Transformer inrush current identification and early warning integrated method based on multi-sensor fusion

Abstract

The invention discloses a transformer inrush current identification and early warning integrated method based on multi-sensor fusion, and relates to the technical field of transformer equipment, and the method comprises the following steps: S1, carrying out the distributed deployment of multiple sensors and the construction of a self-adaptive sensing network; s2, performing multi-dimensional feature correction and time alignment; s3, fuzzy characteristics are extracted, and a complex nonlinear relation in the operation state of the transformer is captured; s4, fusing the data after the fuzzy characteristics are extracted; s5, constructing an inrush current identification model, and carrying out the precise identification of the inrush current of the transformer; and S6, carrying out real-time early warning. According to the transformer inrush current identification and early warning integrated method based on multi-sensor fusion, dynamic extraction and deep learning modeling of multi-modal features are realized by introducing a distributed multi-sensor fusion technology, and real-time dynamic threshold calculation, reinforcement learning optimization and Bayesian fault probability inference are combined, so that the transformer inrush current identification and early warning integrated method is realized. And integration of inrush current identification and early warning of the transformer is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer equipment, and specifically provides an integrated method for transformer inrush current identification and early warning based on multi-sensor fusion. Background Art

[0002] Transformer inrush current is a common transient phenomenon in the operation of power systems. Due to its similar characteristics to fault current, it is prone to cause misjudgment and misoperation, thereby affecting power supply stability and equipment safety. At present, traditional inrush current identification methods mainly rely on the physical property analysis of a single sensor (such as the waveform of exciting current) or fixed empirical thresholds, and it is difficult to cope with environmental interference and working condition changes in complex power systems. In addition, some methods process inrush current identification and early warning independently and fail to achieve real-time linkage, resulting in lag in fault prediction and early warning. Summary of the Invention

[0003] Aiming at the deficiencies of the existing technology, the present invention provides an integrated method for transformer inrush current identification and early warning based on multi-sensor fusion to solve the problems raised in the above background art.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] In a first aspect, an embodiment of the present invention provides an integrated method for transformer inrush current identification and early warning based on multi-sensor fusion, including the following steps:

[0006] S1. Perform distributed deployment of multi-sensors and construction of an adaptive perception network, and collect sensor data;

[0007] S2. Perform multi-dimensional feature correction and time alignment on the collected sensor data;

[0008] S3. Extract fuzzy characteristics from the corrected data to capture complex non-linear relationships in the transformer operating state;

[0009] S4. Fuse the data after extracting fuzzy characteristics and generate a multi-modal fusion feature vector;

[0010] S5. Construct an inrush current identification model, input the multi-modal fusion feature vector, and accurately identify the transformer inrush current;

[0011] S6. Based on the inrush current identified by the inrush current identification model, analyze the precursor characteristics of the abnormal state and perform real-time early warning.

[0012] To further optimize this technical solution, in step S1, in the power equipment site where the transformer is located, sensors including current, voltage, magnetic field, and temperature are arranged based on distributed deployment; each node in the adaptive perception network is interconnected through a low-latency wireless protocol to ensure that the sensors work in real-time synchronization under dynamic working conditions.

[0013] To further optimize this technical solution, in step S2, a dynamic multi-dimensional feature mapping model is constructed to perform feature alignment of sensor data across time and dimensions;

[0014] In the dynamic multi-dimensional feature mapping model, it is assumed that there are sensors, and the recorded data matrix is , where , represents the feature dimension recorded by the sensor;

[0015] The mapped feature is defined as .

[0016] To further optimize this technical solution, the dynamic multi-dimensional feature mapping model includes:

[0017] Time correction:

[0018] Based on interpolation and dynamic weight optimization, the time alignment is performed, and the corrected data is:

[0019] ;

[0020] Among them,

[0021] represents the clock drift of sensor , which is dynamically updated based on the global synchronization mechanism;

[0022] is the correction coefficient, which is adjusted according to the real-time performance of the sensor;

[0023] Spatial deviation correction:

[0024] Based on the linear and non-linear combination model, the correction is:

[0025] ;

[0026] Among them,

[0027] is the linear correction matrix of sensor , which is obtained by least squares estimation;

[0028] is the non-linear adjustment factor, which is optimized through the local deviation model;

[0029] represents a non - linear correction function;

[0030] Scale normalization:

[0031] For the scale differences of multi - dimensional features, the normalization formula is:

[0032] ;

[0033] where,

[0034] and are the mean and standard deviation of the data collected by the sensor respectively, and are calculated dynamically based on a sliding window;

[0035] Final mapping:

[0036] The mapped feature is calculated as:

[0037] ;

[0038] where,

[0039] is the weight matrix, which is optimized and generated based on mutual - information constraints to ensure the maximization of the correlation between features.

[0040] To further optimize this technical solution, in the dynamic multi - dimensional feature mapping model:

[0041] Time correction : After obtaining the sensor data, first correct the clock drift by dynamically adjusting and to ensure the synchronization of each sensor data in the time dimension;

[0042] Spatial deviation correction : Perform spatial correction on the data after time correction through the linear correction matrix and the non - linear function to correct the deviation caused by the position or environmental factors of the sensor;

[0043] Scale normalization : After spatial correction, normalize the data. The and calculated by the sliding window adapt to the changes in sensor performance dynamically to ensure the data balance in the unified feature space;

[0044] Feature mapping : Finally, through the weight matrix Feature mapping is performed on the normalized data, and the weight matrix is generated by optimizing mutual information to ensure high feature correlation and low redundancy, providing an optimized multi-dimensional input for the follow-up.

[0045] To further optimize this technical solution, in step S3, by performing eigen-decomposition on the corrected data and using the fuzzy logic analysis method, a rule-based fuzzy feature library is constructed. The specific steps include:

[0046] Using the fuzzy clustering algorithm to divide the eigenvalues of current, voltage, and magnetic field into multiple fuzzy sets, including "normal", "critical", and "abnormal";

[0047] Designing fuzzy rules between different fuzzy sets to quantify the state transition probability of each feature;

[0048] Introducing a fuzzy weight adjustment mechanism based on the entropy weight method to make the rule library have self-adaptability to the importance of features.

[0049] To further optimize this technical solution, in step S4, based on the data fusion of the graph neural network GNN, the sensor data is modeled as a graph structure;

[0050] The sensor nodes are used as the vertices of the graph, and the associations between different modal features are used as the edges of the graph. Through the hierarchical propagation mechanism of GNN, global fusion features are extracted;

[0051] At the same time, using the multi-head attention mechanism to enhance the weights of key features and endowing higher discriminative ability to important parameters;

[0052] Finally, a multi-modal fusion feature vector is generated as the input of the subsequent inrush current identification model.

[0053] To further optimize this technical solution, in step S5, the input of the inrush current identification model is the multi-modal fusion feature vector output by step S4 , dynamically analyzing the complex relationships between features to generate the inrush current classification result;

[0054] The inrush current identification model includes local feature extraction, time-dependent modeling, dynamic optimization, and classification decision-making.

[0055] To further optimize this technical solution, in the inrush current identification model:

[0056] Local feature extraction:

[0057] Convolutional operations are used to extract the local characteristics of the feature vector, and the feature is represented as:

[0058] ;

[0059] Among them,

[0060] is the feature matrix after the -layer convolution;

[0061] and are the convolution kernel and the bias term, respectively;

[0062] is the non-linear activation function;

[0063] Temporal Dependency Modeling:

[0064] Perform temporal modeling on the feature sequence extracted by convolution:

[0065] ;

[0066] where

[0067] is the hidden state at time ;

[0068] represents the feature matrix extracted by convolution as the input, and the feature vector generated by convolution, realizing the temporal dynamic modeling of the convolution feature sequence;

[0069] , , are 's mapping weight, the mapping weight of the previous moment, and the overall bias term, respectively;

[0070] is the activation function;

[0071] Dynamic Optimization:

[0072] Introduce reinforcement learning to dynamically adjust the classification threshold , and optimize the reward function:

[0073] ;

[0074] where

[0075] is the reward value;

[0076] represents the model accuracy, which is the proportion of correctly classified samples in the total samples;

[0077] and are the false positive and false negative rates, respectively, reflecting the proportion of negative samples misjudged as positive samples, that is, the probability of false alarms, It reflects the proportion of positive samples misjudged as negative samples, that is, the probability of missed reports;

[0078] is the penalty coefficient;

[0079] Classification decision:

[0080] Based on the output of local feature extraction and time-dependent modeling, classification is performed:

[0081] ;

[0082] Among them,

[0083] is the classification result, including "normal inrush current" and "abnormal inrush current";

[0084] , are the weights and biases of the output layer.

[0085] To further optimize this technical solution, in step S6, real-time early warning includes:

[0086] Precursor feature extraction;

[0087] Dynamic threshold calculation;

[0088] Fault probability inference;

[0089] Real-time early warning decision.

[0090] In a second aspect, an embodiment of the present invention provides a computer device, including a memory and a processor. The memory stores a computer program, where: when the computer program instructions are executed by the processor, the steps of a method for integrated identification and early warning of transformer inrush current based on multi-sensor fusion as described in the first aspect of the present invention are implemented.

[0091] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, where: when the computer program instructions are executed by the processor, the steps of a method for integrated identification and early warning of transformer inrush current based on multi-sensor fusion as described in the first aspect of the present invention are implemented.

[0092] Compared with the prior art, the present invention provides a method for integrated identification and early warning of transformer inrush current based on multi-sensor fusion, having the following beneficial effects:

[0093] The integrated method for transformer inrush current identification and early warning based on multi-sensor fusion realizes the dynamic extraction of multi-modal features and deep learning modeling by introducing distributed multi-sensor fusion technology, and combines real-time dynamic threshold calculation, reinforcement learning optimization and Bayesian fault probability inference to achieve the integration of transformer inrush current identification and early warning. Compared with the existing technology, this method has higher inrush current identification accuracy and robustness under complex working conditions. At the same time, the early warning function can respond to system changes in real time, significantly improving the intelligent level and safety, effectively reducing the false detection and missed detection rates, and ensuring the reliability and stability of power grid operation. Brief Description of the Drawings

[0094] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings 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.

[0095] Figure 1 It is a schematic flow chart of an integrated method for transformer inrush current identification and early warning based on multi-sensor fusion proposed by the present invention. Detailed Embodiments

[0096] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the following will describe the detailed embodiments of the present invention with reference to the drawings of the specification.

[0097] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0098] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure or characteristic that can be included in at least one implementation of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or selectively exclusive embodiments from other embodiments.

[0099] Embodiment 1:

[0100] Refer to Figure 1 , which is the first embodiment of the present invention. This embodiment provides an integrated method for transformer inrush current identification and early warning based on multi-sensor fusion, including the following steps:

[0101] S1. Perform distributed deployment of multi-sensors and construction of an adaptive perception network, and collect sensor data.

[0102] In this embodiment, in the power equipment site where the transformer is located, sensors including current, voltage, magnetic field, and temperature are arranged through reasonable distributed deployment. Each sensor node can integrate an adaptive sensing module, which uses an embedded algorithm to dynamically adjust the sampling frequency and resolution to adapt to real-time environmental changes. For example, when a large current fluctuation is detected, the sensor will automatically switch to the high-frequency sampling mode.

[0103] Each node in the adaptive sensing network is interconnected through a low-latency wireless protocol to achieve coordinated operation between sensors, while ensuring data quality, so as to ensure the real-time synchronous operation of sensors under dynamic working conditions.

[0104] S2. Perform multi-dimensional feature correction and time alignment on the collected sensor data.

[0105] In this embodiment, problems such as clock drift, spatial position deviation, and physical parameter scale differences may exist in the data obtained from the distributed sensor network. Multi-dimensional features not only include data at different time steps, but also include spatial and physical parameter differences between sensors. These features may reflect multiple physical information related to inrush phenomena at the same time point. Therefore, a dynamic multi-dimensional feature mapping model is constructed to perform cross-time and cross-dimensional feature alignment on the sensor data, and map the data of different sensors to a unified feature space, laying a standardized data foundation for subsequent inrush identification.

[0106] In the dynamic multi-dimensional feature mapping model, it is assumed that there are sensors, and the recorded data matrix is where , represents the feature dimension recorded by the sensor. Multi-dimensional features are composed of multiple feature dimensions. For example, if a sensor provides both current and temperature data at the same time, then the multi-dimensional features of this sensor include two main dimensions of current and temperature, and there may be multiple sub-dimensions under each dimension.

[0107] In the mapping process of multi-dimensional features, we will correct, standardize, and align each feature dimension to ensure that different dimension data obtained by different sensors can be accurately fused and reflect the complete system state.

[0108] Define the mapped feature as .

[0109] The dynamic multi-dimensional feature mapping model includes:

[0110] Time correction:

[0111] The time alignment is based on interpolation and dynamic weight optimization, and the corrected data is:

[0112] ;

[0113] Among them,

[0114] represents the current time step, which is used to identify the data collected by the sensor at that moment and is used for interpolation calculation during the time correction process;

[0115] represents the sensor clock drift, which is dynamically updated based on the global synchronization mechanism;

[0116] is the correction coefficient, which is adjusted according to the real-time performance of the sensor.

[0117] Spatial deviation correction:

[0118] Based on the linear and non-linear combination model, it is corrected to:

[0119] ;

[0120] Among them,

[0121] is the linear correction matrix of the sensor obtained by least squares estimation;

[0122] is the non-linear adjustment factor, which is optimized through the local deviation model;

[0123] represents the non-linear correction function (such as polynomial fitting or kernel function).

[0124] Scale normalization:

[0125] For the scale difference of multi-dimensional features, the normalization formula is:

[0126] ;

[0127] Among them,

[0128] and are the mean and standard deviation of the data collected by the sensor respectively, which are dynamically calculated based on the sliding window.

[0129] Final mapping:

[0130] The mapped feature is calculated as:

[0131] ;

[0132] Among them,

[0133] is a weight matrix, which is optimized and generated based on mutual information constraints to ensure the maximization of the correlation between features.

[0134] Furthermore, in the dynamic multi-dimensional feature mapping model:

[0135] Time correction : After obtaining the sensor data, first correct the clock drift by dynamically adjusting and to ensure the synchronization of sensor data in the time dimension. For example, if a certain sensor has a large time deviation, its is preferentially adjusted to match the global time reference. In practical applications, and are adjusted using an adaptive feedback mechanism: will be dynamically updated according to the real-time signal-to-noise ratio and data fluctuation of the sensor (reflecting its measurement stability), while is corrected in real time based on the deviation between the global synchronization reference (such as GPS or the master clock) and the local clock of the sensor.

[0136] Spatial deviation correction : After time correction, the data is spatially corrected through a linear correction matrix and a non-linear function to correct the deviation caused by the position or environment of the sensor. For example, subsequently, through the optimization of the correction matrix , the position information of the sensor (such as the installation coordinates) and environmental parameters (such as temperature, electromagnetic interference level) are incorporated into the spatial deviation correction, so as to ensure the effective association and unified compensation between the data after time correction and the spatial correction.

[0137] Scale normalization : After spatial correction, the data is normalized, and the and calculated by the sliding window dynamically adapt to the changes in sensor performance to ensure the data balance in the unified feature space.

[0138] Feature mapping : Finally, the normalized data is feature-mapped through the weight matrix . The weight matrix is generated by mutual information optimization to ensure high feature correlation and low redundancy, providing optimized multi-dimensional input for the subsequent process.

[0139] S3. Extract the fuzzy characteristics of the corrected data to capture the complex non-linear relationships in the transformer operating state.

[0140] In this embodiment, by performing feature decomposition on the corrected data and adopting a fuzzy logic analysis method, a rule-based fuzzy feature library is constructed. The specific steps include:

[0141] The fuzzy clustering algorithm is used to divide the characteristic values ​​of current, voltage and magnetic field into multiple fuzzy sets, including "normal", "critical" and "abnormal".

[0142] Design fuzzy rules between different fuzzy sets to quantify the state transition probability of each feature.

[0143] The state transition probability is used to describe the possibility of transitioning from the current state to the next state. It is calculated based on historical data and real-time observation data. Its quantitative basis mainly includes feature membership, sample statistical distribution, and time series change trend. The degree of transition between different states is measured by fuzzy membership function. In the specific calculation, the frequency of observation data under different states is counted, and fuzzy rules are combined to ensure the rationality and continuity of state transition. The adjustment of state transition probability can be combined with real-time data update to adapt to environmental changes and improve recognition accuracy.

[0144] Fuzzy rules include:

[0145] Definition of fuzzy sets;

[0146] According to the operating status of the transformer, the input characteristics (such as current, voltage, magnetic field strength, etc.) are divided into multiple fuzzy sets. Each fuzzy set is represented in the form of linguistic variables, for example:

[0147] Current: low (L), medium (M), high (H);

[0148] Voltage: normal (N), high (PH), low (PL);

[0149] Magnetic field strength: weak (W), medium (M), strong (S).

[0150] Design of fuzzy rules;

[0151] Fuzzy rules take the form of "if...then..." and determine the relationship between fuzzy sets through logical reasoning. For example:

[0152] Rule 1: If the current is high (H) and the voltage is low (PL), then the transformer status is abnormal (abnormal set).

[0153] Rule 2: If the current is medium (M) and the magnetic field strength is medium (M), then the transformer status is normal (normal set).

[0154] Rule 3: If the voltage is high (PH) and the magnetic field strength is strong (S), then the transformer state is critical (critical set).

[0155] Each rule realizes state inference through fuzzy logic operations (such as the max-min method or the weighted average method).

[0156] In order to dynamically adjust the weights of different features in the fuzzy rules, the entropy weight method is used to quantify the importance of features. A fuzzy weight adjustment mechanism based on the entropy weight method is introduced to make the rule base adaptable to the importance of features.

[0157] In practical applications, the fuzzy rule base will be dynamically adjusted according to the change of weights. For example:

[0158] When the current entropy value is low (the feature is stable), the weight of the current feature decreases, and the rule depends more on the voltage or magnetic field strength.

[0159] When the voltage entropy value increases (the feature fluctuates violently), the weight of the voltage feature increases, and the rule emphasizes more on the influence of voltage.

[0160] Through fuzzy feature decomposition, the complex states during transformer operation can be expressed in the fuzzy space, providing non-linear input features for the subsequent inrush current identification model.

[0161] S4. Fuse the data after extracting fuzzy features and generate a multi-modal fusion feature vector.

[0162] In this embodiment, in step S4, based on the data fusion of the graph neural network GNN, the sensor data is regarded as a graph structure for modeling.

[0163] The sensor nodes are used as the vertices of the graph, and the associations between different modal features are used as the edges of the graph. Through the hierarchical propagation mechanism of GNN, global fusion features are extracted: the original data of each sensor (such as current, voltage, temperature, etc.) are processed by standardization and dimensionality reduction and used as the initial feature embeddings of the nodes to form the input feature matrix. The graph neural network (GNN) is used to perform multi-layer feature propagation on the graph structure data to obtain global fusion features.

[0164] At the same time, the multi-head attention mechanism is used to enhance the weights of key features and endow higher discriminative ability to important parameters. The multi-head attention mechanism is introduced to enhance the weights of key modal features:

[0165] ;

[0166] Each attention head focuses on the importance of different modal features;

[0167] represents the enhanced feature of node .

[0168] Finally, a multi-modal fusion feature vector is generated , which serves as the input for the subsequent inrush current identification model.

[0169] Integrate the spatio-temporal dynamic features of all nodes into a fusion feature vector through a fully connected layer and pooling operations :

[0170] ;

[0171] It can be global max pooling or average pooling, which is used to extract global features.

[0172] denotes the multi-modal fusion feature vector, which contains sensor spatial correlation and temporal dynamic characteristics.

[0173] S5. Build an inrush current identification model, input the multi-modal fusion feature vector, and accurately identify the transformer inrush current.

[0174] In this embodiment, the input of the inrush current identification model is the multi-modal fusion feature vector output in step S4 , dynamically analyze the complex relationships between features, and generate an inrush current classification result;

[0175] The inrush current identification model includes local feature extraction, time dependence modeling, dynamic optimization, and classification decision-making.

[0176] Furthermore, in the inrush current identification model:

[0177] Local feature extraction:

[0178] Convolutional operations extract the local characteristics of the feature vector, and the feature is represented as:

[0179] ;

[0180] Among them,

[0181] is the feature matrix after the -th layer of convolution;

[0182] and are the convolutional kernel and bias term respectively;

[0183] is a non-linear activation function (such as ReLU).

[0184] Time dependence modeling:

[0185] Perform time modeling on the feature sequence extracted by convolution:

[0186] ;

[0187] Among them,

[0188] is the hidden layer state at time ;

[0189] represents the feature matrix extracted by convolution as the input and the feature vector generated by convolution, realizing the temporal dynamic modeling of the convolution feature sequence;

[0190] , , are respectively the mapping weight of , the mapping weight of the previous moment, and the overall bias term;

[0191] is the activation function;

[0192] Dynamic optimization:

[0193] Introduce reinforcement learning to dynamically adjust the classification threshold , and optimize the reward function:

[0194] ;

[0195] Among them,

[0196] is the reward value;

[0197] represents the model accuracy rate, which is the proportion of correctly classified samples in the total samples;

[0198] and are respectively the false positive rate and the false negative rate, reflecting the proportion of negative samples misjudged as positive samples, that is, the probability of false alarm, reflecting the proportion of positive samples misjudged as negative samples, that is, the probability of missed alarm;

[0199] is the penalty coefficient.

[0200] The classification threshold functions to determine the decision boundary when the model performs inrush identification, affecting the false positive rate (FP) and the false negative rate (FN). In the reinforcement learning module, the classification threshold is not fixed, but dynamically adjusted. By optimizing the reward function R, it maximizes the accuracy rate A while minimizing false alarms (FP) and missed alarms (FN). When the false alarm rate is too high, appropriately increase the threshold to reduce false alarms; when the missed alarm rate is too high, decrease To improve sensitivity. In this way, the classification threshold can be adaptively adjusted in different operating environments to ensure the reliability and stability of recognition.

[0201] Classification decision:

[0202] Based on the output of local feature extraction and time-dependent modeling, classification is performed:

[0203] ;

[0204] Among them,

[0205] is the classification result, including "normal inrush current" and "abnormal inrush current";

[0206] is the hidden state vector at the final time step T, representing the feature extraction and memory results of the model over the entire time series. It synthesizes the information of all historical time steps and is used as the input for the final classification decision. After being transformed by the fully connected layer, it is used to determine the inrush current category.

[0207] , are the weights and biases of the output layer.

[0208] When this model is used, it includes:

[0209] Local feature extraction (CNN): Multi-modal fusion features Through the convolutional kernel local spatial patterns are extracted. It is particularly important for the recognition of inrush current characteristics (such as amplitude changes and specific frequency band responses). Nonlinear activation emphasizes key features and eliminates redundant information.

[0210] Time-dependent modeling (LSTM): The feature sequence output by CNN is passed to the LSTM unit to capture long-term time-dependent relationships, such as the dynamic changes of the inrush current waveform. Through time series modeling, it is possible to better distinguish transient inrush current and steady-state inrush current.

[0211] Dynamic optimization: Involves the dynamic optimization of the classification threshold Reinforcement learning adjusts according to the real-time classification results to ensure adaptability under different working conditions. For example, when the environmental interference is large, the reward function will reduce the weight of the false positive rate to improve accuracy.

[0212] Classification decision: The joint output of CNN and LSTM is calculated by the fully connected layer to obtain the classification probability, and the inrush current state is determined based on the maximum probability value.

[0213] S6. Analyze the precursor characteristics of the abnormal state based on the inrush current identification model and conduct real-time warning.

[0214] In this embodiment, the real-time warning includes:

[0215] Precursor feature extraction;

[0216] Use the method combining time domain and frequency domain to extract precursor features :

[0217] ;

[0218] Where:

[0219] is the precursor feature vector;

[0220] is the precursor weight matrix;

[0221] is the multi-modal fusion feature of the sensor at the current moment;

[0222] represents the frequency domain feature extracted by fast Fourier transform;

[0223] is the frequency domain feature adjustment factor.

[0224] Through generate the multi-modal feature of the current sensor, and combine with FFT to obtain the frequency domain feature, forming the precursor feature vector . This combination method can capture the weak feature changes before the inrush current occurs, such as the periodic fluctuation of the current amplitude.

[0225] The multi-modal feature refers to the multi-dimensional data collected from multiple different sensors. Through feature extraction and fusion methods, it constitutes the multi-modal fusion feature for inrush current identification. These features come from different types of sensors, such as current sensors, temperature sensors, vibration sensors, etc. Each sensor provides data of different modalities, and after fusion, the accuracy and robustness of inrush current identification can be improved. It mainly includes:

[0226] Time domain features: such as the mean, variance, skewness, peak value, rise time of the current waveform, etc.;

[0227] Frequency domain features: such as harmonic components, main frequency, spectral energy distribution, etc.;

[0228] Spatial features: the signal correlation between different sensors, such as the current ratio of different measurement points at the same time;

[0229] Time series features: Use RNN / LSTM for time modeling to capture the patterns of signal changes over time.

[0230] Furthermore, frequency domain features refer to the performance of the signal in the frequency domain. It reflects the periodicity and harmonic components of the signal and can reveal the change trend of specific frequency components when inrush current occurs. It mainly includes:

[0231] Main frequency component: The main frequency component of the signal;

[0232] Harmonic ratio: The energy ratio of the fundamental wave to the higher harmonics;

[0233] Spectrum energy distribution: The energy ratio of different frequency bands, such as 0–50Hz, 50–150Hz, 150–500Hz, etc.;

[0234] Spectrum entropy: Measures the complexity of the signal spectrum. The higher the entropy value, the more complex the signal frequency distribution.

[0235] Frequency domain features use the fast Fourier transform (FFT) to convert the time-domain signal into a frequency-domain signal and calculate the above frequency domain features.

[0236] Dynamic threshold calculation;

[0237] Threshold Dynamic adjustment, the expression is:

[0238] ;

[0239] Where:

[0240] Is the initial threshold;

[0241] Is the dynamic adjustment coefficient;

[0242] And Are the mean and standard deviation of the precursor features (calculated by a dynamic sliding window).

[0243] During the warning process, the system performs standardized calculation on the volatility of the precursor features (based on the mean and standard deviation of the sliding window). The dynamically adjusted threshold Can adapt to different operating conditions, such as load fluctuations or environmental impacts.

[0244] Fault probability inference;

[0245] Based on Bayes' formula, calculate the real-time fault probability :

[0246] ;

[0247] Where:

[0248] is the probability of observing precursor characteristics in a fault state ;

[0249] is the prior fault probability;

[0250] is the total probability of the characteristics.

[0251] Using Bayesian inference method combined with real-time data to calculate the probability of fault occurrence . In the case where the inrush precursor characteristics are obvious, this value will increase rapidly, enhancing the sensitivity of the model.

[0252] Real-time early warning decision-making;

[0253] Early warning signal Output rules of:

[0254] ;

[0255] Among them:

[0256] indicates triggering an early warning;

[0257] indicates the normal state.

[0258] According to the fault probability and the threshold for comparison, output a real-time early warning signal. If , then trigger the inrush early warning mechanism and transmit the signal to execute subsequent protection actions.

[0259] Embodiment 2:

[0260] This embodiment also provides a computer device, applicable to a situation of an integrated method for inrush current identification and early warning of a transformer based on multi-sensor fusion, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement an integrated method for inrush current identification and early warning of a transformer based on multi-sensor fusion as proposed in the above embodiment.

[0261] This embodiment also provides a storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements an integrated method for inrush current identification and early warning of a transformer based on multi-sensor fusion as proposed in the above embodiment.

[0262] The computer device may be a terminal, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner, and the wireless manner can be implemented through WIFI, carrier networks, NFC (Near Field Communication), or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads provided on the housing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0263] If a function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs, etc., which can store program codes.

[0264] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or in combination with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device.

[0265] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection (electronic device) having one or more wirings, a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which a program can be printed, as the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing it as appropriate, and then storing it in a computer memory.

[0266] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or a combination of the following techniques known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), and the like.

[0267] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. An integrated method for transformer inrush current identification and warning based on multi-sensor fusion, characterized in that, It includes the following steps: S1. Conduct distributed deployment of multi-sensors and construction of an adaptive perception network, and collect sensor data; S2. Perform multi-dimensional feature correction and time alignment on the collected sensor data; S3. Extract fuzzy characteristics from the corrected data to capture complex non-linear relationships in the transformer operating state; S4. Fuse the data after extracting fuzzy characteristics and generate a multi-modal fusion feature vector; S5. Construct an inrush current identification model, input the multi-modal fusion feature vector, and accurately identify the transformer inrush current; S6. Based on the inrush current identified by the inrush current identification model, analyze the precursor characteristics of abnormal states and conduct real-time early warning.

2. The integrated method for transformer inrush current identification and early warning based on multi-sensor fusion according to claim 1, characterized in that, In step S1, in the power equipment site where the transformer is located, sensors including current, voltage, magnetic field, and temperature are arranged based on distributed deployment; each node in the adaptive perception network is interconnected through a low-latency wireless protocol to ensure real-time synchronous operation of the sensors under dynamic working conditions.

3. The integrated method for transformer inrush current identification and warning based on multi-sensor fusion according to claim 1, wherein In step S2, a dynamic multi-dimensional feature mapping model is constructed to perform cross-time and cross-dimension feature alignment on the sensor data; In the dynamic multi-dimensional feature mapping model, it is assumed that there are sensors, and the recorded data matrix is , where , represents the feature dimension recorded by the sensor; Define the mapped feature as .

4. The integrated method for transformer inrush current identification and early warning based on multi-sensor fusion according to claim 3, wherein, The dynamic multi-dimensional feature mapping model includes: Time correction: Time alignment is based on interpolation and dynamic weight optimization, and the corrected data is: ； where, Indicates the clock drift of the sensor and is dynamically updated based on the global synchronization mechanism; is a correction coefficient, adjusted according to the real-time performance of the sensor; Spatial deviation correction: Based on a linear and non-linear combined model, the correction is: ； where, is the linear correction matrix of the sensor obtained by least squares estimation; is a non-linear adjustment factor optimized by a local deviation model; represents a non-linear correction function; Scale normalization: For the scale differences of multi-dimensional features, the normalization formula is: ； where, and are the mean and standard deviation of the data collected by the sensor respectively, and are calculated dynamically based on a sliding window; Final mapping: Mapped features Calculated as: ； where, is a weight matrix, which is optimized and generated based on mutual information constraints to ensure the maximization of the correlation between features.

5. The integrated method for transformer inrush current identification and warning based on multi-sensor fusion according to claim 4, characterized in that, In the dynamic multi-dimensional feature mapping model: Time Calibration : After obtaining sensor data, first correct the clock drift by dynamically adjusting and to ensure that the sensor data is synchronized in the time dimension; Spatial deviation correction : Perform spatial correction on the data after time correction through a linear correction matrix and a non-linear function to correct the deviation caused by the position or environmental factors of the sensor; Scale normalization : After spatial correction, the data is normalized, and the and dynamically adapts to changes in sensor performance to ensure data balance in a unified feature space; Feature mapping : Finally, the normalized data is subjected to feature mapping through the weight matrix generated by optimizing mutual information, ensuring high feature correlation and low redundancy, and providing optimized multi-dimensional input for subsequent processing.

6. The integrated method for transformer inrush current identification and warning based on multi-sensor fusion according to claim 1, characterized in that, In step S3, through feature decomposition of the corrected data and using the fuzzy logic analysis method, a rule-based fuzzy characteristic library is constructed. The specific steps include: Using the fuzzy clustering algorithm to divide the characteristic values of current, voltage, and magnetic field into multiple fuzzy sets, including "normal", "critical", and "abnormal"; Design fuzzy rules between different fuzzy sets to quantify the state transition probabilities of each feature; Introduce a fuzzy weight adjustment mechanism based on the entropy weight method to make the rule library adaptable to the importance of features.

7. The integrated method for transformer inrush current identification and early warning based on multi-sensor fusion according to claim 1, characterized in that In step S4, based on the data fusion of the graph neural network GNN, the sensor data is regarded as a graph structure for modeling; The sensor nodes are used as the vertices of the graph, and the associations between different modal features are used as the edges of the graph. Through the hierarchical propagation mechanism of GNN, global fusion features are extracted; At the same time, use the multi-head attention mechanism to enhance the weights of key features and give higher discriminative ability to important parameters; Finally, generate a multi-modal fusion feature vector as the input for the subsequent inrush current identification model.

8. The integrated method for transformer inrush current identification and early warning based on multi-sensor fusion according to claim 1, characterized in that In the step S5, the input of the inrush current identification model is the multi-modal fusion feature vector output by the step S4 , dynamically analyze the complex relationships between features, and generate an inrush current classification result; The inrush current identification model includes local feature extraction, time-dependent modeling, dynamic optimization, and classification decision-making.

9. The integrated method for transformer inrush current identification and early warning based on multi-sensor fusion according to claim 8, characterized in that In the inrush current identification model: Local feature extraction: Convolutional operations extract the local characteristics of the feature vector, and the feature representation is: ； where, is the feature matrix after the -th layer of convolution; and are the convolution kernel and the bias term, respectively; is a non-linear activation function; Time-dependent modeling: Perform time modeling on the feature sequence extracted by convolution: ； where, is the hidden layer state at time ; Represents the feature matrix extracted by convolution As the input, and the generated feature vector through convolution is used to realize the temporal dynamic modeling of the convolutional feature sequence; , , are respectively the mapping weight, the mapping weight at the previous moment, and the overall bias term; is an activation function; Dynamic optimization: Introduce reinforcement learning to dynamically adjust the classification threshold , and optimize the reward function: ； where, is the reward value; Indicates the model accuracy, which is the proportion of correctly classified samples to the total samples; and are the false positive rate and false negative rate respectively, reflecting the proportion of negative samples misjudged as positive samples, that is, the probability of false alarm, reflecting the proportion of positive samples misjudged as negative samples, that is, the probability of missed alarm; is the penalty coefficient; Classification decision-making: Based on the outputs of local feature extraction and time-dependent modeling, classification is performed: ； where, For classification results, including "normal inrush current" and "abnormal inrush current"; , are the weights and biases of the output layer.

10. The integrated method for transformer inrush current identification and warning based on multi-sensor fusion according to claim 1, characterized in that, In step S6, the real-time early warning includes: Precursor feature extraction; Dynamic threshold calculation; Fault probability inference; Real-time early warning decision-making.

Images