A transformer partial discharge type identification method

Abstract

This invention relates to the field of power equipment testing technology, and more particularly to a method for identifying partial discharge types in transformers. The method first acquires partial discharge pulse signals using an ultra-high frequency sensor, and obtains a standard time-domain signal sequence through denoising and normalization preprocessing. Then, it extracts two types of features in parallel: first, it uses an improved adaptive noise complete set empirical mode decomposition to obtain key intrinsic mode components, and calculates multi-scale permutation entropy to form a first feature subset; second, it converts the signal into a time-frequency distribution image, and extracts high-dimensional deep features through a pre-trained deep convolutional neural network to form a second feature subset. Finally, it uses kernel canonical correlation analysis to fuse the features, and after dimensionality reduction, obtains a low-dimensional discriminative feature vector, which is input into a cascaded forest ensemble classifier containing a channel attention module to complete the identification. This invention mines features from multiple dimensions, improving feature discrimination and recognition accuracy, and has strong anti-interference capabilities, providing reliable support for transformer fault diagnosis and ensuring the stable operation of power systems.

Description

Technical Field

[0001] This invention relates to the field of power equipment testing technology, specifically a method for identifying the type of partial discharge in a transformer. Background Technology

[0002] As a core piece of equipment in the power system, the operating status of transformers directly affects the safety, stability, and reliability of power supply. Partial discharge is a typical phenomenon in the evolution of insulation defects inside transformers. Different types of partial discharge correspond to different insulation fault modes. Accurately identifying the type of partial discharge is a key step in judging the insulation status of transformers and locating potential faults, which is of great significance for ensuring the long-term safe operation of transformers.

[0003] As power systems evolve towards higher voltage and larger capacity, the operating environment of transformers is becoming increasingly complex. Their internal insulation structures are susceptible to electrical aging, thermal aging, and mechanical stress, leading to various types of partial discharge. Traditional partial discharge detection methods often rely on ultrasonic sensors or high-frequency current sensors. These sensors suffer from limited bandwidth and weak anti-interference capabilities, making it difficult to accurately capture weak partial discharge pulse signals. Furthermore, the acquired signals are easily affected by transformer operating noise and external electromagnetic interference, resulting in a low signal-to-noise ratio, which complicates subsequent feature extraction and type identification.

[0004] In terms of feature extraction, most existing technologies employ single-dimensional feature extraction methods, either extracting simple statistical features such as peak values ​​and mean values ​​from time-domain signals or focusing solely on frequency-domain information analysis, failing to fully explore the intrinsic and external morphological characteristics of the signal. Single-dimensional features often suffer from insufficient representational capabilities, making it difficult to comprehensively reflect the differentiated characteristics of different types of partial discharges, resulting in limited feature discrimination and affecting subsequent identification accuracy.

[0005] Feature fusion and dimensionality reduction also have significant drawbacks. Traditional methods often fuse different features by directly concatenating them, failing to effectively uncover the potential relationships between feature subsets. The fused features tend to contain a large amount of redundant information, increasing computational complexity. Furthermore, commonly used dimensionality reduction methods easily lose key discriminative information during the feature reduction process, further weakening the effectiveness of the features.

[0006] In the classification and recognition stage, existing technologies mostly use a single base learner for classification. Such classification models are poorly adaptable to complex and ever-changing partial discharge signals and have insufficient generalization ability. When faced with the coexistence of multiple discharge types or the distortion of discharge signals, they are prone to misjudgment and missed judgment. The recognition accuracy and reliability are difficult to meet the needs of actual engineering.

[0007] In summary, existing methods for identifying partial discharge types in transformers have many shortcomings in terms of the comprehensiveness of signal acquisition quality feature extraction, the effectiveness of feature fusion, and the robustness of classification models. They are unable to accurately and efficiently complete the task of identifying partial discharge types and cannot provide reliable support for transformer insulation fault diagnosis. There is an urgent need for a technical solution that can take into account both the comprehensiveness and effectiveness of signal quality feature fusion and the accuracy of identification to solve the above problems. Summary of the Invention

[0008] The purpose of this invention is to provide a method for identifying the type of partial discharge in a transformer, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying the type of partial discharge in a transformer includes the following steps: S1. Signal Acquisition and Preprocessing: Partial discharge pulse signals inside the transformer are acquired by an ultra-high frequency sensor, and the pulse signals are preprocessed by denoising and normalization to obtain a standard time-domain signal sequence. S2. Multi-dimensional Feature Construction: For standard time-domain signal sequences, the following two types of feature extraction processes are executed in parallel: S2-1, Intrinsic Mode Feature Extraction Process: The standard time-domain signal sequence is decomposed using an adaptive signal decomposition method to obtain multiple intrinsic mode components; based on predetermined screening criteria, key components are selected from the multiple intrinsic mode components; the nonlinear dynamic characteristic parameters of the key components are calculated to form the first feature subset; S2-2, Time-Frequency Image Feature Extraction Process: Convert the standard time-domain signal sequence into a time-frequency distribution image; use a deep convolutional neural network model to extract high-dimensional depth features from the time-frequency distribution image to form a second feature subset; S3. Hybrid Feature Fusion and Dimensionality Reduction: The first feature subset and the second feature subset are fused to form a hybrid feature vector, and the hybrid feature vector is then subjected to dimensionality reduction to obtain a low-dimensional discriminative feature vector. S4. Intelligent Discharge Type Recognition: The low-dimensional discriminative feature vector is input into a pre-trained ensemble classifier. The ensemble classifier obtains the final partial discharge type recognition result by fusing the outputs of multiple heterogeneous base learners. The first feature subset and the second feature subset characterize the discharge characteristics from two complementary dimensions: the intrinsic mode and the external time-frequency morphology of the signal. They are then combined through the fusion and dimensionality reduction in step S3 to influence the classification decision in step S4.

[0010] As a preferred embodiment, in step S2-1, the adaptive signal decomposition method is an improved adaptive noise complete set empirical mode decomposition method, wherein the amplitude of the added white noise is adaptively adjusted according to the envelope entropy of the standard time domain signal sequence.

[0011] As a preferred option, in step S2-1, the predetermined screening criterion is the kurtosis-correlation coefficient joint criterion, which is as follows: calculate the correlation coefficient of each intrinsic mode component with the standard time domain signal sequence and its own kurtosis value, and screen out the components whose correlation coefficient and kurtosis value are both greater than the corresponding preset threshold as key components.

[0012] As a preferred approach, in step S2-1, the nonlinear dynamic characteristic parameter is the multi-scale permutation entropy; the calculation of the nonlinear dynamic characteristic parameter of the key component specifically involves calculating the permutation entropy value of each key component at multiple time scales, which together constitute the first feature subset.

[0013] As a preferred embodiment, step S2-2, converting the standard time-domain signal sequence into a time-frequency distribution image, specifically includes: performing continuous wavelet transform on the standard time-domain signal sequence to obtain a time-frequency matrix; converting the time-frequency matrix into a grayscale image and performing size normalization processing to obtain a time-frequency distribution image.

[0014] As a preferred option, in step S2-2, the deep convolutional neural network model is a model pre-trained on a large image dataset, whose last fully connected layer is removed, and the output of its last convolutional or pooling layer is used as a high-dimensional deep feature.

[0015] As a preferred approach, in step S3, the first feature subset and the second feature subset are fused. Specifically, a method based on kernel canonical correlation analysis is used to map the two feature subsets to the correlation subspace and then concatenate them to form a hybrid feature vector. The hybrid feature vector is then subjected to dimensionality reduction processing. Specifically, a local preservation projection or t-distributed random neighborhood embedding method is used to reduce the hybrid feature vector to a preset dimension.

[0016] As a preferred approach, in step S4, the ensemble classifier is a cascaded forest structure, which includes multi-granularity scanning layers and cascaded decision tree forest layers. The output results of multiple heterogeneous base learners are fused. Specifically, in each cascaded layer, the output features or category distribution vectors of different types of decision tree forests within that layer are combined and passed to the next layer. In the cascaded layers of the cascaded forest structure, a channel attention module is integrated to recalibrate the channel weights of the feature vectors passed from the previous layer, and then input the weighted features into the decision tree forest of this layer for processing.

[0017] As can be seen from the technical solution provided by the present invention above, the transformer partial discharge type identification method provided by the present invention has the following beneficial effects: The signal acquisition and preprocessing stage uses an ultra-high frequency sensor combined with adaptive threshold wavelet denoising and normalization processing, which effectively filters out external electromagnetic interference and internal noise during transformer operation, ensuring the purity and standardization of the acquired signal, providing high-quality data support for subsequent feature extraction, and avoiding interference from inferior signals on the recognition results. Multi-dimensional feature construction extracts features from two complementary dimensions: intrinsic mode features and time-frequency image features. It captures the nonlinear dynamics of the signal and covers the high-dimensional structural information in the time-frequency domain, greatly improving the comprehensiveness and discriminative ability of the features and solving the problem of insufficient single-dimensional feature representation. Hybrid feature fusion and dimensionality reduction employs kernel canonical correlation analysis to achieve deep fusion of feature subsets, fully exploring the potential correlations between features of different dimensions. Dimensionality reduction is then achieved through local preservation projection or t-distributed random neighborhood embedding methods, which remove redundant information and reduce computational complexity while preserving key discriminative features to the maximum extent, thus balancing feature quality and computational efficiency. The ensemble classifier adopts a cascaded forest structure and integrates a channel attention module. By fusing multi-granularity scanning and hierarchical heterogeneous base learners, it fully leverages the advantages of ensemble learning, improves the model's adaptability to complex discharge types and recognition accuracy, and the introduction of the channel attention mechanism further strengthens the role of key features, effectively reduces the false positive rate, and makes the recognition results more reliable. The entire methodology is logically rigorous and the steps are seamlessly connected. It balances identification accuracy, computational efficiency, and practical deployment feasibility, enabling rapid and accurate identification of transformer partial discharge types. This provides a scientific basis for transformer fault diagnosis and maintenance, helps to promptly identify potential hazards, reduces the risk of power outages, and ensures the safe and stable operation of the power system. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the steps in the transformer partial discharge type identification method of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] To better understand the above technical solutions, the following will provide a detailed description of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, this embodiment of the invention provides a method for identifying the type of partial discharge in a transformer, comprising the following steps: S1. Signal Acquisition and Preprocessing: Partial discharge pulse signals inside the transformer are acquired by an ultra-high frequency sensor, and the pulse signals are preprocessed by denoising and normalization to obtain a standard time-domain signal sequence. S2. Multi-dimensional Feature Construction: For standard time-domain signal sequences, the following two types of feature extraction processes are executed in parallel: S2-1, Intrinsic Mode Feature Extraction Process: The standard time-domain signal sequence is decomposed using an adaptive signal decomposition method to obtain multiple intrinsic mode components; based on predetermined screening criteria, key components are selected from the multiple intrinsic mode components; the nonlinear dynamic characteristic parameters of the key components are calculated to form the first feature subset; S2-2, Time-Frequency Image Feature Extraction Process: Convert the standard time-domain signal sequence into a time-frequency distribution image; use a deep convolutional neural network model to extract high-dimensional depth features from the time-frequency distribution image to form a second feature subset; S3. Hybrid Feature Fusion and Dimensionality Reduction: The first feature subset and the second feature subset are fused to form a hybrid feature vector, and the hybrid feature vector is then subjected to dimensionality reduction to obtain a low-dimensional discriminative feature vector. S4. Intelligent Discharge Type Recognition: The low-dimensional discriminative feature vector is input into a pre-trained ensemble classifier. The ensemble classifier obtains the final partial discharge type recognition result by fusing the outputs of multiple heterogeneous base learners. The first feature subset and the second feature subset characterize the discharge characteristics from two complementary dimensions: the intrinsic mode and the external time-frequency morphology of the signal. They are then combined through the fusion and dimensionality reduction in step S3 to influence the classification decision in step S4.

[0022] In this embodiment, the core function of step S1 is to accurately capture the partial discharge pulse signal within the transformer. Through targeted denoising and normalization processing, the influence of interference signals and amplitude differences is eliminated, generating a standardized time-domain signal sequence, providing a high-quality data foundation for subsequent multi-dimensional feature construction. The detailed steps are as follows: S1-1: Data Acquisition System Setup and Parameter Configuration A high-frequency partial discharge signal acquisition system was constructed, with core components including a high-frequency sensor, a signal conditioning module, a data acquisition card, and a host computer. The high-frequency sensor should be a broadband sensor with a bandwidth covering 300MHz to 3GHz. This bandwidth range matches the main frequency range of the high-frequency electromagnetic pulse signal generated by the transformer's partial discharge, ensuring distortion-free signal acquisition. The sensor is deployed at a pre-set detection window on the transformer tank wall, and conductive adhesive is used to seal the sensor to the tank wall to reduce signal attenuation and external electromagnetic interference. Configure the signal conditioning module parameters, setting the preamplifier gain to 40dB and the bandpass filter cutoff frequency to 300MHz to 3GHz for initial filtering of low-frequency interference and high-frequency noise. The data acquisition card parameters must meet signal sampling requirements, with a sampling rate of 10GS / s and quantization bits of 16 bits to ensure complete recording of waveform details and amplitude information of the pulse signal. Install the data acquisition software on the host computer, setting the acquisition trigger mode to external trigger and the trigger threshold to 20mV to prevent low-amplitude noise signals from falsely triggering the acquisition process. S1-2: Partial Discharge Pulse Signal Acquisition: The data acquisition system is started to monitor the transformer's operating status synchronously. Signal acquisition begins after the transformer has been running stably under rated load conditions for 30 minutes. During the acquisition process, the data acquisition card receives signals processed by the sensors and conditioning module in real time. The acquisition duration is set to 10 minutes to ensure that partial discharge pulse signals at different stages can be captured. A multi-channel synchronous acquisition method is used during the acquisition process, and the data of each channel is stored independently in binary raw data format to avoid signal distortion during the format conversion process. Anti-interference measures must be taken during the data acquisition process. The grounding resistance of the acquisition system should be controlled within 4Ω. Shielded cables should be used to connect the sensors and the acquisition equipment, and the cable length should not exceed 5m to reduce electromagnetic interference during transmission. At the same time, data acquisition should be avoided during the operation of high-frequency equipment near transformers to further reduce the impact of external environmental interference on the acquired signal. S1-3: Adaptive threshold wavelet denoising: For the acquired raw pulse signal, an adaptive threshold wavelet denoising method is used to eliminate residual interference. First, the raw signal is decomposed by wavelet, and the db4 wavelet is selected as the wavelet basis function. The number of decomposition layers is set to 5, resulting in one approximation coefficient and five detail coefficients. The approximation coefficient corresponds to the low-frequency useful components of the signal, and the detail coefficients contain the high-frequency components of noise and partial discharge signal. The adaptive threshold for the detail coefficients of each layer is calculated. This threshold is obtained by the logarithmic correlation between the noise standard deviation of the corresponding layer detail coefficient and the number of sampling points of the original signal. The noise standard deviation is calculated using the fifth layer detail coefficients. Specifically, the median of the absolute values ​​of all detail coefficients in the fifth layer is taken and then divided by 0.6745. Based on the calculated thresholds for each layer, the detail coefficients of the corresponding layers are thresholded. The processing rule is that when the absolute value of a detail coefficient is greater than the corresponding threshold, the coefficient is retained and the threshold is subtracted before retaining the original sign; when the absolute value of a detail coefficient is less than or equal to the corresponding threshold, the coefficient is set to zero. The processed approximation coefficients and the detail coefficients of each layer are subjected to inverse wavelet transform to reconstruct the denoised partial discharge pulse signal, thus effectively removing the interference signal. S1-4: Signal normalization processing: The denoised signal is normalized to eliminate the influence of signal amplitude differences under different discharge intensities, so that the signal amplitude is uniformly mapped to the range of 0 to 1. The min-max normalization method is adopted. The specific processing method is to subtract the minimum amplitude of the signal from the value of each sampling point of the denoised signal, and then divide by the difference between the maximum amplitude and the minimum amplitude to achieve the standardization conversion of the amplitude. During the normalization process, it is necessary to traverse all sampling points of the denoised signal, first determine the minimum and maximum amplitude of the signal, and then complete the above calculation point by point to ensure that the amplitude normalization conversion is completed for each sampling point. S1-5: Standard Time-Domain Signal Sequence Generation: The normalized signals are arranged in chronological order of sampling time to form a standard time-domain signal sequence. The sequence length is consistent with the number of sampling points of the original acquired signal, and each sequence element corresponds to the standardized signal amplitude at a sampling time. The generated standard time-domain signal sequence is validated, and the validation criteria include whether the signal amplitude range is within the interval of 0 to 1 and whether there are missing or outlier values ​​in the sequence. If outlier values ​​are found, linear interpolation is used for correction. If the proportion of missing values ​​exceeds 5%, steps S1-2 to S1-4 are repeated for signal acquisition and preprocessing. After the validation is passed, the standard time-domain signal sequence is stored in the data server to provide data input for the multi-dimensional feature construction in step S2.

[0023] In this embodiment, the core function of step S2 is to extract comprehensive and discriminative features in parallel from the two complementary dimensions of the intrinsic mode and the external time-frequency morphology of the standard time-domain signal sequence, constructing a multi-dimensional feature set, which provides a solid feature foundation for subsequent hybrid feature fusion and dimensionality reduction; the detailed steps are as follows: S2-1: Intrinsic Modal Feature Extraction Process: This process focuses on mining the intrinsic dynamic characteristics of the signal. Through adaptive decomposition, key component screening, and nonlinear feature calculation, a first feature subset that can characterize the essential laws of the discharge signal is constructed. S2-1-1: Adaptive signal decomposition: An improved adaptive noise-complete ensemble empirical mode decomposition method is employed to decompose a standard time-domain signal sequence. During the decomposition process, white noise is added to the standard time-domain signal sequence. The amplitude of the white noise is not fixed but adaptively adjusted based on the envelope entropy of the standard time-domain signal sequence. Envelope entropy reflects the complexity of the signal; a high envelope entropy indicates high signal complexity, and the white noise amplitude is appropriately increased to improve the completeness of the decomposition. Conversely, a low envelope entropy indicates low signal complexity, and the white noise amplitude is decreased to avoid excessive interference with the useful signal. Through this adaptive addition of white noise, the standard time-domain signal sequence undergoes multiple rounds of decomposition and ensemble averaging, ultimately obtaining multiple independent intrinsic mode components covering different frequency components. S2-1-2: Key Component Screening: Key components are selected from multiple intrinsic mode components based on a joint criterion of kurtosis and correlation coefficient. First, the correlation coefficient between each intrinsic mode component and the original standard time-domain signal sequence is calculated. The correlation coefficient measures the similarity between two signals; a higher value indicates that the intrinsic mode component contains more useful information from the original signal. Simultaneously, the kurtosis value of each intrinsic mode component is calculated. The kurtosis value reflects the pulse characteristics of the signal. Partial discharge signals have obvious pulse characteristics, and a higher kurtosis value indicates that the component is more likely to contain key information about the discharge signal. Correlation coefficient thresholds and kurtosis value thresholds are pre-set. The correlation coefficient and kurtosis value of each intrinsic mode component are compared with the corresponding thresholds, and only components with both correlation coefficients and kurtosis values ​​greater than the corresponding thresholds are retained. These retained components are the key components that can effectively characterize the characteristics of partial discharge. S2-1-3: Calculation of nonlinear dynamic characteristic parameters: Using multi-scale permutation entropy as a nonlinear dynamic characteristic parameter, the nonlinear dynamic characteristics of each key component are calculated. Multi-scale permutation entropy can reflect the nonlinear dynamic characteristics of the signal at different time scales and reveal the inherent laws of the signal. In the calculation process, firstly, multiple different time scales are set for each key component, and the selection of time scales needs to cover the possible dynamic change range of the signal. For each time scale, the key component is scale-transformed, and then the amplitude permutation law of the signal at that scale is analyzed according to the calculation logic of permutation entropy to obtain the permutation entropy value at the corresponding time scale. The permutation entropy values ​​of each key component at all set time scales are integrated, and the permutation entropy values ​​of all key components together constitute the first feature subset. S2-2: Time-frequency image feature extraction process: This process converts the time-domain signal into a time-frequency distribution image, and utilizes the strong feature extraction capability of deep convolutional neural networks to mine high-dimensional deep features of the signal in the time-frequency domain, and constructs a second feature subset. S2-2-1: Time-frequency distribution image transformation: First, a continuous wavelet transform is performed on the standard time-domain signal sequence. By selecting appropriate wavelet basis functions, the continuous wavelet transform maps the standard time-domain signal sequence to a two-dimensional time-frequency plane, obtaining a time-frequency matrix that simultaneously reflects the signal's time and frequency distribution. Each element in the time-frequency matrix represents the signal's amplitude at the corresponding time and frequency point. Then, the obtained time-frequency matrix is ​​converted into a grayscale image. By mapping different amplitude intensities in the time-frequency matrix to different grayscale levels, higher amplitude intensity corresponds to a grayscale value closer to white, and lower amplitude intensity corresponds to a grayscale value closer to black, thus forming an intuitive time-frequency grayscale image. Finally, the grayscale image is normalized. According to the input requirements of the subsequent deep convolutional neural network model, the image is adjusted to a fixed width and height to ensure uniform image size and meet the model input standards, ultimately obtaining the time-frequency distribution image. S2-2-2: High-dimensional deep feature extraction: A deep convolutional neural network model pre-trained on a large image dataset is selected for feature extraction. These pre-trained models, trained on a large amount of image data, possess powerful general feature extraction capabilities and can effectively identify complex textures, structures, and other features in images. Before use, the model needs to be adaptively adjusted by removing the last fully connected layer, as fully connected layers are primarily used for classification decisions in specific tasks, while this step aims to extract high-dimensional deep features rather than direct classification. After adjustment, the time-frequency distribution image, after size normalization, is input into the deep convolutional neural network model. The model extracts features from the image through layer-by-layer convolution and pooling operations, ultimately obtaining the output of the last convolutional or pooling layer. This output is the high-dimensional deep feature of the time-frequency distribution image, which comprehensively represents the key information of the time-frequency image and constitutes the second feature subset. The two feature extraction processes described above are executed in parallel without interfering with each other. They obtain features from the intrinsic nature and external morphology of the signal, respectively. The first feature subset and the second feature subset together constitute a multi-dimensional feature set, providing comprehensive and complementary feature support for the subsequent hybrid feature fusion and dimensionality reduction steps.

[0024] In this embodiment, the core function of step S3 is to effectively fuse the first feature subset and the second feature subset, explore the potential correlation between the two feature subsets, integrate complementary feature information, and at the same time, eliminate redundant information and reduce feature dimensionality through dimensionality reduction processing to obtain a low-dimensional discriminative feature vector that is both representative and concise, providing efficient feature input for subsequent intelligent identification of discharge types; the detailed steps are as follows: S3-1: Hybrid Feature Fusion Feature fusion is performed using a kernel canonical correlation analysis method. First, kernel mapping is applied to the first and second feature subsets respectively. The kernel function maps the two feature subsets from their original feature spaces to a high-dimensional Hilbert space, mining the linear correlation between the two feature subsets in the high-dimensional space. Then, the two mapped feature subsets are projected onto a common correlation subspace, maximizing the linear correlation of the projected feature vectors and thus fully preserving the complementary information and intrinsic relationship between the two feature subsets. Finally, the two feature subsets projected onto the correlation subspace are concatenated sequentially, combining the originally separate feature vectors into a unified feature vector, forming a hybrid feature vector containing multi-dimensional complementary information. S3-2: Feature Dimensionality Reduction Processing For the fused hybrid feature vector, either local preservation projection or t-distributed random neighborhood embedding is used for dimensionality reduction. When choosing the local preservation projection method, this method preserves the local neighborhood relationships between data samples during the dimensionality reduction process, ensuring that the dimensionality-reduced feature vector can still reflect the local structural information of the samples in the original hybrid features, avoiding the loss of key discriminative information. When choosing the t-distributed random neighborhood embedding method, this method maps the high-dimensional hybrid feature vector to a low-dimensional space by constructing the probability distribution relationship between samples, while maintaining the similarity structure between samples, which is particularly suitable for handling the dimensionality reduction needs of nonlinear features. During the dimensionality reduction process, the target dimension is pre-set based on the input requirements of the subsequent integrated classifier and the computational resource limitations of the actual application scenario. The hybrid feature vector is then processed using the aforementioned dimensionality reduction method to gradually remove redundant information from the features, retaining the core features that play a crucial role in discharge type identification. Ultimately, the dimension of the hybrid feature vector is reduced to the preset dimension, resulting in a low-dimensional discriminative feature vector. This vector integrates key information on the intrinsic modes and external time-frequency morphology of the signal, and also features low dimensionality and high computational efficiency, effectively supporting subsequent discharge type identification tasks.

[0025] In this embodiment, the core function of step S4 is to utilize a pre-trained ensemble classifier to perform deep processing on the low-dimensional discriminative feature vector. By fusing the outputs of multiple heterogeneous base learners, the advantages of ensemble learning are fully leveraged to achieve accurate identification of partial discharge types and output reliable identification results. The detailed steps are as follows: S4-1: Loading and Initializing the Ensemble Classifier First, the pre-trained ensemble classifier is loaded. This classifier adopts a cascaded forest structure, with core components including multi-granularity scanning layers and cascaded decision tree forest layers, and a channel attention module is integrated in the cascaded decision tree forest layers. During loading, the pre-trained weight parameters, network structure configuration, and various hyperparameters of the classifier need to be imported simultaneously to ensure that the functions of each component of the classifier are activated normally. In the initialization phase, the input interface of the classifier is adapted to ensure that the low-dimensional discriminative feature vector can be input smoothly. At the same time, the state of each layer component is reset to avoid interference from historical running data to the current recognition task. S4-2: Feature transformation of multi-granularity scanning layers: The low-dimensional discriminative feature vector is input into the multi-granularity scanning layer of the ensemble classifier. This layer performs multi-scale scanning on the input feature vector through sliding windows of different sizes, extracting multiple sub-feature vectors of different granularities. The size of each sliding window is pre-set according to the dimension of the low-dimensional discriminative feature vector and the feature requirements for discharge type identification, ensuring that the key information regions in the feature vector are covered. Subsequently, all the extracted sub-feature vectors are input into the basic decision tree forest built into this layer for preliminary feature transformation. Through the splitting and aggregation operations of the decision tree forest, feature representation information at different granularities is extracted to form a multi-granularity feature set, providing rich feature input for the deep processing of subsequent cascaded layers. S4-3: Cascaded decision tree forest layer-by-layer processing: Multi-granularity feature sets are sequentially fed into cascaded decision tree forest layers. Each layer's processing flow includes two core steps: channel attention weighting and heterogeneous base learner operation. First, the feature vectors passed from the previous layer enter the channel attention module of this layer. This module recalibrates the weight of each channel by analyzing the contribution of each channel of the feature vector to the identification of the discharge type. The contribution evaluation is based on the channel importance parameters learned during the pre-training process. Channels with significant identification effects are given higher weights, while redundant or interfering channels are given lower weights. The weight calibration highlights key feature information and suppresses invalid information. After the channel weights are calibrated, the weighted feature vectors are input into multiple heterogeneous base learners in this layer. These heterogeneous base learners include different types of decision tree forests, such as random forests and extreme random forests. Each learner has unique feature learning preferences and decision logic. Each heterogeneous base learner processes the weighted feature vectors independently and outputs its own feature representation or category distribution information through internal decision tree splitting, voting and other operations. S4-4: Hierarchical Output Fusion and Transmission In each cascaded decision tree forest layer, the outputs of all heterogeneous base learners within that layer are fused. The fusion method involves concatenating the class distribution vectors output by each learner, or calculating the comprehensive class distribution vector of that layer using average pooling. This ensures that the decision information from different learners is fully integrated, compensating for the limitations of a single learner. The fused comprehensive feature vector or class distribution vector is then used as the input data for the next layer and passed sequentially to all cascaded layers. Through multiple rounds of deep processing, the discriminative ability of the feature representation is gradually optimized. S4-5: Final Discharge Type Decision: After the feature vector is processed by the last cascaded decision tree forest layer, the layer outputs the final comprehensive category distribution vector. This vector contains the predicted probability corresponding to each type of partial discharge. Based on the category distribution vector, the category with the highest predicted probability is selected as the final partial discharge type identification result. At the same time, the confidence score corresponding to this identification result is output. The confidence score is the specific value of the highest predicted probability, which is used to reflect the reliability of the identification result. Finally, the partial discharge type identification result and the confidence score are output together, providing a direct basis for the diagnosis and treatment of transformer partial discharge faults.

[0026] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying the type of partial discharge in a transformer, characterized in that: Includes the following steps: S1. Signal Acquisition and Preprocessing: Partial discharge pulse signals inside the transformer are acquired by an ultra-high frequency sensor, and the pulse signals are preprocessed by denoising and normalization to obtain a standard time-domain signal sequence. S2. Multi-dimensional Feature Construction: For standard time-domain signal sequences, the following two types of feature extraction processes are executed in parallel: S2-1, Intrinsic Mode Feature Extraction Process: The standard time-domain signal sequence is decomposed using an adaptive signal decomposition method to obtain multiple intrinsic mode components; based on predetermined screening criteria, key components are selected from the multiple intrinsic mode components. Calculate the nonlinear dynamic characteristic parameters of the key components to form the first feature subset; S2-2, Time-Frequency Image Feature Extraction Process: Convert the standard time-domain signal sequence into a time-frequency distribution image; use a deep convolutional neural network model to extract high-dimensional depth features from the time-frequency distribution image to form a second feature subset; S3. Hybrid Feature Fusion and Dimensionality Reduction: The first feature subset and the second feature subset are fused to form a hybrid feature vector, and the hybrid feature vector is then subjected to dimensionality reduction to obtain a low-dimensional discriminative feature vector. S4. Intelligent Discharge Type Recognition: The low-dimensional discriminative feature vector is input into a pre-trained ensemble classifier. The ensemble classifier obtains the final partial discharge type recognition result by fusing the outputs of multiple heterogeneous base learners. The first feature subset and the second feature subset characterize the discharge characteristics from two complementary dimensions: the intrinsic mode and the external time-frequency morphology of the signal. They are then combined through the fusion and dimensionality reduction in step S3 to influence the classification decision in step S4.

2. The method for identifying the type of partial discharge in a transformer according to claim 1, characterized in that: In step S2-1, the adaptive signal decomposition method is an improved adaptive noise complete set empirical mode decomposition method, wherein the amplitude of the added white noise is adaptively adjusted according to the envelope entropy of the standard time domain signal sequence.

3. The method for identifying the type of partial discharge in a transformer according to claim 2, characterized in that: In step S2-1, the predetermined screening criterion is the kurtosis-correlation coefficient joint criterion, which is as follows: calculate the correlation coefficient of each intrinsic mode component with the standard time domain signal sequence and its own kurtosis value, and screen out the components whose correlation coefficient and kurtosis value are both greater than the corresponding preset threshold as key components.

4. The method for identifying the type of partial discharge in a transformer according to claim 1, characterized in that: In step S2-1, the nonlinear dynamic characteristic parameter is the multi-scale permutation entropy; the nonlinear dynamic characteristic parameter of the key component is calculated, specifically: the permutation entropy value of each key component at multiple time scales is calculated, which together constitute the first feature subset.

5. The method for identifying the type of partial discharge in a transformer according to claim 1, characterized in that: In step S2-2, the standard time-domain signal sequence is converted into a time-frequency distribution image, which specifically includes: performing continuous wavelet transform on the standard time-domain signal sequence to obtain a time-frequency matrix; converting the time-frequency matrix into a grayscale image and performing size normalization processing to obtain a time-frequency distribution image.

6. The method for identifying the type of partial discharge in a transformer according to claim 1, characterized in that: In step S2-2, the deep convolutional neural network model is a model pre-trained on a large image dataset. Its last fully connected layer is removed, and the output of its last convolutional or pooling layer is used as a high-dimensional deep feature.

7. The method for identifying the type of partial discharge in a transformer according to claim 1, characterized in that: In step S3, the first feature subset and the second feature subset are fused. Specifically, a method based on kernel canonical correlation analysis is used to map the two feature subsets to the correlation subspace and then concatenate them to form a hybrid feature vector. The hybrid feature vector is then subjected to dimensionality reduction processing. Specifically, a local preservation projection or t-distributed random neighborhood embedding method is used to reduce the hybrid feature vector to a preset dimension.

8. The method for identifying the type of partial discharge in a transformer according to claim 1, characterized in that: In step S4, the ensemble classifier is a cascaded forest structure, which includes multi-granularity scanning layers and cascaded decision tree forest layers. It integrates the outputs of multiple heterogeneous base learners. Specifically, in each cascaded layer, it combines the output features or category distribution vectors of different types of decision tree forests within that layer and passes them to the next layer. In the cascaded layers of the cascaded forest structure, a channel attention module is integrated to recalibrate the channel weights of the feature vectors passed from the previous layer, and then input the weighted features into the decision tree forest of this layer for processing.

Images