Deep analysis and positioning method and device for abnormal operation cost of communication machine room

Through the semi-supervised IARN interpretable neural network algorithm, the accuracy and coverage problems in the detection of abnormal operation cost of communication room are solved, accurate abnormal data and cause analysis is achieved, and audit risks and operation costs are reduced.

CN118917725APending Publication Date: 2024-11-08CHINA MOBILE GRP GUANGDONG CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410964255.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

When detecting abnormal operating costs of communication computer rooms, the existing technology lacks standardized and standardized detection processes, and cannot fully cover all abnormal problem scenarios, resulting in low audit risks and accuracy.

Method used

The semi-supervised IARN can interpret neural network integration algorithm, and by obtaining computer room operation cost data sets of multiple indicators, abnormal identification and cause analysis, the impact weights of each indicator are calculated, and the causes of abnormalities are accurately positioned.

Benefits of technology

It improves the accuracy of abnormal data discovery, accurately to the minimum granularity of indicators, reduces audit risks and operation costs, and achieves a more comprehensive abnormal data and cause analysis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118917725A_ABST
    Figure CN118917725A_ABST
Patent Text Reader

Abstract

The invention discloses a deep analysis and positioning method and device for abnormal operation cost of a communication machine room, and the method comprises the steps: obtaining a machine room ID operation cost data set containing a plurality of indexes; inputting the machine room ID operation cost data set into the trained IARN anomaly recognition model to carry out anomaly recognition detection so as to output a real-time anomaly index; inputting the real-time abnormal index into the trained IARN abnormal reason analysis model for operation so as to output a real-time index abnormal value; and calculating the influence weight of each index based on the real-time index abnormal value, and positioning the cause of abnormality according to the influence weight of each index. According to the method, all data features of the original data can be reserved, more abnormal data can be identified, and more specific abnormal reasons can be quickly analyzed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication services, and in particular to a method and device for deeply analyzing and locating abnormal operation costs of a communication room. Background Art

[0002] In the operation of the communication room, in addition to equipment factors, various factors such as business, contract, sharing status, room location, area, external environment, etc. will cause anomalies. Some of these anomalies are explicit and can be directly discovered or recorded, while others are implicit and difficult to be discovered. Some are delayed and cannot be discovered in time, forming audit risks.

[0003] In order to avoid audit risks and reduce operating costs, it is important to find anomalies from a wide range of cost data, analyze the causes of anomalies and trace and update them, which has become an important part of daily network operation cost control. The existing detection method of abnormal cost data is mainly to analyze the billing items of a single cost type independently, screen them according to the threshold, and screen out higher or lower data, which is considered abnormal data. Taking electricity costs as an example, the current steps for detecting abnormal electricity costs are as follows: each reimbursement point reports the current expenses; the headquarters summarizes the current expenses and historical expenses of each city and takes out the median; the expenses of each reimbursement point are calculated with the median. When the deviation exceeds a certain threshold, the current expenses of the reimbursement point are considered abnormal expenses; output each abnormal reimbursement point ID according to each city. Finally, the next step is carried out through the associated bill ID and reimbursement point ID. Similarly, other data such as service fees and rental fees must also be analyzed and viewed separately to screen out problematic data.

[0004] According to the existing technical methods, there are mainly the following shortcomings, such as: the existing technical solution is in the form of screening according to each individual billing indicator, that is, establishing an abnormal standard for each indicator, sorting according to the indicator value or screening according to the threshold, and discovering abnormal network operation cost data. However, due to the wide variety of related cost billing items and billing indicators, it is impossible to establish audit standards for all billing indicators, and there is no combination influence rule between the billing items. Only core indicators are used for detection, which has certain limitations, such as the abnormality of increased power consumption when the equipment is damaged. Because there are too many similar combinations, relying solely on experts for detection and judgment cannot completely cover all types of abnormal problem scenarios, resulting in abnormal cost expenditures and audit risks; in the existing technical solutions, the verification of abnormal network operation cost data depends on expert experience. There is a lack of a standardized and normalized detection process; the current abnormal standards are based on the city-level fees. When the overall fees of some computer rooms differ greatly, it is easy to make misjudgments and the accuracy is not high. Summary of the invention

[0005] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0006] To this end, the present invention proposes a method for in-depth analysis and positioning of abnormal operating costs of communication rooms, and develops a semi-supervised IARN interpretable neural network integration algorithm, which improves the accuracy of detecting anomalies and can locate the causes in detail, accurately to the smallest indicator granularity, so as to comprehensively and accurately find out the abnormal data of network operating costs and the causes of the anomalies, thereby reducing costs while reducing audit risks.

[0007] Another object of the present invention is to provide a device for in-depth analysis and positioning of abnormal operating costs in a communication room.

[0008] To achieve the above-mentioned purpose, the present invention proposes, on one hand, a method for deeply analyzing and locating abnormal operation costs of a communication room, comprising:

[0009] Obtain a data set of operating costs for computer room IDs that includes multiple indicators;

[0010] Input the computer room ID operating cost data set into the trained IARN anomaly recognition model to perform anomaly recognition detection to output real-time anomaly indicators;

[0011] Inputting the real-time abnormal index into the trained IARN abnormal cause analysis model for calculation to output the real-time index abnormal value;

[0012] The influence weight of each indicator is calculated based on the real-time indicator abnormal value, and the cause of the abnormality is located according to the influence weight of each indicator.

[0013] The method for in-depth analysis and location of abnormal operation costs of a communication room according to an embodiment of the present invention may also have the following additional technical features:

[0014] In one embodiment of the present invention, the computer room ID operating cost data set includes network operating cost data and comprehensive capital data; after obtaining the computer room ID operating cost data set including multiple indicators, the method further includes:

[0015] The abnormal data in the computer room ID operating cost data set is normalized.

[0016] In one embodiment of the present invention, before inputting the computer room ID operating cost data set into the trained IARN anomaly recognition model, the method further includes training the IARN anomaly recognition model, including:

[0017] Obtain the training data set corresponding to the normalized data set of computer room ID operation cost;

[0018] Construct an IARN anomaly recognition model and optimize model parameters to obtain the optimal IARN anomaly recognition model;

[0019] Inputting the training data set into the optimal IARN anomaly recognition model to map the data points corresponding to all indicators in the training data set to a high-dimensional identical distribution space through a Gaussian kernel function to obtain a hypersphere;

[0020] The distance from each data point in the training data set to the center of the hypersphere is calculated, and when it is judged that the distance is greater than a preset distance threshold, the indicator is marked as abnormal to obtain an abnormal marking result of the training data set, and a trained IARN anomaly recognition model is obtained based on the abnormal marking result.

[0021] In one embodiment of the present invention, before inputting the real-time abnormality indicator into the trained IARN abnormality cause analysis model, the method further includes:

[0022] Constructing an IARN abnormality cause analysis model; wherein the IARN abnormality cause analysis model includes at least a convolutional layer and an attention mechanism layer;

[0023] Inputting the anomaly annotation result into the IARN anomaly cause analysis model, so as to use the convolution layer to perform a convolution operation on the indicator features in the anomaly annotation result to extract local features;

[0024] The attention mechanism layer is used to weight the local features according to the importance of the input features, and the weighted features are calculated and analyzed to output abnormal indicators, so as to obtain a trained IARN abnormal cause analysis model based on the abnormal indicators.

[0025] In one embodiment of the present invention, locating the cause of the abnormality according to the influence weights of the various indicators includes:

[0026] Establish an abnormality cause comparison table according to the preset standard and the impact weight; wherein the abnormality cause comparison table includes abnormal value code, abnormal value type, weight algorithm, related indicators, abnormal description and possible abnormal cause;

[0027] Performing a matching operation on the impact weight and the abnormal cause comparison table to obtain a matching operation result;

[0028] The possible reasons for the abnormality of the indicator are automatically output according to the matching operation result.

[0029] To achieve the above-mentioned purpose, the present invention further provides a device for analyzing and locating abnormal operation costs of a communication room, comprising:

[0030] Operation data acquisition module, used to obtain the data set of computer room ID operation cost containing multiple indicators;

[0031] An abnormality indicator identification module, used for inputting the computer room ID operation cost data set into the trained IARN abnormality identification model to perform abnormality identification detection to output real-time abnormality indicators;

[0032] An indicator value calculation module, used for inputting the real-time abnormal indicator into the trained IARN abnormal cause analysis model for calculation to output the real-time indicator abnormal value;

[0033] The abnormality cause locating module is used to calculate the influence weight of each indicator based on the abnormal value of the real-time indicator, and locate the cause of the abnormality according to the influence weight of each indicator.

[0034] The in-depth analysis and positioning method and device for abnormal operation cost of a communication room in an embodiment of the present invention performs overall analysis and processing of various network data and uses an explainable network model to analyze the causes of abnormal network operation costs, so as to comprehensively and accurately find the abnormal data of network operation costs and the causes of the abnormalities.

[0035] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0037] Figure 1 is a flow chart of a method for in-depth analysis and location of abnormal operation costs of a communication room according to an embodiment of the present invention;

[0038] Figure 2 is a flow chart of another method for in-depth analysis and location of abnormal operation costs of a communication room according to an embodiment of the present invention;

[0039] Figure 3 is a schematic diagram of a cost data anomaly identification and processing process according to an embodiment of the present invention;

[0040] Figure 4 2 is a schematic diagram of the training results of the IARN anomaly recognition model according to an embodiment of the present invention;

[0041] Figure 5 2 is a schematic diagram of test results of the IARN anomaly recognition model according to an embodiment of the present invention;

[0042] Figure 6 is a schematic diagram of a cost abnormality cause analysis and processing process according to an embodiment of the present invention;

[0043] Figure 7It is a structural schematic diagram of a device for deeply analyzing and locating abnormal operation costs of a communication room according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0045] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0046] The following describes the method and device for in-depth analysis and positioning of abnormal operation costs of a communication room according to an embodiment of the present invention with reference to the accompanying drawings.

[0047] The following introduces technical terms that the present invention may be applied to, as shown in Table 1.

[0048] Table 1

[0049]

[0050] Figure 1 is a flow chart of a method for in-depth analysis and location of abnormal operation costs of a communication room according to an embodiment of the present invention. Figure 1 As shown, the method includes but is not limited to the following steps:

[0051] S1, obtain the data set of computer room ID operation cost containing multiple indicators.

[0052] It can be understood that the room ID operating cost data set obtained by the present invention includes network operating cost data and comprehensive capital data, and the corresponding indicators are data indicators contained in the network operating cost data and comprehensive capital data. A variety of indicators can include electricity price, electricity volume, electricity fee, monthly rent, area, service fee, region, number of days, scene division, property ownership, house type, resource type and power supply method, etc.

[0053] Furthermore, the present invention also performs pre-processing on the abnormal data in the computer room ID operation cost data set, such as normalization processing.

[0054] It is understandable that before normalization, it is necessary to ensure that the outliers in the data set have been properly handled, otherwise the presence of outliers may affect the effect of normalization. In addition, the choice of normalization method should be determined based on the distribution characteristics of the data and the needs of subsequent analysis and modeling. Therefore, in the analysis of operating cost data, appropriate normalization helps to eliminate the dimensional effects between different cost items, making different cost categories comparable and better revealing the underlying laws behind the data.

[0055] S2, input the data set of computer room ID operation cost into the trained IARN anomaly recognition model for anomaly recognition detection to output real-time anomaly indicators.

[0056] It is understandable that before using the network to identify abnormal indicators, the present invention needs to effectively train the model.

[0057] It can be understood that the IARN anomaly recognition model in the embodiment of the present invention is an IARN explainable neural network anomaly recognition model. During the training time, the normalized computer room ID operating cost data set is input into the IARN explainable neural network anomaly recognition module for model training.

[0058] Specifically, a large number of sample sets of normalized network operation cost data obtained by the present invention are used for model training and testing. During training, deep feature extraction is performed based on indicators and values ​​such as ID, electricity price, electricity quantity, electricity bill, monthly rent, area, service fee, region, number of days, scene division, property ownership, house type, resource type, and power supply mode. 67% of the data are randomly selected for training, and the remaining 33% are used as a validation set. All data are mapped to a high-dimensional identical distribution space through a Gaussian kernel function, and a hypersphere can be obtained. The sphere contains all normal data. If it is mapped to the sphere, it is normal data, and normal labels are marked in the data list. If it is mapped outside the sphere, it is abnormal data, and abnormal labels are marked in the data list to obtain abnormal labeling results. Thus, a trained IARN abnormality recognition model is obtained.

[0059] Furthermore, after obtaining the trained IARN explainable neural network anomaly recognition model, the real-time normalized data set of computer room ID operation cost is input into the trained IARN explainable neural network anomaly recognition model to obtain the real-time anomaly index.

[0060] The embodiment of the present invention inputs the real-time normalized data into the trained IARN interpretable neural network anomaly recognition model. Due to its interpretability, IARN can not only make predictions or classification decisions, but also clearly point out which features play a key role in judging anomalies. The model calculates the corresponding anomaly score or probability index based on the input cost data. If the model is based on a threshold, a score exceeding a certain threshold will be regarded as an anomaly; if it is based on probability, whether it is an anomaly or not depends on whether the predicted probability exceeds the set threshold. Finally, the anomaly index calculated in real time is fed back to the monitoring system in real time. When an anomaly is found, an early warning or further troubleshooting measures are immediately triggered to ensure that the operating costs of the computer room are within a normal and controllable range. Through such a process, AI technology can be used in real time and effectively to monitor and identify anomalies in the operating costs of the computer room, assist decision-making and improve operation and maintenance efficiency.

[0061] S3, input the real-time abnormal index into the trained IARN abnormal cause analysis model for calculation to output the real-time index abnormal value.

[0062] It is understandable that before the real-time abnormality index is input into the trained IARN abnormality cause analysis model, the IARN abnormality cause analysis model needs to be trained. The IARN abnormality cause analysis model is the IARN interpretable neural network abnormality cause analysis module of an embodiment of the present invention.

[0063] Specifically, the data with anomalies in the above anomaly annotation results are selected and input into the anomaly cause analysis model of the IARN interpretable neural network for model training. Indicators such as ID, electricity price, electricity consumption, electricity fee, monthly rent, service fee, region, number of days, scene division, property ownership, house type, resource type, power supply method, and annotation results are selected for deep feature extraction. Convolution is used to process the data. Each convolution layer has 64 convolution kernels of size [1*3]. These convolution kernels will slide on the input cost data. Since the data is one-dimensional data, a convolution kernel of length [1*3] is used to perform convolution operations on each block to output local features. Different features are weighted according to the importance of the input. Finally, when an anomaly occurs, the cause classification feature data can be used for analysis to see which feature caused the anomaly.

[0064] Furthermore, the output of the IARN network anomaly cause analysis module is used for anomaly identification. The input sample can be judged as abnormal by multi-classification. Finally, by analyzing the weight of the IARN mechanism, it can be determined which features play a key role in the generation of anomalies, thereby finding out the causes of anomalies.

[0065] In actual use, the detected real-time abnormal indicators are input into the IARN network abnormality cause analysis module for calculation, and the abnormal values ​​of each indicator are output.

[0066] Through such convolutional layer processing, the embodiment of the present invention can effectively extract the temporal and spatial features contained in the cost data even if it is not in the form of an image, and provide richer feature input for subsequent tasks such as anomaly detection and trend prediction. At the same time, combined with the pooling layer and other fully connected layers, a powerful neural network model can be constructed to deeply understand and analyze the operating cost data.

[0067] S4, calculates the impact weight of each indicator based on the indicator abnormal value, and locates the cause of the abnormality according to the impact weight of each indicator.

[0068] Specifically, based on the abnormal index of each feature marked above, the influence weight of each index is calculated, and the weight = index value / the sum of all index values ​​of the data.

[0069] After calculating the impact weights of each abnormal indicator through the above model, the indicator with a higher weight is likely to be the main reason for the abnormal overall operating cost. Combining business knowledge and actual conditions, in-depth analysis of these high-weight abnormal indicators can further locate the specific reasons for the abnormality, such as new equipment, suspected power theft, equipment damage or equipment withdrawal from the network or no business, contract termination, but rental fees are still being billed, etc. Therefore, based on the impact weight and specific business scenarios, manual review and confirmation of the abnormal cause can be carried out, and corresponding improvement measures and solutions can be formulated to reduce the probability or degree of abnormal operating costs in the future.

[0070] The in-depth analysis and positioning method for abnormal operation costs of a communication room in an embodiment of the present invention combines the semi-supervised learning advantages of the SVDD algorithm and its modeling capabilities for unlabeled data, and the capabilities of ResNet50 and Attention for feature importance analysis, and formulates a semi-supervised IARN interpretable neural network integration algorithm to get rid of the influence of human factors and cover as many indicators as possible; it identifies abnormal data in a wider range and enriches the types of abnormal scenarios; at the same time, it improves the accuracy of discovering anomalies, and can locate the causes in detail, accurately to the smallest indicator granularity, so as to comprehensively and accurately find the abnormal data of network operation costs and the causes of the anomalies, thereby reducing costs and improving benefits while reducing audit risks.

[0071] Figure 2 Another method for analyzing and locating abnormal operation costs of a communication room according to an embodiment of the present invention is as follows: Figure 2 As shown:

[0072] In one embodiment of the present invention, after obtaining network operation cost data in the network operation cost system, including electricity fee, rental fee, tower service fee reimbursement data for a certain period, the comprehensive capital data is associated with the computer room ID as the computer room ID operation cost data set. The network operation cost data is shown in Table 2.

[0073] Table 2

[0074]

[0075] In one embodiment of the present invention, after data access, some of the above data needs to be processed, such as deleting abnormal data and supplementing blank data. The string type data is encoded according to the rules, such as encoding the region as shown in Table 3, encoding the scene as shown in Table 4, and encoding the property ownership as shown in Table 5.

[0076] Table 3

[0077] area Tianhe District Haizhu District Baiyun District Panyu District Liwan District Yuexiu District Nansha District …… coding 1 2 3 4 5 6 7 ……

[0078] Table 4

[0079]

[0080] Table 5

[0081] Ownership owner China Tower China Mobile …… coding 1 2 3 ……

[0082] Specifically, by integrating the electricity charges, rental fees, tower service fees and comprehensive financing system data in the network operating cost system, a comprehensive data set of computer room ID operating costs was constructed, which accurately shows the operating cost status of each computer room.

[0083] The processed data is shown in Table 6, which shows the data after the operating cost and comprehensive capital are associated and aggregated:

[0084] Table 6

[0085]

[0086] Furthermore, since the numerical differences between input data such as electricity, rent, area, etc. are too large, in order to eliminate the dimensional differences between different features and make different features comparable, the above-mentioned output abnormal data needs to be normalized.

[0087] In one embodiment of the present invention, the data is mapped to [0, 1] using Min-Max Normalization, and the calculation formula is as follows:

[0088]

[0089] Among them, each dimension feature is defined as xi , i=1,2,...P, then the eigenvalue of the kth sample is x ik ,k=1,2,...,n,where min(x i ) and max(x i ) are the features x i The minimum and maximum values ​​among all samples.

[0090] After processing, it is converted into the normalized total data set in Table 7, that is, the normalized data set M of the computer room ID operating cost:

[0091] Table 7

[0092]

[0093] Thus, a normalized data set of computer room ID operating costs containing multiple indicators is obtained, which can be used for subsequent model training and testing.

[0094] In one embodiment of the present invention, a normalized data set of operating costs of a computer room ID containing multiple indicators is input into an IARN interpretable neural network anomaly recognition module for model training. This may include obtaining a sample set corresponding to the normalized data set of operating costs of a computer room ID, dividing the sample set into a training data set, constructing an IARN anomaly recognition model, and optimizing model parameters to obtain an optimal IARN anomaly recognition model; inputting the training data set into the optimal IARN anomaly recognition model to map the data points corresponding to all indicators in the training data set to a high-dimensional identical distribution space through a Gaussian kernel function to obtain a hypersphere; calculating the distance from each data point in the training data set to the center of the hypersphere, and when it is determined that the distance is greater than a preset distance threshold, marking the indicator as abnormal to obtain an abnormal marking result of the training data set, and obtaining a trained IARN anomaly recognition model based on the abnormal marking result.

[0095] Specifically, Figure 3 As shown, the abnormality recognition data training process of the IARN interpretable neural network abnormality recognition module of the present invention is as follows:

[0096] 1) The normalized data set M of the computer room ID operation cost containing multiple indicators is input into the IARN explainable neural network anomaly recognition module. The IARN explainable neural network anomaly recognition module includes a deep support vector data description network, a residual connection network, an attention mechanism enhanced network, a dynamic clustering network, an explanatory output layer network structure, etc.

[0097] 2) Model parameter tuning, set two parameter combinations, two optimizer and learning rate parameter combinations; parameter set 1: [2 layers, 64 neurons], [3 layers, 128 neurons], [4 layers, 256 neurons], parameter set 2: [3 layers, 64 neurons], [3 layers, 128 neurons], [4 layers, 128 neurons]; optimizer and learning rate parameter sets: [Adam optimizer, learning rate is 0.001], [SGD optimizer, learning rate is 0.01]; By trying different parameter combinations to tune, evaluate the impact of different deep network structures on model performance, and select different optimizer and learning rate parameters to evaluate the effect of the model.

[0098] 3) The optimal parameter combination is [3 layers, 128 neurons], [SGD optimizer, learning rate 0.01], and the appropriate solution is as follows: the center R is 0.79745794, and the radius r is 0.046698399566880044.

[0099] 4) Using the optimal IARN interpretable neural network anomaly recognition module obtained in the previous step, calculate the distance h from each sample point to the center of the hypersphere in the normalized data set M of the computer room ID operation cost containing multiple indicators; when the distance h is greater than the preset distance threshold r, the sample is marked as abnormal; thus, the anomaly labeling result of the total data set can be obtained. Based on the anomaly labeling result, a trained IARN interpretable neural network anomaly recognition module is obtained.

[0100] The training results are as follows Figure 4 The anomaly labeling results are shown in Table 8:

[0101] Table 8

[0102]

[0103]

[0104] The embodiment of the present invention adopts the IARN interpretable neural network anomaly recognition module to discover hidden problems and patterns in the data through in-depth analysis and processing of electricity charges, rental fees, and tower service fee data, such as whether there are abnormal data points, whether there are data errors, whether there are abnormal events, etc., thereby improving the accuracy and reliability of data analysis and decision-making.

[0105] Furthermore, we introduced the expert-annotated data set for testing. Based on the test of 480 normal and abnormal test sets, we conducted statistical analysis on the output data. The test results are as follows: Figure 5 shown.

[0106] Furthermore, the data comparison between the manual annotation of the test set and the output of the ARN model is shown in Table 9:

[0107] Table 9

[0108] Manual annotation of test set IARN Model Output Data volume 480 480 Correct data 458 451 Abnormal data 22 27 Abnormality rate 4.58% 5.625%

[0109] Based on the test results, the data under the horizontal line that exceeds the center of the circle are the anomalies. As can be seen from Table 9, IARN outputs 5 more abnormal data than manual annotation, but there is 1 data that IARN did not find abnormal, and the actual deviation is 6. After manual comparison, these 6 data do have abnormal cost expenditure problems.

[0110] It can be known that the use of IARN will significantly increase the detection rate of abnormal cost expenditures, from 4.58% to 5.625%, with an actual effect improvement of 22.8%.

[0111] Furthermore, the data with the anomaly annotated is taken and input into the anomaly cause analysis module of the IARN interpretable neural network for training. This may include inputting the anomaly annotation results obtained from the training into the IARN anomaly cause analysis model, using the convolution layer to perform convolution operations on the indicator features in the anomaly annotation results to extract local features; using the attention mechanism layer to weight the local features according to the importance of the input features, and calculating and analyzing the weighted features to output anomaly indicators, so as to obtain a trained IARN anomaly cause analysis model.

[0112] Specifically, Figure 6 As shown in the figure, the above anomaly annotation results are input into the IARN explainable neural network anomaly cause analysis module, which includes a convolution layer, an attention mechanism layer, an autoencoder layer, and a clustering layer network structure. Convolution is used to process data. Each convolution layer has 64 convolution kernels of size 3x3. These convolution kernels slide on the input cost data and perform convolution operations on each 3x3 block (i.e., a local area in the cost data). Through continuous convolution, the neural network can better capture local features in the cost data such as sudden growth, seasonal fluctuations, and trend changes (short-term trends or periodic patterns of time series data, edges, corners or other local structures of spatial data, anomalies or anomaly patterns in cost data, key inflection points or change trends in data), and can extract more features. By introducing the attention mechanism for different local features, the model can weight different features according to the importance of the input features (after the convolution operation, the attention mechanism layer weights the features). This process is automatically performed in the forward propagation stage of model training. The model learns which features contribute more to anomaly recognition and focuses on these key features. Such a mechanism helps the model to quickly locate the cause of anomalies and improves the accuracy and efficiency of anomaly detection.

[0113] Furthermore, all the data detected as abnormal, after the module operation and analysis, output the abnormal values ​​of each feature (ID, electricity price, electricity consumption, electricity fee, monthly rent, area, service fee) and mark them. As shown in Table 10:

[0114] Table 10

[0115]

[0116] Furthermore, based on the abnormal index of each feature marked above (Table 10), the weight of each index is calculated, weight = index value / sum of all index values ​​of the data, and the calculation formula is as follows:

[0117]

[0118] Here we define each dimension feature as x i , i=1,2,...n, then the eigenvalue of the jth sample is x ij ,j=1,2,...,m.

[0119] Take the first data in Table 10 as an example: the power index weight of this data = 0.458800955 / (0.002006394+0.458800955+…+0.554568126)≈43.97%, as shown in the power index grid of the computer room with ID 1 in Table 11.

[0120] Each data feature is calculated separately, and Table 11 shows the impact weight of each indicator on the overall output:

[0121] Table 11

[0122] ID Electricity Prices Power Electricity Monthly rent area Service Fee 1 0.19% 43.97% 0.71% 0.07% 1.92% 53.14% 2 8.09% 26.47% 3.77% 2.31% 23.58% 35.78% 3 1.94% 67.01% 8.39% 3.60% 11.37% 7.69% 4 0.46% 45.39% 1.22% 1.04% 1.99% 49.89% …… …… …… …… …… …… ……

[0123] Furthermore, the present invention establishes an abnormal cause comparison table according to preset standards and impact weights; wherein the abnormal cause comparison table includes abnormal value codes, abnormal value types, weight algorithms, related indicators, abnormal descriptions and possible abnormal causes; performs matching operations on the impact weights and the abnormal cause comparison table to obtain matching operation results; and automatically outputs the possible abnormal causes that lead to the indicator abnormality according to the matching operation results.

[0124] Specifically, the present invention refers to the influencing weights of each indicator in Table 11, and at the same time associates the influencing factors such as the air conditioning power and the rated power of the main equipment of the comprehensive capital data of the previous and next months within the account period, and establishes the "Abnormal Cause Comparison Table" based on expert experience and the following standards:

[0125] 1) Only one indicator is the main influencing indicator. For example, if only the power indicator of a certain data is abnormal, it means that the power problem caused this abnormality, and the cause of this indicator is recorded as an abnormality;

[0126] 2) If two indicators are abnormal, compare and analyze what the two indicators are, and record the abnormal reasons according to the different weights of the indicators. If the electricity price + electricity quantity indicator is abnormal and the electricity price weight is greater than the electricity quantity weight, the electricity price is the main reason. If the electricity price + monthly rent indicator is abnormal and the monthly rent weight is greater than the electricity price weight, record the monthly rent as the main abnormality and the others as subordinate reasons.

[0127] 3) And so on.

[0128] The entire abnormal cause comparison table of the embodiment of the present invention is divided into 6 parts:

[0129] Abnormal value coding: Each abnormality is given a unique code, which is convenient for reference and output;

[0130] Outlier type: differentiated according to the number of main indicators that cause the anomaly;

[0131] Weight algorithm: Divide the indicators according to their impact weight on the anomaly. IARN can interpret the data output by the neural network anomaly cause analysis module, which needs to be matched with the algorithm rules to output the anomaly cause. The enumeration in the table is a reference value and can be adjusted according to actual conditions;

[0132] Involved indicators: What are the main indicators that affect the anomaly?

[0133] Abnormal description: Consider abnormal indicators and weights, and fill in the abnormal problem description after manual analysis;

[0134] Possible causes of the exception: Based on the exception description, the detailed causes of the exception are analyzed. The causes involve various possible analyses such as environment, equipment, and behavior, providing guidance for cost analysis.

[0135] The above rules automatically output the abnormal reason comparison table as shown in Table 12:

[0136] Table 12

[0137]

[0138]

[0139] As the training sample data increases, the table can be continuously supplemented and improved.

[0140] Furthermore, based on the output result table 11, a matching operation is performed according to the "Abnormal Cause Comparison Table" in Table 12 to obtain the corresponding matching operation result. When a row of data in Table 11 satisfies a certain algorithm in the comparison table, the possible abnormal cause that caused the abnormal indicator is automatically output according to the matching operation result. The system will automatically output the possible abnormal cause that caused the abnormality according to the comparison table.

[0141] For example, the abnormal indicator of a data output in Table 11 is power consumption and the impact weight is 60%. Table 13 can be used as a reference for the abnormal list page of a certain period of data output:

[0142] Table 13

[0143]

[0144] In summary, the present invention uses a semi-supervised IARN interpretable neural network ensemble algorithm and a spherical distribution model to identify abnormal data, and realizes abnormal identification and abnormal cause analysis for electricity charges, rental fees, tower service fee data and comprehensive capital data in network operating costs. First, the collected data set is cleaned and normalized to eliminate dirty data in the data and dimensional differences between different dimensions. Then, an abnormal interpretable identification model is established in the IARN interpretable neural network. Using the learned spherical distribution, IARN can detect another class from one class and find threatening abnormal behaviors. The IARN interpretable neural network transforms data features and weights, constructs an abnormal cause comparison table, and quickly locates the cause to avoid cost management risks. The present invention can retain all data features of the original data, identify more abnormal data, and analyze more specific abnormal causes. It has better interpretability than traditional manual analysis or early neural networks, and has higher data processing efficiency. It can clearly explain the decision-making process made by the model, greatly saving computing power costs.

[0145] In summary, the communications industry is one of the country's basic pillar industries, and its operating costs are huge. However, due to the rapid development of the industry, unbalanced regional development, complex and diverse equipment, and poor data circulation, abnormal problems often occur, and it is difficult to locate the causes of the abnormalities, which in turn leads to a large increase in operating costs. The present invention abandons the traditional manual statistical method and proposes a new solution based on AI, making abnormal data detection more comprehensive and accurate, and giving statistical results for the proportion of factors causing abnormalities, thereby proposing guiding directions for analyzing and solving problems, and effectively solving the abnormal costs of operators. Therefore, the use of the present invention can significantly reduce the abnormal costs of communication operators, thereby improving the profitability of operators.

[0146] According to the in-depth analysis and positioning method of abnormal operation cost of communication room according to the embodiment of the present invention, through the semi-supervised IARN interpretable neural network ensemble algorithm, and the technical method of using the spherical distribution model to identify abnormal data, it is possible to realize abnormal identification and abnormal cause analysis for the electricity fee, rental fee, tower service fee data and comprehensive capital data in the network operation cost, and automatically detect abnormal data through the artificial intelligence system, so that more and more accurate abnormal data can be detected. At the same time, compared with the existing technology, the present invention gets rid of the limitation of human subjectivity, can find more abnormal causes, and explain the influencing factors causing the abnormality. Compared with the traditional method, the present invention is more intelligent, more accurate, and more labor-saving.

[0147] In order to implement the above embodiment, Figure 7 As shown, this embodiment also provides a communication room abnormal operation cost in-depth analysis and positioning device 10, which includes an operation data acquisition module 100, an abnormal indicator identification module 200, an indicator value calculation module 300 and an abnormal cause positioning module 400.

[0148] The operation data acquisition module 100 is used to obtain a data set of operation costs of a computer room ID including multiple indicators;

[0149] The abnormality indicator identification module 200 is used to input the computer room ID operation cost data set into the trained IARN abnormality identification model to perform abnormality identification detection to output real-time abnormality indicators;

[0150] The indicator value calculation module 300 is used to input the real-time abnormality indicator into the trained IARN abnormality cause analysis model for calculation to output the real-time indicator abnormality value;

[0151] The abnormality cause location module 400 is used to calculate the influence weight of each indicator based on the real-time indicator abnormal value, and locate the cause of the abnormality according to the influence weight of each indicator.

[0152] Furthermore, the room ID operation cost data set includes network operation cost data and comprehensive capital data; after the operation cost acquisition module 100, it also includes:

[0153] The preprocessing module is used to normalize the abnormal data in the computer room ID operating cost data set.

[0154] Furthermore, before the abnormal indicator identification module 200, an identification model training module is also included, which is used to:

[0155] Obtain the training data set corresponding to the normalized data set of computer room ID operation cost;

[0156] Construct an IARN anomaly recognition model and optimize model parameters to obtain the optimal IARN anomaly recognition model;

[0157] The training data set is input into the optimal IARN anomaly recognition model to map the data points corresponding to all indicators in the training data set to the high-dimensional identical distribution space through the Gaussian kernel function to obtain a hypersphere;

[0158] The distance from each data point in the training data set to the center of the hypersphere is calculated. When it is judged that the distance is greater than a preset distance threshold, the indicator is marked as abnormal to obtain an abnormal labeling result of the training data set, and a trained IARN anomaly recognition model is obtained based on the abnormal labeling result.

[0159] Furthermore, before the indicator value calculation module 300, it also includes: a cause analysis model training module, which is used to:

[0160] Constructing an IARN abnormality cause analysis model; wherein the IARN abnormality cause analysis model at least includes a convolutional layer and an attention mechanism layer;

[0161] The anomaly annotation results are input into the IARN anomaly cause analysis model to perform convolution operations on the indicator features in the anomaly annotation results using a convolution layer to extract local features;

[0162] The attention mechanism layer is used to weight the local features according to the importance of the input features, and the weighted features are calculated and analyzed to output abnormal indicators, so as to obtain a trained IARN abnormal cause analysis model based on the abnormal indicators.

[0163] Furthermore, the above-mentioned abnormality cause location module 400 is also used for:

[0164] Establish an abnormality cause comparison table according to the preset standard and the impact weight; wherein the abnormality cause comparison table includes abnormal value code, abnormal value type, weight algorithm, related indicators, abnormal description and possible abnormal cause;

[0165] Performing a matching operation on the impact weight and the abnormal cause comparison table to obtain a matching operation result;

[0166] The possible reasons for the abnormality of the indicator are automatically output according to the matching operation result.

[0167] According to the in-depth analysis and positioning device for abnormal operation cost of a communication room according to an embodiment of the present invention, a semi-supervised IARN interpretable neural network is used, and a spherical distribution model is used to identify abnormal data, so that abnormal identification and abnormal cause analysis can be performed for electricity charges, rental charges, tower service charges and comprehensive capital data in network operation costs, so that abnormal data can be automatically detected through an artificial intelligence system, and more and more accurate abnormal data can be detected. At the same time, compared with the existing technology, the present invention has gotten rid of the limitation of human subjectivity, can find more abnormal causes, and explain the influencing factors causing the abnormality. Compared with traditional methods, the present invention is more intelligent, more accurate, and more labor-saving.

[0168] In order to implement the above embodiments, the present application also proposes a non-temporary computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method described in the above embodiments is implemented.

[0169] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0170] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

Claims

1. A method for in-depth analysis and positioning of abnormal operation costs of a communication room, characterized in that: include: Obtain a data set of operating costs for computer room IDs that includes multiple indicators; Input the computer room ID operating cost data set into the trained IARN anomaly recognition model to perform anomaly recognition detection to output real-time anomaly indicators; Inputting the real-time abnormal index into the trained IARN abnormal cause analysis model for calculation to output the real-time index abnormal value; The influence weight of each indicator is calculated based on the real-time indicator abnormal value, and the cause of the abnormality is located according to the influence weight of each indicator.

2. The method according to claim 1, characterized in that The computer room ID operation cost data set includes network operation cost data and comprehensive capital data; after obtaining the computer room ID operation cost data set including multiple indicators, the method further includes: The abnormal data in the computer room ID operating cost data set is normalized.

3. The method according to claim 2, characterized in that Before inputting the computer room ID operating cost data set into the trained IARN anomaly recognition model, the method further includes training the IARN anomaly recognition model, including: Obtain the training data set corresponding to the normalized data set of computer room ID operation cost; Construct an IARN anomaly recognition model and optimize model parameters to obtain the optimal IARN anomaly recognition model; Inputting the training data set into the optimal IARN anomaly recognition model to map the data points corresponding to all indicators in the training data set to a high-dimensional identical distribution space through a Gaussian kernel function to obtain a hypersphere; The distance from each data point in the training data set to the center of the hypersphere is calculated, and when it is judged that the distance is greater than a preset distance threshold, the indicator is marked as abnormal to obtain an abnormal marking result of the training data set, and a trained IARN anomaly recognition model is obtained based on the abnormal marking result.

4. The method according to claim 3, characterized in that Before inputting the real-time abnormality indicator into the trained IARN abnormality cause analysis model, the method further includes: Constructing an IARN abnormality cause analysis model; wherein the IARN abnormality cause analysis model includes at least a convolutional layer and an attention mechanism layer; Inputting the anomaly annotation result into the IARN anomaly cause analysis model, so as to use the convolution layer to perform a convolution operation on the indicator features in the anomaly annotation result to extract local features; The attention mechanism layer is used to weight the local features according to the importance of the input features, and the weighted features are calculated and analyzed to output abnormal indicators, so as to obtain a trained IARN abnormal cause analysis model based on the abnormal indicators.

5. The method according to claim 1, characterized in that: The locating the cause of the abnormality according to the influence weight of each indicator includes: Establish an abnormality cause comparison table according to the preset standard and the impact weight; wherein the abnormality cause comparison table includes abnormal value code, abnormal value type, weight algorithm, related indicators, abnormal description and possible abnormal cause; Performing a matching operation on the impact weight and the abnormal cause comparison table to obtain a matching operation result; The possible reasons for the abnormality of the indicator are automatically output according to the matching operation result.

6. A device for analyzing and locating abnormal operation costs in a communication room, characterized in that: include: The operating cost acquisition module is used to obtain the operating cost data set of the computer room ID containing multiple indicators; An abnormality indicator identification module, used for inputting the computer room ID operation cost data set into the trained IARN abnormality identification model to perform abnormality identification detection to output real-time abnormality indicators; An indicator value calculation module, used for inputting the real-time abnormal indicator into the trained IARN abnormal cause analysis model for calculation to output the real-time indicator abnormal value; The abnormality cause locating module is used to calculate the influence weight of each indicator based on the abnormal value of the real-time indicator, and locate the cause of the abnormality according to the influence weight of each indicator.

7. The device according to claim 6, characterized in that The room ID operation cost data set includes network operation cost data and comprehensive capital data; after the operation cost acquisition module, it also includes: The preprocessing module is used to normalize the abnormal data in the computer room ID operating cost data set.

8. The device according to claim 7, characterized in that Before the abnormal indicator identification module, there is also an identification model training module for: Obtain the training data set corresponding to the normalized data set of computer room ID operation cost; Construct an IARN anomaly recognition model and optimize model parameters to obtain the optimal IARN anomaly recognition model; Inputting the training data set into the optimal IARN anomaly recognition model to map the data points corresponding to all indicators in the training data set to a high-dimensional identical distribution space through a Gaussian kernel function to obtain a hypersphere; The distance from each data point in the training data set to the center of the hypersphere is calculated, and when it is judged that the distance is greater than a preset distance threshold, the indicator is marked as abnormal to obtain an abnormal marking result of the training data set, and a trained IARN anomaly recognition model is obtained based on the abnormal marking result.

9. The device according to claim 8, characterized in that Before the indicator value calculation module, it also includes: a cause analysis model training module, which is used to: Constructing an IARN abnormality cause analysis model; wherein the IARN abnormality cause analysis model includes at least a convolutional layer and an attention mechanism layer; Inputting the anomaly annotation result into the IARN anomaly cause analysis model, so as to use the convolution layer to perform a convolution operation on the indicator features in the anomaly annotation result to extract local features; The attention mechanism layer is used to weight the local features according to the importance of the input features, and the weighted features are calculated and analyzed to output abnormal indicators, so as to obtain a trained IARN abnormal cause analysis model based on the abnormal indicators.

10. The device according to claim 6, characterized in that The abnormal cause location module is also used for: Establish an abnormality cause comparison table according to the preset standard and the impact weight; wherein the abnormality cause comparison table includes abnormal value code, abnormal value type, weight algorithm, related indicators, abnormal description and possible abnormal cause; Performing a matching operation on the impact weight and the abnormal cause comparison table to obtain a matching operation result; The possible reasons for the abnormality of the indicator are automatically output according to the matching operation result.