Computer control system and method based on cloud platform

Abstract

The invention relates to the technical field of cloud platform computer control, and discloses a computer control system and method based on a cloud platform. The system comprises a multi-source data acquisition module which acquires heterogeneous data in real time and divides the heterogeneous data into time sequence segments; the data fusion preprocessing module is used for removing noise and screening key features to generate a standardized data stream; the dynamic decision control module outputs a control instruction set based on a deep reinforcement learning algorithm; the abnormal mode recognition module is used for recognizing anomalies by adopting spectral clustering and an isolated forest model; the resource flexible scheduling module is used for optimizing resource scheduling by using an improved genetic algorithm; and a safety audit module is also arranged to guarantee instruction safety. According to the system and the method, heterogeneous data of the cloud platform can be effectively acquired and processed, accurate decision control is realized, abnormity is accurately identified, resources are reasonably scheduled, system security, stability and resource utilization rate are improved, and complex service requirements of the cloud platform are met.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud platform computer control, and specifically to a computer control system and method based on a cloud platform. Background Art

[0002] With the rapid development of cloud computing technology, cloud platforms are increasingly widely used in various fields. Enterprises and organizations have migrated a large number of their operations to cloud platforms to obtain flexible computing resources, efficient data storage, and convenient service delivery capabilities. However, the complexity and dynamism of the cloud platform environment also bring a series of technical challenges.

[0003] In terms of data collection, there are a large number of heterogeneous data sources in the cloud platform, including various device operation parameters, complex and ever-changing network traffic matrices, and diverse user operation trajectories. Traditional data collection methods are difficult to achieve real-time and comprehensive collection of these heterogeneous data. For example, in a large data center, there are a large number of devices, and the collection frequencies and data formats of the operation parameters of different devices vary greatly. If a unified collection mode is adopted, either the high-precision collection requirements of some devices cannot be met, or resource waste will be caused due to over-collection. Moreover, network traffic surges occur frequently, and existing collection means are difficult to accurately capture the instantaneous changes in traffic, easily losing key information and affecting subsequent analysis and optimization of network conditions.

[0004] Data processing and fusion are another difficult problem. The raw data collected often contains a large amount of noise, and the data feature dimensions are complex and numerous, and some of the dimensions have a low correlation with actual application requirements. If the unprocessed data is directly used for analysis and decision-making, it will lead to large deviations in analysis results and inaccurate decisions. Taking equipment fault prediction as an example, the equipment operation data without noise removal may cause the prediction model to misjudge normal fluctuations as fault signals, while redundant feature dimensions will increase the complexity of model training, extend the training time, and reduce the real-time performance and reliability of the model.

[0005] In terms of system decision control, the dynamic characteristics of the cloud platform require the control system to make accurate and timely decisions based on real-time data. Traditional control strategies are usually based on pre-set rules and cannot adapt to the complex and ever-changing workloads and environmental changes in the cloud platform. For example, in the scenario of cloud server resource allocation, when the business traffic suddenly increases, the traditional fixed allocation strategy cannot quickly adjust resources, resulting in slow or even timed-out responses to some user requests, seriously affecting the user experience and the normal operation of the business.

[0006] Anomaly recognition is also an important issue faced by cloud platforms. Due to the huge amount of data and complex data patterns in cloud platforms, existing anomaly detection algorithms are difficult to quickly and accurately identify anomaly patterns. Once an anomaly occurs, if it cannot be detected and processed in time, it may trigger a chain reaction, resulting in a decline in system performance, data loss, or even system paralysis. For example, in a cloud storage system, if the anomalies of storage nodes cannot be detected in time, it may cause data corruption or loss, bringing huge losses to users.

[0007] Resource scheduling is a key link for the efficient operation of cloud platforms. In cloud platforms, computing resources, storage resources, network resources, etc. need to be reasonably allocated to numerous users and services. However, existing resource scheduling algorithms often only consider a single objective, such as pursuing the maximum throughput or minimum cost, and cannot comprehensively balance multiple objectives such as load balancing, energy consumption cost, and task completion delay. This leads to a situation where some servers may be overloaded while some servers are idle in actual applications, not only causing resource waste but also affecting the performance and stability of the overall system. Summary of the Invention

[0008] The purpose of the present invention is to provide a computer control system and method based on a cloud platform to solve the problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solution: A computer control system based on a cloud platform, the system includes: A multi-source data acquisition module, which collects heterogeneous data in real time through distributed edge nodes, including device operation parameters, network traffic matrices, and user operation trajectories, and divides the data into time series segments using a sliding time window; A data fusion and preprocessing module, which is used to remove the noise components of the time series segments, screen the key feature dimensions by combining mutual information entropy, and generate a fused standardized data stream; A dynamic decision-making control module, which constructs a policy network based on the deep reinforcement learning algorithm, inputs the standardized data stream and outputs a control instruction set, and updates the network weights through policy gradients to optimize the instruction generation logic; An anomaly pattern recognition module, which uses spectral clustering algorithm to divide the spatial distribution of the standardized data stream, combines the isolation forest model to calculate the local anomaly factor, and generates anomaly pattern labels and confidence scores; A resource elastic scheduling module, which performs multi-objective optimization on computing resources through an improved genetic algorithm, dynamically adjusts the virtual machine deployment strategy and bandwidth allocation ratio, and generates a resource scheduling plan.

[0010] Preferably, the specific implementation of the multi-source data acquisition module includes: In the data acquisition stage, the original waveform data of the device operation parameters is obtained through the sensor array of the edge node, and the time series decomposition algorithm is used to extract the trend term and the periodic term; Perform traffic shaping on the network traffic matrix to eliminate the time offset error of the burst traffic, and generate a traffic feature fingerprint through the mapping of the hash function; Divide the user operation trajectory into discrete event sequences according to the operation type, and use the state transition matrix to describe the association probability between events; If the normalized error of the time series segment exceeds the preset tolerance threshold, trigger the data resampling process and record the abnormal acquisition log.

[0011] Preferably, the specific process of the data fusion preprocessing module includes: Perform discrete wavelet transform on the time series segment, extract the high-frequency noise component through multi-scale decomposition, and use the soft threshold filtering algorithm to suppress the noise energy; Calculate the mutual information entropy value between each feature dimension and the target variable, and retain the feature dimensions with entropy values higher than the preset screening threshold; Perform tensor splicing on the screened feature dimensions and the traffic feature fingerprint, and smooth the local fluctuations of the data stream through the moving average algorithm to generate a standardized data stream.

[0012] Preferably, the deep reinforcement learning algorithm of the dynamic decision control module includes: Construct a dual network architecture of the policy network and the value network, and the output layer of the policy network uses the Softmax function to generate the probability distribution of the control instruction; Design the reward function as a linear combination of the system stability index and the resource consumption rate, and update the parameters of the value network through the time difference error; Synchronize the weight parameters of the policy network and the target network every preset period. If the variance of the reward function for consecutive multiple iterations exceeds the stability threshold, reset the exploration rate parameter.

[0013] Preferably, the implementation of the spectral clustering algorithm of the abnormal pattern recognition module includes: Calculate the similarity matrix of the standardized data stream, and obtain the eigenvector space through the Laplacian matrix transformation; Perform K-means clustering on the eigenvectors, divide the spatial distribution area of the data stream, and calculate the density ratio of each area to mark the sparse area; Input the data in the sparse area into the isolation forest model, calculate the local outlier factor through the path length, and generate an abnormal pattern label if the factor value exceeds the dynamic threshold.

[0014] Preferably, the method for determining the dynamic threshold includes: Statistically calculate the mean and standard deviation of historical local anomaly factors, and fit a Gaussian distribution model based on the data distribution within the sliding window; Calculate the initial static threshold through the quantile method, and adjust the threshold offset according to the data skewness coefficient of the current window; If no anomaly label is detected in multiple consecutive windows, reduce the threshold sensitivity exponentially; if anomaly labels are continuously detected, increase the threshold sensitivity linearly.

[0015] Preferably, the improved genetic algorithm of the resource elastic scheduling module includes: Design the chromosome coding scheme as a binary combination of virtual machine deployment locations and bandwidth allocation ratios, and set the adaptive adjustment rules for the crossover probability and mutation probability; Construct a multi-objective fitness function, including the weighted sum of the load balance degree, energy consumption cost, and task completion delay; Adopt the elitist retention strategy to avoid the loss of high-quality solutions, and screen the non-dominated solution set through the Pareto front after each generation of evolution; If it is detected that the convergence speed of the fitness function is lower than the preset threshold, trigger the population diversity enhancement mechanism.

[0016] Preferably, the specific implementation of the adaptive adjustment rule includes: Dynamically adjust the crossover probability according to the dispersion degree of the individual fitness in the population. The higher the dispersion degree, the lower the crossover probability; Calculate the adjustment coefficient of the mutation probability through the fitness difference between the current generation and the parent generation. The smaller the difference, the higher the mutation probability; If it is detected that the population falls into a local optimum for more than the preset number of generations, reset the crossover and probability parameters to their initial values.

[0017] Preferably, the system further includes: A security audit module, which is used to perform compliance verification on the control instruction set, verify the legality of the instruction source using digital signature technology, and record the operation logs through blockchain deposit technology; If an unsigned instruction or inconsistent deposit hash value is detected, freeze the current control instruction and trigger a security warning signal.

[0018] Preferably, the present invention further includes a computer control method based on a cloud platform, which is applied to the above-mentioned computer control system based on a cloud platform. The method includes the following steps: Use the distributed edge nodes to utilize the multi-source data acquisition module to collect heterogeneous data such as device operation parameters, network traffic matrices, and user operation trajectories in real time, and use a sliding time window to divide the collected data into time series segments; The time series segments are processed by the data fusion preprocessing module to remove the noise components therein, and the key feature dimensions are screened by combining mutual information entropy to generate a fused standardized data stream; Based on the deep reinforcement learning algorithm, a policy network of the dynamic decision control module is constructed. The standardized data stream is input into the policy network to output a control instruction set, and the network weights are updated by policy gradient to optimize the instruction generation logic; The spectral clustering algorithm is used to divide the spatial distribution of the standardized data stream, and the local outlier factor is calculated by combining the isolation forest model. The outlier pattern label and confidence score are generated by the outlier pattern recognition module; The improved genetic algorithm is used to perform multi-objective optimization on the computing resources by the resource elastic scheduling module, dynamically adjust the virtual machine deployment policy and bandwidth allocation ratio, and generate a resource scheduling plan.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: The computer control system and method based on the cloud platform provided by the present invention show significant beneficial effects in multiple aspects. In the data acquisition and processing link, the multi-source data acquisition module realizes the real-time acquisition of heterogeneous data such as device operation parameters, network traffic matrices, and user operation trajectories with the help of distributed edge nodes. The data is segmented into time series segments by a sliding time window, laying a foundation for subsequent precise processing. Moreover, during the acquisition process, specific processing methods are adopted for different types of data. For example, trend terms and periodic terms are extracted from device operation parameters, making the data features more prominent, which helps to analyze the long-term change trend and periodic law of the device operation state. The data fusion preprocessing module removes noise components and screens key feature dimensions, improving the data quality. The discrete wavelet transform and soft threshold filtering algorithm are used to suppress noise, which can effectively avoid the influence of noise interference on the data analysis results; the feature dimensions are screened by combining mutual information entropy to remove redundant information, which not only reduces the data processing volume but also improves the data effectiveness, enabling the subsequent decision control module to make decisions based on more accurate data.

[0020] The dynamic decision control module constructs a policy network based on the deep reinforcement learning algorithm, and optimizes the instruction generation logic by updating the network weights through policy gradient. In this way, a control instruction set can be dynamically generated according to the real-time data of the cloud platform. Compared with the traditional fixed-rule decision-making method, it has stronger adaptability. For example, when the business load of the cloud platform changes dynamically, this module can adjust the server resource allocation strategy in real time to ensure that the system can operate stably and efficiently under different load conditions, improve the system response speed and throughput, reduce the average response time of user requests, and enhance the user experience. At the same time, the designed reward function comprehensively considers the system stability index and resource consumption rate, balances the system performance and resource utilization efficiency, avoids excessive resource consumption, and realizes the reasonable utilization of resources.

[0021] The abnormal pattern recognition module combines the spectral clustering algorithm and the isolation forest model, which can accurately divide the spatial distribution of the standardized data stream, accurately calculate the local anomaly factor, and generate abnormal pattern labels and confidence scores. Compared with traditional anomaly detection methods, it has higher detection accuracy and lower false alarm rate. In the complex data environment of the cloud platform, it can timely detect abnormal situations such as equipment failures and network attacks, take measures in advance to handle them, effectively ensure the security and stability of the system, and reduce the system downtime and data loss caused by anomalies.

[0022] The resource elastic scheduling module uses an improved genetic algorithm to perform multi-objective optimization on computing resources, and dynamically adjusts the virtual machine deployment strategy and bandwidth allocation ratio. By constructing a multi-objective fitness function that includes load balance degree, energy consumption cost, and task completion delay, the comprehensive balance of multiple objectives is achieved. For example, while ensuring the timely completion of tasks, it reduces system energy consumption, improves resource utilization rate, avoids excessive idle or overload of server resources, enhances the overall resource management efficiency of the cloud platform, and reduces operating costs. Moreover, it adaptively adjusts the crossover probability and mutation probability, and triggers the population diversity enhancement mechanism, effectively avoiding the algorithm falling into local optima, improving the search ability of the algorithm, and ensuring the generation of a better resource scheduling plan.

[0023] The security audit module performs compliance verification on the control instruction set, uses digital signature technology to verify the legality of the instruction source, and records operation logs through blockchain deposit technology. This greatly enhances the security and traceability of the system, prevents the execution of illegal instructions, and ensures the security of data and services in the cloud platform. Once a security problem occurs, the operation records can be quickly and accurately traced, which is convenient for timely discovery and handling of security risks, and provides a strong guarantee for the safe and stable operation of the cloud platform. Description of the Drawings

[0024] Figure 1 It is the working principle diagram of the computer control system based on the cloud platform described in the present invention; Figure 2 It is the flowchart of the deep reinforcement learning algorithm of the dynamic decision control module; Figure 3 It is the flowchart of the spectral clustering algorithm of the abnormal pattern recognition module; Figure 4 It is the flowchart of the dynamic threshold determination of the abnormal pattern recognition module. Detailed Embodiments

[0025] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0026] Please refer to Figures 1 - 4 , the present invention provides a computer control system based on a cloud platform, which aims to achieve efficient acquisition, processing, decision control, anomaly recognition, and reasonable scheduling of various types of data in the cloud platform. Specifically, the system includes the following core modules working together: Multi-source data acquisition module: Heterogeneous data is collected in real time by means of distributed edge nodes. This data covers device operation parameters, network traffic matrices, and user operation trajectories. After the data is collected, it is segmented into time series segments using a sliding time window, providing basic data units for subsequent processing.

[0027] Data fusion and preprocessing module: This module is responsible for removing the noise components in the time series segments, screening out the key feature dimensions by combining mutual information entropy, and finally generating a fused and standardized data stream, making the data more usable and better supporting subsequent analysis and decision-making.

[0028] Dynamic decision control module: A policy network is constructed based on the deep reinforcement learning algorithm, taking the standardized data stream as the input and outputting a control instruction set. During operation, the network weights are continuously updated through policy gradients to optimize the instruction generation logic, thereby achieving dynamic and precise control of the system.

[0029] Anomaly pattern recognition module: The spectral clustering algorithm is used to divide the spatial distribution of the standardized data stream, and then the local outlier factor is calculated in combination with the isolation forest model, and then the anomaly pattern labels and corresponding confidence scores are generated to timely detect anomalies in the system and ensure the stable operation of the system.

[0030] Resource elastic scheduling module: The improved genetic algorithm is used to perform multi-objective optimization on computing resources, dynamically adjust the virtual machine deployment strategy and bandwidth allocation ratio, and finally generate a reasonable resource scheduling plan to improve resource utilization and meet the needs of different services.

[0031] The present invention will be further described below in conjunction with Embodiments 1 to 6: Embodiment 1: In the data acquisition stage, the original waveform data of the device operation parameters is obtained through the sensor array deployed on the edge nodes. Since the original waveform data may contain various trends and periodic components, in order to better analyze and utilize this data, a time series decomposition algorithm is adopted for processing. For example, parameters such as temperature and voltage during the device operation change over time, and the time series decomposition algorithm can decompose them into a trend term and a periodic term. The trend term reflects the change trend of the device operation parameters over a long period of time. For instance, as the device usage years increase, the trend of the operation temperature gradually rising; the periodic term reflects the fluctuation law of the parameter within a fixed period, such as the voltage fluctuation caused by the periodic load change of some devices.

[0032] For the network traffic matrix, in order to eliminate the time offset error caused by burst traffic, traffic shaping processing is performed. Burst traffic may cause abnormal fluctuations in network traffic over time, affecting the judgment of the true situation of network traffic. After traffic shaping processing, a traffic feature fingerprint is generated through hash function mapping. The hash function can map complex traffic data into a feature fingerprint with a fixed length, facilitating subsequent storage and comparison. For example, common MD5 hash function or SHA - 256 hash function can be used for this purpose.

[0033] The user operation trajectories are divided into discrete event sequences according to the operation types, such as user login, click, query and other operations. A state transition matrix is used to describe the association probability between these events. Assume that the main operations of the user in the system are three types: login (A), browse page (B), and conduct search (C). The state transition matrix can represent the probability of transitioning from one operation state to another. For example, the probability of transitioning from the login state A to the browse page state B is P(A→B), and the probability of transitioning from the browse page state B to the conduct search state C is P(B→C), etc.

[0034] In addition, if the normalized error of the time series segment exceeds the preset tolerance threshold, this means that the collected data may be abnormal. At this time, the data resampling process will be triggered to re - collect the relevant data to ensure the accuracy of the data. Meanwhile, an abnormal collection log is recorded, detailing information such as the time when the abnormality occurred and the data types involved, facilitating subsequent analysis and troubleshooting of data abnormalities.

[0035] Example 2: The data fusion pre - processing module first performs discrete wavelet transform on the time series segment. Discrete wavelet transform is a multi - resolution analysis method that can decompose the signal at different frequency scales. Through multi - scale decomposition, the high - frequency noise components in the time series segment can be extracted. For example, in the time series segment of device operation parameters, the high - frequency noise may be caused by the electromagnetic interference of the device or the error of the measuring instrument.

[0036] Next, the soft threshold filtering algorithm is used to suppress the noise energy. The principle of the soft threshold filtering algorithm is that for wavelet coefficients, when the absolute value of the coefficient is less than a certain threshold, it is set to 0; when the absolute value of the coefficient is greater than or equal to the threshold, it is shrunk to near the threshold. Suppose the wavelet coefficient is , and the threshold is . The coefficient after soft threshold filtering is calculated as follows:

[0037] where is the sign function. When , ; when , ; when , .

[0038] Then, the mutual information entropy value between each feature dimension and the target variable is calculated. Mutual information entropy is used to measure the correlation between two variables, and the feature dimensions with entropy values higher than the preset screening threshold are retained. For example, when analyzing the relationship between equipment operation parameters and equipment failures, the mutual information entropy values of some feature dimensions such as equipment temperature and rotation speed with equipment failures are relatively high, indicating that they have a strong correlation with equipment failures, and these feature dimensions will be retained.

[0039] Finally, the screened feature dimensions are tensor-concatenated with the flow feature fingerprint, and the local fluctuations of the data stream are smoothed by the moving average algorithm. The moving average algorithm can smooth the data stream, remove the short-term fluctuations in the data, and make the data trend more obvious. Suppose the data stream is , the moving window size is , then the data stream after moving average is calculated as follows: , where . Through such processing, a standardized data stream is finally generated.

[0040] Example 3: The dynamic decision control module constructs a policy network based on the deep reinforcement learning algorithm, adopting a dual-network architecture, namely the policy network and the value network. The output layer of the policy network uses the Softmax function to generate the probability distribution of control instructions. The Softmax function can transform the output of the policy network into a probability form, so that each control instruction has a corresponding occurrence probability. Suppose the output of the policy network is , representing scores corresponding to different control instructions. After being processed by the Softmax function, the probability of the th control instruction The calculation formula is:

[0041] The designed reward function is a linear combination of the system stability index and the resource consumption rate. The system stability index can be measured by parameters such as the response time and throughput of the system, while the resource consumption rate reflects the consumption of computing resources, storage resources, etc. during the operation of the system. Assume that the system stability index is , and the resource consumption rate is , then the reward function can be expressed as:

[0042] where is the weight coefficient, and its value range is between , which is used to adjust the relative importance of the system stability index and the resource consumption rate in the reward function.

[0043] During the training process, the parameters of the value network are updated through the temporal difference error. The temporal difference error refers to the difference between the estimated value and the actual value in the current state. The weight parameters of the policy network and the target network are synchronized every preset period, which can make the parameters of the target network relatively stable and avoid excessive fluctuations during the training process. If the variance of the reward function exceeds the stability threshold for several consecutive iterations, it indicates that the current exploration rate may be inappropriate, which will lead to unstable training. At this time, the exploration rate parameter is reset to readjust the balance between exploration and exploitation, so as to ensure that the policy network can better learn the optimal policy and output a more reasonable control instruction set.

[0044] Example 4: The implementation process of the spectral clustering algorithm in the abnormal mode recognition module is relatively complex. First, calculate the similarity matrix of the standardized data stream. The similarity matrix is used to measure the similarity degree between data points in the dataset, and usually methods such as Euclidean distance and cosine similarity can be used for calculation. Assume that there are data points in the standardized data stream, and the element of the similarity matrix can be expressed as:

[0045] where represents the function for calculating the similarity between and .

[0046] Then, obtain the eigenvector space through the Laplacian matrix transformation. The Laplacian matrix is a commonly used graph theory tool that can reflect the connection relationship between data points. The Laplacian matrix Related to the similarity matrix After the Laplacian matrix transformation, the eigenvector space can be obtained.

[0047] Next, perform K-means clustering on the eigenvectors to divide the spatial distribution area of the data stream. K-means clustering is a commonly used clustering algorithm that divides data points into clusters, so that the data points within the clusters have a high similarity, and the data points between the clusters have a low similarity. After dividing the spatial distribution area, calculate the density ratio of each area to mark the sparse areas. The density ratio can be obtained by calculating the ratio of the number of data points in the area to the volume of the area. Sparse areas are usually areas with fewer data points, and the data in these areas is more likely to be abnormal data.

[0048] Input the data in the sparse area into the Isolation Forest model, and calculate the Local Outlier Factor through the path length. The Isolation Forest model is a tree-based anomaly detection algorithm that isolates data points by constructing binary trees. The Local Outlier Factor (LOF) is used to measure the anomaly degree of a data point. Assume that the data point The path length in the Isolation Forest model is The calculation of the Local Outlier Factor involves comparing with the path lengths of neighboring points. If the factor value exceeds the dynamic threshold, an anomaly pattern label is generated.

[0049] The method for determining the dynamic threshold is as follows: First, calculate the mean and standard deviation of the historical Local Outlier Factors, and fit a Gaussian distribution model according to the data distribution within the sliding window. The Gaussian distribution model can better describe the distribution characteristics of the data. Calculate the initial static threshold through the quantile method. The quantile method determines the threshold according to the position of the data in the distribution. Then, adjust the threshold offset according to the data skewness coefficient of the current window. The data skewness coefficient reflects the degree of asymmetry of the data distribution. If the data skewness is large, it means that the data distribution has a certain bias, and the threshold needs to be adjusted accordingly. If no anomaly labels are detected in multiple consecutive windows, it means that the current threshold may be too high, and the threshold sensitivity is reduced exponentially; if anomaly labels are continuously detected, the threshold sensitivity is increased linearly to more accurately identify abnormal data.

[0050] Example 5: The improved genetic algorithm of the resource elastic scheduling module plays an important role in realizing the multi-objective optimization of computing resources. First, design the chromosome encoding scheme as a binary combination of the virtual machine deployment location and the bandwidth allocation ratio. For example, assume there are virtual machines and available deployment locations, as well as different bandwidth allocation levels. The chromosome can be represented as a length of The binary string. Among them, the first bits represent the deployment location of the virtual machine, and the last bits represent the bandwidth allocation ratio.

[0051] Set the adaptive adjustment rules for the crossover probability and mutation probability. Dynamically adjust the crossover probability according to the dispersion degree of the individual fitness in the population. The higher the dispersion degree, the lower the crossover probability. The dispersion degree of individual fitness can be measured by calculating the variance of the individual fitness in the population. Assume that the individual fitness in the population is , and the mean is , and the variance The calculation formula is:

[0052] When the variance is large, it indicates that there are large differences among individuals in the population. At this time, reduce the crossover probability to avoid excessive destruction of excellent individuals.

[0053] Calculate the adjustment coefficient of the mutation probability through the fitness difference between the current generation and the parent generation. The smaller the difference, the higher the mutation probability. Assume that the average fitness of the current generation population is , and the average fitness of the parent generation population is , and the mutation probability adjustment coefficient can be expressed as:

[0054] The mutation probability can be adjusted according to the adjustment coefficient, such as , where is the initial mutation probability.

[0055] If it is detected that the population has fallen into local optimum for more than the preset number of generations, reset the crossover and mutation probability parameters to the initial values to reintroduce population diversity and avoid premature convergence of the algorithm.

[0056] Construct a multi-objective fitness function, including the weighted sum of the load balance degree, energy consumption cost, and task completion delay. Assume that the load balance degree is , the energy consumption cost is , the task completion delay is , and the multi-objective fitness function can be expressed as:

[0057] Among them, , , are weight coefficients, and , and they adjust the relative importance of the load balance degree, energy consumption cost, and task completion delay in the fitness function according to actual needs.

[0058] The elitist retention strategy is adopted to avoid the loss of high-quality solutions, that is, the optimal individuals in the current generation are directly retained in the next generation. After each generation of evolution, the non-dominated solution set is screened through the Pareto front. The Pareto front refers to a set of solutions in a multi-objective optimization problem that cannot be simultaneously outperformed by other solutions in all objectives. By screening the non-dominated solution set, a set of optimal solutions that achieve a balance among different objectives can be obtained. If it is detected that the convergence rate of the fitness function is lower than the preset threshold, it indicates that the search efficiency of the algorithm may decrease. At this time, the population diversity enhancement mechanism is triggered, such as introducing new individuals or mutating existing individuals to increase the population diversity and improve the search ability of the algorithm, and finally generate a better resource scheduling scheme.

[0059] Example 6: In the entire computer control system, the security audit module is mainly responsible for the compliance verification of the control instruction set. In the actual operation process, digital signature technology is used to verify the legitimacy of the instruction source. Digital signature technology is based on the public key cryptosystem. The sender uses its own private key to sign the control instruction, and the receiver uses the sender's public key to verify the signature.

[0060] At the same time, the operation log is recorded through blockchain evidence storage technology. Blockchain is a distributed ledger technology with characteristics such as immutability and traceability. The operation log is recorded on the blockchain, and each operation record contains information such as the operation time, operation content, and operator. These records are packaged into blocks and linked in chronological order. When querying the operation record, the corresponding operation log can be obtained through the query interface of the blockchain according to relevant conditions (such as time range, operation type, etc.).

[0061] If an unsigned instruction or inconsistent deposit hash value is detected, it means that there may be a security risk in the control instruction. At this time, the system will freeze the current control instruction to prevent its further execution and prevent possible security vulnerabilities from being exploited. At the same time, a security warning signal is triggered to notify the system administrator or relevant security personnel for handling. The security warning signal can be sent in various ways, such as email notification, SMS reminder, or displaying prominent warning information on the system management interface, etc., so as to timely discover and handle security problems in the system and ensure the stable operation of the entire cloud platform-based computer control system.

[0062] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0063] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A computer control system based on a cloud platform, characterized in that, Including: A multi-source data acquisition module that real-time collects heterogeneous data through distributed edge nodes, including device operation parameters, network traffic matrices, and user operation trajectories, and divides the data into time-series segments using a sliding time window; A data fusion preprocessing module for removing the noise components of the time-series segments, screening key feature dimensions by combining mutual information entropy, and generating a fused standardized data stream; A dynamic decision-making and control module that constructs a policy network based on a deep reinforcement learning algorithm, inputs the standardized data stream and outputs a control instruction set, and updates the network weights through policy gradients to optimize the instruction generation logic; An abnormal pattern recognition module that uses a spectral clustering algorithm to divide the spatial distribution of the standardized data stream, combines an isolation forest model to calculate the local abnormal factor, and generates abnormal pattern labels and confidence scores; A resource elastic scheduling module that performs multi-objective optimization on computing resources through an improved genetic algorithm, dynamically adjusts the virtual machine deployment strategy and bandwidth allocation ratio, and generates a resource scheduling plan.

2. The computer control system according to claim 1, wherein The specific implementation of the multi-source data acquisition module includes: In the data acquisition stage, the original waveform data of device operation parameters is obtained through the sensor array of the edge node, and the trend term and periodic term are extracted using a time series decomposition algorithm; Perform traffic shaping on the network traffic matrix to eliminate the time offset error of burst traffic, and generate a traffic feature fingerprint through hash function mapping; Divide the user operation trajectory into discrete event sequences according to the operation type, and use a state transition matrix to describe the association probability between events; If the normalized error of the time-series segment is detected to exceed the preset tolerance threshold, trigger the data resampling process and record the abnormal acquisition log.

3. The computer control system according to claim 1, wherein, The specific process of the data fusion preprocessing module includes: Perform discrete wavelet transform on the time-series segment, extract high-frequency noise components through multi-scale decomposition, and use a soft threshold filtering algorithm to suppress the noise energy; Calculate the mutual information entropy value between each feature dimension and the target variable, and retain the feature dimensions with entropy values higher than the preset screening threshold; Perform tensor splicing on the screened feature dimensions and the traffic feature fingerprint, and smooth the local fluctuations of the data stream through a moving average algorithm to generate a standardized data stream.

4. The computer control system according to claim 1, wherein The deep reinforcement learning algorithm of the dynamic decision-making and control module includes: Construct a dual-network architecture of a policy network and a value network, and the output layer of the policy network uses the Softmax function to generate the probability distribution of control instructions; Design the reward function as a linear combination of the system stability index and the resource consumption rate, and update the parameters of the value network through the time difference error; Synchronize the weight parameters of the policy network and the target network every preset period. If the variance of the reward function for consecutive multiple iterations exceeds the stability threshold, reset the exploration rate parameter.

5. The computer control system according to claim 1, characterized in that, The implementation of the spectral clustering algorithm of the abnormal pattern recognition module includes: Calculate the similarity matrix of the standardized data stream, and obtain the eigenvector space through Laplacian matrix transformation; Perform K-means clustering on the eigenvectors, divide the spatial distribution area of the data stream, and calculate the density ratio of each area to mark the sparse areas; Input the data in the sparse area into the isolation forest model, calculate the local abnormal factor through the path length, and generate an abnormal pattern label if the factor value exceeds the dynamic threshold.

6. The computer control system according to claim 5, characterized in that, The method for determining the dynamic threshold includes: Statistically calculate the mean and standard deviation of historical local anomaly factors, and fit a Gaussian distribution model according to the data distribution within the sliding window; Calculate the initial static threshold through the quantile method, and adjust the threshold offset according to the data skewness coefficient of the current window; If no anomaly label is detected in multiple consecutive windows, reduce the threshold sensitivity exponentially; if anomaly labels are continuously detected, increase the threshold sensitivity linearly.

7. The computer control system according to claim 1, wherein The improved genetic algorithm of the resource elastic scheduling module includes: Design the chromosome coding scheme as a binary combination of virtual machine deployment locations and bandwidth allocation ratios, and set the adaptive adjustment rules for the crossover probability and mutation probability; Construct a multi-objective fitness function, including the weighted sum of the load balance degree, energy consumption cost, and task completion delay; Adopt the elitist retention strategy to avoid the loss of high-quality solutions, and screen the non-dominated solution set through the Pareto front after each generation of evolution; If it is detected that the convergence speed of the fitness function is lower than the preset threshold, trigger the population diversity enhancement mechanism.

8. The computer control system according to claim 7, wherein The specific implementation of the adaptive adjustment rules includes: Dynamically adjust the crossover probability according to the dispersion degree of the individual fitness in the population. The higher the dispersion degree, the lower the crossover probability; Calculate the adjustment coefficient of the mutation probability through the fitness difference between the current generation and the parent generation. The smaller the difference, the higher the mutation probability; If it is detected that the population falls into the local optimum for more than the preset number of generations, reset the crossover and probability parameters to the initial values.

9. The computer control system according to claim 1, wherein It also includes: A security audit module, which is used to perform compliance verification on the control instruction set, verify the legality of the instruction source using digital signature technology, and record the operation log through blockchain evidence storage technology; If an unsigned instruction or inconsistent evidence storage hash value is detected, freeze the current control instruction and trigger a security warning signal.

10. A computer control method based on a cloud platform, applied to the computer control system according to any one of claims 1-9, characterized in that, It includes the following steps: Use the distributed edge nodes to collect heterogeneous data such as device operation parameters, network traffic matrices, and user operation trajectories in real time through the multi-source data collection module, and use a sliding time window to divide the collected data into time series segments; Use the data fusion preprocessing module to process the time series segments, remove the noise components therein, screen the key feature dimensions in combination with mutual information entropy, and generate a fused standardized data stream; Build a policy network for the dynamic decision control module based on the deep reinforcement learning algorithm, input the standardized data stream into the policy network, output the control instruction set, and update the network weights through policy gradients to optimize the instruction generation logic; Use the spectral clustering algorithm to divide the spatial distribution of the standardized data stream, calculate the local anomaly factor in combination with the isolation forest model, and generate anomaly pattern labels and confidence scores through the anomaly pattern recognition module; Use the improved genetic algorithm to perform multi-objective optimization on the computing resources through the resource elastic scheduling module, dynamically adjust the virtual machine deployment strategy and bandwidth allocation ratio, and generate a resource scheduling plan.

Images