Enterprise digital management method and system based on big data processing

Abstract

This invention discloses a method and system for enterprise digital management based on big data processing, belonging to the field of enterprise digital management technology. It constructs a multi-dimensional heterogeneous time-series data tensor; obtains a high signal-to-noise ratio (SNR) operational feature sequence; constructs a dynamic spatiotemporal business association graph; performs spatial topological feature aggregation between multi-hop nodes and temporal evolution feature extraction based on long short-term memory networks, forming a multi-scale global state evolution feature representation; generates a comprehensive state decision feature map; outputs and automatically issues corresponding sets of dynamic elastic allocation of underlying computing resources and network isolation control instructions for abnormal nodes, completing closed-loop scheduling management of enterprise digital management. This invention significantly improves the accuracy of underlying feature extraction and the utilization rate of computing resources.

Description

Technical Field

[0001] This invention relates to the field of enterprise digital management technology, and in particular to an enterprise digital management method and system based on big data processing. Background Technology

[0002] With the deepening of enterprise digital transformation and the widespread application of microservice architecture, the operational data generated by enterprise business systems is showing a trend of cross-domain, multimodal and massive explosion. Centralized monitoring and management platforms based on big data processing have become the standard architecture for maintaining the stable operation of enterprise underlying infrastructure and upper-level business.

[0003] Traditional monitoring data acquisition methods typically rely on probes independently deployed in various hardware or middleware components. The heterogeneous data from multiple sources exhibits severe disorder and dispersion in timestamps and sampling frequencies. This makes it difficult for management systems to perform precise multi-dimensional time-series alignment when faced with massive concurrent data streams. Furthermore, collecting and processing all raw data containing significant amounts of normal background noise consumes substantial storage and computing resources and can easily mask weak anomalous signals in the early stages of system mutations, resulting in extremely low signal-to-noise ratios for feature extraction.

[0004] As the complexity of enterprise business applications becomes increasingly modularized, ensuring efficient global scheduling becomes more and more difficult. Traditional system status assessment mechanisms often treat the underlying physical servers and the upper-layer business logic in isolation. Relying solely on static physical network topology to check the health of isolated nodes cannot effectively capture the tightness of business collaboration between microservices caused by dynamic calls. When faced with cascading failures across nodes or hidden resource exhaustion, it is difficult to accurately extract the resonance characteristics of spatial topology evolution trends and potential risks between multi-hop physical nodes.

[0005] Traditional server resource scheduling and anomaly management mechanisms are typically designed for passive response environments, relying on static threshold rules preset by human experience. They cannot comprehensively consider the global state evolution of heterogeneous time-series data. Rule-matching mechanisms based on lagging indicators cannot perform forward-looking elastic allocation of underlying computing power, struggle to balance load balancing and decision diversity under complex global conditions, and are unable to autonomously and rapidly output network-level isolation control commands for specific abnormal nodes when facing unknown anomalies. This results in delays in enterprise digital scheduling and significant waste of idle resources. Summary of the Invention

[0006] One objective of this invention is to propose a digital management method and system for enterprises based on big data processing. This invention significantly improves the accuracy of underlying feature extraction and the utilization rate of computing resources.

[0007] A digital management method for enterprises based on big data processing according to an embodiment of the present invention includes: Collect cross-domain multimodal operation data streams from various physical business nodes and underlying servers of the enterprise, perform time-series alignment and spatial dimension mapping based on dynamic time windows on the cross-domain multimodal operation data streams, and construct multi-dimensional heterogeneous time-series data tensors; For multi-dimensional heterogeneous time-series data tensors, the local information entropy difference of different data dimensions is calculated, and adaptive feature filtering and redundancy removal processing based on information entropy threshold is performed to obtain high signal-to-noise ratio running feature sequences. Based on the high signal-to-noise ratio running feature sequence, the feature vector of the current active business flow is extracted, and it is mapped to the pre-configured enterprise underlying server physical communication topology to construct a dynamic spatiotemporal business association graph. The dynamic spatiotemporal business association map is input into a pre-constructed spatiotemporal graph convolutional neural network to perform spatial topological feature aggregation between multi-hop nodes and temporal evolution feature extraction based on long short-term memory network, forming a multi-scale global state evolution feature representation. The multi-scale global state evolution feature representation and the high signal-to-noise ratio running feature sequence are cross-attentionally concatenated in the channel dimension to generate a comprehensive state decision feature map. The comprehensive state decision feature map is then input into a pre-trained deep reinforcement learning decision model to solve the policy. The model outputs and automatically issues the corresponding set of dynamic elastic allocation of underlying computing resources and network isolation control instructions for abnormal nodes, thus completing the enterprise's digital closed-loop scheduling management.

[0008] Optionally, the process of collecting cross-domain multimodal operational data streams from various physical business nodes and underlying servers of the enterprise, performing time-series alignment and spatial dimension mapping based on dynamic time windows on the cross-domain multimodal operational data streams, and constructing a multi-dimensional heterogeneous time-series data tensor includes: By collecting operational data from each physical business node and underlying server using data probes, raw multimodal data streams with disordered and discrete timestamps are obtained. Extract the first derivative of historical network traffic to calculate the fluctuation variance and correction coefficient, and dynamically generate an adaptive current time window width; Based on the current time window width, local mean smoothing downsampling or cubic spline interpolation compensation is performed on the original multimodal data stream to obtain a strictly aligned multimodal synchronization time series; Extract the physical network address of the multimodal synchronization time sequence, query the asset database to map it to the logical business node number, so as to establish a four-dimensional coordinate system of space, time, mode and feature; Based on the four-dimensional coordinate system, multi-dimensional splicing and stacking operations are performed on the mapped multimodal synchronous time series to instantiate and construct a multi-dimensional heterogeneous time series data tensor.

[0009] Optionally, for multi-dimensional heterogeneous time-series data tensors, the step of calculating the local information entropy differences of different data dimensions and performing adaptive feature filtering and redundancy removal based on information entropy thresholds to obtain high signal-to-noise ratio running feature sequences includes: The time-dimensional slice sliding is performed on the multi-dimensional heterogeneous time-series data tensor, and the empirical probability distribution of each segment is reconstructed using the Gaussian kernel function to obtain a continuous probability density curve. Discretization calculus and logarithmic operations are performed on the probability density curve to calculate the Shannon entropy, a local information that precisely quantifies the degree of data disorder. By combining the global information entropy mean and standard deviation, the local information Shannon entropy value is subjected to elastic multiple calculation to dynamically generate a real-time updated information entropy discrimination threshold; Information entropy is used to identify thresholds to compare and judge multi-dimensional heterogeneous time series tensors. Principal component analysis is performed to reduce the dimension of low-entropy stable data and remove it, while high-entropy abrupt data is completely preserved to form a highly concentrated and pure feature set. The clean feature set is cleaned and reorganized in the cache array, and the output is a high signal-to-noise ratio running feature sequence with no false positives and true negatives.

[0010] Optionally, the step of extracting feature vectors of currently active service flows based on high signal-to-noise ratio running feature sequences, and mapping these vectors to pre-configured enterprise underlying server physical communication topology to construct a dynamic spatiotemporal service association graph includes: The high signal-to-noise ratio running feature sequence is input into the multi-head self-attention module, the probability weight distribution of historical state and current state is calculated and weighted summed, and the feature vector of the current active business flow is extracted. Read the underlying static physical communication topology data to obtain the adjacency matrix and edge attribute matrix that record physical links and bandwidth delay parameters; The feature vector of the current active business flow is accurately mapped to the physical node coordinates corresponding to the adjacency matrix, generating a rich attribute node carrying a high-dimensional running state; Features of adjacent rich attribute nodes are extracted and concatenated, input into a feedforward neural network, and combined with the edge attribute matrix to calculate service relevance, generating a comprehensive edge weight value that integrates transmission limit and service concurrency. By using rich attribute nodes as graph attributes and comprehensive edge weights as connection strengths, and persistently instantiating them in a graph database, a dynamic spatiotemporal business association graph is constructed.

[0011] Optionally, the step of inputting the dynamic spatiotemporal service association map into a pre-constructed spatiotemporal graph convolutional neural network to perform spatial topological feature aggregation between multi-hop nodes and temporal evolution feature extraction based on long short-term memory networks to form a multi-scale global state evolution feature representation includes: The dynamic spatiotemporal business association map is used as a time step snapshot and input into a spatiotemporal graph convolutional neural network framework that incorporates a multi-hop resonance attenuation factor; In the spatial dimension, the inner product score of the central node and its multi-order neighbor nodes is calculated and the topological decay coefficient is applied. Then, graph attention weighted summation is performed to complete the topological aggregation and obtain the spatial feature matrix. In the time dimension, the spatial feature matrix is ​​concatenated with the historical state and then input into the long short-term memory network. After cross-clock cycle calculations through the forget gate, input gate and output gate, the newly added temporal hidden state is extracted. The newly extracted temporal hidden states from different receptive fields are spliced ​​and aligned according to the channel dimension, and deep fusion is performed to obtain a multi-scale global state evolution feature representation.

[0012] Optionally, the step of cross-attention concatenating the multi-scale global state evolution feature representation with the high signal-to-noise ratio running feature sequence in the channel dimension to generate a comprehensive state decision feature map, inputting the comprehensive state decision feature map into a pre-trained deep reinforcement learning decision model for policy solving, outputting and automatically issuing corresponding control commands to complete the enterprise's digital closed-loop scheduling management, includes: The deep multi-scale global state evolution feature representation and the shallow high signal-to-noise ratio running feature sequence are cross-attention covariance evaluated and weighted and spliced ​​to generate a comprehensive state decision feature map. The comprehensive state decision feature map is used as an environmental observation variable and input into the maximum entropy multi-agent deep reinforcement learning decision model with a reward term regularized by distribution entropy value. Through forward propagation calculation of the decision model, continuous floating-point parameters representing the resource scaling ratio and discrete classification logic values ​​representing network isolation determination are output, forming a hybrid decision result. The hybrid decision results are automatically translated and encapsulated according to API specifications and routing rules to generate a corresponding set of resource allocation and isolation control instructions. It automatically sends a set of resource allocation and isolation control commands to the host machine and core switching equipment for forced execution, realizing autonomous and unmanned enterprise digital closed-loop scheduling and management.

[0013] A big data-based enterprise digital management system includes: The time-series data construction module collects cross-domain multimodal operation data streams from various physical business nodes and underlying servers of the enterprise to construct multi-dimensional heterogeneous time-series data tensors. The data processing module calculates the local information entropy differences of different data dimensions for multi-dimensional heterogeneous time-series data tensors, and performs adaptive feature filtering and redundancy removal based on information entropy thresholds to obtain high signal-to-noise ratio running feature sequences. The graph construction module extracts the feature vectors of the current active business flow based on the high signal-to-noise ratio running feature sequence, and maps them to the pre-configured physical communication topology of the enterprise's underlying server to construct a dynamic spatiotemporal business association graph. The global state evolution module inputs the dynamic spatiotemporal business association map into the pre-constructed spatiotemporal graph convolutional neural network, performs spatial topological feature aggregation between multi-hop nodes and temporal evolution feature extraction based on long short-term memory network, and forms a multi-scale global state evolution feature representation. The decision feature map module performs cross-attention concatenation of the multi-scale global state evolution feature representation and the high signal-to-noise ratio running feature sequence in the channel dimension to generate a comprehensive state decision feature map. The execution module inputs the comprehensive state decision feature map into the pre-trained deep reinforcement learning decision model to solve the strategy, outputs and automatically issues the corresponding set of dynamic elastic allocation of underlying computing resources and network isolation control instructions for abnormal nodes, and completes the enterprise's digital closed-loop scheduling management.

[0014] The beneficial effects of this invention are: (1) This invention introduces a mechanism for dynamically adjusting the width of the time window and calculating the difference in local information entropy of different data dimensions. This allows for adaptive adjustment of the sampling and filtering strategies of the multimodal data streams on the underlying server based on the surge or stability of enterprise business traffic, improving the signal-to-noise ratio and alignment accuracy of the original data. When faced with massive amounts of high-concurrency, heterogeneous operating data filled with normal background noise, it can accurately extract key turning point signals of system state changes. This invention dynamically updates the discrimination threshold through local information entropy evaluation, ensuring that non-contributing background noise is removed without losing key anomaly precursor features, significantly improving the accuracy of underlying feature extraction and the utilization rate of computational resources.

[0015] (2) This invention aggregates spatial topological features with distance decay factors and extracts temporal evolution features from multi-hop nodes in a dynamic spatiotemporal business association graph, enabling the model to focus more on the potential risks of cross-node cascading failures and the dynamic trend of resource depletion. This invention can perform multi-level neighbor weighting on the topological feature graph that integrates business concurrency intensity between nodes, dynamically capturing hidden danger information accumulated over a long historical span, enabling the network to amplify attention to cascading avalanche risks in complex enterprise microservice architectures, thereby more accurately depicting the macro-evolution view of the entire system's upstream and downstream, significantly improving the accuracy of system bottleneck and anomaly root cause location when processing high-dimensional state data of large-scale clusters, especially in cross-domain long-link calls and deep microservice dependency scenarios.

[0016] (3) This invention combines jump residual connection and cross attention mechanism, and uses a maximum entropy multi-agent flexible action evaluation deep reinforcement learning model to make forward-looking autonomous decisions on the dynamic elastic allocation of enterprise underlying computing resources and the network isolation of abnormal nodes. It uses the cross attention mechanism to integrate the deep macro-evolution trend of the system with the precise waveform of the shallow micro high-frequency fluctuation, and combines a reinforcement learning algorithm with entropy regularization reward to dynamically optimize the global load balancing strategy. This avoids the model from getting stuck in a deadlock state of local suboptimal solution in the complex resource scheduling space, and predicts and automatically issues the optimal microsecond-level computing power hot-plug instructions and access control black hole routing rules. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an enterprise digital management method and system based on big data processing proposed in this invention. Detailed Implementation

[0018] Example 1: Reference Figure 1 A digital management method for enterprises based on big data processing includes: Collect cross-domain multimodal operation data streams from various physical business nodes and underlying servers of the enterprise, perform time-series alignment and spatial dimension mapping based on dynamic time windows on the cross-domain multimodal operation data streams, and construct multi-dimensional heterogeneous time-series data tensors; In this embodiment, cross-domain multimodal operation data streams from various physical business nodes and underlying servers of the enterprise are collected. Temporal alignment and spatial dimension mapping based on dynamic time windows are performed on the cross-domain multimodal operation data streams to construct a multi-dimensional heterogeneous time-series data tensor, including: In the physical business network topology of an enterprise, data probes deployed in various application containers, virtual machines, physical machines, and network switching devices are used to continuously collect cross-domain multimodal operation data streams. These cross-domain multimodal operation data streams include the instruction execution cycle frequency of the central processing unit, the swap-in and swap-out rate of memory pages, the waiting depth of the disk I / O queue, the number of bytes received and sent per second by the network interface card, and the concurrent request response latency of business applications. Due to the physical deviation of the sampling hardware clock crystal oscillators of different data sources and the different degrees of network transmission congestion, the collected cross-domain multimodal operation data streams are disordered and discrete in terms of timestamps. To eliminate temporal misalignment between multimodal data, the system initializes a baseline time axis and dynamically calculates the width of the time window based on the variance of the total data throughput of the entire business system. Specifically, the calculation method is as follows: extract the average absolute value of the first discrete derivative of the network traffic of all physical nodes in the past historical baseline period, divide the average value by the preset system stable operation tolerance constant, take the reciprocal of the division result as the width correction coefficient of the current dynamic time window, and multiply the correction coefficient by the preset base time step to obtain the dynamically adaptive current time window width. This makes the time window narrower during the surge period of business traffic to capture high-frequency transient micro-features, and wider during the trough period of stable business traffic to reduce the computing load of the underlying server. After determining the width of the dynamic time window, with the reference time axis as the alignment target, a non-uniform sampling time alignment operation is performed on all modal data falling within each dynamic time window. For data streams with sampling frequencies higher than the reference time axis, a local mean smoothing downsampling strategy based on inverse distance weight decay is adopted. For multiple high-frequency data points within the time window, the time difference between each high-frequency data point and the reference timestamp is calculated. The time difference is input into the Gaussian kernel function to calculate the distance weight of each high-frequency data point. Then, the values ​​of the high-frequency data points and their corresponding distance weights are multiplied and summed. This summation result is used as the downsampled value after alignment. For data streams with sampling frequencies lower than the reference time axis, or time points where data is missing due to network packet loss, an interpolation compensation strategy based on cubic spline curve fitting is adopted. The two nearest valid historical data points before and after the missing time point are found, and a piecewise cubic polynomial equation system is constructed. By solving the linear equation system that ensures the continuity of the first and second derivatives at the nodes, a smooth transition interpolation is calculated to fill the missing timestamp position, resulting in a multimodal synchronization time sequence that is completely and strictly aligned in the time dimension. The system performs spatial dimension mapping operations, extracts the underlying hardware media access control address or Internet protocol address corresponding to each collected data stream, queries the pre-built enterprise global asset configuration management database, resolves and maps the physical network address to a logical business node number with a unique identifier, uses the logical business node number as the index coordinate of the spatial dimension, uses the aligned base timestamp as the index coordinate of the time dimension, uses different data types such as CPU load and memory load as the index coordinate of the modality dimension, and finally uses the specific numerical value as the content of the feature dimension. All aligned sequences are concatenated and pushed onto a multidimensional array in memory space. A multidimensional heterogeneous time series data tensor containing four dimensions—number of nodes, number of time steps, number of modalities, and number of features—is instantiated in the server's main memory area as the absolutely standardized input base for subsequent state analysis.

[0019] For multi-dimensional heterogeneous time-series data tensors, the local information entropy difference of different data dimensions is calculated, and adaptive feature filtering and redundancy removal processing based on information entropy threshold is performed to obtain high signal-to-noise ratio running feature sequences. In this embodiment, for multi-dimensional heterogeneous time-series data tensors, the local information entropy differences of different data dimensions are calculated, and adaptive feature filtering and redundancy removal processing based on information entropy thresholds are performed to obtain high signal-to-noise ratio running feature sequences, including: After acquiring the multi-dimensional heterogeneous time-series data tensor, in order to eliminate redundant normal background data that is periodic, repetitive, and does not contribute significantly to decision-making, the system starts a feature denoising pipeline based on information theory, without losing key anomaly precursor features. The system slices and slides along the time dimension of the multi-dimensional heterogeneous time series data tensor. For any specific logical business node and specific data mode, it extracts the time series numerical set within an analysis period. The kernel density estimation algorithm is used to reconstruct the empirical probability distribution function of the time series numerical set. The specific steps are as follows: Gaussian smoothing kernel function of preset width is placed on each discrete data point of the time series numerical set. All kernel functions are linearly superimposed and summed in the entire numerical value space. The summation is then divided by the total number of data points to obtain a continuous probability density curve that reflects the fluctuation of the probability of the data of that node and that mode in the current time period. Based on the probability density curve, the system further calculates the local information entropy. It performs an approximate discretization operation using equal-step calculus within a continuous numerical interval, extracting the occurrence probability value for each discrete interval. For each occurrence probability value, its logarithm (base of the natural constant) is calculated. This logarithm is then multiplied by the occurrence probability value itself. The products from all discrete intervals are then summed, and a negative sign is added to the final sum. This yields the local information Shannon entropy value of the modality data within the current slice window. This entropy value precisely quantifies the degree of disorder and unpredictability of the current server state data in a physical sense. When the server is in stable standby or operating regularly, the data exhibits periodicity, and the local information entropy is extremely low. However, when the server encounters a distributed denial-of-service attack scan at the network layer or frequent garbage collection due to internal memory leaks, the data fluctuations exhibit a chaotic state, and the local information entropy spikes dramatically. The system collects the local information entropy values ​​of all nodes and all modalities within the entire cluster in real time, calculates the global information entropy mean and standard deviation, and adds the global information entropy mean to the preset elastic control multiple multiplied by the global information entropy standard deviation to obtain a dynamically updated information entropy discrimination threshold. The system traverses each modal time series in the multi-dimensional heterogeneous time-series data tensor, calculating the absolute value of the local information entropy difference between two adjacent sliding windows. If the absolute value of the information entropy difference is less than the dynamically updated information entropy discrimination threshold, it indicates that the data reflects the operating status of the underlying physical devices. If the data is extremely stable and no state transition occurs, the system will trigger a redundancy removal mechanism. Principal component analysis will be used on this smooth sequence to calculate the eigenvalues ​​and eigenvectors of the sequence's covariance matrix. The system will retain the principal component projection vectors corresponding to the largest eigenvalue whose cumulative contribution rate exceeds a preset lower limit, and discard the remaining higher-order minor components, thereby achieving deep data dimensionality reduction and storage space release. If the absolute value of the information entropy difference is greater than or equal to the information entropy discrimination threshold, it indicates that this data segment contains a key turning point signal of a sudden change in the system state. The system will retain all high-frequency original features within this time period without any compression or deletion. Through the fine-grained information entropy difference assessment and selective principal component projection dimensionality reduction operations for each node, each modality, and each time window, the system cleans and reorganizes the original, massive data tensor filled with normal background noise into a dataset that highly condenses the system's state evolution mutation points and early weak features of anomalies. Finally, it outputs a high signal-to-noise ratio running feature sequence in the cache array, which is free of false information.

[0020] Based on the high signal-to-noise ratio running feature sequence, the feature vector of the current active business flow is extracted, and it is mapped to the pre-configured enterprise underlying server physical communication topology to construct a dynamic spatiotemporal business association graph. In this embodiment, based on the high signal-to-noise ratio running feature sequence, the feature vector of the currently active business flow is extracted, and it is mapped to the pre-configured enterprise underlying server physical communication topology to construct a dynamic spatiotemporal business association graph, including: To obtain high signal-to-noise ratio running feature sequences, and to extract active business flow features that can characterize the real interaction state between the current microservice components, the system initializes a feature extraction module based on multi-head self-attention. The feature sequence of each business node in the current time window is used as the input query vector, and the sequence in the historical time window is used as the input key vector and value vector. The dot product similarity matrix between the query vector and the key vector is calculated, and the similarity score is converted into a probability weight distribution between zero and one through an exponential normalization function. Then, the probability weight distribution and the value vector are weighted and summed to capture the implicit influence of the historical state on the current state. The lengthy time sequence of each node is condensed into a comprehensive feature vector of fixed dimension length that contains the current resource consumption pattern, which serves as the feature vector of the current active business flow. The system reads the pre-configured static server physical communication topology data from the enterprise's underlying infrastructure management platform. This topology data consists of a binary connection adjacency matrix and an edge attribute matrix. The binary connection adjacency matrix records whether there is a physical fiber optic or twisted-pair cable connection between the motherboard network cards on each server rack in the data center through the top-level switch or core router. If a physical link exists, the corresponding element in the matrix is ​​one; otherwise, it is zero. The edge attribute matrix records the maximum designed bandwidth capacity and the reference physical delay time for optical signal transmission for each physical link. To upgrade the physical topology into a dynamic spatiotemporal business association graph that reflects the operational status of services, the system performs a dynamic mapping and fusion operation of node features and edge attributes. Node mapping is performed by precisely assigning the current active business flow feature vector of each extracted logical business node to the physical server node coordinates corresponding to one in the binary connection adjacency matrix. This makes each graph node in the topology not only represent a hardware chassis, but also a rich attribute node carrying real-time high-frequency fluctuation characteristics of the machine's current processor and high-dimensional business operation status information such as memory fragmentation. An improved graph attention algorithm is introduced to dynamically map and construct edge weights. The system considers the presence or absence of physical connections and also examines the business collaboration tightness between nodes. For any two adjacent nodes with physical connections, the mapped feature vectors are concatenated and input into a single-layer fully connected feedforward neural network. The output of the neural network is processed by a linear rectified activation function with leakage and then obtained by exponential normalization to obtain an initial business relevance attention score. The system multiplies the business relevance attention score with the corresponding physical bandwidth capacity in the edge attribute matrix and then divides it by the physical delay time to obtain a comprehensive edge weight value that integrates the intensity of business concurrent requests and the underlying physical transmission limit. A directed graph data structure, configured with the current active business flow feature vector as the internal attribute of the node and the comprehensive edge weight value as the connection strength between the nodes, is persistently instantiated in the graph database. Since this graph data structure not only includes the spatial physical location arrangement of the devices, but also integrates the business flow features and attention weights that dynamically evolve over time, a dynamic spatiotemporal business association graph that can comprehensively reflect the system's operational context is finally constructed.

[0021] The dynamic spatiotemporal business association map is input into a pre-constructed spatiotemporal graph convolutional neural network to perform spatial topological feature aggregation between multi-hop nodes and temporal evolution feature extraction based on long short-term memory network, forming a multi-scale global state evolution feature representation. In this embodiment, the dynamic spatiotemporal service association graph is input into a pre-constructed spatiotemporal graph convolutional neural network to perform spatial topological feature aggregation between multi-hop nodes and temporal evolution feature extraction based on long short-term memory networks, forming a multi-scale global state evolution feature representation, including: The dynamic spatiotemporal business association graph is taken as a time step graph snapshot and continuously input into the system's pre-built improved spatiotemporal graph convolutional neural network framework, which integrates a multi-hop graph attention mechanism and a gated recurrent unit. To perform multi-hop node topology feature aggregation in the spatial dimension, traditional graph convolution can only capture the features of directly adjacent nodes, which cannot cope with cross-node cascading avalanche failures caused by excessively long call chains in enterprise-level microservice architectures. Therefore, a graph attention aggregation strategy that introduces a multi-hop resonance attenuation factor is adopted. For any central target node in the graph, the system not only extracts the features of its first-order neighbor nodes with direct physical connections, but also extracts the features of its second-order and third-order neighbor nodes through a breadth-first search algorithm. When calculating the attention weights between the central target node and its multi-order neighbor nodes, the system not only calculates the inner product score after concatenating the feature vectors, but also multiplies the inner product score by a preset topological distance attenuation factor. The coefficient, the topological distance decay coefficient, is a power function with the natural constant as the base and the shortest hop count between nodes multiplied by a negative penalty parameter as the exponent. This ensures that the influence of distant nodes on the central node decreases exponentially, but can still transmit weak abnormal resonance signals when cascading failures occur. After calculating the decay attention weights of all multi-order neighbors, the feature vectors of all neighbor nodes are linearly weighted and summed using the normalized weights, thereby completing the spatial feature update of the central target node. This ensures that the updated node features include its own resource status and also aggregate the potential risk status of the entire related business chain, completing a deep spatial topological feature aggregation graph convolution operation. To capture the deterioration or self-healing trend of the system state over time, the current full-map node feature matrix after spatial feature aggregation is input into a temporal evolution feature extraction module based on a Long Short-Term Memory (LSTM) network. The LSM network contains three core gated microcomputation units: a forget gate, an input gate, and an output gate. The system concatenates the hidden state vector from the previous time step with the spatial feature matrix of the current completed graph convolution column-wise. The concatenated data is multiplied by the forget gate weight matrix and a bias constant is added. Then, it is mapped to the zero-to-one interval using a sigmoid nonlinear activation function to obtain the forgetting control threshold vector. This forgetting control threshold vector determines how much of the system's accumulated hidden danger information over a long historical period is present. A small percentage should be retained, and a certain percentage should be discarded as expired features. The same concatenated data is multiplied again by the input gate weight matrix and the candidate cell state weight matrix, respectively, and processed by the sigmoid activation function and the hyperbolic tangent activation function. The outputs of the two are multiplied element-wise to extract the most valuable new state mutation information at the current time. The historical information retained by the forget gate and the new information extracted by the input gate are vector-added to complete the cross-clock cycle update of the core cell state. The control vector generated by the output gate is multiplied element-wise with the result of the updated cell state processed by the hyperbolic tangent function to obtain the final hidden state output of the short-term memory network at the current time step. The system uses stacked graph convolutional kernels with receptive fields of different sizes and multiple recurrent network layers to capture microsecond-level transient network jitter features and hour-level slow memory leakage features in parallel. It then splices and aligns the spatiotemporal joint hidden state vectors extracted at different scales according to the channel dimension, and finally outputs a multi-scale global state evolution feature representation that deeply integrates the microservice spatial call dependency risk and resource temporal depletion trend.

[0022] Multi-scale global state evolution feature representation and high signal-to-noise ratio operation feature sequence are cross-attention spliced ​​in the channel dimension to generate a comprehensive state decision feature map. The comprehensive state decision feature map is input into a pre-trained deep reinforcement learning decision model to solve the policy, and outputs and automatically issues the corresponding set of dynamic elastic allocation of underlying computing power resources and network isolation control instructions for abnormal nodes, thus completing the enterprise's digital closed-loop scheduling management.

[0023] In this embodiment, the multi-scale global state evolution feature representation and the high signal-to-noise ratio running feature sequence are cross-attentionally concatenated at the channel dimension to generate a comprehensive state decision feature map. This comprehensive state decision feature map is then input into a pre-trained deep reinforcement learning decision model for policy solving. The model outputs and automatically issues corresponding sets of dynamic elastic allocation of underlying computing resources and network isolation control commands for abnormal nodes, completing the enterprise's digital closed-loop scheduling management, including: To prevent the gradient vanishing of initial local weak anomalous signals after multiple layers of abstraction and transformation in deep neural networks, the system introduces a skip residual connection and cross-attention splicing mechanism. The multi-scale global state evolution feature representation located in the deepest layer of the network is used as the query guidance matrix, and the high signal-to-noise ratio running feature sequence located in the shallow layer of the network, which retains the original accurate waveform, is used as the key value mapping matrix. The cross-attention covariance matrix between the two is calculated to evaluate the matching correlation between the deep macro-evolution trend and the shallow micro high-frequency fluctuation. Based on the matching correlation, the two sets of features are weighted and fused according to the feature dimensions to generate a comprehensive state decision feature map that takes into account both the macro view of the global architecture and the micro view of the underlying hardware. The comprehensive state decision feature map is used as a variable in the environmental state observation space and input into the pre-trained maximum entropy multi-agent flexible action evaluation deep reinforcement learning decision model. The reinforcement learning model includes an executor neural network entity responsible for action generation and a commentator neural network entity responsible for value evaluation. In order to encourage the model to explore better global load balancing strategies in the complex enterprise resource scheduling space and avoid deadlock in local suboptimal solutions, the reward function of the algorithm not only includes positive feedback reward value for throughput improvement, negative penalty deduction value for network latency exceeding the standard, and negative penalty deduction value for resource idle waste, but also adds an entropy regularization reward term for the policy output distribution. That is, the more random and diverse the action probability distribution output by the executor neural network is, the higher the value of the regularization reward term. After receiving the comprehensive state decision feature map, the executor neural network performs nonlinear forward propagation calculations through a multilayer perceptron, outputting two types of control decision values ​​with different dimensions at its output layer: the first type is a continuous floating-point value, used to represent the dynamic increase or decrease ratio of the number of virtual central processing unit cores allocated to each business application container, as well as the dynamic expansion or reduction of the upper limit of container running memory in bytes; the second type is a discrete classification logic value, used to determine whether a node has been infected with a malicious cryptocurrency operation program or has fallen into an irreversible crash loop, requiring immediate triggering of hard blocking and isolation at the physical network layer. The system's internal instruction translation microservice module intercepts the output decision values, automatically translates and encapsulates the first type of continuous floating-point values ​​into elastic scaling control scripts that can perform real-time hot-swap allocation of system resource groups according to the preset cluster orchestration engine application programming interface specifications, and automatically translates and encapsulates the second type of discrete classification logic values ​​into access control list blocking black hole routing rules that are issued to the underlying software-defined network controller according to the device port mapping relationship in the configuration management database. These two parts constitute the final set of control instructions. The system automatically sends a set of control commands to the kernel-mode control processes of the host operating systems of each physical server and the control plane of the core network switching equipment through the Secure Shell Protocol or Remote Procedure Call Protocol. This enables the system to autonomously complete millisecond-level dynamic and elastic allocation of complex digital underlying computing resources and millisecond-level network micro-isolation of abnormally infected nodes without human intervention. This fully realizes the enterprise's digital closed-loop scheduling and management from multimodal data perception to intelligent decision-making and physical hardware execution.

[0024] A big data-based enterprise digital management system includes: The time-series data construction module collects cross-domain multimodal operation data streams from various physical business nodes and underlying servers of the enterprise to construct multi-dimensional heterogeneous time-series data tensors. The data processing module calculates the local information entropy differences of different data dimensions for multi-dimensional heterogeneous time-series data tensors, and performs adaptive feature filtering and redundancy removal based on information entropy thresholds to obtain high signal-to-noise ratio running feature sequences. The graph construction module extracts the feature vectors of the current active business flow based on the high signal-to-noise ratio running feature sequence, and maps them to the pre-configured physical communication topology of the enterprise's underlying server to construct a dynamic spatiotemporal business association graph. The global state evolution module inputs the dynamic spatiotemporal business association map into the pre-constructed spatiotemporal graph convolutional neural network, performs spatial topological feature aggregation between multi-hop nodes and temporal evolution feature extraction based on long short-term memory network, and forms a multi-scale global state evolution feature representation. The decision feature map module performs cross-attention concatenation of the multi-scale global state evolution feature representation and the high signal-to-noise ratio running feature sequence in the channel dimension to generate a comprehensive state decision feature map. The execution module inputs the comprehensive state decision feature map into the pre-trained deep reinforcement learning decision model to solve the strategy, outputs and automatically issues the corresponding set of dynamic elastic allocation of underlying computing resources and network isolation control instructions for abnormal nodes, and completes the enterprise's digital closed-loop scheduling management.

[0025] Example 2: During a period of rapid increase in business concurrency, the system of this invention at the bottom layer of the enterprise's core hybrid cloud data center is collecting real-time cross-domain multimodal data streams from a cluster of 2,500 physical nodes. The implementer observed through a kernel-level probe in the system backend that the overall volume of the cluster exhibited a high degree of imbalance.

[0026] The probe intercepted multimodal raw data originating from a specific application container on the Rack-04 logical rack unit. In the intercepted raw data packets, the virtual CPU's instruction execution cycle frequency reached 3.5 billion times per second, the memory page swapping rate surged from 50 times per second to 1200 times per second within two seconds, and the number of bytes received and sent per second on the physical network interface card eth0 corresponding to this node exhibited violent sawtooth fluctuations. Due to a physical deviation in the host machine's hardware clock crystal oscillator, the timestamp for reported memory load was 0.45 milliseconds, while the timestamp for reported network throughput was 0.52 milliseconds, with sampling frequencies of 400 Hz and 600 Hz respectively, demonstrating significant disorder and dispersion in the data.

[0027] Faced with this timing misalignment, the system immediately triggered a timing alignment and spatial dimension mapping mechanism based on a dynamic time window. The system extracted the average absolute value of the first-order discrete derivative of all network traffic within a historical baseline period, calculating the average value to be 1500 megabytes per second. The system divided this average value by a preset system smooth operation tolerance constant, obtaining a ratio of 3.0. The reciprocal 0.33 was used as a correction coefficient, multiplied by the base time step of 100 milliseconds, thus precisely shrinking and locking the current data alignment dynamic time window for that node to 33.3 milliseconds. Within this extremely narrow time window, for network throughput data with a sampling frequency as high as 600 Hz, the system performed inverse distance weighted decay local mean smoothing downsampling, smoothing multiple high-frequency oscillation points into a single aligned value representing 650 megabytes per second. For memory paging rates resulting from data loss due to congestion, the system searched for historical valid data 0.21 milliseconds and 0.85 milliseconds before and after the missing time point, solved a system of cubic polynomial equations, calculated smooth interpolation at 890 times per second, and filled in the missing positions. The system queries the global configuration database and maps the IP address 10.52.8.105 to a unique logical business node number, Node-Cart-8105. In the server's main memory area, the system successfully pushes and instantiates a multi-dimensional heterogeneous time-series data tensor with dimensions of 2500 (number of nodes) × 100 (number of time steps) × 8 (number of modalities) × 64 (number of features).

[0028] The system initiates local information entropy difference calculation and redundancy removal processing for multi-dimensional heterogeneous time-series data tensors. The system slides slices along the time dimension, and for the memory paging rate modal data of the Node-Cart-8105 node, it estimates the kernel density by placing a Gaussian smoothing kernel function, reconstructing the probability density curve. By performing an approximate discretization operation of the probability density curve using equal-step calculus, the system calculates that the local information Shannon entropy value of this node within the current time window is as high as 5.87. In contrast, the local information entropy value of the Node-Static-8102 node, which carries the static image retrieval service on the same rack, is only 1.05. The system collects the global information entropy mean and standard deviation in real time, multiplies them by a preset elasticity factor of 1.5, and calculates the current dynamic discrimination threshold to be 3.425. Since the entropy value of Node-Static-8102 is far below the threshold, the system immediately triggers principal component analysis for dimensionality reduction, compressing its 64-dimensional feature sequence to retain only the projection vectors of the top 3 principal components with a cumulative contribution rate of over 90%, discarding the extremely large amount of normal background data; while the entropy value of Node-Cart-8105 far exceeds the threshold of 3.425, the system determines that it has undergone a state transition, completely preserving all of its 64-dimensional high-frequency original fluctuation features, thereby outputting a cleaned high signal-to-noise ratio running feature sequence in the cache array.

[0029] Based on the above sequence, the system begins to construct a dynamic spatiotemporal service association graph. The system inputs the feature sequence of Node-Cart-8105 into a multi-head self-attention module, calculates the dot product similarity between the query vector and the historical key vector, and extracts a 128-dimensional comprehensive feature vector of the currently active service flow. The system reads the underlying communication topology data and finds that Node-Cart-8105 has physical connections with upstream Node-Order-9201 and downstream Node-DB-5510 through a top-level switch. The system concatenates the feature vectors of adjacent nodes and inputs them into a single-layer fully connected feedforward neural network. After processing with a linear rectified activation function with leakage, the system calculates a service relevance attention score of 0.94 between Node-Cart-8105 and Node-DB-5510. The system multiplies 0.94 by the physical bandwidth and divides by the latency, obtaining a comprehensive edge weight value as high as 470,000. Thus, a dynamic spatiotemporal service association graph containing 2,500 attribute-rich nodes and tens of thousands of dynamically weighted directed edges is successfully instantiated in the graph database.

[0030] The system inputs this graph snapshot into a spatiotemporal graph convolutional neural network framework. Using Node-Cart-8105 as the central target node, the system performs multi-hop spatial topological feature aggregation. The system not only extracts features from Node-DB-5510, which is directly connected to it, but also extracts features from its neighbors (logic IP 10.52.12.11) outside the third-order distance using breadth-first search. The system calculates the third-order topological distance decay coefficient, obtaining a decay weight of 0.548. Through linear weighted summation, it is found that the high-frequency memory oscillation features of Node-Cart-8105 have triggered a weak anomalous resonance integral along the call chain on Node-Pay-1211 outside the third hop. The data enters the Long Short-Term Memory (LSTM) network layer. The forget gate outputs a control vector close to zero, completely discarding the node's stable historical state from the previous five minutes, while the input gate outputs an extremely high activation value, extracting the new mutation information of the current memory leak spreading to the payment node outside the third hop. After concatenation along the channel dimensions, the system ultimately outputs a 512-dimensional multi-scale global state evolution feature representation.

[0031] In the split second before the disaster caused physical system failure, the system executed cross-attention splicing and reinforcement learning decision-making. The system combined 512-dimensional macroscopic evolutionary features with 64-dimensional microscopic high signal-to-noise ratio sequences (by performing cross-attention covariance calculation and weighted splicing) to generate a 768-dimensional comprehensive state decision feature map. This feature map was then input into a maximum entropy multi-agent flexible action evaluation deep reinforcement learning model. In this specific scenario, cutting off order nodes according to conventional logic would lead to widespread errors. To avoid getting trapped in this local suboptimal solution, the system, driven by entropy-regularized rewards, explored an extremely precise multi-dimensional strategy.

[0032] The output layer of the executor neural network ultimately outputs a tensor containing a set of floating-point values ​​and categorical logical values ​​with precise instructions: continuous output is [+2.5, +4096], and discrete output is [1, 0, 1]. The instruction translation microservice within the system immediately intercepts this tensor and automatically translates it. For Node-Pay-1211, which is outside the three-hop limit and is under resonance pressure, the system automatically generates an expansion script via API, instantly adding 2.5 cores to its virtual CPU and hard-expanding its maximum running memory by 4096MB (4GB). Meanwhile, for the culprit Node-Cart-8105, given its discrete output, it is determined to be trapped in an irreversible resource exhaustion loop, and the translator directly issues an access control list black hole routing rule. After receiving the set of control commands from the SSH protocol, the underlying software-defined network controller rewrote the switch's flow table within 0.03 seconds, forcibly blocking the forwarding of all packets destined for IP 10.52.8.105 and ports 8080 and 3306, and automatically load-balanced the new traffic to backup healthy container nodes. The entire process took less than 1.2 seconds. Without any intervention, the system completely prevented a global avalanche caused by a hidden memory leak in a deep microservice and successfully completed the enterprise's digital closed-loop scheduling management.

[0033] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A digital management method for enterprises based on big data processing, characterized in that, include: Collect cross-domain multimodal operation data streams from various physical business nodes and underlying servers of the enterprise, perform time-series alignment and spatial dimension mapping based on dynamic time windows on the cross-domain multimodal operation data streams, and construct multi-dimensional heterogeneous time-series data tensors; For multi-dimensional heterogeneous time-series data tensors, the local information entropy difference of different data dimensions is calculated, and adaptive feature filtering and redundancy removal processing based on information entropy threshold is performed to obtain high signal-to-noise ratio running feature sequences. Based on the high signal-to-noise ratio running feature sequence, the feature vector of the current active business flow is extracted, and it is mapped to the pre-configured enterprise underlying server physical communication topology to construct a dynamic spatiotemporal business association graph. The dynamic spatiotemporal business association map is input into a pre-constructed spatiotemporal graph convolutional neural network to perform spatial topological feature aggregation between multi-hop nodes and temporal evolution feature extraction based on long short-term memory network, forming a multi-scale global state evolution feature representation. The multi-scale global state evolution feature representation and the high signal-to-noise ratio running feature sequence are cross-attentionally concatenated in the channel dimension to generate a comprehensive state decision feature map. The comprehensive state decision feature map is input into a pre-trained deep reinforcement learning decision model to solve the policy, and the corresponding set of dynamic elastic allocation of underlying computing resources and network isolation control instructions for abnormal nodes are output and automatically issued to complete the enterprise's digital closed-loop scheduling management.

2. The enterprise digital management method based on big data processing according to claim 1, characterized in that, The process involves collecting cross-domain multimodal operational data streams from various physical business nodes and underlying servers of the enterprise, performing time-series alignment and spatial dimension mapping based on dynamic time windows on these data streams, and constructing a multi-dimensional heterogeneous time-series data tensor, including: By collecting operational data from each physical business node and underlying server using data probes, a raw multimodal data stream with disordered and discrete timestamps is obtained. Extract the first derivative of historical network traffic to calculate the fluctuation variance and correction coefficient, and dynamically generate an adaptive current time window width; Based on the current time window width, local mean smoothing downsampling or cubic spline interpolation compensation is performed on the original multimodal data stream to obtain a strictly aligned multimodal synchronization time series; Extract the physical network address of the multimodal synchronization time sequence, query the asset database to map it to the logical business node number, so as to establish a four-dimensional coordinate system of space, time, mode and feature; Based on the four-dimensional coordinate system, multi-dimensional splicing and stacking operations are performed on the mapped multimodal synchronous time series to instantiate and construct a multi-dimensional heterogeneous time series data tensor.

3. The enterprise digital management method based on big data processing according to claim 2, characterized in that, For multi-dimensional heterogeneous time-series data tensors, the local information entropy differences of different data dimensions are calculated, and adaptive feature filtering and redundancy removal processing based on information entropy thresholds are performed to obtain high signal-to-noise ratio running feature sequences, including: The time-dimensional slice sliding is performed on the multi-dimensional heterogeneous time-series data tensor, and the empirical probability distribution of each segment is reconstructed using the Gaussian kernel function to obtain a continuous probability density curve. By performing discretization calculus and logarithmic operations on the probability density curve, the Shannon entropy value, a local information that precisely quantifies the degree of data disorder, is calculated. By combining the global information entropy mean and standard deviation, the local information Shannon entropy value is subjected to elastic multiple calculation to dynamically generate a real-time updated information entropy discrimination threshold; Information entropy is used to identify thresholds to compare and judge multi-dimensional heterogeneous time series tensors. Principal component analysis is performed to reduce the dimensionality of low-entropy stable data and remove it, while high-entropy abrupt data is completely preserved to form a highly concentrated and pure feature set. The clean feature set is cleaned and reorganized in the cache array, and the output is a high signal-to-noise ratio running feature sequence with no false positives and true negatives.

4. The enterprise digital management method based on big data processing according to claim 3, characterized in that, The process involves extracting feature vectors from currently active service flows based on high signal-to-noise ratio running feature sequences, mapping these vectors to pre-configured enterprise underlying server physical communication topology, and constructing a dynamic spatiotemporal service association graph, including: The high signal-to-noise ratio running feature sequence is input into the multi-head self-attention module, the probability weight distribution of historical state and current state is calculated and weighted summed, and the feature vector of the current active business flow is extracted. Read the underlying static physical communication topology data to obtain the adjacency matrix and edge attribute matrix that record physical links and bandwidth delay parameters; The feature vector of the current active business flow is accurately mapped to the physical node coordinates corresponding to the adjacency matrix, generating a rich attribute node carrying a high-dimensional running state; Features of adjacent rich attribute nodes are extracted and concatenated, input into a feedforward neural network, and combined with the edge attribute matrix to calculate service relevance, generating a comprehensive edge weight value that integrates transmission limit and service concurrency. By using rich attribute nodes as graph attributes and comprehensive edge weights as connection strengths, and persistently instantiating them in a graph database, a dynamic spatiotemporal business association graph is constructed.

5. The enterprise digital management method based on big data processing according to claim 4, characterized in that, The process involves inputting a dynamic spatiotemporal service association graph into a pre-constructed spatiotemporal graph convolutional neural network to perform spatial topological feature aggregation between multi-hop nodes and temporal evolution feature extraction based on a long short-term memory network, forming a multi-scale global state evolution feature representation, including: The dynamic spatiotemporal business association map is used as a time step snapshot and input into a spatiotemporal graph convolutional neural network framework that incorporates a multi-hop resonance attenuation factor; In the spatial dimension, the inner product score of the central node and its multi-order neighbor nodes is calculated and the topological decay coefficient is applied. Then, graph attention weighted summation is performed to complete the topological aggregation and obtain the spatial feature matrix. In the time dimension, the spatial feature matrix is ​​concatenated with the historical state and then input into the long short-term memory network. After cross-clock cycle calculations through the forget gate, input gate and output gate, the newly added temporal hidden state is extracted. The newly extracted temporal hidden states from different receptive fields are spliced ​​and aligned according to the channel dimension, and deep fusion is performed to obtain a multi-scale global state evolution feature representation.

6. The enterprise digital management method based on big data processing according to claim 5, characterized in that, The process of inputting the comprehensive state decision feature map into a pre-trained deep reinforcement learning decision model for policy solving, outputting and automatically issuing corresponding control commands, and completing enterprise digital closed-loop scheduling management includes: The deep multi-scale global state evolution feature representation and the shallow high signal-to-noise ratio running feature sequence are cross-attention covariance evaluated and weighted and spliced ​​to generate a comprehensive state decision feature map. The comprehensive state decision feature map is used as an environmental observation variable and input into the maximum entropy multi-agent deep reinforcement learning decision model with a reward term regularized by distribution entropy value. Through forward propagation calculation of the decision model, continuous floating-point parameters representing the resource scaling ratio and discrete classification logic values ​​representing network isolation determination are output, forming a hybrid decision result. The hybrid decision results are automatically translated and encapsulated according to API specifications and routing rules to generate a corresponding set of resource allocation and isolation control instructions. It automatically sends a set of resource allocation and isolation control commands to the host machine and core switching equipment for forced execution, realizing autonomous and unmanned enterprise digital closed-loop scheduling and management.

7. A big data-based enterprise digital management system, used to execute the big data-based enterprise digital management method according to any one of claims 1-6, characterized in that, include: The time-series data construction module collects cross-domain multimodal operation data streams from various physical business nodes and underlying servers of the enterprise to construct multi-dimensional heterogeneous time-series data tensors. The data processing module calculates the local information entropy differences of different data dimensions for multi-dimensional heterogeneous time-series data tensors, and performs adaptive feature filtering and redundancy removal based on information entropy thresholds to obtain high signal-to-noise ratio running feature sequences. The graph construction module extracts the feature vectors of the current active business flow based on the high signal-to-noise ratio running feature sequence, and maps them to the pre-configured physical communication topology of the enterprise's underlying server to construct a dynamic spatiotemporal business association graph. The global state evolution module inputs the dynamic spatiotemporal business association map into the pre-constructed spatiotemporal graph convolutional neural network, performs spatial topological feature aggregation between multi-hop nodes and temporal evolution feature extraction based on long short-term memory network, and forms a multi-scale global state evolution feature representation. The decision feature map module performs cross-attention concatenation of the multi-scale global state evolution feature representation and the high signal-to-noise ratio running feature sequence in the channel dimension to generate a comprehensive state decision feature map. The execution module inputs the comprehensive state decision feature map into the pre-trained deep reinforcement learning decision model to solve the strategy, outputs and automatically issues the corresponding set of underlying computing power resources dynamic elastic allocation and abnormal node network isolation control instructions, and completes the enterprise's digital closed-loop scheduling management.

Images