Machine hard disk fault detection system and method and medium

Abstract

The invention discloses a machine hard disk fault detection system and method and a medium, mainly relates to the technical field of hard disk faults, and is used for solving the problem that an existing scheme is not suitable for learning model training under the conditions that the number of produced data samples is small and the data samples are different from public data set distribution; in addition, the operation and maintenance of the server hard disk in the prior art are often to replace the faulted hard disk after the hard disk breaks down. Comprising a data acquisition module used for automatically detecting and recording S.M.A.R.T parameters of the hard disk according to a preset period; serialized time information is obtained; taking the S.M.A.R.T parameter and the S.M.A.R.T parameter as a group of prediction data; the fault prediction module is used for inputting prediction data into the trained Transform model to obtain extracted feature data; and inputting the feature data and the S.M.A.R.T parameters into an XGBoost model to obtain a prediction result.

Description

Technical Field

[0001] The present application relates to the technical field of hard disk fault detection, and in particular to a machine hard disk fault detection system, method, and medium. Background Art

[0002] Hard drive failure prediction is a critical task for maintaining data security and stable system operations. Existing hard drive failure prediction approaches are primarily categorized into traditional machine learning and deep learning approaches. Traditional machine learning relies on manual feature engineering of data upfront. While the model is simple and the number of input samples does not need to be large, the prediction results are highly dependent on the quality of the feature engineering. Deep learning approaches, on the other hand, use models such as LSTM or Transformer to extract high-dimensional features from raw data. These models are more complex but offer greater predictive power.

[0003] However, in actual implementation, deep learning models require a large number of training samples, making them unsuitable for situations where a company's production data samples are small and have distribution differences from publicly available datasets. Furthermore, when maintaining server hard drives, company operations personnel often replace them after a drive failure. Consequently, a failed hard drive may not be replaced in time, leading to server issues.

[0004] Therefore, there is an urgent need for a machine hard disk fault detection system, method and medium to solve the problem that the existing solution has a small number of production data samples and the distribution is different from the public data set, which is not suitable for learning model training; in addition, the existing operation and maintenance of server hard disks often involves replacing the faulty hard disk after the hard disk fails and crashes. Summary of the Invention

[0005] In response to the above-mentioned deficiencies in the prior art, the present application provides a machine hard disk fault detection system, method and medium to solve the problem that the existing solution has a small number of production data samples and a distribution difference with the public data set, which is not suitable for learning model training; in addition, the existing operation and maintenance of server hard disks often involves replacing the faulty hard disk after the hard disk fails and crashes.

[0006] In a first aspect, the present application provides a machine hard disk fault detection system, the system comprising: The data acquisition module is used to automatically detect and record the SMART parameters of the hard disk according to a preset period; the SMART parameters in the current detection period are processed in time serialization according to the time nodes to obtain serialized time information; and the SMART parameters and SMART parameters are used as a set of prediction data; The model training module is used to randomly extract a preset minimum amount of prediction data from the prediction data as initial training data; label the initial training data, input the labeled initial training data into the small sample expansion program to obtain training data of a preset scale; and input the preset scale of training data into the Transformer model to obtain a trained Transformer model; The fault prediction module is used to input the prediction data into the trained Transformer model to obtain the extracted feature data; the feature data and SMART parameters are input into the XGBoost model to obtain the prediction results; The risk warning module is used to send the prediction results and the corresponding hard disk number to the preset maintenance terminal.

[0007] In one implementation of the present application, the data acquisition module includes a timing processing unit, Used to convert the time data in SMART parameters into preset format data; Using progressive importance time coding, the weight of the time data in a preset format is adjusted so that the weight of the time data closest to the current time point is the largest, thereby obtaining serialized time information.

[0008] In one implementation of the present application, the model training module includes a sample verification unit, Used to input the predicted data into the small sample expansion program to obtain a certain amount of synthetic data; Convert the predicted data and the synthesized data into vector data; Calculate the distance between the predicted data and the synthesized data vector; Determine that the synthetic data with a distance less than a preset minimum distance is qualified training data; Delete the synthetic data whose distance is greater than or equal to the preset minimum distance.

[0009] In one implementation of the present application, the fault prediction module includes a data processing unit, The Encoder layer is used to input the predicted data into the trained Transformer model; the Encoder layer presets N multi-head attention mechanisms and feedforward neural networks; Extract the features output by the Encoder in the trained Transformer model as feature data.

[0010] In one implementation of the present application, the system further includes: an accuracy training module, It is used to randomly extract a preset minimum number of prediction data from the prediction data as initial training data; mark the initial training data, input the marked initial training data into the small sample expansion program to obtain training data of a preset scale; input the preset scale training data into the XGBoost model to obtain a trained XGBoost model.

[0011] In a second aspect, the present application provides a method for detecting a machine hard disk failure, the method comprising: Automatically detect and record the hard disk's SMART parameters according to a preset cycle; perform time serialization processing on the SMART parameters in the current detection period according to the time node to obtain serialized time information; and use the SMART parameters and SMART parameters as a set of prediction data; Randomly extracting a preset minimum amount of prediction data from the prediction data as initial training data; labeling the initial training data, inputting the labeled initial training data into a small sample expansion program to obtain training data of a preset scale; inputting the preset scale of training data into the Transformer model to obtain a trained Transformer model; Input the predicted data into the trained Transformer model to obtain the extracted feature data; input the feature data and SMART parameters into the XGBoost model to obtain the prediction results; The prediction result and the corresponding hard disk number are sent to the preset maintenance terminal.

[0012] In one implementation of the present application, the SMART parameters in the current detection time period are subjected to time serialization processing according to time nodes to obtain serialized time information, specifically including: Convert the time data in SMART parameters into preset format data; Using progressive importance time coding, the weight of the time data in a preset format is adjusted so that the weight of the time data closest to the current time point is the largest, thereby obtaining serialized time information.

[0013] In one implementation of the present application, the labeled initial training data is input into a small sample expansion program to obtain training data of a preset size, specifically including: Input the predicted data into the small sample expansion program to obtain a certain amount of synthetic data; Convert the predicted data and the synthesized data into vector data; Calculate the distance between the predicted data and the synthesized data vector; Determine that the synthetic data with a distance less than a preset minimum distance is qualified training data; Delete the synthetic data whose distance is greater than or equal to the preset minimum distance.

[0014] In one implementation of the present application, the predicted data is input into a trained Transformer model to obtain extracted feature data, specifically including: The predicted data is input into the Encoder layer of the trained Transformer model; the Encoder layer presets N multi-head attention mechanisms and feedforward neural networks; Extract the features output by the Encoder in the trained Transformer model as feature data.

[0015] In a third aspect, the present application provides a non-volatile computer storage medium having computer instructions stored thereon, which, when executed, implement a machine hard disk failure detection method as described above.

[0016] Those skilled in the art will appreciate that this application has at least the following beneficial effects: This application provides a machine hard drive fault detection system, method, and medium. The data acquisition module is responsible for automatically detecting and recording the hard drive's SMART (Self-Monitoring, Analysis, and Reporting Technology) parameters at a preset interval. These parameters include, but are not limited to, hard drive temperature, read and write error rates, and the number of reallocated sectors, providing a comprehensive picture of the drive's health. The module also performs time-series processing on the SMART parameters, generating serialized time information to provide temporal data support for subsequent model training.

[0017] Model Training Module: First, a minimum number of prediction data is randomly sampled from the prediction data as initial training data and labeled to distinguish between normal and faulty samples. Then, a small sample expansion program is used to amplify this initial training data to generate training data of a predetermined size, addressing the limited number of production data samples. Finally, this amplified training data is fed into the Transformer model for training. Leveraging the Transformer model's powerful sequence modeling capabilities, the model learns how hard drive SMART parameters change over time, ultimately yielding a trained Transformer model.

[0018] Fault Prediction Module: This module feeds new prediction data (including SMART parameters and serialization time information) into a trained Transformer model to extract feature data. This feature data contains in-depth information about the drive's health. This feature data, along with the original SMART parameters, is then fed into the XGBoost model, leveraging its efficient classification capabilities to predict drive failures.

[0019] Risk Alert Module: Based on the fault prediction module's predictions, the hard drive numbers predicted to be faulty and their predictions are sent to a pre-set maintenance terminal, such as the operator's phone, computer, or dedicated maintenance management system, to provide timely warnings and accurate fault location. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for the description. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 This is a schematic diagram of the internal structure of a machine hard disk fault detection system provided in an embodiment of the present application.

[0022] Figure 2 This is a flow chart of a machine hard disk failure detection method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0023] It should be understood by those skilled in the art that the embodiments described below are merely preferred embodiments of the present disclosure and do not imply that the present disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely intended to explain the technical principles of the present disclosure and are not intended to limit the scope of protection of the present disclosure. Based on the preferred embodiments provided by the present disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of the present disclosure.

[0024] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0025] The technical solutions proposed in the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0026] This application Figure 1 A machine hard disk fault detection system is provided in the embodiment of the present application. Figure 1 As shown, the system provided in the embodiment of the present application mainly includes: The data acquisition module 110 is used to automatically detect and record the SMART parameters of the hard disk according to a preset period; perform time serialization processing on the SMART parameters in the current detection time period according to the time node to obtain serialized time information; and use the SMART parameters and SMART parameters as a set of prediction data.

[0027] It should be noted that the core function of data acquisition module 110 is to periodically acquire hard drive SMART parameters (such as temperature, read / write error rate, and sector remapping count). These parameters are generated in real time by the hard drive's built-in self-monitoring technology and reflect the drive's health status. Data is automatically scanned and recorded at a preset period (e.g., daily or hourly) by invoking the hard drive interface (e.g., SCSI or SATA commands), ensuring coverage of workload variations over time. For example, high-frequency acquisition can capture transient anomalies (such as motor current fluctuations), while low-frequency acquisition is suitable for long-term trend analysis (such as platter wear).

[0028] The aforementioned use of SMART parameters and SMART parameters as a set of prediction data can be specifically implemented as follows: Feature fusion: A single data point contains the original parameter value (such as the current temperature value) and time series features (such as the temperature change rate over the past 24 hours), forming a multidimensional input vector. Normalization: Z-score normalization is performed on parameters of different dimensions (such as temperature in °C and motor starts and stops in counts) to prevent model bias due to differences in numerical ranges. Label association: If subsequent model training requires supervised learning, the data set must be associated with historical fault records (e.g., labeled "normal" or "warning") to supervise model training.

[0029] The data acquisition module 110 includes a timing processing unit for converting the time data in the SMART parameters into data in a preset format; Using progressive importance time coding, the weight of the time data in a preset format is adjusted so that the weight of the time data closest to the current time point is the largest, thereby obtaining serialized time information.

[0030] It should be noted that the core function of the timing processing unit is to convert the time data in the SMART parameters (such as the acquisition timestamp) into standardized data in a preset format. The specific implementation includes: Timestamp alignment: Converts raw time data (such as "2025-03-03 10:15:30") to a unified format (such as Unix timestamp or ISO 8601 format) to eliminate time format differences between different systems / devices.

[0031] ‌Time Zone Synchronization‌: Adjust time data according to a preset time zone (such as UTC+8) to ensure time consistency across server clusters across regions.

[0032] ‌Example‌: Original time data: 2025-03-03T10:15:30+08:00; After normalization: 1709360130 (Unix timestamp).

[0033] The saliency of recent data is enhanced through a dynamic weight allocation strategy n. Specific techniques include: Exponential decay model: Assign a weight coefficient to each time point. The larger the time difference, the lower the weight.

[0034] Sliding Window Mechanism: Only data from the last N time windows (e.g., the past 24 hours) is retained. The weight of data within the window increases linearly, and the weight of data outside the window is reset to zero.

[0035] Combine the normalized time data with the weighted parameters to form a multidimensional time series feature vector: Feature dimension expansion: Based on the original SMART parameters, a new time weight field (such as "time decay factor") is added for subsequent model training.

[0036] Dynamic feature concatenation: Concatenates multi-period data in chronological order to generate a structured sequence such as [parameter value_1, weight_1, parameter value_2, weight_2, ..., parameter value_n, weight_].

[0037] The model training module 120 is used to randomly extract a preset minimum amount of prediction data from the prediction data as initial training data; mark the initial training data, input the marked initial training data into a small sample expansion program to obtain training data of a preset scale; and input the preset scale of training data into the Transformer model to obtain a trained Transformer model.

[0038] As an example, this module may be specifically: 1. Initial training data is randomly extracted: The module uses a random sampling strategy to extract a preset minimum number of initial training data (e.g., 1,000 groups) from the prediction data to ensure the diversity and representativeness of the data distribution and avoid model overfitting due to local data deviation. The specific implementation includes: Uniform sampling: Uses random sampling without replacement to cover SMART parameter combinations in different time periods to reduce sampling bias.

[0039] Dynamic Adjustment: If the amount of data after extraction is still insufficient (e.g., a shortage of hard drive failure samples), oversampling techniques (such as SMOTE) can be used to supplement rare category data.

[0040] 2. Marking Processing and Small Sample Expansion: Label and enhance the initial training data to solve the model generalization problem in small sample scenarios: Label Definition: Data is labeled as "normal" or "faulty" based on SMART parameter thresholds (such as temperature > 60°C or a sudden increase in sector error rate), forming the labeling system required for supervised learning.

[0041] 3. Small sample extension program: Feature space expansion: Generate synthetic data by adding Gaussian noise and time series shift (e.g., ±5% time offset) to expand the sample size.

[0042] Semi-supervised learning: Utilizes unlabeled data (such as unlabeled SMART history records) for self-training to improve the model’s adaptability to unseen data. 5

[0043] ‌Example‌: The initial 1,000 sets of data can be expanded to 5,000 sets after noise injection and time shifting, covering more potential failure modes‌3.

[0044] 4. Transformer model training: The expanded training data is fed into the Transformer model, and its self-attention mechanism is used to capture the long-term dependencies of SMART parameters: Input encoding: Concatenate the time-series SMART parameters (such as temperature and error rate) and the time-decay weights into a high-dimensional vector as the model input.

[0045] Multi-head attention layer: Calculates the correlation strength between different parameters (such as the synchronous change of temperature rise and motor load) and identifies the combination of fault characteristics.

[0046] Output Optimization: Adopts the cross-entropy loss function and Adam optimizer, dynamically adjusts the learning rate to accelerate convergence, and prevents overfitting through the Dropout layer.

[0047] In addition, in order to prevent abnormalities in the training data, the present application may include a sample verification unit in the model training module 120 to perform data verification. The specific process is as follows: Used to input the predicted data into the small sample expansion program to obtain a certain amount of synthetic data; Convert the predicted data and the synthesized data into vector data; Calculate the distance between the predicted data and the synthesized data vector; Determine that the synthetic data with a distance less than a preset minimum distance is qualified training data; Delete the synthetic data whose distance is greater than or equal to the preset minimum distance.

[0048] Specifically, synthetic data is generated through data augmentation techniques (such as adding Gaussian noise and time shifting) or generative adversarial networks (GANs) to address the problem of insufficient original prediction data. For example, virtual samples with slight perturbations are generated based on the original SMART parameter sequence to simulate potential failure modes. The predicted and synthetic data are mapped into high-dimensional vectors using a pre-trained embedding layer or feature extraction model (such as BERT or ResNet), forming a measurable feature space. The similarity between the predicted data vector and the synthetic data vector is calculated using Euclidean distance or cosine similarity. The smaller the distance, the closer the synthetic data distribution is to the real data. A preset minimum distance (such as a Euclidean distance threshold of 0.5) is set to retain only synthetic data with a distance below the threshold, while anomalous synthetic samples that deviate from the real distribution are discarded. For example, if a synthetic data point exceeds the distance threshold due to excessive noise, it will be discarded.

[0049] The fault prediction module 130 is used to input the prediction data into the trained Transformer model to obtain extracted feature data; input the feature data and SMART parameters into the XGBoost model to obtain the prediction results; The fault prediction module 130 includes a data processing unit for inputting the prediction data into the Encoder layer of the trained Transformer model; wherein the Encoder layer presets N multi-head attention mechanisms and feedforward neural networks; extracting the features output by the Encoder in the trained Transformer model as feature data.

[0050] For example, consider inputting a SMART parameter sequence (temperature, error rate, read / write latency) into the TransformerEncoder layer to generate a 128-dimensional feature vector. The time series feature data extracted by the Transformer is concatenated with the original SMART parameters (such as the current temperature and the cumulative number of errors) to form a mixed feature vector that is then fed into the XGBoost model. XGBoost uses the gradient boosting tree algorithm to process structured features, combining the abstract time series features with the original statistical parameters to improve prediction accuracy.

[0051] In order to improve the accuracy of the model, the feature data and SMART parameters are input into the XGBoost model and the XGBoost model is trained at the same time, the accuracy training module, It is used to randomly extract a preset minimum number of prediction data from the prediction data as initial training data; mark the initial training data, input the marked initial training data into the small sample expansion program to obtain training data of a preset scale; input the preset scale training data into the XGBoost model to obtain a trained XGBoost model.

[0052] The risk warning module 140 is used to send the prediction result and the corresponding hard disk number to a preset maintenance terminal.

[0053] It should be noted that the risk alert module 140 binds the hard drive number to the prediction results (such as failure probability and health score) to ensure that alert information can be traced back to the specific physical device. For example, by mapping the hard drive number with metadata such as the device serial number and deployment location through a database, the efficiency of operation and maintenance positioning is improved. Structured data (JSON format) is pushed to the maintenance terminal using HTTP / REST or MQTT protocols, and is compatible with mainstream operation and maintenance platforms (such as Zabbix and Prometheus). Multiple alert thresholds (such as warning, critical, and emergency) are set based on the prediction results, and different notification methods (email, SMS, API callback) are triggered for different levels. For example, when the failure probability is ≥80%, an SMS message is sent to the operation and maintenance personnel, and when it is ≥95%, a work order is automatically created. A time window (such as 30 minutes) is set for repeated alerts for the same hard drive number to avoid frequent notification interference.

[0054] Based on the above description, this application provides a machine hard drive fault detection system, method, and medium. The data acquisition module is responsible for automatically detecting and recording the hard drive's SMART (Self-Monitoring, Analysis, and Reporting Technology) parameters according to a preset period. These parameters include, but are not limited to, hard drive temperature, read and write error rate, and the number of reallocated sectors, providing a comprehensive reflection of the hard drive's health. Furthermore, this module performs time-series processing on the SMART parameters, generating serialized time information to provide temporal data support for subsequent model training. First, a preset minimum number of prediction data is randomly extracted from the prediction data as initial training data and labeled to distinguish between normal and faulty samples. Then, a small sample expansion program is used to amplify this initial training data to generate training data of a preset size, addressing the issue of a small number of production data samples. Finally, this amplified training data is fed into a Transformer model for training. Leveraging the Transformer model's powerful sequence modeling capabilities, the model learns how the hard drive's SMART parameters change over time, resulting in a trained Transformer model. The new prediction data (including SMART parameters and serialization time information) is input into the trained Transformer model to extract feature data. These feature data contain in-depth information about the health status of the hard disk. Subsequently, the feature data and the original SMART parameters are input into the XGBoost model together, and the prediction results of the hard disk failure are obtained by using the efficient classification ability of the XGBoost model. According to the prediction results of the fault prediction module, the hard disk number predicted to be faulty and its prediction results are sent to the preset maintenance terminal, such as the operation and maintenance personnel's mobile phone, computer or special operation and maintenance management system, to achieve timely alarm and accurate positioning of the fault. This solves the problem that the existing solution has a small number of production data samples and is different from the distribution of the public data set, which is not suitable for learning model training, and the existing operation and maintenance of server hard disks often involves replacing the faulty hard disk after the hard disk fails and goes down.

[0055] In addition, the embodiment provides a method for detecting a machine hard disk failure, such as Figure 2 As shown, the method provided in the embodiment of the present application mainly includes the following steps: Step 210: Automatically detect and record the SMART parameters of the hard disk according to a preset period; perform time serialization processing on the SMART parameters in the current detection time period according to the time nodes to obtain serialized time information; and use the SMART parameters and the SMART parameters as a set of prediction data.

[0056] It's important to note that this step periodically acquires the drive's SMART parameters (such as temperature, read / write error rate, and sector remapping count). These parameters are generated in real time by the drive's built-in self-monitoring technology and reflect the drive's health status. Data is automatically scanned and recorded at a preset interval (e.g., daily or hourly) by invoking the drive interface (such as SCSI or SATA commands), ensuring coverage of workload variations over time. For example, high-frequency acquisition can capture transient anomalies (such as motor current fluctuations), while low-frequency acquisition is suitable for long-term trend analysis (such as platter wear).

[0057] In some embodiments, the SMART parameters and the SMART parameters are used as a set of prediction data, and the specific implementation includes: Timestamp alignment: Converts raw time data (such as "2025-03-03 10:15:30") to a unified format (such as Unix timestamp or ISO 8601 format) to eliminate time format differences between different systems / devices.

[0058] ‌Time Zone Synchronization‌: Adjust time data according to a preset time zone (such as UTC+8) to ensure time consistency across server clusters across regions.

[0059] ‌Example‌: Original time data: 2025-03-03T10:15:30+08:00; After normalization: 1709360130 (Unix timestamp).

[0060] The saliency of recent data is enhanced through a dynamic weight allocation strategy n. Specific techniques include: Exponential decay model: Assign a weight coefficient to each time point. The larger the time difference, the lower the weight.

[0061] Sliding Window Mechanism: Only data from the last N time windows (e.g., the past 24 hours) is retained. The weight of data within the window increases linearly, and the weight of data outside the window is reset to zero.

[0062] Combine the normalized time data with the weighted parameters to form a multidimensional time series feature vector: Feature dimension expansion: Based on the original SMART parameters, a new time weight field (such as "time decay factor") is added for subsequent model training.

[0063] Dynamic feature concatenation: Concatenates multi-period data in chronological order to generate a structured sequence such as [parameter value_1, weight_1, parameter value_2, weight_2, ..., parameter value_n, weight_].

[0064] For the SMART parameters in the current detection time period, time serialization processing is performed according to the time nodes to obtain serialized time information, which can be specifically: Convert the time data in SMART parameters into preset format data; Using progressive importance time coding, the weight of the time data in a preset format is adjusted so that the weight of the time data closest to the current time point is the largest, thereby obtaining serialized time information.

[0065] Step 220: randomly extract a preset minimum amount of prediction data from the prediction data to serve as initial training data; label the initial training data, input the labeled initial training data into a small sample expansion program to obtain training data of a preset scale; and input the preset scale of training data into a Transformer model to obtain a trained Transformer model.

[0066] Input the labeled initial training data into the small sample expansion program to obtain training data of a preset size, including: Input the predicted data into the small sample expansion program to obtain a certain amount of synthetic data; Convert the predicted data and the synthesized data into vector data; Calculate the distance between the predicted data and the synthesized data vector; Determine that the synthetic data with a distance less than a preset minimum distance is qualified training data; Delete the synthetic data whose distance is greater than or equal to the preset minimum distance.

[0067] Step 230: Input the predicted data into the trained Transformer model to obtain extracted feature data; input the feature data and SMART parameters into the XGBoost model to obtain the prediction results.

[0068] For example, consider inputting a SMART parameter sequence (temperature, error rate, read / write latency) into the TransformerEncoder layer to generate a 128-dimensional feature vector. The time series feature data extracted by the Transformer is concatenated with the original SMART parameters (such as the current temperature and the cumulative number of errors) to form a mixed feature vector that is then fed into the XGBoost model. XGBoost uses the gradient boosting tree algorithm to process structured features, combining the abstract time series features with the original statistical parameters to improve prediction accuracy.

[0069] Input the predicted data into the trained Transformer model to obtain the extracted feature data, which can be specifically: The predicted data is input into the Encoder layer of the trained Transformer model; wherein, the Encoder layer presets N multi-head attention mechanisms and feedforward neural networks; the features output by the Encoder in the trained Transformer model are extracted as feature data.

[0070] Step 240: Send the prediction result and the corresponding hard disk number to a preset maintenance terminal.

[0071] In some embodiments, this step binds the drive ID to the prediction results (such as failure probability and health score) to ensure that alarm information can be traced back to the specific physical device. For example, a database can be used to map the drive ID with metadata such as the device serial number and deployment location to improve O&M location efficiency. Structured data (JSON format) is pushed to the maintenance terminal using HTTP / REST or MQTT protocols, compatible with mainstream O&M platforms (such as Zabbix and Prometheus). Multiple alarm thresholds (such as warning, critical, and emergency) are set based on the prediction results, and different notification methods (email, SMS, API callback) are triggered for different levels. For example, a SMS message is sent to O&M personnel when the failure probability is ≥80%, and a work order is automatically created when it is ≥95%. A time window (such as 30 minutes) is set for repeated alarms for the same drive ID to avoid frequent notification interruptions.

[0072] Thus far, the technical solutions of the present disclosure have been described in conjunction with the foregoing multiple embodiments. However, it is easy for those skilled in the art to understand that the scope of protection of the present disclosure is not limited to these specific embodiments. Without departing from the technical principles of the present disclosure, those skilled in the art may split and combine the technical solutions in the above-mentioned various embodiments, and may also make equivalent changes or replacements to the relevant technical features. Any changes, equivalent replacements, improvements, etc. made within the technical concepts and / or technical principles of the present disclosure will fall within the scope of protection of the present disclosure.

Claims

1. A machine hard disk fault detection system, characterized in that: The system comprises: The data acquisition module is used to automatically detect and record the SMART parameters of the hard disk according to a preset period; the SMART parameters in the current detection period are processed in time serialization according to the time nodes to obtain serialized time information; and the SMART parameters and SMART parameters are used as a set of prediction data; The model training module is used to randomly extract a preset minimum amount of prediction data from the prediction data as initial training data; label the initial training data, input the labeled initial training data into the small sample expansion program to obtain training data of a preset scale; and input the preset scale of training data into the Transformer model to obtain a trained Transformer model; The fault prediction module is used to input the prediction data into the trained Transformer model to obtain the extracted feature data; the feature data and SMART parameters are input into the XGBoost model to obtain the prediction results; The risk warning module is used to send the prediction results and the corresponding hard disk number to the preset maintenance terminal.

2. The machine hard disk fault detection system according to claim 1, characterized in that: The data acquisition module includes a timing processing unit, Used to convert the time data in SMART parameters into preset format data; Using progressive importance time coding, the weight of the time data in a preset format is adjusted so that the weight of the time data closest to the current time point is the largest, thereby obtaining serialized time information.

3. The machine hard disk fault detection system according to claim 1, characterized in that: The model training module includes a sample verification unit, Used to input the predicted data into the small sample expansion program to obtain a certain amount of synthetic data; Convert the predicted data and the synthesized data into vector data; Calculate the distance between the predicted data and the synthesized data vector; Determine that the synthetic data with a distance less than a preset minimum distance is qualified training data; Delete the synthetic data whose distance is greater than or equal to the preset minimum distance.

4. The machine hard disk fault detection system according to claim 1, characterized in that: The fault prediction module includes a data processing unit, The Encoder layer is used to input the predicted data into the trained Transformer model; the Encoder layer presets N multi-head attention mechanisms and feedforward neural networks; Extract the features output by the Encoder in the trained Transformer model as feature data.

5. The machine hard disk fault detection system according to claim 1, characterized in that: The system also includes: a precision training module, It is used to randomly extract a preset minimum number of prediction data from the prediction data as initial training data; mark the initial training data, input the marked initial training data into the small sample expansion program to obtain training data of a preset scale; input the preset scale training data into the XGBoost model to obtain a trained XGBoost model.

6. A method for detecting a machine hard disk failure, characterized in that: The method comprises: Automatically detect and record the hard disk's SMART parameters according to a preset cycle; perform time serialization processing on the SMART parameters in the current detection period according to the time node to obtain serialized time information; and use the SMART parameters and SMART parameters as a set of prediction data; Randomly extracting a preset minimum amount of prediction data from the prediction data as initial training data; labeling the initial training data, inputting the labeled initial training data into a small sample expansion program to obtain training data of a preset scale; inputting the preset scale of training data into the Transformer model to obtain a trained Transformer model; Input the predicted data into the trained Transformer model to obtain the extracted feature data; input the feature data and SMART parameters into the XGBoost model to obtain the prediction results; The prediction result and the corresponding hard disk number are sent to the preset maintenance terminal.

7. The machine hard disk failure detection method according to claim 6, characterized in that: For the SMART parameters in the current detection time period, time serialization processing is performed according to the time nodes to obtain serialized time information, including: Convert the time data in SMART parameters into preset format data; Using progressive importance time coding, the weight of the time data in a preset format is adjusted so that the weight of the time data closest to the current time point is the largest, thereby obtaining serialized time information.

8. The method for detecting a machine hard disk failure according to claim 6, wherein: Input the labeled initial training data into the small sample expansion program to obtain training data of a preset size, including: Input the predicted data into the small sample expansion program to obtain a certain amount of synthetic data; Convert the predicted data and the synthesized data into vector data; Calculate the distance between the predicted data and the synthesized data vector; Determine that the synthetic data with a distance less than a preset minimum distance is qualified training data; Delete the synthetic data whose distance is greater than or equal to the preset minimum distance.

9. The machine hard disk failure detection method according to claim 6, characterized in that: Input the predicted data into the trained Transformer model to obtain the extracted feature data, including: The predicted data is input into the Encoder layer of the trained Transformer model; the Encoder layer presets N multi-head attention mechanisms and feedforward neural networks; Extract the features output by the Encoder in the trained Transformer model as feature data.

10. A non-volatile computer storage medium, characterized in that Computer instructions are stored thereon, and when the computer instructions are executed, a machine hard disk failure detection method according to any one of claims 6 to 9 is implemented.

Images