A high voltage control cabinet with partial discharge detection device

Abstract

The invention discloses a high-voltage control cabinet with a partial discharge detection device, relates to the field of local monitoring technology, and solves the problems of detection sensitivity and positioning accuracy of existing high-voltage control cabinets; the invention dynamically identifies and eliminates noise interference through an adaptive digital filtering algorithm and a deep convolutional neural network, thereby ensuring the accuracy of signal analysis; performs type recognition and intensity evaluation on extracted features through an adaptive multidimensional spectrum clustering discharge recognition algorithm to effectively distinguish different types of discharges, thereby improving the accuracy of recognition and the reliability of intensity evaluation; realizes high-precision three-dimensional positioning of partial discharge sources through a three-dimensional precise positioning module, thereby significantly improving the accuracy and stability of positioning; uses big data analysis and machine learning prediction models to predict future failure trends of equipment, thereby reducing the operating risks caused by failures and accelerating the speed of problem locating; the invention greatly improves the sensitivity and positioning accuracy of partial discharge detection in existing high-voltage control cabinets.

Description

Technical Field

[0001] The present invention relates to the technical field of local monitoring, and more particularly to a high-voltage control cabinet with a partial discharge detection device. Background Art

[0002] With the rapid development of power systems and the increasing requirements for power supply reliability, high-voltage control cabinets, as key equipment in power systems, are of vital importance for safe and stable operation. Partial discharge (PD) is one of the main reasons for the degradation of insulation performance of high-voltage control cabinets, which may eventually lead to equipment failure or even power grid accidents. Therefore, partial discharge detection devices that can monitor and accurately diagnose partial discharge in real time have emerged and become one of the key technologies to ensure the long-term stable operation of high-voltage control cabinets. The development of partial discharge detection technology aims to avoid potential large-scale power outages and economic losses by detecting insulation defects at an early stage and taking preventive maintenance measures.

[0003] At present, partial discharge detection technology has become an indispensable part of high-voltage control cabinets. For example, the detection technology based on transient earth voltage (TEV) identifies partial discharge by monitoring the transient voltage signal generated by partial discharge, and has been widely used in actual high-voltage control cabinets. In addition, ultra-high frequency (UHF) sensing technology uses ultra-high frequency electromagnetic waves generated by partial discharge for detection, effectively avoiding on-site interference and improving the sensitivity and anti-interference ability of detection. The application of these technologies has significantly improved the ability to monitor the internal insulation status of high-voltage control cabinets and provided important data support for preventive maintenance.

[0004] However, this method still has disadvantages. For example, the TEV signal may be affected by the internal structure and material properties of the switch cabinet, resulting in a complex signal propagation path and difficulty in accurately determining the discharge location. In addition, in a complex electromagnetic environment, the TEV signal may be confused with other electromagnetic interference signals, affecting the accuracy of detection; UHF detection may be affected by the shielding effect of the metal shell, resulting in low positioning accuracy of the local discharge point and incomplete signal capture.

[0005] Therefore, the present invention discloses a high-voltage control cabinet with a partial discharge detection device. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention discloses a high-voltage control cabinet with a partial discharge detection device. The present invention can improve the detection sensitivity and positioning accuracy of the local discharge points of the existing high-voltage control cabinet with the partial discharge detection device.

[0007] In order to achieve the above technical effects, the present invention adopts the following technical solutions:

[0008] A high-voltage control cabinet with a partial discharge detection device comprises a data acquisition module and a cloud-based intelligent control platform; the data acquisition module is used to capture partial discharge signals and environmental data inside and outside the high-voltage control cabinet, and transmit them to the cloud-based intelligent control platform; the data acquisition module realizes the acquisition and perception of partial discharge signals in the high-voltage control cabinet through a multi-dimensional sensor array and a signal synchronization method; the multi-dimensional sensor array covers different types of discharge signals by deploying ultrasonic sensors, ultra-high frequency sensors, high-frequency current transformers, acoustic emission sensors, gas detection sensors and photoelectric sensors; the signal synchronization method ensures the clock synchronization of all sensor data through the network time protocol NTP;

[0009] The cloud-based intelligent control platform is used to receive and process partial discharge signals and environmental data from a multi-dimensional sensor array, and to perform real-time monitoring, intelligent diagnosis, health prediction, precise positioning and remote control of the high-voltage control cabinet; the cloud-based intelligent control platform includes an intelligent noise reduction module, a partial discharge feature deep mining module, a three-dimensional precise positioning module, a health prediction and analysis module and a cloud management module;

[0010] The intelligent noise reduction module is used to enhance the purity of the partial discharge signal and environmental data captured by the data acquisition module; the intelligent noise reduction module removes the noise interference generated by the environment and the system itself through an adaptive digital filtering algorithm and a noise model learning method; the noise model learning method learns historical noise data through a deep convolutional neural network and constructs a noise model to predict and eliminate noise; the denoised data is output to the partial discharge feature deep mining module;

[0011] The partial discharge feature deep mining module is used to realize discharge type recognition and intensity evaluation based on the partial discharge signal and environmental data enhanced by the intelligent noise reduction module, and obtain discharge type recognition and intensity evaluation results; the partial discharge feature deep mining module extracts features from the partial discharge signal through deep feature joint learning and adaptive feature enhancement network; the deep feature joint learning realizes cross-modal integration and enhancement of features through multimodal fusion, deformable convolution network and attention mechanism; the adaptive feature enhancement network realizes dynamic enhancement of partial discharge features through gating mechanism, feature enhancement method and progressive network structure;

[0012] The three-dimensional precise positioning module and the health prediction and analysis module are used to locate the spatial position of the partial discharge source based on the discharge type identification and intensity assessment results, evaluate the health status of the equipment, and predict potential faults; the three-dimensional precise positioning module realizes the three-dimensional spatial positioning of the partial discharge source through time difference arrival positioning TDOA, signal strength analysis and acoustic imaging; the time difference arrival positioning TDOA calculates the distance through the time difference of sensor signals; the signal strength analysis assists positioning through signal strength differences; the acoustic imaging further refines the positioning results through sound wave reflection and interference effects; the health prediction and analysis module conducts in-depth mining of historical monitoring data through big data analysis and machine learning prediction models, evaluates the health status of equipment and predicts future fault trends;

[0013] The cloud management module is used to realize the operation interface display, data storage and management, real-time intelligent alarm notification, and remote operation, monitoring and configuration management of the high-voltage control cabinet monitoring platform.

[0014] As a further technical solution of the present invention, the data acquisition module includes a sensor array integrated unit, a signal synchronization unit and a data transmission unit; the sensor array integrated unit includes an ultrasonic sensor unit, an ultra-high frequency sensor unit, a high-frequency current mutual inductance unit, an acoustic emission sensor unit, a gas detection sensor unit and a photoelectric sensor unit; the ultrasonic sensor unit realizes directional capture and positioning of ultrasonic signals through a beam forming method; the ultra-high frequency sensor unit adopts broadband amplification to enhance the receiving ability of ultra-high frequency electromagnetic waves; the high-frequency current mutual inductance unit measures and converts transient signals in high-voltage current through the principle of electromagnetic induction; the acoustic emission sensor unit captures tiny stress waves caused by partial discharge through acoustic emission detection; the gas detection sensor unit detects the gas inside the control cabinet through a chemical sensor The photoelectric sensing unit detects the light signal changes caused by partial discharge by adopting the photoelectric conversion mechanism; the signal synchronization unit includes a time synchronization subunit and a data synchronization subunit; the time synchronization subunit ensures the consistency of the signal time tag through the network time protocol NTP; the network time protocol NTP ensures the sensor time synchronization through the time calibration mechanism, the hierarchical time server structure and the timestamp; the data synchronization subunit realizes the parallel acquisition and synchronous storage of data through the synchronous data acquisition card; the data transmission unit includes a data encoding subunit and a communication interface subunit; the data encoding subunit improves the reliability and efficiency of data transmission through error control coding; the communication interface subunit uploads the monitoring data to the cloud server in real time through the Internet of Things communication protocol MQTT.

[0015] As a further technical solution of the present invention, the working steps of the intelligent noise reduction module in a high-voltage control cabinet with a partial discharge detection device are as follows:

[0016] Step 1: Data reception and preprocessing;

[0017] The real-time collected data of the data acquisition module is obtained through the data interface, and the collected signal is pre-processed through digital filtering and signal conditioning methods; the digital filtering removes high-frequency noise through a low-pass filtering mechanism; the signal conditioning method realizes signal amplitude adjustment and digitization through an amplifier and an analog-to-digital converter;

[0018] Step 2: Time-frequency analysis and feature extraction;

[0019] The characteristics of the local discharge signal and the background noise are distinguished by wavelet transform and short-time Fourier transform. The wavelet transform can capture the time-varying characteristics of the signal through multi-scale decomposition and sliding window mechanism; the short-time Fourier transform can analyze the frequency components of the signal through local window processing.

[0020] Step 3: Adaptive digital filtering;

[0021] Based on the result of feature extraction, the filter coefficients are dynamically adjusted according to the signal characteristics through the minimum mean square error and recursive least squares algorithm. The minimum mean square error and recursive least squares algorithm suppress non-partial discharge noise in real time through the gradient descent method and the fast convergence algorithm to improve the signal purity;

[0022] Step 4: Noise model learning;

[0023] Automatically identify noise patterns from signals and construct noise models through supervised learning algorithms and feature selection strategies; the supervised learning algorithms optimize model parameters through iterative learning of training set data; the feature selection strategies screen feature signals through statistical testing methods;

[0024] Step 5: Noise prediction and elimination;

[0025] As the noise model is updated in real time, noise is removed from the original signal through signal subspace estimation, and the noise model realizes the prediction output of noise through the inference engine of the machine learning model; the signal subspace estimation realizes the separation of signal and noise through singular value decomposition SVD;

[0026] Step 6: Performance evaluation and feedback;

[0027] The quantitative evaluation of noise reduction effect and the dynamic adjustment of system parameters are realized through performance evaluation indicators and feedback control mechanism; the performance evaluation indicators measure the noise reduction effect through mathematical statistical methods, and the feedback control mechanism realizes adaptive adjustment of parameters through a closed-loop control system; the performance evaluation indicators include signal-to-noise ratio and mean square error;

[0028] Step 7: Model self-learning and updating;

[0029] The continuous learning and knowledge updating of the noise model for new data are achieved through online learning algorithms and model fusion strategies; the model fusion strategy achieves model knowledge fusion and optimization through an integrated learning method.

[0030] 1) As a further technical solution of the present invention, the signal is preliminarily purified and standardized by the signal preprocessing layer; the signal preprocessing layer removes the DC offset and baseline drift in the signal by high-order statistical analysis and spectrum purification algorithm, and uses normalization processing to unify the signal amplitude to the same scale; the high-order statistical analysis identifies and removes non-Gaussian noise by cumulative estimation; the spectrum purification algorithm filters out interference in specific frequency bands by frequency band filtering method; the normalization processing ensures data consistency by maximum and minimum value scaling;

[0031] 2) learning and extracting local features in the signal through the convolution layer; the convolution layer slides on the input signal through the convolution operation, and the weight of each convolution kernel is automatically adjusted through learning to match the specific features, and the shared weight is used to reduce the number of parameters to improve efficiency and generalization ability;

[0032] 3) The feature enhancement and noise reduction layer is used to strengthen the signal features while weakening the noise impact; based on feature extraction, the feature enhancement and noise reduction layer enhances the relevant features of the partial discharge signal through sparse coding and principal component analysis, and reconstructs the signal through the autoencoder to minimize the reconstruction error;

[0033] 4) Reducing the dimension of the feature map through the pooling layer;

[0034] 5) The extracted features are integrated through the fully connected layer to perform high-level abstraction; the fully connected layer learns the complex relationship between features through multi-layer neuron full connection, and enhances the model's expressiveness through the ReLU activation function;

[0035] 6) predicting noise characteristics through the model layer; the model layer simultaneously predicts the noise type, intensity and trend through multi-task learning, and performs probability prediction through Bayesian network;

[0036] 7) According to the comparison between the model prediction results and the actual noise situation, the model parameters are dynamically adjusted through the optimization layer; the optimization layer calculates the loss function gradient through the back propagation algorithm, and updates the network weights through the optimization algorithm SGD.

[0037] As a further technical solution of the present invention, the local discharge feature deep mining module includes a feature integration submodule, a feature enhancement submodule, and a discharge evaluation submodule; the feature integration submodule includes a multimodal fusion unit, a feature mapping unit, and an attention mechanism unit; the multimodal fusion unit realizes the integration of different sensor data through cross-modal association learning; the feature mapping unit realizes the spatial adaptation of input data through the spatial transformation network STN to adapt to features of different shapes; the spatial transformation network STN dynamically adjusts the shape and size of the convolution kernel according to signal characteristics through an adaptive convolution mechanism; the attention mechanism unit optimizes the feature selection process through an attention weighted strategy; The feature enhancement submodule includes an information screening unit, a weight optimization unit and a progressive network structure unit; the information screening unit dynamically adjusts the information flow of the network through the gated recurrent unit GRU to optimize the feature transmission and enhancement process; the weight optimization unit determines the importance of the feature and weights it through a feature weighting algorithm; the feature weighting algorithm selects features through particle swarm optimization and weights the selected features through the feature importance in the random forest; the progressive network structure unit improves the training effect of the deep network through residual connections; the discharge evaluation submodule performs type recognition and intensity evaluation on the extracted features through an adaptive multidimensional spectral clustering discharge recognition algorithm.

[0038] As a further technical solution of the present invention, the working method of the adaptive multi-dimensional spectrum clustering discharge identification algorithm includes the following steps:

[0039] Step 1: Multi-dimensional feature spectrum embedding;

[0040] The high-dimensional features are mapped to the low-dimensional space through the Laplace eigenmapping function, while retaining the local structural characteristics of the data. The formula expression of the Laplace eigenmapping function is:

[0041]

[0042] In formula (1), h t is the feature matrix after spectral embedding, and the column vector is the representation after the original high-dimensional feature space is mapped to the low-dimensional feature space; W is the degree matrix, where the elements D on the diagonal of the degree matrix W are ii represents the degree of node i, that is, the number of edges connected to node i; L represents the Laplacian matrix; W -1 / 2 is the semi-inverse matrix of the degree matrix, which is used to balance the degrees of different nodes; L ij represents the connection strength between nodes i and j;

[0043] Step 2: Adaptive weighted spectral clustering;

[0044] According to the spectral embedding features obtained in step 1, an adaptive weighted strategy is used for spectral clustering. The adaptive weighted strategy enhances the clustering effect by dynamically adjusting the weights of categories. The adaptive weighted spectral clustering formula is:

[0045]

[0046] In formula (2), X JL is the clustering weight matrix, which is used to quantify the similarity or distance between samples; B is the basis matrix of the embedding space, where the column vector of the basis matrix B is the eigenvector of the Laplace matrix L, corresponding to the non-zero eigenvalue; T is the eigenvalue diagonal matrix, and the elements on the diagonal of the eigenvalue diagonal matrix T are γ i is the eigenvalue of the Laplace matrix L;

[0047] Step 3: Discharge type identification based on spectrum;

[0048] Using the clustering results obtained in step 2, the discharge type is identified by a spectrum-based method; the spectrum-based method achieves classification by the distance between the spectral features and the known types; the spectral distance formula for discharge type identification is:

[0049]

[0050] In formula (3), x sp is the spectral distance, which is used to measure the distance between the sample and the center of the known discharge type in the spectral embedding space; γ j is the spectral feature value of the jth feature of the sample; C i is the center of the i-th known discharge type; h t represents the feature matrix after spectrum embedding; u ij 2 Represents the average of the spectral feature values ​​of the jth feature of the i-th class of known samples;

[0051] Step 4: Multi-scale evaluation of discharge intensity;

[0052] On the basis of the discharge type identification, the discharge intensity is quantitatively analyzed by a multi-scale evaluation method; the multi-scale evaluation method evaluates the severity of the discharge by combining the spectral features of different time scales; the discharge intensity evaluation formula is:

[0053]

[0054] In formula (4), P g is the discharge intensity index; P ref is the reference discharge intensity, used to standardize the discharge intensity evaluation benchmark value; k is the number of scales; Q represents the time point considered at each scale; e ktis the spectral eigenvalue of the kth scale at time t, which is used to measure the characteristics of the discharge at time t; w kt is the weighting coefficient at scale k and time t, which is used to adjust the contribution of different scales and time points to the discharge intensity assessment;

[0055] Step 5: Adaptive iterative optimization;

[0056] Finally, the spectral embedding and clustering parameters are continuously adjusted through the adaptive iterative optimization process to adapt to new data changes and improve recognition accuracy; the iterative optimization formula is:

[0057]

[0058] In formula (5), f new are the updated model parameters; f are the model parameters, including the learnable parameters in the process of spectral clustering, discharge type identification and intensity evaluation; m is the learning rate, which is used to control the step size of parameter update; p is the state-action distribution under the current strategy; y is the strategy network, which is used to measure the probability of taking an action in a given state; f old Current model parameters; Based on the current parameter f old The gradient of the error function e is used to indicate the direction and magnitude of parameter update.

[0059] As a further technical solution of the present invention, the three-dimensional precise positioning module includes a signal acquisition submodule, a data processing submodule and an advanced analysis submodule; the signal acquisition submodule includes a sensor network unit and a time synchronization unit; the sensor network unit uses a distributed ultra-wideband electromagnetic wave sensor and a high-frequency response ultrasonic sensor to capture transient signals generated by partial discharge; the time synchronization unit uses an atomic clock as a master clock source and uses a time protocol PTP IEEE 1588 synchronized sensor network units collect data time tags; the data processing submodule includes a TDOA calculation unit and a signal strength analysis unit; based on the timestamp information of the received signal, the TDOA calculation unit infers the distance from the discharge source to each sensor from the time difference through Kalman filtering, thereby building a preliminary framework of the spatial position; the signal strength analysis unit extracts directional information from the signal strength attenuation through an attenuation model and a pattern recognition algorithm; the advanced analysis submodule performs acoustic imaging processing by integrating the acoustic wave sensor array with the image reconstruction engine; the acoustic wave sensor array captures the interference and diffraction effects of the discharge sound waves, and the sound pressure level data of each point is collected and sent to the image reconstruction engine; the image reconstruction engine generates an acoustic image of the discharge area through an inverse scattering algorithm based on deep learning.

[0060] As a further technical solution of the present invention, the inverse scattering algorithm based on deep learning includes an input layer, a data layer, a feature layer, an inverse scattering coding layer, a circulation layer, a decoding and reconstruction layer, and an output layer; the working steps of the inverse scattering algorithm based on deep learning are:

[0061] S1. Data preparation: receiving the original sound pressure level data collected by the acoustic wave sensor array through the input layer, and performing denoising, normalization and spatiotemporal alignment on the input data through the data layer to eliminate the influence of environmental noise and sensor differences;

[0062] S2, feature learning: extracting multi-scale features from the input data through the feature layer to learn local and global patterns in the sound pressure level data; the feature layer learns spatial features and time series patterns in the sound pressure level data through multi-layer convolution kernels;

[0063] S3, inverse scattering modeling: inversely inferring the spatial distribution of the discharge source from the detected acoustic wave field through the inverse scattering coding layer; the inverse scattering coding layer infers the spatial distribution information of the discharge source from the sound pressure level characteristics through deep learning model learning;

[0064] S4, spatiotemporal feature integration: For time-varying discharge signals, the cycle layer integrates time series information to enhance the model's understanding of the dynamic process of discharge; the cycle layer captures the changing rules of the discharge process over time through a recurrent neural network to enhance the understanding of dynamic partial discharge behavior;

[0065] S5, image reconstruction: The encoded feature information is converted back into an image form through the decoding and reconstruction layer, and based on the feature map output by the previous layer, the decoding and reconstruction layer reconstructs the acoustic image of the discharge area through an upsampling operation;

[0066] S6, output and optimization: output the reconstructed discharge area image through the output layer to display the discharge source position, shape and intensity distribution, and adjust the network parameters through back propagation to optimize the model performance;

[0067] Among them, S3 guides the inverse scattering model through the inverse scattering loss function to correctly infer the discharge image from the sound pressure level data and minimize the predicted output S e (x u ) and the actual discharge image y u The difference between them; the formula expression of the inverse scattering loss function is:

[0068]

[0069] In formula (6), x u Represents the input sound pressure level data, used for inverse scattering processing; y u Represents the real discharge image data; Se represents the inverse scattering model, defined by parameter e, which is used to reconstruct the discharge image from the sound pressure level data; R(e) represents the regularization term, which is used to prevent overfitting; n represents the regularization coefficient, which is used to balance the inverse scattering reconstruction error and model complexity;

[0070] In the spatiotemporal feature integration step, the input information at time t and the state at the previous moment are integrated through the cell state update function to achieve efficient capture of the temporal features; the formula expression of the cell state update function is:

[0071]

[0072] In formula (7), E t Represents the cell state at the current time step t, which is used to store long-term dependency information; E t-1 Represents the time state of the previous time step; r t-1 is the hidden state of the previous time step, used to carry historical information; x t Represents the input data of the current time step; tanh() represents the hyperbolic tangent activation function, which is used to limit the output range and introduce nonlinearity; f t represents the forget gate; i t represents the input gate; W d represents the weight matrix, which is used to map the input and the previous hidden state to the new candidate state; represents the element-wise multiplication operation, which is used for the gating mechanism; B c Represents the bias term, which is used to adjust the candidate state;

[0073] In the image reconstruction step, the feature map n after encoder compression and potential feature processing is converted into tg The reconstructed image G that is upsampled and converted back to the original image size guides the model to optimize the quality of image reconstruction; the formula expression of the deconvolution layer output function is:

[0074]

[0075] In formula (8), G is the reconstructed image output; n fjg represents the deconvolution kernel weight, which is used to upsample and reconstruct the image from the feature map; β represents the activation function, which is used to introduce nonlinearity and limit the output range; n tg represents the feature map after being processed by the encoder and latent features; B c Represents the bias term, which is used to adjust the output.

[0076] As a further technical solution of the present invention, the health prediction and analysis module first conducts in-depth mining of historical monitoring data through big data analysis, and integrates data streams from different sensors and monitoring points through data fusion methods to construct an equipment status data set; then, based on the equipment status data set, the health prediction and analysis module identifies abnormal patterns in the data through clustering analysis and anomaly detection algorithms, and analyzes the historical performance and real-time data of the equipment through a machine learning prediction model to identify health degradation patterns and fault precursors; wherein, the data fusion method extracts features from multi-source data through feature engineering and principal component analysis, and captures the trends and periodic changes of equipment status over time through time series analysis; the feature engineering transforms and enhances the original data through feature selection and feature construction to generate a model learning feature set; the machine learning prediction model processes time series data through a long short-term memory network to capture long-term dependencies, and constructs an integration of multiple weak prediction models through an ensemble learning method, and in addition combines online learning mechanisms and incremental learning to enable the model to adapt to real-time changes in equipment status, continuously update the knowledge base, and improve the timeliness of predictions.

[0077] As a further technical solution of the present invention, the cloud management module includes a user interface submodule, a data management submodule, an intelligent alarm submodule and a remote control submodule; the user interface submodule realizes data display and user operation and control of the platform through interactive visualization methods and responsive layout; the data management submodule realizes storage and management of large-scale data through a distributed database system; the distributed database system realizes data storage through data sharding and load balancing; based on the output of the partial discharge feature deep mining module, the three-dimensional precise positioning module and the health prediction analysis module; the intelligent alarm submodule realizes timely transmission of alarm information through a rule engine, message push and email / SMS notification service; the message push performs real-time information push through long polling; the remote control submodule includes a control instruction unit and a device management unit; the control instruction unit realizes remote operation and control of the high-voltage control cabinet equipment through a remote procedure call protocol, and the device management unit realizes remote management and maintenance of the equipment through equipment status monitoring and configuration management; the equipment status monitoring realizes real-time monitoring of the equipment status through heartbeat detection and log analysis; the configuration management realizes remote update and management of the equipment configuration through version control and configuration update mechanism.

[0078] Positive beneficial effects:

[0079] The present invention uses an adaptive digital filtering algorithm to remove environmental and system noise through an intelligent noise reduction module, combines a deep convolutional neural network to learn historical noise patterns, builds a noise prediction model, dynamically identifies and eliminates noise interference, significantly improves the purity of partial discharge signals, and ensures the accuracy of signal analysis. By adopting deep feature joint learning, integrating multimodal data, and using deformable convolutional networks and attention mechanisms to enhance feature extraction across modalities, while adaptive feature enhancement networks dynamically enhance key partial discharge features, effectively distinguish different types of discharges, and improve the accuracy of recognition and the reliability of intensity assessment. Combining time difference arrival positioning TDOA, signal strength positioning RSSI and acoustic imaging technology, multi-dimensional cross-validation, TDOA uses time difference to accurately calculate spatial distance, RSSI assists positioning through signal strength differences, and acoustic imaging further refines the positioning results. The three work together to achieve high-precision three-dimensional positioning of partial discharge sources, significantly improving the accuracy and stability of positioning. Using big data analysis and machine learning prediction models, in-depth mining of historical monitoring data can not only evaluate the current health status of the equipment, but also predict the future failure trend of the equipment prospectively, providing a scientific basis for preventive maintenance and effectively reducing the operation risk caused by failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work, among which:

[0081] Figure 1 A framework diagram of a high-voltage control cabinet with a partial discharge detection device according to the present invention;

[0082] Figure 2 It is a working step diagram of the adaptive multi-dimensional spectrum clustering discharge identification algorithm of the present invention;

[0083] Figure 3 A diagram showing the working steps of the intelligent noise reduction module of the present invention;

[0084] Figure 4 This is a working principle architecture diagram of the adaptive multi-dimensional spectrum clustering discharge identification algorithm of the present invention. DETAILED DESCRIPTION

[0085] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0086] like Figure 1-Figure 4 As shown, a high-voltage control cabinet with a partial discharge detection device includes a data acquisition module and a cloud-based intelligent control platform; the data acquisition module is used to capture partial discharge signals and environmental data inside and outside the high-voltage control cabinet, and transmit them to the cloud-based intelligent control platform; the data acquisition module realizes the acquisition and perception of partial discharge signals in the high-voltage control cabinet through a multi-dimensional sensor array and a signal synchronization method; the multi-dimensional sensor array covers different types of discharge signals by deploying ultrasonic sensors, ultra-high frequency sensors, high-frequency current transformers, acoustic emission sensors, gas detection sensors and photoelectric sensors; the signal synchronization method ensures the clock synchronization of all sensor data through the network time protocol NTP;

[0087] The cloud-based intelligent control platform is used to receive and process partial discharge signals and environmental data from a multi-dimensional sensor array, and to perform real-time monitoring, intelligent diagnosis, health prediction, precise positioning and remote control of the high-voltage control cabinet; the cloud-based intelligent control platform includes an intelligent noise reduction module, a partial discharge feature deep mining module, a three-dimensional precise positioning module, a health prediction and analysis module and a cloud management module;

[0088] The intelligent noise reduction module is used to enhance the purity of the partial discharge signal and environmental data captured by the data acquisition module; the intelligent noise reduction module removes the noise interference generated by the environment and the system itself through an adaptive digital filtering algorithm and a noise model learning method; the noise model learning method learns historical noise data through a deep convolutional neural network and constructs a noise model to predict and eliminate noise; the denoised data is output to the partial discharge feature deep mining module;

[0089] The partial discharge feature deep mining module is used to realize discharge type recognition and intensity evaluation based on the partial discharge signal and environmental data enhanced by the intelligent noise reduction module, and obtain discharge type recognition and intensity evaluation results; the partial discharge feature deep mining module extracts features from the partial discharge signal through deep feature joint learning and adaptive feature enhancement network; the deep feature joint learning realizes cross-modal integration and enhancement of features through multimodal fusion, deformable convolution network and attention mechanism; the adaptive feature enhancement network realizes dynamic enhancement of partial discharge features through gating mechanism, feature enhancement method and progressive network structure;

[0090] The three-dimensional precise positioning module and the health prediction and analysis module are used to locate the spatial position of the partial discharge source based on the discharge type identification and intensity assessment results, evaluate the health status of the equipment, and predict potential faults; the three-dimensional precise positioning module realizes the three-dimensional spatial positioning of the partial discharge source through time difference arrival positioning TDOA, signal strength analysis and acoustic imaging; the time difference arrival positioning TDOA calculates the distance through the time difference of sensor signals; the signal strength analysis assists positioning through signal strength differences; the acoustic imaging further refines the positioning results through sound wave reflection and interference effects; the health prediction and analysis module conducts in-depth mining of historical monitoring data through big data analysis and machine learning prediction models, evaluates the health status of equipment and predicts future fault trends;

[0091] The cloud management module is used to realize the operation interface display, data storage and management, real-time intelligent alarm notification, and remote operation, monitoring and configuration management of the high-voltage control cabinet monitoring platform.

[0092] In a specific embodiment, the working process of a high-voltage control cabinet with a partial discharge detection device is as follows:

[0093] m1. Data acquisition stage: Using the data acquisition module, the multi-dimensional sensor array synchronously captures the partial discharge signals and environmental data inside and outside the high-voltage control cabinet. These sensors include ultrasonic sensors, ultra-high frequency sensors, high-frequency current transformers, acoustic emission sensors, gas detection sensors and photoelectric sensors to ensure comprehensive coverage of different types of discharge signals.

[0094] m2. Signal synchronization: Through the signal synchronization method, specifically the Network Time Protocol (NTP), the precise synchronization of all sensor data in time is ensured, providing accurate timing information for subsequent data processing.

[0095] m3. Data transmission: The collected data is transmitted to the cloud-based intelligent control platform, which is responsible for data storage, processing and display.

[0096] m4. Intelligent noise reduction processing: In the cloud-based intelligent control platform, the intelligent noise reduction module first processes the data, removes noise interference through an adaptive digital filtering algorithm, and learns historical noise data through a deep convolutional neural network to build a noise model to further predict and eliminate noise.

[0097] m5. Feature extraction and deep mining: The PD feature deep mining module performs in-depth analysis on the purified signal and extracts the features of the discharge signal through deep feature joint learning and adaptive feature enhancement network. Multimodal fusion, deformable convolutional network and attention mechanism realize cross-modal integration and enhancement of features.

[0098] m6. Three-dimensional precise positioning: The three-dimensional precise positioning module uses time difference arrival positioning TDOA, signal strength positioning RSSI and acoustic imaging technology to accurately locate the spatial position of the partial discharge source.

[0099] m7. Discharge type identification and intensity assessment: Utilize the output of the partial discharge feature deep mining module to identify the discharge type and assess the discharge intensity, providing quantitative analysis of the health status of the equipment.

[0100] m8. Health prediction analysis: The health prediction analysis module uses big data analysis and machine learning prediction models to evaluate the current health status of the equipment and predict possible failure trends in the future.

[0101] m9. Intelligent alarm and remote control: The cloud-based intelligent control platform provides intelligent alarm functions based on the results of the health prediction analysis module, and notifies the operation and maintenance personnel in a timely manner when potential failure trends are detected. At the same time, the platform supports remote control functions, allowing operators to remotely process or adjust the status of equipment.

[0102] m10. Cloud management module coordination: The cloud management module is responsible for coordinating the operation of the entire platform, including data management, user access control, system maintenance and updates, etc.

[0103] m11. User interface display: Ultimately, all processed data, analysis results, and prediction information are displayed to users through the user interface, providing an intuitive and easy-to-use interactive environment.

[0104] In the above embodiment, the data acquisition module includes a sensor array integrated unit, a signal synchronization unit and a data transmission unit; the sensor array integrated unit includes an ultrasonic sensor unit, an ultra-high frequency sensor unit, a high-frequency current mutual inductance unit, an acoustic emission sensor unit, a gas detection sensor unit and a photoelectric sensor unit; the ultrasonic sensor unit realizes directional capture and positioning of ultrasonic signals through a beam forming method; the ultra-high frequency sensor unit adopts broadband amplification to enhance the receiving ability of ultra-high frequency electromagnetic waves; the high-frequency current mutual inductance unit measures and converts transient signals in high-voltage current through the principle of electromagnetic induction; the acoustic emission sensor unit captures tiny stress waves caused by partial discharge through acoustic emission detection; the gas detection sensor unit detects the gas composition inside the control cabinet through chemical sensors changes; the photoelectric sensing unit adopts a photoelectric conversion mechanism to detect the changes of optical signals caused by partial discharge; the signal synchronization unit includes a time synchronization subunit and a data synchronization subunit; the time synchronization subunit ensures the consistency of the time tag of the signal through the network time protocol NTP; the network time protocol NTP ensures the time synchronization of the sensor through a time calibration mechanism, a hierarchical time server structure and a timestamp; the data synchronization subunit realizes parallel data acquisition and synchronous storage through a synchronous data acquisition card; the data transmission unit includes a data encoding subunit and a communication interface subunit; the data encoding subunit improves the reliability and efficiency of data transmission through error control coding; the communication interface subunit uploads the monitoring data to the cloud server in real time through the Internet of Things communication protocol MQTT.

[0105] In a specific embodiment, ultrasonic sensors are capable of detecting high-frequency sound waves generated by partial discharges. These sensors are very effective at capturing the tiny sound waves generated by discharges in insulating materials, which can help determine the intensity and possible location of the discharge. Ultra-high frequency (UHF) sensors are used to capture the high-frequency electromagnetic waves generated by partial discharges. Because the frequency of these signals is very high, UHF sensors can largely avoid low-frequency electromagnetic interference, thereby providing more accurate data. Infrared thermal imaging sensors are used to detect temperature anomalies in equipment caused by partial discharges. By monitoring temperature changes, such sensors can provide early warning before the discharge causes significant damage. Acoustic emission sensors are used to detect and record stress waves generated within materials due to partial discharges. These sensors are particularly effective in capturing signs of cracks or other physical damage to the insulation layer. Photosensitive sensors can detect abnormal light inside the control cabinet, such as light caused by arcs or partial discharges. These sensors can help identify discharge events, especially in the visible light range. Partial discharges sometimes decompose the surrounding air or insulation materials, producing specific gases, such as ozone or nitrogen oxides. Gas analysis sensors can detect these chemical changes, providing another means of detecting partial discharges. Monitoring temperature and humidity inside the control cabinet is also important, as these factors can affect insulation performance and the possibility of partial discharges. Humidity and temperature sensors provide environmental data to help more fully assess the insulation condition of the facility. High-frequency current transformers detect high-frequency current pulses caused by partial discharge and are suitable for discharge monitoring at locations such as cable terminals and joints.

[0106] In the above embodiment, the working steps of the intelligent noise reduction module in a high-voltage control cabinet with a partial discharge detection device are as follows:

[0107] Step 1: Data reception and preprocessing;

[0108] The real-time collected data of the data acquisition module is obtained through the data interface, and the collected signal is pre-processed through digital filtering and signal conditioning methods; the digital filtering removes high-frequency noise through a low-pass filtering mechanism; the signal conditioning method realizes signal amplitude adjustment and digitization through an amplifier and an analog-to-digital converter;

[0109] Step 2: Time-frequency analysis and feature extraction;

[0110] The characteristics of the local discharge signal and the background noise are distinguished by wavelet transform and short-time Fourier transform. The wavelet transform can capture the time-varying characteristics of the signal through multi-scale decomposition and sliding window mechanism; the short-time Fourier transform can analyze the frequency components of the signal through local window processing.

[0111] Step 3: Adaptive digital filtering;

[0112] Based on the result of feature extraction, the filter coefficients are dynamically adjusted according to the signal characteristics through the minimum mean square error and recursive least squares algorithm. The minimum mean square error and recursive least squares algorithm suppress non-partial discharge noise in real time through the gradient descent method and the fast convergence algorithm to improve the signal purity;

[0113] Step 4: Noise model learning;

[0114] Automatically identify noise patterns from signals and construct noise models through supervised learning algorithms and feature selection strategies; the supervised learning algorithms optimize model parameters through iterative learning of training set data; the feature selection strategies screen feature signals through statistical testing methods;

[0115] Step 5: Noise prediction and elimination;

[0116] As the noise model is updated in real time, noise is removed from the original signal through signal subspace estimation, and the noise model realizes the prediction output of noise through the inference engine of the machine learning model; the signal subspace estimation realizes the separation of signal and noise through singular value decomposition SVD;

[0117] Step 6: Performance evaluation and feedback;

[0118] The quantitative evaluation of noise reduction effect and the dynamic adjustment of system parameters are realized through performance evaluation indicators and feedback control mechanism; the performance evaluation indicators measure the noise reduction effect through mathematical statistical methods, and the feedback control mechanism realizes adaptive adjustment of parameters through a closed-loop control system; the performance evaluation indicators include signal-to-noise ratio and mean square error;

[0119] Step 7: Model self-learning and updating;

[0120] The continuous learning and knowledge updating of the noise model for new data are achieved through online learning algorithms and model fusion strategies; the model fusion strategy achieves model knowledge fusion and optimization through an integrated learning method.

[0121] In a specific embodiment, the adaptive digital filter algorithm is based on signal processing theory, and dynamically adjusts the filter parameters by designing filters that adapt to different signal characteristics to achieve noise suppression within a specific frequency range. The algorithm analyzes the signal spectrum in real time, automatically identifies and reduces the noise in the non-PD signal frequency band, while retaining useful PD signal components. The adaptive adjustment of the filter parameters depends on the real-time statistical characteristics of the signal, such as power spectrum density estimation, to ensure that a good filtering effect can be maintained under different working conditions. The noise model learning method uses a deep convolutional neural network CNN to learn historical noise data. CNN can automatically extract features from noise data through multiple layers of convolution kernels, and learn the difference between noise and normal signals during the training process. Through a large number of labeled noise sample training, the network can construct a complex noise model that can not only identify the characteristics of noise, but also predict the type and intensity of noise that may be encountered in the future, so as to achieve more accurate noise prediction and elimination. In the specific implementation, the noise interference in the environment and the system is removed by intelligent noise reduction, which significantly improves the signal-to-noise ratio of the PD signal, making the subsequent PD feature analysis more accurate. In addition, the probability of false alarms and missed alarms is reduced, the reliability and accuracy of partial discharge detection is improved, and potential faults in power equipment are discovered in a timely manner. Secondly, accurate partial discharge detection helps to achieve early warning of equipment status, making maintenance work more targeted, avoiding unnecessary maintenance operations, and reducing operating costs.

[0122] In the above embodiment, the deep convolutional neural network includes a signal preprocessing layer, a convolution layer, a feature enhancement and noise reduction layer, a pooling layer, a fully connected layer, a model layer and an optimization layer; the working method of the deep convolutional neural network includes:

[0123] 1) The signal preprocessing layer performs preliminary purification and standardization on the signal; the signal preprocessing layer removes the DC offset and baseline drift in the signal through high-order statistical analysis and spectrum purification algorithm, and uses normalization processing to unify the signal amplitude to the same scale; the high-order statistical analysis identifies and removes non-Gaussian noise through cumulative estimation; the spectrum purification algorithm filters out interference in specific frequency bands through frequency band filtering method; the normalization processing ensures data consistency through maximum and minimum value scaling;

[0124] 2) learning and extracting local features in the signal through the convolution layer; the convolution layer slides on the input signal through the convolution operation, and the weight of each convolution kernel is automatically adjusted through learning to match the specific features, and the shared weight is used to reduce the number of parameters to improve efficiency and generalization ability;

[0125] 3) The feature enhancement and noise reduction layer is used to strengthen the signal features while weakening the noise impact; based on feature extraction, the feature enhancement and noise reduction layer enhances the relevant features of the partial discharge signal through sparse coding and principal component analysis, and reconstructs the signal through the autoencoder to minimize the reconstruction error;

[0126] 4) Reducing the dimension of the feature map through the pooling layer;

[0127] 5) The extracted features are integrated through the fully connected layer to perform high-level abstraction; the fully connected layer learns the complex relationship between features through multi-layer neuron full connection, and enhances the model's expressiveness through the ReLU activation function;

[0128] 6) predicting noise characteristics through the model layer; the model layer simultaneously predicts the noise type, intensity and trend through multi-task learning, and performs probability prediction through Bayesian network;

[0129] 7) According to the comparison between the model prediction results and the actual noise situation, the model parameters are dynamically adjusted through the optimization layer; the optimization layer calculates the loss function gradient through the back propagation algorithm, and updates the network weights through the optimization algorithm SGD.

[0130] In a specific embodiment, a deep convolutional neural network extracts local features by convolving a convolution kernel with an input signal. The weights learned by the convolution kernel reflect the local dependencies in the signal. Nonlinearity is introduced using the ReLU activation function, allowing the network to learn more complex feature representations. The pooling operation uses maximum pooling to reduce the spatial dimension of the features while increasing the invariance to signal displacement. The fully connected layer generates a noise prediction by integrating the features extracted by the convolution layer and the pooling layer. The loss function evaluates the difference between the predicted output and the actual noise. The optimization algorithm is used to update the network weights according to the gradient of the loss function to optimize the network performance.

[0131] In a high-voltage control cabinet with a partial discharge detection device, DCNN can effectively separate noise from complex signals, significantly improve the purity of partial discharge signals, and provide clearer data for subsequent analysis. The noise features extracted through deep learning help to more accurately identify and extract the characteristics of partial discharge signals and improve the accuracy of discharge detection. In addition, the pure signal after noise reduction helps the three-dimensional precise positioning module to more accurately determine the location of the partial discharge source and improve the positioning accuracy. Secondly, DCNN has good generalization ability and can adapt to different noise environments and signal conditions to ensure the consistency and effectiveness of the noise reduction effect. At the same time, the intelligent noise reduction module reduces manual intervention, realizes the automation and intelligence of the signal processing process, and improves work efficiency.

[0132] In the above embodiment, the local discharge feature deep mining module includes a feature integration submodule, a feature enhancement submodule, and a discharge evaluation submodule; the feature integration submodule includes a multimodal fusion unit, a feature mapping unit, and an attention mechanism unit; the multimodal fusion unit realizes the integration of different sensor data through cross-modal association learning; the feature mapping unit realizes the spatial adaptation of input data through the spatial transformation network STN to adapt to features of different shapes; the spatial transformation network STN dynamically adjusts the shape and size of the convolution kernel according to the signal characteristics through an adaptive convolution mechanism; the attention mechanism unit optimizes the feature selection process through an attention weighted strategy; The feature enhancement submodule includes an information screening unit, a weight optimization unit and a progressive network structure unit; the information screening unit dynamically adjusts the information flow of the network through the gated recurrent unit GRU to optimize the feature transmission and enhancement process; the weight optimization unit determines the importance of the feature and weights it through a feature weighting algorithm; the feature weighting algorithm selects features through particle swarm optimization and weights the selected features through the feature importance in the random forest; the progressive network structure unit improves the training effect of the deep network through residual connections; the discharge evaluation submodule performs type recognition and strength evaluation on the extracted features through an adaptive multidimensional spectral clustering discharge recognition algorithm.

[0133] In a specific embodiment, the technical core of the partial discharge feature deep mining module lies in the fusion application of deep learning and signal processing. First, multimodal fusion technology integrates signals from different sensors, such as ultrasound, ultra-high frequency, etc., to obtain a comprehensive partial discharge feature view. Then, the deformable convolutional network dynamically adjusts the convolution kernel according to the specific form of the signal to adapt to features of different shapes, enhancing the flexibility and accuracy of feature extraction. Further, the attention mechanism is introduced to strengthen the network's focus on key discharge features in the signal, and improve the distinguishability of feature representation by weighting important features. In addition, the adaptive feature enhancement network uses a gating mechanism to control the information flow, and through feature enhancement methods and progressive network structures, it realizes the dynamic enhancement of partial discharge features to ensure the effective transmission of key information in the network. Specifically, the process of implementing the partial discharge feature deep mining module is as follows: First, the partial discharge signal is collected through a multi-dimensional sensor array in the high-voltage control cabinet, and is pre-processed through pre-processing steps such as filtering, denoising and normalization to prepare for input to the deep mining module. Then, the deep feature joint learning technology is used, combined with multimodal fusion, deformable convolutional network and attention mechanism, to extract features from the pre-processed signal. In the adaptive feature enhancement network, features are further refined through gating mechanisms and feature enhancement methods to enhance key discharge features in the signal. Then, the feature integration submodule integrates the extracted and enhanced features to form a comprehensive feature vector in preparation for discharge evaluation. The discharge evaluation submodule uses a classification algorithm to identify the discharge type and uses a regression model to evaluate the discharge intensity. Next, the discharge evaluation results are analyzed in detail, and the model parameters are adjusted according to the actual discharge situation to achieve more accurate discharge feature mining. Finally, the module is continuously optimized through machine learning methods, including hyperparameter tuning and model structure improvement, to adapt to new data and improve performance.

[0134] In specific implementation, the partial discharge feature deep mining module greatly improves the accuracy and reliability of discharge detection. Through in-depth analysis of complex signals, it can effectively distinguish different types of discharge events and accurately evaluate the discharge intensity. Secondly, the module enhances the system's adaptability to new situations through adaptive learning and dynamic feature enhancement, ensuring long-term stable operation. In addition, the implementation of the module also improves the preventive maintenance capability of the equipment. Through continuous monitoring and evaluation of discharge characteristics, potential faults can be predicted, thereby achieving timely maintenance and repair, reducing unexpected downtime, and extending equipment life.

[0135] In the above embodiment, the working method of the adaptive multi-dimensional spectrum clustering discharge identification algorithm includes the following steps:

[0136] Step 1: Multi-dimensional feature spectrum embedding;

[0137] The high-dimensional features are mapped to the low-dimensional space through the Laplace eigenmapping function, while retaining the local structural characteristics of the data. The formula expression of the Laplace eigenmapping function is:

[0138]

[0139] In formula (1), h t is the feature matrix after spectral embedding, and the column vector is the representation after the original high-dimensional feature space is mapped to the low-dimensional feature space; W is the degree matrix, where the elements D on the diagonal of the degree matrix W are ii represents the degree of node i, that is, the number of edges connected to node i; L represents the Laplacian matrix; W -1 / 2 is the semi-inverse matrix of the degree matrix, which is used to balance the degrees of different nodes; L ij represents the connection strength between nodes i and j;

[0140] Step 2: Adaptive weighted spectral clustering;

[0141] According to the spectral embedding features obtained in step 1, an adaptive weighted strategy is used for spectral clustering. The adaptive weighted strategy enhances the clustering effect by dynamically adjusting the weights of categories. The adaptive weighted spectral clustering formula is:

[0142]

[0143] In formula (2), X JL is the clustering weight matrix, which is used to quantify the similarity or distance between samples; B is the basis matrix of the embedding space, where the column vector of the basis matrix B is the eigenvector of the Laplace matrix L, corresponding to the non-zero eigenvalue; T is the eigenvalue diagonal matrix, and the elements on the diagonal of the eigenvalue diagonal matrix T are γ i is the eigenvalue of the Laplace matrix L;

[0144] Step 3: Discharge type identification based on spectrum;

[0145] Using the clustering results obtained in step 2, the discharge type is identified by a spectrum-based method; the spectrum-based method achieves classification by the distance between the spectral features and the known types; the spectral distance formula for discharge type identification is:

[0146]

[0147] In formula (3), x sp is the spectral distance, which is used to measure the distance between the sample and the center of the known discharge type in the spectral embedding space; γ j is the spectral feature value of the jth feature of the sample; C i is the center of the i-th known discharge type; h t represents the feature matrix after spectrum embedding; uij 2 Represents the average of the spectral feature values ​​of the jth feature of the i-th class of known samples;

[0148] Step 4: Multi-scale evaluation of discharge intensity;

[0149] On the basis of the discharge type identification, the discharge intensity is quantitatively analyzed by a multi-scale evaluation method; the multi-scale evaluation method evaluates the severity of the discharge by combining the spectral features of different time scales; the discharge intensity evaluation formula is:

[0150]

[0151] In formula (4), P g is the discharge intensity index; P ref is the reference discharge intensity, used to standardize the discharge intensity assessment benchmark value; K is the number of scales; Q represents the time point considered at each scale; e kt is the spectral eigenvalue of the kth scale at time t, which is used to measure the characteristics of the discharge at time t; w kt is the weighting coefficient at scale k and time t, which is used to adjust the contribution of different scales and time points to the discharge intensity assessment;

[0152] Step 5: Adaptive iterative optimization;

[0153] Finally, the spectral embedding and clustering parameters are continuously adjusted through the adaptive iterative optimization process to adapt to new data changes and improve recognition accuracy; the iterative optimization formula is:

[0154]

[0155] In formula (5), f new are the updated model parameters; f are the model parameters, including the learnable parameters in the process of spectral clustering, discharge type identification and intensity evaluation; m is the learning rate, which is used to control the step size of parameter update; p is the state-action distribution under the current strategy; y is the strategy network, which is used to measure the probability of taking an action in a given state; f old Current model parameters; Based on the current parameter f old The gradient of the error function E is used to indicate the direction and magnitude of parameter update.

[0156] In a specific embodiment, the core of the adaptive multi-dimensional spectral clustering discharge identification algorithm is to use multi-dimensional feature space and spectral clustering theory to analyze complex and nonlinear partial discharge signal patterns. The working method of the algorithm is: first, a series of key features are extracted from the collected partial discharge signal, such as time domain statistics (such as mean, variance), frequency domain features (such as spectrum energy distribution), waveform shape parameters, etc., to form a high-dimensional feature vector. These features can fully reflect the dynamic characteristics and physical nature of the partial discharge signal. Then, the algorithm automatically adjusts the preprocessing strategy (such as filtering, noise reduction) according to the signal quality to ensure the accuracy of subsequent analysis. This step is the embodiment of "adaptive" and can optimize the signal processing process for different environments and equipment states. Subsequently, the high-dimensional feature space is segmented using the spectral clustering method. Spectral clustering is different from traditional distance-based clustering such as K-means. It is based on graph theory and Laplace matrix and can more effectively handle the clustering problem of non-convex and high-dimensional data sets. By calculating the graph Laplace eigenvalues ​​corresponding to the eigenvectors, the algorithm can automatically discover the natural structure in the data and classify similar partial discharge patterns into one category.

[0157] In practical applications, the partial discharge signals in the high-voltage control cabinet are first continuously monitored and collected at high speed. After preliminary filtering, the collected data is input into the adaptive multi-dimensional spectral clustering recognition system. The system automatically performs feature extraction, adaptive preprocessing, and clusters the feature vectors through the spectral clustering algorithm. The algorithm output includes but is not limited to the identification label of the partial discharge type (such as surface discharge, internal discharge, etc.) and the intensity evaluation of each type of discharge. Operators or intelligent monitoring systems can quickly locate the source of the fault based on this information, evaluate the health status of the equipment, and take maintenance measures in time. The algorithm significantly improves the automation level and diagnostic accuracy of partial discharge detection. Through the adaptive mechanism, the algorithm can adapt to signal changes under different working conditions and reduce false alarms and missed alarms. The application of the multi-dimensional spectral clustering method overcomes the limitations of traditional methods in dealing with complex partial discharge patterns and improves the robustness and accuracy of classification. This not only helps to detect potential power equipment failures in advance and extend the service life of equipment, but also reduces the economic losses caused by unplanned downtime and ensures the safe and stable operation of the power grid.

[0158] Specifically, Experiment 1 is designed to verify the superiority of the adaptive multi-dimensional spectral clustering discharge recognition algorithm over the traditional K-means algorithm in the detection of partial discharge in high-voltage control cabinets. By increasing the data complexity and professional parameters, the performance of the algorithm in practical applications is more comprehensively evaluated, including recognition accuracy, robustness, processing speed, and adaptability to different discharge types.

[0159] The data source is the continuous monitoring of partial discharge signals at different locations in the high-voltage control cabinet within a week, covering various scenarios from normal operation, mild discharge to severe discharge. The extracted features include time domain features (root mean square value (RMS), peak voltage, peak-to-peak value, rise time and pulse repetition rate), frequency domain features (spectral energy distribution, center frequency and specific frequency components), phase features (phase angle and phase correlation), and waveform shape indicators (pulse width and shape factor); among them, the adaptive multidimensional spectral clustering algorithm adopts a dynamically adjusted preprocessing strategy, combined with spectral clustering of multidimensional feature space, to achieve fine identification and intensity quantification of partial discharge types. The traditional K-means algorithm is used as a benchmark, with fixed parameter settings, and clustering is performed only based on a simplified feature set (such as several key parameters in the time domain and frequency domain). In the partial discharge detection of high-voltage control cabinets, in order to conduct a comprehensive and detailed analysis, professional parameters can be extracted from the following dimensions to construct a more complex and accurate feature space:

[0160] Performance evaluation indicators include: Recognition accuracy: the proportion of correctly identified discharge events to total events. False alarm rate: the proportion of normal events that are mistakenly identified as discharges. Missed alarm rate: the proportion of discharge events that are not correctly identified. Processing time: the time required from data input to result output. Robustness evaluation: the consistency of performance under different noise levels and signal strengths. The experimental results comparison table is shown in Table 1:

[0161] Table 1 Experimental data comparison table

[0162]

[0163] As shown in Table 1, the adaptive algorithm showed higher recognition accuracy and strength assessment accuracy under all test conditions, especially in complex discharge types (such as suspended discharge) and weak signal detection. Secondly, the adaptive algorithm is more stable in different noise environments, reducing the impact of external factors on the recognition results, and showing good robustness. Although the adaptive algorithm takes a little longer to process, its significant advantages in improving diagnostic accuracy and reducing false alarm rate have improved the overall detection efficiency and maintenance decision quality.

[0164] In the above embodiment, the three-dimensional precise positioning module includes a signal acquisition submodule, a data processing submodule and an advanced analysis submodule; the signal acquisition submodule includes a sensor network unit and a time synchronization unit; the sensor network unit uses distributed ultra-wideband electromagnetic wave sensors and high-frequency response ultrasonic sensors to capture transient signals generated by local discharge; the time synchronization unit uses an atomic clock as a master clock source to synchronize the time tags of data collected by the sensor network unit through the time protocol PTP IEEE1588; the data processing submodule includes a TDOA calculation unit and a signal strength analysis unit; based on the timestamp information of the received signal, the TDOA calculation unit uses Kalman filtering to infer the distance from the discharge source to each sensor from the time difference, thereby constructing a preliminary framework of the spatial position; the signal strength analysis unit extracts direction information from the signal strength attenuation through an attenuation model and a pattern recognition algorithm; the advanced analysis submodule performs acoustic imaging processing by integrating an acoustic wave sensor array with an image reconstruction engine; the acoustic wave sensor array captures the interference and diffraction effects of the discharge sound waves, and the sound pressure level data of each point is collected and sent to the image reconstruction engine; the image reconstruction engine generates an acoustic image of the discharge area through an inverse scattering algorithm based on deep learning.

[0165] In a specific embodiment, the three-dimensional precise positioning module integrates the three technologies of time difference arrival positioning, signal strength analysis and acoustic imaging, and realizes the precise determination of the spatial position of the partial discharge source through a highly coordinated working mechanism. Time difference arrival positioning (TDOA) uses multiple sensors arranged at different positions to capture the partial discharge signal at the same time, calculates the distance difference between the discharge source and the sensor according to the time difference of the signal reaching each sensor, and determines the spatial coordinates of the discharge source by multilateral measurement in combination with geometric principles. This method relies on the known value of the signal propagation speed in the medium and precise time synchronization technology. The signal strength decays with increasing distance, and the signal strength analysis (SAA) can assist in estimating the relative distance of the partial discharge source by comparing the signal strength received by different sensors. Although a single signal strength is not enough for precise positioning, it can be combined with other positioning methods to provide additional positioning clues, especially to enhance the robustness of positioning in complex environments. Acoustic imaging (AI) uses ultrasonic sensors to capture the sound wave signals caused by partial discharge, and generates sound field images by analyzing the reflection and interference effects of the sound waves to further refine the location of the partial discharge source. Acoustic imaging can provide detailed layout information of the discharge area, and is particularly suitable for characterizing the local characteristics and directionality of the discharge source.

[0166] In a specific embodiment, the three-dimensional precise positioning module is first deployed in multiple sensor networks around the high-voltage control cabinet through the signal acquisition submodule. These sensors include electrical measurement sensors (for TDOA and signal strength analysis) and ultrasonic sensors (for acoustic imaging). When a local discharge event occurs, the sensor network instantly captures the relevant signals and transmits the data to the data processing submodule. After receiving the raw data, the data processing submodule first uses high-precision clock synchronization data to calculate the time difference of signal arrival between different sensors, and applies the TDOA algorithm to preliminarily estimate the three-dimensional coordinates of the local discharge source. Subsequently, combined with the signal strength analysis results, the preliminary positioning results are corrected and optimized. Finally, the acoustic imaging technology analyzes the propagation mode of the sound wave and makes subtle adjustments to the position of the local discharge source to achieve refined spatial positioning. The three-dimensional precise positioning module significantly improves the accuracy of local discharge source positioning by comprehensively using a variety of positioning technologies, which helps to quickly lock the fault point and reduce troubleshooting time and maintenance costs. In addition, the precise positioning capability enables potential faults to be discovered in the early stages of partial discharge development, making it possible to implement predictive maintenance and avoid sudden equipment failures caused by partial discharge. Secondly, the integrated positioning system is easy to integrate with the intelligent monitoring platform to achieve automation and remote management of partial discharge detection and management, thereby improving the level of intelligent management of power facilities.

[0167] In the above embodiment, the inverse scattering algorithm based on deep learning includes an input layer, a data layer, a feature layer, an inverse scattering coding layer, a circulation layer, a decoding and reconstruction layer, and an output layer; the working steps of the inverse scattering algorithm based on deep learning are:

[0168] S1. Data preparation: receiving the original sound pressure level data collected by the acoustic wave sensor array through the input layer, and performing denoising, normalization and spatiotemporal alignment on the input data through the data layer to eliminate the influence of environmental noise and sensor differences;

[0169] S2, feature learning: extracting multi-scale features from the input data through the feature layer to learn local and global patterns in the sound pressure level data; the feature layer learns spatial features and time series patterns in the sound pressure level data through multi-layer convolution kernels;

[0170] S3, inverse scattering modeling: inversely inferring the spatial distribution of the discharge source from the detected acoustic wave field through the inverse scattering coding layer; the inverse scattering coding layer infers the spatial distribution information of the discharge source from the sound pressure level characteristics through deep learning model learning;

[0171] S4, spatiotemporal feature integration: For time-varying discharge signals, the cycle layer integrates time series information to enhance the model's understanding of the dynamic process of discharge; the cycle layer captures the changing rules of the discharge process over time through a recurrent neural network to enhance the understanding of dynamic partial discharge behavior;

[0172] S5, image reconstruction: The encoded feature information is converted back into an image form through the decoding and reconstruction layer, and based on the feature map output by the previous layer, the decoding and reconstruction layer reconstructs the acoustic image of the discharge area through an upsampling operation;

[0173] S6, output and optimization: output the reconstructed discharge area image through the output layer to display the discharge source position, shape and intensity distribution, and adjust the network parameters through back propagation to optimize the model performance;

[0174] Among them, S3 guides the inverse scattering model through the inverse scattering loss function to correctly infer the discharge image from the sound pressure level data and minimize the predicted output S e (x u ) and the actual discharge image y u The difference between them; the formula expression of the inverse scattering loss function is:

[0175]

[0176] In formula (6), x u Represents the input sound pressure level data, used for inverse scattering processing; y u Represents the real discharge image data; S e represents the inverse scattering model, defined by parameter e, which is used to reconstruct the discharge image from the sound pressure level data; R(e) represents the regularization term, which is used to prevent overfitting; n represents the regularization coefficient, which is used to balance the inverse scattering reconstruction error and model complexity;

[0177] In the spatiotemporal feature integration step, the input information at time t and the state at the previous moment are integrated through the cell state update function to achieve efficient capture of the temporal features; the formula expression of the cell state update function is:

[0178]

[0179] In formula (7), E t Represents the cell state at the current time step t, which is used to store long-term dependency information; E t-1 Represents the time state of the previous time step; r t-1 is the hidden state of the previous time step, used to carry historical information; x t Represents the input data of the current time step; tanh() represents the hyperbolic tangent activation function, which is used to limit the output range and introduce nonlinearity; f t represents the forget gate; i t represents the input gate; W d represents the weight matrix, which is used to map the input and the previous hidden state to the new candidate state; represents the element-wise multiplication operation, which is used for the gating mechanism; B cRepresents the bias term, which is used to adjust the candidate state;

[0180] In the image reconstruction step, the feature map n after encoder compression and potential feature processing is converted into tg The reconstructed image G that is upsampled and converted back to the original image size guides the model to optimize the quality of image reconstruction; the formula expression of the deconvolution layer output function is:

[0181]

[0182] In formula (8), G is the reconstructed image output; n fjg represents the deconvolution kernel weight, which is used to upsample and reconstruct the image from the feature map; β represents the activation function, which is used to introduce nonlinearity and limit the output range; n tg represents the feature map after being processed by the encoder and latent features; B c Represents the bias term, which is used to adjust the output.

[0183] In a specific embodiment, the inverse scattering algorithm based on deep learning is an innovative image reconstruction technology that uses deep neural networks (DNN) to analyze acoustic wave signals and reconstruct the acoustic image of the discharge area. The essence of the inverse scattering problem is to infer the spatial distribution of the emission source from the received signal, which is usually an ill-conditioned inverse problem. Traditional methods are difficult to deal with signal aliasing and noise interference in complex environments. The introduction of deep learning technology, especially convolutional neural network and recurrent neural network structures, provides powerful nonlinear mapping capabilities for dealing with such complex inverse problems. The core of the technology is to train a model so that it can learn the mapping relationship from acoustic wave signals to acoustic images. The model input is the acoustic wave data collected by the sensor array, and the output is a two-dimensional or three-dimensional image of the discharge area. The network weights are optimized by back propagation so that the model gradually approaches the true inverse scattering solution. The specific implementation process is as follows:

[0184] Data preparation: Collect or simulate a large number of acoustic data sets, covering partial discharge signals at different locations, intensities, and environmental conditions, and annotate the corresponding discharge images as training targets.

[0185] Network construction: Design a deep learning architecture containing multiple convolutional layers, pooling layers, fully connected layers, and possible recurrent layers to adapt to the time series characteristics and spatial structure information of sound wave data.

[0186] Training model: Use the labeled dataset to train the network. During this period, the difference between the predicted image and the real image is evaluated through the loss function, and the network parameters are continuously adjusted to minimize the loss.

[0187] Verification and optimization: Test the model performance on an independent validation set, and adjust the network structure or training strategy based on indicators such as accuracy and recall until the predetermined performance standards are met.

[0188] Deployment application: The trained model is embedded into the partial discharge detection device of the high-voltage control cabinet, which receives sensor data in real time and generates acoustic images of the discharge area through algorithms to assist in diagnosis.

[0189] In specific implementation, the deep learning algorithm can extract features from complex acoustic signals, effectively separate noise, and achieve high-precision image reconstruction even in noisy environments and complex structures, thereby improving the accuracy of localizing partial discharge sources. Specifically, Experiment 2 is designed to evaluate the performance of the deep learning inverse scattering algorithm in partial discharge detection of high-voltage control cabinets. First, a partial discharge acoustic wave signal data set containing multiple samples and multiple scenes is constructed, including but not limited to single-point discharge, multi-point discharge, signals of different frequencies, different intensities, and different noise levels, as well as signals in high-voltage equipment of different materials and structures. The data set must contain real image labels as the basis for supervised learning. Next, an advanced deep learning framework (such as TensorFlow or PyTorch) is used to design a composite network model, combining CNN for feature extraction, RNN or Transformer to process time series data, adding an attention mechanism to enhance the ability to capture key information and optimize the inverse scattering process. Then, positioning error, image reconstruction similarity (SSIM), running time, model size, etc. are selected as the main evaluation indicators to ensure that the accuracy, efficiency, and resource usage of the algorithm are comprehensively considered. The experimental data record table is shown in Table 2:

[0190] Table 2 Performance evaluation of partial discharge image reconstruction based on deep learning inverse scattering algorithm

[0191]

[0192] In Data Table 2, the discharge type indicates the category of the discharge event, including single point, multi-point (such as double point), weak signal discharge, etc. Signal strength (dBm) indicates the power level of the partial discharge signal, expressed in decibel milliwatts. Noise level (dB) indicates the background noise intensity, which is used to evaluate the performance of the algorithm in a noisy environment. The number of algorithm iterations indicates the number of iterations of deep learning model training, which affects the convergence and performance of the model. Positioning error (mm) indicates the deviation between the position of the partial discharge source in the reconstructed image and the actual position, which measures the positioning accuracy. Image reconstruction SSIM represents the structural similarity index, which measures the similarity between the reconstructed image and the real image, ranging from 0 to 1. The closer the value is to 1, the higher the image quality.

[0193] In the above embodiment, the health prediction analysis module first conducts in-depth mining of historical monitoring data through big data analysis, and integrates data streams from different sensors and monitoring points through data fusion methods to construct an equipment status data set; then, based on the equipment status data set, the health prediction analysis module identifies abnormal patterns in the data through clustering analysis and anomaly detection algorithms, and analyzes the historical performance and real-time data of the equipment through a machine learning prediction model to identify health degradation patterns and precursors to failures; wherein, the data fusion method extracts features from multi-source data through feature engineering and principal component analysis, and captures the trends and periodic changes of equipment status over time through time series analysis; the feature engineering transforms and enhances the original data through feature selection and feature construction to generate a model learning feature set; the machine learning prediction model processes time series data through a long short-term memory network to capture long-term dependencies, and constructs an integration of multiple weak prediction models through an integrated learning method, and in addition combines online learning mechanisms and incremental learning to enable the model to adapt to real-time changes in equipment status, continuously update the knowledge base, and improve the timeliness of predictions.

[0194] In a specific embodiment, the health prediction analysis module uses labeled historical data (including normal equipment operation data and known fault data) to train the model. The model learns how to predict possible problems with the equipment in the future based on current and past monitoring data patterns. The newly collected data is input into the trained model, and the model outputs the health score or failure probability of the equipment. At the same time, the predictive performance of the model is evaluated by cross-validation, AUC-ROC curve and other methods to ensure the reliability of the prediction results. In the specific implementation, the health prediction analysis module is integrated with the existing monitoring system to realize real-time transmission and processing of data. When the model predicts that the health status of the equipment is declining or a failure is about to occur, the early warning system is automatically triggered to notify the maintenance team. By predicting failures in advance, we switch from passive response to active prevention, reduce unexpected downtime, extend equipment life, and reduce maintenance costs.

[0195] In the above embodiment, the cloud management module includes a user interface submodule, a data management submodule, an intelligent alarm submodule and a remote control submodule; the user interface submodule realizes data display and user operation and control of the platform through interactive visualization methods and responsive layout; the data management submodule realizes storage and management of large-scale data through a distributed database system; the distributed database system realizes data storage through data sharding and load balancing; based on the output of the partial discharge feature deep mining module, the three-dimensional precise positioning module and the health prediction analysis module; the intelligent alarm submodule realizes timely transmission of alarm information through a rule engine, message push and email / SMS notification service; the message push performs real-time information push through long polling; the remote control submodule includes a control instruction unit and a device management unit; the control instruction unit realizes remote operation and control of the high-voltage control cabinet equipment through the remote procedure call protocol, and the device management unit realizes remote management and maintenance of the equipment through equipment status monitoring and configuration management; the equipment status monitoring realizes real-time monitoring of the equipment status through heartbeat detection and log analysis; the configuration management realizes remote update and management of the equipment configuration through version control and configuration update mechanism.

[0196] In a specific embodiment, the cloud management module adopts a cloud service model (IaaS, PaaS, SaaS) to provide computing resources, platform services and software applications. In addition, the cloud management module applies data mining and machine learning techniques to analyze and process a large amount of collected partial discharge data. In the specific implementation, the operation and maintenance efficiency of high-voltage equipment is improved through centralized management and monitoring. Secondly, systematic data management improves the availability and analysis capabilities of data. At the same time, real-time monitoring and intelligent alarm mechanisms can respond to potential faults and anomalies in a timely manner, reducing the risk of accidents.

[0197] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. A high voltage control cabinet with a partial discharge detection device, characterized in that: It includes a data acquisition module and a cloud-based intelligent control platform; the data acquisition module is used to capture the local discharge signals and environmental data inside and outside the high-voltage control cabinet, and transmit them to the cloud-based intelligent control platform; the data acquisition module realizes the acquisition and perception of the local discharge signals in the high-voltage control cabinet through a multi-dimensional sensor array and a signal synchronization method; the multi-dimensional sensor array covers different types of discharge signals by deploying ultrasonic sensors, ultra-high frequency sensors, high-frequency current transformers, acoustic emission sensors, gas detection sensors and photoelectric sensors; the signal synchronization method ensures the clock synchronization of all sensor data through the network time protocol NTP; The cloud-based intelligent control platform is used to receive and process partial discharge signals and environmental data from a multi-dimensional sensor array, and to perform real-time monitoring, intelligent diagnosis, health prediction, precise positioning and remote control of the high-voltage control cabinet; the cloud-based intelligent control platform includes an intelligent noise reduction module, a partial discharge feature deep mining module, a three-dimensional precise positioning module, a health prediction and analysis module and a cloud management module; The intelligent noise reduction module is used to enhance the purity of the partial discharge signal and environmental data captured by the data acquisition module; the intelligent noise reduction module removes the noise interference generated by the environment and the system itself through an adaptive digital filtering algorithm and a noise model learning method; the noise model learning method learns historical noise data through a deep convolutional neural network and constructs a noise model to predict and eliminate noise; the denoised data is output to the partial discharge feature deep mining module; The partial discharge feature deep mining module is used to realize discharge type identification and intensity evaluation based on the partial discharge signal and environmental data after the enhancement processing of the intelligent noise reduction module, and obtain the discharge type identification and intensity evaluation results; the partial discharge feature deep mining module extracts features from the partial discharge signal through deep feature joint learning and adaptive feature enhancement network; the deep feature joint learning realizes cross-modal integration and enhancement of features through multimodal fusion, deformable convolution network and attention mechanism; the adaptive feature enhancement network realizes dynamic enhancement of partial discharge features through gating mechanism, feature enhancement method and progressive network structure; the partial discharge feature deep mining module includes feature integration submodule, feature enhancement submodule and discharge evaluation submodule; the feature integration submodule includes multimodal fusion unit, feature mapping unit and attention mechanism unit; the multimodal fusion unit realizes the integration of different sensor data through cross-modal association learning; the feature mapping unit realizes spatial adaptation of input data through spatial transformation network STN to adapt to features of different shapes; the spatial transformation network STN dynamically adjusts the shape and size of the convolution kernel according to signal characteristics through adaptive convolution mechanism; the attention mechanism unit optimizes the feature selection process through attention weighting strategy; The feature enhancement submodule includes an information screening unit, a weight optimization unit and a progressive network structure unit; the information screening unit dynamically adjusts the information flow of the network through the gated recurrent unit GRU to optimize the transmission and enhancement process of the feature; the weight optimization unit determines the importance of the feature and weights it through the feature weighting algorithm; the feature weighting algorithm selects features through particle swarm optimization and weights the selected features through the feature importance in the random forest; the progressive network structure unit improves the training effect of the deep network through residual connection; the discharge evaluation submodule performs type recognition and strength evaluation on the extracted features through the adaptive multi-dimensional spectrum clustering discharge recognition algorithm; The three-dimensional precise positioning module and the health prediction and analysis module are used to locate the spatial position of the partial discharge source based on the discharge type identification and intensity assessment results, and evaluate the health status of the equipment and predict potential faults; the three-dimensional precise positioning module realizes the three-dimensional spatial positioning of the partial discharge source through time difference arrival positioning TDOA, signal strength analysis and acoustic imaging; the time difference arrival positioning TDOA calculates the distance through the time difference of sensor signals; the signal strength analysis assists positioning through signal strength difference; the acoustic imaging further refines the positioning results through sound wave reflection and interference effects; the health prediction and analysis module deeply mines historical monitoring data through big data analysis and machine learning prediction models, evaluates the health status of the equipment and predicts future fault trends; the three-dimensional precise positioning module includes a signal acquisition submodule, a data processing submodule and an advanced analysis submodule; the signal acquisition submodule includes a sensor network unit and a time synchronization unit; the sensor network unit uses a distributed ultra-wideband electromagnetic wave sensor and a high-frequency response ultrasonic sensor to capture the transient signal generated by partial discharge; the time synchronization unit uses an atomic clock as the master clock source and uses the time protocol PTP The IEEE1588 synchronous sensor network unit collects the time tag of the data; the data processing submodule includes a TDOA calculation unit and a signal strength analysis unit; based on the timestamp information of the received signal, the TDOA calculation unit uses Kalman filtering to infer the distance from the discharge source to each sensor from the time difference, thereby building a preliminary framework of the spatial position; the signal strength analysis unit extracts the direction information from the signal strength attenuation through the attenuation model and pattern recognition algorithm; the advanced analysis submodule performs acoustic imaging processing by integrating the acoustic wave sensor array and the image reconstruction engine; the acoustic wave sensor array captures the interference and diffraction effects of the discharge sound waves, and the sound pressure level data of each point is collected and sent to the image reconstruction engine; the image reconstruction engine generates an acoustic image of the discharge area through an inverse scattering algorithm based on deep learning; the inverse scattering algorithm based on deep learning includes an input layer, a data layer, a feature layer, an inverse scattering coding layer, a circulation layer, a decoding and reconstruction layer, and an output layer; the working steps of the inverse scattering algorithm based on deep learning are: S1. Data preparation: receiving the original sound pressure level data collected by the acoustic wave sensor array through the input layer, and performing denoising, normalization and spatiotemporal alignment on the input data through the data layer to eliminate the influence of environmental noise and sensor differences; S2, feature learning: extracting multi-scale features from the input data through the feature layer to learn local and global patterns in the sound pressure level data; the feature layer learns spatial features and time series patterns in the sound pressure level data through multi-layer convolution kernels; S3, inverse scattering modeling: inversely inferring the spatial distribution of the discharge source from the detected acoustic wave field through the inverse scattering coding layer; the inverse scattering coding layer infers the spatial distribution information of the discharge source from the sound pressure level characteristics through deep learning model learning; S4, spatiotemporal feature integration: For time-varying discharge signals, the time series information is integrated through the recurrent layer to enhance the model's understanding of the discharge dynamic process; The recurrent layer captures the changing pattern of the discharge process over time through a recurrent neural network, thereby enhancing the understanding of dynamic partial discharge behavior; S5, image reconstruction: The encoded feature information is converted back into an image form through the decoding and reconstruction layer, and based on the feature map output by the previous layer, the decoding and reconstruction layer reconstructs the acoustic image of the discharge area through an upsampling operation; S6, output and optimization: output the reconstructed discharge area image through the output layer to display the discharge source position, shape and intensity distribution, and adjust the network parameters through back propagation to optimize the model performance; Among them, S3 guides the inverse scattering model through the inverse scattering loss function to correctly infer the discharge image from the sound pressure level data and minimize the predicted output S e (x u ) and the actual discharge image y u The difference between them; the formula expression of the inverse scattering loss function is: In formula (6), x u Represents the input sound pressure level data, used for inverse scattering processing; y u Represents the real discharge image data; S e represents the inverse scattering model, defined by parameter e, which is used to reconstruct the discharge image from the sound pressure level data; R(e) represents the regularization term, which is used to prevent overfitting; n represents the regularization coefficient, which is used to balance the inverse scattering reconstruction error and model complexity; In the spatiotemporal feature integration step, the input information at time t and the state at the previous moment are integrated through the cell state update function to achieve efficient capture of the temporal features; the formula expression of the cell state update function is: In formula (7), E t Represents the cell state at the current time step t, which is used to store long-term dependency information; E t-1 Represents the time state of the previous time step; r t-1 is the hidden state of the previous time step, used to carry historical information; x t Represents the input data of the current time step; tanh() represents the hyperbolic tangent activation function, which is used to limit the output range and introduce nonlinearity; f t represents the forget gate; i t represents the input gate; W d represents the weight matrix, which is used to map the input and the previous hidden state to the new candidate state; represents the element-wise multiplication operation, which is used for the gating mechanism; B c represents the bias term, which is used to adjust the candidate state; In the image reconstruction step, the feature map n after encoder compression and potential feature processing is converted into tg The reconstructed image G that is upsampled and converted back to the original image size guides the model to optimize the quality of image reconstruction; the formula expression of the deconvolution layer output function is: In formula (8), G is the reconstructed image output; n fjg represents the deconvolution kernel weight, which is used to upsample and reconstruct the image from the feature map; β represents the activation function, which is used to introduce nonlinearity and limit the output range; n tg represents the feature map after being processed by the encoder and latent features; B c Represents the bias term, which is used to adjust the output; The cloud management module is used to realize the operation interface display, data storage and management, real-time intelligent alarm notification, and remote operation, monitoring and configuration management of the high-voltage control cabinet monitoring platform.

2. A high voltage control cabinet with a partial discharge detection device according to claim 1, characterized in that: The data acquisition module includes a sensor array integrated unit, a signal synchronization unit and a data transmission unit; the sensor array integrated unit includes an ultrasonic sensor unit, an ultra-high frequency sensor unit, a high-frequency current mutual inductance unit, an acoustic emission sensor unit, a gas detection sensor unit and a photoelectric sensor unit; the ultrasonic sensor unit realizes directional capture and positioning of ultrasonic signals through a beam forming method; the ultra-high frequency sensor unit adopts broadband amplification to enhance the receiving ability of ultra-high frequency electromagnetic waves; the high-frequency current mutual inductance unit measures and converts transient signals in high-voltage current through the principle of electromagnetic induction; the acoustic emission sensor unit captures tiny stress waves caused by partial discharge through acoustic emission detection; the gas detection sensor unit detects changes in gas composition inside the control cabinet through chemical sensors; The photoelectric sensing unit adopts a photoelectric conversion mechanism to detect the change of the optical signal caused by partial discharge; the signal synchronization unit includes a time synchronization subunit and a data synchronization subunit; the time synchronization subunit ensures the consistency of the time tag of the signal through the network time protocol NTP; the network time protocol NTP ensures the time synchronization of the sensor through a time calibration mechanism, a hierarchical time server structure and a timestamp; the data synchronization subunit realizes parallel data acquisition and synchronous storage through a synchronous data acquisition card; the data transmission unit includes a data encoding subunit and a communication interface subunit; the data encoding subunit improves the reliability and efficiency of data transmission through error control coding; the communication interface subunit uploads the monitoring data to the cloud server in real time through the Internet of Things communication protocol MQTT.

3. A high voltage control cabinet with a partial discharge detection device according to claim 1, characterized in that: The working steps of the intelligent noise reduction module in a high-voltage control cabinet with a partial discharge detection device are as follows: Step 1: Data reception and preprocessing; The real-time collected data of the data acquisition module is obtained through the data interface, and the collected signal is pre-processed through digital filtering and signal conditioning methods; the digital filtering removes high-frequency noise through a low-pass filtering mechanism; the signal conditioning method realizes signal amplitude adjustment and digitization through an amplifier and an analog-to-digital converter; Step 2: Time-frequency analysis and feature extraction; The characteristics of the local discharge signal and the background noise are distinguished by wavelet transform and short-time Fourier transform. The wavelet transform can capture the time-varying characteristics of the signal through multi-scale decomposition and sliding window mechanism; the short-time Fourier transform can analyze the frequency components of the signal through local window processing. Step 3: Adaptive digital filtering; Based on the result of feature extraction, the filter coefficients are dynamically adjusted according to the signal characteristics through the minimum mean square error and recursive least squares algorithm. The minimum mean square error and recursive least squares algorithm suppress non-partial discharge noise in real time through the gradient descent method and the fast convergence algorithm to improve the signal purity; Step 4: Noise model learning; Automatically identify noise patterns from signals and construct noise models through supervised learning algorithms and feature selection strategies; the supervised learning algorithms optimize model parameters through iterative learning of training set data; the feature selection strategies screen feature signals through statistical testing methods; Step 5: Noise prediction and elimination; As the noise model is updated in real time, noise is removed from the original signal through signal subspace estimation, and the noise model realizes the prediction output of noise through the inference engine of the machine learning model; the signal subspace estimation realizes the separation of signal and noise through singular value decomposition SVD; Step 6: Performance evaluation and feedback; The quantitative evaluation of noise reduction effect and the dynamic adjustment of system parameters are realized through performance evaluation indicators and feedback control mechanism; the performance evaluation indicators measure the noise reduction effect through mathematical statistical methods, and the feedback control mechanism realizes adaptive adjustment of parameters through a closed-loop control system; the performance evaluation indicators include signal-to-noise ratio and mean square error; Step 7: Model self-learning and updating; The continuous learning and knowledge updating of the noise model for new data are achieved through online learning algorithms and model fusion strategies; the model fusion strategy achieves model knowledge fusion and optimization through an integrated learning method.

4. A high voltage control cabinet with a partial discharge detection device according to claim 1, characterized in that: The deep convolutional neural network includes a signal preprocessing layer, a convolution layer, a feature enhancement and noise reduction layer, a pooling layer, a fully connected layer, a model layer and an optimization layer; the working method of the deep convolutional neural network includes: 1) The signal preprocessing layer performs preliminary purification and standardization on the signal; the signal preprocessing layer removes the DC offset and baseline drift in the signal through high-order statistical analysis and spectrum purification algorithm, and uses normalization processing to unify the signal amplitude to the same scale; the high-order statistical analysis identifies and removes non-Gaussian noise through cumulative estimation; the spectrum purification algorithm filters out interference in specific frequency bands through frequency band filtering method; the normalization processing ensures data consistency through maximum and minimum value scaling; 2) learning and extracting local features in the signal through the convolutional layer; The convolution layer slides on the input signal through the convolution operation, and the weight of each convolution kernel is automatically adjusted to match the specific features through learning, and the shared weight is used to reduce the number of parameters to improve efficiency and generalization ability; 3) The feature enhancement and noise reduction layer is used to strengthen the signal features while weakening the noise impact; based on feature extraction, the feature enhancement and noise reduction layer enhances the relevant features of the partial discharge signal through sparse coding and principal component analysis, and reconstructs the signal through the autoencoder to minimize the reconstruction error; 4) Reducing the dimension of the feature map through the pooling layer; 5) The extracted features are integrated through the fully connected layer to perform high-level abstraction; the fully connected layer learns the complex relationship between features through multi-layer neuron full connection, and enhances the model's expressiveness through the ReLU activation function; 6) predicting noise characteristics through the model layer; the model layer simultaneously predicts the noise type, intensity and trend through multi-task learning, and performs probability prediction through Bayesian network; 7) According to the comparison between the model prediction results and the actual noise situation, the model parameters are dynamically adjusted through the optimization layer; the optimization layer calculates the loss function gradient through the back propagation algorithm, and updates the network weights through the optimization algorithm SGD.

5. The high voltage control cabinet with a partial discharge detection device according to claim 1, characterized in that: The health prediction and analysis module first conducts in-depth mining of historical monitoring data through big data analysis, and integrates data streams from different sensors and monitoring points through data fusion methods to construct an equipment status data set; then, based on the equipment status data set, the health prediction and analysis module identifies abnormal patterns in the data through clustering analysis and anomaly detection algorithms, and analyzes the historical performance and real-time data of the equipment through a machine learning prediction model to identify health degradation patterns and fault precursors; wherein, the data fusion method extracts features from multi-source data through feature engineering and principal component analysis, and captures the trends and periodic changes of equipment status over time through time series analysis; the feature engineering transforms and enhances the original data through feature selection and feature construction to generate a model learning feature set; the machine learning prediction model processes time series data through a long short-term memory network to capture long-term dependencies, and constructs an integration of multiple weak prediction models through an ensemble learning method, and in addition combines online learning mechanisms and incremental learning to enable the model to adapt to real-time changes in equipment status, continuously update the knowledge base, and improve the timeliness of predictions.

6. A high voltage control cabinet with a partial discharge detection device according to claim 1, characterized in that: The cloud management module includes a user interface submodule, a data management submodule, an intelligent alarm submodule and a remote control submodule; the user interface submodule realizes data display and user operation and control of the platform through interactive visualization methods and responsive layout; the data management submodule realizes the storage and management of large-scale data through a distributed database system; the distributed database system realizes data storage through data sharding and load balancing; based on the output of the partial discharge feature deep mining module, the three-dimensional precise positioning module and the health prediction analysis module; The intelligent alarm submodule realizes timely transmission of alarm information through rule engine, message push and email / SMS notification service; the message push performs real-time information push through long polling; The remote control submodule includes a control instruction unit and a device management unit; the control instruction unit realizes remote operation and control of the high-voltage control cabinet equipment through the remote procedure call protocol, and the device management unit realizes remote management and maintenance of the equipment through equipment status monitoring and configuration management; the equipment status monitoring realizes real-time monitoring of the equipment status through heartbeat detection and log analysis; the configuration management realizes remote update and management of equipment configuration through version control and configuration update mechanism.

Images