Early fault active defense device and method for converter transformer based on artificial intelligence judgment

Abstract

An active defense device and method for early fault of a converter transformer based on artificial intelligence judgment, collect ultra-high frequency / high frequency electromagnetic signals and mechanical vibration sound wave signals inside the transformer; extract time domain features and frequency domain features of each signal; adopt a ResNet-KNN hybrid model with a fusion attention mechanism, take a multi-dimensional feature fusion vector of the ultra-high frequency / high frequency signals as input, and output a partial discharge phase confidence vector; a multi-source fusion judgment module calculates an ultra-high frequency fault degree, a high frequency fault degree and an acoustic fault degree, then dynamically calculates fusion weights of the three fault degrees based on the phase confidence vector, and obtains a comprehensive fault index by weighted summation; an early warning module sends a graded early warning according to a discharge phase classification result and a comprehensive fault judgment result; an execution module executes a tripping active defense action based on the comprehensive fault index and the discharge phase confidence. The present application realizes whole-process management and control of partial discharge of the converter transformer from early warning to active defense.

Description

Technical Field

[0001] This invention relates to the field of converter transformer fault monitoring technology, specifically to an active defense device and method for converter transformer fault early warning and identification based on artificial intelligence. The core of this invention is to achieve early and accurate early warning and fault identification of partial discharge in converter transformers through artificial intelligence algorithms. Background Technology

[0002] As a core component of ultra-high voltage direct current (UHVDC) transmission projects, the operating status of converter transformers directly affects the safety and stability of the power grid. Partial discharge (PD) is an early and critical sign of insulation degradation in converter transformers. It specifically refers to the phenomenon where, under the influence of an electric field, discharge occurs in certain areas of the transformer's insulation medium without penetrating between conductors. This type of discharge continuously erodes the insulation material, leading to a gradual decline in insulation performance. If not detected and intervened in time, it will eventually cause insulation breakdown, potentially triggering a transformer explosion and resulting in significant economic losses and the risk of power grid outages.

[0003] Existing partial discharge detection technologies for converter transformers mainly include oil chromatography, ultrasonic testing, and ultra-high frequency (UHF) testing. Oil chromatography, as the mainstream chemical detection method, relies on detecting the characteristic gases dissolved in transformer oil to determine the equipment status. However, it has significant shortcomings: firstly, its timeliness is extremely poor; the characteristic gases generated by partial discharge require a long time to dissolve and accumulate before reaching a detectable concentration, often indicating that the fault has already progressed to a middle or late stage by the time an anomaly is detected; secondly, its sensitivity is insufficient, making it difficult to capture trace amounts of gas generated by early, weak discharges, thus failing to provide early warning. Single ultrasonic testing relies on mechanical vibration signals, making it susceptible to interference from equipment operating noise, and its sensitivity to early, weak discharge signals is insufficient. Crucially, while conventional UHF testing can capture electromagnetic signals in the 300MHz-1.5GHz frequency band generated by partial discharges and possesses strong anti-interference capabilities, it still relies on traditional signal analysis methods without incorporating artificial intelligence algorithms. This makes it unable to accurately quantify the discharge development stage and even more difficult to accurately identify early, weak faults, resulting in a high false alarm rate and a lack of early warning capabilities—a core technical bottleneck.

[0004] Furthermore, existing technologies generally lack a comprehensive judgment mechanism combining "artificial intelligence analysis + multi-source signal fusion." Relying solely on single sensor signals and traditional analysis methods for fault diagnosis not only fails to avoid misjudgments caused by occasional interference but also cannot leverage AI algorithms to uncover hidden characteristics of early-stage faults. This results in an inability to achieve end-to-end control from early warning to proactive defense based on the severity of the fault. Therefore, developing an intelligent device with artificial intelligence judgment at its core, possessing capabilities for accurate early identification, phased judgment, and multi-source fusion defense, has become a critical requirement for ensuring the safe operation of converter transformers. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as low sensitivity, weak early identification capability, high false alarm rate, and lack of active defense capability, this invention provides an active defense device for converter transformer fault early warning and identification based on artificial intelligence. It collects signals through ultra-high frequency and high frequency sensors, achieves accurate classification of discharge stages through artificial intelligence algorithms, and issues defense commands by combining multi-sensor fusion judgment, realizing full-process control of partial discharge of converter transformers from early warning to active defense.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention discloses an active defense device for early faults of converter transformers based on artificial intelligence judgment, including a sensor module, a signal preprocessing module, an artificial intelligence analysis module, a multi-source fusion judgment module, an early warning module, and an execution module; The sensor module includes an ultra-high frequency sensor, a high frequency sensor, and an acoustic sensor, which are used to collect ultra-high frequency electromagnetic signals, high frequency electromagnetic signals, and mechanical vibration acoustic signals inside the transformer, respectively. The signal preprocessing module is used to preprocess the ultra-high frequency electromagnetic signals, high frequency electromagnetic signals and mechanical vibration acoustic signals collected by the sensor module to obtain the feature vectors of each signal. Different wavelet decomposition strategies are adopted for different sampling rates and effective frequency bands of different signals, and notch filtering and adaptive noise cancellation algorithms are superimposed to achieve multi-level noise reduction. The artificial intelligence analysis module adopts a ResNet-KNN hybrid model with an attention mechanism. It takes the pre-processed ultra-high frequency electromagnetic signal and the multi-dimensional feature fusion vector obtained by splicing the feature vectors of the high frequency electromagnetic signal as input, and outputs the partial discharge stage and the corresponding confidence vector P. The multi-source fusion judgment module calculates the degree of UHF fault based on the feature vectors of each signal and the confidence vector output by the artificial intelligence analysis module. High frequency failure level Harmony and acoustic failure degree Then, based on the stage confidence vector P, the fusion weights of the three types of fault severity are calculated, and the weighted sum is used to obtain the comprehensive fault index F; The early warning module is connected to the artificial intelligence analysis module and the multi-source fusion judgment module, and issues graded early warnings based on the confidence level of the partial discharge stage and the comprehensive fault index. The execution module performs proactive tripping defense actions based on the comprehensive fault index and the confidence level of the partial discharge stage.

[0007] More preferably, The ultra-high frequency sensor detects frequencies ranging from 300MHz to 1.5GHz, while the high frequency sensor detects frequencies ranging from 3 to 30MHz, working together to acquire electromagnetic signals of different frequencies generated by partial discharge inside the transformer. The acoustic sensor detects frequencies ranging from 10 to 60kHz and is used to acquire acoustic signals of mechanical vibrations that accompany partial discharge.

[0008] More preferably, The signal preprocessing module extracts time-domain and frequency-domain features for ultra-high frequency electromagnetic signals. It obtains marginal spectral features of the signal through successive variational mode decomposition combined with Hilbert transform. After dimensionality reduction by principal component analysis, these features are merged with the time-domain and frequency features extracted from the aforementioned ultra-high frequency electromagnetic signals. This is then expanded into a 512-dimensional ultra-high frequency electromagnetic signal feature vector through a fully connected mapping layer. For high frequency electromagnetic signals, it extracts time-domain and frequency-domain features and obtains time-frequency matrix features through short-time Fourier transform. After dimensionality reduction by principal component analysis, these time-domain matrix features are merged with the time-domain and frequency-domain features extracted from the aforementioned high frequency electromagnetic signals. This is then expanded into a 512-dimensional high frequency electromagnetic signal feature vector through a fully connected mapping layer. For acoustic signals, it extracts five time-domain features, including mean, standard deviation, peak value, kurtosis, and skewness, and frequency-domain features, including total sound pressure level, dominant frequency, and spectral flatness. It then divides the signal into eight standard frequency bands according to 1 / 3 octave bands to calculate the sound pressure level, obtaining a 13-dimensional acoustic signal feature vector.

[0009] More preferably, The signal preprocessing module performs 8-layer decomposition using the db8 wavelet for ultra-high frequency signals with a sampling rate of 5 GHz, retaining the cD2 and cD3 layers containing the discharge main frequency and discarding other layers; for high frequency signals with a sampling rate of 250 MHz, it performs 6-layer decomposition using the db6 wavelet, retaining the cD3, cD4, and cD5 layers and discarding other layers; for acoustic signals with a sampling rate of 2 MHz, it performs 5-layer decomposition using the db4 wavelet, retaining the cD5 and cA5 layers and discarding the cD1-cD4 layers. Notch filtering for ultra-high frequency signals with a center frequency of 900MHz and a bandwidth of 100MHz.

[0010] More preferably, The artificial intelligence analysis module employs a ResNet-KNN hybrid model that incorporates an attention mechanism, specifically including: A ResNet network is used as the base feature extraction network. A coordinate attention module is introduced into the residual block to enhance the model's attention to key features. The deep features output by the ResNet network are input into the KNN classifier to construct a ResNet-KNN hybrid model.

[0011] More preferably, UHF fault severity and the degree of high frequency failure Calculate according to the following formulas: ; ; The stage confidence level output by the artificial intelligence analysis module; ,Right now eigenvectors of ultra-high frequency electromagnetic signals The gradient magnitude; ,Right now eigenvectors of high-frequency electromagnetic signals The gradient magnitude; c is one of the early, middle or late stages.

[0012] More preferably, Acoustic Failure Degree Calculate according to the following formula:

[0013] in, This represents the magnitude of the acoustic eigenvector. It is the 95th percentile of the magnitude of the acoustic feature vectors of all training samples.

[0014] More preferably, The multi-source fusion judgment module calculates the comprehensive fault index F according to the following formula:

[0015] in, For ultra-high frequency weights, For high-frequency weights, The acoustic weights are summed to 1. The value of F ranges from 0 to 1.5. The fault levels are determined based on the value of F: 0 ≤ F < 0.3 is normal, 0.3 ≤ F < 0.6 is a level 1 warning, 0.6 ≤ F < 0.9 is a level 2 warning, and F ≥ 0.9 is a level 3 warning.

[0016] More preferably, The calculation of the ultra-high frequency weight, high frequency weight, and acoustic weight in the multi-source fusion judgment module is as follows:

[0017] The original weights are normalized to obtain the final fused weights. :

[0018]

[0019]

[0020]

[0021] in, , , These are the original weights for ultra-high frequency, high frequency, and acoustics, respectively. , , These represent the confidence levels for the early, middle, and late stages of discharge, respectively, output by the artificial intelligence analysis module.

[0022] More preferably, The early warning module adopts a hierarchical early warning mechanism. When the artificial intelligence analysis module determines that the early discharge or the comprehensive fault index has reached the first-level early warning threshold, a first-level early warning is issued; when it determines that the mid-term discharge or the comprehensive fault index has reached the second-level early warning threshold, a second-level early warning is issued; and when the comprehensive fault index has reached the third-level early warning threshold, a third-level early warning is issued, and the execution module action is triggered at the same time.

[0023] More preferably, The execution module includes a signal isolation unit and a trip drive unit. The signal isolation unit is used to achieve electrical isolation between the control signal and the power system. After receiving the trip command from the multi-source fusion judgment module, the trip drive unit drives the circuit breaker corresponding to the converter transformer to operate. The execution module is equipped with a dual confirmation mechanism: only when the artificial intelligence analysis module outputs the later confidence level... When the comprehensive fault index F ≥ 0.95 and the comprehensive fault index F ≥ 0.9 for 5 consecutive sampling cycles, a trip command is output.

[0024] Secondly, this invention discloses an active defense method for early faults in converter transformers based on the aforementioned active defense device, the method comprising the following steps: S1. The sensor module collects ultra-high frequency electromagnetic signals, high frequency electromagnetic signals, and mechanical vibration acoustic signals inside the transformer, respectively. S2. The signal preprocessing module preprocesses the signals acquired by the sensor module to obtain the feature vectors of ultra-high frequency electromagnetic signals, high frequency electromagnetic signals and acoustic signals. S3. Input the multi-dimensional feature vector obtained by splicing the feature vectors of ultra-high frequency electromagnetic signals and high frequency electromagnetic signals into the trained artificial intelligence analysis module, and output the partial discharge stage and the corresponding confidence vector P. S4. The multi-source fusion judgment module calculates the degree of UHF fault based on the feature vectors of each signal and the stage confidence vector output by the artificial intelligence analysis module. High frequency failure level Harmony and acoustic failure degree The fusion weights of the three types of fault severity are dynamically calculated based on the stage confidence vector P, and the comprehensive fault index F is obtained by weighted summation. S5. The early warning module issues graded early warnings based on the discharge stage output by the artificial intelligence analysis module and the comprehensive fault index calculated by the multi-source fusion judgment module. S6. The execution module performs proactive tripping defense actions based on the comprehensive fault index and the confidence level of the discharge stage.

[0025] More preferably, In S3, the artificial intelligence analysis module uses the ResNet network as the basic feature extraction network and introduces a coordinate attention module into the residual block to enhance the model's attention to key features. The deep features output by the ResNet network are input into the KNN classifier to construct a ResNet-KNN hybrid model. The hybrid model is trained using a comprehensive loss function, which combines a cross-entropy loss term and a regularity loss term. The regularity loss term is obtained by calculating the deviation between the actual feature values ​​and the evolution fitting curve of the discharge stage.

[0026] Thirdly, the present invention discloses a computer device, including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the steps of the active defense method.

[0027] Fourthly, the present invention discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the active defense method.

[0028] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention employs a combination of ultra-high frequency (UHF) and high frequency (HF) sensors to acquire signals, and integrates a ResNet-KNN hybrid model with a coordinate attention mechanism to achieve accurate classification of early, middle, and late partial discharges with a classification accuracy of 96.42%. The high frequency sensor is particularly effective in capturing early weak discharge signals, solving the problem that traditional technologies struggle to identify early weak discharges and providing a precise basis for early fault intervention. 2. By employing differentiated wavelet decomposition strategies and feature extraction methods for different signals through the signal preprocessing module, and combining the fully connected mapping layer, heterogeneous features are unified into 512-dimensional high-frequency feature vectors and 512-dimensional high-frequency feature vectors, laying the foundation for subsequent fusion. 3. The fault degree calculation method based on gradient contribution decomposition is adopted, which allocates the later confidence of the model output according to the contribution ratio of ultra-high frequency and high frequency to the decision. The independent contribution of the two signals can be separated without modifying the model structure. The physical meaning is clear and the calculation is simple and efficient. 4. A dynamic fusion weighting mechanism based on stage confidence is adopted, which adaptively adjusts the fusion weights of ultra-high frequency, high frequency, and acoustic according to the current discharge stage. In the early stage, the high frequency weight is the highest to enhance early identification, in the later stage, the ultra-high frequency weight is the highest to ensure accuracy, and when the stage is ambiguous, the acoustic weight is increased to assist in judgment, thus realizing intelligent fusion. 5. Construct a full-chain proactive defense system of "Level 1 warning - Level 2 warning - Level 3 warning - tripping", and combine it with a dual confirmation mechanism (AI confidence + comprehensive fault index + duration period) to significantly reduce the probability of false tripping and significantly improve the operational safety of converter transformers. Attached Figure Description

[0029] Figure 1 This is a logic flowchart of the implementation of the present invention; Figure 2 This is a system architecture diagram of the active defense device of the present invention; Figure 3 This is a flowchart of the training process for the artificial intelligence model of this invention; Figure 4 This is a structural diagram of the artificial intelligence model of this invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0031] See appendix Figure 1 and attached Figure 2 The present invention discloses an active defense device for early warning and identification of converter transformer faults based on artificial intelligence, comprising a sensor module, a signal preprocessing module, an artificial intelligence analysis module, a multi-source fusion judgment module, an early warning module, and an execution module.

[0032] The sensor module includes an ultra-high frequency sensor, a high frequency sensor, and an acoustic sensor; the ultra-high frequency sensor has a detection frequency band covering 300MHz-1.5GHz, the high frequency sensor has a detection frequency band covering 3-30MHz, and the two work together to collect electromagnetic signals of different frequency bands generated by partial discharge inside the transformer; the acoustic sensor has a detection frequency band covering 10-60kHz, and is used to collect the mechanical vibration acoustic signals generated by partial discharge.

[0033] The signal preprocessing module is connected to the sensor module and is used to perform noise reduction and feature extraction on ultra-high frequency electromagnetic signals, high frequency electromagnetic signals and acoustic signals. The signal preprocessing module adopts differentiated wavelet decomposition strategies for different sampling rates and effective frequency bands of different signals, and superimposes notch filtering and adaptive noise cancellation algorithms to achieve multi-level noise reduction. The signal preprocessing module extracts features from three types of signals: For ultra-high frequency electromagnetic signals, time-domain and frequency-domain features are extracted, and marginal spectral features are obtained by successive variational mode decomposition combined with Hilbert transform. After dimensionality reduction by principal component analysis, these features are merged with the time-domain and frequency features extracted from the aforementioned ultra-high frequency electromagnetic signals, and then expanded into a 512-dimensional feature vector of ultra-high frequency electromagnetic signals through a fully connected mapping layer; For high frequency electromagnetic signals, time-domain and frequency-domain features are extracted, and time-frequency matrix features are obtained by short-time Fourier transform. After dimensionality reduction by principal component analysis, these time-domain matrix features are merged with the time-domain and frequency-domain features extracted from the aforementioned high frequency electromagnetic signals, and then expanded into a 512-dimensional feature vector of high frequency electromagnetic signals through a fully connected mapping layer; For acoustic signals, five time-domain features, including mean, standard deviation, peak value, kurtosis, and skewness, and frequency-domain features, including total sound pressure level, dominant frequency, and spectral flatness, are extracted. The sound pressure level is calculated by dividing the signal into eight standard frequency bands according to 1 / 3 octave bands, resulting in a 13-dimensional feature vector of acoustic signals.

[0034] Furthermore, the signal preprocessing module employs differentiated wavelet decomposition strategies for different signals: for ultra-high frequency electromagnetic signals with a sampling rate of 5 GHz, 8-layer decomposition is performed using the db8 wavelet, retaining the cD2 and cD3 layers containing the discharge main frequency and discarding other layers; for high frequency electromagnetic signals with a sampling rate of 250 MHz, 6-layer decomposition is performed using the db6 wavelet, retaining the cD3, cD4, and cD5 layers and discarding other layers; for acoustic signals with a sampling rate of 2 MHz, 5-layer decomposition is performed using the db4 wavelet, retaining the cD5 and cA5 layers and discarding the cD1-cD4 layers; after noise reduction, a notch filter with a center frequency of 900 MHz and a bandwidth of 100 MHz is superimposed on the ultra-high frequency electromagnetic signals to filter out communication band interference; finally, LMS adaptive filtering is used for ultra-high frequency electromagnetic signals and high frequency electromagnetic signals with no-load environmental noise as the reference signal, and LMS adaptive filtering is used for acoustic signals with transformer shell vibration signal as the reference.

[0035] Furthermore, the feature extraction and mapping of the signal preprocessing module are as follows: after feature extraction, the UHF electromagnetic signal obtains 264-dimensional original features, which are then expanded to 512 dimensions through a mapping layer composed of two cascaded fully connected networks. The first layer expands the 264 dimensions to 384 dimensions and performs batch normalization, ReLU activation, and Dropout regularization, while the second layer expands the 384 dimensions to 512 dimensions and performs batch normalization and ReLU activation. After feature extraction, the HF electromagnetic signal obtains 262-dimensional original features, which are then expanded to 512 dimensions using a mapping layer with the same structure. After feature extraction, the acoustic signal obtains 13-dimensional original features, which are directly used as the feature vector of the acoustic signal.

[0036] The artificial intelligence analysis module is connected to the signal preprocessing module and adopts a ResNet-KNN hybrid model with a fused coordinate attention mechanism. It uses a 1024-dimensional fused feature vector (composed of a UHF 512-dimensional feature vector and a HF 512-dimensional feature vector) as input to achieve accurate classification of partial discharge stages. The output of the artificial intelligence analysis module includes the discharge stage classification result and the corresponding stage confidence vector. , representing the predicted probabilities for the early, middle, and late stages, respectively; The partial discharge stage is divided into an early stage, a middle stage and a late stage. The early stage is a weak discharge caused by minor insulation defects, the middle stage is a stable discharge caused by accelerated insulation deterioration, and the late stage is a strong discharge caused by insulation breakdown. Furthermore, the model training process of the artificial intelligence analysis module includes the following steps: S1. Data Acquisition: A partial discharge simulation test platform for converter transformers was built to simulate the partial discharge process under different insulation defects. Discharge signals at different stages were collected by UHF and HF sensors in tandem, and acoustic signals were recorded simultaneously as auxiliary verification data to construct a sample dataset. Each sample was labeled with the corresponding discharge stage label and sensor feature data. S2. Data Preprocessing: The 512-dimensional UHF feature vector after noise reduction, feature extraction and mapping is concatenated with the 512-dimensional HF feature vector at a 1:1 ratio to form a 1024-dimensional fused feature vector. After normalization, it is divided into training set, validation set and test set in a 7:1:2 ratio. The 13-dimensional acoustic feature vector is saved separately for subsequent multi-source fusion verification. S3. Model Construction: The 1D-ResNet50 network is used as the basic feature extraction network. Coordinate attention modules are inserted after the four feature extraction stages to enhance the model's attention to key features at different scales. The 2048-dimensional deep features output by the ResNet network are input into the KNN classifier to construct a ResNet-KNN hybrid model. S4. Model Training: The hybrid model is trained using a comprehensive loss function. The cross-entropy loss term With regular loss terms We get the weighted sum, where Take 0.6, Take 0.4; in a more preferred scheme, , ; Let be the initial weight for the cross-entropy loss term, preferably set to 0.5. To adjust the step size, a value of 0.2 is preferred. The time parameter for the current discharge phase. The longest discharge cycle in the historical data of converter partial discharge is available. The cross-entropy loss term uses label smoothing technology with a smoothing coefficient of 0.096; the regularity loss term is obtained by calculating the squared Euclidean distance between the actual 2048-dimensional feature values ​​of the sample and the theoretical values ​​of the discharge stage evolution curve, ensuring that the model follows the physical evolution law of discharge while learning data features. S5. Model Optimization: The model performance was verified using a test set. The hyperparameters of the model were optimized using a grid search method and five-fold cross-validation. Finally, an artificial intelligence analysis model with a classification accuracy of 96.42% in the discharge stage was obtained.

[0037] The multi-source fusion judgment module is connected to both the artificial intelligence analysis module and the signal preprocessing module. It receives the stage confidence vector P output by the artificial intelligence analysis module, and simultaneously receives the preprocessed UHF 512-dimensional feature vector, high-frequency 512-dimensional feature vector, and acoustic 13-dimensional feature vector. The multi-source fusion judgment module first calculates the UHF fault degree based on the gradient contribution decomposition method. and the degree of high frequency failure Calculating the degree of acoustic faults based on the feature modulus normalization method Then, based on the stage confidence vector P, the fusion weights of the three types of sensors are dynamically calculated, and finally the weighted sum is used to obtain the comprehensive fault index F, thereby realizing adaptive multi-source fusion judgment. Furthermore, the fault degree calculation of the multi-source fusion judgment module is specifically as follows: for the UHF 512-dimensional feature vector... and high-frequency 512-dimensional feature vectors Stage confidence level based on the output of the artificial intelligence analysis module (c can be selected as early, middle or late stage depending on the needs), calculate the degree of failure. and :

[0038]

[0039] in, ,Right now eigenvectors of ultra-high frequency electromagnetic signals The gradient magnitude reflects the degree of contribution of ultra-high frequency frequencies to the model's decision-making. ,Right now eigenvectors of high-frequency electromagnetic signals The gradient magnitude reflects the degree of contribution of high frequencies to the model's decision-making. For acoustic 13-dimensional feature vectors Calculate the degree of failure based on the characteristic modulus. :

[0040] in, This represents the magnitude of the acoustic eigenvector. It is the 95th percentile of the magnitude of the acoustic feature vectors of all training samples.

[0041] Furthermore, the dynamic weight calculation of the multi-source fusion judgment module is specifically as follows: based on the stage confidence vector Calculate the original weights of the three types of sensors:

[0042] The original weights are normalized to obtain the final fused weights. :

[0043]

[0044]

[0045]

[0046] in, , , These are the original weights for ultra-high frequency, high frequency, and acoustics, respectively. , , These represent the confidence levels for the early, middle, and late stages of discharge, respectively, output by the artificial intelligence analysis module.

[0047] The early warning module is connected to the artificial intelligence analysis module and the multi-source fusion judgment module, and issues graded early warnings based on the discharge stage classification results and the comprehensive fault index F. The execution module is connected to the multi-source fusion judgment module and is used to receive the defense instructions output by the multi-source fusion judgment module and execute active defense actions such as tripping.

[0048] The overall failure index F is obtained by weighted summation of the failure levels of the three sensors and the dynamic fusion weights:

[0049] The value of F ranges from 0 to 1.5. The fault level is divided according to the value of F: 0≤F<0.3 is normal, 0.3≤F<0.6 is a level 1 warning, 0.6≤F<0.9 is a level 2 warning, and F≥0.9 is a level 3 warning.

[0050] The early warning module adopts a hierarchical early warning mechanism: when the artificial intelligence analysis module determines that the early discharge or the comprehensive fault index has reached the first-level early warning threshold, a first-level early warning is issued; when it determines that the mid-term discharge or the comprehensive fault index has reached the second-level early warning threshold, a second-level early warning is issued; when the comprehensive fault index has reached the third-level early warning threshold, a third-level early warning is issued, and the execution module action is triggered at the same time.

[0051] The execution module includes a signal isolation unit and a trip drive unit. The signal isolation unit is used to achieve electrical isolation between the control signal and the power system to ensure operational safety. After receiving the trip command from the multi-source fusion judgment module, the trip drive unit drives the circuit breaker corresponding to the converter transformer to operate, achieving rapid tripping. The execution module is equipped with a dual confirmation mechanism: only when the artificial intelligence analysis module outputs the later confidence level... A trip command is only output when the comprehensive fault index F ≥ 0.95 and continues to be ≥ 0.9 for 5 sampling cycles.

[0052] The device also includes a data storage and communication module, which stores the sensor's raw data, preprocessed data, and model analysis results in a hierarchical manner according to "raw signal - preprocessed data - analysis results," and uploads the data to the background monitoring system via the network to achieve remote monitoring and data traceability.

[0053] This invention also discloses an active defense method for early faults in converter transformers based on the aforementioned active defense device, the method comprising the following steps: S1. The sensor module collects ultra-high frequency electromagnetic signals, high frequency electromagnetic signals, and mechanical vibration acoustic signals inside the transformer, respectively. S2. The signal preprocessing module preprocesses the signals acquired by the sensor module to extract the time-domain and frequency-domain features of each signal. After noise reduction, the feature vectors of ultra-high frequency electromagnetic signals, high frequency electromagnetic signals, and acoustic signals are obtained. S3. Input the multi-dimensional feature fusion vector obtained by splicing the feature vectors of ultra-high frequency electromagnetic signals and high frequency signals into the trained artificial intelligence analysis module, and output the partial discharge stage and confidence vector P. The artificial intelligence analysis module uses the ResNet network as the basic feature extraction network and introduces a coordinate attention module into the residual block to enhance the model's attention to key features. The deep features output by the ResNet network are input into the KNN classifier to construct a ResNet-KNN hybrid model. The hybrid model is trained using a comprehensive loss function, which combines a cross-entropy loss term and a regularity loss term. The regularity loss term is obtained by calculating the deviation between the actual feature values ​​and the evolution fitting curve of the discharge stage. S4. Calculate the degree of UHF fault through the multi-source fusion judgment module. High frequency fault degree Harmony and acoustic failure degree The fusion weights of the three types of fault severity are dynamically calculated based on the stage confidence vector P, and the comprehensive fault index F is obtained by weighted summation. S5. The early warning module issues graded early warnings based on the discharge stage and the comprehensive fault index. S6. The execution module performs proactive tripping defense actions based on the comprehensive fault index and the confidence level of the discharge stage.

[0054] Example 1: This embodiment is based on the test and application platform built on the converter transformer of a ±200kV converter station. The selection and parameters of each sensor are as follows: The DN125 acoustic-electric integrated telescopic sensor is a UHF and acoustic dual-mode sensor designed specifically for transformer partial discharge monitoring. It integrates UHF electromagnetic wave signal monitoring and ultrasonic signal monitoring technologies, and can comprehensively capture electromagnetic wave signals and ultrasonic signals caused by insulation defects inside the transformer. The UHF sensor has a center frequency of about 700MHz, a bandwidth of 300MHz-1.5GHz, and an average equivalent height of more than 8mm, making it suitable for complex internal working conditions of transformers; the acoustic sensor has a detection frequency band of 10kHz-80kHz and a peak sensitivity of 90dB (V / (m / s)); the high-frequency sensor is an inductively coupled sensor of model XD10T6, with a detection frequency band of 150kHz-80MHz, a sensitivity of ≤0.1pC, a transmission impedance of >20mV / mA, and a matching impedance of 50Ω, which focuses on capturing low-frequency electromagnetic radiation signals of early weak discharge. The three work together, with the ultra-high frequency and high frequency sensors synchronously acquiring electromagnetic signals, and the time synchronization error is negligible. The acoustic sensor delay is ≤35ms, ensuring signal timing consistency.

[0055] The signal preprocessing module employs a multi-level noise reduction strategy, processing each signal according to its sampling rate and effective frequency band: For ultra-high frequency signals (sampling rate 5GHz, effective frequency band 300MHz-1.5GHz): 8-layer decomposition is performed using the dB8 wavelet, retaining the cD2 layer (625-1250MHz) and cD3 layer (312.5-625MHz) containing the discharge main frequency, while discarding other layers; a notch filter with a center frequency of 900MHz, bandwidth of 100MHz, and notch depth of 30dB is superimposed to filter out communication band interference; finally, LMS adaptive filtering using idle ambient noise as the reference signal is used to eliminate environmental electromagnetic interference; For high frequency signals (sampling rate 250MHz, effective frequency band 3-30MHz): ... The db6 wavelet is used for 6-level decomposition, retaining the cD3 layer (15.625-31.25MHz), cD4 layer (7.8125-15.625MHz), and cD5 layer (3.90625-7.8125MHz), while discarding other layers. LMS adaptive filtering using unloaded ambient noise as the reference signal is used to eliminate environmental electromagnetic interference. For acoustic signals (sampling rate 2MHz, effective frequency band 10-60kHz): db4 wavelet is used for 5-level decomposition, retaining the cD5 layer (31.25-62.5kHz) and cA5 layer (0-31.25kHz), while discarding cD1-cD4 layers. LMS adaptive filtering using transformer shell vibration signal as the reference is used to eliminate mechanical vibration interference.

[0056] After noise reduction, feature extraction was performed on the three types of signals: For the UHF signal, four-dimensional time-domain features (pulse amplitude, rise time, pulse repetition rate, peak factor) and four-dimensional frequency-domain features (spectral peak value, center frequency, bandwidth, spectral entropy) were extracted. The signal was then decomposed into six intrinsic mode functions through successive variational mode decomposition. Hilbert transform was applied to the dominant components to obtain marginal spectral features. These marginal spectral features were then subjected to principal component analysis for dimensionality reduction, resulting in 256-dimensional data. The total 264-dimensional feature vectors were input into a fully connected mapping layer, which consists of two cascaded fully connected networks. The first fully connected network layer expands the 264 dimensions to 384 dimensions and then sequentially processes them through batch normalization, ReLU activation function, and Dropout regularization (dropout rate 0.3) to enhance feature representation and prevent overfitting. The second fully connected network further expands the 384-dimensional feature vector to 512-dimensional features. After batch normalization and ReLU activation, it finally outputs a 512-dimensional mapped feature vector. For high-frequency signals, it extracts 3-dimensional time-domain features (pulse duration, amplitude variance, and average amplitude) and 3-dimensional frequency-domain features (dominant frequency, spectral energy percentage, and harmonic order). It then generates a 256×256-dimensional time-frequency matrix through short-time Fourier transform, which is reduced to 256-dimensional through principal component analysis. The high-frequency features total 262-dimensional. Similarly, a fully connected network is used to transform the 262-dimensional features. The 10-dimensional feature vector is mapped to a 512-dimensional feature vector. For the acoustic signal, 3-dimensional time-domain features (pulse amplitude, pulse duration, peak factor) and 2-dimensional frequency-domain features (dominant frequency, spectral energy percentage) are extracted. The 10-60kHz frequency range is divided into 8 standard frequency bands (center frequencies of 10kHz, 12.5kHz, 16kHz, 20kHz, 25kHz, 31.5kHz, 40kHz, and 50kHz) according to 1 / 3 octave bands. The sound pressure level of each frequency band is calculated to obtain 8-dimensional octave band spectral features, resulting in a total of 13 acoustic features. Finally, the ultra-high frequency features and high frequency features are normalized and then concatenated at a 1:1 ratio to obtain the 1024-dimensional input feature vector of the model in the artificial intelligence analysis. This vector is then divided into training, validation, and test sets at a 7:1:2 ratio. The 13-dimensional features of the acoustic signal are saved separately for auxiliary verification in the subsequent multi-source fusion judgment module.

[0057] The artificial intelligence analysis module employs a ResNet-KNN hybrid model that integrates coordinate attention mechanisms. Its training process is as follows: Figure 3As shown, the specific structure is as follows: First, an input adaptation layer is constructed to reshape the 1024-dimensional input features into a shape of [n, 1, 1024]. Then, through a 1-dimensional convolutional layer Conv1D and batch normalization and ReLU activation function, and then through a 1-dimensional max pooling layer, the output shape is [n, 64, 256]. Based on this, an improved 1D-ResNet50 backbone network is constructed, comprising four feature extraction stages. The first stage consists of three stacked bottleneck residual blocks. Each bottleneck residual block contains a three-layer structure: 1×1 convolutional dimensionality reduction, 3×3 convolutional spatial feature extraction, and 1×1 convolutional dimensionality enhancement, with residual connections introduced. The output shape of this stage is [n, 256, 256], with a receptive field of approximately 20 time points, corresponding to the detailed features of a single discharge pulse. After this stage, the first coordinate attention module is inserted, generating temporal dimension weights through a temporal attention branch and channel dimension weights through a channel attention branch. The input features are multiplied element-wise with the two weights to achieve precise localization enhancement of the single pulse position. The second stage consists of four stacked bottleneck residual blocks. The first residual block is set with a stride of 2 for downsampling, and the output shape is [n, 512, 128]. The first stage consists of approximately 80 receptive field points, corresponding to short-duration patterns composed of multiple pulses. A second coordinate attention module is inserted after this stage to enhance the focus on pulse repetition patterns. The third stage consists of six stacked bottleneck residual blocks. The first residual block is set with a step size of 2 to achieve downsampling, and the output shape is [n, 1024, 64]. The receptive field is approximately 300 time points, corresponding to the discharge variation trend with the power frequency cycle. A third coordinate attention module is inserted after this stage to enhance the focus on key points of the development trend. The fourth stage consists of three stacked bottleneck residual blocks. The first residual block is set with a step size of 2 to achieve downsampling, and the output shape is [n, 2048, 32]. The receptive field is approximately 1000 time points, almost covering the entire signal length, corresponding to the overall stage characteristics of the discharge. A fourth coordinate attention module is inserted after this stage to enhance the focus on key features for stage discrimination. After the four feature extraction stages, a global average pooling layer is used to average the time dimension, compressing the 2048 feature channels of each sample into a 2048-dimensional feature vector. This vector integrates multi-scale information from local details to overall trends and undergoes four attention enhancements to highlight key features related to discharge stage discrimination. Finally, a KNN classifier is constructed, using the 2048-dimensional feature vector as input and constructing a feature library from the feature vectors of all samples in the training set. For the test sample, the Euclidean distance between it and each sample in the feature library is calculated. The K samples with the smallest distances (K=15, determined through grid search optimization) are selected, and the discharge stage labels of these K samples are statistically analyzed. A majority voting method is used to determine the final classification result, outputting early, middle, or late discharge. The model structure is as follows. Figure 4 As shown.

[0058] The training of this model uses a comprehensive loss function, which is obtained by weighted summation of the cross-entropy loss term and the regularity loss term. Its expression is as follows:

[0059] in, This is the weight of the cross-entropy loss term, with a value of 0.6; The weight of the regularity loss term is 0.4; For cross-entropy loss term, This is the regularity loss term. The cross-entropy loss term uses label smoothing to convert hard labels (one-hot encoding) into soft labels, preventing the model from becoming overconfident and overfitting. For the true class, its target probability is no longer 1, but... For other categories, the target probability is no longer 0, but... ,in For smoothing coefficients, Number of categories ( (corresponding to early discharge, mid-discharge, and late discharge) During this training process The value is 0.096. The formula for calculating the cross-entropy loss after label smoothing is:

[0060] in, The total number of samples, Number of categories ( ), The target probability after label smoothing. Predict the probability that the i-th sample belongs to class c for the model.

[0061] The regularity loss term is used to constrain the features learned by the model to conform to the physical laws of discharge evolution. During the development of partial discharge from its early to middle and then late stages, the sample distribution in the feature space should exhibit a continuous evolution trend, rather than a random distribution. Therefore, based on extensive experimental data and the physical mechanism of discharge, a library of discharge stage evolution curves is fitted. This library describes the theoretical changes in key feature parameters of different models from the early to late stages. The regularity loss term is obtained by calculating the deviation between the actual feature values ​​of the samples and the theoretical values ​​of the evolution curves; its expression is:

[0062] in, Let be the deep feature vector extracted from the i-th sample using the ResNet network, with a dimension of 2048. The theoretical feature vector is obtained by interpolating the quantized value of the discharge stage of the i-th sample onto the evolution curve. It has a dimension of 2048, and N is the total number of samples.

[0063] The training process employed the Adam optimizer with an initial learning rate of 0.001. A cosine annealing strategy was used to dynamically adjust the learning rate, with a maximum of 100 iterations. The batch size was set to 32, and the training run consisted of 200 iterations. In each training round, forward propagation was used to calculate the comprehensive loss, followed by backpropagation to update the network parameters, ensuring that the model's feature distribution conformed to the physical laws of discharge evolution while maintaining accurate classification.

[0064] After model training, the model performance was evaluated using a test set. Five-fold cross-validation was performed using a grid search method, dividing the training set into five parts. Four parts were used for training and one part for validation in rotation, and the average accuracy of the five validations was used as the evaluation metric. Finally, the hyperparameter combination that achieved the highest accuracy on the validation set was selected as the model parameters. The resulting AI analysis model achieved an overall classification accuracy of 96.42% on the test set, with a 94.69% accuracy rate for identifying early-stage discharges. This represents a 4.3 percentage point improvement over the model without high-frequency signal input, validating the effectiveness and innovation of the present invention.

[0065] The multi-source fusion judgment module receives the preprocessed UHF 512-dimensional feature vector, HF 512-dimensional feature vector, and acoustic 13-dimensional feature vector. First, it calculates the feature normalization value to characterize the discharge severity reflected by each sensor signal. For the UHF feature vector... The degree of its failure The calculation formula is as follows:

[0066] in, The confidence level c represents the corresponding stage output of the ResNet-KNN model (c = early stage, middle stage, late stage). ,Right now For UHF input The gradient magnitude reflects the degree of contribution of ultra-high frequency frequencies to the model's decision-making. ,Right now For high frequency input The gradient magnitude reflects the contribution of high-frequency features to the model's decision-making; for high-frequency feature vectors... The degree of its failure The calculation formula is as follows:

[0067] The formula has the same meaning as above.

[0068] Since acoustic signals are not involved in the training of the ResNet-KNN model, their fault degree Based on independent calculation of feature magnitude, the formula is as follows:

[0069] in, , representing the magnitude of the acoustic eigenvector. The statistical constant for the training set is the 95th percentile of the magnitude of the acoustic feature vectors of all training samples.

[0070] After calculating the severity of the fault , , Subsequently, the multi-source fusion judgment module dynamically calculates the weights of the three types of sensors during fusion based on the stage confidence vector P. The calculation principle is as follows:

[0071] in, , , The confidence vector representing the output of the ResNet-KNN model. To ensure that the sum of the three weights is 1, normalization is performed:

[0072]

[0073]

[0074]

[0075] The comprehensive fault index F is obtained by weighting and summing the fault levels of the three sensors with the dynamic fusion weights:

[0076] The magnitude of the F value reflects the overall severity of the transformer fault, as detailed below:

[0077] In the early warning module, during the first-level early warning (early discharge), the yellow indicator light on site flashes at a frequency of 2Hz, the buzzer emits a prompt tone at 1-second intervals, and at the same time, a pop-up warning (including discharge stage and characteristic parameters) is sent to the background monitoring system via the IEC61850 protocol; during the second-level early warning (mid-term discharge), the orange indicator light stays on, the buzzer sounds continuously, and the background system marks it as an "urgent defect"; during the third-level early warning (late discharge), the red indicator light flashes at a high frequency (10Hz), a high-pitched alarm is emitted, and the execution module is triggered to take action.

[0078] After receiving the trip command, the execution module drives the relay to operate via DC power, outputting a trip signal to the high-voltage side circuit breaker of the converter transformer, with an operation time of ≤50ms. Simultaneously, it feeds back the operation status to the multi-source fusion judgment module, forming a "command-execution-feedback" closed loop. To avoid false trips, a dual confirmation mechanism is implemented: a trip command is only output when the AI ​​model confidence level is ≥95% and the comprehensive fault index is ≥0.9 for five consecutive sampling periods.

[0079] The data storage and communication module uses a 2TB industrial-grade hard drive, storing data in layers: raw signal - preprocessed data - analysis results. Raw signals are stored at 10-minute intervals per file, while preprocessed data and analysis results (including feature vectors, discharge stages, fault indices, etc.) are stored at 1-minute intervals per record. The data retention period can be set via the backend. Communication is via wired Ethernet interface (RJ45) to the station control layer, using a Layer 2 protocol with a data transmission rate ≥1000Mbps. The backend monitoring system, developed based on SophiicCloud, features real-time data display, historical data query, fault trend analysis, and report generation. It can intuitively display the signal waveforms, spectrum diagrams, and comprehensive fault index change curves of the three sensors.

[0080] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0081] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0082] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0083] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. An active defense device for early faults in converter transformers based on artificial intelligence judgment, comprising a sensor module, a signal preprocessing module, an artificial intelligence analysis module, a multi-source fusion judgment module, an early warning module, and an execution module; characterized in that: The sensor module includes an ultra-high frequency sensor, a high frequency sensor, and an acoustic sensor, which are used to collect ultra-high frequency electromagnetic signals, high frequency electromagnetic signals, and mechanical vibration acoustic signals inside the transformer, respectively. The signal preprocessing module is used to preprocess the ultra-high frequency electromagnetic signals, high frequency electromagnetic signals and mechanical vibration acoustic signals collected by the sensor module to obtain the feature vectors of each signal. Different wavelet decomposition strategies are adopted for different sampling rates and effective frequency bands of different signals, and notch filtering and adaptive noise cancellation algorithms are superimposed to achieve multi-level noise reduction. The artificial intelligence analysis module adopts a ResNet-KNN hybrid model with an attention mechanism. It takes the pre-processed ultra-high frequency electromagnetic signal and the multi-dimensional feature fusion vector obtained by splicing the feature vectors of the high frequency electromagnetic signal as input, and outputs the partial discharge stage and the corresponding confidence vector P. The multi-source fusion judgment module calculates the degree of UHF fault based on the feature vectors of each signal and the confidence vector output by the artificial intelligence analysis module. High frequency fault degree Harmony and acoustic failure degree Then, based on the stage confidence vector P, the fusion weights of the three types of fault severity are calculated, and the weighted sum is used to obtain the comprehensive fault index F; ； ； ； ， ， ； ； ， ； ； The stage confidence level output by the artificial intelligence analysis module; ,Right now eigenvectors of ultra-high frequency electromagnetic signals The gradient magnitude; ,Right now eigenvectors of high-frequency electromagnetic signals The gradient magnitude; c is one of the early, middle, or late stages; , , These are the original weights for ultra-high frequency, high frequency, and acoustics, respectively. , , These are the confidence levels for the early, middle, and late stages of discharge output by the artificial intelligence analysis module, respectively. The early warning module is connected to the artificial intelligence analysis module and the multi-source fusion judgment module, and issues graded early warnings based on the confidence level of the partial discharge stage and the comprehensive fault index. The execution module performs proactive tripping defense actions based on the comprehensive fault index and the confidence level of the partial discharge stage.

2. The active defense device for early faults of converter transformers according to claim 1, characterized in that: The ultra-high frequency sensor detects frequencies ranging from 300MHz to 1.5GHz, while the high frequency sensor detects frequencies ranging from 3 to 30MHz, working together to acquire electromagnetic signals of different frequencies generated by partial discharge inside the transformer. The acoustic sensor detects frequencies ranging from 10 to 60kHz and is used to acquire acoustic signals of mechanical vibrations that accompany partial discharge.

3. The active defense device for early faults of converter transformers according to claim 2, characterized in that: The signal preprocessing module extracts time-domain and frequency-domain features for ultra-high frequency electromagnetic signals. It obtains marginal spectral features of the signal through successive variational mode decomposition combined with Hilbert transform. After dimensionality reduction by principal component analysis, these features are merged with the time-domain and frequency features extracted from the aforementioned ultra-high frequency electromagnetic signals. This is then expanded into a 512-dimensional ultra-high frequency electromagnetic signal feature vector through a fully connected mapping layer. For high frequency electromagnetic signals, it extracts time-domain and frequency-domain features and obtains time-frequency matrix features through short-time Fourier transform. After dimensionality reduction by principal component analysis, these time-domain matrix features are merged with the time-domain and frequency-domain features extracted from the aforementioned high frequency electromagnetic signals. This is then expanded into a 512-dimensional high frequency electromagnetic signal feature vector through a fully connected mapping layer. For acoustic signals, it extracts five time-domain features, including mean, standard deviation, peak value, kurtosis, and skewness, and frequency-domain features, including total sound pressure level, dominant frequency, and spectral flatness. It then divides the signal into eight standard frequency bands according to 1 / 3 octave bands to calculate the sound pressure level, obtaining a 13-dimensional acoustic signal feature vector.

4. The active defense device for early faults of converter transformers according to claim 3, characterized in that: The signal preprocessing module performs 8-layer decomposition using the db8 wavelet for ultra-high frequency signals with a sampling rate of 5 GHz, retaining the cD2 and cD3 layers containing the discharge main frequency and discarding other layers; for high frequency signals with a sampling rate of 250 MHz, it performs 6-layer decomposition using the db6 wavelet, retaining the cD3, cD4, and cD5 layers and discarding other layers; for acoustic signals with a sampling rate of 2 MHz, it performs 5-layer decomposition using the db4 wavelet, retaining the cD5 and cA5 layers and discarding the cD1-cD4 layers. Notch filtering for ultra-high frequency signals with a center frequency of 900MHz and a bandwidth of 100MHz.

5. The active defense device for early faults of converter transformers according to claim 1, characterized in that: The artificial intelligence analysis module employs a ResNet-KNN hybrid model that incorporates an attention mechanism, specifically including: A ResNet network is used as the base feature extraction network. A coordinate attention module is introduced into the residual block to enhance the model's attention to key features. The deep features output by the ResNet network are input into the KNN classifier to construct a ResNet-KNN hybrid model.

6. The active defense device for early faults of converter transformers according to claim 1, characterized in that: Acoustic Failure Degree Calculate according to the following formula: in, This represents the magnitude of the acoustic eigenvector. It is the 95th percentile of the magnitude of the acoustic feature vectors of all training samples.

7. The active defense device for early faults of converter transformers according to claim 1 or 6, characterized in that: The value of F ranges from 0 to 1.

5. The fault level is divided according to the value of F: 0≤F<0.3 is normal, 0.3≤F<0.6 is a level 1 warning, 0.6≤F<0.9 is a level 2 warning, and F≥0.9 is a level 3 warning.

8. The active defense device for early faults of converter transformers according to claim 1, characterized in that: The early warning module adopts a tiered early warning mechanism. When the artificial intelligence analysis module determines that the early discharge or comprehensive fault index has reached the first-level early warning threshold, a first-level early warning is issued; when it determines that the mid-term discharge or comprehensive fault index has reached the second-level early warning threshold, a second-level early warning is issued. When the comprehensive fault index reaches the level 3 warning threshold, a level 3 warning is issued, and the execution module action is triggered.

9. The active defense device for early faults of converter transformers according to claim 8, characterized in that: The execution module includes a signal isolation unit and a trip drive unit. The signal isolation unit is used to achieve electrical isolation between the control signal and the power system. After receiving the trip command from the multi-source fusion judgment module, the trip drive unit drives the circuit breaker corresponding to the converter transformer to operate. The execution module is equipped with a dual confirmation mechanism: only when the artificial intelligence analysis module outputs the later confidence level... When the comprehensive fault index F ≥ 0.95 and the comprehensive fault index F ≥ 0.9 for 5 consecutive sampling cycles, a trip command is output.

10. A method for proactive defense against early faults in converter transformers based on the device described in any one of claims 1-9, characterized in that, The method includes the following steps: S1. The sensor module collects ultra-high frequency electromagnetic signals, high frequency electromagnetic signals, and mechanical vibration acoustic signals inside the transformer. S2. The signal preprocessing module preprocesses the signals acquired by the sensor module to obtain the feature vectors of ultra-high frequency electromagnetic signals, high frequency electromagnetic signals and acoustic signals. S3. Input the multi-dimensional feature vector obtained by splicing the feature vectors of ultra-high frequency electromagnetic signals and high frequency electromagnetic signals into the trained artificial intelligence analysis module, and output the partial discharge stage and the corresponding confidence vector P. S4. The multi-source fusion judgment module calculates the degree of UHF fault based on the feature vectors of each signal and the stage confidence vector output by the artificial intelligence analysis module. High frequency fault degree Harmony and acoustic failure degree The fusion weights of the three types of fault severity are dynamically calculated based on the stage confidence vector P, and the comprehensive fault index F is obtained by weighted summation. S5. The early warning module issues graded early warnings based on the discharge stage output by the artificial intelligence analysis module and the comprehensive fault index calculated by the multi-source fusion judgment module. S6. The execution module performs proactive tripping defense actions based on the comprehensive fault index and the confidence level of the discharge stage.

11. The active defense method for early faults in converter transformers according to claim 10, characterized in that: In S3, the artificial intelligence analysis module uses the ResNet network as the basic feature extraction network and introduces a coordinate attention module into the residual block to enhance the model's attention to key features. The deep features output by the ResNet network are input into the KNN classifier to construct a ResNet-KNN hybrid model. The hybrid model is trained using a comprehensive loss function, which combines a cross-entropy loss term and a regularity loss term. The regularity loss term is obtained by calculating the deviation between the actual feature values ​​and the evolution fitting curve of the discharge stage.

12. A computer device, comprising a memory, a processor, and a computer program stored in the memory; characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 10-11.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program performs the steps of the method according to any one of claims 10-11.

Images