Method, device and system for operation control of cold data storage system

Abstract

The application provides a cold data storage system operation control method, device and system, and belongs to the technical field of storage device management, and solves the problems of invalid power consumption of a hard disk and lack of intelligent power supply management in cold data storage. The method comprises the following steps: obtaining operation data of a storage system; preprocessing the operation data to obtain a structured data set; performing format processing on the structured data set to obtain a time series data set; inputting the time series data set into a prediction model for processing to obtain storage system state data; and determining the storage system operation data according to the storage system state data. The scheme realizes intelligent management of power supply of a cold data storage system, saves energy and reduces the failure rate.

Description

Technical Field

[0001] This invention relates to the field of storage device management technology, and in particular to an operation control method, apparatus and system for a cold data storage system. Background Technology

[0002] In current cold data storage systems, hard drives generally use continuous power supply mode, resulting in a large number of idle hard drives still consuming 60% to 70% of peak power. Existing technologies mainly manage power supply through timed switching or simple hibernation strategies, which have significant drawbacks: First, they cannot dynamically adjust power supply based on data access frequency; for example, there is no differentiated power supply between cold data hard drives (access frequency ≤ 2 times / cycle) and active hard drives. Second, hard drives maintain a high power consumption state even when idle, resulting in energy waste. Finally, there is a lack of real-time monitoring and early warning mechanisms for hard drive health parameters (temperature, vibration, current), leading to a persistently high failure rate.

[0003] Existing solutions such as Hierarchical Storage (HSM) only migrate data but do not optimize power supply, while traditional power management modules, lacking predictive capabilities, suffer from response latency in the minutes range, failing to meet real-time access requirements. Therefore, there is an urgent need for an operation control method that integrates intelligent prediction and dynamic power adjustment to solve energy efficiency and reliability issues. Summary of the Invention

[0004] This invention provides an operation control method, device, and system for a cold data storage system, which solves the problems of ineffective power consumption of hard drives and lack of intelligent power supply management in cold data storage.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] This invention provides an operation control method for a cold data storage system, comprising:

[0007] Obtain operational data from the storage system;

[0008] The running data is preprocessed to obtain a structured dataset;

[0009] The structured dataset is formatted to obtain a time-series dataset;

[0010] The time-series dataset is input into the prediction model for processing to obtain the storage system state data;

[0011] Based on the storage system status data, determine the storage system operating data.

[0012] Optionally, obtain the storage system's operational data, including:

[0013] Based on sensors installed in the cold data storage system, operational data of the storage system is acquired, including at least one of hard disk temperature, current, power, vibration, and access frequency.

[0014] Optionally, the runtime data is preprocessed to obtain a structured dataset, including:

[0015] The missing value imputation process is performed on the running data to obtain the imputed running data;

[0016] The filled-in running data is subjected to outlier detection processing to obtain outlier-detected running data;

[0017] The deduplication process is performed on the operational data after the anomaly detection to obtain the deduplicated operational data;

[0018] The deduplicated running data is then subjected to noise reduction processing to obtain a structured dataset.

[0019] Optionally, the structured dataset is formatted to obtain a time-series dataset, including:

[0020] The structured dataset is processed to unify its types, resulting in data with a unified type.

[0021] The data of the uniform type are subjected to feature scaling to obtain feature-scaled data.

[0022] The scaled data is reconstructed in chronological order to obtain time-series reconstructed data.

[0023] The reconstructed time series data is encoded to obtain a time series dataset.

[0024] Optionally, the time-series dataset is input into a prediction model for processing to obtain stored system state data, including:

[0025] The time-series dataset is input into the prediction model for processing to obtain storage system state data. The prediction model obtains prediction results of multiple trees by independently predicting the time-series dataset using multiple decision trees, and integrates the prediction results of the multiple trees by majority voting to obtain storage system state data.

[0026] Optionally, the prediction model is obtained through the following training process:

[0027] Retrieve historical operational data from the storage system;

[0028] The historical operational data is preprocessed to obtain a historical structured dataset;

[0029] The historical structured dataset is formatted to obtain the sample dataset;

[0030] The sample dataset is processed to obtain a training dataset and a test sample set;

[0031] Input the training dataset into the prediction model to obtain the training result data;

[0032] The training results data are compared with the test sample set, and the kernel function, penalty parameter, and kernel function parameter are adjusted according to the comparison results to obtain the prediction model.

[0033] Optionally, based on the storage system status data, the storage system operating data is determined, including:

[0034] The storage system status data is compared with a preset value to obtain a comparison result;

[0035] Based on the comparison results, the control signal is determined;

[0036] The operating state of the storage system is controlled according to the control signal.

[0037] This invention also provides an operation control device for a cold data storage system, comprising:

[0038] The acquisition module is used to acquire the operating data of the storage system;

[0039] The processing module is used to preprocess the running data to obtain a structured dataset; format the structured dataset to obtain a time-series dataset; and input the time-series dataset into a prediction model for processing to obtain storage system state data.

[0040] The determination module is used to determine the storage system operating data based on the storage system status data.

[0041] This invention also provides an operation control system for a cold data storage system, comprising: a backplane, a controller, and a sensor module. The controller is electrically connected to the backplane and the sensor module, respectively. The backplane is provided with multiple hard disk interfaces, which are electrically connected to multiple hard disks. When the operation control system of the cold data storage system is running, it executes the above-described method.

[0042] This invention also provides a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when run by the processor, executes the above-described method.

[0043] The technical solution of the present invention has at least the following effects:

[0044] The above-mentioned solution of the present invention obtains the operating data of the storage system; preprocesses the operating data to obtain a structured dataset; formats the structured dataset to obtain a time-series dataset; inputs the time-series dataset into a prediction model for processing to obtain storage system status data; and determines the storage system operating data based on the storage system status data, thereby realizing intelligent power supply management for cold data hard drives, saving energy and reducing the failure rate. Attached Figure Description

[0045] Figure 1 This is a flowchart of the operation control method of the cold data storage system provided in the embodiments of the present invention;

[0046] Figure 2 This is a structural diagram of the operation control device of the cold data storage system provided in an embodiment of the present invention;

[0047] Figure 3 This is a structural diagram of the operation control system of the cold data storage system provided in the embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of the structure of the computing device provided in an embodiment of the present invention. Detailed Implementation

[0049] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0050] like Figure 1 As shown, an embodiment of the present invention proposes an operation control method for a cold data storage system, comprising:

[0051] Step 11: Obtain the operating data of the storage system;

[0052] Step 12: Preprocess the running data to obtain a structured dataset;

[0053] Step 13: Format the structured dataset to obtain a time-series dataset;

[0054] Step 14: Input the time series dataset into the prediction model for processing to obtain the storage system state data;

[0055] Step 15: Determine the storage system operating data based on the storage system status data.

[0056] In step 11 of this embodiment, the operational data mainly comes from high-precision sensor modules installed on the hard drive. These sensors are responsible for real-time monitoring of various operating status parameters of the hard drive, including but not limited to key parameters such as temperature, current, power, and vibration. The sensor modules continuously collect data, and a timing module is set to read data at fixed time intervals (e.g., per second) to ensure high-frequency data acquisition, thereby capturing subtle changes in the hard drive's status.

[0057] In step 12, the running data is preprocessed, specifically including: missing value processing, outlier detection, redundant data processing, and data integration, thereby integrating multi-source heterogeneous data from different sensors to form a structured dataset in a unified format, which facilitates subsequent processing and analysis.

[0058] In step 13, the various data types in the structured dataset are uniformly converted into a consistent data type (such as integers, floating-point numbers, dates, etc.) to ensure the accuracy and efficiency of subsequent processing. Specifically, this includes: feature scaling, which standardizes numerical features to convert them into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating dimensional differences between different features; time series construction, which uses windowing techniques to divide the original time series data into multiple time window samples, each containing data points within a certain time range to meet the processing requirements of time series models; and label encoding and one-hot encoding, which performs label encoding or one-hot encoding on categorical features, converting them into a numerical form understandable by machine learning models.

[0059] In step 14, the Random Forest algorithm is used as the main prediction model. This model improves the accuracy and robustness of predictions by constructing multiple decision trees and employing Bagging ensemble techniques. During model training, the Random Forest model is trained using a high-quality dataset, and hyperparameters are tuned to improve model performance. Cross-validation is used during training to evaluate the model's accuracy and generalization ability. The formatted time-series dataset is input into the trained Random Forest model for processing. The model dynamically predicts the future state of the storage system based on historical and real-time data, including hard drive access patterns and power consumption changes.

[0060] In step 15, the storage system status data output by the predictive model is analyzed in depth to identify the access patterns of active and inactive hard drives. Based on the status data analysis results, the power supply strategy for each hard drive is dynamically adjusted. For inactive hard drives, their power supply is switched to a low-power mode to reduce energy consumption; for active hard drives, their full power consumption state is maintained to ensure optimal performance. A real-time feedback mechanism is established to ensure that the system can adapt and update the power supply strategy in a timely manner when the hard drive status changes (such as changes in environment or usage frequency), thereby maintaining the overall performance and energy efficiency of the system. At the same time, the predictive model is periodically retrained and optimized based on feedback information to improve its accuracy and adaptability.

[0061] The above embodiments of the present invention realize intelligent power supply management for cold data hard drives, saving energy and reducing the failure rate.

[0062] In an optional embodiment of the present invention, step 11 may include:

[0063] Step 111: Obtain the operating data of the storage system based on the sensors installed in the cold data storage system. The operating data includes at least one of hard disk temperature, current, power, vibration, and access frequency.

[0064] In this embodiment, various types of sensors are installed in the cold data storage system to comprehensively monitor the operating status of the hard drive. These sensors include, but are not limited to, temperature sensors, current sensors, power sensors, vibration sensors, and access frequency monitoring modules, wherein:

[0065] (1) The temperature sensor is used to monitor the operating temperature of the hard drive in real time to ensure that it is within a safe operating temperature range and to prevent damage to the hard drive due to overheating.

[0066] (2) Current sensors are used to monitor the current consumption of hard drives, assess their energy efficiency, and detect current fluctuations, which helps to identify potential power supply problems.

[0067] (3) Power sensors are used to directly measure the power consumption of hard drives, providing direct data support for energy efficiency analysis.

[0068] (4) Vibration sensors are used to detect the vibration of the hard drive when it is working. Abnormal vibration may indicate mechanical problems or improper installation of the hard drive, which can help to provide early warning of failure.

[0069] (5) The access frequency monitoring module is used to record the hard drive access frequency through software or hardware, distinguish between active data and cold data, and provide a basis for power supply strategy adjustment.

[0070] The sensor module continuously collects various operational data from the hard drive, and a timing module is set to read the data at fixed time intervals (e.g., per second). This high-frequency data acquisition ensures the real-time nature and accuracy of the data, enabling the capture of subtle changes in the hard drive's status.

[0071] The acquired data is transmitted via a high-speed data bus (such as I...) 2 The data is transmitted via the C-bus to the processing unit of the PowerManagement AI controller, ensuring both speed and security of data transmission.

[0072] After initial integration, the raw data collected by the sensor module is transmitted to Power Management AI for further processing. The integration process includes data alignment and timestamp synchronization to ensure data integrity and consistency.

[0073] In an optional embodiment of the present invention, step 12 may include:

[0074] Step 121: Perform missing value imputation on the running data to obtain the imputed running data;

[0075] Step 122: Perform outlier detection processing on the filled-in running data to obtain outlier-detected running data;

[0076] Step 123: Perform deduplication on the running data after anomaly detection to obtain deduplicated running data;

[0077] Step 124: Perform noise reduction processing on the deduplicated running data to obtain a structured dataset.

[0078] In step 121 of this embodiment, the collected operational data is first examined to identify data records with missing values. Missing values ​​may be caused by sensor malfunctions, data transmission errors, or mismatched sampling intervals. An appropriate imputation strategy is selected based on the proportion and distribution of the missing values. For a small number of missing values, methods such as mean imputation, median imputation, or forward imputation (filling with the previous valid value) can be used. For a large number of consecutive missing values, more complex interpolation methods or model predictions need to be considered. The selected imputation strategy is applied to impute the missing values, generating a complete imputed dataset. The imputed dataset should retain the distribution characteristics and statistical properties of the original data as much as possible.

[0079] In step 122, outliers refer to data points that deviate significantly from the overall data distribution due to reasons such as sensor failure, data recording errors, or extreme events.

[0080] Outliers can be identified using the Z-score method. The Z-score method identifies outliers by calculating the standard deviation distance between a data point and the mean, and is suitable for data that is approximately normally distributed. For detected outliers, the appropriate options are to delete, replace, or retain and mark them, depending on the specific circumstances. If the outlier is due to a measurement error, it should be deleted directly; if the outlier contains useful information (such as extreme events), it can be considered for replacement with a reasonable value or retention and marking for subsequent analysis.

[0081] In steps 123, check if there are identical records or record pairs in the dataset. These duplicates are caused by data transmission errors, excessively small sampling intervals, or data merging. Based on the proportion and distribution of duplicate data, select an appropriate deduplication strategy. For identical records, duplicates can be directly deleted; for records with some identical fields, you can choose to retain the latest record, the earliest record, or merge records, depending on business needs. Apply the selected deduplication strategy to process the duplicate data, generating a deduplicated dataset. The deduplicated dataset should ensure the uniqueness of each record to avoid introducing bias in subsequent analysis.

[0082] In step 124, noise refers to the portion of the dataset that is unrelated to the overall trend or fluctuates randomly, caused by factors such as sensor accuracy limitations, environmental interference, or data transmission errors. Methods such as random forest algorithms, wavelet transforms, or moving averages can be used to identify and remove noise features. Random forest algorithms reduce the impact of noise by constructing multiple decision trees and combining their prediction results; wavelet transforms decompose the signal into different frequency components, removing high-frequency noise; and moving averages smooth data fluctuations by calculating the average value of local data. The selected denoising method is applied to process the deduplicated data to generate a structured dataset. A structured dataset has a clear data structure and consistent statistical properties, facilitating subsequent data analysis and modeling. Simultaneously, the denoising process should preserve as much of the original information and feature distribution of the data as possible.

[0083] In an optional embodiment of the present invention, step 13 may include:

[0084] Step 131: Perform type unification processing on the structured dataset to obtain data with unified type;

[0085] Step 132: Perform feature scaling on the data of the uniform type to obtain feature-scaled data;

[0086] Step 133: Reconstruct the scaled data according to time order to obtain time series reconstructed data;

[0087] Step 134: Encode the reconstructed time series data to obtain a time series dataset.

[0088] In step 131 of this embodiment, the data type of each field in the structured dataset is first checked to identify different data types, such as numeric, character, and date / time. Based on subsequent analysis needs, data type conversion rules are formulated. For example, character-type date / time fields are uniformly converted to date / time type; potentially mixed-type fields (such as those with some numeric and some character data) are uniformly converted to either numeric or character type to ensure data type consistency. The conversion rules are then applied to convert the type of each field in the structured dataset, generating a dataset with a unified data type. The converted dataset should eliminate analysis errors caused by inconsistent data types, improving the accuracy and efficiency of data processing.

[0089] In step 132, the range of each numerical feature in the unified dataset is analyzed to identify features with dimensional differences or excessively large numerical ranges. These features, caused by variations in sensor accuracy and measurement units, may lead to unstable model training or biased results if directly used in subsequent analysis. An appropriate scaling method is selected based on the feature distribution, such as min-max scaling or standardization. Min-max scaling scales feature values ​​to the [0, 1] interval, suitable for cases where feature value distribution is relatively uniform; standardization transforms feature values ​​into a distribution with a mean of 0 and a standard deviation of 1, suitable for cases where feature values ​​have a large degree of dispersion. The selected scaling method is applied to scale the numerical features, generating a scaled dataset. The scaled dataset should eliminate the influence of dimensional differences and excessively large numerical ranges, improving the stability and accuracy of model training.

[0090] In step 133, the time field is identified in the feature-scaled dataset. This field should uniquely identify the time point or time period of each record. The dataset is then sorted according to the time field, ensuring that each record is arranged in chronological order. Within the sorted dataset, the data is reconstructed based on the analytical requirements. For example, data from consecutive time points can be combined into time series segments for subsequent time series prediction or pattern recognition; alternatively, data from different time scales can be aggregated to generate higher-level time series data. The reconstructed dataset should have a clear time series structure to facilitate subsequent time series analysis.

[0091] In step 134, non-numerical features or features requiring special processing, such as categorical variables and text labels, are analyzed in the time series reconstruction dataset. These features cannot be directly used for time series analysis or model training and require encoding. Appropriate encoding methods are selected based on the feature type, such as one-hot encoding, label encoding, or embedding encoding. One-hot encoding is suitable for situations with few categorical variables, converting each categorical variable into a binary vector; label encoding converts categorical variables into integer labels, suitable for situations with many categorical variables but unordered categories; embedding encoding learns to map categorical variables to vector representations in a low-dimensional space, suitable for situations where semantic relationships between categorical variables need to be preserved. The selected encoding method is applied to encode the non-numerical features, generating the time series dataset. The encoded time series dataset should eliminate the impact of non-numerical features on time series analysis and model training, improving data usability and analytical accuracy. Simultaneously, the encoding process should preserve as much information and distribution characteristics of the original features as possible.

[0092] In an optional embodiment of the present invention, step 14 may include:

[0093] Step 141: Input the time series dataset into the prediction model for processing to obtain storage system state data. The prediction model obtains the prediction results of multiple trees by independently predicting the time series dataset using multiple decision trees, and integrates the prediction results of the multiple trees by majority voting to obtain storage system state data.

[0094] In step 141 of this embodiment, the prediction model adopts an ensemble learning approach based on multiple decision trees. Decision trees, as a commonly used machine learning algorithm, can construct a tree structure for prediction by progressively dividing data features. Each decision tree independently learns and predicts the input time-series dataset, generating a prediction result. Ensemble learning, by combining the prediction results of multiple decision trees, can effectively reduce the overfitting or underfitting problems of a single decision tree, improving the overall prediction accuracy and robustness. In this embodiment, the majority voting ensemble method is used, that is, the category that appears most frequently in the prediction results of multiple decision trees is taken as the final prediction result. This method is suitable for prediction scenarios of classification problems and can accurately determine the state category of the storage system. The specific processing flow includes:

[0095] (1) Time window slicing;

[0096] Time series dataset Data is segmented according to a preset window size; the time-series dataset contains standardized features of the hard drive within the time window, such as temperature, current, vibration, and access frequency. The data structure can be represented as follows:

[0097]

[0098] in, For standardized time series datasets; For the first m 1 eigenvector;

[0099] Window length W =10min, generate window samples:

[0100]

[0101] in, For window samples, ; m The number of features; W The length of the window;

[0102] (2) Prediction of a single tree;

[0103] In a random forest, each decision tree makes independent predictions, randomly selects a subset of features, and generates predicted labels based on the CART algorithm. The number of features used for each tree is:

[0104]

[0105] in, The number of features used for each tree, The proportion of features considered for each tree; It is a floor function;

[0106] (3) Statistically analyze the prediction results of all trees;

[0107] All tree prediction results can be represented as:

[0108]

[0109] in, For the first n The predicted results for each tree;

[0110] The prediction results of all decision trees in the random forest (total) n (trees)

[0111] For each tree's predicted label, determine: if the tree's output label is cold data, then the counter... Increment by 1; if the tree output label is active, the counter remains unchanged. The final total number of votes for the cold data is obtained:

[0112]

[0113] Divide the number of votes for cold data by the total number of trees: ;

[0114] in, The percentage of trees that support cold data labels;

[0115] like If the value is ≥0.5 (meaning ≥50% of the trees support cold data), then the final label is determined to be cold data. ;

[0116] in, For final tags;

[0117] like If the value is less than 0.5 (meaning less than 50% of the trees support cold data), then the final label is determined to be active. ;

[0118] (4) Threshold determination;

[0119] like If the actual access frequency is less than or equal to cold_threshold, then the data is considered cold data. ;

[0120] like If the access frequency is greater than or equal to active_threshold, then the device is considered active. ;

[0121] Where cold_threshold is the maximum number of accesses to cold data; active_threshold is the minimum number of accesses to active hard drives; To store system status data.

[0122] Through the above steps, the time-series dataset is input into the prediction model for processing. By leveraging the ensemble learning advantages of multiple decision trees, accurate and reliable storage system status data is finally obtained, providing strong data support for the operation and control of cold data storage systems.

[0123] In an optional embodiment of the present invention, the prediction model is obtained through the following training process:

[0124] Step 1411: Obtain historical operating data of the storage system;

[0125] Step 1412: Preprocess the historical operation data to obtain a historical structured dataset;

[0126] Step 1413: Format the historical structured dataset to obtain the sample dataset;

[0127] Step 1414: Extract the sample dataset to obtain the training dataset and the test sample set;

[0128] Step 1415: Input the training dataset into the prediction model to obtain the training result data;

[0129] Step 1416: Compare the training result data with the test sample set, and adjust the kernel function, penalty parameter and kernel function parameter according to the comparison result to obtain the prediction model.

[0130] In step 1411 of this embodiment, the source of historical operational data is first identified, including monitoring modules, log files, and related management databases within the storage system. During long-term operation, these data sources continuously record various aspects of system operation information, such as storage device temperature, load, read / write operation frequency, disk usage, and power consumption. Automated tools or scripts are used to periodically collect historical operational data from these data sources. For example, by writing scheduled tasks, utilizing system-provided API interfaces or directly reading log files, the required data is stored in a designated data warehouse according to a unified format. Simultaneously, the stability and integrity of the data collection process are ensured to prevent data loss or corruption, thereby guaranteeing the accuracy and reliability of subsequent training.

[0131] In step 1412, check for missing values ​​in the historical operating data. For a small number of missing values, methods such as mean imputation, median imputation, or forward imputation can be used to fill in the missing values. For example, if the temperature data of the storage device is missing within a certain time period, it can be imputed by averaging the temperature values ​​of adjacent time points. For a large number of consecutive missing values, the cause of the missing values ​​needs to be analyzed. If it is due to unrecoverable situations such as data acquisition failure, the data record can be deleted to ensure the integrity and accuracy of the data. Statistical methods or machine learning algorithms are used to detect outliers in the historical operating data. For example, the Z-score method is used to calculate the standard deviation distance between the data point and the mean. When the distance exceeds a certain threshold, it is judged as an outlier. For the detected outliers, their rationality needs to be judged in conjunction with business knowledge. If it is due to measurement errors or data entry errors, it can be deleted or corrected. If the outlier reflects the extreme operating state of the system, it can be retained and marked for subsequent analysis and processing. Check for duplicate records in the dataset, which may be caused by duplicate storage during data acquisition or transmission. By comparing the values ​​of each field in the records, duplicate data is identified and removed to ensure the uniqueness of each record and avoid interference with the model during subsequent training. Random forest, wavelet transform, or moving average methods are used to denoise the historical data. The random forest algorithm reduces noise by constructing multiple decision trees and combining their predictions; wavelet transform decomposes the signal into different frequency components, eliminating high-frequency noise; and moving average smooths data fluctuations by calculating the average value of local data. After denoising, the resulting structured historical dataset more clearly reflects the inherent patterns of the storage system's operation.

[0132] In step 1413, the data type of each field in the historical structured dataset is checked. Character-type date and time fields are uniformly converted to date and time types, and mixed-type fields (such as those with some numeric and some character elements) are uniformly converted to either numeric or character types to ensure data type consistency for subsequent data processing and analysis. The range of numeric features is analyzed. For features with differences in units or excessively large numerical ranges, scaling methods such as min-max scaling or standardization are used. Min-max scaling scales feature values ​​to the [0, 1] interval, suitable for cases where feature value distribution is relatively uniform; standardization converts feature values ​​to a distribution with a mean of 0 and a standard deviation of 1, suitable for cases where feature values ​​have a large degree of dispersion. Feature scaling eliminates the influence of units between different features, improving the stability and accuracy of model training. The time field in the historical structured dataset is identified, and the dataset is sorted according to the time field to ensure that each record is arranged in chronological order. Then, the data is reconstructed according to the analysis requirements. For example, data from consecutive time points are combined into time series segments, or data from different time scales are aggregated to generate higher-level time series data. The reconstructed dataset has a clear time series structure, which facilitates subsequent time series analysis and model training. For non-numerical features or features requiring special processing in the dataset, such as categorical variables and text labels, one-hot encoding, label encoding, or embedding encoding are used. One-hot encoding is suitable for cases with few categorical variables, converting each categorical variable into a binary vector; label encoding converts categorical variables into integer labels, suitable for cases with many categorical variables but no order between categories; embedding encoding learns to map categorical variables to vector representations in a low-dimensional space, suitable for cases where semantic relationships between categorical variables need to be preserved. The encoded sample dataset eliminates the influence of non-numerical features on model training, improving data usability and analytical accuracy.

[0133] In step 1414, the ratio of the training dataset to the test sample set is determined based on the dataset size and model training requirements. Generally, the training dataset accounts for 70% to 80%, and the test sample set accounts for 20% to 30%, ensuring that the model has sufficient data for learning and that model performance can be effectively evaluated using the test sample set. Random sampling is used to extract data from the sample dataset, ensuring that the distribution of the training and test sample sets is consistent with the original sample dataset, avoiding model training bias or inaccurate evaluation due to unreasonable data partitioning. For example, random sampling can be implemented using the `random` library in Python or the `train_test_split` function in the `scikit-learn` library.

[0134] In step 1415, a suitable prediction model architecture is selected, such as an ensemble learning model based on multiple decision trees (e.g., random forest). The model parameters are initialized, including the kernel function type, the number of decision trees, and the tree depth. These initial parameters can be set based on empirical values ​​or simple experiments, and subsequent optimization will further improve model performance. The training dataset is input into the initialized prediction model, which learns and trains based on the features and labels of the input data. During training, the model continuously adjusts its internal parameters to minimize prediction error. For example, in a random forest model, each decision tree independently learns and predicts on the training data, and the prediction accuracy is improved by combining the results of multiple decision trees; in a gradient boosting tree model, prediction error is gradually reduced by iteratively building new decision trees. After training is complete, the training result data is obtained, including the model's predictions on the training data and the corresponding error metrics.

[0135] In step 1416, the trained model is used to predict on the test sample set to obtain the test prediction results. The test prediction results are compared with the true labels of the test sample set, and the model's performance metrics, such as accuracy, recall, F1 score, and mean squared error (MSE), are calculated to evaluate the model's generalization ability on unknown data. Based on the performance evaluation results, the model's problems and shortcomings are analyzed. If the model exhibits underfitting, consider increasing the number of decision trees, adjusting the tree depth, or optimizing the kernel function parameters; if the model exhibits overfitting, appropriately increase the penalty parameter, reduce the number of decision trees, or perform regularization. By continuously adjusting the kernel function, penalty parameter, and kernel function parameters, the model is retrained and its performance is evaluated until the model achieves satisfactory performance metrics on the test sample set. After multiple parameter adjustments and model training, when the model's performance on the test sample set stabilizes and meets the expected requirements, the final prediction model is determined. This model can learn the operating patterns of the storage system from historical operating data, accurately predict the storage system state, and provide reliable decision support for the operation and control of the cold data storage system.

[0136] In an optional embodiment of the present invention, step 15 may include:

[0137] Step 151: Compare the storage system status data with preset values ​​to obtain comparison results;

[0138] Step 152: Determine the control signal based on the comparison results;

[0139] Step 153: Control the operating state of the storage system according to the control signal.

[0140] In step 151 of this embodiment, the storage system status data is compared with a preset value to determine whether the hard drive is currently active. If the hard drive does not receive an access request within a predetermined time, the system will automatically mark it as "idle".

[0141] In steps 152 and 153, when the hard drive is identified as being in an "idle" state, the system determines a control signal based on the comparison result and immediately switches the power supply of the hard drive to a low-power mode to reduce current and power consumption and extend the lifespan of the hard drive. When an access request (such as reading or writing data) is detected for the corresponding hard drive, the system responds quickly and immediately switches the hard drive to a full-power state to ensure that the hard drive operates at optimal performance.

[0142] A specific embodiment of the operation control method for the cold data storage system provided in this invention is as follows:

[0143] Step 1, Data Preparation. Real-time acquisition of raw monitoring parameters such as hard drive temperature, current, power, vibration, and access frequency, with a sampling period of... T =30s;

[0144] (1) Missing value handling: Remove records with a large number of missing values ​​and handle them according to `deletion_threshold`. Imputation is performed using the mean, or other missing values ​​are filled using the forward imputation method.

[0145] (2) Outlier detection: Outliers are detected using the Z-score method and marked according to the set "z_score_threshold". The preliminary model obtained from training is used to predict the data and compare it with the actual values. Points with high residuals are marked as outliers.

[0146] (3) Duplicate data processing: The dataset is deduplicated to ensure that each record appears only once, and this is checked by a unique identifier.

[0147] (4) Noise filtering: The random forest algorithm is used to identify and remove noise features with low weights to improve the accuracy of the model.

[0148] Step 2, data formatting.

[0149] (1) Unified data type: Convert the collected data in various formats into a consistent data type (such as integer, floating point, date, etc.) to ensure that subsequent processing is error-free.

[0150] (2) Feature scaling: Standardize the data to ensure that the data is converted into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0151] (3) Time series construction: Use windowing techniques to convert the original time series data into multiple time window samples to adapt to the time series model.

[0152] (4) Label encoding and one-hot encoding: Label encoding and one-hot encoding are performed on the classification features to convert the category data into a numerical form that can be understood by the machine learning model.

[0153] Step 3, model training.

[0154] (1) Application of Random Forest Algorithm: Random forest algorithm is used as the main machine learning method to optimize power management for cold data storage of hard disk.

[0155] (2) Overview of Random Forest Algorithm: Random forest improves the accuracy and robustness of the model by constructing multiple decision trees and employing the Bagging (Bootstrap Aggregating) technique. Its steps are as follows:

[0156] 1) Dataset Sampling: Multiple subsets of samples are randomly selected from the original training dataset using the bootstrap method. Each decision tree is trained from a different subset of data.

[0157] 2) Feature selection: When splitting at a node of each decision tree, a subset of features is randomly selected to reduce the risk of overfitting.

[0158] 3) Constructing a decision tree: Use basic decision tree algorithms such as ID3, C4.5, or CART to split the tree, without pruning during training.

[0159] 4) Ensemble prediction: Vote on all constructed decision trees and select the category with the most votes as the final prediction result, or calculate the average of all trees for the regression problem.

[0160] (3) Advantages of Random Forest:

[0161] 1) High-dimensional data processing capability: It performs well when processing high-dimensional features, automatically selecting the features most relevant to the target variable and avoiding the influence of redundant features.

[0162] 2) Noise resistance: The ensemble effect of multiple trees makes the algorithm more robust to noise and outliers in the data.

[0163] 3) Interpretability: Provides feature importance scores to help users understand the features the model relies on.

[0164] (4) Integration and application in Power Management AI:

[0165] 1) Data cleaning and post-processing: Ensure the quality of input data, including missing value handling, outlier detection, and redundant data removal, which directly affects the training efficiency and prediction accuracy of the random forest model.

[0166] 2) Data formatting: Ensure that the input features are provided to the random forest algorithm in a suitable form for feature labeling or one-hot encoding.

[0167] 3) Model training: Train the random forest model using a high-quality dataset, perform hyperparameter tuning to improve model performance; use cross-validation (set "cv_folds") to evaluate the model's accuracy and generalization ability.

[0168] 4) Real-time prediction: The trained random forest model is used to make dynamic predictions on real-time monitoring data; PowerManagement AI adjusts the power supply mode of the hard drive based on the prediction results, switching inactive hard drives to a low-power state while maintaining the optimal power supply for active hard drives.

[0169] 5) Feedback and retraining: Monitor the prediction results and actual system performance in real time, collect feedback data, and retrain and optimize the random forest model regularly based on the feedback information to improve the accuracy and adaptability of the model.

[0170] Step 4, Dynamic power regulation module.

[0171] The Dynamic Power Conditioning module is a crucial component of Power Management AI. Its purpose is to optimize power management for cold data storage hard drives based on real-time monitoring data and historical access patterns. Through intelligent power supply decisions, the system can effectively improve energy efficiency, reduce operating costs, and ensure system reliability.

[0172] 4.1 Intelligent Power Supply Decision

[0173] Intelligent power supply decision-making focuses on ensuring that the hard drive dynamically adjusts its power supply status according to demand. The steps to implement this decision are as follows:

[0174] (1) Real-time status monitoring:

[0175] 1) Data reception: Power Management AI receives real-time data from the intelligent monitoring module, including information such as hard drive temperature, power consumption, operating status, and access requests.

[0176] 2) Status determination: The system determines whether the hard drive is currently active. If the hard drive does not receive an access request within a predetermined time, the system will automatically mark it as "idle".

[0177] (2) Dynamically adjust power supply:

[0178] 1) Low-power mode switching: When the hard drive is identified as "idle", Power Management AI will immediately switch the power supply of the hard drive to a low-power mode to reduce current and power consumption and extend the life of the hard drive.

[0179] 2) Restore full power state: Once Power Management AI detects an access request (such as reading or writing data) for the corresponding hard drive, the system will respond quickly and immediately switch the hard drive to full power state to ensure that the hard drive operates at its best performance.

[0180] (3) Response time optimization:

[0181] 1) Response time monitoring: Power Management AI must accurately monitor the current status of the hard drive and maintain a small latency to quickly adjust the power supply status when a user request occurs.

[0182] 2) Preprocessing mechanism: To reduce response time, the system can pre-power the hard drive that is about to be accessed, ensuring the smooth completion of user requests.

[0183] 4.2 Historical Data Analysis

[0184] The purpose of historical data analysis is to improve the accuracy of smart power supply decisions, help identify access patterns for cold data storage, and optimize power supply strategies. The steps involved in this analysis include:

[0185] (1) Historical data collection:

[0186] 1) Data aggregation: Power Management AI aggregates hard drive access records, temperature data, and power consumption data over a period of time (such as several weeks or months) to form a dataset.

[0187] 2) Layered data storage: Data is stored in layers to ensure that data at each layer is classified according to frequency and importance, so as to facilitate fast access and processing.

[0188] (2) Access pattern recognition:

[0189] 1) Pattern recognition algorithm: Use clustering algorithms (such as K-means, hierarchical clustering) to analyze historical data and identify access patterns between active hard drives and cold data.

[0190] 2) Trend Analysis: Use time series analysis tools (such as ARIMA, LSTM) to predict access trends of cold data and update power supply strategies in a timely manner.

[0191] (3) Optimize power supply strategy:

[0192] 1) Rule-based power supply strategy: Based on the identified access patterns, a set of power supply rules for cold data is constructed, and the hard drive is automatically adjusted to a low-power mode.

[0193] 2) Customized power management: Allows users to set access expectations for specific time periods, and Power Management AI will adjust itself based on these settings.

[0194] The operation control method for the cold data storage system proposed in this invention achieves the following technical effects by integrating real-time sensor monitoring, predictive models, and dynamic power control modules:

[0195] (1) Energy efficiency optimization: Based on hard drive access frequency prediction, idle hard drives are dynamically switched to low power mode, thus achieving energy saving;

[0196] (2) Improved lifespan and reliability: Reduced power supply duration extends hard drive lifespan; Real-time monitoring of parameters such as temperature, current, and vibration, along with linked fault warnings, reduces failure rate;

[0197] (3) Enhanced system stability: The model is continuously optimized by adopting a feedback mechanism, combined with a rapid isolation strategy for faulty hard drives, to ensure data security and reduce operation and maintenance costs.

[0198] like Figure 2 As shown, this embodiment of the invention also provides an operation control device 20 for a cold data storage system, comprising:

[0199] Module 21 is used to acquire the operating data of the storage system;

[0200] Processing module 22 is used to preprocess the running data to obtain a structured dataset; format the structured dataset to obtain a time series dataset; and input the time series dataset into a prediction model for processing to obtain storage system state data.

[0201] The determination module 23 is used to determine the storage system operating data based on the storage system status data.

[0202] Optionally, module 21 is specifically used for:

[0203] Based on sensors installed in the cold data storage system, operational data of the storage system is acquired, including at least one of hard disk temperature, current, power, vibration, and access frequency.

[0204] Optionally, processing module 22 is specifically used for:

[0205] The missing value imputation process is performed on the running data to obtain the imputed running data;

[0206] The filled-in running data is subjected to outlier detection processing to obtain outlier-detected running data;

[0207] The deduplication process is performed on the operational data after the anomaly detection to obtain the deduplicated operational data;

[0208] The deduplicated running data is then subjected to noise reduction processing to obtain a structured dataset.

[0209] Optionally, the processing module 22 is also specifically used for:

[0210] The structured dataset is processed to unify its types, resulting in data with a unified type.

[0211] The data of the uniform type are subjected to feature scaling to obtain feature-scaled data.

[0212] The scaled data is reconstructed in chronological order to obtain time-series reconstructed data.

[0213] The reconstructed time series data is encoded to obtain a time series dataset.

[0214] Optionally, the processing module 22 is also specifically used for:

[0215] The time-series dataset is input into the prediction model for processing to obtain storage system state data. The prediction model obtains prediction results of multiple trees by independently predicting the time-series dataset using multiple decision trees, and integrates the prediction results of the multiple trees by majority voting to obtain storage system state data.

[0216] Optionally, the prediction model is obtained through the following training process:

[0217] Retrieve historical operational data from the storage system;

[0218] The historical operational data is preprocessed to obtain a historical structured dataset;

[0219] The historical structured dataset is formatted to obtain the sample dataset;

[0220] The sample dataset is processed to obtain a training dataset and a test sample set;

[0221] Input the training dataset into the prediction model to obtain the training result data;

[0222] The training results data are compared with the test sample set, and the kernel function, penalty parameter, and kernel function parameter are adjusted according to the comparison results to obtain the prediction model.

[0223] Optionally, module 23 is specifically used for:

[0224] The storage system status data is compared with a preset value to obtain a comparison result;

[0225] Based on the comparison results, the control signal is determined;

[0226] The operating state of the storage system is controlled according to the control signal.

[0227] It should be noted that this device is a device corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0228] like Figure 3 As shown, this embodiment of the invention also provides an operation control system for a cold data storage system, including: a backplane, a controller, and a sensor module. The controller is electrically connected to the backplane and the sensor module respectively. The backplane is provided with multiple hard disk interfaces, and the multiple hard disk interfaces are electrically connected to multiple hard disks. When the operation control system for the cold data storage system is running, it executes the above-described operation control method for the cold data storage system.

[0229] In this embodiment, the operation and control system of the cold data storage system consists of three main components: a backplane main body, a power management intelligent controller (Power Management AI), and a sensor module. Each component provides the necessary support and guarantee for the realization of the overall system function.

[0230] (1) Back panel main body

[0231] 1) The backplane is the physical foundation of the entire storage system, responsible for connecting each hard drive and providing necessary power management.

[0232] 2) Multiple Hard Drive Interfaces: The back panel features 24 hard drive interfaces, each capable of connecting to hard drives of different brands and models. The interfaces utilize a standardized design, supporting SAS / SATA / NVMe protocols to ensure hard drive compatibility and scalability. Each hard drive interface integrates an independent power control module, capable of adjusting power according to the hard drive's real-time needs, effectively reducing power consumption.

[0233] 3) Power Control Module: Each power control module is responsible for monitoring and adjusting the current and voltage of the corresponding hard drive to ensure that the hard drive operates in an optimized state. The embedded control circuit of the module controls the current and voltage through (I... 2 The C-bus communication protocol enables real-time response to power adjustment commands issued by the Power Management AI, improving system flexibility and responsiveness. The power control module employs a high-efficiency power conversion design to reduce energy consumption and heat generation, thereby enhancing system reliability and durability.

[0234] (2) Power Management AI

[0235] 1) Power Management AI is the intelligent brain of the system, responsible for data processing, decision making, and intelligent management.

[0236] 2) Embedded Microprocessor: Power Management AI employs a high-performance embedded microprocessor (ARM Cortex series multi-core processor), featuring fast data processing capabilities and low power consumption. The microprocessor is responsible for performing tasks such as data acquisition, analysis, visualization, and power supply decisions, ensuring efficient system operation and timely response to external requests.

[0237] 3) Machine Learning Module: Power Management AI incorporates a machine learning module that processes real-time monitoring data and learns historical access patterns to predict future hard drive usage. Based on modeling algorithms (such as regression analysis, decision trees, and neural networks), the module analyzes access frequency and application trends, dynamically adjusting power supply strategies to achieve effective energy management.

[0238] 4) Real-time decision-making and feedback mechanism: The Power Management AI system can monitor the status of each hard drive in real time and automatically adjust the power supply based on preset power supply strategies and real-time data. The feedback mechanism ensures that when the status of the hard drives changes (such as changes in environment or usage frequency), Power Management AI can adapt and update the power supply strategy in a timely manner to maintain system performance.

[0239] (3) Sensor module

[0240] 1) The sensor module is responsible for acquiring hard drive operating status data in real time, providing decision-making basis for Power Management AI.

[0241] 2) Real-time monitoring sensors: This module is equipped with multiple high-precision sensors specifically designed to monitor key parameters of the hard drive, such as temperature, current, power, and vibration, ensuring real-time and reliable data. Temperature sensors monitor the hard drive's operating temperature to ensure it remains within a safe range and prevent overheating damage; current sensors monitor power consumption and assess operating efficiency; vibration sensors detect abnormal vibrations and provide timely warnings of potential faults.

[0242] 3) Communication Interface: The sensor module communicates via a standardized communication interface (such as I...). 2 (e.g., C, SPI, etc.) exchange data with PowerManagement AI to quickly transmit status data, ensuring efficiency and stability.

[0243] 4) Data Integration and Transmission: The collected data is integrated by the sensor module, optimizing the transmission process and reducing communication burden. Power Management AI employs customizable sampling frequencies and data refresh rates to adapt to different working environments and application requirements.

[0244] The Power Management AI design of this invention aims to create an intelligent monitoring and management system that optimizes power management for cold data storage by monitoring hard drive status in real time and intelligently analyzing historical data. The main functions of Power Management AI include, but are not limited to:

[0245] 1) Real-time monitoring: Monitor hard drive temperature, power consumption, running time and access frequency to ensure that the hard drive operates in a safe and optimal state.

[0246] 2) Data prediction and learning: Using machine learning algorithms to predict future hard drive usage patterns and dynamically adjust power supply strategies to maximize energy efficiency.

[0247] 3) System Coordination and Adjustment: Coordinate the power supply distribution of the multi-hard drive system to ensure the overall energy efficiency and stability of the system.

[0248] Intelligent Monitoring Module: As a fundamental component of Power Management AI, the intelligent monitoring module is responsible for collecting and processing hard drive operating status data in real time. Its main features are as follows:

[0249] (1) Monitoring content:

[0250] 1) Temperature monitoring: Provides real-time temperature data to prevent hard drive damage caused by overheating.

[0251] 2) Power consumption monitoring: Accurately monitor the current and power consumption of each hard drive to assess energy efficiency.

[0252] 3) Running time statistics: Calculate the total running time of the hard drive to assess its lifespan and maintenance requirements.

[0253] 4) Access frequency recording: Tracks and records the access frequency of each hard drive to identify active and cold data.

[0254] (2) Monitoring process:

[0255] 1) Data Acquisition: Sensors are placed on each hard drive interface to continuously collect temperature, power, and operating status data. A timing module is configured to read data at defined time intervals (e.g., per second) to ensure high-frequency data acquisition.

[0256] 2) Data transmission: The acquired data is transmitted via a high-speed data bus I. 2 The data is fed into the Power Management AI processing unit, ensuring fast and secure data transmission.

[0257] 3) Data storage: Real-time monitoring data is stored using memory cache or internal database to provide a foundation for subsequent data analysis and learning.

[0258] (3) AI machine learning and training

[0259] 1) The parameter definitions are shown in Table 1.

[0260] Table 1 Parameter Definition Table

[0261]

[0262] 2) Learning Algorithm: The learning algorithm is the core part of Power Management AI, responsible for analyzing historical and real-time data to optimize power management.

[0263] The core objectives include:

[0264] Access pattern prediction: By analyzing historical usage data, actively used hard drives and cold storage hard drives (defined as hard drives that have not been read or written to in the past 3 days) are dynamically identified, and power supply strategies are formulated accordingly.

[0265] Intelligent power supply adjustment: Based on historical and real-time data analysis, the system flexibly adjusts the power supply status of each hard drive to maximize energy utilization and extend hard drive lifespan. It also learns from the system's cold data processing capabilities to identify which hard drives contain cold data.

[0266] The fault diagnosis and early warning module is an additional feature of Power Management AI, designed to improve system reliability and availability. By monitoring the health status of hard drives in real time, it promptly detects faults and issues early warnings, thereby reducing the risk of data loss and ensuring stable system operation.

[0267] Health monitoring is used to assess the operating status of the hard drive in real time. Its implementation process is as follows:

[0268] (1) Monitoring parameters:

[0269] 1) Temperature monitoring: Using built-in temperature sensors, the temperature data of each hard drive is collected in real time to ensure that it is within the specified safe operating temperature range.

[0270] 2) Current monitoring: Monitors the real-time value of hard drive current, provides feedback on power usage, and detects power consumption fluctuations.

[0271] 3) Vibration monitoring: Install vibration sensors to capture vibration data generated by the hard drive during operation. Abnormal vibration may indicate a problem with the mechanism or improper installation.

[0272] (2) Data acquisition process:

[0273] 1) Sensor deployment: Multiple sensors are placed symmetrically above each hard drive to monitor temperature, current and vibration.

[0274] 2) Real-time data collection: Power Management AI collects real-time monitoring data from each hard drive periodically (e.g., every second) and transmits it via a bus (e.g., I / O). 2 (C or SPI) transmits data to the central processing unit.

[0275] 3) Data storage: The collected real-time monitoring data is cached in memory or a database for subsequent analysis.

[0276] The intelligent fault warning function detects hard drive failures promptly and takes appropriate action based on health status monitoring results. Its implementation process is as follows:

[0277] (1) Anomaly detection:

[0278] 1) Threshold comparison: Set and maintain safety thresholds for each monitoring data for each hard drive to ensure that each parameter (such as temperature, current, vibration) is within a safe range.

[0279] 2) Real-time monitoring: Power Management AI compares real-time data with set thresholds. Once any parameter is found to exceed the safety threshold, the system will automatically mark it as "abnormal".

[0280] (2) Fault response mechanism:

[0281] 1) Alert notification: Once an anomaly is detected, Power Management AI will immediately send an alert notification to the system administrator via SMS, email or push message.

[0282] 2) Isolate the faulty hard drive: Quickly disconnect the power to the faulty hard drive to ensure the stability of the entire system.

[0283] (3) Fault recording and analysis:

[0284] 1) Fault Log: Records detailed information for each fault warning, including fault type, occurrence time, monitoring data, etc., for subsequent analysis.

[0285] 2) Fault cause analysis: Use data analysis tools to classify and summarize the hard drive's fault modes.

[0286] The operation and control system of the cold data storage system proposed in this invention realizes intelligent power supply decision-making and proactive fault early warning, thereby improving energy efficiency and user data security.

[0287] like Figure 4 As shown, this embodiment of the invention also provides a computing device 40, including a processor 41, a memory 42, and a program or instructions stored in the memory 42 and executable on the processor 41. When the program or instructions are executed by the processor 41, they implement the various processes of the above-described cold data storage system operation control method embodiment and achieve the same technical effect. To avoid repetition, they will not be described again here. It should be noted that the computing device in this embodiment of the invention includes the above-described mobile electronic devices and non-mobile electronic devices.

[0288] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0289] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0290] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0291] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0292] In addition, the functional units in the various embodiments of the present invention 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.

[0293] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0294] Furthermore, it should be noted that in the apparatus and method of the present invention, it is obvious that the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent solutions of the present invention. Moreover, the steps performing the above series of processes can naturally be executed in the order described, but are not necessarily required to be executed in chronological order; some steps can be executed in parallel or independently of each other. Those skilled in the art will understand that all or any step or component of the method and apparatus of the present invention can be implemented in any computing device (including processors, storage media, etc.) or network of computing devices, in hardware, firmware, software, or a combination thereof. This is something that those skilled in the art can achieve by using their basic programming skills after reading the description of the present invention.

[0295] Therefore, the object of the present invention can also be achieved by running a program or a set of programs on any computing device. The computing device can be a known general-purpose device. Therefore, the object of the present invention can also be achieved simply by providing a program product containing program code for implementing the method or apparatus. That is, such a program product also constitutes the present invention, and the storage medium storing such a program product also constitutes the present invention. Obviously, the storage medium can be any known storage medium or any storage medium developed in the future. It should also be noted that in the apparatus and method of the present invention, it is obvious that the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent to the present invention. Furthermore, the steps for performing the above series of processes can naturally be performed in the order described, but are not necessarily required to be performed in chronological order. Some steps can be performed in parallel or independently of each other.

[0296] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for controlling the operation of a cold data storage system, characterized in that, include: Obtain the operating data of the storage system; The running data is preprocessed to obtain a structured dataset; The structured dataset is formatted to obtain a time-series dataset; The time-series dataset is input into the prediction model for processing to obtain the storage system state data; Based on the storage system status data, determine the storage system operating data; Specifically, the time-series dataset is input into a prediction model for processing to obtain stored system state data, including: The time-series dataset is input into a prediction model for processing to obtain storage system state data. The prediction model obtains prediction results from multiple trees by independently predicting the time-series dataset, and then integrates these prediction results through majority voting to obtain the storage system state data. The specific processing flow includes: Time series dataset Data is segmented according to a preset window size; the time-series dataset contains multiple standardized features of the hard disk within the time window, including temperature, current, vibration, and access frequency, and the data structure is as follows: in, For standardized time series datasets; For the first m 1 eigenvector; m The number of features; Given a preset window length, generate a window sample: in, For window samples, ; W The length of the window; In a random forest, each decision tree makes independent predictions, randomly selects a subset of features, and generates predicted labels based on a pre-defined algorithm. The number of features used for each tree is: in, The number of features used for each tree, The proportion of features considered for each tree; It is a floor function; All tree prediction results are represented as follows: in, For the first n The predicted results for each tree; Iterate through the prediction results of all decision trees in the random forest, totaling... n Tree; For each tree's predicted label, determine: if the tree's output label is cold data, then the counter... Increment by 1; if the tree output label is active, the counter remains unchanged; obtain the total number of votes for the cold data: Divide the number of votes for cold data by the total number of trees: ;in, The percentage of trees that support cold data labels; like If ≥0.5, or ≥50% of the trees support cold data, then the final label is determined to be cold data. ;in, For final tags; like If the value is less than 0.5, meaning less than 50% of the trees support cold data, then the final label is determined to be active. ; like If the actual access frequency is less than or equal to cold_threshold, then the data is considered cold data. ; like If the access frequency is greater than or equal to active_threshold, then the device is considered active. ; Where cold_threshold is the maximum number of accesses to cold data; active_threshold is the minimum number of accesses to active hard drives; To store system status data; The prediction model is obtained through the following training process: Retrieve historical operational data from the storage system; The historical operational data is preprocessed to obtain a historical structured dataset; The historical structured dataset is formatted to obtain the sample dataset; The sample dataset is processed to obtain a training dataset and a test sample set; Input the training dataset into the prediction model to obtain the training result data; The training result data is compared with the test sample set, and the kernel function, penalty parameter and kernel function parameter are adjusted according to the comparison result to obtain the prediction model; The storage system operation data is determined based on the storage system status data, including: The storage system status data is compared with a preset value to obtain a comparison result; Based on the comparison results, the control signal is determined; The operating state of the storage system is controlled according to the control signal.

2. The operation control method for the cold data storage system according to claim 1, characterized in that, Obtain operational data from the storage system, including: Based on sensors installed in the cold data storage system, operational data of the storage system is acquired, including at least one of hard disk temperature, current, power, vibration, and access frequency.

3. The operation control method for the cold data storage system according to claim 1, characterized in that, The runtime data is preprocessed to obtain a structured dataset, including: The missing value imputation process is performed on the running data to obtain the imputed running data; The filled-in running data is subjected to outlier detection processing to obtain outlier-detected running data; The deduplication process is performed on the operational data after the anomaly detection to obtain the deduplicated operational data; The deduplicated running data is then subjected to noise reduction processing to obtain a structured dataset.

4. The operation control method for the cold data storage system according to claim 1, characterized in that, The structured dataset is formatted to obtain a time-series dataset, including: The structured dataset is processed to unify its types, resulting in data with a unified type. The data of the uniform type are subjected to feature scaling to obtain feature-scaled data. The scaled data is reconstructed in chronological order to obtain time-series reconstructed data. The reconstructed time series data is encoded to obtain a time series dataset.

5. An operation control device for a cold data storage system, characterized in that, include: The acquisition module is used to acquire the operating data of the storage system; The processing module is used to preprocess the running data to obtain a structured dataset; The structured dataset is formatted to obtain a time-series dataset; The time-series dataset is input into the prediction model for processing to obtain the storage system state data; The determination module is used to determine the storage system operating data based on the storage system status data; Specifically, the time-series dataset is input into a prediction model for processing to obtain stored system state data, including: The time-series dataset is input into a prediction model for processing to obtain storage system state data. The prediction model obtains prediction results from multiple trees by independently predicting the time-series dataset, and then integrates these prediction results through majority voting to obtain the storage system state data. The specific processing flow includes: Time series dataset Data is segmented according to a preset window size; the time-series dataset contains multiple standardized features of the hard disk within the time window, including temperature, current, vibration, and access frequency, and the data structure is as follows: in, For standardized time series datasets; For the first m 1 eigenvector; m The number of features; Given a preset window length, generate a window sample: in, For window samples, ; W The length of the window; In a random forest, each decision tree makes independent predictions, randomly selects a subset of features, and generates predicted labels based on a pre-defined algorithm. The number of features used for each tree is: in, The number of features used for each tree, The proportion of features considered for each tree; It is a floor function; All tree prediction results are represented as follows: in, For the first n The predicted results for each tree; Iterate through the prediction results of all decision trees in the random forest, totaling... n Tree; For each tree's predicted label, determine: if the tree's output label is cold data, then the counter... Increment by 1; if the tree output label is active, the counter remains unchanged; obtain the total number of votes for the cold data: Divide the number of votes for cold data by the total number of trees: ;in, The percentage of trees that support cold data labels; like If ≥0.5, or ≥50% of the trees support cold data, then the final label is determined to be cold data. ;in, For final tags; like If the value is less than 0.5, meaning less than 50% of the trees support cold data, then the final label is determined to be active. ; like If the actual access frequency is less than or equal to cold_threshold, then the data is considered cold data. ; like If the access frequency is greater than or equal to active_threshold, then the device is considered active. ; Where cold_threshold is the maximum number of accesses to cold data; active_threshold is the minimum number of accesses to active hard drives; To store system status data; The prediction model is obtained through the following training process: Retrieve historical operational data from the storage system; The historical operational data is preprocessed to obtain a historical structured dataset; The historical structured dataset is formatted to obtain the sample dataset; The sample dataset is processed to obtain a training dataset and a test sample set; Input the training dataset into the prediction model to obtain the training result data; The training result data is compared with the test sample set, and the kernel function, penalty parameter and kernel function parameter are adjusted according to the comparison result to obtain the prediction model; Based on the storage system status data, determine the storage system operating data, including: The storage system status data is compared with a preset value to obtain a comparison result; Based on the comparison results, the control signal is determined; The operating state of the storage system is controlled according to the control signal.

6. An operation control system for a cold data storage system, comprising: The system comprises a backplane, a controller, and a sensor module, wherein the controller is electrically connected to the backplane and the sensor module respectively, and the backplane is provided with multiple hard disk interfaces, which are electrically connected to multiple hard disks. The system is characterized in that, when the cold data storage system's operation control system is running, it executes the method described in any one of claims 1 to 4.

7. A computing device, characterized in that, include: A processor, a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method as described in any one of claims 1 to 4.

Images