An abnormal recognition method for power supply service work orders based on deep belief network

Abstract

The present application discloses a method for identifying abnormal power supply service work orders based on a deep belief network, which includes decomposing the power supply work order quantity matrix composed of the training set into a coefficient matrix through non-negative matrix factorization, and calculating the abnormal feature indicators of power supply work orders based on statistical laws, temporal characteristics, and spatial similarity respectively; constructing an abnormal power supply work order identification model based on a deep belief network, using the abnormal feature indicators of power supply work orders as inputs and the abnormal classification labels of power supply work orders as outputs, performing model training and parameter tuning, and identifying abnormal power supply work orders according to the trained deep belief network to output the abnormal identification results. The characteristics of abnormal work orders different from normal work orders are extracted from three aspects: statistical laws, temporal characteristics, and spatial similarity, and an abnormal power supply work order identification model based on a deep belief network is constructed, which can effectively and accurately identify abnormal work orders, and is of great significance for improving the processing efficiency of power supply work orders and improving the quality of power supply services.

Description

Technical Field

[0001] The present invention relates to the field of power supply service anomaly recognition, and particularly to a method for recognizing anomalies in power supply service work orders based on a deep belief network. Background Art

[0002] With the continuous popularization and deepening of the power marketization process, power users' requirements for power supply services and power supply quality are constantly increasing. As a bridge for communication between power grid companies and power users, the "95598" power supply service hotline is becoming an important way to effectively handle customer demands and improve the quality of power supply services. However, the data volume involved in the "95598" operator work orders is large, and "bad data" with recording errors is likely to occur during the collection of power supply work orders, resulting in a large difference between the statistical results of the number of power supply work orders and the actual situation. Therefore, effectively identifying anomalies in power supply work order data is of great significance for improving the quality and efficiency of power supply work order processing and discovering weak links in power supply services. At the same time, with the development of power customer service call centers, the massive work order data resources generated by power users provide reliable data support for the recognition of anomalies in power supply work orders.

[0003] Currently, most of the research on the recognition of anomalies in power supply work orders identifies abnormal work orders from the perspective of time series prediction through the correlation between the number of power supply work orders and time. In fact, the reasons for the change in the number of power supply work orders are relatively complex and diverse. Considering the influence of various factors comprehensively plays a key role in improving the accuracy of anomaly recognition. In addition, the construction of anomaly recognition features for power supply work orders usually cannot rely solely on the change in the number of work orders of the power supply company itself, but should also consider the correlation with the change in the number of work orders of other power supply companies.

[0004] How to make full use of the massive power supply work order data and select scientific and effective methods to identify abnormal power supply work orders is an issue that needs to be deeply studied in the current recognition and early warning of anomalies in power supply work orders. It can be seen that the existing methods for recognizing anomalies in power supply work orders need to be improved. Summary of the Invention

[0005] An embodiment of the present application proposes a method for recognizing anomalies in power supply service work orders based on a deep belief network, extracts the features that distinguish abnormal work orders from normal work orders from three aspects: statistical laws, temporal characteristics, and spatial similarity, and constructs a power supply work order anomaly recognition model based on a deep belief network, which can effectively and accurately identify abnormal work orders.

[0006] Specifically, a method for recognizing anomalies in power supply service work orders based on a deep belief network proposed by an embodiment of the present application includes:

[0007] Step 1: Input the "95598" power supply work order data to form a power supply work order quantity matrix, and randomly divide the power supply work order data into a training set and a test set;

[0008] Step 2: Decompose the power supply work order quantity matrix composed of the training set into a feature matrix and a coefficient matrix through non-negative matrix factorization, calculate the similarity distance between the column vectors of the decomposed coefficient matrix, identify the abnormal power supply work orders in the training set, and correct the abnormal work orders;

[0009] Step 3: Calculate the abnormal feature indicators of power supply work orders based on statistical laws, temporal characteristics, and spatial similarity respectively;

[0010] Step 4: Build an abnormal power supply work order recognition model based on a deep belief network, use the abnormal feature indicators of power supply work orders as input and the abnormal classification labels of power supply work orders as output, perform model training and parameter tuning, and identify abnormal power supply work orders according to the trained deep belief network, and output the abnormal recognition results.

[0011] Optionally, step 2 includes:

[0012] For a power supply work order quantity matrix Y composed of N power supply companies and M-day power supply work order data N×M , non-negative matrix factorization finds two decomposition matrices W N×R and L R×M such that the product of matrix W N×R and matrix L R×M is approximately equivalent to the original power supply work order quantity matrix Y N×M , that is:

[0013] Y N×M ≈W N×R ·L R×M ;

[0014] In the formula, matrix W N×R is called the feature matrix, L R×M is called the coefficient matrix, and each element in matrix W N×R and L R×M is non-negative, that is, the power supply work order quantity matrix of the training set is expressed as a weighted combination of feature vectors through non-negative matrix factorization;

[0015] Construct the minimization objective function expression as:

[0016]

[0017] In the formula, D KL is the information loss cost representing the non-negative matrix factorization process; y ij represents the power supply work order quantity represented by the i-th row and j-th column of the power supply work order quantity matrix composed of the training set; W and L represent the feature matrix and the coefficient matrix respectively; the feature matrix and the coefficient matrix are solved by iterative multiplication update rules. Assuming convergence after p iterations, the feature matrix W is obtainedp and the coefficient matrix L p ;

[0018] Based on the changes of the column vectors of the coefficient matrix within a specific time interval T, initially identify the abnormal work orders in the historical data. The expression for judging the abnormal work orders is:

[0019] ||L p，t || > max{(1 + ξ)||L p，t-T ||, (1 + ξ)||L p，t+T ||};

[0020] In the formula, L p,t represents the column vector of the coefficient matrix corresponding to time t; ξ represents the maximum threshold for the change of the column vector of the coefficient matrix when judging abnormal work orders;

[0021] On this basis, correct the abnormal work orders in the historical data. The expression is:

[0022]

[0023] Through the feature matrix W p and the corrected coefficient matrix L p * , reconstruct the preprocessed power supply work order quantity matrix for constructing the abnormal feature index of the power supply work order. Its calculation formula is:

[0024]

[0025] Optionally, in step 3, considering the statistical law of the change of the power supply work order quantity to construct the abnormal feature index of the power supply work order, including:

[0026] For the newly input work order quantity to be identified, determine its rejection region according to the statistical hypothesis test, and thus calculate the abnormal feature index of the power supply work order based on the statistical probability as:

[0027]

[0028] In the formula, Y i,t represents the work order quantity to be identified newly input by power supply company i on the t-th day; represents the abnormal feature index of the power supply work order based on the statistical probability corresponding to the work order to be identified; α is the significance level.

[0029] Optionally, in step 3, considering the time series of the change of the power supply work order quantity to construct the abnormal feature index of the power supply work order, including:

[0030] Analyze the variation law of the number of power supply work orders in the time domain through a long short-term memory neural network, predict its trend in the short term in the future, and construct an expression for the abnormal feature index of power supply work orders based on time series, which is:

[0031]

[0032] In the formula, represents the abnormal feature index of the power supply work order based on the LSTM neural network of power supply company i on the t-th day; δ is the correction margin of the time series prediction error.

[0033] Optionally, in step 3, constructing the abnormal feature index of the power supply work order by considering the spatial similarity of the change in the number of power supply work orders includes:

[0034] Quantify the spatial acceptance ability through the coefficient of variation, and its calculation formula is:

[0035]

[0036] In the formula, B i represents the coefficient of variation of power supply company i; it can be seen that when the coefficient of variation is large, it indicates that the fluctuation degree of the number of power supply work orders is large, and the number of work orders of this power supply company is more easily affected by neighboring power supply companies;

[0037] The spatial influence is quantified by the distance between power supply companies and the similarity degree between the changes in the number of their work orders, that is:

[0038]

[0039]

[0040]

[0041] In the formula, d ij represents the distance between power supply company i and power supply company j; respectively represent the average values of the number of work orders of power supply company i and power supply company j within time T; represents the spatial influence between power supply company i and power supply company j;

[0042] On this basis, the estimated value of the work order of power supply company i on the t-th day based on spatial similarity is expressed as:

[0043]

[0044] In the formula, N * represents all power supply companies adjacent to power supply company i; ΔY j,tDenote the change in the number of work orders of the j - th power supply company adjacent to power supply company i on the t - th day compared with the previous day; thus, the abnormal feature index of power supply work orders based on spatial similarity is determined as

[0045]

[0046] In the formula, Denote the abnormal feature index of power supply work orders based on spatial similarity of power supply company i on the t - th day; ε is the correction margin of the prediction error based on spatial similarity.

[0047] Optionally, step 3 includes:

[0048] The specific steps for constructing the power supply work order anomaly recognition model are as follows. Among them, steps 1) - 6) are the unsupervised pre - training layer by layer of RBM, and steps 7) - 8) are the fine - tuning process of the BP algorithm;

[0049] 1) Standardize the abnormal feature indexes of power supply work orders respectively;

[0050] 2) Input the standardized abnormal feature indexes of power supply work orders as the initial state of the first visible layer V (1) of RBM;

[0051] 3) Calculate the activation probability of each neuron in the first hidden layer H (1) of RBM, that is:

[0052]

[0053] In the formula, p(h j (1) |V (1) ) represents the activation probability of the j - th hidden layer neuron when the state of the visible layer neurons is given; f() is the activation function, and in this invention, tanh is used as the activation function;

[0054] 4) Through the Gibbs sampling technique, sample and reconstruct the visible layer and the hidden layer respectively. The reconstructed visible layer and hidden layer are represented as and

[0055] 5) Update the biases of the neurons in the visible layer and the hidden layer of RBM and the weights between them, and their expressions are respectively:

[0056]

[0057]

[0058]

[0059] In the formula, λ is the learning rate; ΔBv(1) is the bias update value of the visible layer in the first RBM; ΔB h(1) is the bias update value of the hidden layer in the first RBM; Δω (1) is the update value of the weight between the visible layer and the hidden layer in the first RBM;

[0060] 6) Repeat steps 2)-5), and train each RBM layer by layer separately and unsupervised to obtain its weights and biases;

[0061] 7) According to the results of RBM pre-training, use the tanh function as the activation function to estimate the states of each hidden layer and the output layer as follows:

[0062]

[0063]

[0064] In the formula, is the estimated activation state of the j-th hidden layer neuron in the n-th RBM, is the estimated corresponding output value;

[0065] 8) Determine the loss function of the DBN network; then use the abnormal classification label of the power supply work order as the model output, and use the BP algorithm to perform supervised fine-tuning on the weights and biases of the DBN network.

[0066] Optionally, it also includes using the F-measure metric to evaluate the abnormal recognition method, including:

[0067] According to the prediction results and actual situations of power supply work order abnormal recognition, it is divided into four categories: true positive, false positive, true negative, and false negative;

[0068] Record the number of samples of the above four categories as TP, FP, TN, and FN respectively. When measuring the results of abnormal recognition, the commonly used measurement features are precision and recall, and the calculation formulas are respectively:

[0069]

[0070]

[0071] In the formula, P is the precision rate, indicating the proportion of abnormal work orders among the recognized work orders; R is the recall rate, indicating the proportion of all abnormal work orders that are recognized.

[0072] Optionally, it also includes:

[0073] Use the F-measure to evaluate the performance of abnormal recognition, and its expression is:

[0074]

[0075] In the formula, β represents the relative importance of recall and precision; when β > 1, recall has a greater impact; when 0 < β < 1, precision has a greater impact; in the present invention, β = 1.25 is taken as the standard for evaluating the performance of the anomaly recognition model.

[0076] Beneficial effects:

[0077] Preliminarily identify and correct the anomaly points in the historical data to improve the quality of the training samples, and further improve the accuracy of anomaly recognition. Construct the anomaly feature indicators of the power supply work orders, and extract the features of the abnormal work orders different from the normal work orders from three aspects: statistical law, time series, and spatial similarity. Construct an anomaly recognition model for power supply work orders based on a deep belief network to quickly and accurately identify abnormal work orders, which is of great significance for improving the processing efficiency of power supply work orders and improving the quality of power supply services. Description of the drawings

[0078] In order to more clearly illustrate the technical solutions of the present application, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0079] Figure 1 It is a flowchart of a method for identifying anomalies in power supply service work orders based on a deep belief network proposed in an embodiment of the present application;

[0080] Figure 2 It is an anomaly recognition model for power supply work orders based on DBN proposed in an embodiment of the present application;

[0081] Figure 3 It is the result of preprocessing power supply work order data based on NMF proposed in an embodiment of the present application. Specific implementation manners

[0082] To make the structure and advantages of the present application clearer, the structure of the present application will be further described below with reference to the drawings.

[0083] Specifically, Figure 1 It is a flowchart of a method for identifying anomalies in power supply service work orders based on a deep belief network according to the present invention, and the method includes the following steps:

[0084] Step 1: Input the 95598 power supply work order data to form a power supply work order quantity matrix, and randomly divide the power supply work order data to obtain a training set and a test set;

[0085] Step 2: Decompose the power supply work order quantity matrix composed of the training set into a feature matrix and a coefficient matrix through non-negative matrix factorization, calculate the similarity distance between the column vectors of the decomposed coefficient matrix, identify the abnormal power supply work orders in the training set, and correct the abnormal work orders;

[0086] Step 3: Calculate the abnormal feature indicators of power supply work orders based on statistical laws, timeliness, and spatial similarity respectively;

[0087] Step 4: Build an abnormal power supply work order recognition model based on a deep belief network, use the abnormal feature indicators of power supply work orders as input and the abnormal classification labels of power supply work orders as output, perform model training and parameter tuning, and identify abnormal power supply work orders according to the trained deep belief network, and output the abnormal recognition results.

[0088] In implementation, preprocess the power supply work order data based on non-negative matrix factorization, including:

[0089] To initially identify and correct the abnormal work order data in the training set to improve the accuracy of abnormal recognition, preprocess the power supply work order data based on non-negative matrix factorization. For a power supply work order quantity matrix Y composed of N power supply companies and M-day power supply work order data N×M , non-negative matrix factorization finds two decomposition matrices W N×R and L R×M , such that the product of matrix W N×R and matrix L R×M is approximately equivalent to the original power supply work order quantity matrix Y N×M , that is, Y N×M ≈W N×R ·L R×M .

[0090] In the formula, matrix W N×R is called the feature matrix (basis matrix); L R×M is called the coefficient matrix (weight matrix). Each element in matrix W N×R and L R×M is non-negative, that is, the power supply work order quantity matrix of the training set can be expressed as a weighted combination of feature vectors through non-negative matrix factorization.

[0091] The implementation of non-negative matrix factorization can be expressed as a constrained optimization problem. Under the constraint that each element in the feature matrix and the coefficient matrix is required to be non-negative, the goal is to minimize the information loss cost of the preprocessing process for power supply work order data, which can be expressed as:

[0092]

[0093] In the formula, D KL represents the information loss cost of the non-negative matrix factorization process; y ijDenote the number of power supply work orders represented by the element in the \(i\)-th row and \(j\)-th column of the matrix composed of the training set. \(W\) and \(L\) represent the feature matrix and the coefficient matrix respectively. The solutions of the feature matrix and the coefficient matrix can be iteratively completed by the multiplicative update rule. Assume that after \(p\) iterations, convergence is achieved, and the feature matrix \(W\) can be obtained p and the coefficient matrix \(L\) p . Generally speaking, the change trends of the number of normal power supply work orders in the short term should have certain similarities, while abnormal work orders show significantly more than recent work orders. Through non-negative matrix factorization, the matrix of the number of power supply work orders is equivalent to the "splicing combination" of various groups of features affecting the change of the number of power supply work orders through the coefficient matrix, that is, the column vectors of the coefficient matrix \(L\) p characterize the importance of various groups of features affecting the change of the number of power supply work orders. Therefore, through the change of the column vectors of the coefficient matrix within a specific time interval \(T\), abnormal work orders in historical data can be initially identified. The expression for judging abnormal work orders is:

[0094] \(\|L\) p，t \|>\max\{(1 + \xi)\|L\) p，t-T \|,(1 + \xi)\|L\) p，t+T \| \};

[0095] In the formula, \(L\) p,t represents the column vector of the coefficient matrix corresponding to time \(t\); \(\xi\) represents the maximum threshold for the change of the column vector of the coefficient matrix when judging abnormal work orders. On this basis, abnormal work orders in historical data can be corrected to improve the accuracy of subsequent abnormal work order recognition for power supply work orders, that is:

[0096]

[0097] Through the feature matrix \(W\) p and the corrected coefficient matrix \(L\) p * , the matrix of the number of power supply work orders after preprocessing can be reconstructed for the construction of abnormal feature indicators of power supply work orders. Its calculation formula is:

[0098]

[0099] Considering the statistical law of the change of the number of power supply work orders, construct abnormal feature indicators of power supply work orders, including:

[0100] The number of power supply work orders is affected by various relatively independent factors such as weather, policies, and business types, and its daily distribution usually has contingency and randomness. According to the central limit theorem, when a large number of repeated collections of daily power supply work order data are carried out, the results usually show a certain statistical regularity. The statistical regularity of the distribution of the number of power supply work orders is not just a simple superposition of the distribution characteristics of daily power supply work order data, but the inevitability of the power supply work order data collection process. Therefore, in the present invention, the number of work orders of the power supply company every day is regarded as a random event, and its long-term mathematical distribution can be assumed to be a normal distribution. According to the parameter estimation theory, parameter estimation is carried out on the number of power supply work orders within a certain period of time to obtain the mean and variance of the power supply work orders.

[0101] For the newly input number of work orders to be recognized, according to statistical hypothesis testing, its rejection region can be determined, and thus the abnormal characteristic index of the power supply work order based on statistical probability can be calculated as:

[0102]

[0103] In the formula, Y i,t represents the number of newly input work orders to be recognized by the power supply company i on the t-th day; represents the abnormal characteristic index of the power supply work order based on statistical probability corresponding to the work order to be recognized; α is the significance level

[0104] Considering the time series of the change in the number of power supply work orders, the abnormal characteristic index of the power supply work order is constructed, including:

[0105] Natural conditions such as climate and temperature that affect the change in the number of power supply work orders usually have a certain rhythm. Therefore, the number of power supply work orders will remain relatively stable in the short term and have a certain periodicity in the long term. On this basis, the change law of the number of power supply work orders in the time domain can be analyzed by a long short-term memory (LSTM) neural network, and its trend in the short term in the future can be predicted, so as to construct the abnormal characteristic index of the power supply work order based on time series.

[0106]

[0107] In the formula, represents the abnormal characteristic index of the power supply work order based on the LSTM neural network of the power supply company i on the t-th day; δ is the correction margin of the time series prediction error.

[0108] Considering the spatial similarity of the change in the number of power supply work orders, the abnormal characteristic index of the power supply work order is constructed, including:

[0109] In adjacent regions, due to similar climate and policy impacts, the changing trends of the number of similar work orders also have a certain degree of similarity. Therefore, the number of work orders of the target power supply company can be predicted based on the changing trends of the number of work orders of other power supply companies that are spatially adjacent, and then an abnormal feature index of power supply work orders can be constructed.

[0110] For different power supply companies, their sensitivities to changes in the number of work orders of adjacent power supply companies are not exactly the same. The present invention uses spatial acceptance ability to characterize the sensitivity of the target power supply company to changes in the number of work orders of adjacent power supply companies when the number of work orders of adjacent power supply companies changes. A strong spatial acceptance ability indicates that the change in the number of work orders of adjacent power supply companies has a greater impact on the target power supply company. Therefore, the spatial acceptance ability can be quantified by the coefficient of variation, and its calculation formula is:

[0111]

[0112] In the formula, B i represents the coefficient of variation of power supply company i. It can be seen that when the coefficient of variation is large, it indicates that the fluctuation degree of the number of power supply work orders is large, and the number of work orders of this power supply company is more easily affected by adjacent power supply companies.

[0113] On the other hand, the impacts of changes in the number of work orders of adjacent power supply companies on the target power supply company are not exactly the same. The present invention uses spatial influence to characterize the degree of influence of spatially adjacent power supply companies on the target power supply company. The smaller the distance between power supply companies and the more similar the changing trends of the number of work orders, the greater the spatial influence between them. Therefore, the spatial influence between power supply companies can be quantified by the distance between them and the similarity degree between the changing trends of their work orders, that is:

[0114]

[0115]

[0116]

[0117] In the formula, d ij represents the distance between power supply company i and power supply company j; respectively represent the average values of the number of work orders of power supply company i and power supply company j within time T; represents the spatial influence between power supply company i and power supply company j.

[0118] On this basis, the estimated value of the work order of power supply company i on the t-th day based on spatial similarity can be expressed as:

[0119]

[0120] In the formula, N* Denote all power supply companies adjacent to power supply company i; ΔY j,t Denote the change in the number of work orders of the j-th power supply company adjacent to power supply company i on the t-th day compared with the previous day. Thus, the abnormal feature index of power supply work orders based on spatial similarity can be determined as follows:

[0121]

[0122] In the formula, Denote the abnormal feature index of power supply work orders based on spatial similarity of power supply company i on the t-th day; ε is the correction margin of the prediction error based on spatial similarity.

[0123] An abnormal identification model of power supply work orders based on a deep belief network is constructed. By comprehensively considering the abnormal feature index of power supply work orders, the identification of abnormal work orders can be realized, including:

[0124] The deep belief network (DBN) is a type of deep learning model, and its basic structure is composed of stacked restricted Boltzmann machines (RBMs). The RBM is a model based on an energy function, which consists of visible layer neurons and hidden layer neurons with randomness. The visible layer neurons are fully connected to the hidden layer neurons, and there are no connections between neurons in the same layer. In the present invention, power supply work orders are divided into two categories: normal and abnormal. By constructing a DBN model for power supply work order classification, abnormal work orders can be identified.

[0125] The abnormal identification model of power supply work orders is composed of n RBMs. For the k-th RBM, hereinafter, Denote the activation state of each neuron in the visible layer; Denote the activation state of each neuron in the hidden layer; Denote the bias of the visible layer neurons; Denote the bias of the hidden layer neurons; ω ij Denote the weight connecting the i-th visible layer neuron and the j-th hidden layer neuron. The construction process of the abnormal identification model of power supply work orders mainly includes the pre-training of RBM and the fine-tuning of the back-propagation (BP) algorithm, as Figure 2 shown.

[0126] The specific steps for constructing the abnormal identification model of power supply work orders are as follows. Among them, steps 1)-6) are the layer-by-layer unsupervised pre-training of RBM, and steps 7)-8) are the fine-tuning process of the BP algorithm.

[0127] 1) Standardize the abnormal feature index of power supply work orders respectively to reduce the numerical differences between indicators.

[0128] 2) Input the abnormal feature indicators of the standardized power supply work order as the initial state of the first visible layer V of the RBM. (1) of the RBM.

[0129] 3) Calculate the activation probabilities of each neuron in the first hidden layer H (1) of the RBM, that is:

[0130]

[0131] In the formula, p(h j (1) |V (1) ) represents the activation probability of the j-th hidden layer neuron when the state of the visible layer neurons is given; f() is the activation function, and in this invention, tanh is used as the activation function.

[0132] 4) Respectively sample and reconstruct the visible layer and the hidden layer through the Gibbs sampling technique. The reconstructed visible layer and hidden layer are respectively represented as and

[0133] 5) Update the biases of the visible layer and hidden layer neurons in the RBM and the weights between them. Their expressions are respectively:

[0134]

[0135]

[0136]

[0137] In the formula, λ is the learning rate; ΔB v(1) is the bias update value of the visible layer in the first RBM; ΔB h(1) is the bias update value of the hidden layer in the first RBM; Δω (1) is the weight update value between the visible layer and the hidden layer in the first RBM.

[0138] 6) Repeat steps 2) - 5), and layer by layer train each RBM separately and unsupervised to obtain its weights and biases.

[0139] 7) According to the results of the RBM pre-training, use the tanh function as the activation function to estimate the states of each hidden layer and the output layer, as follows:

[0140]

[0141]

[0142] In the formula, is the estimation of the activation state of the j-th hidden layer neuron in the n-th RBM, is the estimation of the corresponding output value.

[0143] 8) Determine the loss function of the DBN network. Then, take the abnormal classification label of the power supply work order as the model output, and use the BP algorithm to perform supervised fine-tuning on the weights and biases of the DBN network as follows:

[0144]

[0145]

[0146]

[0147] where Loss represents the loss function of the DBN network; respectively represent the fine-tuning amounts of the weights and biases of the DBN network.

[0148] It also includes using the "F" metric to evaluate the performance of anomaly recognition, including:

[0149] According to the differences between the prediction results and the actual situations of power supply work order anomaly recognition, it can be divided into four categories: true positive, false positive, true negative, and false negative, as shown in Table 1.

[0150] Table 1 Confusion matrix of power supply service anomaly recognition results

[0151]

[0152] Denote the number of examples of the above four categories as TP, FP, TN, and FN respectively. When measuring the results of anomaly recognition, common measurement features such as precision (also known as accuracy) and recall (also known as recall rate) are often used, and their calculation formulas are as follows:

[0153]

[0154]

[0155] where P is the precision rate, representing the proportion of abnormal work orders among the recognized work orders; R is the recall rate, representing the proportion of all abnormal work orders that are recognized.

[0156] In fact, when performing anomaly recognition on power supply work orders, usually more attention is paid to identifying as many abnormal work orders as possible under the condition that the recognition cost is controlled within a certain range. Therefore, the present invention uses the "F" metric to evaluate the performance of anomaly recognition, and its expression is:

[0157]

[0158] In the formula, β represents the relative importance of recall and precision. When β > 1, recall has a greater impact; when 0 < β < 1, precision has a greater impact. In the present invention, β = 1.25 is taken as the standard for evaluating the performance of the anomaly recognition model.

[0159] To further understand the present invention, taking the data of a certain type of power supply work orders of 79 power supply companies under the jurisdiction of Zhejiang Province for 540 days as an example, anomaly recognition of power supply work orders is carried out. A total of 43,279 power supply work order data are collected. To improve the accuracy of the evaluation results, the power supply work order data are randomly divided into training / test sets according to the ratio of 70% / 30%, and this is repeated 10 times and the average value is taken as the result of the algorithm evaluation.

[0160] First, preprocess the power supply work order data. Taking one of the power supply companies as an example, the preprocessing result of the power supply work order data based on NMF is as Figure 3 shown.

[0161] It can be seen that abnormal power supply work orders generally exist in the historical data (only a part of the abnormal intervals are marked in the figure), and the preprocessing process of the power supply work orders based on non-negative matrix factorization can effectively correct the abnormal work orders in the original data. In addition, the preprocessing process is equivalent to "filtering" the power supply work orders. On the premise of ensuring the original information as much as possible, by discarding redundant feature information, the noise influence in the process of collecting power supply work order data can be reduced.

[0162] Then, calculate the work order anomaly characteristic indicators of each power supply company and use them as the input of the power supply work order anomaly recognition model based on the deep belief network. Taking three of the power supply companies as an example, the work order anomaly characteristic indicators and the power supply work order anomaly recognition results based on the deep belief network on a certain day are shown in Table 2. Among them, R1, R2, and R3 respectively represent the work order anomaly characteristic indicators constructed based on statistical laws, time series prediction, and spatial similarity; the work order anomaly recognition threshold is taken as 0.5, and when the model output exceeds this threshold, the corresponding power supply work order is determined to be abnormal.

[0163] Table 2 Power supply work order anomaly recognition results based on the deep belief network

[0164]

[0165] As can be seen from the above results, for power supply company A, its R1, R2, and R3 are all relatively large, so there is a high probability that the work order data of power supply company A on that day is abnormal; for power supply company B, its R1, R2, and R3 are all relatively small, that is, the work order data on that day will not be identified as abnormal; for power supply company C, although its abnormal recognition eigenvalue R1 is small, due to the obvious abnormal recognition features represented by R2 and R3, the work order data on that day is finally determined to be abnormal.

[0166] When processing the same amount of work order data, the method of the present invention is compared with other methods, including the manual screening method and the power supply work order abnormal recognition method based on LSTM time series prediction. The specific results are shown in Table 4.

[0167] Table 2 Comparison between the method of the present invention and other methods

[0168] Required time / s P R F value Manual screening / 0.447 0.895 0.643 Time series prediction method 142 0.909 0.263 0.364 The method of the present invention 175 0.661 0.816 0.747

[0169] As can be seen from the above table, although the manual screening method can ensure a high recall rate and F value, the time and performance required by it largely depend on the screening conditions selected by humans; the time series prediction method has a high precision rate and the least time consumption, but its recall rate is low, resulting in a low F value. This is because in actual situations, the abnormal recognition of power supply work orders needs to be comprehensively considered in combination with specific natural factors and social factors such as temperature changes and policy release situations. If only the change trend of the number of power supply work orders is predicted from the time scale to identify abnormal work orders, it is easy to cause missed judgments. Although the recall rate and precision rate of the method of the present invention are not the best, by comprehensively considering various abnormal recognition features, as many abnormal work orders as possible can be identified on the premise of ensuring a certain efficiency. Therefore, its F value is the highest among the three methods.

[0170] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention, and any reference signs in the claims should not be regarded as limiting the claims involved.

Claims

1. A method for identifying anomalies in power supply service work orders based on a deep belief network, characterized in that Including: Step 1: Input 95598 power supply work order data, form a power supply work order quantity matrix, and randomly divide the power supply work order data to obtain a training set and a test set. Step 2: Decompose the power supply work order quantity matrix composed of the training set into a feature matrix and a coefficient matrix through non-negative matrix factorization, calculate the similarity distance between the column vectors of the decomposed coefficient matrix, identify abnormal power supply work orders in the training set, and correct the abnormal work orders. Step 3: Calculate the abnormal feature indicators of power supply work orders based on statistical laws, temporal characteristics, and spatial similarity respectively. Step 4: Construct an abnormal power supply work order recognition model based on a deep belief network, use the abnormal feature indicators of power supply work orders as input and the abnormal classification labels of power supply work orders as output, perform model training and parameter tuning, and identify abnormal power supply work orders according to the trained deep belief network, and output the abnormal recognition results. In the said Step 3, when constructing the abnormal feature indicators of power supply work orders considering the spatial similarity of the change in the quantity of power supply work orders, it includes: Quantify the spatial acceptance ability through the coefficient of variation, and its calculation formula is: where B i represents the coefficient of variation of power supply company i. It can be seen that when the coefficient of variation is large, it indicates that the fluctuation degree of the number of power supply work orders is large, and the number of work orders of this power supply company is more likely to be affected by neighboring power supply companies. The spatial influence is quantified by the distance between power supply companies and the similarity degree between the changes in their work order quantities, that is: where d ij represents the distance between power supply company i and power supply company j; respectively represent the average number of work orders of power supply company i and power supply company j within time T; represents the spatial influence between power supply company i and power supply company j; On this basis, the work order estimated value of power supply company i on the t-th day based on spatial similarity is expressed as: where N * represents all power supply companies adjacent to power supply company i; ΔY j,t represents the change in the number of work orders of the j-th power supply company adjacent to power supply company i on the t-th day compared with the previous day; Y i,t-1 is the number of work orders to be identified newly input by power supply company i on the (t - 1)-th day; thus, the abnormal feature index of power supply work orders based on spatial similarity is determined as follows: In the formula, represents the abnormal feature index of the power supply work order based on spatial similarity of power supply company i on the t-th day; ε is the correction margin of the prediction error based on spatial similarity; Y i,t represents the number of newly input work orders to be identified by power supply company i on the t-th day.

2. The abnormal recognition method for power supply service work orders based on a deep belief network according to claim 1, wherein The said Step 2 includes: For a power supply work order quantity matrix Y composed of N power supply companies and M-day power supply work order data N×M , non-negative matrix factorization finds two factorization matrices W N×R and L R×M such that the product of matrix W N×R and matrix L R×M is approximately equivalent to the original power supply work order quantity matrix Y N×M , that is Y N×M ≈W N×R ·L R×M ; In the formula, matrix W N×R is called the feature matrix, L R×M is called the coefficient matrix. Each element in matrix W N×R and L R×M is non-negative, that is, the power supply work order quantity matrix of the training set is represented as a weighted combination of feature vectors through non-negative matrix factorization; Construct a minimization objective function expression, which is In the formula, D KL represents the information loss cost of the non - negative matrix factorization process; y ij represents the power supply work order quantity indicated by the element in the i - th row and j - th column of the power supply work order quantity matrix composed of the training set; W and L respectively represent the feature matrix and the coefficient matrix; the solution of the feature matrix and the coefficient matrix is completed by iterative multiplicative update rules. Assuming convergence after p iterations, the feature matrix W p and the coefficient matrix L p ; Based on the change situation of the column vectors of the coefficient matrix within a specific time interval T, initially identify the abnormal work orders in the historical data, and the expression for judging the abnormal work orders is: ‖L p,t ‖ > max{(1 + ξ)‖L p,t-T ‖, (1 + ξ)‖L p,t+T ‖}; where L p,t represents the column vector corresponding to the coefficient matrix with respect to time t; ξ represents the maximum threshold for the change in the column vector of the coefficient matrix when judging abnormal work orders; On this basis, correct the abnormal work orders in the historical data, and the expression is: Through the feature matrix W p and the corrected coefficient matrix L p * , the preprocessed power supply work order quantity matrix is reconstructed for the construction of abnormal feature indicators of power supply work orders. The calculation formula is as follows:

3. A method for identifying anomalies in power supply service work orders based on a deep belief network according to claim 1, characterized in that, In the said Step 3, when constructing the abnormal feature indicators of power supply work orders considering the statistical law of the change in the quantity of power supply work orders, it includes: For the newly input work order quantity to be identified, determine its rejection region according to statistical hypothesis testing, and thus calculate the abnormal feature indicator of power supply work orders based on statistical probability as: where Y i,t represents the number of work orders to be identified newly input by power supply company i on the t-th day; represents the abnormal characteristic index of power supply work orders based on statistical probability corresponding to the work orders to be identified; α is the significance level; and are respectively the mean and variance of the number of work orders of power supply company i within a certain period of time.

4. A method for identifying anomalies in power supply service work orders based on a deep belief network according to claim 1, characterized in that, In the said Step 3, when constructing the abnormal feature indicators of power supply work orders considering the temporal characteristics of the change in the quantity of power supply work orders, it includes: Analyze the change law of the power supply work order quantity in the time domain through a long short-term memory neural network, predict its trend in the short term in the future, and construct an expression for the abnormal feature indicator of power supply work orders based on temporal characteristics as: Where Y i,t represents the number of work orders to be identified newly input by power supply company i on the t-th day; represents the abnormal feature index of power supply work orders of power supply company i on the t-th day based on the LSTM neural network; δ is the correction margin of the time series prediction error.

5. The method for identifying abnormal power supply service work orders based on a deep belief network according to claim 1, characterized in that The said Step 3 includes: The specific steps for constructing the abnormal power supply work order recognition model are as follows. Among them, Steps 1)-6) are the layer-by-layer unsupervised pre-training of RBM, and Steps 7)-8) are the fine-tuning process of the BP algorithm. 1) Standardize the abnormal feature indicators of power supply work orders respectively. 2) Input the abnormal characteristic indicators of the standardized power supply work order as the initial state of the first visible layer V of RBM (1) ; 3) Calculate the activation probabilities of each neuron in the first RBM hidden layer H (1) That is: In the formula, represents the probability that the j-th hidden layer neuron is activated when the state of the visible layer neuron is given; f() is the activation function, and the tanh function is used as the activation function; is the activation state of the i-th neuron in the visible layer V of the first RBM (1) ; is the bias of the j-th neuron in the visible layer of the first RBM; is the weight connecting the i-th visible layer neuron and the j-th hidden layer neuron in the first RBM; N v is the number of neurons in the visible layer of the first RBM; 4) Sampling and reconstruction are respectively performed on the visible layer and the hidden layer through the Gibbs sampling technique, and the reconstructed visible layer and hidden layer are respectively represented as and 5) Update the biases of the visible layer and hidden layer neurons in RBM and the weights between them, and their expressions are respectively where λ is the learning rate; ΔB v(1) is the bias update value of the visible layer in the first RBM; ΔB h(1) is the bias update value of the hidden layer in the first RBM; Δω (1) is the update value of the weight between the visible layer and the hidden layer in the first RBM; 6) Repeat Steps 2)-5), and train each RBM layer by layer separately and unsupervised to obtain its weights and biases. 7) According to the results of RBM pre-training, use the tanh function as the activation function to estimate the states of each hidden layer and the output layer, as follows: Wherein, is the estimation of the activation state of the j-th hidden layer neuron in the n-th RBM, and is the estimation of the corresponding output value; is the activation state of the i-th neuron in the visible layer V (n) of the n-th RBM; is the bias of the j-th neuron in the hidden layer of the n-th RBM; is the weight of the connection between the i-th visible layer neuron and the j-th hidden layer neuron in the n-th RBM; is the number of neurons in the visible layer of the n-th RBM; 8) Determine the loss function of the DBN network; then use the abnormal classification labels of power supply work orders as the model output, and use the BP algorithm to perform supervised fine-tuning on the weights and biases of the DBN network.

6. The abnormal identification method for power supply service work orders based on a deep belief network according to claim 1, wherein It also includes a method for evaluating anomaly recognition using the F - measure index, including: According to the differences between the predicted results and the actual situations of power supply work order anomaly recognition, it is divided into four categories: true positive, false positive, true negative, and false negative; Denote the number of samples of the above four categories as TP, FP, TN, and FN respectively. When measuring the results of anomaly recognition, the commonly used measurement features are precision and recall, and their calculation formulas are respectively In the formula, P is the precision, indicating the proportion of abnormal work orders among the recognized work orders; R is the recall, indicating the proportion of all abnormal work orders that are recognized.

7. A method for identifying anomalies in power supply service work orders based on a deep belief network according to claim 6, characterized in that, It also includes: Using the F - measure to evaluate the performance of anomaly recognition, and its expression is: In the formula, β represents the relative importance of recall and precision; when β > 1, recall has a greater impact; when 0 < β < 1, precision has a greater impact; select β = 1.25 as the standard for evaluating the performance of the anomaly recognition model.

Images