A service system reliability analysis system and prediction method based on hybrid neural network and attention mechanism
Through a service system reliability analysis system based on hybrid neural network and attention mechanism, Prometheus monitoring technology and deep learning algorithms are used to extract the software and hardware feature information of the service system, and the insufficient accuracy of the service system reliability analysis in the existing technology is solved. The problem is to achieve high accuracy reliability prediction and dynamic display.
Patent Information
- Application Number
- CN202211447386.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-18
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2042-11-18
AI Technical Summary
It is difficult for the prior art to effectively analyze and predict the reliability of service systems in dynamic and large-scale environments, especially under high complexity and diverse user needs, and it is impossible to accurately monitor and optimize the failures and resource allocation of service systems.
Using a service system reliability analysis system based on hybrid neural network and attention mechanism, the software and hardware environment information is obtained in real time through Prometheus cloud native monitoring technology, data preprocessing and timing data analysis are performed, and multi-grained LSTM and multi-core convolutional neural network extraction is used. Characteristic information, combined with attention mechanism to make reliability prediction.
It improves the accuracy of service system reliability analysis, can directly predict service levels under the premise of multiple classifications, meet the agreement requirements between service providers and users, and provide dynamic real-time rendering reliability information display.
Smart Images

Figure CN115811473B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an analysis system, in particular to a service system reliability analysis system based on a hybrid neural network and an attention mechanism, and belongs to the field of service computing. Background Art
[0002] With the surge in global collaborative needs and the in-depth practice of service computing technology, service-oriented has become the mainstream trend of constructing complex software. As the core technology for building service-oriented systems, service composition not only needs to meet the functional needs of users, but also needs to ensure the quality of service (QoS). The premise of service quality assurance is to be able to find problems, that is, it is necessary to implement service system reliability measurement technology. However, with the development of service technology and the complexity and diversity of user needs, the construction problems of service systems are increasingly showing characteristics such as dynamic, large-scale and diversified preferences, which in turn brings a series of new challenges to the reliability measurement technology of service systems from the aspects of service system fault monitoring configuration, service system fault monitoring optimization scale and service system fault monitoring optimization target. The adaptive evolution mechanism of intelligent application systems is studied, and the characteristics of network resource call, data processing and interaction of service components in complex intelligent systems of human-machine groups in a highly dynamic and changeable big data environment are explored. The correlation between them and the operating environment, user load and throughput is analyzed, and the law of adaptive construction of intelligent application systems in dynamic environments is studied. The online resource feature information of service systems is extracted, and a refined resource feature portrait of service systems is constructed to collect high-quality data of service system runtime for tasks such as service system reliability measurement and prediction.
[0003] Therefore, in order to effectively monitor the reliability trends of component services in service systems, developing new algorithms for service system reliability analysis has important theoretical significance and practical application value. Summary of the invention
[0004] The present invention aims at the problems existing in the prior art and provides a service system reliability analysis system based on a hybrid neural network and an attention mechanism. The technical solution provides a new way of thinking for reliability analysis of the service system at runtime, which makes full use of the characteristics of system reliability contained in the software and hardware environment of the service system at runtime and improves the accuracy of reliability analysis of the service system.
[0005] In order to achieve the above object, the technical solution of the present invention is as follows: a service system reliability analysis system based on a hybrid neural network and an attention mechanism, characterized in that the analysis system includes
[0006] Data acquisition module: uses Promethus cloud-native monitoring technology to acquire and store a large amount of runtime timing information of the software and hardware environment in the service system in real time;
[0007] Database storage module: stores the historical QoS data of the service system acquired in real time into the database and stores it persistently;
[0008] Data preprocessing module: preprocesses the timing information of the service system hardware and software environment during runtime, including data minimum and maximum normalization;
[0009] Time series data reliability analysis module: uses a hybrid neural network and attention mechanism algorithm to mine the time series data of the software and hardware status of the service system during operation, and performs system reliability analysis and prediction on this type of time series information;
[0010] Service system reliability information display module: displays the analyzed reliability information in a visual form;
[0011] The control relationship between each module is as follows: the real-time QoS data of the service system is obtained through the data acquisition model, the QoS data of the service system is written to the database, the QoS data of the service system is extracted from the database, the data preprocessing module performs preprocessing operations on the extracted data, and the preprocessed service historical QoS data is put into the reliability analysis and prediction model trained in advance to obtain the service reliability level classification corresponding to the future effective prediction period, and the predicted service reliability situation is thrown into the information display module, the future reliability prediction level information of the service is combined with the historical QoS information of the service system obtained through the Web interface, and the combined information is dynamically rendered in real time in the information display model.
[0012] The system uses Prometheus cloud-native monitoring technology to obtain and store the service system's software and hardware timing information in real time, and then uses a hybrid neural network to discover the time series data characteristics of high-dimensional indicators when the service system is running, and pays real-time attention to the service system's software and hardware information, that is, the trend of its service operation reliability.
[0013] A method for predicting service system reliability based on a hybrid neural network and an attention mechanism, the method comprising the following steps:
[0014] Step 1: Data minimum and maximum normalization;
[0015] We use the data minimum and maximum normalization method to achieve data normalization of service QoS time series information and standardize service
[0016] Service response time, service throughput, and service reliability to [0, 1];
[0017]
[0018]
[0019]
[0020] Where responseTime*, throughput*, reliability* are responseTime, throughput,
[0021] The result after reliability normalization. responseTime max , throughput max ,reliability max For Service
[0022] The maximum value of service response time, service throughput, and service reliability in the system QoS data;
[0023] responseTime min , throughput min , reliability min is the service response time in the service system QoS data,
[0024] The minimum value of service throughput and service reliability.
[0025] Step 2: Time series data window division;
[0026] The minimum and maximum normalized data are divided into time series data windows, and the data is divided according to the order of historical service system QoS, leading time, and service system reliability data in the future effective prediction period. The service system reliability data in the future effective prediction period window is averaged to obtain the service system reliability classification label corresponding to the future effective prediction period window. The time series length of the service history QoS data is set to t, and the dimension of the service QoS is n. After data normalization and time series data window division, a t*n service history QoS matrix Q={q1, q2, ..., q t} and a reliability level classification sequence L = {l1, l2, ..., l t}, where q i ={responseTime i , thoughtput i , reliability i}
[0027] Step 3: Use multi-granularity LSTM to extract service QoS time series features;
[0028] LSTM with 128, 64, and 32 neural units is used to extract the characteristics of service QoS time, Q = {q1, q2, ..., q t} is used as the input of the LSTM model, and the output can be obtained as R j = {r j1 , r j2 , ..., r jt}(j=32,64,128)
[0029] Step 4: Use the attention mechanism to extract service QoS sensitive information;
[0030] In the entire service QoS time slice, calculate Attention score The calculated attention score The softmax function is used to calculate the attention distribution value. Finally, the feature vector R′ of the service QoS after attention extraction is calculated.
[0031]
[0032]
[0033] Step 5: Service QoS feature integration;
[0034] Fusion of service QoS time series characteristics R′ at each granularity 32 , R′ 64 , R′ 128 The calculation formula is as follows:
[0035] M = concat(R′ 32 , R′ 64 , R′ 128 )
[0036] Step 6: Use a multi-core convolutional neural network to extract service QoS feature profile information;
[0037] This patent uses 3x64 and 2x32 convolution kernels, and the specific convolution calculation formula is as follows:
[0038] C 64 =W 3*64 M[i:i+63](i=1,2,...,t)
[0039] C 64 After batch normalization and relu activation function, we get C64
[0040] C 32 =W 2*32 C′ 64[i:i+31](i=1,2,...,t)
[0041] C 32 After batch normalization and relu activation function, we get C′ 32
[0042] Step 7: Maximum convolution pooling service QoS feature matrix;
[0043] R″′=max(C′ 32 )
[0044] Step 8: Use the fully connected layer to buffer the service reliability classification features;
[0045] R″′=w3R″+b″′
[0046]
[0047] Step 9: Calculation of loss function;
[0048] Finally, we use sparse categorical cross entropy as the loss function of AHN_SR, which is calculated as:
[0049]
[0050] Compared with the prior art, the present invention has the following advantages. On the one hand, the technical solution can supplement the defects of the existing research on service system reliability analysis, and from a new perspective, provide a service system reliability analysis method based on hybrid neural network and attention mechanism for maintaining service system reliability. Compared with the existing service system reliability analysis method, the present invention integrates the advantages of hybrid neural network and the key information capture of attention mechanism, and makes full use of the software and hardware environment information of the service system during operation, so as to have higher accuracy in service system reliability analysis. This scheme uses a multi-classification prediction method for service system reliability. Compared with the past service system reliability regression prediction method, the algorithm based on hybrid neural network and attention mechanism proposed in the present invention is the first to predict the service system reliability under the premise of multi-classification of service levels. It uses multi-granularity LSTM neural network, attention mechanism, and multi-core CNN neural network to extract service operation characteristics in the service system QoS portrait. Compared with the traditional service system reliability regression prediction method, the method proposed in the present invention can directly obtain which service reliability level the current service belongs to, and directly judge whether the service meets the service level agreement between the service provider and the service user. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 It is a flow chart of the service system reliability prediction method based on hybrid neural network and attention mechanism;
[0052] Figure 2 This is a framework diagram of a service system reliability prediction method based on a hybrid neural network and attention mechanism;
[0053] Figure 3 It is a framework diagram of the service system reliability analysis system. DETAILED DESCRIPTION
[0054] In order to deepen the understanding of the present invention, the present embodiment is described in detail below with reference to the accompanying drawings.
[0055] Example 1: See Figure 1 , a service system reliability analysis system based on hybrid neural network and attention mechanism, the analysis system includes
[0056] Data acquisition module: uses Promethus cloud-native monitoring technology to acquire and store a large amount of runtime timing information of the software and hardware environment in the service system in real time;
[0057] Database storage module: stores the historical QoS data of the service system acquired in real time into the database and stores it persistently.
[0058] Data preprocessing module: preprocesses the timing information of the service system hardware and software environment during runtime, including data minimum and maximum normalization;
[0059] Time series data reliability analysis module: uses an algorithm based on a hybrid neural network and an attention mechanism to mine the time series data of the software and hardware status of the service system during operation, and performs system reliability analysis and prediction on this type of time series information;
[0060] Service system reliability information display module: displays the analyzed reliability information in a visual form.
[0061] The control relationship between each module is as follows: the real-time QoS data of the service system is obtained through the data acquisition model, the QoS data of the service system is written to the database, the QoS data of the service system is extracted from the database, the data preprocessing module performs preprocessing operations on the extracted data, and the preprocessed service historical QoS data is put into the reliability analysis and prediction model trained in advance to obtain the service reliability level classification corresponding to the future effective prediction period, and the predicted service reliability situation is thrown into the information display module, the future reliability prediction level information of the service is combined with the historical QoS information of the service system obtained through the Web interface, and the combined information is dynamically rendered in real time in the information display model.
[0062] The system uses Prometheus cloud-native monitoring technology to obtain and store the service system's software and hardware timing information in real time, and then uses a hybrid neural network to discover the time series data characteristics of high-dimensional indicators when the service system is running, and pays real-time attention to the service system's software and hardware information, that is, the trend of its service operation reliability.
[0063] Embodiment 2: A service system reliability prediction method based on a hybrid neural network and an attention mechanism, the method comprising the following steps:
[0064] Step 1: Data minimum and maximum normalization
[0065] We use the data min-max normalization method to achieve data normalization of service QoS time series information, and standardize service response time, service throughput, and service reliability to [0, 1].
[0066]
[0067]
[0068]
[0069] Among them, responseTime*, throughput*, reliability* are the normalized results of responseTime, throughput, and reliability. max , throughput max ,reliability max It is the maximum value of the service response time, service throughput, and service reliability in the service system QoS data. min , throughput min , reliability min It is the minimum value of the service response time, service throughput, and service reliability in the service system QoS data.
[0070] Step 2: Time series data window division
[0071] The minimum and maximum normalized data is divided into time series data windows, and the data is divided according to the order of historical service system QoS, leading time, and service system reliability data within the future effective prediction period. The service system reliability data within the future effective prediction period window is averaged to obtain the corresponding service system reliability classification label within the future effective prediction period window. We set the time series length of the service history QoS data to t and the dimension of the service QoS to n. After data normalization and time series data window division, we obtain a t*n service history QoS matrix Q={q1, q2, ..., q t} and a reliability level classification sequence L = {l1, l2, ..., l t}, where q i ={responseTime i , thoughtput i , reliability i}
[0072] Step 3: Use multi-granularity LSTM to extract service QoS time series features
[0073] LSTM with 128, 64, and 32 neural units is used to extract the characteristics of service QoS time. Q = {q1, q2, ..., q t}As the input of the LSTM model, the output result can be obtained
[0074] Step 4: Use attention mechanism to extract service QoS sensitive information
[0075] In the entire service QoS time slice, calculate Attention score The calculated attention score The softmax function is used to calculate the attention distribution value. Finally, the feature vector R′ of the service QoS after attention extraction is calculated.
[0076]
[0077]
[0078] Step 5: Service QoS feature integration
[0079] Fusion of service QoS time series characteristics R′ at each granularity 32 , R′ 64 , R′ 128 The calculation formula is as follows:
[0080] M = concat(R′32 , R′ 64 , R′ 128 )
[0081] Step 6: Use a multi-core convolutional neural network to extract service QoS feature profile information
[0082] This patent uses 3x64 and 2x32 convolution kernels, and the specific convolution calculation formula is as follows:
[0083] C 64 =W 3*64 M[i:i+63](i=1,2,...,t)
[0084] C 64 After batch normalization and relu activation function, we get C′ 64
[0085] C 32 =W 2*32 C′ 64 [i:i+31](i=1,2,...,t)
[0086] C 32 After batch normalization and relu activation function, we get C′ 32
[0087] Step 7: Max convolution pooling service QoS feature matrix
[0088] R″′=max(C′ 32 )
[0089] Step 8: Use a fully connected layer to buffer service reliability classification features
[0090] R″′=w3R″+b″′
[0091]
[0092] Step 9: Loss Function
[0093] Finally, we use sparse classification cross entropy as the loss function of AHN_SR. The calculation formula is:
[0094]
[0095] Through the above description, we can conclude that the service system reliability prediction method and system based on hybrid neural network and attention mechanism proposed in the present invention can make full use of the multi-dimensional runtime information in the time series data information, thereby obtaining higher classification accuracy and having high application value.
[0096] It should be noted that the above embodiments are not intended to limit the protection scope of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.
Claims
1. A service system reliability analysis system based on hybrid neural network and attention mechanism to achieve service system reliability prediction, It is characterized in that The analysis system includes Data acquisition module: Use Promethus cloud-native monitoring technology to acquire and store the runtime information of the software and hardware environment in the service system in real time; Database storage module: stores the historical QoS data of the service system acquired in real time into the database and stores it persistently; Data preprocessing module: preprocesses the service system software and hardware environment runtime information, including data minimum and maximum normalization processing; Time series data reliability analysis module: uses a hybrid neural network and attention mechanism algorithm to mine the time series data of the software and hardware status of the service system during operation, and performs system reliability analysis and prediction on the time series data; Service system reliability information display module: displays the analyzed reliability information in a visual form; The control relationship between each module: the real-time QoS data of the service system is obtained through the data acquisition module, the service system QoS data is stored in the database, the service system QoS data is extracted from the database, the data preprocessing module performs preprocessing operations on the extracted data, and the preprocessed service history QoS data is put into the reliability analysis and prediction model trained in advance to obtain the service reliability level classification corresponding to the future effective prediction period, and the predicted service reliability situation is thrown into the information display module, the service future reliability prediction level information and the service system history QoS information obtained through the Web interface are combined, and the combined information is dynamically rendered in real time in the information display module; it is used to realize the reliability prediction of the service system; The prediction The following steps are involved: Among them, step 1: data minimum and maximum normalization, as follows: Use the data minimum and maximum normalization method to achieve data normalization of service QoS time series information, and standardize service response time, service throughput, and service reliability to [0,1]; where responseTime * ,throughput * ,reliability * is the normalized result of responseTime, throughput, and reliability. max ,throughput max ,reliability max It is the maximum value of service response time, service throughput, and service reliability in the service system QoS data. responseTime min ,throughput min ,reliability min is the service response time in the service system QoS data, The minimum value of service throughput and service reliability. Step 2: Time series data window division, as follows: divide the time series data window for the minimum and maximum normalized data, divide the data according to the order of historical service system QoS, leading time, and service system reliability data in the future effective prediction period, average the service system reliability data in the future effective prediction period window, and obtain the corresponding service system reliability classification label in the future effective prediction period window. Set the time series length of the service history QoS data to t, and the dimension of the service QoS to n. After data normalization and time series data window division, a t*n service history QoS matrix Q={q 1 ,q 2 ,...,q t } and a reliability level classification sequence L = {l 1 ,l 2 ,...,l t }, where q i ={responseTime i ,thoughtput i ,reliability i }; Step 3: Use multi-granularity LSTM to extract service QoS time series features; The details are as follows: LSTM with 128, 64, and 32 neural units is used to extract the characteristics of service QoS time, Q = {q 1 ,q 2 ,...,q t }As the input of the LSTM model, the output result can be obtained Step 4: Use the attention mechanism to extract service QoS sensitive information; the details are as follows: In the entire service QoS time slice, calculate Attention score The calculated attention score Use the softmax function to calculate the attention distribution value. Finally, calculate the feature vector R′ of the service QoS after attention extraction. Step 5: Service QoS feature integration, as follows: Fusion of service QoS time series characteristics R′ at each granularity 32 ,R′ 64 ,R′ 128 , the calculation formula is as follows: M=concat(R′ 32 ,R′ 64 ,R′ 128 ); Step 6: Use a multi-core convolutional neural network to extract service QoS feature profile information; the details are as follows: Using 3x64 and 2x32 convolution kernels, the specific convolution calculation formula is as follows: C 64 =W 3*64 M[i:i+63] (i=1,2,…,t) C 64 After batch normalization and relu activation function, we get C′ 64 C 32 =W 2*32 C′ 64 [i:i+31] (i=1,2,…,t) C 32 After batch normalization and relu activation function, we get C′ 32 ; Step 7: Maximum convolution pooling service QoS feature matrix; the details are as follows: R″′=max(C′ 32 ), Step 8: Use the fully connected layer to buffer the service reliability classification features; R″′=w 3 R″+b″′ Step 9: Calculation of loss function; details are as follows: Finally, sparse classification cross entropy is used as the loss function of AHN_SR, and the calculation formula is:
