DenseBlock-based partial discharge feature fusion extraction method and system

Abstract

The invention discloses a DenseBlock-based partial discharge feature fusion extraction method and system, and relates to the technical field of electrical equipment partial discharge detection, and the method comprises the steps: collecting original data through a first test, and carrying out the first classification; the classified atlas is preprocessed; and constructing a first identification model, and performing classification identification validity evaluation. According to the method, the original data is collected through the first test and the PRPD atlas is generated, partial discharge characteristic information can be effectively reserved, data redundancy and complexity are reduced by preprocessing the classified PRPD atlas, the DenseNet model is constructed by introducing the DenseBlock module into the convolutional neural network, the expression ability of the first recognition model is improved, and the recognition efficiency is improved. The information transmission is improved, the over-fitting phenomenon is effectively reduced, and the accuracy and reliability of fault identification are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of partial discharge detection of electrical equipment, and in particular to a partial discharge feature fusion extraction method and system based on DenseBlock. Background Art

[0002] The partial discharge problem of electrical equipment in the distribution network system is one of the most common faults. The insulation performance of electrical equipment may deteriorate due to changes in the external environment, aging of internal materials, and mechanical vibration, which may lead to partial discharge. This phenomenon greatly damages the insulation performance of transformer equipment, and different types of discharges have different degrees of harm to the equipment. Therefore, timely and accurate identification and classification of partial discharge phenomena in transformers are of great significance to the stable operation of the power system and the long-term reliability of transformers.

[0003] At present, the detection of partial discharge of transformers mainly uses ultra-high frequency method, ultrasonic method, optical detection method and oil-soluble gas analysis method, etc. These traditional methods can detect the partial discharge phenomenon of transformers and collect discharge pulse information, and then cooperate with artificial intelligence algorithms such as convolutional neural networks to automatically identify the type of partial discharge.

[0004] However, most current methods often face challenges in high-dimensional data processing and feature extraction when processing the collected discharge pulse information. These traditional algorithms may lose some important information during feature selection and classification, resulting in low recognition accuracy. In addition, the complexity of the model and the overfitting phenomenon during the training process also limit its practical application effect, resulting in large final recognition errors. Summary of the invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is: the existing transformer partial discharge detection method has the problems of feature extraction information loss, low recognition accuracy, over-fitting of the recognition model, and how to improve the recognition accuracy of partial discharge fault features.

[0007] To solve the above technical problems, the present invention provides the following technical solutions: a local discharge feature fusion extraction method based on DenseBlock, comprising collecting original data through a first test and performing a first classification; preprocessing the classified graph; constructing a first recognition model, and performing classification and recognition effectiveness evaluation.

[0008] As a preferred solution of the DenseBlock-based partial discharge feature fusion extraction method described in the present invention, wherein: the acquisition of raw data through the first test includes generating a PRPD spectrum.

[0009] As a preferred solution of the partial discharge feature fusion extraction method based on DenseBlock described in the present invention, wherein: the first classification includes eliminating useless PRPD spectra and classifying high-quality PRPD spectra.

[0010] As a preferred solution of the partial discharge feature fusion extraction method based on DenseBlock described in the present invention, the preprocessing includes normalizing and graying the classified PRPD spectrum.

[0011] As a preferred solution of the partial discharge feature fusion extraction method based on DenseBlock described in the present invention, wherein: the preprocessing also includes dividing the preprocessed PRPD spectrum into a training set and a test set.

[0012] As a preferred solution of the DenseBlock-based partial discharge feature fusion extraction method described in the present invention, wherein: the construction of the first recognition model includes using a DenseBlock module and a convolutional neural network CNN to construct a DenseNet model to identify partial discharge fault characteristics in electrical equipment, and iteratively train and optimize the first recognition model through a training set.

[0013] As a preferred solution of the partial discharge feature fusion extraction method based on DenseBlock described in the present invention, the classification and recognition effectiveness evaluation includes inputting the test set into the trained first recognition model, outputting the classification result, comparing the output result with the true label of the test set, calculating the classification accuracy, F1 score, precision and recall rate, and generating a performance analysis report to ensure the generalization ability and recognition accuracy of the first recognition model.

[0014] Another object of the present invention is to provide a local discharge feature fusion extraction system based on DenseBlock, which can reduce the redundancy and complexity of the data by preprocessing the classified graph, so that the first recognition model can more effectively learn the key features in the data, thereby solving the problem of low recognition accuracy in the current transformer local discharge detection technology.

[0015] As a preferred solution of the local discharge feature fusion extraction system based on DenseBlock described in the present invention, it includes: a data acquisition module, a data preprocessing module, and a model building and evaluation module; the data acquisition module is used to collect original data through a first test and perform a first classification; the data preprocessing module is used to preprocess the classified map; the model building and evaluation module is used to build a first recognition model and perform classification and recognition effectiveness evaluation.

[0016] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a step of a local discharge feature fusion extraction method based on DenseBlock.

[0017] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a local discharge feature fusion extraction method based on DenseBlock.

[0018] Beneficial effects of the present invention: The DenseBlock-based partial discharge feature fusion extraction method provided by the present invention collects original data and generates a PRPD spectrum through a first test, thereby ensuring the comprehensiveness and reliability of the measured data, being able to effectively retain partial discharge feature information, reducing the risk of information loss, and classifying the PRPD spectrum. By eliminating useless spectra and retaining high-quality spectra for further analysis, the effectiveness of subsequent first recognition model training and feature extraction is guaranteed, and important features are ensured not to be omitted. By normalizing and graying the classified PRPD spectrum, the redundancy and complexity of the data are reduced, so that the subsequent first recognition model can more effectively learn the key features in the data, thereby improving the recognition accuracy. By introducing the DenseBlock module into the convolutional neural network to construct the DenseNet model, each layer can directly access the features of all previous layers, thereby enhancing the fluidity of information and reducing the risk of gradient disappearance. The feature reuse mechanism of DenseNet improves the expression ability of the first recognition model and improves information transmission while reducing the number of parameters. This design effectively reduces the overfitting phenomenon and improves the accuracy and reliability of fault identification. The present invention achieves better results in terms of accuracy, safety and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0020] Figure 1 An overall flow chart of a local discharge feature fusion extraction method based on DenseBlock provided in the first embodiment of the present invention.

[0021] Figure 2 A schematic flow chart of a partial discharge feature fusion extraction method based on DenseBlock provided in the second embodiment of the present invention.

[0022] Figure 3 A DenseBlock module schematic diagram of a DenseBlock-based partial discharge feature fusion extraction method provided in the second embodiment of the present invention.

[0023] Figure 4 A module schematic diagram of a local discharge feature fusion extraction system based on DenseBlock provided in the third embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0025] Example 1, reference Figure 1 , is an embodiment of the present invention, and provides a local discharge feature fusion extraction method based on DenseBlock, comprising:

[0026] S1: Collect original data through the first experiment and perform the first classification.

[0027] Furthermore, collecting raw data through the first experiment includes generating a PRPD spectrum.

[0028] Among them, the PRPD spectrum is a two-dimensional spectrum formed by combining the partial discharge pulse with the power frequency phase, one of the coordinate axes represents the power frequency phase The other axis represents the discharge quantity (q), which is often used to visualize and analyze the partial discharge characteristics of electrical equipment, which helps in fault diagnosis and assessment of equipment health status.

[0029] Collecting the original data through the first test may be collecting the partial discharge signal of the transformer through a high voltage excitation signal, or may be collecting the partial discharge signal of the transformer using a spectrum analysis method, or other methods capable of collecting the partial discharge signal of the transformer.

[0030] It should be noted that, in the embodiment of the present application, collecting raw data through the first test is to collect the partial discharge signal of the transformer through the high voltage excitation signal, and the specific steps are:

[0031] By applying a high voltage excitation signal, the insulation material of the transformer may produce partial discharge under certain working conditions, affecting the long-term stability and safety of the equipment. Apply a high voltage to the transformer, such as using a booster to increase the voltage to 150% of the rated value, to ensure that potential partial discharge can be stimulated. During this process, highly sensitive sensors such as ultrasonic sensors and current transformers are installed to capture current spikes and acoustic signals caused by partial discharge, which usually have weak amplitudes and are difficult to identify. Next, the data acquisition system is used to record the captured discharge signals in real time, and digitally process them to construct a phase-amplitude distribution diagram (PRPD spectrum). Further analysis of the PRPD spectrum can extract characteristic information of the discharge, such as discharge frequency, discharge amplitude and location, thereby providing a reliable data basis for subsequent fault diagnosis.

[0032] It should be noted that, in an optional embodiment, collecting raw data through the first test is to collect partial discharge signals of the transformer using a spectrum analysis method, and the specific steps are:

[0033] The use of spectrum analysis methods to collect partial discharge signals of transformers involves in-depth analysis of the current or voltage signals generated during the operation of the transformer. Through high-precision spectrum analysis instruments, the current curves generated by the equipment under normal working conditions are regularly recorded to accurately capture the high-frequency fluctuation characteristics in the signal. The signal amplitude changes under equipment operating conditions are often very weak, usually at the millivolt level, but through spectrum analysis, characteristic peaks in the frequency domain can be identified, which correspond to different types of partial discharge activities. By calculating and drawing the phase-amplitude distribution diagram (PRPD spectrum), the frequency information is associated with the time characteristics of the discharge activity, which can effectively obtain the frequency range of partial discharge and the fluctuation characteristics of the discharge activity, providing a quantitative basis for the degradation assessment of insulation materials and the prediction of equipment failures.

[0034] Furthermore, the first classification includes eliminating useless PRPD spectra and classifying high-quality PRPD spectra.

[0035] Among them, eliminating useless PRPD spectra includes removing useless spectra with unclear features or affected by noise.

[0036] The classification of high-quality PRPD spectra includes classifying the remaining PRPD spectra after eliminating useless PRPD spectra into suspended discharge, corona discharge, insulation discharge, surface discharge, etc. according to the type of partial discharge that may occur in the equipment, and labeling each type for identification.

[0037] The first classification may be performed to manually eliminate useless PRPD spectra based on threshold setting, or to eliminate invalid spectra interfered by noise based on signal processing technology, or other methods that can eliminate useless PRPD spectra.

[0038] It should be noted that in the embodiment of the present application, the first classification is performed based on threshold setting to manually eliminate useless PRPD spectra, and the specific steps are:

[0039] First, analysts should define multiple indicators, such as amplitude, frequency, and waveform characteristics, and establish clear standards to screen PRPD spectra. Specifically, the amplitude threshold is set to 0.1 millivolts. If a spectrum fails to reach this amplitude during the signal process, it can be marked as invalid. In addition, the analysis of frequency components also needs attention. For example, spectra with a main frequency below 500Hz are regarded as low-quality spectra that may contain noise. Then, for residual spectra that meet the standards, they are classified through software tools, and the spectra are divided into categories such as suspended discharge, corona discharge, insulation discharge, and surface discharge according to their discharge characteristics. At the same time, specific labels are marked for each category to provide a basis for subsequent fault diagnosis and trend analysis.

[0040] It should be noted that, in an optional embodiment, the first classification is performed by eliminating invalid spectra with noise interference based on signal processing technology, and the specific steps are:

[0041] First, the signal is decomposed by wavelet transform to extract detailed information in different frequency ranges. This method can effectively distinguish low-frequency signals from high-frequency noise, and then filter out the noise. Subsequently, Fourier transform can be used as a tool to further analyze the signal, and the main signal components can be identified through the spectrum diagram. After denoising, the preset feature extraction algorithm, such as the peak position and amplitude change of the signal, is used to identify high-quality spectra with obvious features. Finally, these high-quality PRPD spectra are integrated and classified according to different partial discharge types, including suspended discharge, corona discharge, insulation discharge and surface discharge, and each discharge type is labeled to achieve the purpose of effective identification, judgment and detailed electrical discharge behavior analysis.

[0042] It should also be noted that the collection of raw data in the first test, including the generation of PRPD spectra, can achieve a preliminary assessment of the state of electrical equipment and visualization of fault characteristics. Specifically, by applying a high-voltage excitation signal, the local discharge phenomenon in the transformer insulation material is stimulated, thereby capturing weak current spikes and acoustic signals. This highly sensitive data acquisition method can reveal the potential failure risk of electrical equipment under high voltage conditions, thereby providing an important basis for subsequent fault diagnosis. Through the generated PRPD spectra, analysts can clearly identify characteristic information such as discharge frequency, amplitude, and discharge geographical location. The construction of the PRPD spectrum not only enhances the understanding of the local discharge phenomenon, but also lays a data foundation for subsequent classification and model training, promotes efficient fault monitoring and equipment health status assessment, and thus achieves timely maintenance and reduces the risk of failure.

[0043] S2: Preprocess the classified PRPD map.

[0044] Furthermore, the preprocessing includes normalizing and gray-scaling the classified PRPD atlas.

[0045] The normalization and grayscale processing includes scaling the classified PRPD maps to the same size, normalizing and grayscale the images, and grayscale processing, which is expressed as:

[0046]

[0047] Among them, A ij is the gray value of the pixel at the image coordinate H(i,j), n ij is the number of discharges at the image coordinate H(i,j), n max is the maximum number of discharges in the image coordinate H(i,j).

[0048] It should be noted that the preprocessing also includes dividing the preprocessed PRPD atlas into a training set and a test set.

[0049] It should also be noted that preprocessing, including normalization and grayscale conversion, can ensure that the data input to the model always remains consistent and comparable. Specifically, by scaling the maps to the same size and performing grayscale conversion, the standardization of the data is ensured, thereby providing a balanced input basis for subsequent model training. In addition, the preprocessing step includes dividing the maps into training sets and test sets to ensure data independence during training and objectivity of model evaluation. This segmentation method can effectively reduce the risk of overfitting and enhance the generalization ability of the model, thereby ensuring the accuracy and reliability of subsequent classification and recognition, thereby improving the performance and application scope of the electrical equipment fault recognition model.

[0050] S3: Construct the first recognition model and evaluate the effectiveness of classification recognition.

[0051] Furthermore, constructing the first recognition model includes using a DenseBlock module and a convolutional neural network (CNN) to construct a DenseNet model to identify partial discharge fault characteristics in electrical equipment, and iteratively training and optimizing the first recognition model through a training set.

[0052] The first identification model may be constructed by using a DenseBlock module and a convolutional neural network (CNN) to construct a DenseNet model, or may be an adaptive fault identification model based on reinforcement learning, or other methods capable of identifying partial discharge fault characteristics in electrical equipment.

[0053] It should be noted that, in the embodiment of the present application, the first recognition model is constructed by using a DenseBlock module and a convolutional neural network CNN to construct a DenseNet model, which specifically includes:

[0054] The DenseBlock module is the core part of the DenseNet network. It improves the efficiency of information flow and gradient propagation through dense connections, allowing the network to reuse features more efficiently, reduce the gradient vanishing problem, and improve the parameter utilization efficiency of the network.

[0055] In the DenseBlock module, the output of each layer is not only passed to the next layer, but also concatenated with the output of all previous layers to form a new input. Assuming that DenseBlock has n layers and the input feature is I0, the specific steps include:

[0056] A1. For the nth layer, the output of all previous layers is used as the input of the nth layer, expressed as:

[0057] I n =Concat(x0,x1,...,x n-1 )

[0058] Among them, I n is the input of the nth layer, Concat is the concatenation of the outputs of all previous layers in the channel dimension, x n-1 is the output of the n-1th layer.

[0059] A2. Input I of the nth layer n After batch normalization, ReLU activation function and convolution operation, the output x is obtained l , expressed as:

[0060]

[0061] Among them, Conv is the convolution operation of the CNN model on the data, ReLU is the activation function, and BN(I n ) and 0, BN(·) is used to normalize the output in the network.

[0062] A3, repeat steps A1 and A2, the output x of the nth layer n It will be used as one of the inputs of the subsequent layers and continue to be spliced, expressed as:

[0063] I n+1 =Concat(x0,x1,...,x n )

[0064] This connection method allows each layer to receive the feature outputs of all previous layers, concatenate features layer by layer, and continuously increase the input dimension, forming a "dense connection" mode.

[0065] Furthermore, after adding the DenseBlock module to CNN, the DenseNet model is iteratively trained with the training set to initialize the model parameters and training hyperparameters, including learning rate, batch size, etc., and the model is optimized using the cross entropy loss function to calculate the error between the model output and the true label; the network parameters are adjusted through the back propagation algorithm to reduce the training error and achieve the best training effect.

[0066] It should be noted that, in an optional embodiment, constructing the first recognition model is an adaptive fault recognition model based on reinforcement learning, and the specific steps are:

[0067] First, by setting a clear reward function to measure the recognition accuracy, when the model correctly identifies the fault features, it will receive a positive reward, and misidentification will be punished. This feedback mechanism prompts the model to improve its discrimination ability in each learning iteration, so that after multiple consecutive iterations, the recognition accuracy can be increased to more than 95%. Furthermore, by establishing a structure that combines a multi-layer neural network with reinforcement learning, each layer can self-update feature selection when receiving a new signal. This dynamic feature selection capability ensures that the model can be flexibly adjusted to different environmental noises and equipment states. For example, in a high-noise environment, the model prioritizes the robustness of the feature rather than the detailed information of the original signal to reduce the impact of noise interference on the recognition results. Ultimately, the introduction of reinforcement learning enables the model to maintain at least 90% recognition effectiveness in a variety of scenarios, can adapt to the emergence of new fault modes, make the fault detection process of electrical equipment more intelligent and efficient, and greatly reduce maintenance costs and time.

[0068] It should also be noted that the effectiveness evaluation of classification and recognition includes inputting the test set into the trained first recognition model, outputting the classification results, comparing the output results with the true labels of the test set, calculating the classification accuracy, F1 score, precision and recall, and generating a performance analysis report to ensure the generalization ability and recognition accuracy of the first recognition model.

[0069] It should also be noted that by using the DenseBlock module and the convolutional neural network CNN to build the DenseNet model, the accuracy and efficiency of fault identification can be effectively improved. The dense connection method of DenseNet enables each layer to fully reuse the features of the previous layer, thereby reducing the complexity and training time of the model, while avoiding the common gradient vanishing problem. This structure is particularly suitable for identifying partial discharge features, because each layer transmits information from all previous layers, ensuring the richness and expressiveness of the features. After the training is completed, the effectiveness of classification and recognition is evaluated, and the accuracy of the model is rigorously quantitatively analyzed through the test set, including classification accuracy, F1 score, precision and recall rate and other indicators, to ensure the reliability and effectiveness of the model in practical applications. This process not only provides transparency of model performance, but also provides a basis for subsequent applications and further optimization, which helps to achieve accurate electrical equipment fault identification, reduce maintenance and operating costs, and improve the safety and stability of the overall system.

[0070] Example 2, reference Figure 2-Figure 3 , which is an embodiment of the present invention, provides a local discharge feature fusion extraction method based on DenseBlock. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0071] like Figure 2 As shown, it is a flow chart of a partial discharge feature fusion extraction method based on DenseBlock in this embodiment. First, the transformer pulse discharge information is collected and a PRPD spectrum is generated, and then the data is screened and preprocessed to ensure data quality. Next, the processed data is divided into a training set and a test set to prepare for the training and evaluation of the model. Then, a convolutional neural network (CNN) model is constructed and a DensenBlock module is added to improve performance. Subsequently, the model framework structure is adjusted and the effectiveness of the model is evaluated. According to the evaluation results, it is decided whether the model needs to be further optimized. If no optimization is required, the final model is saved.

[0072] Specifically, first, high-sensitivity sensors, including ultrasonic sensors and current transformers, are set up to capture weak current spikes and acoustic signals caused by partial discharge. The data acquisition system records, stores and digitally processes in real time to generate a PRPD spectrum. The two coordinate axes of the spectrum represent the power frequency phase and the discharge amount, respectively, which can intuitively reflect the partial discharge characteristics of the equipment.

[0073] According to the set standards, the first classification is performed. In this embodiment, the standards for eliminating useless spectra include: the amplitude threshold is 0.1 millivolts, if the signal does not reach this amplitude, it is marked as invalid; the detection standard of the frequency component is that the main frequency is higher than 500Hz to identify potential noise interference. The software tool is used to classify the spectra that meet the standards, and according to their discharge characteristics, they are divided into suspended discharge, corona discharge, insulation discharge and surface discharge, and each category is marked for subsequent analysis and research.

[0074] After classification, the high-quality PRPD atlas was preprocessed. The preprocessing included scaling the classified atlas to 256×256 pixels and graying it. Each pixel in the image was normalized to a value between 0 and 1, while ensuring that the mean of the input feature was 0 and the standard deviation was 1. Then, the preprocessed atlas was randomly split into a training set and a test set in a ratio of 7:3 to maintain the randomness and diversity of the data in the subsequent model training.

[0075] In the model construction stage, a convolutional neural network (CNN) combined with a DenseBlock module is used to build a DenseNet model. The principle of the DenseBlock module is as follows Figure 3 As shown, the efficiency of feature information flow and gradient propagation is improved by dense connection, the risk of gradient disappearance is reduced, and the network's ability to reuse features is improved. The configuration of model training hyperparameters includes setting the batch size to 32, the initial learning rate to 0.02, and a total of 100 training cycles. In order to evaluate the difference between the model prediction results and the actual labels, the cross entropy loss function is used as the classification loss. The optimization algorithm uses random gradient, and the goal is to minimize the loss function. In order to dynamically adjust the learning rate and optimize the training effect, the model introduces a learning rate scheduler StepLR, which decays the learning rate by 0.9 times after every 10 training cycles. Through the dynamic learning rate update strategy, the convergence speed and classification accuracy of the model are effectively improved. In addition, this embodiment also adds global average pooling and Dropout mechanisms to effectively prevent overfitting while reducing the number of parameters, further enhancing the generalization ability of the model.

[0076] Finally, the test set is input into the trained DenseNet model, the classification results are output, and the classification accuracy, F1 score, precision and recall are calculated according to the true labels. A performance analysis report is generated to evaluate the effectiveness and generalization ability of the constructed model.

[0077] Table 1 Comparison of model effects

[0078] The DenseBlock module model is not added Add DenseBlock module model F1 score 90.897％ 98.864％ Accuracy 92.371％ 98.897％ Recall 91.480％ 98.875％ Loss value 7.904 1.113

[0079] Through the above steps, Table 1 shows the specific performance of the model before and after adding the DenseBlock module. By comparing the experimental results, it can be seen that the DenseNet model with the DenseBlock module is superior to the model without DenseBlock in various performance indicators. The introduction of the DenseBlock module greatly improves the feature reuse capability and information fluidity of the model, so that the F1 score is increased from 90.897% to 98.864%, and the precision and recall rates are 98.897% and 98.875% respectively, which improves the accuracy and comprehensiveness of fault identification. At the same time, the loss value is greatly reduced from 7.904 to 1.113, indicating that the DenseNet model has not only made a qualitative leap in classification accuracy, but also effectively reduced errors and overfitting phenomena, ensuring the generalization ability of the model in different types of partial discharge.

[0080] This example fully proves that the DenseBlock module has shown outstanding advantages in optimizing gradient propagation, reducing parameter redundancy and improving model performance. The local discharge feature fusion extraction method based on DenseBlock proposed in the present invention has extremely high application value in local discharge fault detection.

[0081] Example 3, reference Figure 4 , which is an embodiment of the present invention, provides a local discharge feature fusion extraction system based on DenseBlock, including a data acquisition module, a data preprocessing module, and a model building and evaluation module.

[0082] The data acquisition module is used to collect original data through the first experiment and perform the first classification; the data preprocessing module is used to preprocess the classified graph; the model building and evaluation module is used to build the first recognition model and perform classification and recognition effectiveness evaluation.

[0083] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0084] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0085] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0086] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc. It should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not limited. Although the present invention is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

[0087] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A local discharge feature fusion extraction method, characterized in that: include: Collecting original data through the first experiment and performing the first classification; Preprocess the classified graphs; Construct the first recognition model and evaluate the effectiveness of classification recognition.

2. The local discharge feature fusion extraction method according to claim 1, characterized in that: The collecting of raw data through the first experiment includes generating a spectrum.

3. The local discharge feature fusion extraction method according to claim 2, characterized in that: The first classification includes classifying the atlas.

4. The local discharge feature fusion extraction method according to claim 3, characterized in that: The preprocessing includes performing data processing on the classified graphs.

5. The local discharge feature fusion extraction method according to claim 4, characterized in that: The preprocessing further includes dividing the preprocessed atlas into data sets.

6. The local discharge feature fusion extraction method according to claim 5, characterized in that: The construction of the first recognition model includes using a DenseBlock module and a convolutional neural network (CNN) to construct a DenseNet model to identify partial discharge fault characteristics in electrical equipment, and iteratively training and optimizing the first recognition model through a training set.

7. The local discharge feature fusion extraction method according to claim 6, characterized in that: The classification and recognition effectiveness evaluation includes inputting the test set into the trained first recognition model, outputting the classification result, comparing the output result with the true label of the test set, calculating the classification accuracy, F1 score, precision and recall rate, and generating a performance analysis report to ensure the generalization ability and recognition accuracy of the first recognition model.

8. A system using the partial discharge feature fusion extraction method according to any one of claims 1 to 7, characterized in that: It includes data acquisition module, data preprocessing module, model building and evaluation module; The data collection module is used to collect raw data through a first test and perform a first classification; The data preprocessing module is used to preprocess the classified graphs; The model building and evaluation module is used to build a first recognition model and perform classification and recognition effectiveness evaluation.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the partial discharge feature fusion extraction method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the partial discharge feature fusion extraction method according to any one of claims 1 to 7 are implemented.

Images