Internal resistance big data driven security operation and maintenance method and system

Abstract

The invention discloses a security operation and maintenance method and system driven by internal resistance big data, and the method comprises the steps: collecting the historical and real-time internal resistance data of a target device, and generating a structured internal resistance time sequence data set through a dynamic time warping algorithm; performing multi-scale feature extraction and high-dimensional space mapping on the internal resistance time sequence data set, and identifying an internal resistance abnormal mode cluster implied in data distribution by adopting an adaptive density clustering algorithm; performing multi-mode fault correlation analysis based on the internal resistance abnormal mode cluster to generate a fault prediction map; and generating a differentiated operation and maintenance strategy set according to the fault prediction map, and driving the operation and maintenance management platform to execute a corresponding operation and maintenance instruction. According to the embodiment of the invention, the timeliness and accuracy of fault early warning can be improved, the operation and maintenance cost is reduced, and safe and stable operation of equipment is guaranteed.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of operation and maintenance, in particular to a safety operation and maintenance method and system driven by internal resistance big data. BACKGROUND

[0002] In key fields such as power, communication and new energy batteries, the reliability and safety operation of equipment are crucial. As one of the core parameters representing the health status of equipment, the change trend of internal resistance is closely related to potential faults. Existing operation and maintenance methods mostly rely on threshold monitoring of conventional parameters such as voltage, current and temperature, or simple threshold judgment and trend analysis of internal resistance data. Such methods are difficult to effectively mine the deep evolution rules and early abnormal signs of internal resistance data in the time dimension, and lack the ability to predict complex and gradual faults. At the same time, traditional methods usually treat internal resistance data in isolation and fail to deeply integrate and analyze it with multi-dimensional operating conditions of equipment, resulting in high false and missed alarm rates, and the inability to effectively evaluate the evolution path and risk level of faults, which makes operation and maintenance decisions lack foresight and precision, and the efficiency of operation and maintenance resource allocation is low. SUMMARY

[0003] The purpose of the present application is to provide a safety operation and maintenance method and system driven by internal resistance big data, to solve the problems in the prior art, and to improve the timeliness and accuracy of fault early warning, reduce operation and maintenance costs and ensure the safe and stable operation of equipment.

[0004] One embodiment of the present application provides a safety operation and maintenance method driven by internal resistance big data, which comprises: Collecting historical and real-time internal resistance data of a target device, and combining multi-dimensional operating state parameters, generating a set of internal resistance time series data after normalization by a dynamic time warping algorithm; Performing multi-scale feature extraction and high-dimensional space mapping on the set of internal resistance time series data, constructing a set of internal resistance feature vectors, and identifying internal resistance abnormal pattern clusters hidden in the data distribution using an adaptive density clustering algorithm; Performing multi-modal fault correlation analysis based on the internal resistance abnormal pattern clusters, predicting the evolution path and risk level of potential faults in combination with the operating history and operating condition labels of the equipment, and generating a fault prediction map; Generating a set of differentiated operation and maintenance strategies according to the fault prediction map, and driving an operation and maintenance management platform to execute corresponding operation and maintenance instructions.

[0005] Another embodiment of the present application provides a safety operation and maintenance system driven by internal resistance big data, which comprises: A collection module for collecting historical and real-time internal resistance data of a target device, and combining multi-dimensional operating state parameters, generating a set of internal resistance time series data after normalization by a dynamic time warping algorithm; The construction module is configured to perform multi-scale feature extraction and high-dimensional space mapping on the internal resistance time series data set, construct an internal resistance feature vector set, and identify an internal resistance abnormal pattern cluster hidden in the data distribution by using an adaptive density clustering algorithm; The analysis module is configured to perform multi-modal fault correlation analysis based on the internal resistance abnormal pattern cluster, combine device operation history and working condition labels to predict an evolution path and a risk level of a potential fault, and generate a fault prediction graph; The operation and maintenance module is configured to generate a set of differential operation and maintenance strategies according to the fault prediction graph, and drive an operation and maintenance management platform to execute corresponding operation and maintenance instructions.

[0006] Another embodiment of the present application provides a storage medium having a computer program stored therein, wherein the computer program is configured to execute the method described in any one of the above embodiments when running.

[0007] Another embodiment of the present application provides an electronic device comprising a memory and a processor, wherein the memory has a computer program stored therein, and the processor is configured to execute the computer program to perform the method described in any one of the above embodiments.

[0008] Compared with the prior art, the safe operation and maintenance method driven by internal resistance big data provided by the present application can improve the timeliness and accuracy of fault early warning, reduce operation and maintenance costs, and ensure safe and stable operation of the device. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 A hardware structure block diagram of a computer terminal of the safe operation and maintenance method driven by internal resistance big data provided by the embodiment of the present application is shown in the figure. Figure 2 A flowchart of the safe operation and maintenance method driven by internal resistance big data provided by the embodiment of the present application is shown in the figure. Figure 3 A structure diagram of the safe operation and maintenance system driven by internal resistance big data provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0010] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be explained as a limitation of the present application.

[0011] Figure 1 A hardware structure block diagram of a computer terminal of the safe operation and maintenance method driven by internal resistance big data provided by the embodiment of the present application is shown in the figure. Figure 1 As shown in the figure, the computer device comprises a processor, a memory and a network interface connected through a system bus, wherein the memory can comprise a non-volatile storage medium and an internal memory.

[0012] Referring to Figure 2Embodiments of the present application provide a large data-driven internal resistance safety operation method, which can include the following steps: S201, collecting historical and real-time internal resistance data of the target device, and combining multi-dimensional running state parameters to generate a set of normalized internal resistance time series data through a dynamic time warping algorithm; Specifically, the internal resistance data of the target device can be collected in real time through a distributed data collection node, and at least multi-dimensional running state parameters including voltage, current and temperature are collected to generate a multi-source original monitoring data set; First of all, the hardware configuration and deployment logic of the "distributed data collection node" are determined, and the node needs to be deployed at the key position of the target device (for example, a 10kV distribution transformer, which is a common power equipment that needs to be monitored) to ensure that the data covers the core state of the equipment operation. The node selects an edge computing collection terminal (model NIcDAQ-9178, supporting 8 modular input channels, working temperature -40~70℃, suitable for harsh industrial field environment), a total of 3 nodes are deployed: No. 1 node is installed at the high-voltage side of the transformer terminal, responsible for collecting internal resistance and voltage data; No. 2 node is installed in the low-voltage side of the outgoing line cabinet, collecting current data; No. 3 node is installed at the top of the transformer oil tank, collecting temperature data, and the nodes are connected through industrial Ethernet (PROFINET protocol, transmission rate 100Mbps, delay ≤10ms) to realize synchronous data transmission.

[0013] The internal resistance data is collected by a high-precision direct current internal resistance tester, and the collection principle is "four-wire measurement method" (eliminating the influence of wire resistance on measurement), and the collection frequency is set to 1 time / minute. Because the internal resistance of the transformer changes slowly with the load and temperature (the daily change is ≤0.5mΩ), too high frequency will cause data redundancy, and 1 time / minute can capture the change rule and control the data volume (1440 pieces per day).

[0014] The multi-dimensional running state parameter collection equipment and parameters are as follows: the voltage is collected by a Hall voltage sensor, the collection frequency is 1 time / second (the voltage is greatly affected by the power grid fluctuation, and needs to be captured at high frequency); the current is collected by a closed-loop Hall current sensor (model ACS758-200A, range 0~200A, accuracy ±0.3%FS, error ≤0.6A), the collection frequency is 1 time / second (synchronous with voltage, convenient for subsequent power calculation); the temperature is collected by a PT100 platinum resistance sensor (model WZP-230, measurement range -200~800℃, accuracy ±0.1℃), the collection frequency is 1 time / 10 seconds (the temperature change lags behind the voltage and current, and the 10-second interval can balance the accuracy and efficiency).

[0015] The multi-source raw monitoring data set is stored in the format of "timestamp (format YYYYMMDDHHMMSSfff) + device ID + internal resistance (unit mΩ) + voltage (unit V) + current (unit A) + temperature (unit ℃)", and the example data is as follows: "20251001080000123, TR-001, 150.2mΩ, 380.5V, 120.3A, 45.2℃" "20251001080001456, TR-001, -, 380.4V, 120.1A, 45.3℃" (where "-" indicates that the internal resistance data at this moment has not been collected, and there is a missing value), and the data set is uploaded to the local database (MySQL, storage capacity 1TB, retaining 30 days of raw data) of the edge node in real time, forming a multi-source raw monitoring data set.

[0016] The multi-source raw monitoring data set is processed for data cleaning and outlier processing, and a sliding window filtering algorithm is used to smooth the data fluctuations to generate a pre-processed multi-dimensional monitoring data set. Data cleaning needs to solve the problems of "missing values" and "outliers", and ensure data integrity and accuracy. Missing value processing adopts "classification filling strategy": internal resistance data missing (such as the internal resistance of 08:00:01 in the above example is "-"), because the internal resistance changes continuously, the "forward filling method" is used - fill with 150.2mΩ at the previous valid time (08:00:00); voltage and current data missing (such as the current is "-" due to sensor transient failure at a certain time), adopt "linear interpolation method" - take each 3 valid data points before and after the missing time (such as 120.3A at 08:00:00 and 119.9A at 08:00:02), calculate the missing value (current at 08:00:01 ≈ 120.1A) through linear fitting; temperature data missing, adopt "mean filling method" - take the mean temperature of the same time period (around 08:00) in the past 7 days (such as the mean temperature of 08:00 in the past 7 days is 45.1℃) to fill, and the missing data rate after filling is ≤0.1%.

[0017] The abnormal value processing adopts "physical threshold method + 3 sigma principle" double judgment: the physical threshold is set based on the rated parameters of the equipment, the normal operating range of the transformer resistance is 140-160 mΩ (lower than 140 mΩ may be due to loose wiring, higher than 160 mΩ may be due to winding aging), the normal voltage range is 360-400 V (exceeding is power grid overvoltage / undervoltage), the normal current range is 100-150 A (exceeding is overload), and the normal temperature range is 40-60°C (exceeding is heat dissipation failure), and exceeding the range is marked as abnormal; the 3 sigma principle is for data within the physical threshold, the mean value μ and the standard deviation σ of nearly 1000 valid data of a parameter are calculated, and if the data point deviates from μ by more than 3σ (|x-μ|>3σ), it is determined as random noise abnormality. Example: the resistance of nearly 1000 data μ=150 mΩ, σ=2 mΩ, 3σ=6 mΩ, a data point 157 mΩ (although it is within 140-160 mΩ, but 157-150=7>6), it is determined as abnormal, and the "adjacent mean value replacement method" is adopted-the mean value of the two valid data before and after the data point (such as 150.4 mΩ, 150.2 mΩ, 149.9 mΩ, 150.1 mΩ, the mean value is 150.15 mΩ) is replaced.

[0018] The sliding window filtering algorithm is used to smooth the data fluctuations (such as voltage transient jump caused by electromagnetic interference), and the selection of window size N needs to be combined with the data acquisition frequency: the resistance data acquisition frequency is 1 time / minute, N=5 (the window covers 5 minutes, that is, 5 data points) is selected to balance the smoothing effect and lag; the voltage and current acquisition frequency is 1 time / second, N=10 (the window covers 10 seconds) is selected; the temperature acquisition frequency is 1 time / 10 seconds, N=3 (the window covers 30 seconds) is selected. The filtering formula is "x_filtered(t)=(x(t)+x(t-1)+…+x(t-N+1)) / N", wherein x(t) is the original data at time t, and x_filtered(t) is the filtered data.

[0019] Based on the pre-processed multi-dimensional monitoring data set, the dynamic time warping algorithm is used to align the resistance data and operating state parameters with different sampling frequencies, eliminate the time scale difference, and generate a time-aligned multi-dimensional data sequence; The difference in sampling frequency of different parameters will cause "time scale misalignment" (such as resistance 1 time / minute, voltage 1 time / second, 1 minute resistance only 1 data, voltage has 60 data), which cannot be directly associated and analyzed, and needs to be aligned through dynamic time warping (Dynamic Time Warping, DTW) algorithm. The core principle of DTW is to find a "minimum cumulative distance" path by constructing a distance matrix of two time series, mapping sequences of different lengths and different frequencies to the same time scale, and not destroying the time order of the sequence.

[0020] Firstly, determine the sequence pairs to be aligned: take the "internal resistance sequence" as the reference sequence (low sampling frequency, as the time anchor point), and take the "voltage sequence", "current sequence", and "temperature sequence" as the sequences to be aligned (high sampling frequency). For example: take the preprocessed data within 10 minutes, the internal resistance sequence T = [t0, t1, …, t9] (10 data points, interval 1 minute), the voltage sequence V = [v0, v1, …, v599] (600 data points, interval 1 second), and align the V sequence to the 10 time points of the T sequence.

[0021] The DTW algorithm performs the following steps: first, construct the distance matrix D, the matrix element d(i, j) is the Euclidean distance between the i-th data of the internal resistance sequence (r_i) and the j-th data of the voltage sequence (v_j), d(i, j) = |r_i-v_j| (since the units are different, the original numerical value is calculated first, and then normalized), in the example d(0, 0) = |150.2-380.5| = 230.3, d(0, 1) = |150.2-380.4| = 230.2; second, calculate the cumulative distance matrix D_acc, D_acc(i, j) = d(i, j) + min(D_acc(i-1, j), D_acc(i, j-1), D_acc(i-1, j-1)), the initial value D_acc(0, 0) = d(0, 0) = 230.3; third, trace back from the lower right corner of D_acc (D_acc(9, 599)) to the upper left corner (D_acc(0, 0)), find the minimum cumulative distance path, the voltage data on this path is the optimal data aligned with the internal resistance data; fourth, for each internal resistance time point t_i, take the arithmetic mean of all voltage data on the path as the aligned voltage value, in the example t0 (08:00) corresponds to v0~v59 (60 voltage data), the mean ≈ 380.48V, which is the aligned voltage value at t0.

[0022] Similarly, the current sequence (600 data points) and the temperature sequence (60 data points, 1 / 10 seconds) are aligned according to the same logic: the current sequence corresponds to the mean of 60 data at each internal resistance time point, and the temperature sequence corresponds to the mean of 6 data at each internal resistance time point. The final time-aligned multi-dimensional data sequence contains "internal resistance + aligned voltage + aligned current + aligned temperature" at each time point (interval 1 minute), for example: "20251001080000123, TR-001, 150.2mΩ, 380.48V, 120.22A, 45.23℃" "20251001080100789, TR-001, 150.3mΩ, 380.52V, 120.18A, 45.31℃", the time scales of each parameter in the sequence are completely consistent (all 1 / minute), eliminating the correlation deviation caused by the difference in sampling frequency.

[0023] The normalized multi-dimensional data sequence is time-aligned, and the internal resistance data and each operating state parameter are unified to the same dimension, and finally a normalized internal resistance time series data set is generated.

[0024] The time-aligned multi-dimensional data has a "dimension difference" (such as internal resistance unit mΩ, numerical range 140-160; voltage unit V, numerical range 360-400; current unit A, numerical range 100-150; temperature unit ℃, numerical range 40-60), which will cause "numerical large parameter weight too high" (such as voltage 380V, which is much larger than internal resistance 150mΩ, which will mask the influence of internal resistance change) during subsequent feature extraction, and normalization processing is required to map all parameters to the [0, 1] interval and unify the dimension.

[0025] The Min-Max normalization method is selected, and the normalized calculation example is as follows: taking the aligned data point (internal resistance 150.2mΩ, voltage 380.48V, current 120.22A, temperature 45.23℃): Internal resistance normalization: (150.2-140) / (160-140)=10.2 / 20=0.51; voltage normalization: (380.48-360) / (400-360)=0.512; current normalization: (120.22-100) / (150-100)=0.4044; temperature normalization: (45.23-40) / (60-40)=0.2615.

[0026] The normalized internal resistance time series data set is stored in the format of "timestamp+device ID+normalized internal resistance+normalized voltage+normalized current+normalized temperature", and the data set needs to meet the following requirements: the data quantity is consistent with the aligned sequence (1440 per day on average), the mean value of each parameter after normalization is around 0.5 (to avoid data bias), the standard deviation is between 0.1 and 0.2 (to retain the change characteristics of the parameter), and it can be directly used for subsequent multi-scale feature extraction and abnormal pattern recognition.

[0027] S202, multi-scale feature extraction and high-dimensional space mapping are performed on the internal resistance time series data set, an internal resistance feature vector set is constructed, and an adaptive density clustering algorithm is used to identify internal resistance abnormal pattern clusters hidden in the data distribution; Specifically, multi-scale wavelet transform can be performed on the normalized internal resistance time series data set to extract time domain and frequency domain features and generate multi-scale feature vectors. First of all, make clear the specific form of "the regularized internal resistance time series data set" - this set is sampled at 1-minute intervals, each data contains "normalized resistance (such as 0.51), normalized voltage (such as 0.512), normalized current (such as 0.4044), normalized temperature (such as 0.2615)", and the data volume is taken as an example of 24-hour running data of a target device (10kV distribution transformer TR-001), a total of 1440, which needs to be captured through multiscale wavelet transform (Multiscale Wavelet Transform) to capture the feature differences at different time scales. Its core advantage is to balance "time domain locality" and "frequency domain globality", which can extract both the slow trend of the device internal resistance over time (such as the daily drift caused by temperature) and the instantaneous fluctuation (such as the short-term change caused by current surge).

[0028] The db4 wavelet basis (Daubechies-4 wavelet) is selected as the transformation core, because db4 wavelet has 8 support points, balancing between smoothness and computational efficiency - too few support points will lead to rough feature extraction, and too many will increase the computational load; The number of decomposition layers is set to 5, each layer corresponds to a different time scale: 1 layer decomposition corresponds to 2 minutes scale (capture 1-2 minutes of instantaneous fluctuation, such as internal resistance change caused by power grid load surge), 2 layers correspond to 4 minutes scale (capture 2-4 minutes of medium-term change), 3 layers correspond to 8 minutes scale, 4 layers correspond to 16 minutes scale, 5 layers correspond to 32 minutes scale (capture 32 minutes or more long-term trend, such as internal resistance drift caused by slow rise of oil temperature), 5 layers of decomposition can cover the main time dimension of device operation, avoiding missing key features.

[0029] The execution process of multiscale wavelet transform needs to be divided into "decomposition-feature extraction" two steps: first, input the 1440 regularized data into the wavelet transform module in time order, and perform 5-layer tower decomposition through db4 wavelet basis, to get the approximate coefficients (A1-A5, reflecting low-frequency trend) and detail coefficients (D1-D5, reflecting high-frequency fluctuation) of each layer; The second step is to extract time domain and frequency domain features based on the coefficients, and the features need to cover "trend, fluctuation, energy" three types of core information to ensure comprehensiveness.

[0030] Time domain feature extraction is aimed at each layer of coefficients, and 5 types of indicators are selected, such as "mean, variance, peak, valley, kurtosis", the calculation example is as follows (take A5 layer approximate coefficient as an example, this layer corresponds to 32 minutes long-term trend, there are 45 coefficient values in this layer, ranging from 0.48 to 0.55): Mean (μ): reflects the trend center level, the calculation formula is μ=(1 / N)×Σx_i (N=45), substituting the data gives μ=(0.48+0.51+...+0.55) / 45≈0.52, which represents the long-term mean of internal resistance at this scale close to the median of the normal interval; Variance (σ²): Reflects the dispersion degree of the trend, the calculation formula is σ²=(1 / N)×Σ(x_i-μ)², and the calculation result is σ²≈((0.48-0.52)²+...+(0.55-0.52)²) / 45≈0.0008, which indicates that the long-term trend is stable and the fluctuation is small; Peak (x_max) and valley (x_min): Reflect the extreme value of the trend, the peak value of A5 layer is 0.55, the valley value is 0.48, and the difference is 0.07, which does not exceed the normal fluctuation range (0.1); Kurtosis (K): Reflects the steepness of data distribution, the calculation formula is K=[(1 / N)×Σ(x_i-μ)^4] / σ^4, and the calculation result is K≈2.9, which is close to normal distribution (K=3) and has no extreme abnormal value.

[0031] Frequency domain features are based on the Fourier transform results of the detail coefficients of each layer, and two types of indexes, "scale energy and wavelet entropy", are selected. For example, the detail coefficients of D1 layer are taken as an example, corresponding to 2 minutes of high-frequency fluctuation, a total of 720 coefficients: Scale energy (E): Reflects the energy intensity of the fluctuation at this scale, the calculation formula is E=Σ|x_i|², and the calculation result is E≈0.09 (the energy of high-frequency noise is low, which is consistent with the characteristics of stable operation of the device); Wavelet entropy (S): Reflects the uniformity of energy distribution at each scale, the calculation formula is S=-Σ(p_i×lnp_i) (p_i is the proportion of the energy of a certain scale to the total energy), the total energy is the sum of the energies of each layer≈1.62, p_i of D1 layer=0.09 / 1.62≈0.056, p_i of A5 layer=1.25 / 1.62≈0.772, and the calculation result is S≈- (0.772×ln0.772+0.056×ln0.056+...)≈0.41, which indicates that the energy is concentrated in the low-frequency scale (long-term trend) and there is no abnormal energy mutation.

[0032] Finally, for each layer of 5-layer wavelet decomposition, 7 types of features, "mean, variance, peak, valley, kurtosis, energy, and entropy", are extracted, a total of 5×7=35 features; At the same time, considering the correlation between multiple parameters, two cross-parameter features, "internal resistance-temperature correlation coefficient (such as 0.68) and voltage-current correlation coefficient (such as 0.82)", are additionally calculated, totaling 37 features, which constitute a multi-scale feature vector, such as "[0.52, 0.0008, 0.55, 0.48, 2.9, 1.25, 0.41, 0.68, 0.82,...]", the first 7 features correspond to A5 layer, and the subsequent features correspond to A4-D1 layer and cross-parameter features in turn, which completely covers the multi-scale feature dimension of the device operation.

[0033] The multi-scale feature vector is mapped to a high-dimensional space using a self-encoder, and a high-dimensional feature representation vector is generated by extracting deep feature representations through the encoder; An autoencoder is an unsupervised deep learning model, whose core function is to learn deep abstract features of data through an encoder-decoder structure. In this case, the 37-dimensional multi-scale feature vector is mapped to a 64-dimensional high-dimensional space, aiming to amplify subtle feature differences (such as the difference between normal patterns and slightly aging patterns) and provide stronger discrimination for subsequent clustering.

[0034] The network structure of the autoencoder needs to match the feature dimension and the mapping target. The specific design is as follows: Input layer (InputLayer): The number of neurons is set to 37, consistent with the dimension of the multi-scale feature vector. The activation function uses Linear (linear activation) to avoid the initial features being compressed by nonlinearity and ensure the integrity of the original feature information. Encoder (Encoder): It is divided into three hidden layers. The first layer (H1) has 128 neurons (to expand the input features and capture low-order correlations), and the activation function is ReLU (formula f(x) = max(0, x), which alleviates the gradient vanishing problem and is suitable for non-linear feature learning of industrial data). The second layer (H2) has 64 neurons (target high-dimensional dimension, realizing high-dimensional mapping), and the activation function is still ReLU. The third layer (H3) has 64 neurons (to strengthen the stability of deep features), and the activation function is Tanh (mapping feature values to the interval [-1, 1] to reduce the influence of numerical fluctuations). Decoder (Decoder): It is symmetrical with the encoder, with H4 (128 neurons, ReLU) and H5 (37 neurons, Linear). The reconstruction error is used to optimize the feature extraction ability of the encoder. The reconstruction error uses Mean Squared Error (MSE, formula MSE = (1 / N) x Σ(y_i - y_i)², y_i is the input feature, and y_i is the reconstructed feature output by the decoder). Training parameters: The optimizer uses Adam (adaptive moment estimation, faster convergence than SGD, suitable for small batch data), the learning rate is set to 0.001 (too small will lead to too long training period, too large will lead to oscillation), the batch size is set to 32 (balance training efficiency and gradient stability), the training rounds (Epoch) are set to 100, and the training is stopped when the reconstruction error is less than 0.005 for 10 consecutive rounds.

[0035] The model training takes 1000 multi-scale feature vectors as the sample set, and the reconstruction error changes during the training process: the initial error is 0.032 (the features are not fully learned), it decreases to 0.009 after 50 rounds, and it stabilizes at 0.004 (satisfies the convergence condition) after 80 rounds. At this time, the output of the H3 layer of the encoder is the high-dimensional feature representation vector, with a dimension of 64, and each dimension corresponds to a set of deep correlation features. The example vector is "[0.35, -0.18, 0.42, -0.25,..., 0.53]", where the 3rd dimension (0.42) corresponds to "the synergy of long-term resistance trend and temperature trend" (coefficient 0.72, highest weight), and the 12th dimension (-0.25) corresponds to "the negative impact of current transient fluctuations on resistance" (coefficient -0.61). These deep features are difficult for humans to directly extract, but they can effectively distinguish between device states - for example, the 3rd dimension value is usually between 0.3 and 0.5 during normal operation, while the value rises to 0.6-0.8 when the winding is slightly aged, and the feature difference is significant.

[0036] Based on the high-dimensional feature representation vector, a set of resistance feature vectors is constructed, and principal component analysis algorithm is used for feature dimension reduction to retain the main feature components and generate the reduced feature vector set. Although the high-dimensional feature representation vector (64 dimensions) has strong discrimination, it has the problem of "dimension disaster" (large amount of calculation and many redundant features), which needs to be reduced through principal component analysis (PCA). The core is to retain the principal components with high cumulative variance contribution rate through linear transformation, ensuring that the core information of the original features is still retained after dimension reduction.

[0037] First, build a "resistance feature vector set" - arrange 1000 64-dimensional high-dimensional vectors in chronological order to form a 1000x64 matrix, with each row corresponding to a high-dimensional feature of a data point and each column corresponding to a high-dimensional feature dimension. The PCA dimension reduction execution steps are as follows: Data centering: center each column (high-dimensional feature dimension) of the matrix, calculate formula x_centered=x-μ_j (μ_j is the mean of the jth column), eliminate the influence of mean difference on covariance calculation. For example, the 3rd column (resistance-temperature synergy feature) mean μ_j=0.41, and the 0.53 data point is centered to 0.12; Calculate the covariance matrix: The covariance matrix has a dimension of 64x64, and the element C_ij represents the covariance between the i-th column and the j-th column. The formula is C_ij = (1 / (n-1)) x Σ(x_centered_i x_centered_j) (n=1000 is the sample size). For example, the covariance between the 3rd column and the 12th column is C_3,12 ≈ (1 / 999) x Σ(0.12 x (-0.28) +...) ≈ -0.03, indicating a weak negative correlation between the two (when the current fluctuates greatly, the resistance-temperature correlation decreases slightly); Solve the eigenvalues and eigenvectors: Perform eigenvalue decomposition on the covariance matrix to obtain 64 eigenvalues (λ_1 ≥ λ_2 ≥... ≥ λ_64) and corresponding eigenvectors (v_1-v_64). The larger the eigenvalue, the more information the corresponding principal component contains. The calculation results are λ_1=13.2, λ_2=10.5, λ_3=7.8, λ_4=5.6, λ_5=4.1, and Σλ_k=46.8 (the sum of the 64 eigenvalues); Select principal components: Set the "cumulative variance contribution rate ≥ 85%" as the screening standard (industry commonly used threshold, taking into account information retention and dimension reduction), calculate the variance contribution rate of each principal component η_j=λ_j / Σλ_k: η_1=13.2 / 46.8≈28.2%, η_2=10.5 / 46.8≈22.4%, η_3=7.8 / 46.8≈16.7%, η_4=5.6 / 46.8≈12.0%, η_5=4.1 / 46.8≈8.8%, the cumulative contribution rate = 28.2% + 22.4% + 16.7% + 12.0% + 8.8% = 88.1% (≥ 85%), so select the first 5 principal components (m=5); Generate dimension reduction vectors: Multiply the 1000x64 matrix after centering with the 64x5 matrix composed of the first 5 eigenvectors to obtain a 1000x5 dimension reduction feature matrix, each dimension reduction vector has a dimension of 5, and the formula is PC=X_centered x V (X_centered is the centered matrix, and V is the eigenvector matrix).

[0038] The dimension-reduced eigenvectors need to explain the physical meaning of each principal component to ensure that the information is traceable: First principal component (PC1): The "resistance-temperature correlation feature" and "temperature trend feature" coefficients in the eigenvector are 0.75 and 0.65, respectively, indicating that PC1 mainly reflects the "long-term co-variation of resistance and temperature", contributing 28.2% of the information; Second principal component (PC2): The "current fluctuation feature" and "resistance short-term fluctuation feature" coefficients are 0.68 and -0.62, respectively, indicating that PC2 mainly reflects the "short-term influence of current instantaneous change on resistance", contributing 22.4% of the information; The third to fifth principal components correspond to "voltage stability and internal resistance correlation", "synchronization of daily load cycle and internal resistance", and "noise energy distribution", respectively, and collectively contribute 37.5% of the information.

[0039] An example dimension reduction vector is "[0.58, -0.32, 0.29, 0.16, -0.09]", in which PC1=0.58 (normal internal resistance and temperature correlation), PC2=-0.32 (small current fluctuation impact), corresponding to a normal device state; another vector "[1.32, -0.85, 0.73, 0.54, -0.36]" has PC1=1.32 (far beyond the normal range of 0.8), indicating abnormal internal resistance and temperature correlation, and possible winding aging.

[0040] An adaptive density clustering algorithm is designed to automatically adjust clustering parameters according to data distribution characteristics, and to perform clustering analysis on the reduced feature vector set to generate preliminary clustering results. Traditional density clustering algorithms (such as DBSCAN) require manual setting of "neighborhood radius ε" (distance threshold for judging core points) and "minimum core point number MinPts" (minimum number of points in the neighborhood), and improper parameters can easily lead to poor clustering results. The "adaptive density clustering algorithm" designed here automatically optimizes parameters based on data distribution, and the core logic is "K-distance graph determines ε, density distribution determines MinPts".

[0041] 1. Adaptive parameter adjustment Automatic determination of ε: Calculate the Euclidean distance between all vectors after dimension reduction (formula d(x, y)=√Σ(x_i-y_i)²), for each vector, select the K nearest points (K=5, an empirical value of 0.5% of the sample size, 5 for 1000 samples), calculate the K-distance (average of the K distances); arrange all K distances in descending order, draw a K-distance graph (horizontal axis is sample number, vertical axis is K-distance), and the K-distance corresponding to the "slope break point" in the graph is ε. For example, the K-distance of the first 820 samples in the K-distance graph is stable at 0.16-0.27, and the K-distance jumps to 0.43 after the 821st sample, the K-distance corresponding to the slope break point is 0.29, so ε=0.29. Automatic determination of MinPts: Calculate the local density of each vector (number of samples in the ε neighborhood), and count the density distribution histogram, and take the "number of points corresponding to the density peak" as MinPts. For example, the histogram shows that when the local density is 10, the number of samples is the largest (210), indicating that the density of most regions is suitable for MinPts=10, so MinPts=10 is set.

[0042] 2. Clustering execution process The core steps of the adaptive density clustering algorithm are consistent with DBSCAN, but the parameters are automatically optimized values: Core point judgment: For the dimension reduction vector x, if the number of samples in its ε neighborhood (distance ≤ 0.29) is ≥ MinPts = 10, then x is a core point. For example, if a certain vector neighborhood contains 12 samples, it is determined to be a core point; Density reachable judgment: If point y is in the ε neighborhood of core point x, then y is density reachable with x; if point z is density reachable with y, then z is density reachable with x. For example, the neighborhood of core point x contains y, and the neighborhood of y contains z, then z is density reachable with x; Cluster generation: all density reachable core points and non-core points are classified into a cluster, and samples with a point number < MinPts in the ε neighborhood are classified as noise points.

[0043] Taking 1000 dimension reduction vectors as an example, the clustering results are as follows: Cluster 1: contains 820 samples, 790 core points, principal component range PC1 ∈ [0.3, 0.8], PC2 ∈ [-0.5, 0.2], PC3 ∈ [-0.3, 0.3], corresponding to the normal operation mode of the device — sample ratio 82%, consistent with the rule that “normal operation time accounts for the majority”; Cluster 2: contains 130 samples, 105 core points, principal component range PC1 ∈ [0.8, 1.1], PC2 ∈ [-0.8, -0.5], corresponding to the slight abnormal mode — PC1 slightly exceeds the normal range, and it is speculated that the winding is slightly aged; Cluster 3: contains 40 samples, 28 core points, principal component range PC1 ∈ [1.1, 1.5], PC2 ∈ [-1.2, -0.8], corresponding to the moderate abnormal mode — PC1 is significantly out of range, and it is speculated that the winding is locally overheated; Noise points: 10 samples, neighborhood point number < 10, no fixed rule, and it is speculated that the abnormal data is caused by transient interference of the power grid (such as lightning, device start-stop).

[0044] The clustering effect verification adopts “silhouette coefficient”, and the calculation results are as follows: cluster 1 silhouette coefficient 0.83, cluster 2 0.76, cluster 3 0.69, and the overall average is 0.79 (> 0.5, good clustering effect); cluster distance (cluster center Euclidean distance): cluster 1-cluster 2 = 0.68, cluster 2-cluster 3 = 0.85, cluster 1-cluster 3 = 1.52, high separation degree, and no cluster overlap.

[0045] Abnormal mode identification is performed on the preliminary clustering results, the cluster that is obviously separated from other clusters is extracted as the internal resistance abnormal mode cluster, and finally the internal resistance abnormal mode cluster identifier is output.

[0046] This step needs to be accurately identified through “benchmark anchoring, multi-dimensional judgment, screening and grading, and identifier output”, to avoid misjudgment and omission, and the core is to establish an “abnormal judgment matrix” to comprehensively quantify the indicators and device mechanisms.

[0047] 1. Determine normal cluster reference Anchored by "historical data verification" and "cluster feature statistics" in two dimensions: Retrieve the past 3 months of fault-free data (7200) and process them in the same process. 98% (7056) of them belong to cluster 1, and the principal component range of cluster 1 is completely overlapped with the historical normal data; Cluster 1 accounts for 82% of the sample, with uniform local density (variance 0.9), which meets the stability characteristics of normal operation of the device, so cluster 1 is determined as "normal operation mode cluster".

[0048] 2. Multi-dimensional abnormality judgment For cluster 2 and cluster 3, quantify from "separation degree, abnormal feature intensity, fault matching degree": Separation degree: formula S = cluster center distance / average distance within the cluster, S of cluster 2 and cluster 1 = 0.68 / 0.26≈2.62, S of cluster 3 and cluster 1 = 1.52 / 0.29≈5.24 (threshold S>3 is significant separation), so cluster 3 is significantly separated, cluster 2 is moderately separated; Abnormal feature intensity: calculate the proportion of principal components exceeding the normal range, cluster 2 PC1 exceeds 70%, PC2 exceeds 65%, total exceeds 33%; Cluster 3 PC1 exceeds 100%, PC2 exceeds 100%, PC3 exceeds 95%, total exceeds 64% (threshold >50% is high intensity abnormality), so cluster 3 has high abnormality intensity, cluster 2 has moderate abnormality intensity; Fault matching degree: establish an "abnormal cluster-fault mode" association library, calculate the matching degree by cosine similarity (the closer to 1, the more matching), cluster 2 matches "winding slight aging" with a matching degree of 0.75, cluster 3 matches "winding local overheating" with a matching degree of 0.92 (threshold >0.8 is high matching), so cluster 3 matches clearly, cluster 2 matches moderately.

[0049] 3. Screening, grading and identification output Set the comprehensive score = separation degree × 0.4 + abnormal feature intensity × 0.3 + fault matching degree × 0.3, cluster 2 score = 2.62 × 0.4 + 33% × 0.3 + 0.75 × 0.3≈1.38 (50 points system is 55 points), cluster 3 score = 5.24 × 0.4 + 64% × 0.3 + 0.92 × 0.3≈2.57 (50 points system is 86 points); Grading according to the score: 80 points and above is "clear abnormal cluster", 50-79 points is "potential abnormal cluster", so cluster 3 is a clear abnormal cluster, cluster 2 is a potential abnormal cluster.

[0050] The final output is "internal resistance abnormal mode cluster identification", which includes: Unique ID: "AM-TR-001-Explicit Abnormality-20251001" (AM = Abnormal Mode, TR-001 = Device ID, date is the clustering day); Detailed Description: "Abnormal Features: PC1 ∈ [1.1, 1.5] (Severe Abnormality in Resistance-Temperature Coordination), PC2 ∈ [-1.2, -0.8] (Severe Impact of Current Fluctuation); Possible Causes: Local Overheating of Winding (Insulation Layer Aging and Insufficient Heat Dissipation); Impact Range: Resistance may exceed 170 mΩ within 72 hours, causing short circuit risk"; Risk Alert: "Moderate Risk, Suggesting Shutdown for Maintenance within 72 hours, Use Infrared Thermal Imager to Detect Winding Temperature"; Visual Results: Dimensionality Reduction Vector Scatter Plot (Normal Cluster Blue, Explicit Abnormal Cluster Red, Potential Abnormal Cluster Yellow), Label Cluster Center and Range, Intuitively Display Abnormal Distribution.

[0051] S203, based on the resistance abnormal mode cluster, perform multi-modal fault correlation analysis, combine device operation history and working condition label to predict potential fault evolution path and risk level, and generate fault prediction map; Specifically, at least the feature statistics of the distribution range, density, and outlying degree of the data points in the cluster can be extracted to generate an abnormal mode feature description; This step needs to quantitatively extract core statistics reflecting the data distribution characteristics from the identified resistance abnormal mode cluster (such as "Cluster 2: Winding Mild Aging" and "Cluster 3: Winding Local Overheating" in the previous text), providing "data fingerprints" for subsequent fault correlation and avoiding correlation bias caused by relying solely on qualitative descriptions. The extraction of feature statistics needs to be combined with the principal component features (PC1-PC5) after dimensionality reduction and the device operation mechanism to ensure that each statistic has a clear physical meaning.

[0052] 1. Cluster data point distribution range extraction The distribution range reflects the coverage area of the abnormal mode in the feature space, and needs to calculate "maximum value - minimum value - mean value - standard deviation" for the five principal components (PC1-PC5) after dimensionality reduction, among which PC1 (long-term coordination between resistance and temperature) and PC2 (short-term impact of current fluctuation on resistance) are the core principal components and need to be highlighted. Taking "Cluster 3: Winding Local Overheating" (40 samples) as an example: PC1: Maximum 1.5, Minimum 1.1, Mean 1.3, Standard Deviation 0.12, indicating that the abnormal degree of this cluster in the "resistance-temperature coordination" dimension is concentrated in 1.1-1.5 (far beyond the normal cluster range of 0.3-0.8), and the sample distribution is compact (small standard deviation), with stable abnormal characteristics; PC2: Max -0.8, Min -1.2, Mean -1.0, Std 0.08, reflecting "current fluctuation's negative impact on internal resistance" is significant and concentrated, without obvious dispersion; PC3-PC5: PC3 (voltage stability correlation) Mean 0.7, Std 0.05, PC4 (intra-day load cycle) Mean 0.52, Std 0.06, PC5 (noise energy) Mean -0.35, Std 0.04, all showing "high abnormality degree, concentrated distribution" characteristics.

[0053] Taking "Cluster 2: Winding slight aging" (130 samples) as an example, the distribution range difference is obvious: PC1: Max 1.1, Min 0.8, Mean 0.95, Std 0.09, abnormality degree is lower than Cluster 3, and the distribution range is narrower (0.8-1.1); PC2: Max -0.5, Min -0.8, Mean -0.65, Std 0.07, current fluctuation impact is weaker, consistent with the gradual characteristics of "slight aging".

[0054] The calculation of the distribution range needs to exclude noise points (such as the extreme value of PC1=1.6 in Cluster 3, which is determined as noise because it exceeds the mean ± 3 times the standard deviation (1.3±0.36), and is not included in the statistics), to ensure that the results reflect the true abnormal distribution.

[0055] 2. Cluster data point density extraction Density reflects the sample aggregation degree of abnormal patterns, and high density indicates that the abnormal pattern appears frequently in device operation and needs to be prioritized. Two indicators, "local density mean" and "core point proportion", are used for quantification: Local density mean: Calculate the average value of the local density (ε=0.29 number of samples within the neighborhood) of all samples in the cluster, formula is "density mean=(1 / N)×∑ρ_i" (N is the number of samples in the cluster, ρ_i is the local density of the i-th sample). Among the 40 samples of Cluster 3, the local density range is 10-15, the mean is 12.5, indicating that the sample has high aggregation degree in the feature space, and the frequency of the abnormal pattern is stable; the local density range of the 130 samples of Cluster 2 is 8-12, the mean is 9.8, the aggregation degree is lower than Cluster 3, consistent with the rule of "low frequency of slight abnormality"; Core point proportion: The proportion of the number of core points (local density ≥ MinPts=10) to the total number of samples in the cluster, formula is "core point proportion=core point number / N×100%". Cluster 3 has 32 core points, accounting for 80% (32 / 40), indicating that the cluster is mainly composed of stable core samples, and the abnormal pattern is reliable; Cluster 2 has 91 core points, accounting for 70% (91 / 130), the core sample proportion is slightly lower, and there are a small number of edge samples (local density 8-9), which need to be further verified in combination with the working conditions.

[0056] 3. Outlier degree extraction of data points in cluster The outlier degree reflects the deviation of abnormal mode from normal mode, which is quantified by "outlier index". The formula is "outlier index = (Euclidean distance between cluster center and normal cluster center) / (average distance of samples in normal cluster)". The larger the index, the more serious the deviation: Cluster center calculation: Take the mean vector of principal components of all samples in the cluster. The cluster 3 center vector is [1.3, -1.0, 0.7, 0.52, -0.35], and the normal cluster (cluster 1) center vector is [0.55, -0.15, 0.02, 0.05, -0.03]; Euclidean distance calculation: The distance d between cluster 3 and normal cluster center is 1.44; The average distance in normal cluster: The average Euclidean distance between 820 samples in normal cluster is 0.21; Outlier index: The outlier index of cluster 3 is 1.44 / 0.21 ≈ 6.86, which is much larger than the outlier index of cluster 2 (0.81 / 0.21 ≈ 3.86), indicating that the abnormal deviation degree of cluster 3 is 1.78 times that of cluster 2, and the fault risk is higher.

[0057] 4. Generating abnormal mode feature description Integrate the above statistical quantities to generate a description according to the structure of "cluster identification-core principal component distribution-density feature-outlier degree-physical meaning", as follows: Cluster 3 (explicit anomaly, local overheating of winding): "Cluster identification AM-TR-001-explicit anomaly-20251001; Core principal component distribution: PC1 [1.1, 1.5] (mean 1.3, standard deviation 0.12), PC2 [-1.2, -0.8] (mean -1.0, standard deviation 0.08), reflecting severe abnormality of internal resistance and temperature coordination, and intense negative impact of current fluctuation; Density feature: local density mean 12.5, core point proportion 80%, high abnormal mode aggregation, stable occurrence frequency; Outlier index 6.86, significant deviation from normal mode; Physical meaning: insulation layer aging of winding leads to heat accumulation, internal resistance increases at a speed much higher than normal, and short-term local overheating failure is easily caused." Cluster 2 (potential anomaly, slight aging of winding): "Cluster identification AM-TR-001-potential anomaly-20251001; Core principal component distribution: PC1 [0.8, 1.1] (mean 0.95, standard deviation 0.09), PC2 [-0.8, -0.5] (mean -0.65, standard deviation 0.07), reflecting slight abnormality of internal resistance and temperature coordination, and weak impact of current fluctuation; Density feature: local density mean 9.8, core point proportion 70%, moderate abnormal mode aggregation; Outlier index 3.86, moderate deviation from normal mode; Physical meaning: slight oxidation of winding wire, internal resistance increases slowly, no direct fault risk in short term, but may deteriorate in long term." In combination with the fault records and working condition labels in the equipment operation history data, a fault correlation model is constructed by using a graph neural network to analyze the correlation between abnormal patterns and fault types and generate a fault correlation analysis result. This step needs to establish a quantitative correlation between "internal resistance abnormal patterns" and "equipment fault types" to avoid errors caused by subjective experience matching. Graph Neural Network (GNN) is suitable for capturing the complex mapping relationship between abnormalities and faults because it can effectively handle "node-edge" correlation structure data. Here, Graph Attention Network (GAT) is selected to automatically assign correlation weights through an attention mechanism to highlight strong correlation relationships.

[0058] 1. Construct the input graph structure of the graph neural network The node and edge definitions of the graph need to be combined with historical data and working condition labels to ensure that all key elements are covered. Node types: including "abnormal pattern nodes" (such as cluster 2, cluster 3), "fault type nodes" (such as slight winding aging, local winding overheating, winding short circuit, and terminal loosening), and "working condition label nodes" (such as high load peak (current 140-150A), low load valley (current 100-110A), high temperature environment (oil temperature ≥ 50℃), and low temperature environment (oil temperature ≤ 40℃)), a total of 2 abnormal nodes, 4 fault nodes, and 4 working condition nodes, totaling 10 nodes. Edge definition: edges represent the correlation between nodes and are divided into "abnormal-fault edges" (direct correlation between abnormal patterns and fault types), "abnormal-working condition edges" (dependent correlation between abnormal patterns and working conditions), and "fault-working condition edges" (correlation between fault occurrence and working conditions). The initial weight of the edge is based on historical data statistics (such as the proportion of the number of times that cluster 3 and local winding overheating co-occur in a high temperature environment).

[0059] Taking the past three years of historical data of the target equipment (TR-001) as an example, a total of 1200 operation records were statistically analyzed, including 28 faults (12 slight aging, 8 local overheating, 5 short circuit, and 3 terminal loosening). The working condition labels cover 200 days of high load peak, 400 days of low load valley, 150 days of high temperature environment, and 250 days of low temperature environment. Based on this, the initial edge weight is determined: Abnormal-fault edge initial weight: the number of times that cluster 3 and local overheating co-occur is 8 (cluster 3 occurs 10 times and faults occur 8 times), initial weight = 8 / 10 = 0.8; the number of times that cluster 2 and slight aging co-occur is 12 (cluster 2 occurs 17 times and faults occur 12 times), initial weight = 12 / 17 ≈ 0.71; Abnormal-condition edge initial weight: cluster 3 appears 9 times in high-temperature environment (total 10 times), weight = 9 / 10 = 0.9; cluster 2 appears 10 times in load peak (total 17 times), weight = 10 / 17 ≈ 0.59.

[0060] 2. Training and optimization of graph neural network (GAT) The GAT network structure is set as "input layer-attention layer-output layer", and the core is to optimize the edge weight through attention mechanism to improve the correlation accuracy: Input layer: the feature vector of each node is "node type code + historical statistical features", such as the feature vector of abnormal mode nodes containing "abnormal type code (cluster 2 = 01, cluster 3 = 10) + outlier index + density mean", and the feature vector of fault nodes containing "fault type code (local overheating = 100) + fault occurrence frequency + average repair time"; Attention layer: single-head attention mechanism is used to calculate the attention coefficient α_ij of each node to its neighbor nodes, the formula is "α_ij=softmax_j(LeakyReLU(W・[h_i||h_j]))" (W is the weight matrix, h_i, h_j are the feature vectors of nodes i, j, || represents concatenation), the larger the attention coefficient, the stronger the correlation between nodes; Output layer: output the optimized edge weight (attention coefficient × initial weight), the training goal is to minimize the mean square error of "predicted correlation result and actual historical fault correlation", training parameters: learning rate 0.001, batchsize=32, training rounds 50, stop training when the validation set error is less than 0.05.

[0061] Example of optimized edge weight after training: Abnormal-fault edge: cluster 3→local overheating weight 0.92 (optimized from 0.8, as the attention mechanism captures strong matching between the two in the PC1 abnormal range), cluster 3→short circuit weight 0.68 (indirect correlation, local overheating is prone to short circuit), cluster 2→mild aging weight 0.85 (optimized from 0.71, as the correlation is enhanced under load peak working condition), cluster 2→local overheating weight 0.32 (potential correlation, mild aging may worsen); Abnormal-condition edge: cluster 3→high-temperature environment weight 0.95 (optimized from 0.9, high temperature aggravates heat accumulation), cluster 2→load peak weight 0.65 (optimized from 0.59, load peak accelerates wire oxidation).

[0062] 3. Generate fault correlation analysis results Integrate the optimized edge weight and output the results in the structure of "abnormal mode-correlation fault-correlation working condition-correlation strength", for example: Cluster 3 (local overheating of windings) correlation results: Core correlation fault: local overheating of windings (correlation strength 0.92), secondary correlation fault: short circuit of windings (correlation strength 0.68); critical correlation condition: high temperature environment (correlation strength 0.95), secondary correlation condition: peak load (correlation strength 0.62); correlation logic: the PC1 anomaly of cluster 3 (severe imbalance of internal resistance-temperature coordination) is highly matched with the fault characteristics of local overheating (winding temperature ≥120℃), and in the high temperature environment, the heat cannot be dissipated in time, and the probability of the abnormal mode turning into a fault increases by 35%.

[0063] Cluster 2 (Slight Aging of Winding) Correlation Results: Core Correlation Fault: Slight Aging of Winding (Correlation Strength 0.85), Secondary Correlation Fault: Local Overheating of Winding (Correlation Strength 0.32); Key Correlation Condition: Peak Load (Correlation Strength 0.65), Secondary Correlation Condition: Low Temperature Environment (Correlation Strength 0.28); Correlation Logic: The slight anomaly of PC1 in Cluster 2 (slow increase in internal resistance) matches the fault characteristics of slight aging (conductor oxide layer thickness ≤ 0.1 mm). During peak load, the current increases, the conductor heats up and accelerates oxidation, and the anomaly may deteriorate into local overheating within 6 months.

[0064] Based on the results of fault correlation analysis, a hidden Markov model is used to predict the evolution path of potential faults, calculate the probability of occurrence of each path, and generate fault evolution path prediction results. Hidden Markov Models (HMMs) are suitable for handling "state sequence evolution" problems. Here, "abnormal mode → fault state → more severe fault" is regarded as a state transition process. HMMs predict the probability of different evolution paths, providing a basis for risk assessment. The core elements of an HMM include "state set, observation set, transition probability matrix, observation probability matrix, and initial state probability," which need to be determined in conjunction with fault correlation results and historical data.

[0065] 1. Define the core elements of an HMM State set S: includes abnormal states and fault states, a total of 5 states, S1=Cluster 2 (slight aging), S2=Cluster 3 (local overheating), S3=Slight aging fault of winding, S4=Local overheating fault of winding, S5=Short circuit fault of winding. Observation set O: Based on the characteristic statistics of the abnormal pattern, there are 3 observations in total: O1 = outlier index 3-4 (minor anomaly observation), O2 = outlier index 6-7 (serious anomaly observation), O3 = internal resistance > 170mΩ (fault observation). Initial state probability π: Based on the proportion of the current abnormal mode, if the current detected cluster 2 samples account for 70% of the abnormal samples and cluster 3 accounts for 30%, then π=[0.7,0.3,0,0,0] (only abnormal states have an initial probability, and the initial probability of the fault state is 0). Transition probability matrix A: A[i][j] represents the probability of transition from state Si to Sj, calculated based on the past 3 years of failure evolution history data (5 out of 12 mild aging deteriorated to local overheating, 3 out of 8 local overheating deteriorated to short circuit): A[S1→S1]=0.35 (maintain cluster 2 anomaly), A[S1→S3]=0.48 (cluster 2 transforms into mild aging failure), A[S1→S2]=0.17 (cluster 2 deteriorates into cluster 3 anomaly); A[S2→S2]=0.30 (maintain cluster 3 anomaly), A[S2→S4]=0.65 (cluster 3 transforms into local overheating failure), A[S2→S5]=0.05 (cluster 3 directly transforms into short circuit failure); A[S3→S3]=0.60 (maintain mild aging failure), A[S3→S4]=0.40 (mild aging deteriorates into local overheating failure); A[S4→S4]=0.70 (maintain local overheating failure), A[S4→S5]=0.30 (local overheating deteriorates into short circuit failure); A[S5→S5]=1.0 (short circuit failure is the terminal state, no transition); Observation probability matrix B: B[i][k] represents the probability of state Si generating observation Ok, calculated based on the matching degree between the characteristic statistics and the state: B[S1→O1]=0.90 (cluster 2 corresponds to mild anomaly observation), B[S1→O2]=0.10 (cluster 2 occasionally appears serious observation); B[S2→O2]=0.95 (cluster 3 corresponds to serious anomaly observation), B[S2→O3]=0.05 (cluster 3 occasionally appears failure observation); B[S3→O1]=0.85 (mild aging failure corresponds to mild observation), B[S3→O2]=0.15 (mild aging occasionally appears serious observation); B[S4→O2]=0.70 (local overheating failure corresponds to serious observation), B[S4→O3]=0.30 (local overheating appears failure observation); B[S5→O3]=1.0 (short circuit failure corresponds to failure observation).

[0066] 2. Predicting failure evolution path and occurrence probability The forward algorithm is used to calculate the occurrence probability of each state sequence (evolution path), focusing on the path from the initial anomaly state to the failure terminal state: Path 1: S1 (cluster 2) → S3 (mild aging failure) → S4 (local overheating failure) → S5 (short circuit failure); Path 2: S1 (cluster 2) → S3 (mild aging failure) → S3 (maintain mild aging); Path 3: S2 (cluster 3) → S4 (local overheating failure) → S4 (maintain local overheating); Path 4: S2 (cluster 3) → S4 (local overheating failure) → S5 (short circuit failure); Path 5: S1 (Cluster 2) → S1 (Maintain Cluster 2) → S1 (Maintain Cluster 2).

[0067] 3. Generate Fault Evolution Path Prediction Results Output the results in the structure of "Path Priority (Probability from High to Low) - Path Description - Occurrence Probability - Evolution Period", and highlight high-probability paths (probability > 5%): Priority 1: Path S1 → S3 → S3 (Maintain Mild Aging Fault), Occurrence Probability 14.56%, Evolution Period 6-12 months, Characteristics: Cluster 2 abnormality is transformed into a mild aging fault, and is maintained for a long time without obvious deterioration; Priority 2: Path S1 → S1 → S1 (Maintain Cluster 2 Abnormality), Occurrence Probability 6.95%, Evolution Period 3-6 months, Characteristics: Cluster 2 abnormality does not transform, only appears slight fluctuation during high load peak; Priority 3: Path S2 → S4 → S4 (Maintain Local Overheating Fault), Occurrence Probability 2.87%, Evolution Period 1-3 months, Characteristics: Cluster 3 abnormality is quickly transformed into a local overheating fault, which needs to be intervened within 72 hours; Priority 4: Path S1 → S3 → S4 → S5 (Mild Aging → Local Overheating → Short Circuit), Occurrence Probability 2.4%, Evolution Period 12-18 months, Characteristics: Abnormality gradually deteriorates, short circuit risk increases over time; Supplementary Note: Path S2 → S4 → S5 (Local Overheating → Short Circuit) Probability 1.76%, although the probability is low, but the evolution period is short (within 1 month), which needs to be closely watched.

[0068] According to the fault evolution path prediction results, evaluate the risk level of each path, considering the severity of the fault and the probability of occurrence, and generate risk level evaluation results; Risk level assessment needs to consider both the "destructiveness of the fault (severity)" and the "likelihood of occurrence (probability)", to avoid misjudgment caused by focusing on a single dimension (such as balancing the risk of high-probability mild fault and low-probability serious fault). The evaluation process consists of four steps: "defining severity classification-quantifying occurrence probability-calculating risk value-dividing risk level", to ensure that the results are quantifiable and comparable.

[0069] 1. Define Fault Severity Classification Based on the impact of the fault on the device operation, repair cost and downtime, the severity is divided into 5 levels (1st lowest, 5th highest), each level corresponds to clear judgment criteria and quantitative indicators: Level 1 (no risk): no failure, only abnormal mode, no downtime, repair cost 0 yuan, such as maintaining cluster 2 abnormality (path 5); Level 2 (slight risk): slight failure, no need to stop, repair cost ≤5000 yuan, downtime 0 hours, such as slight aging failure (path 2); Level 3 (mild risk): moderate failure, need to plan downtime (1-4 hours), repair cost 5000-20000 yuan, such as local overheating failure (path 3); Level 4 (moderate risk): serious failure, need to stop urgently (within 72 hours), repair cost 20000-50000 yuan, such as local overheating failure with short circuit risk (path 4); Level 5 (emergency risk): fatal failure, need to stop immediately (within 1 hour), repair cost >50000 yuan, may cause safety accidents, such as short circuit failure (end of path 1).

[0070] Taking the end state of each path as an example, the severity is marked as follows: Path 2 end S3 (slight aging): Level 2; Path 5 end S1 (cluster 2): Level 1; Path 3 end S4 (local overheating): Level 3; Path 1 end S5 (short circuit): Level 5; Path 4 end S5 (short circuit): Level 5.

[0071] 2. Quantify the failure probability interval Combined with the path occurrence probability, the probability is divided into 3 intervals for risk calculation: High probability: P>10%, corresponding to "extremely likely to occur"; Medium probability: 3%≤P≤10%, corresponding to "likely to occur"; Low probability: P<3%, corresponding to "occasionally occur".

[0072] Mark the probability interval of each path: Path 2 (14.56%): high probability; Path 5 (6.95%): medium probability; Path 3 (2.87%): low probability; Path 1 (2.4%): low probability; Path 4 (1.76%): low probability.

[0073] 3. Calculate risk value and divide risk level The formula "risk value = severity × probability weight × risk coefficient" is used for calculation, where "probability weight" is assigned according to the interval (high probability = 1.0, medium probability = 0.6, low probability = 0.3), and "risk coefficient" is the importance coefficient of the device (target device TR-001 is a key power distribution device, coefficient = 1.2, non-key device coefficient = 1.0). The higher the risk value, the higher the risk level: Path 2 (S3, Level 2, High Probability): Risk Value = 2 x 1.0 x 1.2 = 2.4; Path 5 (S1, Level 1, Medium Probability): Risk Value = 1 x 0.6 x 1.2 = 0.72; Path 3 (S4, Level 3, Low Probability): Risk Value = 3 x 0.3 x 1.2 = 1.08; Path 1 (S5, Level 5, Low Probability): Risk Value = 5 x 0.3 x 1.2 = 1.8; Path 4 (S5, Level 5, Low Probability): Risk Value = 5 x 0.3 x 1.2 = 1.8.

[0074] According to the risk value, 4 risk levels are divided, corresponding to different response priorities: Low risk: Risk Value <1.0, Priority 4 (No immediate treatment needed, monthly monitoring), such as Path 5 (0.72); Lower risk: 1.0 ≤ Risk Value <2.0, Priority 3 (Quarterly assessment, as needed), such as Path 3 (1.08), Path 1 (1.8), Path 4 (1.8); Medium risk: 2.0 ≤ Risk Value <3.0, Priority 2 (Monthly assessment, planned treatment), such as Path 2 (2.4); High risk: Risk Value ≥3.0, Priority 1 (Immediate treatment, 24-hour intervention), no path meets.

[0075] 4. Generate risk level assessment results Output in the structure of "Path-End State-Severity-Probability Interval-Risk Value-Risk Level-Response Priority", for example: Path 2 (S1→S3→S3): End state mild aging failure, severity level 2, probability interval high probability (14.56%), risk value 2.4, risk level medium risk, response priority 2; Suggestions: Increase resistance monitoring by 1 time per month, conduct winding insulation test every 3 months, develop a plan for maintenance shutdown within 6 months to avoid deterioration.

[0076] Path 5 (S1→S1→S1): End state cluster 2 anomaly, severity level 1, probability interval medium probability (6.95%), risk value 0.72, risk level low risk, response priority 4; Suggestions: Maintain regular monitoring frequency (1 time / minute), quarterly data summary, no additional intervention needed.

[0077] Path 3 (S2→S4→S4): End state local overheating failure, severity level 3, probability interval low probability (2.87%), risk value 1.08, risk level lower risk, response priority 3; Suggestions: Complete infrared thermal imaging detection within 72 hours, evaluate winding temperature distribution, if temperature ≥120℃, immediately arrange for emergency shutdown.

[0078] Path 1 (S1→S3→S4→S5) and Path 4 (S2→S4→S5): End state short circuit fault, severity level 5, probability interval low probability (2.4% / 1.76%), risk value 1.8, risk level low risk, response priority 3; suggestion: check winding insulation resistance once a week, if insulation resistance <100MΩ, stop immediately to prevent short circuit accident.

[0079] Integrate the results of fault correlation analysis, fault evolution path prediction, and risk level evaluation, and construct a multi-dimensional fault prediction map to finally output a complete fault prediction map.

[0080] The multi-dimensional fault prediction map needs to visualize the integration of the three types of information: correlation, evolution path, and risk level. A hierarchical structure is adopted to ensure that maintenance personnel can intuitively understand the fault correlation, possible evolution direction, and risk level of abnormal patterns. The map needs to include four modules: basic information area, correlation analysis area, path prediction area, and risk assessment area. Each module needs to be labeled with data source and confidence level to ensure traceability of the results.

[0081] 1. Basic information area of the map Located at the top of the map, it marks the core basic information to ensure clear context: Target device information: device ID (TR-001), device type (10kV distribution transformer), monitoring period (20251001-20251002), data source (real-time monitoring data + 3 years of historical fault records); Abnormal pattern overview: current number of abnormal patterns identified (2: cluster 2, cluster 3), abnormal sample proportion (140 abnormal samples, accounting for 9.7% of the total 1440 samples), data confidence (feature extraction confidence 0.92, clustering confidence 0.89).

[0082] 2. Correlation analysis area of the map Located on the left side of the map, it uses a graphical description of "nodes with weighted edges" (text instead of graphics) to show the correlation between abnormal patterns, faults, and working conditions: Abnormal pattern node: marked with a red box as "cluster 2 (mild aging, outlier index 3.86)" and "cluster 3 (local overheating, outlier index 6.86)", with core feature statistics marked below the node; Fault node: marked with a blue ellipse as "mild aging fault (repair cost 5k)" and "local overheating fault (repair cost 20k)" and "short circuit fault (repair cost 50k)", with severity level marked below the node; Working condition node: marked with a green diamond as "high load peak (140-150A)" and "high temperature environment (≥50℃)", with association frequency marked below the node; Weighted edges: described as "abnormal node → fault node (weight value)" "abnormal node → working condition node (weight value)", such as "cluster 2 → slight aging failure (0.85)" "cluster 3 → high temperature environment (0.95)", the greater the weight value, the more prominent the edge description (such as bold font).

[0083] 3. Atlas path prediction area Located in the middle of the atlas, the evolution paths are arranged vertically according to "path priority", each path is described with "arrow sequence + probability", high probability paths are marked in orange, and low probability paths are marked in gray: Priority 1 (medium risk): cluster 2 → slight aging failure → maintain slight aging (probability 14.56%), evolution period marked "6-12 months"; Priority 2 (low risk): cluster 2 → maintain cluster 2 anomaly (probability 6.95%), evolution period marked "3-6 months"; Priority 3 (lower risk): cluster 3 → local overheating failure → maintain local overheating (probability 2.87%), evolution period marked "1-3 months"; Priority 4 (lower risk): cluster 2 → slight aging → local overheating → short circuit (probability 2.4%), evolution period marked "12-18 months"; Priority 5 (lower risk): cluster 3 → local overheating → short circuit (probability 1.76%), evolution period marked "within 1 month".

[0084] 4. Atlas risk assessment area Located on the right side of the atlas, arranged horizontally according to "risk level", each risk level corresponds to a path and a response suggestion, distinguished by color (low risk green, lower risk yellow, medium risk orange): Low risk (green): path 5 (risk value 0.72), suggestion "regular monitoring, quarterly summary"; Lower risk (yellow): paths 3, 1, 4 (risk value 1.08-1.8), suggestion "infrared detection, weekly insulation check"; Medium risk (orange): path 2 (risk value 2.4), suggestion "monthly monitoring, 6-month plan downtime"; Risk summary: current device overall risk level "lower-medium risk", no high-risk path, key intervention nodes "within 72 hours (cluster 3), within 6 months (cluster 2)".

[0085] 5. Output complete failure prediction atlas The atlas bottom is marked with "atlas update frequency (updated automatically every 24 hours)" "data deviation range (probability deviation ±3%, risk value deviation ±0.2)" "operation responsibility person contact information", ensuring the timeliness and operability of the atlas. The final output atlas can be displayed through the HMI interface of the DCS system, supporting clicking on the node to view detailed data (such as clicking on the cluster 3 node, the cluster's feature statistics and historical failure cases are popped up), providing intuitive and comprehensive decision-making basis for subsequent operation strategy formulation.

[0086] S204, generate a set of differentiated operation and maintenance strategies according to the fault prediction map, and drive the operation and maintenance management platform to execute corresponding operation and maintenance instructions.

[0087] Specifically, the risk level and fault type information in the fault prediction map can be parsed, and the corresponding basic operation and maintenance strategy is matched from the operation and maintenance strategy library to generate a preliminary operation and maintenance strategy set. First, the core decision information of the fault prediction map needs to be extracted completely. The map clearly labels two types of key abnormal patterns: one is "Cluster 2 (light aging of winding)", the fault type is "slowly rising internal resistance caused by oxidation of winding conductor", the risk level is "medium risk", the evolution period is 6-12 months, and the key associated working condition is "high load peak (current 140-150A)"; the other is "Cluster 3 (local overheating of winding)", the fault type is "heat accumulation caused by insulation layer aging", the risk level is "low risk", the evolution period is 1-3 months, and the key associated working condition is "high temperature environment (oil temperature ≥ 50℃)". The parsing process needs to ensure that the key constraint "associated working condition" is not missed - the development speed of the same fault under different working conditions is significantly different (such as the monthly change of internal resistance of Cluster 2 is 1.5 times that of low load valley), which needs to be used as an important basis for subsequent strategy matching.

[0088] The operation and maintenance strategy library needs to be organized according to the "risk level-fault type-working condition label" three-dimensional structure to ensure the accuracy of matching. Each basic strategy in the library contains four core contents: "monitoring frequency, detection item, early warning threshold, basic disposal measure", and the parameters need to be set based on industry standards and historical operation and maintenance experience of the equipment: The basic strategy for the "medium risk-light aging of winding-load peak" scene (strategy number S-002): ① Monitoring frequency: monthly internal resistance special monitoring (sampling frequency 1 time / minute, lasting 24 hours), weekly routine operation parameter (voltage, current, temperature) monitoring; ② Detection item: every 3 months 1 time winding insulation resistance test (using KEITHLEY2450 tester, range 0-1000MΩ, accuracy ±0.01MΩ, threshold ≥100MΩ), every 6 months 1 time infrared thermal imaging detection (FLIRT650sc thermal imager, resolution 640×480, temperature measurement range -20~120℃, winding temperature threshold ≤110℃); ③ Early warning threshold: send a short message warning when the monthly change of internal resistance is greater than 0.5mΩ, and trigger an audible and light alarm when the insulation resistance is less than 100MΩ; ④ Basic disposal: arrange operation and maintenance personnel to review data within 24 hours after warning to exclude measurement errors.

[0089] Basic strategy for the "lower risk - localized overheating of windings - high temperature environment" scenario (strategy number S-008): ① Monitoring frequency: internal resistance monitoring once every 2 weeks, oil temperature monitoring once a day (sampling frequency once every 10 minutes); ② Inspection items: infrared thermal imaging inspection once a month (focusing on scanning the middle area of ​​the winding, temperature threshold ≤120℃), cooling system function check once every 2 months (fan speed ≥2600rpm, no dust accumulation on heat sink); ③ Warning threshold: warning when oil temperature is ≥55℃ for 1 hour or internal resistance change >0.3mΩ; ④ Basic handling: clean the cooling system within 48 hours after the warning, and replace the fan filter if necessary.

[0090] The strategy matching adopts a "dual keyword retrieval": first, candidate strategies are located by "risk level + fault type" (e.g., "medium risk + slight aging" locates S-002 and S-003), and then the optimal strategy is selected by "related operating conditions" (e.g., cluster 2 is associated with "load peak", excluding S-003 which is adapted to "load trough", and determining S-002). The final generated preliminary operation and maintenance strategy set includes 2 core strategies (S-002 and S-008) and 1 supplementary strategy (S-015, for "multiple abnormal mode superposition" scenarios, such as when cluster 2 and cluster 3 occur at the same time, priority is given to handling local overheating). The set needs to label the "applicable scenario, effective time, and confidence level" of each strategy (e.g., S-002 has a confidence level of 0.92, based on the operation and maintenance effect statistics of 12 similar faults in the past 3 years) to ensure that subsequent optimization has a clear basis.

[0091] Based on real-time equipment operating status and historical maintenance effect data, reinforcement learning algorithms are used to optimize and adjust the initial set of maintenance strategies to generate optimized maintenance strategies. The core advantage of reinforcement learning (RL) is its ability to dynamically adapt to changes in equipment state through a trial-and-error reward mechanism. Here, the Deep Deterministic Policy Gradient (DDPG) algorithm is chosen because it is suitable for handling continuous action spaces (such as continuous adjustments to monitoring frequency and warning thresholds), avoiding the parameter jump problem of discrete algorithms. Algorithm design must clearly define the three core elements: agent, environment, and reward function, ensuring that the optimization direction aligns with the operational goals (fault prevention, cost control, and energy saving).

[0092] 1. Definition of core elements Agent: Deployed in the operation and maintenance management system as a "policy optimization module". Its core function is to output adjusted policy parameters (such as monitoring frequency and early warning threshold) and input environmental status feedback.

[0093] Environment: Contains real-time running state and historical operation effect data of the device. Real-time state parameters include: current 142A (high load peak interval), oil temperature 48°C (close to high temperature threshold), internal resistance 152.3 mΩ (0.3 mΩ higher than last month); historical effect data includes: results of the past 10 times of executing S-002 strategy (8 times met the standard, 2 times triggered warning due to current exceeding 145A leading to internal resistance change exceeding 0.5 mΩ, average disposal cost after warning 2000 yuan), energy consumption data under different monitoring frequencies (5 kWh of energy consumption per month, 8 kWh of energy consumption every 2 weeks, 15 kWh of energy consumption every week).

[0094] Reward Function: Comprehensive measure of "fault prevention effect, operation cost, energy consumption", formula is R = 0.4 x R_fault + 0.3 x R_cost + 0.3 x R_energy, where: R_fault (fault prevention score): If no warning is triggered after strategy execution, score 100; if warning is triggered but no fault occurs, score 80; if fault occurs, score 0 (e.g. cluster 2 if not monitored in time leading to deterioration of aging, score 0); R_cost (cost control score): Based on labor and equipment costs, benchmark cost is 2000 yuan (S-002 single execution cost), 15 points for every 10% reduction, 10 points deducted for every 10% increase (e.g. optimized cost reduced to 1800 yuan, 15 points); R_energy (energy saving score): Benchmark energy consumption is 5 kWh / month, 20 points for every 1 kWh reduction, 10 points deducted for every 1 kWh increase (e.g. optimized energy consumption 6 kWh, 10 points deducted).

[0095] 2. Optimization process and results Algorithm training iterates 30 times, initial strategy is the original parameters of S-002 and S-008, at the 10th iteration, key rules are found: when current > 140A, the internal resistance daily change of cluster 2 (0.02 mΩ / day) is 1.5 times that of current ≤ 140A (0.013 mΩ / day), and the temperature deviation of infrared detection will increase by 5% (due to current heating interference). Based on this, the agent begins to adjust the parameters: Optimization for S-002: ① Monitoring frequency: when current > 140A, the internal resistance monitoring frequency is adjusted from "once a month" to "once every 2 weeks", and when current ≤ 140A, it remains once a month; the regular parameter monitoring is changed from "once a week" to "once a day" (only when current > 140A); ② Warning threshold: the internal resistance monthly change is tightened from 0.5mΩ to 0.4mΩ (due to increased risk during current peak), and the insulation resistance threshold remains 100MΩ (not affected by current); ③ Detection items: in high temperature environment (oil temperature ≥ 50℃), the infrared detection cycle is shortened from 6 months to 4 months (to avoid heat accelerated aging).

[0096] Optimization for S-008: ① Monitoring frequency: when oil temperature ≥ 50℃, oil temperature monitoring is changed from "once a day" to "once every 2 hours", and internal resistance monitoring is changed from "once every 2 weeks" to "once a week"; ② Cooling system inspection: fan speed threshold is increased from 2600rpm to 2700rpm (stronger heat dissipation is required in high temperature environment), and cleaning cycle is shortened from 2 months to 1 month.

[0097] The reward function score of the optimized strategy is improved from the initial 75 points (S-002), 72 points (S-008) to 92 points (optimized S-002), 89 points (optimized S-008). The reasons for the improvement are: ① R_fault is increased from 80 points to 100 points (no warning after 2 rounds of iteration after adjustment); ② R_cost is reduced from 85 points to 80 points (cost increases by 5% due to increased monitoring frequency), but R_energy is increased from 60 points to 90 points (only in high-risk working conditions, increase the frequency, and maintain low energy consumption in low-risk conditions), the comprehensive benefit is significant. The final output optimization strategy needs to be marked with "adjustment basis" (such as "shorten the monitoring period when current > 140A, based on historical data that the aging speed increases by 50% under this working condition"), to ensure traceability.

[0098] According to the optimized operation and maintenance strategy, specific executable operation and maintenance instructions are generated, including maintenance time, maintenance content and required resources, to obtain a set of operation and maintenance instructions; Executable instructions need to convert the "abstract parameters" of the optimization strategy into "concrete operation tasks", each instruction needs to meet the "time feasible, content clear, resource available" three requirements, to avoid ambiguous expressions (such as "repair as soon as possible" needs to be clear as "2025-10-12 02:00-04:00"). Instruction generation needs to be combined with device production plan (to avoid affecting normal power supply), resource availability (such as device calibration status, personnel scheduling), safety specifications (such as high-voltage device maintenance requires power-off permission).

[0099] 1. Instructions for optimized S-002 (Cluster 2: slight aging of winding) (Instruction ID: Cmd-TR001-20251005-001) Maintenance time: October 12, 2025, 02:00-04:00 (Choose the low load period in the early morning on weekends, the average load of the equipment is only 30% of the peak, the impact on production is minimal after shutdown; submit a power outage application 72 hours in advance to ensure permission from power dispatching) Maintenance content: ① Winding insulation resistance test: use KEITHLEY2450 insulation resistance tester, disconnect the high voltage side power supply before testing, ground discharge for 10 minutes; test points are the high voltage side A, B, C three-phase terminals and ground terminals, each phase test time is 1 minute, record data (require each phase ≥100MΩ, three-phase deviation ≤5%); ② Infrared thermal imaging detection: use FLIRT650sc thermal imager, lens distance from winding surface 1.5 meters, scanning range covers the entire winding area (including joints, leads), focus on marking areas with temperature >100℃, record the highest temperature (require ≤110℃); ③ Internal resistance retest: use the same type of internal resistance tester, measure the internal resistance of phase A on the high voltage side, compare with the same period data last month (152.0mΩ), calculate the change (require ≤0.4mΩ); ④ Data archiving: enter the test data into the equipment operation and maintenance file (system number file-TR001-202510), generate monthly operation and maintenance report.

[0100] Required resources: ① Equipment: KEITHLEY2450 tester 1, FLIRT650sc thermal imager 1, high voltage detector 1 set; ② Personnel: 2 certified high voltage electricians, 1 detection engineer; ③ Consumables: anhydrous ethanol, insulating gloves; ④ Time allocation: power off discharge for 10 minutes, insulation test for 30 minutes, infrared detection for 40 minutes, internal resistance retest for 20 minutes, data archiving for 20 minutes, total time 120 minutes (reserve 20 minutes for unexpected situations).

[0101] 2. For the optimized S-008 (Cluster 3: Winding local overheating) instruction (instruction ID: Cmd-TR001-20251005-002) Maintenance time: October 08, 2025, 01:00-03:00 (Local overheating risk is high, shorten the execution cycle to within 1 month; choose Wednesday morning to avoid weekend maintenance peak to ensure sufficient resources) Inspection content: ① Oil temperature and internal resistance monitoring: continuous monitoring of oil temperature for 30 minutes (record every 5 minutes, ≤55℃), synchronous measurement of internal resistance (change amount ≤0.3mΩ); ② Cooling system inspection: disassemble the cooling fan filter, clean the dust with compressed air (pressure 0.4MPa), test the fan speed (use Instr-012 speed tester, range 0-5000rpm, ≥2700rpm); ③ Winding temperature scanning: infrared thermal imager scans the middle of the winding (local overheating high-risk area), records the temperature distribution (the highest temperature ≤120℃), if the temperature >115℃, apply heat-conducting silicone grease (model ST-80, thermal conductivity 1.2W / (m・K)); ④ System debugging: restart the cooling system, monitor for 10 minutes, confirm that the fan is running normally and the oil temperature is not abnormally high.

[0102] Required resources: based on S-002 resources, add ① speed tester 1, ② compressed air machine 1, ③ heat-conducting silicone grease 1; add 1 mechanical maintenance worker, total time 120 minutes.

[0103] 3. Instruction set integration and verification The final operation and maintenance instruction set contains the above 2 instructions, which needs to pass the "feasibility verification": ① Time verification: confirm that there is no other equipment maintenance plan during the maintenance period, and the power dispatching has approved the power-off permit; ② Resource verification: all equipment is within the calibration validity period, and the personnel scheduling has no conflict; ③ Safety verification: the maintenance process meets the existing specifications, such as power-off, discharge, and electric verification steps are complete. The instruction set is stored in XML format, each instruction contains "ID, device ID, execution time, content, resources, priority, and responsible person" fields, which is convenient for subsequent API transmission.

[0104] Through the API interface, the operation and maintenance instruction set is sent to the operation and maintenance management platform, which automatically executes the corresponding operation and maintenance tasks and monitors the execution status in real time, finally completing the execution and feedback of the operation and maintenance instruction.

[0105] API interface is the core channel connecting "strategy generation system" and "operation and maintenance management platform", which needs to ensure the safety, real-time and reliability of transmission. Here we use RESTful API architecture (consistent with the general standard of industrial system), combined with Token authentication and data encryption, to avoid tampering or leakage of instructions. The operation and maintenance management platform selects Siemens OpcenterExecution (industrial operation and maintenance management system), which supports automatic generation of work orders, real-time tracking of status and data archiving.

[0106] 1. API interface design and instruction sending Interface parameters: Request method is POST, request header contains: ① Authorization: Bearer Token (valid for 24 hours, obtain Token through login interface before each sending, Token format is eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9...); ② Content-Type: application / xml (adapt platform data format); ③ Timeout: 30s (ensure that the timeout retransmission mechanism works). The request body is the instruction set XML data.

[0107] Sending and confirmation mechanism: After the policy generation system sends the instruction, the platform returns the "reception confirmation" within 5 seconds (status code 200 OK, containing the reception timestamp); if the confirmation is not received within the timeout, the system automatically retransmits (up to 3 times, with an interval of 10s), and if the retransmission fails, an email alert is triggered to the operation and maintenance supervisor.

[0108] 2. Operation and maintenance platform task execution and state monitoring Work order generation: After the platform receives the instruction, it automatically parses the XML data and generates 2 operation and maintenance work orders (work order number consistent with instruction ID), assigns them to "operation and maintenance group-G03", and pushes them to relevant personnel through the system APP (e.g. electrician-012 receives "work order Cmd-TR001-20251005-001 will be executed on October 12, 02:00, please prepare the equipment").

[0109] Real-time state tracking: The platform updates the task progress in real time through the "status code", and the status code is defined as follows: ① 00 (preparation): the work order has been assigned, but the personnel have not confirmed; ② 01 (in execution): the personnel have arrived at the scene and started the repair; ③ 02 (data uploading): the repair is completed and the test data is being uploaded; ④ 03 (completed): the data is archived and the work order is closed; ⑤ 99 (abnormal): problems occur during execution (e.g. device failure, data exceeding standard). The status update frequency is 1 minute / second, and the policy generation system can query it in real time through API.

[0110] Abnormal handling: If an abnormality occurs during execution (e.g. the insulation test data of Cmd-TR001-20251005-001 is 95MΩ < 100MΩ), the platform automatically triggers the "abnormal process": ① suspend the work order and send an alert to the detection engineer; ② the engineer reviews the data (eliminate test instrument error), confirms the abnormality, generates a "supplementary instruction" (Cmd-TR001-20251005-003, content is "replace the winding insulation material, execute within 24 hours"), priority is raised to 1; ③ the supplementary instruction is sent according to the same API process, and the original work order is closed after the supplementary instruction is completed.

[0111] 3. Execution feedback and data archiving Feedback content: after the completion of the work order (status code 03), the platform generates an "execution report", including: ① execution result (such as "insulation resistance 125MΩ, qualified; infrared maximum temperature 105℃, qualified"); ② resource consumption (device usage time 120 minutes, personnel working hours 3 person-hours); ③ abnormal record (none / has, including processing process). The report is fed back to the strategy generation system through API, and is archived to the device operation and maintenance file (associated with historical data, such as comparing the resistance change trend last month).

[0112] Effect evaluation: the strategy generation system updates the historical effect library of reinforcement learning based on the feedback data (such as after the execution of Cmd-TR001-20251005-001, the resistance monthly change of cluster 2 is reduced to 0.2mΩ, and the reward function score is increased by 5 points again), which provides data support for the next round of strategy optimization.

[0113] Through the above process, a closed loop of "instruction sending-execution-monitoring-feedback" is realized, which ensures the accurate landing of operation and maintenance tasks, and finally achieves the safe operation and maintenance goal of "preventing faults, reducing costs and improving device reliability".

[0114] Another embodiment of the present application provides a safe operation and maintenance system driven by internal resistance big data, as shown in Figure 3 , the system can include: The acquisition module 301 is used for acquiring historical and real-time internal resistance data of the target device, and combining multi-dimensional running state parameters to generate a set of internal resistance time series data after time warping through a dynamic time warping algorithm; The construction module 302 is used for multi-scale feature extraction and high-dimensional space mapping of the internal resistance time series data set, constructing an internal resistance feature vector set, and identifying internal resistance abnormal mode clusters hidden in the data distribution by using an adaptive density clustering algorithm; The analysis module 303 is used for multi-modal fault correlation analysis based on the internal resistance abnormal mode clusters, combining device operation history and working condition labels to predict the evolution path and risk level of potential faults, and generating a fault prediction graph; The operation and maintenance module 304 is used for generating a set of differentiated operation and maintenance strategies according to the fault prediction graph, and driving the operation and maintenance management platform to execute corresponding operation and maintenance instructions.

[0115] The above embodiments according to the drawings explain the structure, features and effects of the present application in detail, and the above description is only the preferred embodiment of the present application, but the present application is not limited by the drawings shown, any change or modification made according to the idea of the present application, or equivalent embodiment with equivalent changes, as long as it is within the scope of the present application.

Claims

1. A security operation and maintenance method driven by big data with internal resistance, characterized in that, The method includes: Historical and real-time internal resistance data of the target device are collected and combined with multi-dimensional operating status parameters to generate a normalized internal resistance time series data set through a dynamic time warping algorithm; Multi-scale feature extraction and high-dimensional space mapping are performed on the internal resistance time series data set to construct an internal resistance feature vector set, and an adaptive density clustering algorithm is used to identify the internal resistance anomaly pattern clusters hidden in the data distribution. Based on the aforementioned internal resistance anomaly pattern cluster, multimodal fault correlation analysis is performed. Combined with equipment operating history and operating condition labels, the evolution path and risk level of potential faults are predicted, and a fault prediction map is generated. A differentiated set of operation and maintenance strategies is generated based on the fault prediction map, and the operation and maintenance management platform is driven to execute the corresponding operation and maintenance instructions.

2. The method according to claim 1, characterized in that, The historical and real-time internal resistance data of the target device are collected, and combined with multi-dimensional operating status parameters, a normalized internal resistance time series data set is generated through a dynamic time warping algorithm, including: Through distributed data acquisition nodes, the internal resistance data of the target device is collected in real time, and at the same time, multi-dimensional operating status parameters including at least voltage, current and temperature are collected to generate multi-source raw monitoring datasets. Data cleaning and outlier handling are performed on the multi-source raw monitoring dataset. A sliding window filtering algorithm is used to smooth data fluctuations and generate a preprocessed multi-dimensional monitoring dataset. Based on the preprocessed multidimensional monitoring dataset, a dynamic time warping algorithm is used to align the internal resistance data and operating status parameters at different sampling frequencies, eliminate time scale differences, and generate a time-aligned multidimensional data sequence. The time-aligned multidimensional data sequence is normalized to unify the internal resistance data and various operating state parameters to the same dimension, and finally a normalized internal resistance time series data set is generated.

3. The method according to claim 2, characterized in that, The process involves multi-scale feature extraction and high-dimensional spatial mapping of the internal resistance time-series data set to construct an internal resistance feature vector set. An adaptive density clustering algorithm is then used to identify internal resistance anomaly pattern clusters hidden within the data distribution, including: Multi-scale wavelet transform is performed on the normalized internal resistance time series data set to extract time-domain and frequency-domain features and generate multi-scale feature vectors; An autoencoder is used to perform high-dimensional spatial mapping on multi-scale feature vectors, and deep feature representations are extracted through the encoder to generate high-dimensional feature representation vectors; A set of internal resistance feature vectors is constructed based on high-dimensional feature representation vectors. Principal component analysis algorithm is used to reduce the dimensionality of the features, retaining the main feature components, and generating a set of dimensionality-reduced feature vectors. An adaptive density clustering algorithm is designed to automatically adjust clustering parameters based on data distribution characteristics, perform clustering analysis on the dimensionality-reduced feature vector set, and generate preliminary clustering results. Anomaly pattern identification is performed on the preliminary clustering results. Clusters that are clearly separated from other clusters are extracted as internal resistance anomaly pattern clusters, and finally, the internal resistance anomaly pattern cluster identifiers are output.

4. The method according to claim 3, characterized in that, The multimodal fault correlation analysis based on the internal resistance anomaly pattern cluster, combined with equipment operating history and operating condition labels, predicts the evolution path and risk level of potential faults, and generates a fault prediction map, including: Extract feature statistics from the internal resistance anomaly pattern cluster, including at least the distribution range, density, and outlier degree of data points within the cluster, and generate an anomaly pattern feature description; By combining fault records and operating condition labels from equipment operation history data, a fault association model is constructed using a graph neural network to analyze the correlation between abnormal patterns and fault types and generate fault association analysis results. Based on the results of fault correlation analysis, a hidden Markov model is used to predict the evolution path of potential faults, calculate the probability of occurrence of each path, and generate fault evolution path prediction results. Based on the failure evolution path prediction results, the risk level of each path is assessed, taking into account the severity of the failure and the probability of occurrence, and a risk level assessment result is generated. By integrating the results of fault correlation analysis, fault evolution path prediction, and risk level assessment, a multidimensional fault prediction map is constructed, and a complete fault prediction map is finally output.

5. The method according to claim 4, characterized in that, The step of generating a differentiated operation and maintenance strategy set based on the fault prediction map and driving the operation and maintenance management platform to execute corresponding operation and maintenance instructions includes: Analyze the risk level and fault type information in the fault prediction map, match the corresponding basic operation and maintenance strategies from the operation and maintenance strategy library, and generate a preliminary operation and maintenance strategy set; Based on real-time equipment operating status and historical maintenance effect data, reinforcement learning algorithms are used to optimize and adjust the initial set of maintenance strategies to generate optimized maintenance strategies. Based on the optimized operation and maintenance strategy, specific executable operation and maintenance instructions are generated. The instructions include maintenance time, maintenance content and required resources, resulting in an operation and maintenance instruction set. The system sends a set of operation and maintenance instructions to the operation and maintenance management platform via the API interface, which then drives the platform to automatically execute the corresponding operation and maintenance tasks and monitors the execution status in real time, ultimately completing the execution and feedback of the operation and maintenance instructions.

6. A security operation and maintenance system driven by big data with internal resistance, characterized in that, The system includes: The acquisition module is used to collect historical and real-time internal resistance data of the target device, and combine it with multi-dimensional operating status parameters to generate a normalized internal resistance time series data set through a dynamic time warping algorithm; The construction module is used to perform multi-scale feature extraction and high-dimensional space mapping on the internal resistance time series data set, construct an internal resistance feature vector set, and use an adaptive density clustering algorithm to identify the internal resistance anomaly pattern clusters hidden in the data distribution. The analysis module is used to perform multimodal fault correlation analysis based on the internal resistance anomaly pattern cluster, combine the equipment operation history and operating condition labels to predict the evolution path and risk level of potential faults, and generate a fault prediction map. The operation and maintenance module is used to generate a differentiated operation and maintenance strategy set based on the fault prediction map and drive the operation and maintenance management platform to execute the corresponding operation and maintenance instructions.

7. The system according to claim 6, characterized in that, The acquisition module is specifically used for: Through distributed data acquisition nodes, the internal resistance data of the target device is collected in real time, and at the same time, multi-dimensional operating status parameters including at least voltage, current and temperature are collected to generate multi-source raw monitoring datasets. Data cleaning and outlier handling are performed on the multi-source raw monitoring dataset. A sliding window filtering algorithm is used to smooth data fluctuations and generate a preprocessed multi-dimensional monitoring dataset. Based on the preprocessed multidimensional monitoring dataset, a dynamic time warping algorithm is used to align the internal resistance data and operating status parameters at different sampling frequencies, eliminate time scale differences, and generate a time-aligned multidimensional data sequence. The time-aligned multidimensional data sequence is normalized to unify the internal resistance data and various operating state parameters to the same dimension, and finally a normalized internal resistance time series data set is generated.

8. The system according to claim 7, characterized in that, The building module is specifically used for: Multi-scale wavelet transform is performed on the normalized internal resistance time series data set to extract time-domain and frequency-domain features and generate multi-scale feature vectors; An autoencoder is used to perform high-dimensional spatial mapping on multi-scale feature vectors, and deep feature representations are extracted through the encoder to generate high-dimensional feature representation vectors; A set of internal resistance feature vectors is constructed based on high-dimensional feature representation vectors. Principal component analysis algorithm is used to reduce the dimensionality of the features, retaining the main feature components, and generating a set of dimensionality-reduced feature vectors. An adaptive density clustering algorithm is designed to automatically adjust clustering parameters based on data distribution characteristics, perform clustering analysis on the dimensionality-reduced feature vector set, and generate preliminary clustering results. Anomaly pattern identification is performed on the preliminary clustering results. Clusters that are clearly separated from other clusters are extracted as internal resistance anomaly pattern clusters, and finally, the internal resistance anomaly pattern cluster identifiers are output.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.

Images