Artificial intelligence-based multi-cloud billing analysis method, system, device, and medium

Abstract

The application is suitable for the field of computer technology, and provides a multi-cloud bill analysis method based on artificial intelligence, which comprises the following steps: collecting original bill data of different cloud platforms and associated resource configuration data and resource consumption data, and generating original data; analyzing the text description of the original data, classifying it into a predefined business semantic category, and generating semantic bill data; constructing a bill knowledge graph based on the semantic bill data, the resource configuration data and the resource consumption data; analyzing the root cause of the cost anomaly based on the bill knowledge graph, and generating cost optimization suggestions for comprehensive cost and performance. The application classifies the original data into a multi-level business semantic label system corresponding to the enterprise, realizes the conversion from the technical resource dimension to the business semantic dimension, realizes the association between business and cost, and finally generates cost optimization suggestions with globality, forward-looking and transparency, so as to realize fine and automatic multi-cloud bill analysis for the enterprise.

Description

Technical Field

[0001] This application belongs to the field of computer technology, and in particular relates to a method, system, device and medium for multi-cloud billing analysis based on artificial intelligence. Background Technology

[0002] With the popularization of cloud computing, more and more enterprises are adopting a multi-cloud strategy, deploying their businesses on different cloud platforms. However, the billing systems, data formats, and preferential methods of various cloud service providers, such as AWS, Azure, Google Cloud, and Alibaba Cloud, are very different, making it difficult for enterprises to analyze massive amounts of heterogeneous cost data.

[0003] Existing multi-cloud billing management solutions mainly focus on solving the problems of data "unification" and "visibility". Specifically, they automatically collect bills by integrating the application programming interfaces (APIs) of various cloud platforms, and standardize the data format based on predefined mapping rules. Finally, they realize the aggregation of cross-cloud costs and the display of basic reports in a unified dashboard.

[0004] The aforementioned multi-cloud billing management solutions have solved the problems of data dispersion and inconsistent formats to some extent, and improved cost "visibility." However, they lack multi-dimensional and fine-grained cost analysis capabilities, cannot be linked to business operations, and cannot automatically identify abnormal costs or predict future expenditures, making it difficult to achieve cost optimization. Manual cost planning is still required. Summary of the Invention

[0005] This application provides a method, system, device, and medium for multi-cloud billing analysis based on artificial intelligence, which can solve one of the problems of the prior art mentioned above.

[0006] In a first aspect, embodiments of this application provide a multi-cloud billing analysis method based on artificial intelligence, including: Collect raw billing data and associated resource configuration and resource usage data from different cloud platforms to generate raw data; The text description of the raw data is parsed and classified into predefined business semantic categories to generate semantic billing data. Based on the semantic billing data, the resource configuration data, and the resource usage data, a billing knowledge graph is constructed. Based on the aforementioned billing knowledge graph, the root causes of cost anomalies are analyzed, and cost optimization suggestions for overall cost and performance are generated.

[0007] Furthermore, the step of parsing the textual description of the original data, classifying it into predefined business semantic categories, and generating semantic billing data includes: Based on a predefined unified data model, the original data is standardized in format and cleaned to generate structured billing data; The text description field in the structured billing data is input into a pre-trained deep learning classifier, which outputs a classification result corresponding to a predefined multi-level business semantic label system to generate semantic billing data.

[0008] Furthermore, the construction of a billing knowledge graph based on the semantic billing data, the resource configuration data, and the resource usage data includes: Define a graph pattern, which includes at least business entity nodes, resource entity nodes, cost entity nodes, and edge types representing the relationships between nodes. The edge types include at least "belong to", "run on", "consume", "associate", and "call / depend on". Based on the semantic billing data, the resource configuration data, and the resource usage data, corresponding graph nodes are automatically created or updated through entity extraction. By analyzing resource tags and configurations through a rules engine and combining them with network traffic log analysis, relationship edges between the nodes of the graph are established to form a billing knowledge graph. Monitor cloud platform resource events and dynamically update the nodes and edges in the billing knowledge graph according to the event type.

[0009] Furthermore, the analysis of the root causes of cost anomalies based on the billing knowledge graph includes: The status of the billing knowledge graph in each billing cycle is fused with the cost data of the corresponding cycle to construct a time series graph sequence. Each time step of the time series graph sequence includes the cost time series features of the corresponding cycle. The time series sequence is trained based on the first prediction model, which is used to learn the normal evolution pattern of the cost time series features over time on the billing knowledge graph. For a new billing cycle, the corresponding time series graph sequence is input into the trained first prediction model to calculate the anomaly score of the bill knowledge graph on the cost entity node, and the anomaly status is determined based on the anomaly score. For the abnormal cost entity node, extract the graph attention weights from the first prediction model, and perform backtracking analysis along the relation edges in the billing knowledge graph to locate the root cause of the abnormality.

[0010] Furthermore, the cost optimization recommendations for generating overall cost and performance include: Based on the billing knowledge graph and the second prediction model, the cost of each node in the billing knowledge graph is predicted in future periods, and a future cost trend map is output. Define an optimization action library that includes configuration and change operations for various cloud platforms, and build a simulation engine based on the topological relationship of the billing knowledge graph to simulate the correlation between the optimization action library and cost and performance. Based on the root cause of cost anomalies, a set of candidate optimization strategies is generated. The global impact of each strategy in the set is simulated by the inference engine. A comprehensive evaluation is performed using a multi-objective evaluation function to generate a personalized optimization strategy report.

[0011] Furthermore, an optimization action library containing various cloud platform configuration and change operations is defined, and based on the topological relationships of the billing knowledge graph, a simulation engine is built to demonstrate the correlation between the optimization action library and cost and performance, including: Define an optimization action library, wherein the atomic optimization actions in the optimization action library include at least adjusting resource configuration, changing billing mode, and starting / stopping services; Establish an impact calculation model to quantify the direct cost and performance impact of atomic optimization actions on target resource nodes; An impact propagation model is established, based on the edge types in the billing knowledge graph, to propagate the direct impact along the graph topology to the associated business entity nodes or resource entity nodes, and to calculate the indirect impact.

[0012] Furthermore, based on the root cause results of cost anomalies, a set of candidate optimization strategies is generated, and the global impact of each strategy in the candidate optimization strategy set is simulated by the inference engine, including: Based on the root cause of the cost anomaly, a first class of candidate strategies is generated for the root cause. The resource usage data is analyzed to identify resource entities whose resource utilization rate is lower than a first threshold or higher than a second threshold, and a second type of candidate strategy for resource configuration adjustment actions is generated. The first type of candidate strategy and the second type of candidate strategy are filtered and sorted to generate a set of candidate optimization strategies; Each strategy in the candidate optimization strategy set is parsed into an ordered sequence of atomic optimization actions; Using the future cost landscape map as the initial state, the atomic optimization action sequence is executed sequentially; For each atomic optimization action, calculate its direct impact on the target node, and calculate its indirect impact on associated nodes based on the relation edges in the billing knowledge graph using a preset impact propagation model. By aggregating all direct and indirect effects, global simulation results are generated for the corresponding candidate optimization strategies.

[0013] Secondly, embodiments of this application provide an artificial intelligence-based multi-cloud billing analysis system, including: The first processing module is used to collect raw billing data from different cloud platforms and their associated resource configuration data and resource usage data to generate raw data. The second processing module is used to parse the text description of the raw data, classify it into predefined business semantic categories, and generate semantic billing data. The third processing module is used to construct a billing knowledge graph based on the semantic billing data, the resource configuration data, and the resource usage data. The fourth processing module is used to analyze the root causes of cost anomalies based on the billing knowledge graph and generate cost optimization suggestions that combine cost and performance.

[0014] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described multi-cloud billing analysis method based on artificial intelligence.

[0015] Fourthly, embodiments of this application provide a computer-readable storage medium, including a computer program stored in the computer-readable storage medium, which, when executed by a processor, implements the aforementioned multi-cloud billing analysis method based on artificial intelligence.

[0016] The beneficial effects of the embodiments in this application compared with the prior art are: This application discloses an AI-based multi-cloud billing analysis method. It utilizes a pre-trained deep learning classifier to perform multi-task semantic parsing on standardized structured billing data, classifying the raw data into a multi-level business semantic tagging system corresponding to the enterprise. This achieves a transformation from a technical resource dimension to a business semantic dimension, establishing a link between business and costs. Simultaneously, based on the semanticization results, a billing knowledge graph is constructed and updated in real-time. This knowledge graph integrates scattered cost data, resource configuration, and business architecture, providing a global view of costs and formally defining the cost generation and propagation paths. Furthermore, the billing knowledge graph is fused with time-series cost data, combined with a first prediction model for anomaly detection, and backtracked analysis along the relational edges of the knowledge graph to pinpoint the root causes of cost anomalies. Then, based on these root causes, a global simulation optimization decision is made, ensuring that the final cost optimization recommendations are global, forward-looking, and transparent, enabling enterprises to achieve refined and automated multi-cloud billing analysis. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a multi-cloud billing analysis method based on artificial intelligence, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a multi-cloud billing analysis system based on artificial intelligence provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0019] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0020] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0021] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0023] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0025] Please see Figure 1 As shown, this invention is a multi-cloud billing analysis method based on artificial intelligence, comprising the following steps: S100: Collect raw billing data and associated resource configuration data and resource usage data from different cloud platforms to generate raw data; In this embodiment, raw data is automatically and periodically retrieved from multiple cloud platforms associated with the enterprise to ensure data integrity and timeliness. Specifically, for cloud service providers that provide standardized billing APIs, such as AWS and Azure, dedicated API adapters are configured. Each API adapter integrates the authentication mechanisms of each cloud platform, such as IAM roles, service principal certificates, and access keys, and encapsulates their unique query syntax and pagination logic. A unified scheduled task engine is used to automatically retrieve detailed billing data for a specified time range at preset periods, usually in JSON or CSV format.

[0026] Furthermore, all collected raw billing data, regardless of source or format, is temporarily stored in a unified temporary storage area, and metadata tags such as source and collection time are marked for the raw billing data of each cloud platform, awaiting subsequent processing.

[0027] In addition, the API adapter also synchronously calls the cloud platform's resource inventory API and monitoring API to collect resource configuration data and resource usage data related to billing. Specifically, through the cloud platform's resource inventory API, such as the AWS ResourceGroups Tagging API and Azure Resource Graph, it periodically obtains a list of resources and their key attributes from all cloud platforms, such as resource ID, service name, service code, cloud vendor, resource type, operation type, region, instance specification, usage type, creation time, resource tags, etc., to form resource configuration data. For example, in one embodiment, for AWS cloud platform service providers, resource tags and metadata are obtained through the AWS Resource Groups Tagging API, combined with the EC2 API to collect information such as instance specification and region. Specifically, {Resource ID: i-1234567890abcdef0, Resource type: EC2 instance, Region: us-west-2, Instance specification: m5.large, Resource tag: {"Business System":"CRM","Environment":"Production"}}. At the same time, it connects to cloud monitoring services, such as Amazon CloudWatch and Azure Monitor, to collect resource usage and performance metrics associated with billing items, such as EC2 instance runtime, S3 storage bytes, and network outflow traffic in AWS cloud services, as well as performance metrics such as CPU utilization and disk IOPS, to provide context for cost attribution and anomaly analysis.

[0028] S200: Parse the text description of the original data, classify it into predefined business semantic categories, and generate semantic billing data; In this embodiment, the original billing data, which is scattered, has different formats, and lacks business meaning, is parsed and transformed into structured data with a unified format and clear business semantics, providing high-quality, semantically rich input data for subsequent construction of billing knowledge graphs and analysis.

[0029] Specifically, by using a pre-trained deep learning classifier, the standardized structured billing data is subjected to multi-task semantic parsing, thereby classifying the original data into a multi-level business semantic label system corresponding to the enterprise, realizing the transformation from the technical resource dimension to the business semantic dimension, and realizing the correlation between business and cost.

[0030] In some embodiments, step S200 above includes: Based on a predefined unified data model, the original data is standardized in format and cleaned to generate structured billing data; The text description field in the structured billing data is input into a pre-trained deep learning classifier, which outputs a classification result corresponding to a predefined multi-level business semantic label system to generate semantic billing data.

[0031] In this embodiment, a predefined unified data model is used to transform raw data from different sources and with varying formats into a unified and clean structured form. The unified data model defines the core field set required for the multi-cloud billing analysis in this application, such as billing period, cloud vendor, service code, resource identifier, resource tag, usage value, usage unit, cost amount, and original description text.

[0032] In addition, a standardization engine is configured, which maintains an independent mapping rule configuration file for each supported cloud platform. By loading the mapping rule configuration file of the corresponding cloud platform, the mapping rules between the corresponding original billing data and the unified data model can be obtained, thereby performing operations such as field extraction and value transformation on the original billing data and loading it into the corresponding fields of the unified data model.

[0033] Furthermore, after mapping the fields in the original billing data to a unified data model, the data also needs to be cleaned, such as removing billing records that are irrelevant to business operations, such as test accounts and internal transfers; attempting to imput missing key fields (such as resource IDs) through related data or rules; and uniformly converting the cost currency to the base currency.

[0034] It is worth noting that for resource configuration data and resource usage data in the cloud platform, the mapping rules configuration file defined by the standardized engine is synchronously mapped to the application quantity value, usage unit, resource identifier, resource tag, etc., and the standardized processing of resource configuration data and resource usage data is realized in a synchronous manner to generate complete structured billing data.

[0035] Specifically, taking resource configuration data as an example, AWS's lineItem / ProductCode and Azure's MeterCategory are mapped to the service_name field in the unified data model through the mapping rule files of the corresponding cloud platforms; AWS's product / instanceType is mapped to instance_type. At the same time, the integrity and consistency of the values ​​of each field are ensured by cleaning the data of each mapped field.

[0036] Furthermore, each standardized bill entry mapped through the aforementioned unified data model is assigned business semantics. Specifically, based on the enterprise's organizational structure and business logic, a multi-level, scalable business semantic tag classification system is established. In one embodiment, the first level is the business line, such as e-commerce and fintech; the second level is the application service, such as the user center and payment gateway; the third level is the environment type, such as production environment, pre-release environment, and test environment; and the fourth level is the cost driver, such as computing processing, data storage, and content distribution.

[0037] In this embodiment, a pre-trained language model based on the Transformer architecture, such as BERT, RoBERTa, or its variants, is used as a feature extractor. The input to this pre-trained language model is a concatenated key resource configuration field, which typically includes service name, operation type, usage type, instance specification, and key business keywords extracted from resource tags, such as "[CLS] Service Name: Amazon Elastic Compute Cloud [SEP] Operation Type: RunInstances [SEP] Instance Specification: m5.2xlarge [SEP] Region: us-west-2 [SEP] Usage Type: BoxUsage [SEP]". Here, [CLS] and [SEP] are standard separators for the Transformer model. By concatenating the above discrete resource data, the model can be transformed into continuous text that it can understand, while preserving the semantic boundaries of the fields.

[0038] Specifically, the service name is a direct mapping of the cloud platform's business function. For example, Amazon Elastic Compute Cloud directly indicates that it is a computing service, and Amazon Relational Database Service indicates that it is a database service. This completes the most basic "service type" judgment and is the core basis for cost driver classification. Operation type reveals the resource's usage method or lifecycle state. For example, RunInstances indicates that the instance is running normally and incurring computing costs, Storage indicates pure storage usage, and DataTransfer-Out-Bytes indicates network outflow. The operation type can be used to distinguish the nature of the cost. For example, for the same EC2 service, RunInstances corresponds to computing costs, while EBS:VolumeUsage corresponds to additional storage costs. Instance specifications indicate the resource's performance level and business importance. For example, the cost and use of m5.2xlarge and t3.micro are vastly different. Large-scale instances are more likely to be used for high-load applications in core production businesses, while small-scale instances may be used for testing or background tasks. This can be used to determine business priority or application services. The region is associated with the geographical location of the business deployment. For example, the selection of regions such as us-east-1 (North America) and cn-north-1 (Beijing, China) may correspond to business lines serving different regional markets, thus assisting in the classification of business lines. The usage type further refines the billing scenario. For example, BoxUsage represents on-demand instances, HeavyUsage represents reserved instances, and SpotUsage represents spot instances. Different usage types not only affect costs but may also be related to the stability requirements of the business scenario. For example, core production businesses may use Spot instances less often, which is helpful in determining the environment type.

[0039] Therefore, the feature extractor can analyze and obtain the semantic tags of the corresponding bill in the multi-level business semantic tag classification system through the above key resource configuration fields. Understandably, the resource tags also need to be concatenated as additional fields into the above input sequence to provide the model with stronger business context clues.

[0040] Furthermore, multiple parallel and independent classification layers are connected to the output of the feature extractor. Each classification layer corresponds to a specific level in the multi-level business semantic label system. Each classification layer is responsible for predicting the business semantic label to which the bill item belongs at that level, which is beneficial for learning business classifications at different granularities.

[0041] Specifically, the structured billing data after format standardization is processed by the aforementioned deep learning classifier, which can infer in real time the semantic tags of each level of business corresponding to the bill, and these tags are then added as new fields to the structured billing data.

[0042] S300. Construct a billing knowledge graph based on the semantic billing data, the resource configuration data, and the resource usage data; In this embodiment, based on the semanticization results of step S200 above, a billing knowledge graph is constructed and updated in real time. This billing knowledge graph integrates scattered cost data, resource configuration and business architecture, provides a global view of costs, and formally defines the cost generation and propagation path.

[0043] In some embodiments, step S300 above includes: Define a graph pattern, which includes at least business entity nodes, resource entity nodes, cost entity nodes, and edge types representing the relationships between nodes. The edge types include at least "belong to", "run on", "consume", "associate", and "call / depend on". Based on the semantic billing data, the resource configuration data, and the resource usage data, corresponding graph nodes are automatically created or updated through entity extraction. By analyzing resource tags and configurations through a rules engine and combining them with network traffic log analysis, relationship edges between the nodes of the graph are established to form a billing knowledge graph. Monitor cloud platform resource events and dynamically update the nodes and edges in the billing knowledge graph according to the event type.

[0044] In this embodiment, the billing knowledge graph includes business entity nodes, resource entity nodes, and cost entity nodes. Business entity nodes correspond to different levels of business semantic tags in the semantic billing data of step S200, and their attributes include business ID, business name, business line, and department. Resource entity nodes correspond to the resource configuration data of step S100, and their attributes include resource ID, cloud vendor, resource type, region, instance specification, and creation time. Cost entity nodes represent a specific billing event or aggregated cost unit, and their attributes include cost ID, billing period, cost amount, usage value, and original service name.

[0045] Furthermore, relationship edges representing the relationships between nodes are established. These edges include types such as Belong to, Run on, Consume, Associate, and Invoke / Dependency. Specifically, Belong to connect hierarchical business entity nodes, such as "Order Service" belonging to the "E-commerce Business Unit" and "Risk Control Engine Service" belonging to the "Payment Core Business Unit"; Run on represents the deployment relationship from a business entity node to a resource entity node, such as "User Authentication Service" running on "VM instance i-abc123" and "User Session Service" running on "ECS instance i-xyz789"; Consume represents the generation relationship from a resource entity node to a cost entity node. For example, "VM instance i-abc123" consumes "October 2023 computing cost $50", and "NAT gateway nat-abc123" consumes "January 2024 network egress cost $1200". This relationship is automatically established when the cost entity node is created through its resource_id attribute. Association represents the affiliation relationship between the cost entity node and the business entity node, such as "October 2023 computing cost $50" being associated with "user authentication service". Call / dependency represents the network or logical dependency relationship between resource entity nodes or between business entity nodes, such as "application server A" calling "database B".

[0046] Specifically, the rules engine parses resource configuration data and scans resource tags. When two resource entities share the same resource tag, it's determined that they jointly serve the same business. A "running on" relationship edge is then established between the corresponding resource entity node and the business entity node. Furthermore, network traffic logs are analyzed to identify resource pairs with communication relationships. Specifically, based on resource usage data, the actual call relationships between resource entity nodes are discovered by analyzing their source IP, destination IP, port, and traffic patterns. For example, if a group of EC2 instances frequently accesses a specific RDS database, a call edge is automatically established between these nodes.

[0047] In this embodiment, the output data streams of each cloud platform are monitored in real time, namely resource events of original billing data, resource configuration data, and resource usage data. When new semantic billing data is generated, a new cost entity node is automatically created and its attributes are filled in accordingly. At the same time, according to the resource ID in the bill, the corresponding resource entity node is found or created in the graph and a consumption edge is established. When the resource configuration data is updated, the attributes of the resource entity node in the billing knowledge graph are updated synchronously, or a new resource node is created.

[0048] S400. Based on the billing knowledge graph, analyze the root causes of cost anomalies and generate cost optimization suggestions for overall cost and performance.

[0049] In this embodiment, the billing knowledge graph is fused with time-series cost data, and anomaly detection is performed using a first prediction model. The analysis is then backtracked along the relational edges of the billing knowledge graph to pinpoint the root cause of the cost anomaly. Finally, optimization decisions are made based on a global simulation of the root cause of the cost anomaly, so that the final cost optimization suggestions are global, forward-looking, and transparent, enabling enterprises to achieve refined and automated multi-cloud billing analysis.

[0050] In some embodiments, the step of analyzing the root causes of cost anomalies based on the billing knowledge graph includes: The status of the billing knowledge graph in each billing cycle is fused with the cost data of the corresponding cycle to construct a time series graph sequence. Each time step of the time series graph sequence includes the cost time series features of the corresponding cycle. The time series sequence is trained based on the first prediction model, which is used to learn the normal evolution pattern of the cost time series features over time on the billing knowledge graph. For a new billing cycle, the corresponding time series graph sequence is input into the trained first prediction model to calculate the anomaly score of the bill knowledge graph on the cost entity node, and the anomaly status is determined based on the anomaly score. For the abnormal cost entity node, extract the graph attention weights from the first prediction model, and perform backtracking analysis along the relation edges in the billing knowledge graph to locate the root cause of the abnormality.

[0051] In this embodiment, based on the billing knowledge graph constructed in step S400 above, time-series cost data is fused to detect cost anomalies, and the root cause is traced along the semantic relationships in the billing knowledge graph for the generation of subsequent cost optimization suggestions.

[0052] Specifically, for each cost entity node in the billing knowledge graph, the cost amount of its historical billing cycles is extracted to form a standardized cost time series. In addition, the resource usage data corresponding to the cost entity node can be associated to form multi-dimensional time series features, such as [cost amount, resource usage data].

[0053] Furthermore, at the end of each billing cycle, a graph snapshot for that moment is generated. This snapshot is specifically the topological state of the billing knowledge graph at that moment, including all its nodes and edges. At the same time, the multi-dimensional time-series features formed by combining the above cost time series with resource usage data, as well as the feature vector of each node, are updated to the cost time-series features corresponding to that billing cycle. Thus, a time-series graph sequence {G_t, G_{t-1}, ...,G_{tn}} arranged in chronological order is constructed, where each G_t captures the network state of "service-resource-cost" at that moment.

[0054] In this embodiment, the aforementioned time series graph sequence composed of historical data without anomalies is used to train the first prediction model to learn the normal evolution pattern of cost time series features over time on the billing knowledge graph. This enables the new billing cycle to determine its anomalies based on the first prediction model and to perform backtracking analysis on relation edges to determine the root cause of the anomaly.

[0055] Specifically, the first prediction model is a temporal graph neural network model, comprising a spatial graph convolutional module and a temporal convolutional module. The spatial graph convolutional module employs a graph attention network or message-passing neural network, operating on a single graph snapshot G_t, updating the representation of each node by aggregating the features of neighboring nodes. For example, the features of a "cost entity node" are influenced by the features of "resource entity nodes" with which it has a consumption relationship and "business entity nodes" with which it has an association relationship, thus understanding the dependence of costs on business and resources. The temporal convolutional module employs a gated recurrent unit or a temporal convolutional network, modeling the feature sequence of each cost entity node along the time dimension, capturing its trend, periodicity, and stationarity over time. By tightly coupling the two modules in an interleaved or parallel manner, the model can simultaneously capture the graph structure that evolves over time and the temporal patterns that propagate on the graph structure. The outputs of the two modules are fused to predict the value of each cost entity node in the next billing cycle t+1.

[0056] More specifically, after a new billing cycle is completed, its snapshot G_t is input into the first prediction model trained above. The first prediction model outputs the predicted value for each cost entity node, i.e., the predicted distribution of costs for the next billing cycle. The actual observed value of the node is compared with the predicted value to calculate the reconstruction error. This error reflects the degree to which the cost entity node deviates from its normal evolution pattern. This error is then standardized and used as the node's anomaly score. When the anomaly score of a cost entity node exceeds a preset dynamic threshold, it is determined to be an abnormal cost event.

[0057] Furthermore, if an abnormal cost entity node is detected, the graph attention weights generated by the spatial graph convolution module during the prediction process of the first prediction model are analyzed. It can be understood that the graph attention weights quantify the contribution of the neighboring nodes to the abnormal state of the cost entity node. Then, the direct neighbor node with the highest contribution to the abnormal cost entity node is selected. Typically, this direct neighbor node is a resource entity node connected to it through a consumption edge. Then, the upstream is traced along the relationship edges such as "running on" and "call / dependency" in the billing knowledge graph, thereby obtaining a propagation path from the deep root cause to the surface cost anomaly.

[0058] Furthermore, by associating the relevant operation logs and performance metrics of the identified root cause node, it is determined whether there are any corresponding anomalies, such as a surge in error rate, configuration change events, or simultaneous spikes in CPU and memory usage. This forms a multi-dimensional chain of evidence, ultimately generating a diagnostic report containing anomaly descriptions, root cause localization, propagation path diagrams, and associated evidence as the root cause result of the cost anomaly.

[0059] In some embodiments, the cost optimization recommendations for generating overall cost and performance include: Based on the billing knowledge graph and the second prediction model, the cost of each node in the billing knowledge graph is predicted in future periods, and a future cost trend map is output. Define an optimization action library that includes configuration and change operations for various cloud platforms, and build a simulation engine based on the topological relationship of the billing knowledge graph to simulate the correlation between the optimization action library and cost and performance. Based on the root cause of cost anomalies, a set of candidate optimization strategies is generated. The global impact of each strategy in the set is simulated by the inference engine. A comprehensive evaluation is performed using a multi-objective evaluation function to generate a personalized optimization strategy report.

[0060] In this embodiment, based on the billing knowledge graph constructed in step S300 above, the cost prediction value of each node in the billing knowledge graph is predicted, thereby generating a future cost situation map, which is used to generate cost optimization suggestions for overall cost and performance.

[0061] Specifically, a temporal feature vector for prediction is constructed for each node in the billing knowledge graph. For cost entity nodes in the billing knowledge graph, their historical cost amount sequence is directly extracted, and their resource usage data sequence is fused to form a multi-dimensional temporal feature V_cost(t) = [cost_t, usage_t, ...]. For business entity nodes, the costs of all associated cost entity nodes under them are dynamically aggregated, that is, the cost amounts of all cost entity nodes connected by association edges, forming the historical total cost temporal sequence V_biz(t) of that business entity node. For resource entity nodes, the cost amounts of all cost entity nodes they generate are aggregated, that is, the cost amounts of all cost entity nodes connected by consumption edges, forming the historical total cost temporal sequence V_res(t) of that resource entity node.

[0062] More specifically, the subgraph structure of each node in the billing knowledge graph is obtained, and the temporal feature vector generated corresponding to each node is used as input features. This input is then fed into the second prediction model for prediction, generating the cost prediction value for each node. Specifically, the second prediction model is a graph embedding temporal prediction model. This model includes a graph encoder using a graph attention network or a graph convolutional network, a temporal encoder for capturing the temporal features of each node, and a predictor for prediction. Specifically, the graph encoder learns the representation of the subgraph structure of each node in the current associated network context through message passing and attention mechanisms, and finally outputs the structure of each node. The embedding vector is captured by the temporal encoder, which captures the historical dynamic temporal embedding vector of each node. In addition, the second prediction model also includes a feature fusion layer, which is used to fuse the structural embedding vector and the temporal embedding vector. The fusion method can be simple concatenation or an attention mechanism that allows the model to determine the focus bias during prediction. Finally, a joint representation vector of the node is formed. In the predictor, the joint representation vector of the target node is the core, and the joint representation vectors of its neighboring nodes are used as conditional inputs. In the predictor, the future cost trend is inferred collaboratively based on the joint representation vector of the target node itself and the joint feature vectors of its associated nodes.

[0063] It is worth noting that during the training process, the second prediction model uses a large number of samples constructed from historical sliding windows to train the model. The loss function usually adopts the smoothed average absolute error of multi-step prediction to ensure the accuracy of predictions at each future time point. Its training goal is to enable the second prediction model to accurately predict the future cost sequence of various nodes in the billing knowledge graph.

[0064] Furthermore, when a new round of optimization simulation is required, the second prediction model takes the latest billing knowledge graph and the time-series feature vectors constructed from the historical time-series data of all nodes up to the present as input. The model then predicts the costs of all nodes in the graph for the next H periods. The prediction results are added as new attributes to the corresponding nodes in the billing knowledge graph; for example, adding the attribute `future_cost_forecast=[month1_cost, month2_cost,...]` to each node. Finally, a future cost landscape map is generated, marked with future predicted costs. This map shows the expected cost distribution for each period, business segment, and resource in the future, without any optimization intervention.

[0065] Furthermore, an optimization action library is constructed, and combined with the billing knowledge graph, an inference engine is built to evaluate the potential impact of executing each atomic optimization action in the optimization action library.

[0066] In some embodiments, the definition of an optimization action library containing various cloud platform configuration and change operations, and the construction of a simulation engine based on the topological relationships of the billing knowledge graph to model the correlation between the optimization action library and cost and performance, includes: Define an optimization action library, wherein the atomic optimization actions in the optimization action library include at least adjusting resource configuration, changing billing mode, and starting / stopping services; Establish an impact calculation model to quantify the direct cost and performance impact of atomic optimization actions on target resource nodes; An impact propagation model is established, based on the edge types in the billing knowledge graph, to propagate the direct impact along the graph topology to the associated business entity nodes or resource entity nodes, and to calculate the indirect impact.

[0067] Specifically, the optimization action library is a rule base containing various atomic optimization actions, including adjusting resource configuration, changing billing models, and starting / stopping services. Adjusting resource configuration includes modifying instance type, storage tier, and database specifications; changing billing models includes switching from on-demand instances to reserved or spot instances; and starting / stopping services includes shutting down non-production environment resources and enabling data lifecycle policies. Furthermore, each atomic optimization action in the optimization action library is configured with a clear meaning and execution parameters, for example: {Action: Change instance type, Target: [Resource ID list], Parameter: Change from c5.xlarge to c5.large}; or {Action: Convert to reserved instance, Target: [Resource ID list], Parameter: 1-year prepaid}.

[0068] Furthermore, the constraint inference of the inference engine in the billing knowledge graph specifically includes the impact calculation model and the impact propagation model. For each atomic optimization action in the optimization action library, the target node of the action needs to be located in the billing knowledge graph, which is usually the resource entity node and the business entity node. Then, the direct and indirect impacts after executing the corresponding atomic optimization action are determined through the impact calculation model and the impact propagation model.

[0069] Specifically, regarding the impact on the computing model, based on the cloud service pricing model and the resource performance model, the direct impact of each atomic optimization action on its directly affected target node is calculated. For example, a downgrade action will directly reduce the future predicted cost of the resource and, according to the performance model, cause its estimated processing capacity to decrease by a percentage. Taking the example of {Action: Change instance type, Target: [Resource ID list], Parameter: Change from c5.xlarge to c5.large} above, downgrading an EC2 instance from c5.xlarge to c5.large will reduce its direct cost by approximately 50% according to the cloud service pricing model, while the estimated CPU computing power will decrease by approximately 30% according to the resource performance model.

[0070] Specifically, for the impact propagation model, impact propagation analysis is performed on the relationship edges such as calls, dependencies, and execution in the related billing knowledge graph. For example, downgrading the specifications of a database instance may affect the response time of all application services that call it. It is worth noting that the impact propagation model quantifies the corresponding related impacts through performance degradation models or by looking up preset SLA rules, and marks them on the affected related nodes.

[0071] In addition, the simulation engine integrates all direct and indirect effects to generate a new future cost and performance landscape diagram after simulating the corresponding atomic optimization action. This diagram is then compared with the future cost landscape diagram before the simulation to determine the optimization result of the atomic optimization action.

[0072] In some embodiments, the generation of a candidate optimization strategy set based on the root cause of cost anomalies, and the simulation of the global impact of each strategy in the candidate optimization strategy set by the inference engine, includes: Based on the root cause of the cost anomaly, a first class of candidate strategies is generated for the root cause. The resource usage data is analyzed to identify resource entities whose resource utilization rate is lower than a first threshold or higher than a second threshold, and a second type of candidate strategy for resource configuration adjustment actions is generated. The first type of candidate strategy and the second type of candidate strategy are filtered and sorted to generate a set of candidate optimization strategies; Each strategy in the candidate optimization strategy set is parsed into an ordered sequence of atomic optimization actions; Using the future cost landscape map as the initial state, the atomic optimization action sequence is executed sequentially; For each atomic optimization action, calculate its direct impact on the target node, and calculate its indirect impact on associated nodes based on the relation edges in the billing knowledge graph using a preset impact propagation model. By aggregating all direct and indirect effects, global simulation results are generated for the corresponding candidate optimization strategies.

[0073] In this embodiment, based on the root cause of cost anomalies, candidate optimization strategies are selected and combined from the optimization action library to generate a set of candidate optimization strategies. The global impact is then determined through the constructed inference engine, and the best cost optimization suggestion is finally determined.

[0074] Specifically, the diagnostic report representing the root cause of abnormal costs generated above is used as a decision factor to generate the first candidate strategy. If the diagnostic report indicates that "the root cause of abnormal costs is the surge in storage due to the failure to enable the automatic backup strategy of a certain database", then "enable and configure the automatic backup lifecycle strategy of the database" is generated as the core candidate action, which can ensure that the optimization can directly target the root cause of the problem.

[0075] Furthermore, resource usage data can be analyzed to identify resource entities with low utilization rates for resource reconfiguration and generate second-candidate strategies. For example, the utilization rates of resource entities such as CPU, memory, and storage IOPS in the resource entity nodes of the periodic analysis graph can be used. If a resource consistently falls below a set first threshold, it is marked as an inefficient resource entity, ultimately generating an inefficient resource list. For each resource entity in the inefficient resource list, candidate actions such as "resource downgrading," "resource hibernation," or "resource reclamation" are generated, such as "downgrading the instance size from m5.large to m5.small," "changing the storage volume type from gp3 to gp2," or "configuring a scheduled shutdown for development environment instances." Similarly, if a resource consistently exceeds a set second threshold, it is marked as a high-load resource entity, ultimately generating a high-load resource list. For each resource entity in the high-load resource list, "performance optimization assessment" or "horizontal scaling" suggestions are generated.

[0076] It is worth noting that, in one embodiment, for the generation of candidate strategies, a pre-defined strategy template library is used, defining validated optimization patterns for different scenarios. Based on the corresponding diagnostic reports and the analysis results of resource usage data, the appropriate strategy template is matched, and specific nodes are filled into the template to form candidate strategies. Furthermore, multiple atomic optimization actions with synergistic effects among the identified candidate strategies can be combined and encapsulated to generate composite strategies. For example, if multiple small database instances under a microservice are identified as being able to be downgraded, a composite strategy of "unified downgrading optimization of the database cluster" is generated, specifying the expected economies of scale.

[0077] Furthermore, for each candidate strategy in the candidate optimization strategy set, simulation and deduction need to be performed through the inference engine to determine its global impact. Specifically, the candidate strategy is parsed into a series of ordered atomic optimization actions, and the initial state of the simulation is set to the future cost situation map generated above. Then, the atomic optimization actions in the candidate strategy are input into the inference engine in sequence, and the direct and indirect impacts are deduced through the inference engine. Then, the direct and indirect impacts are aggregated as the global simulation result of the candidate strategy. Then, the simulated state of each node in the billing knowledge graph is updated, including cost and performance indicators, and finally, the simulated situation map is generated.

[0078] Furthermore, by comparing the future cost landscape diagrams before and after the simulation with the post-simulation landscape diagram, the total global cost savings brought about by the candidate strategies, the cost changes of each business line, and the change matrix of key performance indicators are determined.

[0079] Furthermore, a multi-objective evaluation function is defined to quantify the global impact of each candidate strategy, ultimately generating a personalized optimization strategy report. In one embodiment, the multi-objective evaluation function V(S) = α*[Cost_Saving(S) / Norm_Cost] - β*[Perf_Penalty(S) / Norm_Perf] - γ*[Risk_Score(S) / Norm_Risk], where Cost_Saving(S) represents the cost-saving benefit, indicating the net present value cost saving brought by candidate strategy S; Perf_Penalty(S) represents the performance impact penalty, indicating the weighted sum of the negative impacts of candidate strategy S on business performance indicators. Specifically, for each... For each affected business entity node, its simulated performance metrics are compared with the preset SLA target values ​​of the business entity nodes in the billing knowledge graph to generate a deviation value. This deviation value is then multiplied by the priority weight of the business entity node, and finally summed over all business entity nodes. The priority weight is positively correlated with business priority. Risk_Score(S) is the risk assessment score, representing the risk score of candidate strategy S, including technical risks (such as the probability of interruption when using auction instances), operational risks (change complexity), and business risks (such as impacting critical business SLAs). α, β, and γ are adjustable weighting coefficients reflecting the enterprise's different preferences for cost, performance, and risk. Norm_* is a normalization factor to ensure comparability between targets of different dimensions. Finally, all candidate strategies are ranked in descending order of V(S), and the candidate strategy with the highest V(S) score is selected as the recommended strategy, generating an optimization strategy report.

[0080] Furthermore, the optimization strategy report includes, but is not limited to, an execution summary, strategy details, global impact analysis, root cause analysis, implementation risk assessment, and contingency plans. The execution summary outlines the recommended strategy, its total expected savings, and basic information such as the main business departments involved. The strategy details list the specific optimization actions, target resources, and expected direct savings step-by-step; for composite strategies, the synergistic logic between actions is explained. The global impact analysis uses charts to show cost comparisons before and after simulation, performance change heatmaps, etc. The root cause analysis clearly indicates that the recommended strategy aims to address the root causes of cost anomalies. The implementation risk assessment and contingency plans list the identified major risk points (such as performance impact, service interruption window), risk levels, and suggested mitigation measures, such as "execute during off-peak business periods" and "closely monitor API latency after implementation."

[0081] Please see Figure 2 As shown, the present invention also provides a multi-cloud billing analysis system based on artificial intelligence, the system comprising: First processing module 201: Used to collect raw billing data and associated resource configuration data and resource usage data from different cloud platforms, and generate raw data; The second processing module 202 is used to parse the text description of the original data, classify it into predefined business semantic categories, and generate semantic billing data. The third processing module 203 is used to construct a bill knowledge graph based on the semantic bill data, the resource configuration data, and the resource usage data. The fourth processing module 204 is used to analyze the root causes of cost anomalies based on the billing knowledge graph and generate cost optimization suggestions that combine cost and performance.

[0082] It is understandable that, such as Figure 1 The content shown in the embodiments of the AI-based multi-cloud billing analysis method is applicable to the embodiments of this AI-based multi-cloud billing analysis system. The specific functions implemented in the embodiments of this AI-based multi-cloud billing analysis system are the same as those shown below. Figure 1 The illustrated multi-cloud billing analysis method based on artificial intelligence is the same as that shown in the example, and achieves the same beneficial effects. Figure 1 The beneficial effects achieved by the illustrated AI-based multi-cloud billing analysis method embodiment are also the same.

[0083] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0084] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0085] Please see Figure 3 As shown, this embodiment of the invention also provides a computer device 3, including: a memory 302 and a processor 301, and a computer program 303 stored on the memory 302. When the computer program 303 is executed on the processor 301, it implements the multi-cloud billing analysis method based on artificial intelligence as described in any of the above methods.

[0086] The computer device 3 may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that... Figure 3 The computer device 3 is merely an example and does not constitute a limitation on the computer device 3. It may include more or fewer components than shown in the figure, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0087] The processor 301 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0088] In some embodiments, the memory 302 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 302 may be an external storage device of the computer device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 3. Furthermore, the memory 302 may include both internal and external storage units of the computer device 3. The memory 302 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory 302 can also be used to temporarily store data that has been output or will be output.

[0089] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the AI-based multi-cloud billing analysis method as described in any of the above methods.

[0090] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographic device / computer device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0091] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A multi-cloud billing analysis method based on artificial intelligence, characterized in that, include: Collect raw billing data and associated resource configuration and resource usage data from different cloud platforms to generate raw data; The text description of the raw data is parsed and classified into predefined business semantic categories to generate semantic billing data. Based on the semantic billing data, the resource configuration data, and the resource usage data, a billing knowledge graph is constructed. Based on the aforementioned billing knowledge graph, the root causes of cost anomalies are analyzed, and cost optimization suggestions for overall cost and performance are generated.

2. The method as described in claim 1, characterized in that, The process of parsing the textual description of the raw data, classifying it into predefined business semantic categories, and generating semantic billing data includes: Based on a predefined unified data model, the original data is standardized in format and cleaned to generate structured billing data; The text description field in the structured billing data is input into a pre-trained deep learning classifier, which outputs a classification result corresponding to a predefined multi-level business semantic label system to generate semantic billing data.

3. The method as described in claim 1, characterized in that, The construction of a billing knowledge graph based on the semantic billing data, the resource configuration data, and the resource usage data includes: Define a graph pattern, which includes at least business entity nodes, resource entity nodes, cost entity nodes, and edge types representing the relationships between nodes. The edge types include at least "belong to", "run in", "consume", "associate", and "call / depend on". Based on the semantic billing data, the resource configuration data, and the resource usage data, corresponding graph nodes are automatically created or updated through entity extraction. By analyzing resource tags and configurations through a rules engine and combining them with network traffic log analysis, relationship edges between the nodes of the graph are established to form a billing knowledge graph. Monitor cloud platform resource events and dynamically update the nodes and edges in the billing knowledge graph according to the event type.

4. The method as described in claim 1, characterized in that, The analysis of the root causes of cost anomalies based on the billing knowledge graph includes: The status of the billing knowledge graph in each billing cycle is fused with the cost data of the corresponding cycle to construct a time series graph sequence. Each time step of the time series graph sequence includes the cost time series features of the corresponding cycle. The time series sequence is trained based on the first prediction model, which is used to learn the normal evolution pattern of the cost time series features over time on the billing knowledge graph. For a new billing cycle, the corresponding time series graph sequence is input into the trained first prediction model to calculate the anomaly score of the bill knowledge graph on the cost entity node, and the anomaly status is determined based on the anomaly score. For the abnormal cost entity node, extract the graph attention weights from the first prediction model, and perform backtracking analysis along the relation edges in the billing knowledge graph to locate the root cause of the abnormality.

5. The method as described in claim 1, characterized in that, The cost optimization suggestions for generating overall cost and performance include: Based on the billing knowledge graph and the second prediction model, the cost of each node in the billing knowledge graph is predicted in future periods, and a future cost trend map is output. Define an optimization action library that includes configuration and change operations for various cloud platforms, and build a simulation engine based on the topological relationship of the billing knowledge graph to simulate the correlation between the optimization action library and cost and performance. Based on the root cause of cost anomalies, a set of candidate optimization strategies is generated. The global impact of each strategy in the set is simulated by the inference engine. A comprehensive evaluation is performed using a multi-objective evaluation function to generate a personalized optimization strategy report.

6. The method as described in claim 5, characterized in that, The definition includes an optimization action library containing various cloud platform configuration and change operations. Based on the topological relationships of the billing knowledge graph, a simulation engine is built to model the correlation between the optimization action library and cost and performance, including: Define an optimization action library, wherein the atomic optimization actions in the optimization action library include at least adjusting resource configuration, changing billing mode, and starting / stopping services; Establish an impact calculation model to quantify the direct cost and performance impact of atomic optimization actions on target resource nodes; An impact propagation model is established, based on the edge types in the billing knowledge graph, to propagate the direct impact along the graph topology to the associated business entity nodes or resource entity nodes, and to calculate the indirect impact.

7. The method as described in claim 6, characterized in that, Based on the root cause results of cost anomalies, a set of candidate optimization strategies is generated, and the global impact of each strategy in the candidate optimization strategy set is simulated by the inference engine, including: Based on the root cause of the cost anomaly, a first class of candidate strategies is generated for the root cause. The resource usage data is analyzed to identify resource entities whose resource utilization rate is lower than a first threshold or higher than a second threshold, and a second type of candidate strategy for resource configuration adjustment actions is generated. The first type of candidate strategy and the second type of candidate strategy are filtered and sorted to generate a set of candidate optimization strategies; Each strategy in the candidate optimization strategy set is parsed into an ordered sequence of atomic optimization actions; Using the future cost landscape map as the initial state, the atomic optimization action sequence is executed sequentially; For each atomic optimization action, calculate its direct impact on the target node, and calculate its indirect impact on associated nodes based on the relation edges in the billing knowledge graph using a preset impact propagation model. By aggregating all direct and indirect effects, global simulation results are generated for the corresponding candidate optimization strategies.

8. A multi-cloud billing analysis system based on artificial intelligence, characterized in that, include: The first processing module is used to collect raw billing data from different cloud platforms and their associated resource configuration data and resource usage data to generate raw data. The second processing module is used to parse the text description of the raw data, classify it into predefined business semantic categories, and generate semantic billing data. The third processing module is used to construct a billing knowledge graph based on the semantic billing data, the resource configuration data, and the resource usage data. The fourth processing module is used to analyze the root causes of cost anomalies based on the billing knowledge graph and generate cost optimization suggestions that combine cost and performance.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images