A method for collecting and processing wide frequency band signals of partial discharge of power equipment

Abstract

This invention discloses a method for acquiring and processing wideband partial discharge signals of power equipment, relating to the field of partial discharge detection technology for high-voltage power equipment. The method includes: acquiring three types of partial discharge signals by multi-source signal acquisition through a fusion sensor; performing normalization, sliding window segmentation, and grayscale image mapping on the three types of partial discharge signals respectively to form a three-channel fused image tensor; inputting the fused image tensor into a partial discharge recognition model for partial discharge type identification; calculating discharge characteristic parameters such as maximum value, minimum value, average value, instantaneous value, number of discharges, and discharge duration based on a rolling time window, and dynamically displaying them. This invention improves the accuracy of partial discharge recognition through multi-source fusion signal acquisition and channel attention mechanism, and can also reflect the changing trend and key characteristics of partial discharge signals in real time, making it easier for maintenance personnel to intuitively grasp the equipment status.

Description

Technical Field

[0001] This invention relates to the field of partial discharge detection technology for high-voltage power equipment, and in particular to a method for acquiring and processing wideband partial discharge signals of power equipment. Background Technology

[0002] Partial discharge is a non-full-bridge breakdown phenomenon that occurs in localized areas of the insulation system of power equipment. Its occurrence often indicates insulation defects in the equipment. Partial discharge leads to the gradual deterioration of insulation materials, which may eventually cause equipment failure or even power accidents. Therefore, online detection and condition assessment of partial discharge in power equipment is a key technical aspect for achieving safe operation and precise maintenance of power systems.

[0003] Traditional partial discharge signal identification relies on manual extraction of statistical features in the time and frequency domains, combined with threshold criteria or traditional classifiers for discrimination. Limited by the subjectivity of manual feature selection and the increasing complexity of data, traditional methods have limited performance in situations with complex noise backgrounds and diverse signal types. In recent years, deep learning models such as Convolutional Neural Networks (CNNs) can automatically learn multi-level feature representations from raw data, greatly improving the accuracy and robustness of partial discharge signal identification. However, deep learning models have high requirements for training data volume and generalization ability. In the field of partial discharge in power equipment, there are practical challenges such as limited sample size and class imbalance, affecting the actual application effect of the models. Furthermore, the requirements for real-time performance, dynamism, and visualization of detection terminals in practical engineering applications are increasingly increasing. Detection systems not only need to capture various types of partial discharge signals with high precision and sensitivity, but also need to extract and dynamically display key features such as the maximum, minimum, average, instantaneous values, number of discharges, and discharge duration in real time, providing intuitive and comprehensive data support for equipment condition assessment and maintenance decisions.

[0004] In summary, existing technologies face the following main challenges in the field of partial discharge detection: single sensing methods have limited anti-interference and signal perception capabilities, making them difficult to adapt to complex and ever-changing application scenarios; signal feature extraction and intelligent recognition methods need to further improve their accuracy and generalization capabilities, and reduce their reliance on a large number of labeled samples; the feature extraction and dynamic visualization functions of detection terminals urgently need to be strengthened to achieve comprehensive assessment and intelligent early warning of equipment status. Summary of the Invention

[0005] The purpose of this invention is to provide a method for acquiring and processing wideband partial discharge signals of power equipment, thereby solving the problems of the prior art.

[0006] To achieve the above objectives, the present invention provides a method for acquiring and processing broadband partial discharge signals of power equipment, comprising the following steps: Step 100: Detect partial discharge signals using a fusion sensor to obtain three types of partial discharge signals; The fusion sensor includes an ultrasonic signal acquisition module, a TEV signal acquisition module, and an ultra-high frequency signal acquisition module; Step 200: Preprocess and grayscale map the three partial discharge signals respectively, and then stitch the three processed partial discharge signals together in the channel dimension to obtain a three-channel input tensor. Step 300: Input the three-channel input tensor into the partial discharge identification model and output the predicted partial discharge type; Step 400: Based on the identified partial discharge type, calculate and display the discharge characteristic parameters in real time within the rolling time window. Discharge characteristic parameters include maximum value, minimum value, average value, instantaneous value, number of discharges, and discharge duration.

[0007] Furthermore, the ultrasonic signal acquisition module employs a miniaturized ultrasonic MEMS microphone to acquire partial discharge ultrasonic signals in the 20kHz to 300kHz frequency band. The ultrasonic signal acquisition module couples the mechanical vibration waves on the surface of the power equipment through a rigid contact waveguide structure and converts the analog sound wave signal into a single-bit digital signal through a built-in Σ-Δ modulator.

[0008] Furthermore, the TEV signal acquisition module uses capacitor plates to couple partial discharge signals to acquire partial discharge TEV signals in the 3MHz to 100MHz frequency band; the TEV signal acquisition module captures transient voltage pulses on the surface of the metal casing of the equipment through the principle of electric field induction, and uses a two-stage differential amplifier for signal amplification and an elliptic function filter for bandpass filtering.

[0009] Furthermore, the UHF signal acquisition module uses a log-periodic radiating antenna as a sensing unit to acquire UHF electromagnetic signals of partial discharge in the 300MHz to 1.5GHz frequency band. The UHF signal acquisition module converts the transient electromagnetic radiation energy generated during partial discharge into radio frequency electrical signals through electromagnetic coupling between the antenna and the spatial electromagnetic field, and performs signal processing through a two-stage amplification circuit. The front stage uses a low-noise amplifier for fixed gain amplification, and the rear stage uses a variable gain amplifier for dynamic gain compression.

[0010] Furthermore, the steps in step 200 to obtain the three-channel input tensor include: Step 201: Normalize the sampling sequence of the original one-dimensional partial discharge signal to obtain a normalized signal sequence; Step 202: The normalized signal sequence is segmented using a sliding window, and the segmented signal data are combined into a signal matrix; Step 203: Map each value in the signal matrix to a grayscale image and adjust its size to generate a two-dimensional grayscale image; Step 204: Map the ultrasonic signal, TEV signal and UHF signal to their corresponding two-dimensional grayscale images, and stitch them together in the channel dimension to form a three-channel input tensor.

[0011] Furthermore, the partial discharge recognition model includes an input layer, a convolutional layer, an activation function layer, a pooling layer, a channel attention fusion module, a cross-channel fusion layer, a global average pooling layer, and an SVM classification layer connected in sequence. The input layer receives the three-channel fused image tensor and performs preprocessing such as normalization and mean removal; the convolutional pooling layer extracts local features using convolutional kernels; the activation function layer introduces nonlinearity using the ReLU activation function; the pooling layer performs downsampling through pooling operations; the channel attention fusion module performs adaptive weight allocation for features of different modalities; the cross-channel fusion layer performs cross-channel feature fusion through 1×1 convolution; the global average pooling layer generates a global feature vector through global average pooling; and the SVM classification layer performs multi-class classification on the global feature vector and outputs the partial discharge type identification result.

[0012] Furthermore, the processing steps of the channel attention fusion module include the following: Step 301: Perform global average pooling on the original feature map of each channel to obtain the global average response value of each channel; The original feature map is obtained by passing a three-channel input tensor through convolutional layers, activation function layers, and pooling layers; Step 302: Input the global average response value of each channel into the two-layer fully connected network, and generate channel weights through the Sigmoid activation function; Step 303: Multiply the channel weights with the original feature maps of the corresponding channels one by one to obtain the weighted feature maps.

[0013] Furthermore, the SVM classification layer is specifically a Support Vector Machine (SVM) classifier, and the training objective of the SVM classifier is to solve the following optimization problem: ; The constraints are as follows: ; in, This is the penalty parameter, used to control the balance between maximizing the classification margin and penalizing misclassification. For the first Relaxed variables for each training sample, allowing some samples to violate the interval constraint. For the first The class labels of each training sample. and These are the classification weight vector and classification bias parameter to be optimized, respectively. Indicates transpose. Let represent the global feature vector of the i-th training sample.

[0014] Furthermore, the method for calculating the number of discharges is as follows: the number of times the amplitude of the detection signal rises from below the preset discharge criterion threshold to above the preset discharge criterion threshold, and the number of transitions is counted as the number of discharge events.

[0015] Furthermore, the discharge duration is calculated as follows: the number of sampling points whose signal amplitude is higher than the preset discharge criterion threshold is counted, and the product of the number of sampling points and the sampling interval is taken as the discharge duration.

[0016] Therefore, the present invention employs the above-mentioned method for acquiring and processing wideband partial discharge signals of power equipment, which has the following beneficial effects: (1) A multi-source fusion signal acquisition scheme is adopted, which integrates ultrasonic, TEV and UHF partial discharge detection methods into one, and combines synchronous acquisition and fusion mechanism to effectively complement the advantages and disadvantages of various sensors, achieve high sensitivity of partial discharge signals in a wider frequency range, greatly improve anti-interference ability and adaptability, and meet the application needs of complex power equipment operating environment.

[0017] (2) A signal data mapping method based on normalization and window segmentation is adopted, and a multi-modal channel attention fusion mechanism is introduced into the deep model. On the basis of unified three-channel grayscale image expression, dynamic weighting is performed according to the contribution of different sensors in different discharge modes, highlighting key modal information and suppressing redundant interference, thereby realizing adaptive collaborative processing of multi-source information and significantly enhancing the discriminative power of feature fusion.

[0018] (3) An improved CNN-SVM intelligent recognition model was constructed, which combines the advantages of deep feature extraction of convolutional neural network, the modal adaptive enhancement capability of channel attention and the high robust classification performance of support vector machine, and realizes high-precision recognition and strong generalization capability of partial discharge type. It is particularly suitable for field application scenarios with limited samples, low signal-to-noise ratio and complex interference, and can provide more reliable technical support for equipment status assessment and operation and maintenance.

[0019] (4) A dynamic feature extraction and real-time visualization mechanism based on a rolling time window is introduced, which can not only reflect the changing trend and key features of partial discharge signals in real time, but also make it easier for maintenance personnel to intuitively grasp the equipment status and realize early warning and efficient management of abnormal discharge.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for acquiring and processing broadband partial discharge signals of power equipment according to the present invention; Figure 2 This is a flowchart of the ultrasonic signal acquisition module of the present invention; Figure 3 This is a flowchart of the TEV signal acquisition module of the present invention; Figure 4 This is a flowchart of the ultra-high frequency signal acquisition module of the present invention; Figure 5 A flowchart illustrating the preprocessing and grayscale mapping of the partial discharge signal in this invention; Figure 6 This is a schematic diagram of the partial discharge identification model of the present invention. Detailed Implementation

[0022] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely illustrates selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0023] Please see Figure 1-6 A method for acquiring and processing broadband partial discharge signals of power equipment includes the following steps: Step 100: Detect partial discharge signals using a fusion sensor to obtain three types of partial discharge signals; The fusion sensor includes an ultrasonic signal acquisition module, a TEV signal acquisition module, and an ultra-high frequency signal acquisition module.

[0024] The ultrasonic signal acquisition module employs a miniaturized ultrasonic MEMS microphone to acquire ultrasonic signals from partial discharge in the 20-300kHz range. Mechanical vibration waves from the surface of the power equipment are coupled through a rigid contact waveguide structure. Its optimized acoustic cavity significantly improves sound pressure sensitivity, effectively capturing transient stress waves generated by partial discharge. The MEMS chip incorporates a Σ-Δ modulator to convert analog acoustic signals into single-bit digital PDM (Pulse Density Modulation) signals. This architecture uses oversampling technology to push quantization noise to higher frequencies, and combined with subsequent digital filtering, achieves high signal-to-noise ratio signal reconstruction, avoiding the electromagnetic interference problems introduced by traditional analog amplification links.

[0025] The TEV signal acquisition module uses capacitor plates to couple partial discharge signals, acquiring TEV signals ranging from 3MHz to 100MHz. It captures transient voltage pulses on the surface of the device's metal casing through the principle of electric field induction. The electrode surfaces are coated with a nanoscale dielectric coating, achieving electrical isolation protection while maintaining high coupling efficiency. Subsequent signal amplification is achieved using a two-stage differential amplifier, followed by bandpass filtering with an elliptic function filter. The pre-amplifier stage employs a high input impedance design to reduce the signal source load effect.

[0026] The UHF signal acquisition module is used to acquire UHF signals ranging from 300MHz to 1.5GHz generated by partial discharge radiation. The module employs a log-periodic radiating antenna as the sensing unit, converting the transient electromagnetic radiation energy generated during partial discharge into radio frequency signals through electromagnetic coupling between the antenna and the spatial electromagnetic field. The log-periodic radiating structure, through a gradual design of varying element length and spacing, achieves stable impedance matching and directional stability over a wide frequency range. Figure 1 To ensure consistency and effective reception of UHF partial discharge signals across different frequency bands, microwave absorbing material is loaded within the antenna cavity to suppress multipath effects and standing wave interference caused by reflected electromagnetic waves from the environment, thereby improving the stability and accuracy of UHF signal acquisition. The received RF signal is processed by a two-stage amplification circuit. The first stage uses a low-noise amplifier with fixed gain amplification to reduce the system noise figure and improve the detectability of weak discharge signals. The second stage uses a variable gain amplifier, controlled by a real-time signal power detection circuit, to dynamically compress strong interference signals and prevent saturation distortion in subsequent signal processing circuits.

[0027] The fusion sensor communicates with the partial discharge detection terminal via a single Ethernet communication link.

[0028] Specifically, the three partial discharge signals are uniformly input into the microprocessor chip. The microprocessor chip, as the core of signal processing and control, is used to perform synchronous sampling control, data buffering, data processing, and communication management of the partial discharge signals, and then transmits the processed data to the partial discharge detection terminal.

[0029] Step 200: Preprocess and grayscale map the three partial discharge signals respectively, and then stitch the three processed partial discharge signals together in the channel dimension to obtain a three-channel input tensor. In an embodiment of the present invention, taking a partial discharge signal as an example, the specific steps are as follows: Let the sampling sequence of the original one-dimensional partial discharge signal be: ; in, Indicates the first The amplitude of each partial discharge signal, where N is the total number of partial discharge signals.

[0030] Step 201: Normalize the sampling sequence of the original one-dimensional partial discharge signal to obtain a normalized signal sequence; To reduce the impact of amplitude variations, linear normalization is used. The specific calculation formula is as follows: ; in, and These represent the minimum and maximum amplitudes of the partial discharge signal, respectively. Let be the normalized amplitude of the i-th partial discharge signal. It is compressed to the [0,1] interval. The normalized signal sequence is: .

[0031] Step 202: The normalized signal sequence is segmented using a sliding window, and the segmented signal data are combined into a signal matrix. Let the length of each window be... There are a total of There are 1 window, and the window number is... Each signal data is represented as: ; Indicates the first In the window, the first The normalized signal value corresponding to each position, Represents the normalized i-th The amplitude of a partial discharge signal.

[0032] Generate a signal matrix based on the set window. : ; in, Each row corresponds to a window, reflecting the numerical distribution of a local time segment.

[0033] Step 203: Map each value in the signal matrix to a grayscale image and adjust its size to generate a two-dimensional grayscale image; Mapping each value in the signal matrix to a grayscale image involves scaling the normalized values ​​to [0, 255] to obtain the grayscale value of the corresponding pixel. The grayscale value of the pixel at each position in the initial two-dimensional grayscale image is defined as follows: ; This represents the grayscale value of the pixel at the k-th row and j-th position in the initial two-dimensional grayscale image.

[0034] Adjust the size of the initial 2D grayscale image. ,in , The final two-dimensional grayscale image is obtained, and the grayscale value of the pixel corresponding to each position in the two-dimensional grayscale image is represented as: ; in, Represents coordinates in a two-dimensional grayscale image The pixel grayscale value at that location.

[0035] Step 204: Map the ultrasonic signal, TEV signal and UHF signal to their corresponding two-dimensional grayscale images, and stitch them together in the channel dimension to form a three-channel input tensor; The three-channel input tensor is represented as: Three-channel input tensor in spatial coordinates Place, No. The pixel values ​​in each channel are represented as follows: ;in, These correspond to the channels for ultrasonic signals, TEV signals, and UHF signals, respectively.

[0036] Step 300: Input the three-channel input tensor into the partial discharge identification model and output the predicted partial discharge type; (1) Input layer; The partial discharge identification model accepts the aforementioned three-channel input tensor and performs preprocessing such as normalization and mean removal to eliminate numerical scale differences between different channels and samples, suppress the influence of background bias on the feature extraction process, and make the distribution of each input feature more stable. This accelerates the model's convergence speed and improves the accuracy and robustness of subsequent feature extraction and classification. The specific identification process of the partial discharge identification model to obtain input features includes: (2) Convolutional layer; Convolutional layers extract local features through convolutional kernels, and the calculation formula is as follows: ; in, Indicates the first Number of channels in the layer These are the convolution kernel weight parameters. These are the kernel bias parameters. For the first Layer in position The convolution outputs feature values. Indicates the first Layer Each channel is located in Input features, and These represent the height and width of the convolution kernel, respectively.

[0037] (3) Activation function layer; After the convolution operation, the ReLU activation function is used to introduce non-linearity, and the calculation formula is as follows: ; in, This indicates the operation of retrieving the maximum value. Indicates the first Layer in position The activation value, which is obtained by setting negative values ​​to zero and keeping positive values, introduces nonlinearity to enhance the model's ability to express complex features.

[0038] (4) Pooling layer; Pooling layers spatially downsample the convolutional feature maps, reducing feature dimensionality and enhancing translation invariance. This embodiment uses max pooling, calculated as follows: ; in, Indicates the first [unit] after pooling Layer in position eigenvalues, This indicates that the corresponding input feature map in the pooling operation is in the range of... The set of feature values ​​within the local neighborhood of the starting position, typically or The rectangular neighborhood, Indicates the first Layer in position activation value, This indicates the pooling step size.

[0039] This embodiment employs a three-layer convolutional pooling structure, with the number of convolutional kernels being 64, 128, and 256 respectively, and the kernel size being [missing value]. The pooling core size is .

[0040] The original feature map is finally obtained through convolutional layers, activation function layers, and pooling layers. .

[0041] (5) Channel attention fusion module; To fully utilize multi-source modal information, a channel attention mechanism is introduced after convolutional feature extraction to adaptively weight features of different modalities.

[0042] Step 301: Perform global average pooling on the original feature map of each channel: ; in, Indicates the first The global average response value of each channel. Represents the original feature map. and These represent the height and width of the feature map, respectively. This represents the total number of channels. Global average pooling is performed on all channels in the feature map to obtain the global average response value for each channel.

[0043] Step 302: The global average response value of each channel is used to generate the channel weights of each channel through two fully connected layers and a nonlinear mapping. The calculation formula is as follows: ; in, These are the learnable first and second channel weight matrices, For the Sigmoid function, For the first The channel weights of each channel, and ReLU is the ReLU activation function.

[0044] Step 303: Multiply the channel weights by the original feature maps of the corresponding channels one by one to obtain the weighted feature maps. The calculation formula is as follows: ; in, For the weighted number of Each channel is located in The model can enhance modal features that contribute more to the identification of partial discharge types and suppress redundant or highly interfering modal features through adaptive weight allocation.

[0045] (6) Cross-channel fusion layer; The channel attention-weighted feature maps are further fused across channels using 1×1 convolutions: ; in, and They are respectively The weights and biases of the convolution kernel, This indicates that after cross-channel linear fusion, in spatial location First Characteristic response values ​​of each output channel , Indicates the number of output channels.

[0046] (7) Global average pooling layer; The feature map after cross-channel fusion is then subjected to global average pooling to compress the spatial dimension into a single value, generating a global feature vector. The formula for calculating the global average feature value of each channel in the global feature vector is as follows: ; in, Indicates the first The global average eigenvalues ​​of each channel, and the global eigenvectors are represented as follows: , This indicates transpose.

[0047] (8) SVM classification layer; The SVM classification layer is specifically a Support Vector Machine classifier. For a binary classification problem, the decision function is as follows: ; in, For the classification weight vector, For classification bias parameters, This represents the transpose, used to calculate the inner product between the weight vector and the input feature vector. This represents the global feature vector of the support vector machine relative to the input. The output value of the discriminant function is used to measure the discrimination result of whether a sample belongs to different categories.

[0048] The training objective of a support vector machine classifier is to solve the following optimization problem: ; The constraints are as follows: ; in, This is the penalty parameter, used to control the balance between maximizing the classification margin and penalizing misclassification. For the first Relaxed variables for each training sample, allowing some samples to violate the interval constraint. For the first The class labels of each training sample. and These are the classification weight vector and classification bias parameter to be optimized, respectively. Let represent the global feature vector of the i-th training sample.

[0049] In multi-class partial discharge identification scenarios, the support vector machine classifier employs a one-to-many strategy for multi-class training. Specifically, assuming there are L classes of partial discharge types to be identified, L binary support vector machines are trained. The p-th classifier labels samples of class p as positive and samples of the other classes as negative. During classification, the input samples are fed into the L classifiers respectively, resulting in L decision scores. The class with the highest score is selected as the final identification result.

[0050] In embodiments of the present invention, the partial discharge types to be identified by the system include three categories: tip discharge, air gap discharge, and surface discharge. For a certain set of acquired signals: after CNN extracts features, the output scores of the SVM classification layer are [tip discharge: 0.85; air gap discharge: 0.10; surface discharge: 0.05]. Since tip discharge has the highest score, the final prediction result of the partial discharge identification model is tip discharge.

[0051] Through the above outputs, the system not only completes the classification of partial discharge types, but also combines the score information to determine the confidence level of the prediction, providing a basis for the insulation status assessment of power equipment.

[0052] Step 400: Based on the identified partial discharge type, calculate and display the discharge characteristic parameters in real time within the rolling time window.

[0053] To achieve efficient online monitoring of partial discharge, feature extraction is required from the three partial discharge signals acquired by the fusion sensor, and the discharge characteristic parameters need to be displayed in real time on the partial discharge detection terminal interface. The discharge characteristic parameters include six items: maximum value, minimum value, average value, instantaneous value, number of discharges, and discharge duration. To ensure the timeliness and dynamism of the discharge characteristic parameters, a rolling time window method is used for real-time calculation.

[0054] Let the sampling interval of the signal be... The length of the scrolling time window is At the current time At that time, the time window range is This time window contains There are 1 sampling points, and the sampled values ​​are recorded sequentially as follows: The sampling time for each sampling point is , .in, This represents the number of sampling points within the current time window. This is the discharge criterion threshold.

[0055] (1) Maximum value The maximum value of the signal amplitude within this time window is represented by the following formula: ; (2) Minimum value This represents the minimum signal amplitude within the time window, calculated using the following formula: ; (3) Average value The average value of the signal at all sampling points within the time window is calculated using the following formula: ; (4) Instantaneous value This refers to the latest acquired signal value at the current moment, that is, the value of the last sampling point within the time window. The calculation formula is: ; (5) Number of discharges By setting the discharge criterion threshold The number of times the detection signal rises from below the discharge criterion threshold to above the discharge criterion threshold is used to determine a discharge event. Event indication function. The expression is as follows: ; The formula for calculating the number of discharges within the time window is: ; (6) Discharge duration This refers to a signal amplitude that is higher than the discharge criterion threshold. Total time, duration indicator function The expression is as follows: ; The formula for calculating the discharge duration is: .

[0056] All the above features are based on a rolling time window and are updated in real time. Whenever new sampled data arrives, the window slides forward by one sampling interval. The earliest sampling point is discarded, and the latest data point is added. All feature values ​​are continuously refreshed over time to dynamically reflect the signal status. By real-time acquisition and analysis of three types of partial discharge signals—ultrasonic, TEV, and UHF—six feature quantities are used within a rolling time window: maximum value, minimum value, average value, instantaneous value, number of discharges, and discharge duration. This comprehensively reflects the discharge characteristics and development trend of the equipment. The partial discharge detection terminal dynamically visualizes the above discharge feature parameters and partial discharge type identification results, allowing maintenance personnel to intuitively grasp the equipment's discharge status and trends, and achieve continuous monitoring and early warning of anomalies.

[0057] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for acquiring and processing broadband partial discharge signals of power equipment, characterized in that, Includes the following steps: Step 100: Detect partial discharge signals using a fusion sensor to obtain three types of partial discharge signals; The fusion sensor includes an ultrasonic signal acquisition module, a TEV signal acquisition module, and an ultra-high frequency signal acquisition module; Step 200: Preprocess and grayscale map the three partial discharge signals respectively, and then stitch the three processed partial discharge signals together in the channel dimension to obtain a three-channel input tensor. Step 300: Input the three-channel input tensor into the partial discharge identification model and output the predicted partial discharge type; Step 400: Based on the identified partial discharge type, calculate and display the discharge characteristic parameters in real time within the rolling time window. Discharge characteristic parameters include maximum value, minimum value, average value, instantaneous value, number of discharges, and discharge duration.

2. The method for acquiring and processing wideband partial discharge signals of power equipment according to claim 1, characterized in that, The ultrasonic signal acquisition module uses a miniaturized ultrasonic MEMS microphone to acquire partial discharge ultrasonic signals in the 20kHz to 300kHz frequency band. The ultrasonic signal acquisition module couples the mechanical vibration waves on the surface of the power equipment through a rigid contact waveguide structure and converts the analog sound wave signal into a single-bit digital signal through a built-in Σ-Δ modulator.

3. The method for acquiring and processing wideband partial discharge signals of power equipment according to claim 2, characterized in that, The TEV signal acquisition module uses capacitor plates to couple partial discharge signals to acquire partial discharge TEV signals in the 3MHz to 100MHz frequency band. The TEV signal acquisition module captures transient voltage pulses on the surface of the metal casing of the equipment through the principle of electric field induction, and uses a two-stage differential amplifier to amplify the signal and an elliptic function filter for bandpass filtering.

4. The method for acquiring and processing broadband partial discharge signals of power equipment according to claim 3, characterized in that, The UHF signal acquisition module uses a log-periodic radiating antenna as the sensing unit to acquire UHF electromagnetic signals of partial discharge in the 300MHz to 1.5GHz frequency band. The UHF signal acquisition module converts the transient electromagnetic radiation energy generated during partial discharge into radio frequency electrical signals through electromagnetic coupling between the antenna and the spatial electromagnetic field. The signal is then processed through a two-stage amplification circuit. The first stage uses a low-noise amplifier for fixed gain amplification, while the second stage uses a variable gain amplifier for dynamic gain compression.

5. The method for acquiring and processing broadband partial discharge signals of power equipment according to claim 4, characterized in that, Step 200 involves obtaining the three-channel input tensor, including: Step 201: Normalize the sampling sequence of the original one-dimensional partial discharge signal to obtain a normalized signal sequence; Step 202: The normalized signal sequence is segmented using a sliding window, and the segmented signal data are combined into a signal matrix; Step 203: Map each value in the signal matrix to a grayscale image and adjust its size to generate a two-dimensional grayscale image; Step 204: Map the ultrasonic signal, TEV signal and UHF signal to their corresponding two-dimensional grayscale images, and stitch them together in the channel dimension to form a three-channel input tensor.

6. The method for acquiring and processing broadband partial discharge signals of power equipment according to claim 5, characterized in that, The partial discharge identification model consists of an input layer, a convolutional layer, an activation function layer, a pooling layer, a channel attention fusion module, a cross-channel fusion layer, a global average pooling layer, and an SVM classification layer connected in sequence. The input layer receives the three-channel fused image tensor and performs preprocessing such as normalization and mean removal; the convolutional pooling layer extracts local features using convolutional kernels; the activation function layer introduces nonlinearity using the ReLU activation function; the pooling layer performs downsampling through pooling operations; the channel attention fusion module performs adaptive weight allocation for features of different modalities; the cross-channel fusion layer performs cross-channel feature fusion through 1×1 convolution; the global average pooling layer generates a global feature vector through global average pooling; and the SVM classification layer performs multi-class classification on the global feature vector and outputs the partial discharge type identification result.

7. The method for acquiring and processing broadband partial discharge signals of power equipment according to claim 6, characterized in that, The processing steps of the channel attention fusion module include the following: Step 301: Perform global average pooling on the original feature map of each channel to obtain the global average response value of each channel; The original feature map is obtained by passing a three-channel input tensor through convolutional layers, activation function layers, and pooling layers; Step 302: Input the global average response value of each channel into the two-layer fully connected network, and generate channel weights through the Sigmoid activation function; Step 303: Multiply the channel weights with the original feature maps of the corresponding channels one by one to obtain the weighted feature maps.

8. The method for acquiring and processing broadband partial discharge signals of power equipment according to claim 7, characterized in that, The SVM classification layer is specifically a Support Vector Machine classifier. The training objective of the Support Vector Machine classifier is to solve the following optimization problem: ； The constraints are as follows: ； in, This is the penalty parameter, used to control the balance between maximizing the classification margin and penalizing misclassification. For the first Relaxed variables for each training sample, allowing some samples to violate the interval constraint. For the first The class labels of each training sample. and These are the classification weight vector and classification bias parameter to be optimized, respectively. Indicates transpose. Let represent the global feature vector of the i-th training sample.

9. The method for acquiring and processing broadband partial discharge signals of power equipment according to claim 8, characterized in that, The number of discharge events is calculated as follows: the number of times the amplitude of the detection signal rises from below the preset discharge criterion threshold to above the preset discharge criterion threshold is counted as the number of discharge events.

10. The method for acquiring and processing broadband partial discharge signals of power equipment according to claim 9, characterized in that, The discharge duration is calculated as follows: the number of sampling points whose signal amplitude is higher than the preset discharge criterion threshold is counted, and the product of the number of sampling points and the sampling interval is taken as the discharge duration.

Images