Transformer partial discharge detection method and system based on multi-modal feature combination

Abstract

The invention discloses a transformer partial discharge detection method and system based on multi-modal feature combination, and relates to the technical field of transformer detection, and the key points of the technical scheme are that the method comprises the steps: collecting an electromagnetic wave signal, a pulse current signal and a sound wave signal of a target transformer; respectively extracting an electromagnetic wave time-frequency matrix, a pulse current waveform parameter and a sound wave space coordinate; performing multi-modal feature fusion on the electromagnetic wave time-frequency matrix, the pulse current waveform parameters and the sound wave space coordinates based on a dynamic weighting strategy to generate a joint feature tensor; and inputting the joint feature tensor into a pre-constructed neural network model, identifying a discharge type and optimizing a positioning coordinate. According to the method, multi-modal features are subjected to tensor fusion, cross-modal correlation learning is realized through the neural network model, information islands can be eliminated, and the misjudgment rate is effectively reduced; and electromagnetic wave-current-sound wave signal adaptive fusion is realized through a dynamic weight coefficient, and the signal-to-noise ratio is obviously improved through characteristic complementation in an oil paper insulation discharge scene.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer detection, and more specifically, to a transformer partial discharge detection method and system based on multi-modal feature combination. Background Art

[0002] Transformer partial discharge detection is an important means to evaluate the insulation status of transformers. Partial discharge (PD) is a phenomenon in which the insulation medium inside the transformer undergoes non-penetrating breakdown under the action of an electric field, and the discharge energy is concentrated in micron-level defect areas (such as bubbles and cracks). Studies have shown that when the size of the insulation defect exceeds 50μm, the partial discharge starting voltage decreases by about 30%, becoming a direct sign of insulation degradation.

[0003] The existing technology mainly uses detection methods such as ultra-high frequency (UHF), ultrasonic (AE), and optical pulse (HFCT). The ultra-high frequency detection method is characterized by strong anti-electromagnetic interference ability by coupling the electromagnetic wave signal of the frequency band above 300MHz generated by partial discharge, but the traditional UHF sensor is easily affected by the shell structure due to external installation. The ultrasonic detection method mostly uses piezoelectric sensors to be placed outside the transformer shell, and there are positioning deviations and electromagnetic interference problems caused by the complex sound wave propagation path. The optical pulse detection method captures the optical radiation signal of partial discharge, but it is seriously interfered by the visible light of the environment, and the sensitivity of the independent optical fiber probe is limited, making it difficult to achieve multi-angle signal capture. In order to solve the defects of using a single detection method for transformer partial discharge detection in the existing technology, some existing technologies also record the method of running multiple detection methods in parallel, and outputting the results after one method detects them, but this method requires more computing resources for data processing, and the detection sensitivity is too high, and there is a certain probability of false alarm.

[0004] Therefore, how to study and design a transformer partial discharge detection method and system based on the combination of multi-modal features that can overcome the above-mentioned defects is a problem that we urgently need to solve. Summary of the invention

[0005] In order to address the deficiencies in the prior art, the purpose of the present invention is to provide a transformer partial discharge detection method and system based on multimodal feature combination, which performs tensor fusion of multi-dimensional electromagnetic wave time-frequency matrix, pulse current waveform parameters and acoustic wave space coordinates, and realizes cross-modal association learning through a neural network model, which can eliminate information islands and effectively reduce the misjudgment rate; and realizes adaptive fusion of electromagnetic wave-current-acoustic wave signals through dynamic weight coefficients. In the oil-paper insulation discharge scenario, feature complementarity significantly improves the signal-to-noise ratio.

[0006] The above technical objectives of the present invention are achieved through the following technical solutions: In a first aspect, a transformer partial discharge detection method based on multi-modal feature combination is provided, comprising the following steps: Collect electromagnetic wave signals, pulse current signals and sound wave signals of the target transformer; Extracting the electromagnetic wave time-frequency matrix, the pulse current waveform parameters and the acoustic wave space coordinates from the electromagnetic wave signal, the pulse current signal and the acoustic wave signal respectively; Based on a dynamic weighting strategy, multi-modal feature fusion is performed on the electromagnetic wave time-frequency matrix, pulse current waveform parameters and acoustic wave space coordinates to generate a joint feature tensor; The joint feature tensor is input into a pre-built neural network model to identify the discharge type and optimize the positioning coordinates.

[0007] Furthermore, the process of collecting the electromagnetic wave signal, pulse current signal and sound wave signal of the target transformer is specifically as follows: The electromagnetic wave signals in the frequency band of 300MHz-1.5GHz are collected by the UHF sensor array installed on the transformer housing; The 3MHz-30MHz pulse current signal is collected by a high-frequency current transformer coupled to the neutral point grounding wire; Furthermore, the acoustic vibration signal of 20kHz-300kHz is collected by a piezoelectric sensor array arranged on the surface of the oil tank.

[0008] Furthermore, the multimodal feature fusion expression based on the dynamic weighting strategy is: ; in, represents the joint feature tensor; , , Respectively represent the dynamic weight coefficients of electromagnetic wave signals, pulse current signals and acoustic wave signals, with a value range of [0,1]; Represents the electromagnetic wave time-frequency energy matrix, extracted by wavelet packet decomposition; It represents the pulse current characteristic vector, which is composed of the pulse density, rise time and polarity in the pulse current waveform parameters; Represents the spatial coordinates of the sound wave, which is calculated by the DOA algorithm and / or the TDOA algorithm; Represents tensor concatenation.

[0009] Furthermore, the dynamic weight coefficient is determined as follows: The sum of the dynamic weight coefficients of the electromagnetic wave signal, the pulse current signal and the sound wave signal is constrained to be 1, and the greater the signal-to-noise ratio of the corresponding signal, the greater the dynamic weight coefficient allocated to the corresponding signal.

[0010] Furthermore, the neural network model adopts a lightweight convolutional neural network consisting of an input layer, a hidden layer and an output layer; The number of neurons in the input layer is equal to the sum of the dimensions of the vectors corresponding to the electromagnetic wave time-frequency matrix, the pulse current waveform parameters and the acoustic wave space coordinates; The hidden layer consists of 2 fully connected layers and uses the ReLU activation function; The number of neurons in the output layer is equal to the number of output discharge types, and a Softmax activation function is used.

[0011] Furthermore, the expression for optimizing the positioning coordinates is: ; in, Indicates the optimized positioning coordinates; Indicates the coarse positioning coordinates based on sound waves; Represents the electromagnetic wave positioning coordinates based on the electromagnetic wave attenuation model; Represents the correction coefficient, which is determined by fitting historical data.

[0012] Furthermore, the expression of the electromagnetic wave attenuation model is: ; in, Indicates the electromagnetic wave signal strength at distance d; Indicates the initial signal strength at the discharge source; It represents the attenuation coefficient of transformer oil-paper composite medium, and its value range is 0.05-0.1.

[0013] Furthermore, the method also includes: The electromagnetic wave signal, the pulse current signal and the sound wave signal are subjected to time delay compensation and / or phase alignment.

[0014] Furthermore, the method also includes: According to the pulse density of the pulse current signal N Dynamically adjust the number of wavelet packet decomposition layers: when N When >100 pulses / ms, the number of decomposition layers is ≤12; When 50 pulses / ms ≤ N When ≤100 pulses / ms, 12<decomposition layers<16; when N When <50 pulses / ms, the number of decomposition layers is ≥16.

[0015] Furthermore, the method also includes: Based on the high-frequency energy concentration period in the electromagnetic wave time-frequency matrix, the pulse current analysis time window is adjusted to 1.2-1.5 times the high-frequency energy concentration period; and / or, according to the signal-to-noise ratio of the electromagnetic wave signal Dynamically adjust pulse amplitude threshold : .

[0016] Furthermore, the method also includes: Positioning coordinates based on electromagnetic waves , constraining the spatial search range of the TDOA algorithm to cm; and / or, selecting a matching acoustic wave propagation attenuation model library according to the pulse current polarity ratio; The acoustic wave spatial coordinates are calculated using a TDOA algorithm or a method combining a TDOA algorithm with a DOA algorithm.

[0017] Furthermore, the method also includes: Based on the rough positioning coordinates based on sound waves , adjust the beamforming direction of the UHF sensor array so that the main lobe points to cm area; and / or, based on the acoustic wave propagation path length , correct the propagation distance parameter in the electromagnetic wave attenuation model for: .

[0018] Furthermore, the adjustment process of the beamforming direction is: When the acoustic wave positioning error is less than 2 cm, the beam width is set to 15°; When the acoustic wave positioning error is greater than or equal to 2 cm, the beam width is set to 30°.

[0019] In a second aspect, a transformer partial discharge detection system based on multi-modal feature combination is provided, and the system is used to implement the transformer partial discharge detection method based on multi-modal feature combination as described in any one of the first aspects, including: A signal acquisition module, used to collect electromagnetic wave signals, pulse current signals and sound wave signals of the target transformer; A feature extraction module, used to extract the electromagnetic wave time-frequency matrix, the pulse current waveform parameters and the acoustic wave space coordinates from the electromagnetic wave signal, the pulse current signal and the acoustic wave signal respectively; A feature fusion module, used for performing multi-modal feature fusion on the electromagnetic wave time-frequency matrix, pulse current waveform parameters and acoustic wave space coordinates based on a dynamic weighting strategy to generate a joint feature tensor; The discharge detection module is used to input the joint feature tensor into a pre-built neural network model to identify the discharge type and optimize the positioning coordinates.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. The transformer partial discharge detection method based on multimodal feature combination provided by the present invention performs tensor fusion of multi-dimensional electromagnetic wave time-frequency matrix, pulse current waveform parameters and acoustic wave space coordinates, and realizes cross-modal association learning through a neural network model, which can eliminate information islands and effectively reduce the misjudgment rate; and realizes adaptive fusion of electromagnetic wave-current-acoustic wave signals through dynamic weight coefficients. In the oil-paper insulation discharge scenario, feature complementarity significantly improves the signal-to-noise ratio; 2. The present invention adopts a bidirectional feedback chain to significantly reduce the amount of data processing and significantly improve the processing speed in high-density discharge scenarios; 3. The lightweight convolutional neural network adopted in the present invention has lower single detection time consumption and false alarm rate than the method of running multiple detection methods in parallel. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings: Figure 1 is a flow chart of Embodiment 1 of the present invention; Figure 2 This is a system block diagram of Embodiment 4 of the present invention. DETAILED DESCRIPTION

[0022] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0023] Embodiment 1: Transformer partial discharge detection method based on multi-modal feature combination, such as Figure 1 As shown, the following steps are included: S1: Collect electromagnetic wave signals, pulse current signals and acoustic wave signals of the target transformer; S2: extracting electromagnetic wave time-frequency matrix, pulse current waveform parameters and acoustic wave space coordinates from electromagnetic wave signal, pulse current signal and acoustic wave signal respectively; S3: Based on the dynamic weighting strategy, multi-modal feature fusion is performed on the electromagnetic wave time-frequency matrix, pulse current waveform parameters and acoustic wave spatial coordinates to generate a joint feature tensor; S4: The joint feature tensor is input into the pre-built neural network model to identify the discharge type and optimize the localization coordinates.

[0024] In step S1, electromagnetic wave signals (high frequency, wide frequency band) are generally used to detect high-frequency transient faults such as partial discharge and arc, and can capture early hidden dangers such as insulation degradation and winding deformation. They are sensitive to high-frequency electromagnetic interference and are suitable for electrical faults. The pulse current signal (low frequency, high current) reflects the direct measurement of current and is less affected by mechanical vibration. The acoustic vibration signal can capture the mechanical vibration inside the transformer (such as loose windings, core deformation) or the mechanical waves generated by the bursting of bubbles in the oil. It is sensitive to mechanical faults and complements the electrical signal.

[0025] The present invention selects electromagnetic wave signals, pulse current signals and sound wave signals to characterize different physical quantities, and can cover a variety of faults through complementary physical characteristics.

[0026] In some examples, sensor devices that collect electromagnetic wave signals, pulse current signals, and sound wave signals can be installed in an integrated manner.

[0027] In some examples, a distributed layout can be used to take into account signal crosstalk (such as electromagnetic waves interfering with vibration signals) and limited installation locations (such as being unable to be close to the discharge point and the grounding wire at the same time).

[0028] For example, the process of collecting the electromagnetic wave signal, pulse current signal and sound wave signal of the target transformer is as follows: collecting the electromagnetic wave signal in the frequency band of 300MHz-1.5GHz through the ultra-high frequency sensor array installed on the transformer casing; collecting the 3MHz-30MHz pulse current signal through the high-frequency current transformer coupled to the neutral point grounding wire; and collecting the 20kHz-300kHz sound wave vibration signal through the piezoelectric sensor array arranged on the surface of the oil tank.

[0029] In step S2, the electromagnetic wave signal can be decomposed by wavelet packet decomposition to extract the sub-band energy. For example, the 1.2 GHz signal is decomposed into 16 layers by wavelet packet decomposition, and the size of the electromagnetic wave time-frequency matrix is , where 5 represents the preset frequency dimension, i.e. the number of frequency bands.

[0030] The pulse current waveform parameters extracted from the pulse current signal include pulse density, rise time and polarity. According to the pulse density, rise time and polarity, the pulse current characteristic vector can be constructed. For example, pulse density: N =120 pulses / ms; Rise Time: t r =8ns±0.5ns; Polarity: Positive pulse accounts for 85%; Pulse current characteristic vector .

[0031] In some examples, the spatial coordinates of the sound waves are calculated using a TDOA (Time Difference of Arrival) algorithm, or may be calculated using a DOA (Direction of Arrival) algorithm.

[0032] In some examples, the TDOA algorithm and the DOA algorithm can be combined to calculate the spatial coordinates of the acoustic wave, where the TDOA (time difference of arrival) algorithm is used for coarse positioning, and then fine-tuning is performed within the coarse positioning area to reduce the error. For example, the observation equations of the TDOA algorithm and the DOA algorithm are combined to construct a joint cost function, and then the global optimal solution is solved by an optimization algorithm, such as Levenberg-Marquardt, which is an iterative optimization algorithm for solving nonlinear least squares problems.

[0033] In step S3, the multimodal feature fusion expression based on the dynamic weighting strategy is: ; in, represents the joint feature tensor; , , Respectively represent the dynamic weight coefficients of electromagnetic wave signals, pulse current signals and acoustic wave signals, with a value range of [0,1]; Represents the electromagnetic wave time-frequency energy matrix, extracted by wavelet packet decomposition; It represents the pulse current characteristic vector, which is composed of the pulse density, rise time and polarity in the pulse current waveform parameters; Represents the spatial coordinates of the sound wave, which is calculated by the DOA algorithm and / or the TDOA algorithm; Represents tensor concatenation.

[0034] In some examples, the dynamic weight coefficient can be adjusted according to the signal-to-noise ratio of each signal. For example, the dynamic weight coefficient is determined by taking the sum of the dynamic weight coefficients of the electromagnetic wave signal, the pulse current signal, and the sound wave signal as a constraint of 1, and the larger the signal-to-noise ratio of the corresponding signal, the larger the dynamic weight coefficient assigned to the corresponding signal.

[0035] In some examples, the dynamic weight coefficient can also be adjusted according to the degree of signal distortion. The present invention can quantify the signal distortion by harmonic distortion rate (THD) or waveform similarity (such as cosine similarity) to obtain the degree of signal distortion. If a sensor signal is severely distorted, the corresponding weight can be reduced.

[0036] In step S4, the neural network model adopts a lightweight convolutional neural network consisting of an input layer, a hidden layer and an output layer; wherein the number of neurons in the input layer is equal to the sum of the dimensions of the vectors corresponding to the electromagnetic wave time-frequency matrix, the pulse current waveform parameters and the acoustic wave space coordinates; the hidden layer is composed of 2 fully connected layers, and the ReLU activation function is used; the number of neurons in the output layer is equal to the number of output discharge types, and the Softmax activation function is used.

[0037] For example, the size of the electromagnetic wave time-frequency matrix is , the number of eigenvectors (dimensions) of the expanded eigenvectors is 80. The pulse current eigenvector is constructed based on pulse density, rise time and polarity, so the dimension of its corresponding vector is 3. The acoustic wave spatial coordinates correspond to three axes, so the dimension of its corresponding vector is 3, so the number of neurons in the final input layer is 86.

[0038] Since the present invention outputs various discharge types and their corresponding probabilities, such as , which means that the probabilities of internal discharge, surface discharge, suspended discharge and noise are 0.72, 0.15, 0.08 and 0.05 respectively. If the probability of internal discharge is 72% > 60%, it is judged as internal discharge of oil-paper insulation.

[0039] In some examples, the expressions for optimizing the positioning coordinates are: ; in, Indicates the optimized positioning coordinates; Indicates the coarse positioning coordinates based on sound waves; Represents the electromagnetic wave positioning coordinates based on the electromagnetic wave attenuation model; Represents the correction coefficient, which is determined by fitting historical data.

[0040] It should be noted that the above rough positioning coordinates based on sound waves may also be replaced by coordinates obtained by further processing the rough positioning coordinates based on sound waves.

[0041] In some examples, the electromagnetic wave attenuation model is expressed as: ; Where c represents the electromagnetic wave signal strength at distance d; Indicates the initial signal strength at the discharge source; It represents the attenuation coefficient of transformer oil-paper composite medium, and its value range is 0.05-0.1.

[0042] Embodiment 2: A transformer partial discharge detection method based on the combination of multi-modal features. Compared with Embodiment 1, the present invention also performs time delay compensation and phase alignment on electromagnetic wave signals, pulse current signals and sound wave signals.

[0043] In some examples, delay compensation is implemented by the following formula: ; in, Indicates the propagation time difference between electromagnetic waves and sound waves (unit: seconds), which is used to compensate for signal synchronization errors; Indicates the speed of sound in transformer oil, which is 1400m / s; Represents the propagation speed of electromagnetic waves, taking the speed of light in vacuum as ; It represents the straight-line distance from the discharge source to the sensor (piezoelectric sensor array) (unit: meter), which is preliminarily estimated by the ultrasonic flight time.

[0044] Example 3: A transformer partial discharge detection method based on the combination of multimodal features. Compared with Example 1, the present invention also applies a bidirectional feedback optimization chain to optimize the extraction process of the electromagnetic wave time-frequency matrix, pulse current waveform parameters and acoustic wave space coordinates, which can reduce the data processing amount to a certain extent and improve detection efficiency and accuracy.

[0045] In some examples, the present invention can be based on the pulse density of the pulse current signal. N Dynamically adjust the number of wavelet packet decomposition layers: when N When >100 pulses / ms, the number of decomposition layers is ≤12; when 50 pulses / ms ≤ N ≤100 pulses / ms, 12<decomposition level<16; when N When <50 pulses / ms, the number of decomposition layers is ≥16.

[0046] In some examples, when 50 pulses / ms ≤ N When ≤100 pulses / ms, a linear interpolation model can be used to select the number of decomposition layers.

[0047] For example, the linear interpolation model is expressed as: ; in, Indicates the number of decomposition levels; Represents the rounding function.

[0048] In some examples, the present invention can also adjust the pulse current analysis time window to 1.2-1.5 times of the high-frequency energy concentration period based on the high-frequency energy concentration period in the electromagnetic wave time-frequency matrix. For example, extract the energy mutation time point of the 1.1-1.3GHz frequency band in the electromagnetic wave time-frequency matrix , ; Set the pulse current analysis window to .

[0049] In some examples, the present invention can also be based on the signal-to-noise ratio of the electromagnetic wave signal. Dynamically adjust pulse amplitude threshold : .

[0050] In some examples, the present invention can also locate coordinates based on electromagnetic waves , constraining the spatial search range of the TDOA algorithm to The acoustic wave spatial coordinates are calculated using the TDOA algorithm or a method combining the TDOA algorithm with the DOA algorithm.

[0051] For example, when electromagnetic wave positioning reliability When the acoustic wave search step is set to 1cm; when When the acoustic wave search step is set to 3cm.

[0052] In some examples, the present invention may also select a matching acoustic wave propagation attenuation model library according to the pulse current polarity ratio.

[0053] In some examples, the present invention can also be based on rough positioning coordinates based on sound waves. , adjust the beamforming direction of the UHF sensor array so that the main lobe points to cm area.

[0054] For example, the adjustment process of the beam forming direction is: when the acoustic wave positioning error is less than 2 cm, the beam width is set to 15°; when the acoustic wave positioning error is greater than or equal to 2 cm, the beam width is set to 30°.

[0055] In some examples, the present invention can also be based on the length of the sound wave propagation path , correct the propagation distance parameter in the electromagnetic wave attenuation model for: .

[0056] Experimental verification: 1. Contrast Design 1. Traditional single detection method UHF method: only uses ultra-high frequency electromagnetic wave signals (300MHz-1.5GHz); HFCT method: only uses high frequency pulse current signals (3MHz-30MHz); AE method: only uses acoustic emission signals (20kHz-300kHz).

[0057] 2. Parallel-Multi UHF, HFCT and AE are operated independently, and any method will output the result when it detects discharge.

[0058] 3. The present invention (MMF-Joint) Dynamic fusion of multimodal features + lightweight CNN model.

[0059] 2. Experimental Setup 1. Test scenario Discharge type: internal discharge of oil-paper insulation (PD1), surface discharge (PD2), suspended discharge (PD3); Noise environment: the signal-to-noise ratio (SNR) is divided into three levels: 10dB (high noise), 20dB (medium noise), 30dB (low noise); Discharge density: 50 / 100 / 150pulses / ms; Positioning difficulty: the discharge point is located at different positions of the transformer (near the casing, winding gap, near the core).

[0060] 2. Dataset Real transformer fault data (500 groups) + simulation data (500 groups, covering extreme scenarios); data annotation: discharge type label + high-precision positioning coordinates (laser calibration benchmark).

[0061] 3. Evaluation Metrics Detection accuracy (%): the proportion of correctly identified discharge types; Positioning error (mm): the Euclidean distance of |predicted coordinates-true coordinates|; False alarm rate (%): the probability of misjudging discharge when there is no discharge; Single detection time (ms): the time from signal input to result output; Signal-to-noise ratio gain (dB): the improvement of the SNR of the fused signal compared to a single modality.

[0062] 3. Experimental Results The detection accuracy comparison (SNR=20dB) is shown in Table 1. The MMF-Joint of the present invention improves the average accuracy by more than 6.8% through feature complementarity, especially in complex discharge types (PD3).

[0063] Table 1 Comparison results of detection accuracy

[0064] The comparison results of positioning errors are shown in Table 2. The dynamic coordinate optimization algorithm in the present invention reduces the positioning error by 52.7%, and the improvement is particularly obvious in the area where electromagnetic waves are prone to attenuation (near the iron core).

[0065] Table 2 Comparison results of positioning errors (unit: mm)

[0066] The comparison results of anti-interference ability are shown in Table 3. The dynamic weight strategy adopted by the present invention effectively suppresses the interference of low signal-to-noise ratio signals, and the false alarm rate is reduced by 60%-80%. Table 3 Comparison results of anti-interference ability

[0067] The comparison results of detection efficiency are shown in Table 4. The lightweight CNN model adopted by the present invention makes the MMF-Joint time consumption only 47.8% of the parallel method and reduces the memory usage by 51.6%.

[0068] Table 4 Detection efficiency comparison results

[0069] The results of the effectiveness of dynamic weights are shown in Table 5. The dynamic weight strategy adopted by the present invention can maintain high robustness when the signal quality is uneven through adaptive adjustment of the signal-to-noise ratio.

[0070] Table 5 Dynamic weight validity results

[0071] Embodiment 4: A transformer partial discharge detection system based on multi-modal feature combination, the system is used to implement a transformer partial discharge detection method based on multi-modal feature combination as in any one of Embodiments 1-3, such as Figure 2 As shown, it includes a signal acquisition module, a feature extraction module, a feature fusion module and a discharge detection module.

[0072] Among them, the signal acquisition module is used to collect the electromagnetic wave signal, pulse current signal and sound wave signal of the target transformer; the feature extraction module is used to extract the electromagnetic wave time-frequency matrix, pulse current waveform parameters and sound wave space coordinates from the electromagnetic wave signal, pulse current signal and sound wave signal respectively; the feature fusion module is used to perform multimodal feature fusion on the electromagnetic wave time-frequency matrix, pulse current waveform parameters and sound wave space coordinates based on the dynamic weighting strategy to generate a joint feature tensor; the discharge detection module is used to input the joint feature tensor into the pre-built neural network model to identify the discharge type and optimize the positioning coordinates.

[0073] Working principle: The present invention performs tensor fusion on the multi-dimensional electromagnetic wave time-frequency matrix, pulse current waveform parameters and acoustic wave space coordinates, and implements cross-modal association learning through a neural network model, which can eliminate information islands and effectively reduce the misjudgment rate; and realizes adaptive fusion of electromagnetic wave-current-acoustic wave signals through dynamic weight coefficients. In the oil-paper insulation discharge scenario, feature complementarity significantly improves the signal-to-noise ratio.

[0074] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0075] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0076] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0077] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0078] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A transformer partial discharge detection method based on multi-modal feature combination, characterized in that: The following steps are involved: Collect electromagnetic wave signals, pulse current signals and sound wave signals of the target transformer; Extracting the electromagnetic wave time-frequency matrix, the pulse current waveform parameters and the acoustic wave space coordinates from the electromagnetic wave signal, the pulse current signal and the acoustic wave signal respectively; Based on a dynamic weighting strategy, multi-modal feature fusion is performed on the electromagnetic wave time-frequency matrix, pulse current waveform parameters and acoustic wave space coordinates to generate a joint feature tensor; The joint feature tensor is input into a pre-built neural network model to identify the discharge type and optimize the positioning coordinates.

2. The transformer partial discharge detection method based on multi-modal feature combination according to claim 1 is characterized in that: The process of collecting the electromagnetic wave signal, pulse current signal and sound wave signal of the target transformer is specifically as follows: The electromagnetic wave signals in the frequency band of 300MHz-1.5GHz are collected by the UHF sensor array installed on the transformer housing; The 3MHz-30MHz pulse current signal is collected by a high-frequency current transformer coupled to the neutral point grounding wire; Furthermore, the acoustic vibration signal of 20kHz-300kHz is collected by a piezoelectric sensor array arranged on the surface of the oil tank.

3. The transformer partial discharge detection method based on multi-modal feature combination according to claim 1 is characterized in that: The multimodal feature fusion expression based on dynamic weighting strategy is: ； in, represents the joint feature tensor; , , Respectively represent the dynamic weight coefficients of electromagnetic wave signals, pulse current signals and acoustic wave signals, with a value range of [0,1]; Represents the electromagnetic wave time-frequency energy matrix, extracted by wavelet packet decomposition; It represents the pulse current characteristic vector, which is composed of the pulse density, rise time and polarity in the pulse current waveform parameters; Represents the spatial coordinates of the sound wave, which is calculated by the DOA algorithm and / or the TDOA algorithm; Represents tensor concatenation.

4. The transformer partial discharge detection method based on multi-modal feature combination according to claim 3 is characterized in that: The determination process of the dynamic weight coefficient is specifically as follows: The sum of the dynamic weight coefficients of the electromagnetic wave signal, the pulse current signal and the sound wave signal is constrained to be 1, and the greater the signal-to-noise ratio of the corresponding signal, the greater the dynamic weight coefficient allocated to the corresponding signal.

5. The transformer partial discharge detection method based on multi-modal feature combination according to claim 1 is characterized in that: The neural network model adopts a lightweight convolutional neural network consisting of an input layer, a hidden layer and an output layer; The number of neurons in the input layer is equal to the sum of the dimensions of the vectors corresponding to the electromagnetic wave time-frequency matrix, the pulse current waveform parameters and the acoustic wave space coordinates; The hidden layer consists of 2 fully connected layers and uses the ReLU activation function; The number of neurons in the output layer is equal to the number of output discharge types, and a Softmax activation function is used.

6. The transformer partial discharge detection method based on multi-modal feature combination according to claim 1 is characterized in that: The expression of the optimized positioning coordinates is: ； in, Indicates the optimized positioning coordinates; Indicates the coarse positioning coordinates based on sound waves; Represents the electromagnetic wave positioning coordinates based on the electromagnetic wave attenuation model; Represents the correction coefficient, which is determined by fitting historical data.

7. The transformer partial discharge detection method based on multi-modal feature combination according to claim 6 is characterized in that: The expression of the electromagnetic wave attenuation model is: ； in, Indicates the electromagnetic wave signal strength at distance d; Indicates the initial signal strength at the discharge source; It represents the attenuation coefficient of transformer oil-paper composite medium, and its value range is 0.05-0.

1.

8. The transformer partial discharge detection method based on multi-modal feature combination according to any one of claims 1 to 7, characterized in that: The method further includes: The electromagnetic wave signal, the pulse current signal and the sound wave signal are subjected to time delay compensation and / or phase alignment.

9. The transformer partial discharge detection method based on multi-modal feature combination according to any one of claims 1 to 7, characterized in that: The method further includes: According to the pulse density of the pulse current signal N Dynamically adjust the number of wavelet packet decomposition layers: when N When >100 pulses / ms, the number of decomposition layers is ≤12; When 50 pulses / ms ≤ N When ≤100 pulses / ms, 12<decomposition layers<16; when N When <50 pulses / ms, the number of decomposition layers is ≥16.

10. The transformer partial discharge detection method based on multi-modal feature combination according to any one of claims 1 to 7, characterized in that: The method further includes: Based on the high-frequency energy concentration period in the electromagnetic wave time-frequency matrix, the pulse current analysis time window is adjusted to 1.2-1.5 times the high-frequency energy concentration period; and / or, according to the signal-to-noise ratio of the electromagnetic wave signal Dynamically adjust pulse amplitude threshold : 。 11. The transformer partial discharge detection method based on multi-modal feature combination according to any one of claims 1 to 7, characterized in that: The method further includes: Positioning coordinates based on electromagnetic waves , constraining the spatial search range of the TDOA algorithm to cm; and / or, selecting a matching acoustic wave propagation attenuation model library according to the pulse current polarity ratio; The acoustic wave spatial coordinates are calculated using a TDOA algorithm or a method combining a TDOA algorithm with a DOA algorithm.

12. The transformer partial discharge detection method based on multi-modal feature combination according to any one of claims 1 to 7, characterized in that: The method further includes: Based on the rough positioning coordinates based on sound waves , adjust the beamforming direction of the UHF sensor array so that the main lobe points to cm area; and / or, based on the acoustic wave propagation path length , correct the propagation distance parameter in the electromagnetic wave attenuation model for: .

13. The transformer partial discharge detection method based on multi-modal feature combination according to claim 12 is characterized in that: The adjustment process of the beamforming direction is: When the acoustic wave positioning error is less than 2 cm, the beam width is set to 15°; When the acoustic wave positioning error is greater than or equal to 2 cm, the beam width is set to 30°.

14. A transformer partial discharge detection system based on multi-modal feature combination, characterized in that: The system is used to implement the transformer partial discharge detection method based on multi-modal feature combination as described in any one of claims 1 to 13, comprising: A signal acquisition module, used to collect electromagnetic wave signals, pulse current signals and sound wave signals of the target transformer; A feature extraction module, used to extract the electromagnetic wave time-frequency matrix, the pulse current waveform parameters and the acoustic wave space coordinates from the electromagnetic wave signal, the pulse current signal and the acoustic wave signal respectively; A feature fusion module, used for performing multi-modal feature fusion on the electromagnetic wave time-frequency matrix, pulse current waveform parameters and acoustic wave space coordinates based on a dynamic weighting strategy to generate a joint feature tensor; The discharge detection module is used to input the joint feature tensor into a pre-built neural network model to identify the discharge type and optimize the positioning coordinates.

Images