Time delay sensitive business service quality situation prediction method and device, electronic equipment, readable storage medium and computer program product

Abstract

The invention discloses a delay sensitive business service quality situation prediction method and device, electronic equipment, a readable storage medium and a computer program product, and belongs to the technical field of information. The method comprises the following steps: acquiring operation log data of a base station operation and maintenance center (OMC) at the current moment; performing data cleaning and digital coding on the base station OMC operation log to obtain an OMC log body word vector matrix; performing feature extraction on the OMC log body word vector matrix to obtain a local feature vector and a long-distance feature vector, and performing feature fusion on the local feature vector and the long-distance feature vector to serve as a current input log feature vector; and inputting the current input log feature vector into a pre-constructed situation prediction neural network model, and obtaining a prediction result output by the situation prediction neural network model, the prediction result comprising a base station time delay sensitive business service quality situation scene of a preset duration after the current moment corresponding to the current input log feature vector.

Description

Technical Field

[0001] This application relates to the field of cloud terminal technology, and in particular to a method, apparatus, electronic device, readable storage medium, and computer program product for predicting the service quality of latency-sensitive services. Background Technology

[0002] With the rapid development of 5G technology, latency-sensitive services are facing increasingly stringent requirements and a continuously growing demand. In the current communications field, latency-sensitive services, such as autonomous driving, industrial automation control, and remote medical surgery, require latency at the millisecond level or even lower. The smooth operation of these services heavily relies on a low-latency communication environment, and accurate situation prediction is crucial for ensuring the reliability and stability of latency-sensitive services.

[0003] Existing methods for predicting network service quality status have shortcomings. For example, traditional predictions based on historical key performance indicators have low accuracy, and alarm-based prediction mechanisms are unsuitable for latency-sensitive services. Meanwhile, network optimization platform parameters require frequent modifications, and non-compliant operations affect network stability and user experience. Network management alarms are no longer sufficient. Therefore, improving the accuracy of latency-sensitive task status prediction is a pressing issue for those skilled in the art. Summary of the Invention

[0004] This application provides a method, apparatus, electronic device, readable storage medium, and computer program product for predicting the service quality of time-sensitive services, which can solve the problem of how to improve the accuracy of predicting the service quality of time-sensitive tasks.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows: Firstly, a method for predicting the service quality status of latency-sensitive services is provided, including: Obtain the current base station operation and maintenance center (OMC) operation log data; The base station OMC operation logs are cleaned and digitally encoded to obtain the OMC log body word vector matrix; Feature extraction is performed on the word vector matrix of the OMC log body to obtain local feature vectors and long-distance feature vectors. The local feature vectors and long-distance feature vectors are then fused together and used as the current input log feature vector. The current input log feature vector is input into a pre-constructed situation prediction neural network model, and the prediction result output by the situation prediction neural network model is obtained. The prediction result includes: the base station latency-sensitive service quality situation scenario corresponding to the current input log feature vector after a preset time after the current moment.

[0006] Secondly, a device for predicting the quality of service status of time-delay-sensitive services is provided, comprising: The acquisition module is used to acquire the base station operation and maintenance center (OMC) operation log data at the current moment. The data processing module is used to perform data cleaning and digital encoding on the base station OMC operation log to obtain the OMC log body word vector matrix; The feature extraction module is used to extract features from the word vector matrix of the OMC log body to obtain local feature vectors and long-distance feature vectors, and to use the local feature vectors and long-distance feature vectors as the current input log feature vectors; The situation prediction neural network model module is used to input the current input log feature vector into a pre-constructed situation prediction neural network model and obtain the prediction result output by the situation prediction neural network model. The prediction result includes: the base station latency-sensitive service quality situation scenario corresponding to the current input log feature vector after a preset time.

[0007] Thirdly, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one computer program, which is loaded and executed by the processor to implement the above-described method for predicting the quality of service status of time-sensitive businesses.

[0008] Fourthly, a readable storage medium is provided, wherein at least one computer program is stored in the readable storage medium, the computer program being loaded and executed by a processor to implement the above-described method for predicting the service quality status of latency-sensitive services.

[0009] Fifthly, a computer program product is provided, the computer program product comprising at least one computer program, the computer program being loaded and executed by a processor to implement the latency-sensitive service quality status prediction method provided in the various optional implementations described above.

[0010] The method, apparatus, electronic device, readable storage medium, and computer program product for predicting the service quality status of latency-sensitive services provided in this application embodiment firstly, based on the special structure of base station OMC operation log data, perform data cleaning to obtain an effective log body. To fully extract relevant features from the effective log body, and considering the timeliness requirements of latency-sensitive services for prediction, this application employs a lightweight convolutional neural network and a bidirectional gated recurrent unit to extract key features from the word vector matrix of the OMC log body, solving the problem of lacking feature vector extraction with long-term contextual dependencies. Secondly, a support vector machine-based situation assessment model is constructed to evaluate latency-sensitive situation factors, obtaining labeled data of the service quality status scenario of base station latency-sensitive services, and finding the mapping relationship between OMC operation logs and the service quality status scenario of base station latency-sensitive services. This application embodiment is closer to the low latency and low jitter requirements and characteristics of latency-sensitive services, providing corresponding labeled data of the service quality status scenario of base station latency-sensitive services for the feature vectors of base station OMC operation log data. Thirdly, an attention-based situation prediction network model is constructed, which associates multiple feature vectors of varying lengths and amplifies the influence of features with large factors, improving the accuracy and efficiency of situation prediction.

[0011] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0013] Figure 1 A flowchart of a method for predicting the service quality status of latency-sensitive services provided in an exemplary embodiment of this application is shown. Figure 2 This application illustrates a flowchart of data cleaning and digital encoding of base station OMC operation logs provided in an exemplary embodiment; Figure 3 This application illustrates a schematic diagram of a base station OMC operation log provided in an exemplary embodiment. Figure 4 This illustration shows the structure of a convolutional neural network log local feature extraction model provided in an exemplary embodiment of this application; Figure 5 This illustration shows a schematic diagram of the structure of a BiGRU log long-range feature extraction model provided in an exemplary embodiment of this application; Figure 6 This paper shows a schematic diagram of the base station OMC operation log feature extraction model structure provided in an application example of this application; Figure 7 A schematic diagram illustrating the process of constructing a situation prediction network model provided in an exemplary embodiment of this application is shown; Figure 8 This illustration shows a flowchart of a process for obtaining tag data of a base station's latency-sensitive service quality of service scenario based on a feature vector set of latency-sensitive service situation factors and a situation assessment model, provided by an exemplary embodiment of this application. Figure 9 This illustration shows a flowchart of constructing a situation assessment model using a support vector machine, provided in an exemplary embodiment of this application. Figure 10 This illustration shows a flowchart of a process for generating situation assessment pre-labeled data using a clustering algorithm, provided in an exemplary embodiment of this application. Figure 11 A schematic diagram illustrating the process of constructing a situation prediction neural network model provided in an exemplary embodiment of this application is shown; Figure 12 This illustration shows a flowchart of the process for constructing a Transformer layer provided in an exemplary embodiment of this application; Figure 13 This illustration shows a schematic diagram of the calculation of obtaining the Q, K, and V matrices according to an exemplary embodiment of this application; Figure 14 This illustration shows a schematic diagram of the matching between the q vector and the k vector provided in an exemplary embodiment of this application; Figure 15 This invention provides a schematic diagram of the basic structure of an LSTM according to an exemplary embodiment. Figure 16 This invention provides a schematic diagram of the structure of a BiLSTM network model according to an exemplary embodiment of the present application. Figure 17 This illustration shows a schematic diagram of the structure of a delay-sensitive service quality of service situation prediction device provided in an exemplary embodiment of this application; Figure 18 This invention provides a schematic diagram of the structure of another delay-sensitive service quality situation prediction device according to an exemplary embodiment of this application; Figure 19 A structural block diagram of an electronic device provided in an exemplary embodiment of this application is shown. Detailed Implementation

[0014] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0015] To address the issue of how to fully utilize base station OMC operation logs to uncover their correlation with performance service metrics for latency-sensitive services and improve the accuracy of latency-sensitive task situation prediction, this application provides a method for predicting the service quality situation of latency-sensitive services.

[0016] The following will combine Figures 1 to 16 The method for predicting the service quality of time-sensitive services provided in the embodiments of this application is explained and described in detail. These embodiments are only used to explain this application and do not constitute a limitation thereof.

[0017] Figure 1 A flowchart illustrating a method for predicting the service quality status of latency-sensitive services according to an exemplary embodiment of this application is shown. Figure 1 As shown, the method for predicting the service quality status of latency-sensitive services mainly includes the following steps (S101-S104): S101. Obtain the current base station operation and maintenance center (OMC) operation log data; In this embodiment, predicting future network service quality status based on base station OMC operation logs requires addressing the challenge of extracting useful features from the large amounts of raw echo information returned by network management systems from various vendors. Base station OMC operation logs possess unique structural and semantic characteristics; the command logs are mostly semi-structured text files composed of English words. While the syntax and content of semi-structured text follow certain rules, their organization is relatively flexible. Therefore, more targeted and efficient feature extraction methods are needed to accurately uncover the key information contained within, in order to meet the requirements for predicting the network service quality status of latency-sensitive services.

[0018] In this embodiment, the base station OMC operation log at the current moment is collected to predict the service quality status of the base station's latency-sensitive services at the next moment after a preset time interval, i.e., the future moment. In some embodiments, the base station OMC operation log data at the current moment is obtained through a sliding time window. The size of the sliding time window is a preset duration. At time t, within the sliding time window size, a feature sequence (x1, x2, ..., xn) containing n valid log blocks is extracted. n This refers to the current base station operation and maintenance center (OMC) operation log data.

[0019] S102. Perform data cleaning and digital encoding on the base station OMC operation log to obtain the OMC log body word vector matrix; Figure 2 This application illustrates a schematic diagram of a process for data cleaning and digital encoding of base station OMC operation logs, as shown in an exemplary embodiment of the present application.

[0020] In some embodiments, such as Figure 2 As shown, the base station OMC operation log is cleaned and digitally encoded to obtain the OMC log body word vector matrix, including the following steps (S201-S203): S201. Use regular expression templates to filter out OMC modification instruction log bodies involving modification instructions from base station OMC operation log data; After generating parameter modification methods based on existing strategy algorithms, various network optimization platforms generate instruction scripts through the parameter platform and then complete the parameter modification by issuing them to the network management system. The process includes user login to the network management system, entering the base station, creating and opening the planning area, configuring managed object (MO) data, activating the planning area, executing query instructions, executing modification instructions, exiting the base station, and disconnecting the connection. All these sub-processes are completed in the form of instructions to interact with the network management system.

[0021] Only the parts that execute modification instructions will have a fluctuating impact on the quality of service on the existing network. Therefore, the primary task in feature extraction of network management operation logs is to filter out the valid parts involving modification instructions from a large number of base station OMC operation logs, i.e., data cleaning.

[0022] Taking the original echo from the network management system of *** manufacturer as an example, such as Figure 3 In the base station OMC operation log shown, each operation begins with a timestamp and operation description, and ends with the identifier "---END". Regular expressions are written based on this rule to separate network management operations. Based on the operation description and instruction type (MOD / ADD / RMV / RST / SET / BKP / RESET / BLK) in each instruction operation block, the modification instruction log (BLK) is identified. Successfully executed instruction sets are extracted, and symbols such as parentheses and equal signs, as well as numbers, are removed from the instructions. Finally, the filtered valid instruction logs are reassembled to obtain the OMC modification instruction log body involving modification instructions: BLK CELLLOCALCELLIDCELLADMINSTATECELL_LOSSLESS_BLOCKCELLMIDBLKTIMER.

[0023] S202. The OMC modification instruction log body is digitally encoded using the word frequency dictionary segmentation encoding method to obtain the corresponding digital number of the OMC modification instruction log body. Based on the OMC modification instruction log body, in this embodiment of the application, the modification log body text is converted into a numerical encoding form by word frequency dictionary segmentation encoding method, that is, according to the frequency of the word in all log bodies, it is encoded into a numerical form starting from 1.

[0024] S203. The word embedding method is used to re-encode the numerical number corresponding to the OMC modification instruction log body to obtain the OMC log body word vector matrix.

[0025] Building upon log number encoding, this embodiment employs word embedding to re-encode the numerical identifier into a vector form, thereby transforming a set of text input sequences into a numerical matrix. The log body after word embedding is represented as an OMC log body word vector matrix. ,in, This represents the d-dimensional word vector obtained after the encoded word is embedded.

[0026] Table 1 shows the word vectors obtained after encoding the words and then using word embeddings.

[0027] Table 1.

[0028] As shown in Table 1, the word "BLK" is encoded using word frequency segmentation, resulting in a numerical code of 2. After word embedding, its word vector is (0.01, 0.05, 0.12, 0.04). Combining all the words, the instruction for this occluded cell is transformed into the following word vector matrix:

[0029] S103. Extract features from the word vector matrix of the OMC log body to obtain local feature vectors and long-distance feature vectors, and fuse the local feature vectors and long-distance feature vectors as the current input log feature vectors; In some embodiments, feature extraction of key features from the OMC log body word vector matrix is ​​performed to obtain local feature vectors and long-range feature vectors. This includes: using a convolutional neural network to extract features from the OMC log body word vector matrix to obtain local feature vectors; and using a bidirectional gated recurrent unit to extract features from the OMC log body word vector matrix to obtain long-range feature vectors. To fully extract relevant features from the OMC log body word vector matrix, and considering the timeliness requirements of latency-sensitive services for prediction, this application uses a lightweight convolutional neural network and a bidirectional gated recurrent unit to extract key features from the OMC log body word vector matrix.

[0030] Considering the local dependencies between adjacent words in the base station OMC operation log, such as the fixed usage of single board reset commands RST and BRD, base station reset commands RST and BTSNODE, this embodiment uses a convolutional neural network (CNN) that excels at mining short-range features to generate local feature vectors in the base station OMC log.

[0031] The following will combine Figure 4 The scheme for obtaining local feature vectors in base station OMC logs based on convolutional neural networks, as provided in the embodiments of this application, is explained and described in detail. These embodiments are only used to explain this application and do not constitute a limitation thereof.

[0032] Figure 4 This illustration shows a schematic diagram of the structure of a convolutional neural network log local feature extraction model provided in an exemplary embodiment of this application.

[0033] like Figure 4 As shown, the convolutional neural network log local feature extraction model constructed in this embodiment consists of convolutional layers and pooling layers. The convolutional part comprises three sets of convolutional kernels of different sizes, each set containing two kernels of the same size. The word vector matrix after word embedding is used... As network input, the process by which CNN extracts local feature vector I from the base station OMC logs is as follows: Figure 4 As shown.

[0034] The process of CNN extracting local feature vector I from base station OMC logs includes:

[0035]

[0036] Where Max represents max pooling; Conv represents convolution; and the local feature vector I is obtained by concatenating the convolution results of three sets of convolution kernels.

[0037] In the modification instructions of the base station OMC operation log, the instructions that affect service quality are often issued to the network management system in combination. It is necessary to make a comprehensive judgment based on the key words and the instructions above and below. Although CNN has excellent feature extraction capabilities, it does not perform well in long-distance feature mining. Recurrent Neural Network (RNN) has memory properties and is often used for long-distance text feature extraction. Bidirectional Gated Recurrent Unit (BiGRU) is an improved RNN network that has long-distance feature extraction capabilities and better feature extraction performance, making it more suitable for long-distance feature vector extraction scenarios in base station OMC operation logs.

[0038] The following will combine Figure 5 This application provides a detailed explanation and description of the scheme for obtaining long-range feature vectors from base station OMC logs based on BiGRU. These embodiments are for illustrative purposes only and do not constitute a limitation thereof.

[0039] Figure 5 A schematic diagram of the structure of a BiGRU log long-range feature extraction model provided in an exemplary embodiment of this application is shown.

[0040] like Figure 5 As shown, the BiGRU constructed in this embodiment consists of two gated recurrent units (GRUs) with opposite computation directions, using the OMC log body word vector matrix. As input, the process by which BiGRU extracts the long-range feature vector from the base station OMC operation log is as follows: Figure 5 As shown.

[0041] like Figure 5 As shown, the context feature vector at time t As shown below, the feature vectors captured at all time points are concatenated to represent the context feature h extracted by BiGRU.

[0042] , t

[0043] [ ] Thus, through the above processing and model building of OMC network management operation logs, the local feature vector I and the long-distance feature vector h are obtained respectively.

[0044] In one application example Figure 6 A schematic diagram of the base station OMC operation log feature extraction model structure provided in an exemplary embodiment of this application is shown. Figure 6 As shown, in this base station OMC operation log feature extraction model, the input base station OMC operation log data is cleaned at the log body layer to obtain the OMC modification instruction log body. Then, at the word segmentation layer, the OMC modification instruction log body is digitally encoded using a word frequency dictionary word segmentation encoding method to obtain the corresponding numerical number of the OMC modification instruction log body. At the word embedding layer, word embedding is applied to the numerical number to obtain the OMC log body word vector matrix, which is input to the feature extraction layer. The local feature vector of the base station OMC operation log is extracted by CNN, and the long-distance feature vector of the base station OMC operation log is extracted by BiGRU. At the multi-feature fusion layer, the local feature vector and the long-distance feature vector are fused. Finally, the output layer outputs the current input log feature vector, which serves as the input data for the pre-constructed situation prediction neural network model.

[0045] In this embodiment, the feature extraction of the network management OMC operation log revolves around the special semi-structured architecture of the operation log, taking into account both long-distance and local features, and builds an adapted intelligent model (convolutional neural network and BiGRU model) to extract strong features that can effectively reflect the base station OMC operation log, providing a monitoring data foundation for subsequent latency-sensitive service quality status prediction, and greatly improving the prediction accuracy.

[0046] S104. Input the current input log feature vector into the pre-constructed situation prediction neural network model, and obtain the prediction result output by the situation prediction neural network model. The prediction result includes: the base station latency-sensitive service quality situation scenario corresponding to the current time after the current time and the preset time after the current input log feature vector.

[0047] Based on the base station OMC operation log data obtained at the current time t through a sliding time window in step S101. In some embodiments, the preset duration can be set according to the sliding step size Δt of the sliding time window, thereby predicting the service quality status of base station latency-sensitive services at a future time (i.e., time t+Δt).

[0048] In some embodiments, before inputting the current input log feature vector into the pre-built situation prediction network model, the latency-sensitive service quality situation prediction method provided in this application embodiment further includes: a process of constructing a situation prediction network model.

[0049] Figure 7 This illustration shows a flowchart of constructing a situation prediction network model provided by an exemplary embodiment of this application.

[0050] like Figure 7 As shown, the process for constructing a situation prediction network model provided in this application embodiment includes the following steps (S301-S303): S301. Obtain the target historical base station OMC operation log data at the target historical moment; perform data cleaning, digital encoding and feature extraction on the target historical base station OMC operation log data to obtain the target historical local feature vector and the target historical long-distance feature vector, and fuse the target historical local feature vector and the target historical long-distance feature vector as the first training data. In some embodiments, feature vectors of target historical base station OMC operation log data at target historical time T are obtained using a sliding time window. The size of the sliding time window is a preset duration. Within the sliding time window size at time T, a feature sequence (x1, x2, ..., xn) containing n valid log blocks is extracted. n ), which is the feature vector of the target historical base station OMC operation log data at the target historical moment (including the target historical local feature vector and the target historical long-distance feature vector).

[0051] In some embodiments, the target historical base station OMC operation log data is cleaned, digitally encoded, and feature extracted to obtain the target historical local feature vector and the target historical long-distance feature vector, including the following steps (S401-S405): S401. Use a regular expression template to filter out the target historical OMC modification instruction log body involving modification instructions from the target historical base station OMC operation log data; S402. The target historical OMC modification instruction log body is digitally encoded using the word frequency dictionary segmentation encoding method to obtain the corresponding digital number of the target historical OMC modification instruction log body. S403. The word vector matrix of the target historical OMC log body is obtained by re-encoding the numerical number corresponding to the target historical OMC modification instruction log body using word embedding. S404. Use a convolutional neural network to extract features from the word vector matrix of the target historical OMC log body to obtain the target historical local feature vector. S405. A bidirectional gated cyclic unit is used to extract features from the word vector matrix of the target historical OMC log body to obtain the target historical long-distance feature vector.

[0052] In this embodiment, the method of performing data cleaning, digital encoding, and feature extraction on the target historical base station OMC operation log data to obtain the target historical local feature vector and the target historical long-distance feature vector is the same as the specific implementation method of "performing data cleaning, digital encoding, and feature extraction on the base station OMC operation log data at the current moment to obtain the local feature vector and the long-distance feature vector" in steps S102 to S103. The only difference is that the operation object changes from "the base station OMC operation log data at the current moment" to "the target historical base station OMC operation log data". The specific implementation method will not be described again here. Please refer to the description in the embodiment of steps S102 to S103 for relevant descriptions.

[0053] In this embodiment, the first training data is obtained by extracting local feature vectors and long-distance feature vectors from historical base station OMC operation log data and fusing them, so that it can be used as log sample data for training the situation prediction network model.

[0054] S302. Obtain the feature vector set of historical latency-sensitive service situation factors for the next target historical moment after a preset time period following the target historical moment. Use the support vector machine situation assessment model to obtain the label data of the base station latency-sensitive service quality situation scenario for the next target historical moment, and use it as the second training data. Figure 8This illustration shows a flowchart of obtaining tag data of base station latency-sensitive service quality of service scenario based on the feature vector set of latency-sensitive service situation factors and situation assessment model provided by an exemplary embodiment of this application.

[0055] In some embodiments, the feature vector set of historical latency-sensitive service situation factors for the next target historical moment after a preset time interval is obtained, and a support vector machine situation assessment model is used to obtain the label data of the base station latency-sensitive service quality situation scenario for the next target historical moment, including the following steps (S501-S503): S501. Obtain the historical latency-sensitive business situation factor of the next target historical moment after a preset time period following the acquisition of the target historical moment. This application aims to find the mapping relationship between OMC operation logs and base station latency-sensitive service quality of service (QoS) status scenarios. To this end, this application embodiment uses the target historical base station OMC operation log data obtained through a sliding time window at the target historical time T in step S301. The historical latency-sensitive service status factors for the next target historical time are then obtained. Subsequently, based on the status assessment model, the label data of the base station latency-sensitive service QoS status scenario for the next target historical time can be obtained. Thus, the correspondence between OMC operation log features and future status scenario label data can be obtained.

[0056] In some embodiments, the preset duration can be set according to the sliding step size △T of the sliding time window, and the next target historical time is the time T+△T. In this embodiment, the historical latency-sensitive service status factors at the next target historical time (i.e., the time T+△T) after the preset duration are obtained, and the base station latency-sensitive service quality status scenario at the next target historical time is obtained based on these historical latency-sensitive service status factors.

[0057] In this embodiment, from the perspectives of network performance and service quality, the historical latency-sensitive service status factors include at least one of the following: (1) base station engineering parameters, (2) performance management indicators, (3) private configuration management indicators and (4) public network resource management data indicators.

[0058] The following is a detailed explanation of latency-sensitive business situation factors.

[0059] (1) Base station engineering parameters In some embodiments, base station engineering parameters are periodically acquired, grouped by the respective base station. Base station engineering parameters describe the configuration, physical characteristics, and network architecture of network equipment, and characterize the inherent ability of a base station to provide external service quality. These parameters mainly include at least one of the following: average inter-site spacing, antenna height, power supply stability, and maximum antenna transmit power.

[0060] (2) Performance Management (PM) metrics In some embodiments, one-minute granular PM indicator data is added to the base station's operation. This includes various performance indicator data generated by the network device or system during operation, mainly including: average round-trip time (RTT), QCI2 uplink and downlink packet loss rate, Voice over Long-Term Evolution (VoLTE) uplink and downlink packet loss rate, VoLTE downlink average latency, uplink and downlink throughput, interference level, user-oriented downlink average latency, and the number of Control Channel Elements (CCEs) with aggregation degree n in the Physical Downlink Control Channel (PDCCH). All of the above PM indicator data is obtained from the one-minute granular wide table data update and is used to characterize the real-time operational quality of voice or data services at the base station.

[0061] (3) Private Configuration Management (CM) metrics In some embodiments, proprietary CM indicator parameters from the respective base station manufacturer are added, which are configuration management data unique to a particular manufacturer's base station equipment. For example, taking manufacturer *** as an example, the CM indicator parameters mainly include: manufacturer-specific handover optimization parameters, manufacturer-specific channel quality assessment parameters, manufacturer-specific data scheduling priority parameters, manufacturer-specific interference suppression parameters, etc. The current network values ​​of these proprietary CM indicator parameters are obtained in real-time through a periodic generation strategy on the wireless network optimization parameter platform.

[0062] (4) Public network resource management data indicators In some embodiments, additional public parameters of the New Radio Management (NRM) system for the base station are added, namely network resource management data indicators, mainly including: cell load rate, base station available resource ratio, network congestion index, core network processing latency, etc. The current network values ​​of the above public NRM parameters are obtained in real time through a periodic generation strategy of the wireless network optimization parameter platform.

[0063] In some embodiments, among the historical latency-sensitive service situation factors obtained, there are benefit-type indicators such as VoLTE uplink and downlink packet loss rate, and there are also cost-type indicators such as average RTT round-trip latency. The method provided in this application embodiment further includes: unifying the dimensions of the historical latency-sensitive service situation factors to transform cost-type latency indicators into dimensionless benefit-type indicators of type [0,1].

[0064] In this embodiment, the historical latency-sensitive service situation factors include indicators that can characterize network access capabilities, coverage capabilities, anti-interference capabilities, and perception capabilities, providing multiple perspectives and accurate basic data for the subsequent situation assessment model. This enables the situation assessment model to comprehensively evaluate the label data of the base station latency-sensitive service quality situation scenario from all angles.

[0065] S502. Principal component analysis is used to reduce the dimensionality of historical latency-sensitive business situation factors and reconstruct them to obtain the feature vector set of historical latency-sensitive business situation factors. In some embodiments, considering the high dimensionality of latency-sensitive situational factors, to reduce the computational and storage burden of the situational scenario's labeled data, and to eliminate marginal features with low contribution to situational assessment, principal component analysis is used to reconstruct the dimensionality of the feature vectors of latency-sensitive situational factors before inputting them into the situational assessment model, thus avoiding the subsequent dimensionality explosion problem. Through calculation, this embodiment of the application ultimately obtains a set of historical latency-sensitive business situational factor feature vectors that retain 95% of the original data information and are mutually uncorrelated.

[0066] S503. Input the feature vector set of historical latency-sensitive service situation factors into the situation assessment model of the support vector machine to obtain the label data of the service quality situation scenario of base station latency-sensitive services at the next target historical moment.

[0067] In some embodiments, before inputting the historical latency-sensitive service situation factor feature vector set into the support vector machine situation assessment model, the latency-sensitive service quality situation prediction method provided in this application embodiment further includes a process of constructing a support vector machine situation assessment model. Figure 9 This illustration shows a flowchart of constructing a situation assessment model using a support vector machine, provided in an exemplary embodiment of this application. Figure 9 As shown, the specific steps include (S601-S603): S601. Obtain the feature vector set of historical latency-sensitive business situation factors within a preset historical time period, and use it as the first input data; In this embodiment, the method for obtaining the feature vector set of historical latency-sensitive service status factors within a preset historical time period adopts the same approach as... Figure 8 The implementation method of obtaining the feature vector set of historical latency-sensitive service situation factors after a preset time period following the target historical time in the illustrated embodiment is the same, and will not be repeated here. The feature vector set of historical latency-sensitive service situation factors within the preset historical time period is used as the latency-sensitive service situation factor sample data required for training the situation assessment model of the support vector machine.

[0068] S602. Use a clustering algorithm to generate situation assessment pre-label data for the feature vector set of historical latency-sensitive business situation factors within a preset historical time period. The situation assessment pre-label data is used to initially identify the service quality situation scenario of latency-sensitive business corresponding to the feature vector set of historical latency-sensitive business situation factors, and serves as the second input data. Figure 10 This illustration shows a flowchart of a process for generating pre-labeled situation assessment data using a clustering algorithm, provided in an exemplary embodiment of this application.

[0069] In some embodiments, such as Figure 10 As shown, a clustering algorithm is used to generate pre-labeled situation assessment data from the feature vector set of historical latency-sensitive business situation factors, including the following steps (S701-S704): S701. Calculate the global profile coefficient S of all historical latency-sensitive business situation factor vectors in the historical latency-sensitive business situation factor feature vector set when the number of clusters is i. i The global silhouette coefficient is used to evaluate the clustering effect for a cluster size of i, where i = 1, 2, ..., K. Given that support vector machines are a supervised machine learning algorithm, this embodiment uses an improved K-means clustering algorithm to intelligently label a large amount of unlabeled historical latency-sensitive business situation factor data.

[0070] K-means requires pattern classification based on the differences between the feature vectors of historical latency-sensitive business situation factors at different times, according to certain criteria. The Euclidean distance from the feature vectors of historical latency-sensitive business situation factors to the cluster centers is used as a similarity evaluation index. The feature vectors of historical latency-sensitive business situation factors continuously search for the best cluster, and the cluster centers are continuously updated accordingly, until the squared error function within the cluster reaches its minimum value, or the maximum number of iterations is reached. Among them, the selection of the number of clusters i and the initial cluster centers {X1,X2,...,Xn} greatly affects the quality of clustering.

[0071] In this embodiment, the silhouette coefficient is used as an indicator to evaluate the clustering effect. This indicator combines two measures: separation and cohesion. Cohesion represents the local similarity of sample points (i.e., historical latency-sensitive business situation factors) within the same cluster, and is expressed as the sum of the closeness of the distances between any two historical latency-sensitive business situation factor vectors within the same scenario. Separation represents global dissimilarity, and is expressed as the sum of the proximity of historical latency-sensitive business situation factor vector points in different clusters between the same scenario.

[0072] For example, cohesion can be calculated using the following formula. and resolution :

[0073]

[0074] in, It is the i-th cluster in the current clustering. This indicates the cluster Any historical time delay sensitive situation factor vector in the data, , This indicates the cluster In addition to Other historical time delay sensitive situation factor vectors. Represents the l-th cluster in the current cluster. Any historical time delay sensitive situation factor vector in the vector.

[0075] The historical latency-sensitive service situation factor vector in the feature vector set is calculated using the following formula. Global contour coefficients:

[0076] The expected value of the silhouette coefficients of all n time-delay situation factor vectors is taken as the global silhouette coefficient S of this clustering:

[0077] Following the above method, the global silhouette coefficient S of each cluster is calculated sequentially when the number of clusters is i=1,2,...K. i .

[0078] S702, Select the minimum global profile coefficient S i The corresponding number of clusters is taken as the optimal number of clusters K; In some embodiments, the clustering effect under the condition of the optimal number of clusters K should have the lowest cohesion and the highest separation. Therefore, the global silhouette coefficient S i The minimum number of clusters is the optimal number of clusters.

[0079] S703. Determine the service quality status scenario labels for delay-sensitive services of K base stations; In some embodiments, the service quality status scenarios for latency-sensitive services are initially categorized into five types: low-latency service scenarios, normal-latency service scenarios, busy-latency service scenarios, idle-latency scenarios, and abnormal-latency scenarios. After determining the optimal number of clusters K, for example, by combining manual experience and mathematical models, the optimal number of clusters K = 4, corresponding to the four types of latency-sensitive service quality status scenario labels: low-latency service scenarios, normal-latency service scenarios, busy-latency service scenarios, and abnormal-latency scenarios.

[0080] S704. Using a clustering algorithm, the factors in the historical latency-sensitive service quality of service situation feature vector set are divided into clusters based on the service quality of service situation scenario labels of K base stations. Corresponding situation assessment pre-label data is generated for each cluster. The situation assessment pre-label data includes at least one of the service quality of service situation scenario labels of K base stations.

[0081] The K-means clustering results will vary significantly with changes in the initial cluster centers. In this embodiment, before dividing the clusters, at least K prominent situational factor vectors are selected as initial cluster centers {X1, X2, ..., Xk}. The Euclidean distance from the historical latency-sensitive service situational factor feature vectors to the K cluster centers is used as a similarity evaluation index. The historical latency-sensitive service situational factor feature vectors continuously search for the optimal cluster, assigning the vectors to the clusters with the smallest distances. The cluster centers are recalculated after changes in each cluster, and the cluster centers are continuously updated until the cluster centers no longer change, i.e., the squared error function within the cluster reaches its minimum value, or the maximum number of iterations is reached, at which point the clustering ends. After clustering, the classification results are automatically divided according to the label values ​​of each cluster type, generating corresponding situational assessment pre-label data for each cluster type. The situational assessment pre-label data includes at least one of the K base station latency-sensitive service quality situational scenario labels.

[0082] In this embodiment, based on the improved K-means clustering algorithm, a large amount of unlabeled historical latency-sensitive business situation factor data is clustered and divided to generate corresponding situation assessment pre-label data, so as to serve as the labeled data sample data needed for training the situation assessment model.

[0083] S603. Train the situation assessment model of the support vector machine based on the first input data and the second input data.

[0084] In some embodiments, the Support Vector Machine (SVM) model adopts the SVM model after Particle Swarm Optimization (PSO) kernel parameter optimization, namely the lightweight Polynomial Kernel-Support Vector Machine (PK-PSVM) situation assessment model. The historical latency-sensitive business situation factor feature vector set (first input data) within a preset historical time period and the situation assessment pre-labeled data (second input data) are used as the training data for the situation assessment model of the support vector machine to train the model.

[0085] Finally, in step S503, based on the trained support vector machine situation assessment model, the historical latency-sensitive service situation factor feature vector set is input into the support vector machine situation assessment model to obtain the label data of the base station latency-sensitive service quality situation scenario at the next target historical moment, which serves as the second training data for the situation prediction neural network model. Subsequently, the situation prediction neural network model is trained to achieve the prediction of latency-sensitive service situation at future moments.

[0086] This application aims to find the mapping relationship between OMC operation logs and the service quality status scenario of base station latency-sensitive services. To this end, this application uses the target historical base station OMC operation log data obtained through a sliding time window at the target historical time T in step S301. Simultaneously, to obtain the label data of the base station latency-sensitive service quality status scenario at the next target historical time, this application constructs a lightweight situation assessment model based on improved K-means and PK-PSVM, focusing on historical latency-sensitive service situation factors such as minute-level PM parameters, base station engineering parameters, private configuration management indicators, and public network resource management data indicators. Based on this situation assessment model, the label data of the base station latency-sensitive service quality status scenario at the next target historical time is obtained. This maximizes assessment efficiency and ensures the predictive timeliness of latency-sensitive services while ensuring that principal component influence features are not lost.

[0087] S303. Construct a situation prediction neural network model and train the situation prediction neural network model based on the first training data and the second training data.

[0088] In some embodiments, the situation prediction neural network model includes, in sequence: a data input layer, a Transformer layer, a bidirectional long short-term memory network (BiLSTM) layer, and an output layer; Figure 11 This illustration shows a flowchart of constructing a situation prediction neural network model according to an exemplary embodiment of this application. Figure 11 As shown, the situation prediction neural network model is constructed, including the following steps (S801-S804): S801. Construct a data input layer, wherein the input data of the data input layer includes: first training data and second training data, and the input data is normalized through the data input layer to obtain normalized feature vector data; In some embodiments, a sliding time window is used to extract the sample data required for training and testing the model, including: feature vectors of target historical base station OMC operation log data and the corresponding label data of base station latency-sensitive service quality status scenarios.

[0089] In some embodiments, at time T within the sliding time window, a total feature sequence containing n valid log blocks is extracted. This refers to the feature vector of the target historical base station OMC operation log data at the target historical time T. The sliding time window has a sliding step size of ΔT, and the label data of the base station latency-sensitive service quality status scenario at time T+ΔT is obtained. This refers to the situational scenario label data corresponding to the feature vector of the target historical base station OMC operation log data at time T. Table 2 shows the input log samples and input label samples of the situational prediction neural network model during training.

[0090] Table 2.

[0091] As shown in Table 2 This represents the sliding step size of the window, used to acquire input samples at time T. Then, obtain the process. Situational scene tags after duration , a represents The number of feature vectors in the OMC operation log data executed during the period may fluctuate slightly depending on the amount of parameter modification.

[0092] S802. Construct the Transformer layer, which is based on a multi-head attention mechanism. The Transformer layer allocates attention to log feature vectors that have a significant impact on the prediction of service quality status of base station latency-sensitive services in the normalized feature vector data. In this embodiment, the Transformer model is introduced to amplify the log feature vectors that have a significant impact on latency-sensitive service situations. The attention mechanism is the foundation of the Transformer model, which dynamically adjusts the weights of different features in the input sequence, such as uplink and downlink throughput, interference level, and user-facing downlink average latency, so that the model can focus on feature factors related to latency-sensitive situation prediction. This allows the model to automatically focus on different positions of the feature vector, which helps to improve the overall accuracy and prediction efficiency of the network management OMC situation prediction model.

[0093] The following application example explains the specific process of building a Transformer layer. Figure 12 This illustration shows a schematic diagram of the process for constructing a Transformer layer according to an exemplary embodiment of this application. Figure 12 As shown, the process of building the Transformer layer specifically includes the following steps (S901-S908): S901, Input the normalized feature vector data output from the data input layer into the multi-head attention mechanism module; For example, the normalized feature vector data output by the data input layer is represented as a vector matrix. This serves as the input data for the attention mechanism module. The essence of multi-head attention is the reconstruction of the sequence's embedding matrix by multiple self-attention mechanisms, incorporating global information from the context log features. When the input log feature sequence is long, the self-attention mechanism can effectively capture the dependencies between templates.

[0094] S902 and vector matrix S are learned by their respective learnable matrices. , , Obtain the Query, Key, and Value matrix Q, K, V, as follows: Figure 13 As shown.

[0095] get:

[0096] In this embodiment, These represent log feature vectors respectively. Query vector, These represent log feature vectors respectively. The key vector, These represent log feature vectors respectively. The Value vector; the Q, K, and V matrices are obtained from... It is formed by concatenating vectors, where i∈[1,w]. s1

[0097] S903, Calculate q i k i The correlation score between them.

[0098] This is a crucial step in the entire attention mechanism, in which each q vector must be matched with w k vectors, such as... Figure 14 As shown. Among them, using The vector represents the relevance score vector after matching. By calculating the relevance score, the degree of correlation between log features can be obtained.

[0099] S904, q i k i The correlation scores are scaled dot product and softmax normalization are applied.

[0100] To avoid the gradient vanishing problem during backpropagation via softmax and to improve the generalization ability of the situation prediction model, the correlation score is transformed from a large value to a smaller value by scaling the dot product.

[0101] Softmax normalization normalizes the scores to a range between 0 and 1, generating attention weight coefficients for base station operation log features.

[0102] S905. Use attention weight coefficients to weight the Value vector.

[0103] This indicates that the weighted feature si contains the attention vector. It is the dimension of the key vector.

[0104]

[0105] The formula for converting vector dimension operations into corresponding parallelized matrix operations is as follows:

[0106] S906. Place the original information into h subspaces to obtain the weighted self-attention matrix for each subspace. .

[0107] To capture more log feature information, the original information is placed into h subspaces, enabling the Transformer network to notice the feature information of the input matrix in each of the h subspaces. This is a multi-head attention mechanism, where the number of subspaces, h, is the number of attention heads. Each subspace will then obtain its own self-attention matrix. .

[0108]

[0109] The obtained h Matrix concatenation into a single matrix A = The algorithm restores the data to its original dimension before splitting, thus fusing the information extracted from the h subspaces. The concatenated matrix is ​​then connected to a fully connected layer, undergoes a linear transformation, and weighted to obtain the final output matrix of the h-head attention mechanism. :

[0110] in, This represents the weight matrix corresponding to the fully connected layer.

[0111] S907, Use residual connection to connect the vector matrix S and the matrix transformed by the multi-head attention mechanism. The matrix is ​​obtained by performing concatenation and normalization. ; After the multi-head attention mechanism described above, the vector matrix S is transformed into a matrix containing attention information for h subspaces. Dropout is used to discard a portion of neurons to prevent overfitting, and then residual connections are used to connect the vector matrix S and the matrix transformed by the multi-head attention mechanism. Connect them together, and then pass them through a LayerNorm normalization layer to speed up model convergence, resulting in:

[0112] S908, will The input is processed by a feedforward neural network, and then the matrix is ​​processed using residual connections. sum matrix Connect them together and normalize them using LayerNorm to obtain the final output result O; After step S907, The input is fed into a feedforward neural network, denoted as Feedforward. The feedforward neural network consists of a fully connected network and activation functions. The activation functions introduce nonlinear transformations to enhance the model's performance. Finally, residual connections are used again to connect the matrix... sum matrix Connect them together, and normalize using LayerNorm to obtain the final output result O:

[0113]

[0114] In this embodiment, a Transformer is used as the network for analyzing the operation log features of the network management OMC. Its attention mechanism effectively allocates more attention to key factors affecting the service quality trend of base stations. In this embodiment, based on the original multi-dimensional feature extraction of operation log text, the attention mechanism is used to prioritize features across numerous dimensions, providing a better basis for subsequent prediction of the service quality trend of latency-sensitive services.

[0115] S803. Construct a bidirectional LSTM network layer and perform correlation analysis on the feature vector data output by the Transformer layer through the bidirectional LSTM network layer; RNNs possess strong memory capabilities, giving them an advantage in learning nonlinear features of sequences. In this embodiment, a bidirectional long short-term memory network based on an improved recurrent neural network is selected to receive the feature vector data O output by the Transformer layer, enabling correlation analysis and memory storage of various factors affecting base station service quality.

[0116] Long Short-Term Memory (LSTM) networks improve upon the long-term dependency problem of conventional RNNs by selectively passing information through a gating mechanism. The basic structure of LSTM is shown in the diagram below. Figure 15 As shown.

[0117] in, Indicates the Gate of Oblivion; Indicates the input gate; Indicates the output gate; This represents the input information at the current moment; and These represent the cell output values ​​at the previous and current time points, respectively. and These represent memory units from the previous and current time points, respectively. represents the sigmoid activation function; t represents the tanh activation function.

[0118] Input gate Used to control the degree to which cellular information is updated.

[0119]

[0120]

[0121] Forgotten Gate Used to control the information discarded from state cells.

[0122]

[0123]

[0124] Output gate The amount of information used to control cell output values.

[0125]

[0126]

[0127] In the formula, W and b represent the weight matrix and the bias term, respectively.

[0128] In base station service quality (SQM) status prediction, the current SQM status is often related to the SQM status before and after it. To effectively obtain context-related log information and improve prediction accuracy, this application introduces a bidirectional LSTM (BiLSTM) network, using the feature vector data O output from the Transformer layer as input for wireless network SQM status prediction. BiLSTM is composed of a superposition of forward and reverse LSTMs, and its structure is as follows: Figure 16 As shown.

[0129] like Figure 16 As shown, the forward and reverse LSTM networks calculate the log features of the forward and reverse inputs, respectively, to obtain the hidden layer state outputs. and Then, the vectors of the forward and reverse hidden layer state outputs are concatenated to obtain the final output of the BiLSTM network layer.

[0130]

[0131] LSTM adds a long-term memory module to the traditional RNN short-term memory model. This is beneficial for learning and capturing the correlations between various feature parameters of network management operation modification logs, associating and storing local and long-distance feature information from the logs for analysis. Compared to unidirectional LSTM networks, bidirectional LSTM networks can better capture contextual information, thus more comprehensively and accurately combining this feature information with the situational scenarios of latency-sensitive services' quality of service. BiLSTM further processes the situational scenario label information obtained in the previous log feature construction and time window sample generation stages, laying the groundwork for subsequent situational prediction.

[0132] S804. Construct the output layer. Based on the feature vector data output by the bidirectional LSTM network layer, the base station latency-sensitive service quality status scenario is output through the output layer after a preset time period following the target historical moment.

[0133] In this embodiment, the feature vector of the base station OMC operation log data after deep feature extraction is used as the input vector, and the label data of the service quality situation scenario of the next time delay-sensitive service is used as the corresponding training label data. After normalizing the data dimensions, the input layer data is formed, and a Transformer layer and a BiLSTM prediction model based on a multi-head attention mechanism are constructed. In the situation prediction neural network model, the data input layer is a fully connected neural network, which is responsible for receiving input data; the Transformer layer allocates the attention of the overall model to the features that have a large impact on the prediction of the time delay-sensitive service situation scenario, thereby ignoring features with a smaller impact and improving the accuracy of situation prediction; the BiLSTM network layer is responsible for the correlation extraction of the input features; the output layer is responsible for outputting the label of the service quality situation scenario of the sample log in the next time step.

[0134] The service quality status prediction method for latency-sensitive services provided in this application has the following steps: First, based on the unique structure of base station OMC operation log data, data cleaning is performed to obtain an effective log body. To fully extract relevant features from the effective log body, and considering the timeliness requirements of latency-sensitive services, this application employs a lightweight convolutional neural network and a bidirectional gated recurrent unit to extract key features from the word vector matrix of the OMC log body, solving the problem of extracting feature vectors lacking long-term contextual dependencies. Second, a support vector machine-based situation assessment model is constructed to evaluate latency-sensitive situation factors, obtaining labeled data for the service quality status scenario of base station latency-sensitive services, and finding the mapping relationship between OMC operation logs and the service quality status scenario of base station latency-sensitive services. This application's embodiments are closer to the low latency and low jitter requirements and characteristics of latency-sensitive services, providing corresponding labeled data for the service quality status scenario of base station latency-sensitive services for the feature vectors of base station OMC operation log data. Third, an attention-based situation prediction network model is constructed, associating multiple feature vectors of varying lengths and amplifying the influence of features with large factors, thereby improving the accuracy and efficiency of situation prediction.

[0135] An exemplary embodiment of this application provides a delay-sensitive service quality status prediction device 100. Figure 17 This illustration shows a structural block diagram of a latency-sensitive service quality of service (QoS) situation prediction apparatus provided in an exemplary embodiment of this application. The aforementioned latency-sensitive service QoS situation prediction apparatus can achieve the following: Figures 1 to 16 All or part of the contents of any of the embodiments shown. The following is only a brief description of the structure and function of the latency-sensitive service quality of service situation prediction device. For other matters not covered, please refer to the relevant descriptions in the above-described latency-sensitive service quality of service situation prediction method. The embodiments of this latency-sensitive service quality of service situation prediction device correspond to the embodiments of the above-described latency-sensitive service quality of service situation prediction method. All implementation processes and methods of the above-described method embodiments can be applied to the embodiments of this latency-sensitive service quality of service situation prediction device and can achieve the same technical effect.

[0136] like Figure 17As shown, the latency-sensitive service quality situation prediction device includes: an acquisition module 101, a data processing module 102, a feature extraction module 103, and a situation prediction neural network model module 104. In this embodiment, the acquisition module 101 is used to acquire the base station operation and maintenance center (OMC) operation log data at the current moment; the data processing module 102 is used to clean and digitally encode the base station OMC operation log to obtain the OMC log body word vector matrix; the feature extraction module 103 is used to extract features from the OMC log body word vector matrix to obtain local feature vectors and long-distance feature vectors, and uses the local feature vectors and long-distance feature vectors as the current input log feature vector; the situation prediction neural network model module 104 is used to input the current input log feature vector into a pre-constructed situation prediction neural network model, and obtain the prediction result output by the situation prediction neural network model. The prediction result includes: the base station latency-sensitive service quality situation scenario corresponding to the current input log feature vector after a preset time.

[0137] In some embodiments, the data processing module 102 performs data cleaning and digital encoding on the base station OMC operation logs to obtain an OMC log body word vector matrix: using a regular expression template to filter out OMC modification instruction log bodies involving modification instructions from the base station OMC operation log data; using a word frequency dictionary word segmentation encoding method to digitally encode the OMC modification instruction log bodies to obtain the corresponding numerical numbers of the OMC modification instruction log bodies; and using word embedding to re-encode the numerical numbers corresponding to the OMC modification instruction log bodies to obtain the OMC log body word vector matrix.

[0138] In some embodiments, the feature extraction module 103 extracts key features from the OMC log body word vector matrix in the following ways to obtain local feature vectors and long-distance feature vectors: a convolutional neural network is used to extract features from the OMC log body word vector matrix to obtain local feature vectors; a bidirectional gated recurrent unit is used to extract features from the OMC log body word vector matrix to obtain long-distance feature vectors.

[0139] In some embodiments, the acquisition module 101 is further configured to acquire target historical base station OMC operation log data at the target historical moment; the data processing module 102 is further configured to perform data cleaning and digital encoding on the target historical base station OMC operation log data; and the feature extraction module 103 is further configured to perform feature extraction to obtain target historical local feature vector and target historical long-distance feature vector, and to fuse the target historical local feature vector and target historical long-distance feature vector as the first training data. The acquisition module 101 is also used to acquire the feature vector set of historical latency-sensitive service situation factors of the next target historical moment after a preset time after the target historical moment, and to obtain the label data of the base station latency-sensitive service quality situation scenario of the next target historical moment using the situation assessment model of support vector machine, as the second training data. The situation prediction neural network model module 104 is also used to construct a situation prediction neural network model and complete the training of the situation prediction neural network model based on the first training data and the second training data.

[0140] In some embodiments, the data processing module 102 and the feature extraction module 103 perform data cleaning, digital encoding, and feature extraction on the target historical base station OMC operation log data in the following ways to obtain the target historical local feature vector and the target historical long-distance feature vector: The target historical OMC modification instruction log body involving modification instructions is filtered from the target historical base station OMC operation log data using a regular expression template; the target historical OMC modification instruction log body is digitally encoded using a word frequency dictionary word segmentation encoding method to obtain the corresponding numerical number of the target historical OMC modification instruction log body; the numerical number corresponding to the target historical OMC modification instruction log body is re-encoded using a word embedding method to obtain the target historical OMC log body word vector matrix; a convolutional neural network is used to extract features from the target historical OMC log body word vector matrix to obtain the target historical local feature vector; and a bidirectional gated recurrent unit is used to extract features from the target historical OMC log body word vector matrix to obtain the target historical long-distance feature vector.

[0141] In some embodiments, the acquisition module 101 acquires the feature vector set of historical latency-sensitive service status factors for the next target historical moment after a preset time interval following the target historical moment in the following manner, and uses a support vector machine situation assessment model to obtain the label data of the base station latency-sensitive service quality of service situation scenario for the next target historical moment: acquiring historical latency-sensitive service status factors for the next target historical moment after a preset time interval following the target historical moment, wherein the historical latency-sensitive service status factors include at least one of the following: base station engineering parameters, performance management indicators, private configuration management indicators, and public network resource management data indicators; using principal component analysis to perform dimensionality reduction and reconstruction on the historical latency-sensitive service status factors to obtain the feature vector set of historical latency-sensitive service status factors; inputting the feature vector set of historical latency-sensitive service status factors into the support vector machine situation assessment model to obtain the label data of the base station latency-sensitive service quality of service situation scenario for the next target historical moment.

[0142] In some embodiments, such as Figure 18As shown in the embodiment of this application, the latency-sensitive service quality situation prediction device includes: a situation assessment model module 105, used to acquire a set of historical latency-sensitive service situation factor feature vectors within a preset historical time period as first input data; to generate situation assessment pre-label data by using a clustering algorithm on the set of historical latency-sensitive service situation factor feature vectors within a preset historical time period, wherein the situation assessment pre-label data is used to initially identify the latency-sensitive service quality situation scenario corresponding to the set of historical latency-sensitive service situation factor feature vectors as second input data; and to complete the training of the support vector machine situation assessment model based on the first input data and the second input data.

[0143] In some embodiments, the situation assessment model module 105 generates situation assessment pre-label data for the historical latency-sensitive service situation factor feature vector set using a clustering algorithm in the following manner: It sequentially calculates the global profile coefficient Si of all historical latency-sensitive service situation factor vectors in the historical latency-sensitive service situation factor feature vector set when the number of clusters is i, where the global profile coefficient is used to evaluate the clustering effect when the number of clusters is i, i=1,2,...n; selects the number of clusters corresponding to the smallest global profile coefficient Si as the optimal number of clusters K; determines K base station latency-sensitive service quality of service situation scenario labels; and uses a clustering algorithm to divide the factors in the historical latency-sensitive service situation factor feature vector set into cluster types based on the K base station latency-sensitive service quality of service situation scenario labels, generating corresponding situation assessment pre-label data for each cluster type. The situation assessment pre-label data includes at least one of the K base station latency-sensitive service quality of service situation scenario labels.

[0144] In some embodiments, the situation prediction neural network model sequentially includes: a data input layer, a bidirectional long short-term memory (LSTM) network layer, a Transformer layer, and an output layer. The situation prediction neural network model module 104 constructs the situation prediction neural network model in the following ways: constructing a data input layer, wherein the input data of the data input layer includes: first training data and second training data, and the input data is normalized through the data input layer to obtain normalized feature vector data; constructing a Transformer layer, wherein the Transformer layer is based on a multi-head attention mechanism, and attention is allocated to log feature vectors in the normalized feature vector data that have a significant impact on the prediction of the service quality situation of base station latency-sensitive services; constructing a bidirectional LSTM network layer, and performing correlation analysis on the feature vector data output by the Transformer layer through the bidirectional LSTM network layer; constructing an output layer, and outputting the base station latency-sensitive service quality situation scenario after a preset duration following the target historical time based on the feature vector data output by the bidirectional LSTM network layer.

[0145] The latency-sensitive service quality of service (SQHS) situation prediction device provided in this application firstly cleans the base station OMC operation log data based on its unique structure to obtain an effective log body. To fully extract relevant features from the effective log body, and considering the timeliness requirements of latency-sensitive services, this application employs a lightweight convolutional neural network and a bidirectional gated recurrent unit to extract key features from the word vector matrix of the OMC log body, solving the problem of extracting feature vectors lacking long-term contextual dependencies. Secondly, a support vector machine-based situation assessment model is constructed to evaluate latency-sensitive situation factors, obtaining labeled data for the base station latency-sensitive service SQHS situation scenario, and finding the mapping relationship between the OMC operation log and the base station latency-sensitive service SQHS situation scenario. This application embodiment is closer to the low latency and low jitter requirements and characteristics of latency-sensitive services, providing corresponding labeled data for the base station latency-sensitive service SQHS situation scenario for the feature vectors of the base station OMC operation log data. Thirdly, an attention-based situation prediction network model is constructed, which associates multiple feature vectors of varying lengths and amplifies the influence of features with large factors, improving the accuracy and efficiency of situation prediction.

[0146] Figure 19 This application illustrates a structural block diagram of an electronic device 1000 according to an exemplary embodiment. The electronic device 1000 can be implemented as the aforementioned latency-sensitive service quality of service (QoS) situation prediction device, which can be configured in electronic devices such as smartphones, tablets, laptops, desktop computers, smartwatches, televisions, or servers.

[0147] Typically, electronic device 1000 includes a processor 1001 and a memory 1002.

[0148] Processor 1001 may include one or more processing cores, such as a quad-core processor, a deca-core processor, etc. Processor 1001 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 1001 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 1001 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 1001 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0149] The memory 1002 may include one or more computer-readable storage media, which may be non-transitory. The memory 1002 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 1002 is used to store at least one instruction, which is executed by the processor 1001 to implement all or part of the steps in the latency-sensitive service quality of service situation prediction method shown in the method embodiments of this application.

[0150] Those skilled in the art will understand that Figure 19 The structure shown does not constitute a limitation on the electronic device 1000, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0151] In one exemplary embodiment, a readable storage medium is also provided, which stores a program or instructions that, when executed by a processor, implement all or part of the steps in the above-described method for predicting the quality of service status of latency-sensitive businesses. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, or optical data storage device, etc.

[0152] In one exemplary embodiment, a computer program product is also provided, comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions that, when executed by a computer, cause the computer to perform the aforementioned... Figures 1 to 16 All or part of the steps of the latency-sensitive service quality status prediction method shown in any embodiment.

[0153] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0154] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for predicting the service quality status of time-delay-sensitive services, characterized in that, include: Obtain the current base station operation and maintenance center (OMC) operation log data; The base station OMC operation logs are cleaned and digitally encoded to obtain the OMC log body word vector matrix; Feature extraction is performed on the word vector matrix of the OMC log body to obtain local feature vectors and long-distance feature vectors. The local feature vectors and long-distance feature vectors are then fused together and used as the current input log feature vector. The current input log feature vector is input into a pre-constructed situation prediction neural network model, and the prediction result output by the situation prediction neural network model is obtained. The prediction result includes: the base station latency-sensitive service quality situation scenario corresponding to the current input log feature vector after a preset time after the current moment.

2. The method according to claim 1, characterized in that, The step of cleaning and digitally encoding the base station OMC operation logs to obtain the OMC log body word vector matrix includes: The OMC modification instruction log body involving modification instructions is filtered out from the base station OMC operation log data using a regular expression template; The OMC modification instruction log body is digitally encoded using a word frequency dictionary word segmentation encoding method to obtain the corresponding numerical number of the OMC modification instruction log body; The word vector matrix of the OMC log body is obtained by re-encoding the numerical number corresponding to the OMC modification instruction log body using word embedding.

3. The method according to claim 1 or 2, characterized in that, The feature extraction of key features from the word vector matrix of the OMC log body, resulting in local feature vectors and long-distance feature vectors, includes: The local feature vector is obtained by using a convolutional neural network to extract features from the word vector matrix of the OMC log body; The word vector matrix of the OMC log body is used to extract features by a bidirectional gated loop unit to obtain the long-distance feature vector.

4. The method according to claim 1, characterized in that, Before inputting the current input log feature vector into the pre-built situation prediction network model, the method includes: Obtain the target historical base station OMC operation log data at the target historical moment; perform data cleaning, digital encoding and feature extraction on the target historical base station OMC operation log data to obtain the target historical local feature vector and the target historical long-distance feature vector, and use the target historical local feature vector and the target historical long-distance feature vector as the first training data after feature fusion; The feature vector set of historical latency-sensitive service situation factors is obtained after a preset time period following the target historical time. The label data of the base station latency-sensitive service quality situation scenario at the next target historical time is obtained by using the situation assessment model of support vector machine, and is used as the second training data. The situation prediction neural network model is constructed by training the model based on the first training data and the second training data.

5. The method according to claim 4, characterized in that, The process of cleaning, digitally encoding, and extracting features from the target historical base station OMC operation log data to obtain target historical local feature vectors and target historical long-distance feature vectors includes: The target historical OMC modification instruction log body involving modification instructions is filtered out from the target historical base station OMC operation log data using a regular expression template; The target historical OMC modification instruction log body is digitally encoded using a word frequency dictionary segmentation encoding method to obtain the corresponding numerical number of the target historical OMC modification instruction log body; The word vector matrix of the target historical OMC log body is obtained by re-encoding the numerical number corresponding to the target historical OMC modification instruction log body using word embedding. A convolutional neural network is used to extract features from the word vector matrix of the target historical OMC log body to obtain the target historical local feature vector; A bidirectional gated cyclic unit is used to extract features from the word vector matrix of the target historical OMC log body to obtain the target historical long-distance feature vector.

6. The method according to claim 4, characterized in that, The step of obtaining the historical latency-sensitive service situation factor feature vector set for the next target historical moment after a preset time interval from the target historical moment, and using a support vector machine situation assessment model, yields the label data for the base station latency-sensitive service quality of service situation scenario at the next target historical moment, including: The historical latency-sensitive service situation factor is obtained after a preset time period following the target historical time. The historical latency-sensitive service situation factor includes at least one of the following: base station engineering parameters, performance management indicators, private configuration management indicators, and public network resource management data indicators. Principal component analysis is used to reduce the dimensionality of the historical latency-sensitive service situation factors and reconstruct them to obtain the feature vector set of the historical latency-sensitive service situation factors. The historical latency-sensitive service situation factor feature vector set is input into the situation assessment model of the support vector machine to obtain the label data of the base station latency-sensitive service quality situation scenario at the next target historical moment.

7. The method according to claim 6, characterized in that, Before inputting the set of historical latency-sensitive service situational factor feature vectors into the situational assessment model of the support vector machine, the method further includes: Obtain the feature vector set of historical latency-sensitive business situation factors within a preset historical time period, and use it as the first input data; A clustering algorithm is used to generate situation assessment pre-label data for the historical latency-sensitive service situation factor feature vector set within the historical preset time period. The situation assessment pre-label data is used to initially identify the latency-sensitive service quality situation scenario corresponding to the historical latency-sensitive service situation factor feature vector set, and serves as the second input data. The situation assessment model of the support vector machine is trained based on the first input data and the second input data.

8. The method according to claim 7, characterized in that, The feature vector set of historical latency-sensitive service situation factors is used to generate situation assessment pre-label data using a clustering algorithm, including: Calculate the global profile coefficient S of all historical latency-sensitive service situation factor vectors in the feature vector set of the historical latency-sensitive service situation factor when the number of clusters is i. i The global contour coefficient is used to evaluate the clustering effect for a cluster size of i, where i = 1, 2, ..., n. Choose the minimum global profile coefficient S i The corresponding number of clusters is taken as the optimal number of clusters K; Determine the service quality status scenario labels for latency-sensitive services at K base stations; A clustering algorithm is used to divide the factors in the historical latency-sensitive service situation factor feature vector set into cluster types based on the service quality situation scenario labels of the K base stations. Corresponding situation assessment pre-label data is generated for each cluster type. The situation assessment pre-label data includes at least one of the service quality situation scenario labels of the K base stations.

9. The method according to claim 4, characterized in that, The situation prediction neural network model comprises, in sequence: a data input layer, a bidirectional long short-term memory (LSTM) network layer, a Transformer layer, and an output layer; The construction of the situation prediction neural network model includes: The data input layer is constructed, wherein the input data of the data input layer includes: the first training data and the second training data, and the input data is normalized through the data input layer to obtain normalized feature vector data; The Transformer layer is constructed, wherein the Transformer layer is based on a multi-head attention mechanism, and attention is allocated to log feature vectors in the normalized feature vector data that have a significant impact on the prediction of the service quality status of the base station latency-sensitive service scenario. Construct the bidirectional LSTM network layer, and perform correlation analysis on the feature vector data output by the Transformer layer through the bidirectional LSTM network layer; The output layer is constructed, and based on the feature vector data output by the bidirectional LSTM network layer, the base station latency-sensitive service quality status scenario after a preset time after the target historical moment is output through the output layer.

10. A device for predicting the quality of service status of time-delay-sensitive services, characterized in that, include: The acquisition module is used to acquire the base station operation and maintenance center (OMC) operation log data at the current moment. The data processing module is used to perform data cleaning and digital encoding on the base station OMC operation log to obtain the OMC log body word vector matrix; The feature extraction module is used to extract features from the word vector matrix of the OMC log body to obtain local feature vectors and long-distance feature vectors, and to use the local feature vectors and long-distance feature vectors as the current input log feature vectors; The situation prediction neural network model module is used to input the current input log feature vector into a pre-constructed situation prediction neural network model and obtain the prediction result output by the situation prediction neural network model. The prediction result includes: the base station latency-sensitive service quality situation scenario corresponding to the current input log feature vector after a preset time.

11. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing programs or instructions that can run on the processor, the programs or instructions being executed by the processor to implement the steps of the latency-sensitive service quality of service situation prediction method as described in any one of claims 1 to 9.

12. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions, which, when executed by a processor, implement the steps of the latency-sensitive service quality status prediction method as described in any one of claims 1 to 9.

13. A computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions that, when executed by a computer, cause the computer to perform the steps of the delay-sensitive service quality of service status prediction method as described in any one of claims 1 to 9.

Images