Cloud platform data storage method and storage system

Abstract

The invention relates to the technical field of cloud computing data management, and discloses a cloud platform data storage method and storage system.The method comprises the steps that when a data storage task is created, full-life-cycle monitoring on a storage data object is started, and a data state tracking sequence is generated; performing multi-dimensional feature analysis based on the sequence, and extracting a storage feature vector representing data characteristics; inputting the feature vector into a pre-constructed storage mode decision network to obtain an initial storage mode identifier; according to the identifier, activating a corresponding storage resource evaluation process, and generating a dynamic storage evaluation parameter; and fusing the dynamic storage evaluation parameter and the initial storage mode identifier through a storage strategy generator, and outputting a final storage strategy instruction. According to the method, the data full-life-cycle monitoring characteristics are combined with dynamic resource evaluation, so that the rigidity of a traditional static storage strategy is overcome, and the accuracy and the adaptive capacity of cloud platform storage resource allocation are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of cloud computing data management technology, specifically to a method and system for storing cloud platform data. Background Technology

[0002] Current cloud platform data storage primarily employs fixed strategies or simple rule-based storage methods. Existing technologies have limited monitoring dimensions for data objects, failing to achieve continuous tracking throughout the entire lifecycle. Feature extraction methods are simplistic and cannot fully reflect the multidimensional characteristics of data. Storage decision-making mechanisms are rigid, lacking an intelligent mapping relationship between data features and storage patterns. Resource assessment is static, unable to adapt to dynamic changes in cloud platform resources. The strategy generation process is linear, lacking the ability to integrate and optimize multiple factors. Existing methods need to address key technical issues such as full lifecycle monitoring, intelligent feature parsing, and dynamic decision optimization.

[0003] Traditional cloud storage methods suffer from significant shortcomings in adaptability and efficiency. Lifecycle monitoring coverage is narrow, and key state transition nodes are not fully captured. State sequence data analysis is rudimentary, and temporal feature mining is insufficient. Feature vector construction has a single dimension, and representations of important behavioral patterns are lacking. Decision networks have insufficient training data, resulting in limited pattern recognition accuracy. Resource assessment parameters are updated laggingly and are out of touch with reality. Policy generators have one-sided optimization objectives, failing to balance performance and cost. Instruction generation mechanisms are rigid and unable to cope with complex storage scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for storing cloud platform data, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution provided by the present invention is as follows: A method for storing data on a cloud platform, the method comprising: When creating a data storage task on the cloud platform, start full lifecycle monitoring of the stored data object and generate a data status tracking sequence; Based on the data state tracking sequence, multi-dimensional feature analysis is performed on the stored data object to extract the stored feature vector; The storage feature vector is input into a pre-constructed storage pattern decision network, which outputs an initial storage pattern identifier. Based on the initial storage mode identifier, activate the corresponding storage resource evaluation process and generate dynamic storage evaluation parameters; By combining the dynamic storage evaluation parameters with the initial storage mode identifier, a final storage policy instruction is generated through a storage policy generator.

[0006] Preferably, the step of initiating full lifecycle monitoring of stored data objects and generating a data status tracking sequence includes: When a data write operation is initiated, the data creation timestamp, data source identity information, and initial data size are recorded. Continuously monitor the frequency distribution, access time patterns, and requesting user categories of data read operations; Periodically scan the data content and record the data block integrity check value and data update history; The data creation timestamp, data source identity information, initial data size, frequency distribution, access time pattern, requesting user category, data block integrity verification value, and data update history are integrated and encoded in chronological order to generate a data status tracking sequence.

[0007] Preferably, the step of performing multi-dimensional feature parsing on the stored data object and extracting the stored feature vector includes: The time dimension features are separated from the data state tracking sequence, and the data activity cycle and access interval standard deviation are calculated; Access pattern features are extracted from the data state tracking sequence to identify high-frequency access periods and low-frequency access periods; Data value features are extracted from the data status tracking sequence to assess the importance of data-related businesses and the intensity of data updates; The data activity cycle, access interval standard deviation, high-frequency access period, low-frequency access period, importance of data-related business, and data update intensity are normalized and concatenated to form a fixed-dimensional storage feature vector.

[0008] Preferably, the step of inputting the storage feature vector into a pre-constructed storage pattern decision network and outputting an initial storage pattern identifier includes: The storage mode decision network comprises multiple fully connected layers and a nonlinear activation function; The stored feature vector is input into the input layer of the storage pattern decision network, and then transformed layer by layer through the multiple fully connected layers and nonlinear activation functions; In the output layer of the storage mode decision network, the transformed features are mapped to a variety of predefined storage mode categories, and the storage mode category with the highest output probability is encoded as the initial storage mode identifier.

[0009] Preferably, the activation of the corresponding storage resource evaluation process to generate dynamic storage evaluation parameters includes: Based on the initial storage mode identifier, retrieve the matching set of evaluation rules from the resource evaluation strategy library; Based on the aforementioned evaluation rule set, query in real time the remaining capacity, input / output performance indicators, and network bandwidth status of the cloud platform's currently available storage pool; By combining the access pattern features and data value features in the storage feature vector, the expected resource consumption is calculated; By combining the remaining capacity, input / output performance indicators, network bandwidth status, and expected resource consumption, dynamic storage evaluation parameters are generated.

[0010] Preferably, the step of generating dynamic storage evaluation parameters by comprehensively considering the remaining capacity, input / output performance indicators, network bandwidth status, and expected resource consumption includes: The remaining capacity is standardized by converting the current remaining capacity into a percentage value based on the total capacity of the storage pool and the zero boundary. The input and output performance indicators are standardized, and based on the theoretical maximum throughput and zero boundary, the current input and output performance indicators are converted into performance indicator ratio values. The network bandwidth status is standardized by converting the current network bandwidth status into a bandwidth status ratio value based on the physical bandwidth upper limit and the zero value boundary. The expected resource consumption is standardized by converting the current expected resource consumption into a consumption ratio based on the historical maximum consumption value and the zero value boundary. Based on the initial storage mode identifier, obtain the corresponding remaining capacity weight, performance index weight, bandwidth status weight, and consumption weight from the weight configuration library; The remaining capacity ratio is weighted using the remaining capacity weight to obtain a weighted remaining capacity value; The performance index ratio value is weighted using the performance index weight to obtain a weighted performance index value; The bandwidth state ratio value is weighted using the bandwidth state weight to obtain a weighted bandwidth state value; The consumption ratio value is weighted using the consumption weight to obtain a weighted consumption value; The weighted remaining capacity value, weighted performance index value, weighted bandwidth status value, and weighted consumption value are added together to obtain a comprehensive evaluation value; The comprehensive evaluation value is used as a dynamically stored evaluation parameter.

[0011] Preferably, the step of fusing the dynamic storage evaluation parameters and the initial storage mode identifier to generate a final storage policy instruction through a storage policy generator includes: The dynamic storage evaluation parameters are concatenated with the initial storage mode identifier to form a strategy decision input vector; The policy decision input vector is input to the storage policy generator, which is a trained deep learning model. The storage strategy generator outputs multiple strategy parameters, including target storage medium type, data redundancy backup strategy, data fragment size, and data compression algorithm selection. The aforementioned multiple policy parameters are encapsulated into a final storage policy instruction that can be parsed by the cloud platform storage engine.

[0012] Preferably, the training process of the storage policy generator includes: Collect historical cloud platform data storage tasks and their corresponding storage feature vectors, storage mode identifiers, resource evaluation parameters, and manually labeled optimal storage strategies; The deep learning model is constructed by taking a combination of storage feature vectors, storage mode identifiers and resource evaluation parameters as input, and the optimal storage strategy as the desired output. The parameters of the deep learning model are optimized using the gradient descent algorithm until the error between the model's prediction and the optimal storage strategy is lower than a preset threshold.

[0013] Preferably, the method further includes a feedback optimization step after executing the storage policy instruction: After the final storage strategy instruction is issued to the cloud platform storage engine, actual storage performance data is continuously collected. The actual storage performance data is compared with the expected indicators in the dynamic storage evaluation parameters to generate a performance deviation report; Based on the performance deviation report, the parameters of the storage mode decision network and the storage policy generator are fine-tuned.

[0014] Preferably, the present invention also includes a cloud platform data storage system, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the cloud platform data storage method described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: When a data storage task is created on the cloud platform, full lifecycle monitoring of the stored data objects is initiated, generating a data status tracking sequence. Monitoring covers all stages of data creation, access, modification, migration, and destruction. The status tracking sequence records the state migration trajectory of the data objects using timestamps. The sequence data is updated using an event-driven approach to ensure real-time performance. Through full lifecycle monitoring, the dynamic behavioral characteristics of data objects are fully captured.

[0016] Multidimensional feature analysis is performed on stored data objects based on data state tracking sequences to extract storage feature vectors. Feature analysis employs time series analysis methods to identify key features such as access patterns and popularity changes. Multidimensional features include indicators such as access frequency, data correlation, and value decay rate. The feature vectors are dimensionality-reduced to retain the most distinctive feature combinations. The analysis process involves incremental computation to adapt to changes in data behavior. Through multidimensional feature analysis, a precise profile of the data objects is constructed.

[0017] The storage feature vector is input into a pre-constructed storage pattern decision network, which outputs an initial storage pattern identifier. The decision network employs a deep learning model, trained on historical data to acquire classification capabilities. The network structure includes a multilayer perceptron, supporting non-linear feature mapping. The initial identifier corresponds to the recommended storage level and storage type. The identifier output includes a confidence assessment to improve decision reliability. Through this intelligent decision network, accurate storage pattern recommendations are achieved.

[0018] The initial storage mode identifier activates the corresponding storage resource evaluation process, generating dynamic storage evaluation parameters. The evaluation process designs specific evaluation metrics for different storage modes. Dynamic parameters include real-time data such as current storage pool capacity, IOPS performance, and network bandwidth. Parameter collection frequency is adaptively adjusted based on system load. The evaluation results quantify the availability and cost of storage resources. Through dynamic evaluation, the real-time status of the storage environment is understood.

[0019] By integrating dynamic storage evaluation parameters and initial storage mode identifiers, a storage policy generator produces the final storage policy instructions. The fusion algorithm considers the balance between resource constraints and data requirements. The policy generator employs multi-objective optimization, simultaneously optimizing performance and cost. The instructions include specific storage locations, replication policies, and migration plans. Conflict detection is incorporated into the generation process to ensure policy feasibility. Through intelligent fusion, the optimal storage solution is produced. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the working principle of the cloud platform data storage method described in this invention. Figure 2 A flowchart for generating a data state tracking sequence; Figure 3 This is a flowchart for storing feature vector extraction. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 This invention provides a method and system for storing cloud platform data. The method includes: immediately initiating full lifecycle monitoring of the stored data object when a data storage task is created. This monitoring continuously tracks the changes in various states of the data from writing to deletion and generates a data state tracking sequence. Based on the data state tracking sequence, multi-dimensional feature analysis is performed on the stored data object to extract a storage feature vector. The storage feature vector is input into a pre-constructed storage mode decision network, which outputs an initial storage mode identifier. According to the initial storage mode identifier, the corresponding storage resource evaluation process is activated to generate dynamic storage evaluation parameters. The dynamic storage evaluation parameters and the initial storage mode identifier are fused, and a final storage policy instruction is generated through a storage policy generator. This instruction is directly sent to the cloud platform storage engine for execution.

[0023] Example 1, see Figure 2 In practical implementation, the cloud platform's data storage method initiates full lifecycle monitoring of stored data objects when a data storage task is created, generating a data status tracking sequence. This process involves multiple data collection and integration stages. When a data write operation is initiated, the system records the data creation timestamp, data source identity information, and initial data size. The data creation timestamp is accurate to the millisecond level, the data source identity information includes user and application identifiers, and the initial data size is quantified in bytes. The system continuously monitors the frequency distribution, access time patterns, and requesting user categories for data read operations. The frequency distribution is determined by statistically analyzing the number of accesses per unit time and fitting a distribution model. Access time patterns identify access patterns during the day and night, or weekdays and weekends. Requesting user categories are categorized based on user roles or permission levels. The system periodically scans data content, recording data block integrity check values ​​and data update history. Data block integrity check values ​​are calculated using a hash algorithm, and the data update history records the timestamp and summary of each data modification.

[0024] In some embodiments, the generation of the data status tracking sequence is achieved through time-sequential encoding, in which various types of record information are sorted and concatenated by timestamps. In specific implementations, when integrating data creation timestamps, data source identity information, initial data size, frequency distribution, access time patterns, requesting user categories, data block integrity check values, and data update history, a linear sequence structure is adopted, with each data item accompanied by a timestamp. Optionally, the sequence encoding uses a fixed-length format or a variable-length format. The fixed-length format allocates a fixed number of bytes to each data item, while the variable-length format adds a length identifier. The sequence generation formula is expressed as: ; in: It is a data state tracking sequence. It is the i-th timestamp. These are the corresponding data record items, which include the data creation timestamp, data source identity information, and initial data size.

[0025] It is understandable that the generation of data status tracking sequences relies on a real-time monitoring module and a log system. The monitoring module intercepts data operation requests and extracts parameters. In some embodiments, frequency distribution monitoring is achieved through a sliding time window algorithm, and access time patterns use clustering methods to identify peak and trough periods. Periodic scanning of data content is executed by background task scheduling, and the scanning interval is dynamically adjusted according to data activity. Optionally, data block integrity verification values ​​are calculated using cyclic redundancy check or message digest algorithms, and data update history trajectories are stored in the form of incremental logs. The integration process uses a serialization library to convert heterogeneous data into a unified encoding format, ensuring the parsability and integrity of the sequence.

[0026] Example 2, see Figure 3 In practical implementation, time-dimensional features are extracted from the data status tracking sequence, and the data activity period and access interval standard deviation are calculated. The data activity period is obtained by extracting the minimum and maximum timestamps from the data status tracking sequence and calculating their difference. The access interval standard deviation is calculated based on the timestamp interval sequence of consecutive access events in the data status tracking sequence. The formula for calculating the access interval standard deviation is as follows: ; in: It is the standard deviation of the access interval. It is the total number of access intervals. It is the value of the j-th access interval. This is the average of all access intervals, derived from the timestamp sequence of the data status tracking sequence. Access pattern features are extracted from the data status tracking sequence to identify high-frequency and low-frequency access periods. These periods are defined by analyzing access time patterns within the data status tracking sequence, counting the number of accesses per unit time, and applying frequency thresholds. High-frequency access periods correspond to times when the access frequency exceeds a preset upper threshold, while low-frequency access periods correspond to times when the access frequency is below a preset lower threshold. Data value features are also extracted from the data status tracking sequence to assess the importance of data-related business and the intensity of data updates. The importance of data-related business is assigned a weighted score based on business logic, and the intensity of data updates is quantified by the frequency and magnitude of changes in the historical data update trajectory recorded in the data status tracking sequence.

[0027] In some embodiments, the separation of time-dimensional features involves parsing the timestamp field of the data state tracking sequence, with the data activity period calculated as the absolute time difference between the first and last timestamps within the data lifecycle. The separation of access pattern features uses a sliding window algorithm to scan access records in the data state tracking sequence, and the identification of high-frequency and low-frequency access periods is based on clustering analysis results of historical access data. The evaluation of data value features integrates external business metadata; the importance score of data-related business comes from the business configuration library; and the data update intensity is calculated as the ratio of the number of update events in the data state tracking sequence to the time span.

[0028] Optionally, the calculation of time-dimensional features can incorporate a weighted average method to handle abnormal timestamps, and the identification of access pattern features can adopt a dynamic threshold adjustment mechanism to adapt to changes in access patterns. The evaluation of data value features allows for the integration of comprehensive scores from multiple business indicators, and the quantification of data update intensity supports the use of logarithmic scaling to handle high-frequency update scenarios.

[0029] The data activity cycle, access interval standard deviation, high-frequency access period, low-frequency access period, importance of data-related business, and data update intensity are normalized. Normalization uses a linear transformation to map each feature value to a range between zero and one. The concatenation process connects the normalized feature values ​​into a one-dimensional array in a fixed order, forming a fixed-dimensional storage feature vector. The storage pattern decision network contains multiple fully connected layers and non-linear activation functions. The storage feature vector is input to the input layer of the storage pattern decision network. The number of neurons in the input layer matches the dimension of the storage feature vector. After layer-by-layer transformation through multiple fully connected layers and non-linear activation functions, each fully connected layer performs a linear transformation between the weight matrix and the input vector. The non-linear activation function uses the hyperbolic tangent function to introduce non-linear characteristics. In the storage pattern decision network, the hyperbolic tangent function is used as a non-linear activation function in the output transformation process of each fully connected layer. This function maps the linearly transformed neuron activation values ​​to a continuous range from -1 to 1. In specific implementation, the hyperbolic tangent function is called after each fully connected layer, and its output is used as the input of the next layer.

[0030] In some embodiments, the storage mode decision network comprises multiple fully connected layers with decreasing numbers of neurons. A non-linear activation function is applied after each fully connected layer. The output layer maps the transformed features to multiple predefined storage mode categories, and the storage mode category code with the highest output probability is used as the initial storage mode identifier. The storage mode category code corresponds to the storage type supported by the cloud platform, such as object storage, block storage, or file storage.

[0031] The storage pattern decision network is trained using labeled historical data, with predefined storage pattern categories covering common storage scenarios. The softmax function in the output layer transforms the network output into a probability distribution. The initial storage pattern identifier generation process is integrated into the cloud platform's storage task scheduler for automated decision-making. Optionally, the storage pattern decision network can be deployed as a lightweight model to reduce inference latency, and the choice of non-linear activation function is based on a trade-off between model accuracy and computational cost. The output layer mapping supports a scalable number of categories to adapt to updates in the cloud platform's storage services. The fixed-dimensional design of the storage feature vectors ensures compatibility with the input layer of the storage pattern decision network. The layer-by-layer transformation of the fully connected layers gradually extracts high-level features, and the storage pattern category encoding with the highest probability is selected from the output probability distribution using the argmax function.

[0032] In Example 3, in specific implementation, the corresponding storage resource evaluation process is activated based on the initial storage mode identifier to generate dynamic storage evaluation parameters. This process begins by calling the evaluation rule set matching the initial storage mode identifier from the resource evaluation strategy library. The evaluation rule set defines the parameter query paradigm and calculation logic. Based on the evaluation rule set, the remaining capacity, input / output performance indicators, and network bandwidth status of the currently available storage pool on the cloud platform are queried in real time. The remaining capacity is obtained from the storage pool management interface in bytes. Input / output performance indicators include the number of read / write operations per second and throughput. Network bandwidth status refers to the bit rate of the currently available network bandwidth. Combining the access pattern features and data value features in the storage feature vector, the expected resource consumption is calculated. The access pattern features are used to predict the frequency and scale of future data access, while the data value features affect data priority, thereby adjusting resource reservation. The expected resource consumption is estimated using a linear regression model based on historical consumption data and current feature values.

[0033] In some embodiments, the resource assessment strategy library is stored in key-value pairs, where the key is the storage mode category code and the value is the corresponding assessment rule set. The invocation process is implemented via hash lookup. Real-time query operations use a lightweight agent to periodically poll the storage pool monitoring service. Input and output performance metrics are read from performance counters, and network bandwidth status is queried through the network management interface. The model for calculating expected resource consumption is trained using historical data before deployment to ensure prediction accuracy.

[0034] Optionally, the evaluation rule set can be configured to be dynamically loaded, supporting runtime updates to adapt to policy changes. The real-time query mechanism can replace polling with an event-driven model, reducing system overhead. The expected resource consumption calculation model allows switching to different machine learning algorithms, such as decision trees or support vector machines.

[0035] Dynamic storage evaluation parameters are generated by combining remaining capacity, input / output performance metrics, network bandwidth status, and expected resource consumption. This process involves standardization and weighted summation. Remaining capacity is standardized by converting it into a percentage value based on the total storage pool capacity and a zero-value boundary. This percentage is calculated as the ratio of the current remaining capacity to the total storage pool capacity and is cropped to the range of zero to one. Input / output performance metrics are standardized by converting them into percentage values ​​based on the theoretical maximum throughput and a zero-value boundary. This percentage is calculated as the ratio of the current input / output performance metric reading to the theoretical maximum throughput. Network bandwidth status is standardized by converting it into a percentage value based on the physical bandwidth limit and a zero-value boundary. This percentage is calculated as the ratio of the current available bandwidth to the physical bandwidth limit. Expected resource consumption is standardized by converting it into a percentage value based on the historical maximum consumption and a zero-value boundary. This percentage is calculated as the ratio of the expected resource consumption to the historical maximum consumption. Based on the initial storage mode identifier, the corresponding remaining capacity weight, performance metric weight, bandwidth status weight, and consumption weight are retrieved from the weight configuration library. The weight configuration library stores preset weight values ​​for different storage modes. The remaining capacity ratio is weighted using the remaining capacity weight to obtain a weighted remaining capacity value. The performance metric ratio is weighted using the performance metric weight to obtain a weighted performance metric value. The bandwidth status ratio is weighted using the bandwidth status weight to obtain a weighted bandwidth status value. The consumption ratio is weighted using the consumption weight to obtain a weighted consumption value. The weighted remaining capacity value, weighted performance metric value, weighted bandwidth status value, and weighted consumption value are added together to obtain a comprehensive evaluation value. The formula for calculating the comprehensive evaluation value is as follows: ; in: It is a comprehensive evaluation value. It is the weight of remaining capacity. It is the percentage of remaining capacity. It is the weight of performance indicators. It is a performance indicator ratio value. It is the bandwidth state weight. It is the bandwidth status ratio value. It is the consumption weight. This is the consumption ratio. The comprehensive evaluation value will be used as a dynamically stored evaluation parameter.

[0036] In some embodiments, the normalization process employs a min-max scaling method, with zero-value boundaries set to zero to prevent negative values. Queries to the weight configuration library are indexed based on the initial storage mode identifier, with weight values ​​being floating-point numbers. The weighted summation operation is performed in memory, and the overall evaluation value is output as a single-precision floating-point number.

[0037] It is understandable that standardization ensures the comparability of parameters with different dimensions, and weight configuration reflects the differences in resource dimension preferences among different storage modes. The comprehensive evaluation value quantifies the degree of matching between the current resource status and expected needs. Optionally, standardization can introduce logarithmic transformation to handle skewed data, and weight configuration can be manually adjusted by administrators to optimize strategies. The weighted summation formula can be extended to include more resource dimensions, enhancing the comprehensiveness of the evaluation. It is also understandable that the generation process of dynamic storage evaluation parameters is integrated into the resource evaluation service, and the comprehensive evaluation value is passed as a key indicator to the subsequent strategy generation stage.

[0038] In Example 4, in specific implementation, dynamic storage evaluation parameters are concatenated with the initial storage mode identifier to form a policy decision input vector. The concatenation operation treats the dynamic storage evaluation parameters as a scalar element and joins them end-to-end with the encoded value of the initial storage mode identifier, which is typically an integer encoding. The policy decision input vector is a one-dimensional numerical array. This policy decision input vector is then input to a storage policy generator, which is a trained deep learning model. This model receives the policy decision input vector as input and outputs a set of parameters representing the storage configuration. The storage policy generator outputs multiple policy parameters, including the target storage medium type, data redundancy backup strategy, data shard size, and data compression algorithm selection. These multiple policy parameters are then encapsulated into a final storage policy instruction that can be parsed by the cloud platform storage engine. The encapsulation process serializes each parameter according to a predefined protocol format, such as JSON or Protocol Buffers. The final storage policy instruction contains complete storage configuration information.

[0039] The training process of the storage strategy generator includes collecting historical cloud platform data storage tasks and their corresponding storage feature vectors, storage mode identifiers, resource evaluation parameters, and manually labeled optimal storage strategies. A deep learning model is constructed, taking a combination of storage feature vectors, storage mode identifiers, and resource evaluation parameters as input. The combined input is a vector formed by concatenating these elements. The expected output is the optimal storage strategy, which is labeled by domain experts based on the actual performance of historical storage tasks. The parameters of the deep learning model are optimized using a gradient descent algorithm until the error between the model's prediction and the optimal storage strategy is below a preset threshold. The error is calculated using a loss function, and the preset threshold is set according to the accuracy requirements of the application scenario.

[0040] In some embodiments, the dimension of the policy decision input vector is fixed; for example, the initial storage mode identifier is a one-bit integer, the dynamic storage evaluation parameter is a one-bit floating-point number, and the two are concatenated with the storage feature vector to form a fixed-length input. The storage policy generator can employ a multilayer perceptron or convolutional neural network structure, with its output layer corresponding to discrete or continuous values ​​of multiple policy parameters.

[0041] Optionally, the input to the storage policy generator can include additional contextual features, such as timestamps or user identification information. The deep learning model architecture can be a recurrent neural network to handle temporal dependencies. The gradient descent algorithm can employ stochastic gradient descent or a variant of Adam to accelerate convergence.

[0042] As can be understood (referring to Table 1), training the storage policy generator requires a large-scale, high-quality labeled dataset, and the label consistency of the optimal storage policy is crucial to model performance. The design of the loss function needs to consider the different data types of multiple policy parameters; for example, cross-entropy loss is used for classification tasks, while mean squared error is used for regression tasks.

[0043]

[0044] The model parameter optimization process minimizes the difference between the predicted output and the true label, and the degree of difference is determined by the loss function: ; in: It is the loss value. It is the total number of output strategy parameters. It is the first The actual labeled values ​​of each strategy parameter The model is for the first The predicted values ​​of each policy parameter. The gradient descent algorithm gradually reduces the loss value by calculating the gradient of the loss function with respect to the model parameters and updating the parameters along the negative gradient direction.

[0045] In some embodiments, the training process employs mini-batch gradient descent, where the average gradient is calculated using a batch of training samples in each iteration. Model training continues, with performance periodically evaluated on a validation set. Training stops when the loss value on the validation set falls below a preset threshold. It is understood that the performance of the storage policy generator depends on the coverage and quality of the training data, and the model needs to be retrained periodically using newly generated data to adapt to system changes.

[0046] In Example 5, in practical implementation, after the final storage policy instruction is issued to the cloud platform storage engine, the system initiates a feedback optimization step, continuously collecting actual storage performance data. This actual storage performance data includes storage operation latency, data throughput, storage resource utilization, and error rate metrics. This data is collected in real-time through the cloud platform's built-in monitoring agent and performance counters. The actual storage performance data is compared with the expected metrics in the dynamic storage evaluation parameters to generate a performance deviation report. The comparison process calculates the absolute or relative difference between the actual and expected values. The performance deviation report records the magnitude and direction of the deviation for each performance dimension in a structured document format.

[0047] The formula for generating performance deviation reports is expressed as follows: ; in: It is the mean absolute deviation. It is the total number of performance indicators. It is the actual measured value of the i-th performance index. This is the expected value of the i-th performance metric, derived from the performance target derived from the dynamic storage evaluation parameters. Based on the performance deviation report, the parameters of the storage mode decision network and the storage policy generator are fine-tuned. The fine-tuning process uses the performance deviation as a loss signal and updates the network weights slightly through the backpropagation algorithm.

[0048] In some embodiments, actual storage performance data is collected at fixed time intervals, such as once per minute, and a moving average is calculated to reduce noise. The generation of performance deviation reports triggers the policy optimization process, initiating fine-tuning when the mean absolute deviation exceeds a preset tolerance threshold. Fine-tuning of the storage mode decision network focuses on adjusting the weight matrix of the fully connected layers, while fine-tuning of the storage policy generator updates the internal parameters of its deep learning model.

[0049] Optionally, the collection of actual storage performance data can be event-driven, triggering data collection only after significant storage operations are completed. Performance deviation reports can include trend analysis to predict the likelihood of future performance drift. Fine-tuning can be performed by gradually updating the model through online learning, or by accumulating a certain amount of deviation data through batch learning and then updating it centrally.

[0050] Understandably, continuous performance data collection provides the system with closed-loop feedback capabilities, and performance deviation reports quantify the gap between initial decisions and actual results. Fine-tuning enables the storage pattern decision network and storage policy generator to adapt to dynamic changes in cloud platform workload and resource conditions.

[0051] A cloud platform data storage system includes a memory, a processor, and a computer program stored in the memory and running on the processor. The memory is used to store the computer program and the generated data state tracking sequence, store feature vectors, and dynamically store evaluation parameters. When the processor executes the computer program, it executes instructions and coordinates the operation of data monitoring, feature parsing, decision network reasoning, resource evaluation, and policy generation modules.

[0052] In some embodiments, the memory includes non-volatile memory and volatile memory. The computer program is stored in the non-volatile memory and loaded into the volatile memory during runtime. The processor is a general-purpose central processing unit or a dedicated graphics processing unit, responsible for executing arithmetic and logical operations and control flow instructions in the computer program. The system communicates with other components of the cloud platform through a network interface, receives storage task instructions, and issues final storage policy instructions.

[0053] It is understood that the cloud platform data storage system instantiates the aforementioned method as an executable software component, and the collaborative work of memory and processor enables the automated generation and optimization of data storage strategies. The system's operation is independent of any specific hardware platform, possessing portability. Optionally, the system can be deployed as an independent microservice within the cloud platform, interacting with other services through well-defined application programming interfaces. The configuration of memory and processor can be horizontally scaled up or down according to the actual workload. It is understood that the process of a processor executing a computer program is a cyclical decision-execution-feedback process, and the feedback optimization steps ensure that the storage strategy continuously improves its adaptability and accuracy over time. The system enhances the efficiency and reliability of cloud platform data storage through integrated learning and optimization mechanisms.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for storing data on a cloud platform, characterized in that, Perform the following steps: When creating a data storage task on the cloud platform, start full lifecycle monitoring of the stored data object and generate a data status tracking sequence; Based on the data state tracking sequence, multi-dimensional feature analysis is performed on the stored data object to extract the stored feature vector; The storage feature vector is input into a pre-constructed storage pattern decision network, which outputs an initial storage pattern identifier. Based on the initial storage mode identifier, activate the corresponding storage resource evaluation process and generate dynamic storage evaluation parameters; By combining the dynamic storage evaluation parameters with the initial storage mode identifier, a final storage policy instruction is generated through a storage policy generator.

2. The cloud platform data storage method according to claim 1, characterized in that, The step of initiating full lifecycle monitoring of stored data objects and generating a data status tracking sequence includes: When a data write operation is initiated, the data creation timestamp, data source identity information, and initial data size are recorded. Continuously monitor the frequency distribution, access time patterns, and requesting user categories of data read operations; Periodically scan the data content and record the data block integrity check value and data update history; The data creation timestamp, data source identity information, initial data size, frequency distribution, access time pattern, requesting user category, data block integrity verification value, and data update history are integrated and encoded in chronological order to generate a data status tracking sequence.

3. The cloud platform data storage method according to claim 2, characterized in that, The step of performing multi-dimensional feature parsing on the stored data object and extracting the stored feature vector includes: The time dimension features are separated from the data state tracking sequence, and the data activity cycle and access interval standard deviation are calculated; Access pattern features are extracted from the data state tracking sequence to identify high-frequency access periods and low-frequency access periods; Data value features are extracted from the data status tracking sequence to assess the importance of data-related businesses and the intensity of data updates; The data activity cycle, access interval standard deviation, high-frequency access period, low-frequency access period, importance of data-related business, and data update intensity are normalized and concatenated to form a fixed-dimensional storage feature vector.

4. The cloud platform data storage method according to claim 3, characterized in that, The step of inputting the storage feature vector into a pre-constructed storage pattern decision network and outputting an initial storage pattern identifier includes: The storage mode decision network comprises multiple fully connected layers and a nonlinear activation function; The stored feature vector is input into the input layer of the storage pattern decision network, and then transformed layer by layer through the multiple fully connected layers and nonlinear activation functions; In the output layer of the storage mode decision network, the transformed features are mapped to a variety of predefined storage mode categories, and the storage mode category with the highest output probability is encoded as the initial storage mode identifier.

5. The cloud platform data storage method according to claim 4, characterized in that, The activation of the corresponding storage resource evaluation process generates dynamic storage evaluation parameters, including: Based on the initial storage mode identifier, retrieve the matching set of evaluation rules from the resource evaluation strategy library; Based on the aforementioned evaluation rule set, query in real time the remaining capacity, input / output performance indicators, and network bandwidth status of the cloud platform's currently available storage pool; By combining the access pattern features and data value features in the storage feature vector, the expected resource consumption is calculated; By combining the remaining capacity, input / output performance indicators, network bandwidth status, and expected resource consumption, dynamic storage evaluation parameters are generated.

6. The cloud platform data storage method according to claim 5, characterized in that, The dynamic storage evaluation parameters are generated by comprehensively considering the remaining capacity, input / output performance indicators, network bandwidth status, and expected resource consumption, including: The remaining capacity is standardized by converting the current remaining capacity into a percentage value based on the total capacity of the storage pool and the zero boundary. The input and output performance indicators are standardized, and based on the theoretical maximum throughput and zero boundary, the current input and output performance indicators are converted into performance indicator ratio values. The network bandwidth status is standardized by converting the current network bandwidth status into a bandwidth status ratio value based on the physical bandwidth upper limit and the zero value boundary. The expected resource consumption is standardized by converting the current expected resource consumption into a consumption ratio based on the historical maximum consumption value and the zero value boundary. Based on the initial storage mode identifier, obtain the corresponding remaining capacity weight, performance index weight, bandwidth status weight, and consumption weight from the weight configuration library; The remaining capacity ratio is weighted using the remaining capacity weight to obtain a weighted remaining capacity value; The performance index ratio value is weighted using the performance index weight to obtain a weighted performance index value; The bandwidth state ratio value is weighted using the bandwidth state weight to obtain a weighted bandwidth state value; The consumption ratio value is weighted using the consumption weight to obtain a weighted consumption value; The weighted remaining capacity value, weighted performance index value, weighted bandwidth status value, and weighted consumption value are added together to obtain a comprehensive evaluation value; The comprehensive evaluation value is used as a dynamically stored evaluation parameter.

7. The cloud platform data storage method according to claim 6, characterized in that, The process of fusing the dynamic storage evaluation parameters with the initial storage mode identifier and generating a final storage policy instruction through a storage policy generator includes: The dynamic storage evaluation parameters are concatenated with the initial storage mode identifier to form a strategy decision input vector; The policy decision input vector is input to the storage policy generator, which is a trained deep learning model. The storage strategy generator outputs multiple strategy parameters, including target storage medium type, data redundancy backup strategy, data fragment size, and data compression algorithm selection. The aforementioned multiple policy parameters are encapsulated into a final storage policy instruction that can be parsed by the cloud platform storage engine.

8. The cloud platform data storage method according to claim 7, characterized in that, The training process for the storage policy generator includes: Collect historical cloud platform data storage tasks and their corresponding storage feature vectors, storage mode identifiers, resource evaluation parameters, and manually labeled optimal storage strategies; The deep learning model is constructed by taking a combination of storage feature vectors, storage mode identifiers and resource evaluation parameters as input, and the optimal storage strategy as the desired output. The parameters of the deep learning model are optimized using the gradient descent algorithm until the error between the model's prediction and the optimal storage strategy is lower than a preset threshold.

9. The cloud platform data storage method according to claim 1, characterized in that, The method also includes a feedback optimization step after executing the storage policy instruction: After the final storage strategy instruction is issued to the cloud platform storage engine, actual storage performance data is continuously collected. The actual storage performance data is compared with the expected indicators in the dynamic storage evaluation parameters to generate a performance deviation report; Based on the performance deviation report, the parameters of the storage mode decision network and the storage policy generator are fine-tuned.

10. A cloud platform data storage system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the cloud platform data storage method according to any one of claims 1 to 9.

Images