Intelligent communication power supply current fault classification method and system

Abstract

The invention discloses an intelligent communication power supply current fault classification method. The method comprises the following steps: acquiring original current data of a communication power supply system; preprocessing the original current data; extracting multi-dimensional features from the preprocessed current data; dimensionality reduction and optimization are carried out on the extracted multi-dimensional features by using a pre-trained enhanced Gap-DBN auto-encoder; a hierarchical XGBoost classification model is used to carry out current fault classification on the processed features, and a three-level classification architecture is used to refine classification results step by step. The invention further provides a corresponding system. By implementing the invention, the response timeliness of the current fault in the communication power supply system can be improved, and the accuracy of fault classification is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication power supply fault classification, and particularly relates to an intelligent communication power supply current fault classification method and system. BACKGROUND

[0002] With the rapid development of communication technology and the continuous improvement of network infrastructure, communication equipment puts forward higher requirements for the stability and intelligent management of power supply systems. As an important guarantee for the normal operation of communication systems, the health status of communication power supply systems directly affects the reliability and service quality of communication networks. In actual operation, various current faults are often caused by overload, short circuit, power module aging, power grid fluctuation and other factors. If these faults cannot be identified and classified in a timely and accurate manner, it will not only cause abnormality or even damage to communication equipment, but also may cause communication interruption, resulting in serious economic losses and social impact.

[0003] In the prior art, traditional current fault detection mostly relies on fixed threshold judgment, manual experience or simple alarm mechanism. The threshold set by the fixed threshold judgment method is usually fixed, which is difficult to adapt to the dynamic working characteristics of communication power supply under different loads and different environments. For slowly changing fault types such as gradually increasing faults and periodic changing faults, it is difficult for the fixed threshold to discover them in time. The manual experience judgment method is limited by the completeness and timeliness of related knowledge, and is prone to misjudgment and omission, which cannot meet the needs of the rapid development of power systems. The method based on support vector machine learns the mapping relationship between features and fault types from historical data, which usually relies on manually extracted shallow features. For current faults which are nonlinear, high-dimensional and time-varying, shallow features are difficult to fully capture the deep discriminant information. The existing deep learning method is limited by the traditional deep neural network which lacks optimization design specifically for current fault features, and it is difficult to effectively extract key discriminant information from current signals.

[0004] In the actual communication power supply operating environment, the existing technology cannot adapt to the diversified, nonlinear and complex changing current fault characteristics in the power supply system, and is prone to false positives and false negatives. Especially for fluctuating, periodic or strong hidden fault types, the recognition effect is poor, and it is difficult to deeply mine the complex patterns contained in the current signals. The recognition and classification ability is insufficient for the scene where multiple faults coexist or interact. In addition, the existing technology also faces the problem of lack of standardized fault data set. Traditional research mostly relies on laboratory simulation data or single scene data, and lacks comprehensive representation of diversified fault patterns of communication power supply systems in complex actual environments.

[0005] Therefore, there is an urgent need for a method that can efficiently and accurately classify current anomalies in communication power systems to improve the system's fault response capability and operational safety, and achieve early warning of fault conditions and intelligent operation and maintenance management. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide an intelligent communication power supply current fault classification method and system, which can improve the timeliness of response to current faults in the communication power supply system and improve the accuracy of fault classification.

[0007] To solve the above technical problems, as one aspect of the present invention, a method for classifying current faults of an intelligent communication power supply is provided, which comprises the following steps:

[0008] Collect raw current data of communication power supply system;

[0009] Preprocessing the raw current data includes filling missing values, data transformation, and standardization;

[0010] Extract multidimensional features from the preprocessed current data, including basic statistical features, differential features, movement statistics, phase relationship features, waveform characteristic factors, and frequency domain features;

[0011] Use the pre-trained enhanced Gap-DBN autoencoder to reduce the dimensionality and optimize the extracted multi-dimensional features;

[0012] A hierarchical XGBoost classification model is applied to the processed features to classify current faults. A three-level classification architecture is used to refine the classification results step by step, and the fault type and its probability distribution are output.

[0013] Preferably, the missing value filling adopts column mean filling; the data transformation adopts Yeo-Johnson transformation; and the standardization processing adopts Z-score standardization.

[0014] Preferably, the enhanced Gap-DBN autoencoder includes:

[0015] The encoder and decoder adopt a symmetrical structure, including input layer, encoder hidden layer and decoder hidden layer;

[0016] A BatchNormalization layer and a Dropout layer are added after each hidden layer. The Dropout rate of the Dropout layer is 0.3 to stabilize the training process and prevent overfitting.

[0017] Introducing the Gap mechanism to optimize feature discriminability by maximizing inter-class distance and minimizing intra-class distance;

[0018] The training is performed using a ReLU activation function, an Adam optimizer, and an early stopping strategy.

[0019] Preferably, in the hierarchical XGBoost classification model, a three-level classification architecture is included, wherein:

[0020] The first-level classifier is used to distinguish the intermediate five types of faults from the non-intermediate five types of faults;

[0021] The second-level classifier performs fine classification on the intermediate five types of faults;

[0022] The third-level classifier handles the non-intermediate five types of faults;

[0023] The SMOTE algorithm is used to up-sample the intermediate five types of fault samples by 2.5 times to balance the data.

[0024] Preferably, the intermediate five types of faults include gradually increasing faults, sharp peak faults, continuously low faults, periodic change faults, and leakage current abnormal faults, and the non-intermediate five types of faults are sudden increase faults and three-phase imbalance faults.

[0025] Correspondingly, as another aspect of the present application, an intelligent communication power supply current fault classification system is also provided, which at least includes:

[0026] A data acquisition module for acquiring raw current data of a communication power supply system;

[0027] A preprocessing module for preprocessing the raw current data, including missing value filling, data transformation, and standardization processing;

[0028] A feature extraction module for extracting multi-dimensional features from the preprocessed current data;

[0029] A feature optimization module for reducing and optimizing the extracted multi-dimensional features using a pre-trained enhanced Gap-DBN autoencoder;

[0030] A classification processing module for classifying the processed features using a hierarchical XGBoost classification model, adopting a three-level classification architecture to gradually refine the classification results, and outputting the fault type and its probability distribution.

[0031] Preferably, the preprocessing module includes:

[0032] A missing value filling unit for filling missing values in the raw current data using a column mean filling method;

[0033] A data transformation unit for transforming the raw current data using a Yeo-Johnson transformation method;

[0034] A standardization unit is configured to standardize the original current data by using Z-score standardization.

[0035] Preferably, the features extracted by the feature extraction module include:

[0036] Time domain features: original value, absolute value, first-order difference, second-order difference, moving average;

[0037] Inter-phase features: inter-phase difference and its absolute value;

[0038] Waveform features: peak factor, waveform factor;

[0039] Frequency domain features: amplitudes of the first three frequency components extracted by FFT.

[0040] Preferably, the feature optimization module includes:

[0041] The Gap-DBN autoencoder has a symmetric structure of an encoder and a decoder, including an input layer, an encoder hidden layer, and a decoder hidden layer; and a Batch Normalization layer and a Dropout layer are added after each hidden layer, and the Dropout rate of the Dropout layer is 0.3, to stabilize the training process and prevent overfitting;

[0042] The Gap mechanism unit is configured to maximize the inter-class distance and minimize the intra-class distance.

[0043] The training unit adopts an Adam optimizer, a ReLU activation function, and an early stopping strategy.

[0044] Preferably, the classification processing module is deployed on a cloud server or an on-site edge computing node, and is kept in data synchronization with the data acquisition module through an industrial Ethernet, a 5G network, or a dedicated communication link, to realize a collaborative diagnostic service combining local rapid fault response and cloud intelligent decision support, wherein the classification processing module includes:

[0045] A first-level XGBoost classifier is configured to distinguish between intermediate five-class faults and non-intermediate five-class faults, wherein the intermediate five-class faults include a gradually increasing fault, a sharp peak fault, a continuous low fault, a periodic change fault, and a leakage current abnormal fault, and the non-intermediate five-class faults are a sudden increase fault and a three-phase imbalance fault.

[0046] A second-level XGBoost classifier is configured to finely classify the intermediate five-class faults.

[0047] A third-level XGBoost classifier is configured to classify the non-intermediate five-class faults.

[0048] A data balancing unit is configured to perform 2.5 times up-sampling on the intermediate five-class fault samples by using a SMOTE algorithm.

[0049] Implementing the embodiments of the present application has the following beneficial effects:

[0050] The present application provides a kind of intelligent communication power supply current fault classification method and system.Through multi-dimensional feature extraction and enhanced Gap-DBN self-encoder deep feature learning, combined with three-level hierarchical XGBoost classification strategy, in fault diagnosis accuracy aspect, seven kinds of current fault types (sudden increase type fault, gradually increasing type fault, peak type fault, sustained low type fault, periodic change type fault, leakage current abnormal type fault and three-phase unbalanced type fault) can be accurately distinguished, especially for the identification accuracy of the middle five kinds of faults (gradually increasing type fault, peak type fault, sustained low type fault, periodic change type fault, leakage current abnormal type fault) significantly improved, compared with traditional feature extraction method, class separation degree and feature discriminability are greatly improved, especially solve the technical problem that the middle five kinds of similar faults are difficult to distinguish, class distance is increased by 2.3 times, provide key technical support for realizing high-precision fault classification;

[0051] In the robustness aspect, SMOTE data balance, feature selection and Gap-DBN deep feature optimization technology are used, which effectively suppresses the influence of noise, abnormal points and sample imbalance, so that the model can maintain stable fault recognition performance under different working environments and data distribution, and has excellent generalization ability; compared with traditional method, the class separation degree is improved by 65%, especially the ability to distinguish the middle five kinds of similar faults is improved by 2.3 times.

[0052] In terms of adaptability to complex environment, the method can automatically mine hidden patterns and multi-level features in current signal, support efficient processing of complex situations such as coexistence of multiple faults, weak faults and sudden faults, and realize fine identification of similar fault types through three-level hierarchical classification mechanism. At the same time, modular design makes it easy to extend, which can adapt to more complex application scenarios by adding new fault types and features.

[0053] In summary, by implementing the present application, the response timeliness of current fault in communication power supply system can be improved, and the accuracy of fault classification can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings obtained from these drawings without creative labor are still within the scope of the present application.

[0055] Figure 1A schematic diagram of the main process of an embodiment of a method for classifying current faults of an intelligent communication power supply provided by the present invention;

[0056] Figure 2 for Figure 1 A more detailed flowchart is shown in the figure;

[0057] Figure 3 A waveform diagram of a sudden increase in fault involved in the present invention;

[0058] Figure 4 A waveform diagram of a gradually increasing fault involved in the present invention;

[0059] Figure 5 A waveform diagram of a spike-type fault involved in the present invention;

[0060] Figure 6 Schematic diagram of the waveform of the continuous low-type fault involved in the present invention;

[0061] Figure 7 A waveform diagram of a periodic variant fault according to the present invention;

[0062] Figure 8 A waveform diagram of an abnormal leakage current fault according to the present invention;

[0063] Figure 9 A waveform diagram of a three-phase unbalanced fault according to the present invention;

[0064] Figure 10 A schematic structural diagram of an embodiment of an intelligent communication power supply current fault classification system provided by the present invention;

[0065] Figure 11 for Figure 10 Schematic diagram of the structure of the preprocessing module in;

[0066] Figure 12 A comparison diagram of classification effects in an example of the present invention compared with the prior art;

[0067] Figure 13 This is a performance comparison chart of fault classification using traditional XGBoost in an example of the present invention. DETAILED DESCRIPTION

[0068] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be described in further detail below with reference to the accompanying drawings.

[0069] like Figure 1 FIG. 1 is a schematic diagram showing the main flow of an embodiment of a method for classifying current faults of an intelligent communication power supply provided by the present invention. Figures 2 to 9As shown, in the present embodiment, the method comprises the following steps:

[0070] Step S1, collecting original current data of the communication power supply system to form an original data matrix; specifically, the existing power supply monitoring system of the communication base station, data center and communication room can be directly integrated, high-precision Hall effect current sensors are deployed at key nodes such as power supply feeders, UPS systems and power distribution units, and 24-hour uninterrupted real-time collection of the running state of the power supply system is realized.

[0071] Step S2, preprocessing the original current data, the preprocessing including missing value filling, data transformation and standardization processing;

[0072] Wherein, the missing value filling adopts column mean filling; the specific mode is as follows:

[0073]

[0074] Wherein, n' represents the number of non-missing values in the jth column.

[0075] The data transformation adopts Yeo-Johnson transformation, that is, the skewed data is processed by Yeo-Johnson transformation:

[0076]

[0077] Wherein, the optimal transformation parameter λ is determined by maximizing the log-likelihood function as follows:

[0078]

[0079] The standardization processing adopts Z-score standardization mode, and the formula is as follows:

[0080]

[0081] Wherein, is the variance of the transformed data,

[0082] is the mean of the transformed data, μ is the characteristic mean, and σ is the characteristic standard deviation.

[0083] Step S3, extracting multi-dimensional features from the preprocessed current data, including basic statistical features, difference features, moving statistics, inter-phase relationship features, waveform feature factors and frequency domain features;

[0084] Specifically, the features are extracted from the current data, and the basic statistical features are extracted from the current data:

[0085] Original value: f1=x i

[0086] Absolute value: f2 = |x i |

[0087] Calculate first-order difference features and second-order difference features to capture sudden faults and spike faults.

[0088] First-order difference: f3 = x i -x i-1

[0089] Absolute value of the first-order difference: f4=|x i -x i-1 |

[0090] Second order difference: f5=(x i -x i-1 )-(x i-1 -x i-2 )

[0091] Absolute value of second-order difference: f6=|f5|

[0092] Calculate the moving average of different window sizes to identify periodic and persistent low-level faults. For window size w∈{3,5}, the moving average is calculated as:

[0093]

[0094] Calculate the phase difference and its absolute value, specifically for three-phase unbalance fault and leakage current fault:

[0095] Phase difference:

[0096] Absolute value of phase difference: f 10 =|f9|,f 10′ =|f 9′ |,f 10″ =|f 9″ |

[0097] Calculate waveform characteristic factors to enhance the ability to identify specific fault types:

[0098] Crest Factor: Where σ is the standard deviation

[0099] Form Factor:

[0100] Use FFT transform to extract frequency components and optimize the identification of periodic faults:

[0101]

[0102] Take the amplitude of the first three frequency components: f 13,14,15 =|X k|k∈{0,1,2}, the feature matrix is obtained by the above feature extraction.

[0103] Specifically, in combination Figures 3 to 9 The definition and characteristics of the seven types of faults involved in the present application are as follows:

[0104] Sudden increase type fault: the current value suddenly increases greatly in a very short time (usually <5 seconds), the increase is more than 50% of the normal working current, and it shows a step characteristic of jumping from normal state to high value. Key discriminant feature: first-order difference feature f3=x i -x i-1 The maximum value appears at the fault moment, and the second-order difference f5 shows a sharp peak at the mutation point. The waveform of this type of fault is shown in Figure 3 .

[0105] Gradual increase type fault: the current value increases slowly and continuously in a long time (>30 minutes), the growth rate is usually 0.5%-2% / min, and it shows a gradual change trend. Key discriminant feature: moving average feature f 7,8 It shows a monotonous increasing trend, the first-order difference mean is greater than 0, and the time series linear regression slope is positive. The waveform of this type of fault is shown in Figure 4 .

[0106] Peak type fault: a transient impact with very short duration (<1 second) but high amplitude appears in the current signal, usually in the form of narrow pulse. Key discriminant feature: peak value factor f 11 The maximum value is reached at the peak moment, and the second-order difference f5 shows positive and negative alternation before and after the peak, and the frequency characteristic component f 13,14,15 High frequency component is significantly enhanced. The waveform of this type of fault is shown in Figure 5 .

[0107] Continuous low type fault: the current value is lower than the normal working range for a long time, usually more than 20% lower than the rated current, and the duration is more than 10 minutes. Key discriminant feature: the original feature value f1 is lower than the set threshold for a long time, and the moving average f 7,8 The mean value is significantly low, and the wave factor f 12 is relatively increased due to the decrease of the mean value. The waveform of this type of fault is shown in Figure 6 .

[0108] Periodic change type fault: the current signal shows obvious periodic fluctuation, the fluctuation amplitude is usually 10%-30% of the rated value, and the period ranges from several minutes to several hours. Key discriminant feature: frequency domain feature f 13,14,15 There is an obvious peak at a certain frequency, the autocorrelation function detects significant periodicity, and the moving average variance shows a periodic change mode. The waveform of this type of fault is shown in Figure 7 .

[0109] Leakage current abnormal fault: abnormal leakage current path appears in the system, causing normal working current to appear irregular tiny fluctuations and deviations, usually accompanied by abnormal ground current. Key distinguishing features: high-frequency noise components increase random components in the signal, and ground current monitoring shows abnormal leakage. The waveform of this type of fault is shown in Figure 8 .

[0110] Three-phase imbalance fault: there is a significant amplitude or phase imbalance between three-phase currents, with an imbalance degree exceeding 15%, violating the symmetry principle of three-phase systems. Key distinguishing features: the phase-to-phase difference f9 deviates significantly from zero, the absolute value of the phase-to-phase difference f 10 significantly increases, and the three-phase current variance significantly increases. The waveform of this type of fault is shown in Figure 9 .

[0111] Step S4, using the pre-trained enhanced Gap-DBN autoencoder to reduce and optimize the extracted multi-dimensional features;

[0112] The enhanced Gap-DBN autoencoder comprises:

[0113] An encoder-decoder symmetric structure is adopted, including an input layer, an encoder hidden layer, and a decoder hidden layer;

[0114] A BatchNormalization layer and a Dropout layer are added after each hidden layer, and the Dropout rate of the Dropout layer is 0.3, to stabilize the training process and prevent overfitting;

[0115] The Gap mechanism is introduced to optimize the feature discriminability by maximizing the inter-class distance and minimizing the intra-class distance;

[0116] ReLU activation function, Adam optimizer and early stopping strategy are used for training.

[0117] It can be understood that in the method of the application, the enhanced Gap-DBN autoencoder adopts a symmetric encoder-decoder structure, and the greatest innovation is the introduction of the "Gap" mechanism, which is a feature gap optimization technology specially designed for the problem of distinguishing similar fault types. By maximizing the distance between different fault types in the feature space, the most discriminative feature representation is automatically learned, which not only enlarges the inter-class distance, but also ensures the compactness of the features of the same class, and can dynamically optimize the classification boundary during training, especially for the fine processing of the five similar fault types in the middle.

[0118] Mathematical definition of Gap mechanism:

[0119] Feature gap: Gap ij =||z i -zj ||2

[0120] Intra-class compactness measure:

[0121] Inter-class separation measure:

[0122] Gap optimization objective:

[0123] where is the feature center of the c-th fault class.

[0124] The BatchNormalization layer stabilizes the training process by normalizing the input distribution of each mini-batch, which sequentially calculates batch statistics, normalization, scaling and shifting:

[0125]

[0126] where μ1 and are the batch mean and variance, respectively, is a small constant to prevent division by zero, and γ and β are learnable scaling and shifting coefficients. It solves the internal covariate shift problem in deep network training, enabling the network to stably learn the complex nonlinear features of the current fault.

[0127] The Dropout layer prevents overfitting by randomly inactivating neurons, mathematically represented as:

[0128]

[0129] mask Bernoulli(1-p) is a random mask following Bernoulli distribution, and p=0.3 is the inactivation probability. During the training phase, each neuron is set to 0 with a probability of 30%, and the output of the remaining neurons is scaled by , and during the testing phase, all neurons participate in the calculation without random inactivation. The Dropout rate of 0.3 is a validated effective value in deep learning, which provides sufficient regularization effect without excessively affecting the network learning ability.

[0130] The encoding process of the Gap-DBN mathematical model can be represented as:

[0131] h1=σ(W1x+b1)

[0132] h2=σ(W2h1+b2)

[0133] h3=σ(W3h2+b3)

[0134] The decoding process can be represented as:

[0135]

[0136] where σ denotes the ReLU activation function: σ(z) = max(0, z), W i and b i are the weight matrix and bias vector, respectively. Through multiple layers of ReLU activation functions, complex nonlinear feature transformations are achieved, hidden patterns that are difficult to identify by traditional methods are automatically discovered, BatchNormalization suppresses noise interference, Dropout enhances generalization ability, and the model can stably extract fault features in complex power environments.

[0137] The autoencoder is trained using the Adam optimizer, the learning rate is dynamically adjusted using the ReduceLROnPlateau strategy, and the early stopping strategy is used to prevent overfitting. After training is complete, the encoder part is used to extract optimized features:

[0138] X2 = Encoder(X1) ∈ R n×64 .

[0139] More specifically, the SelectFromModel method is used to select the most important feature subset, and the threshold is set to the median:

[0140] Threshold = Median(Importance(f i )), i ∈ {1, 2,..., d}

[0141] Features with importance greater than the threshold are retained:

[0142] X3 = X2[:, Importance(f i )>Threshold].

[0143] Seven types of current fault are labeled and encoded, including sudden increase fault (label 0), gradual increase fault (label 1), sharp peak fault (label 2), persistent low fault (label 3), periodic change fault (label 4), leakage current abnormality fault (label 5), and three-phase imbalance fault (label 6). LabelEncoder is used for label encoding to ensure that the label starts from 0 and is continuous;

[0144] The distribution of each type of fault sample is calculated:

[0145]

[0146] where N c represents the number of samples of class c, and N represents the total number of samples.

[0147] Set a higher target sampling number for the intermediate five types of faults (labels 1-5):

[0148]

[0149] Step S5, apply the hierarchical XGBoost classification model to the processed features for current fault classification, adopt a three-level classification architecture to gradually refine the classification results, and output the fault type and its probability distribution.

[0150] Among the hierarchical XGBoost classification model, a three-level classification architecture is included, wherein:

[0151] The first-level classifier is used to distinguish between the intermediate five types of faults and the non-intermediate five types of faults; the intermediate five types of faults include gradually increasing faults, sharp peak faults, persistent low faults, periodic change faults, and leakage current abnormal faults, and the non-intermediate five types of faults are sudden increase faults and three-phase imbalance faults.

[0152] The second-level classifier performs fine classification on the intermediate five types of faults.

[0153] The third-level classifier processes the non-intermediate five types of faults.

[0154] Use the SMOTE algorithm to oversample the intermediate five types of fault samples by 2.5 times to balance the data.

[0155] Specifically, in this step, synthetic samples are generated by the SMOTE algorithm to fill the gaps of the original samples, the generated new sample feature values are within a reasonable range, and the feature patterns of the original classes are maintained, which is a smooth expansion relative to the original classes. After SMOTE enhancement, the data quantity difference between different classes of samples is reduced, and the number of samples of each class tends to be balanced.

[0156] Specifically, for the classes that need to increase samples, synthetic samples are generated using the SMOTE algorithm. The principle of the SMOTE algorithm is to perform linear interpolation between minority class samples:

[0157] 1. For a minority class sample x i , randomly select a sample x3

[0158] 2. Generate a synthetic sample: x4=x i +λ×(x3-x i ), where λ∈[0,1] is a random number.

[0159] k-neighbor selection: Given a minority class sample set S={x1,x2,...,x m}, for a sample x i , its k-neighbor set is:

[0160] NNk (x i )={x j1 ,x j2 ,...,x jk},d(x i ,x j1 )≤d(x i ,x j2 )≤...≤d(x i ,x jk )

[0161] Among them, d(x i ,x j ) is the Euclidean distance calculation:

[0162] The hierarchical XGBoost classification model adopts a three-layer classification architecture, especially for the middle five categories of faults. The first-level classifier is used to distinguish between the middle five categories of faults (labels 1 to 5) and the non-middle five categories of faults (labels 0, 6). The second-level classifier is specifically used to finely classify the middle five categories of faults. The third-level classifier is used to process the non-middle five categories of faults.

[0163] It is understandable that XGBoost is an efficient gradient boosting decision tree algorithm that achieves strong classification performance by integrating multiple weak learners. For the tth iteration, the objective function of XGBoost is:

[0164]

[0165] Cumulative forecast:

[0166] Regularization term:

[0167] in is the loss function, f t (x i ) is the predicted value of the t-th tree, T is the number of leaf nodes, w j is the weight of the jth leaf node, γ is the penalty coefficient for the number of leaf nodes, and λ is the L2 regularization coefficient for the leaf weight. A second-order Taylor expansion approximation function is used to improve optimization efficiency. The final classification result is obtained through fine classification using a three-level hierarchical XGBoost classifier.

[0168] Specifically, in this step, we first build a first-level classifier to distinguish the middle five categories from the non-middle five categories:

[0169]

[0170] Set higher sample weights for the middle five categories:

[0171]

[0172] Then the second level special classifier is constructed, which is specially aimed at the intermediate five types of faults:

[0173] X5=X4[y level1 =1]

[0174] y2=y1[y level1 =1]-1(remap to 0~4)

[0175] Finally, the third level classifier is constructed, which is aimed at the non-intermediate five types of faults:

[0176] X6=X4[y level1 =0]

[0177] y3=y1[y level1 =0]

[0178] The non-intermediate labels are remapped:

[0179] y4=LabelMap(y3)

[0180] Then the remapped labels are output corresponding to the fault types, and the final classification result is obtained. The advantages and disadvantages of the classification result are judged by using the classification accuracy, precision, recall and F1 score.

[0181] In summary, in actual examples, the specific steps of using the method of the application to classify the newly collected current data generally include the following steps:

[0182] First, the new current data is feature extracted to obtain a feature vector;

[0183] Then the trained enhanced Gap-DBN encoder is used to optimize the features to obtain the deep feature representation after dimension reduction;

[0184] Then the important feature subset is screened out;

[0185] Subsequently, the first level classifier is used to determine whether the sample belongs to the intermediate five types of faults (gradually increasing type, sharp peak type, continuous low type, periodic change type, leakage current abnormal type);

[0186] Finally, according to the first level classification result, if it belongs to the intermediate five types of faults, the second level special classifier is called for fine classification, and if it belongs to the non-intermediate five types of faults, the third level classifier is used for prediction, and the final fault type and its probability distribution are output.

[0187] It can be understood that the embodiment of the application discloses an intelligent communication power current fault classification method based on Gap-DBN-XGBoost, which effectively solves the problem that similar fault types are difficult to distinguish in current fault classification, especially the identification of the middle five types of faults (gradually increasing type, sharp peak type, continuous low type, periodic change type and leakage current abnormal type fault). Compared with the traditional classification algorithm, deep feature representation can be extracted by the enhanced Gap-DBN autoencoder, combined with a three-level hierarchical classification strategy and a differentiated sampling technology, and seven common current fault types can be accurately identified in a complex power environment. The method realizes rapid fault identification response, helps the fault warning and accurate maintenance of the communication power supply system, significantly reduces the system downtime and maintenance cost, and improves the reliability of the communication network.

[0188] As shown in Figure 10 , a structural schematic diagram of one embodiment of an intelligent communication power current fault classification system provided by the application is shown. In combination with Figure 11 , in the embodiment, the intelligent communication power current fault classification system 1 at least includes:

[0189] A data acquisition module 10 is configured to acquire original current data of a communication power supply system.

[0190] A preprocessing module 11 is configured to preprocess the original current data, including missing value filling, data transformation and standardization processing.

[0191] A feature extraction module 12 is configured to extract multi-dimensional features from the preprocessed current data.

[0192] A feature optimization module 13 is configured to use a pre-trained enhanced Gap-DBN autoencoder to reduce and optimize the extracted multi-dimensional features.

[0193] A classification processing module 14 is configured to apply a hierarchical XGBoost classification model to classify the processed features, adopt a three-level classification architecture to gradually refine the classification results, and output the fault type and its probability distribution.

[0194] As shown in Figure 11 , in one specific example, the preprocessing module 11 includes:

[0195] A missing value filling unit 110 is configured to fill the missing values of the original current data by using the column mean filling method.

[0196] A data transformation unit 111 is configured to transform the original current data by using the Yeo-Johnson transformation method to process the skew data.

[0197] The standardization unit 112 is configured to standardize the original current data by using Z-score standardization.

[0198] In specific examples, the features extracted by the feature extraction module 12 include:

[0199] Time domain features: original value, absolute value, first-order difference, second-order difference, moving average;

[0200] Inter-phase features: inter-phase difference and its absolute value;

[0201] Waveform features: peak factor, waveform factor;

[0202] Frequency domain features: amplitudes of the first three frequency components extracted by FFT.

[0203] In specific examples, the feature optimization module 13 includes:

[0204] The Gap-DBN autoencoder has a symmetrical structure of an encoder and a decoder, including an input layer, an encoder hidden layer, and a decoder hidden layer; and a Batch Normalization layer and a Dropout layer are added after each hidden layer, and the Dropout rate of the Dropout layer is 0.3, to stabilize the training process and prevent overfitting;

[0205] The Gap mechanism unit is configured to maximize the inter-class distance and minimize the intra-class distance.

[0206] The training unit adopts an Adam optimizer, a ReLU activation function, and an early stopping strategy.

[0207] In specific examples, the classification processing module 14 is deployed on a cloud server or an on-site edge computing node, and is kept in data synchronization with the data acquisition module through an industrial Ethernet, a 5G network, or a dedicated communication link, to realize a collaborative diagnosis service combining local rapid fault response and cloud intelligent decision support, wherein the classification processing module includes:

[0208] The first-level XGBoost classifier is configured to distinguish between intermediate five-class faults and non-intermediate five-class faults, wherein the intermediate five-class faults include a gradually increasing fault, a sharp peak fault, a sustained low fault, a periodic change fault, and a leakage current anomaly fault, and the non-intermediate five-class faults are a sudden increase fault and a three-phase imbalance fault.

[0209] The second-level XGBoost classifier is configured to finely classify the intermediate five-class faults.

[0210] The third-level XGBoost classifier is configured to classify the non-intermediate five-class faults.

[0211] The data balancing unit performs 2.5 times up-sampling on the intermediate five types of fault samples by using an SMOTE algorithm.

[0212] For more details, refer to the foregoing description of the Figures 1 to 9 , which will not be repeated here.

[0213] In order to further understand the beneficial effects of the method provided by the present application, the following will combine specific use examples to experimentally illustrate the method described in the present application.

[0214] The experimental content is: collecting the current fault data of the communication base station power supply as the original data for the experiment through the communication power supply monitoring system. This experiment compares the classification effect of the intelligent communication power supply current fault classification method based on Gap-DBN-XGBoost described in the embodiment of the present application with four kinds of mainstream classification methods such as LightGBM, TabNet, BERT and CatBoost, and compares it with the traditional XGBoost single classifier method, analyzes the fault recognition accuracy of classification, and takes the classification accuracy, precision, recall rate, F1 score and classification time as evaluation indexes. The data used in the experiment of the present application is derived from the high-precision current data of the actual operating environment of multiple communication base stations in a certain place in China from 2024 to 2025.

[0215] The same data set, preprocessing process and evaluation index are used to configure the optimal parameters for the two types of methods, avoid affecting the comparison results due to improper parameter setting, extract the current data of the fault type from the current data of a certain place in China from 2024 to 2025 to form the required fault current data set, containing 100,000 sample data, 10,000 fault samples, the training set accounts for 70%, and the test set accounts for 30%. The computer used for comparison test uses Win10 system, Intel(R) Core(TM) i7-10875H CPU@2.30GHz processor, NVIDIA GeForce RTX 2060(6GB), Intel(R) UHD Graphics(128MB) graphics card, and python 3.12.9 environment under 64-bit operating system to test the two types of methods respectively.

[0216] The Gap-DBN autoencoder parameter settings of the method of the present application are as follows: the input neurons of the network structure are set to 45, the neurons of each layer of the encoder are set to [256, 128, 64], the decreasing compression structure, the neurons of each layer of the decoder are set to [128, 256], the symmetric reconstruction structure, the Dropout layer inactivation probability is 0.3, the number of each small batch of samples is set to 64, the maximum training round is 50, the initial learning rate of the Adma optimizer is 0.001, the weight coefficient of the Gap loss function is set to 0.1, and the L2 regularization weight coefficient is set to 0.001.

[0217] Three-level hierarchical XGBoost parameter settings: first-level classifier parameters: decision tree maximum depth 7, sample sampling ratio 0.8, feature sampling ratio 0.8, integrated tree number 300, binary classification; second-level classifier parameters: decision tree maximum depth 8, sample sampling ratio 0.9, more tree number 500, five classification; third-level classifier: decision tree maximum depth 7, standard tree number 300, binary classification. Traditional XGBoost parameters: decision tree maximum depth 8, uniform tree number 500, seven classification.

[0218] The specific implementation process of the method of the application applied to the communication base station power supply current fault data classification is as follows:

[0219] Step one: collect the current fault data of the communication power supply as the original data, including seven fault types (sudden increase type fault, gradual increase type fault, sharp peak type fault, continuous low type fault, periodic change type fault, leakage current abnormal type fault, and three-phase unbalanced type fault), and pre-process the data.

[0220] Step two: extract optimized features from the processed data to obtain a series of related features.

[0221] Step three: build an enhanced Gap-DBN autoencoder, reduce and optimize the extracted features through a multi-layer encoding-decoding structure, use the Adam optimizer for training, and finally obtain the optimized feature representation.

[0222] Step four: use the XGBoost-based feature selection method to evaluate the feature importance, and use the median as the threshold to select the most important feature subset; at the same time, the training data is subjected to SMOTE imbalance processing, and especially the intermediate five types of faults (gradual increase type, sharp peak type, continuous low type, periodic change type and leakage current abnormal type fault) are subjected to 2.5 times up-sampling.

[0223] Step five: build a three-level hierarchical XGBoost classification system, including a first-level classifier to distinguish between intermediate five types of faults and non-intermediate five types of faults, a second-level special classifier for fine classification of the intermediate five types of faults, and a third-level classifier for processing non-intermediate five types of faults.

[0224] Step six: use the hierarchical prediction process to classify the test data, output the fault type and its probability distribution, obtain the final fault classification result, and show the classification effect in the confusion matrix.

[0225] Among them, the classification effect comparison chart is as shown in Figure 12 .

[0226] In order to further verify that the method of the embodiment of the present invention can more accurately identify communication power supply current faults than the traditional classification algorithm, the traditional XGBoost single classifier method is selected for comparison with the method of the embodiment of the present invention. The classification effect and confusion matrix results of the traditional XGBoost method are slightly worse. The classification accuracy, precision, recall rate, F1 score and classification time are used as evaluation indicators. The results are as follows: Figure 13 shown.

[0227] Analysis of the above experimental results shows that: based on the original current data collected by the communication power supply monitoring system, from the time domain, the current waveform characteristics of different fault types are similar, especially the middle five types of faults (gradual increase type, peak type, continuous low type, periodic change type and leakage current abnormal type fault) are difficult to accurately distinguish under traditional methods; the introduction of the Gap mechanism makes the overall classification accuracy of the method of the embodiment of the present invention reach 87.23%, which is 15.76 percentage points higher than the 71.47% of the traditional XGBoost method, and the macro-average F1 score is improved by 14.77 percentage points; especially for the middle five types of faults that are difficult to distinguish, the method of the present invention uses Gap-DBN feature optimization and hierarchical classification strategy to significantly improve the recognition ability of similar fault types; while the traditional XGBoost single classifier method has a low recognition accuracy for the middle five types of faults and is prone to misclassification. From the confusion matrix analysis, it can be seen that there is obvious inter-class confusion. Therefore, the method of the embodiment of the present invention has a good classification effect for communication power supply current faults in complex power environments.

[0228] The intelligent communication power supply current fault classification method based on Gap-DBN-XGBoost proposed in the embodiment of the present invention can effectively solve the problem of insufficient feature expression through multi-dimensional optimized feature extraction and enhanced Gap-DBN autoencoder. The three-level hierarchical XGBoost classification strategy can significantly improve the recognition accuracy of the five difficult-to-distinguish intermediate faults (gradual increase type, peak type, continuous low type, periodic change type and leakage current abnormality type) compared to the traditional single classifier. The differential sampling technology effectively solves the sample imbalance problem and achieves high-precision recognition of seven common current fault types in a complex power environment. This method fully utilizes the advantages of Gap-DBN in feature dimensionality reduction and optimization through the combination of feature engineering and deep learning, which is helpful for accurate fault diagnosis and preventive maintenance of communication power supply systems, significantly reduces system downtime and maintenance costs, and improves the reliability of communication networks.

[0229] The method of the present application can be applied to communication infrastructure power management, data center power infrastructure; or communication room power distribution system. It can be directly integrated into the existing power monitoring system of the communication base station, data center and communication room, and through the deployment of high-precision Hall effect current sensors at key nodes such as power feeders, UPS systems, power distribution units, etc. 24-hour uninterrupted real-time collection of power system operation status is realized. The system uses Gap-DBN algorithm for deep feature extraction and fault mode recognition on local devices. When an abnormal current is detected, the system immediately triggers the fault classification mechanism and pushes detailed current fault types to the operation and maintenance personnel through the network management platform.

[0230] Further, the method of the present application can also be applied to fields such as: intelligent transformation of power grid system; new energy power generation system; health management of industrial equipment; smart construction and smart home; and intelligent transportation; the lightweight Gap-DBN model can be deployed on the field edge computing node or industrial gateway device, responsible for real-time current data processing, feature extraction and preliminary fault identification, to ensure fast response capability; at the same time, the complete deep learning model and big data analysis platform are deployed on the cloud server, responsible for complex fault mode deep analysis, historical data mining, cross-device correlation analysis and algorithm model continuous optimization and updating tasks. The field device keeps data synchronization with the cloud platform through industrial Ethernet, 5G network or special communication link, realizes the collaborative diagnosis service combining local fast fault response and cloud intelligent decision support.

[0231] Implementing the embodiment of the present application has the following beneficial effects:

[0232] The present application provides an intelligent communication power current fault classification method and system. Through multi-dimensional feature extraction and enhanced Gap-DBN autoencoder deep feature learning, combined with three-level hierarchical XGBoost classification strategy, in terms of fault diagnosis accuracy, seven types of current faults (sudden increase fault, gradual increase fault, peak fault, continuous low fault, periodic change fault, leakage current abnormal fault and three-phase imbalance fault) can be accurately distinguished. Especially for the middle five types of faults (gradual increase fault, peak fault, continuous low fault, periodic change fault, leakage current abnormal fault), the recognition accuracy is significantly improved. Compared with the traditional feature extraction method, the class separation degree and feature discriminability are greatly improved. Especially, it solves the technical problem that the middle five similar faults are difficult to distinguish, the class distance is improved by 2.3 times, which provides key technical support for realizing high-precision fault classification.

[0233] In terms of robustness, the SMOTE data balancing, feature selection and Gap-DBN deep feature optimization technology are adopted, the influence of noise, abnormal points and sample imbalance is effectively inhibited, the model can maintain stable fault recognition performance under different working environments and data distribution, and has excellent generalization ability; compared with the traditional method, the class separation degree is improved by 65%, and the distinguishing ability of the similar faults of the middle five classes is improved by 2.3 times.

[0234] In terms of complex environment adaptability, the method can automatically mine hidden patterns and multi-level features in the current signal, support efficient processing of complex situations such as coexistence of multiple faults, weak faults and burst faults, and realize fine identification of similar fault types through a three-level hierarchical classification mechanism. At the same time, the modular design makes it easy to expand, and can adapt to more complex application scenarios by adding new fault types and features.

[0235] In summary, the implementation of the present application can improve the timeliness of current fault response in the communication power supply system and improve the accuracy of fault classification.

[0236] Those skilled in the art will appreciate that embodiments of the present application can be provided as methods, apparatuses, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0237] The present application is described with reference to flowcharts and / or block diagrams according to the method, device (system) and computer program product of the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flowchart and / or block diagram. Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks

[0238] The above disclosure is only one preferred embodiment of the present application, and of course cannot limit the scope of the rights of the present application, so the equivalent changes made according to the claims of the present application still fall within the scope of the present application.

Claims

1. A method for classifying current faults of intelligent communication power supplies, characterized in that: The following steps are involved: Collect raw current data of communication power supply system; Preprocessing the raw current data includes filling missing values, data transformation, and standardization; Extract multidimensional features from the preprocessed current data, including basic statistical features, differential features, movement statistics, phase relationship features, waveform characteristic factors, and frequency domain features; Use the pre-trained enhanced Gap-DBN autoencoder to reduce the dimensionality and optimize the extracted multi-dimensional features; A hierarchical XGBoost classification model is applied to the processed features to classify current faults. A three-level classification architecture is used to refine the classification results step by step, and the fault type and its probability distribution are output.

2. The method according to claim 1, characterized in that The missing value filling adopts column mean filling; the data transformation adopts Yeo-Johnson transformation; and the standardization processing adopts Z-score standardization.

3. The method according to claim 2, characterized in that The enhanced Gap-DBN autoencoder includes: The encoder and decoder adopt a symmetrical structure, including input layer, encoder hidden layer and decoder hidden layer; A BatchNormalization layer and a Dropout layer are added after each hidden layer. The Dropout rate of the Dropout layer is 0.3 to stabilize the training process and prevent overfitting. Introducing the Gap mechanism to optimize feature discriminability by maximizing inter-class distance and minimizing intra-class distance; The ReLU activation function, Adam optimizer and early stopping strategy are used for training.

4. The method according to claim 3, characterized in that In the hierarchical XGBoost classification model, a three-level classification architecture is included, where: The first-level classifier is used to distinguish the middle five types of faults from the non-middle five types of faults; The second-level classifier performs fine classification on the middle five types of faults; The third-level classifier handles the non-intermediate five types of faults; The SMOTE algorithm is used to upsample the middle five types of fault samples by a factor of 2.5 to balance the data.

5. The method according to claim 4, characterized in that The five intermediate fault types include gradually increasing faults, peak faults, continuously low faults, periodic change faults and abnormal leakage current faults. The five non-intermediate fault types include sudden increasing faults and three-phase unbalanced faults.

6. An intelligent communication power supply current fault classification system, characterized in that: At least: Data acquisition module, used to collect raw current data of the communication power supply system; Preprocessing module, used to preprocess the raw current data, including missing value filling, data transformation and standardization; A feature extraction module is used to extract multidimensional features from the preprocessed current data; Feature optimization module, which uses the pre-trained enhanced Gap-DBN autoencoder to reduce the dimensionality and optimize the extracted multi-dimensional features; The classification processing module applies the hierarchical XGBoost classification model to classify the processed features into current faults, adopts a three-level classification architecture to refine the classification results step by step, and outputs the fault type and its probability distribution.

7. The system according to claim 6, characterized in that The pre-processing module comprises: A missing value filling unit is used to fill missing values ​​in the original current data using a column mean filling method; A data conversion unit is used to perform data conversion on the original current data using the Yeo-Johnson transformation method and process the skew data; The normalization unit is used to normalize the raw current data using Z-score normalization.

8. The system according to claim 7, characterized in that The features extracted by the feature extraction module include: Time domain features: original value, absolute value, first-order difference, second-order difference, moving average; Phase characteristics: phase difference and its absolute value; Waveform characteristics: peak factor, form factor; Frequency domain features: the amplitudes of the first three frequency components extracted by FFT.

9. The system according to claim 8, characterized in that The feature optimization module includes: The Gap-DBN autoencoder adopts a symmetrical encoder-decoder structure, including an input layer, an encoder hidden layer, and a decoder hidden layer. A BatchNormalization layer and a Dropout layer are added after each hidden layer. The Dropout rate of the Dropout layer is 0.3 to stabilize the training process and prevent overfitting. Gap mechanism unit, used to maximize the distance between classes and minimize the distance within classes; The training unit uses the Adam optimizer, ReLU activation function and early stopping strategy.

10. The system according to claim 9, characterized in that The classification processing module is deployed on a cloud server or an on-site edge computing node, and synchronizes data with the data acquisition module via industrial Ethernet, 5G network, or a dedicated communication link to achieve a collaborative diagnosis service that combines local rapid fault response with cloud-based intelligent decision support. The classification processing module includes: The first-level XGBoost classifier is used to distinguish between the five intermediate fault categories and the five non-intermediate fault categories. The five intermediate fault categories include gradually increasing faults, spike faults, continuously low faults, periodic change faults, and abnormal leakage current faults. The five non-intermediate fault categories include sudden increase faults and three-phase imbalance faults. The second-level XGBoost classifier is used to finely classify the middle five types of faults; The third-level XGBoost classifier is used to classify the non-intermediate five-category faults; The data balancing unit uses the SMOTE algorithm to upsample the middle five types of fault samples by a factor of 2.5.

Images