The process data includes: mixed pulse signals, independent pulse source signals, and multimodal features. The output results include: judgment results, interpretable output, and secure timestamps.

8. The positioning system according to claim 7, characterized in that, The number of high-frequency current sensors is ≥3, and they are arranged along the flat iron linear array.

9. The positioning system according to claim 7, characterized in that, The processing and judgment module processes the mixed pulse signal through beamforming to enhance the signal and suppress external interference.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one program, which is loaded and executed by a processor to implement the partial discharge identification and localization method as described in any one of claims 1 to 6.

Images