Method for predicting service life of storage device, electronic device, and storage medium
By processing various types of data from storage devices through a multimodal feature fusion layer, the problem of inaccurate prediction of storage device lifespan in existing technologies is solved, achieving more accurate lifespan prediction and avoiding data loss and interruption caused by device failure.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- AXD (ANXINDA) MEMORY TECH CO LTD
- Filing Date
- 2025-06-12
- Publication Date
- 2026-07-21
AI Technical Summary
Existing methods for predicting the lifespan of storage devices are based on electrical data, temperature data, or error logs, which have low accuracy and are difficult to meet the needs of accurate prediction in practical applications.
The multimodal lifespan prediction model fuses internal structural images, sound data, electrical data, temperature data, and error monitoring data to calculate the predicted lifespan of the storage device using a multimodal feature fusion layer and a lifespan prediction decision layer.
It improves the accuracy of storage device lifespan prediction, enabling early prediction of device failures and preventing data loss and business interruptions.
Smart Images
Figure CN120611362B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of next-generation information network industry technology, specifically to a method for predicting the lifespan of storage devices, electronic devices, and storage media. Background Technology
[0002] In today's rapidly developing information technology landscape, storage devices, as core components for data storage and management, are widely used in cloud computing data centers, enterprise server clusters, personal computers, and various mobile smart devices. With the explosive growth of data volume, the workload of storage devices is increasing daily. Therefore, predicting the lifespan of storage devices is crucial for ensuring data security, optimizing equipment maintenance strategies, and reducing operating costs. For example, in large data centers, if the lifespan of storage devices can be predicted in advance, backup storage devices can be replaced earlier, avoiding data loss and business interruptions due to sudden storage device failures. Currently, existing methods for predicting the lifespan of storage devices rely solely on electrical data, temperature data, or error logs. All of this data can be obtained by reading the storage device.
[0003] However, the aging of storage devices is a complex process, affected by a combination of factors such as changes in internal structure, mechanical wear and electrical performance degradation. Predicting lifespan based solely on electrical data, temperature data or error logs of storage devices will result in low accuracy of the predictions and will be difficult to meet the needs of accurate prediction of storage device lifespan in practical applications. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this application provides a method for predicting the lifespan of storage devices, an electronic device, and a storage medium. By fusing internal structural images, sound data, electrical data, temperature data, and error monitoring data in the multimodal feature fusion layer of a multimodal lifespan prediction model, a fused feature vector is obtained. This fused feature vector is then input into the lifespan prediction decision layer to obtain the predicted lifespan of the storage device. This method can calculate the predicted lifespan of the storage device based on multimodal data, thereby improving the accuracy of the prediction and meeting the practical application requirements for precise prediction of the lifespan of storage devices.
[0005] To address the above problems, the present invention provides the following technical solution: In a first aspect, embodiments of this application provide a method for predicting the lifespan of a storage device, comprising: controlling a scanning device to scan the storage device to obtain an image of the internal structure of the storage device; The sound monitoring device is controlled to monitor the sound of the storage device, and the sound data of the storage device during operation is obtained. Acquire electrical data, temperature data, and error monitoring data of the storage device. The electrical data includes voltage and current curves when the storage device is operating. The temperature data includes the operating temperature curve when the storage device is operating. The error monitoring data includes hard disk stop failure records, write error records, erase error records, read error records, and data verification error records. The internal structure image, the sound data, the electrical data, the temperature data, and the error monitoring data are input into the trained multimodal lifespan prediction model; In the multimodal lifespan prediction model, multimodal feature fusion is performed on the internal structure image, the sound data, the electrical data, the temperature data, and the error monitoring data to obtain a fused feature vector. The fused feature vector is then input into the lifespan prediction decision layer to obtain the predicted lifespan of the storage device. The multimodal lifespan prediction model includes a multimodal data preprocessing layer, a multimodal feature fusion layer, a lifespan prediction decision layer, and a calibration layer. The multimodal data preprocessing layer is used to preprocess the internal structure image, the sound data, the electrical data, the temperature data, and the error monitoring data, and extract feature vectors corresponding to multiple data. The multimodal feature fusion layer includes a reliability weight calculation module, a multimodal feature alignment module, an error type association module, and a fusion operation module. The reliability weight calculation module is used to calculate the weights corresponding to the internal structure image, the sound data, the electrical data, the temperature data, and the error monitoring data. The multimodal feature alignment module is used to map multiple input feature vectors to a semantic space and obtain multiple aligned feature vectors. The error type association module is used to associate the multiple aligned feature vectors with error types to obtain multiple associated feature vectors. The fusion operation module is used to perform multimodal feature fusion based on the weights corresponding to the multiple data output by the reliability weight calculation module and the multiple associated feature vectors output by the error type association module to obtain the fused feature vector. The lifetime prediction decision layer is used to calculate the initial predicted lifetime of the storage device based on the fused feature vector; The calibration layer is used to calibrate the initial predicted lifespan of the storage device to obtain the final predicted lifespan.
[0006] In some implementations, the reliability weight calculation module includes a spatiotemporal consistency calculation unit, a prior weight determination unit, and a weight calculation function unit. The spatiotemporal consistency calculation unit is used to calculate the spatiotemporal consistency index corresponding to multiple data points. The prior weight determination unit is used to determine the prior weights corresponding to multiple data points based on multiple physical models. The weight calculation function unit is used to calculate the weight corresponding to each data point based on the spatiotemporal consistency index and prior weights corresponding to multiple data points using a preset weight calculation function. The weight calculation function includes pre-trained adjustment parameters.
[0007] In some implementations, the spatiotemporal consistency calculation unit is used to divide the data into multiple data segments according to a preset time period when the data is two-dimensional data, extract the feature vector of each data segment and calculate the similarity of the feature vectors of two adjacent data segments in each group to obtain a time similarity sequence, and calculate the spatiotemporal consistency index of the data based on the time similarity sequence. The spatiotemporal consistency calculation unit is used to divide the data into multiple spatial units according to a preset spatial unit size when the data is three-dimensional data, extract the feature vector of each spatial unit and calculate the similarity of the feature vectors of two adjacent spatial units in each group to obtain a spatial similarity sequence, and calculate the spatiotemporal consistency index of the data based on the spatial similarity sequence.
[0008] In some implementations, the error type association module includes an error type feature library and a graph neural network unit. The error type feature library includes error types, corresponding feature vectors, associated data types, and physical models. The graph neural network unit is used to construct a multimodal feature interaction graph based on the multiple aligned feature vectors, using the feature vectors in the error type feature library as edge weights. A node in the multimodal feature interaction graph is one of the aligned feature vectors. The graph neural network unit is also used to aggregate the information of adjacent nodes in the multimodal feature interaction graph using graph convolutional layers and update the alignment feature vector represented by each node to obtain multiple associated feature vectors, wherein the associated feature vectors contain constraint information of error type.
[0009] In some implementations, the lifetime prediction decision layer is further configured to input the fused feature vector into a fully connected layer, and input the vector output by the fully connected layer into an activation layer to obtain the probability of each error type occurring in the storage device output by the activation layer.
[0010] In some implementations, the calibration layer is further configured to obtain the device type, usage scenario type, and usage duration of the storage device, obtain the corresponding calibration coefficient from the calibration error information based on the device type, usage scenario type, and usage duration, and calibrate the initial predicted lifespan of the storage device based on the calibration coefficient to obtain the final predicted lifespan.
[0011] In some implementations, the calibration error information is a calibration error table that represents a multidimensional mapping relationship between device type, usage scenario type, usage duration, and calibration coefficient, determined based on training data. In the calibration error table, a combination of device type, usage scenario type, and usage duration interval corresponds to a calibration coefficient.
[0012] In some embodiments, the method further includes: The 3D scanning device is controlled to scan the storage device to obtain the 3D point cloud data of the storage device; The step of inputting the internal structure image, the sound data, the electrical data, the temperature data, and the error monitoring data into the pre-trained multimodal lifespan prediction model includes: The internal structure image, the sound data, the electrical data, the temperature data, the error monitoring data, and the three-dimensional point cloud data are input into the pre-trained multimodal lifespan prediction model.
[0013] Secondly, embodiments of this application provide an electronic device, the electronic device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the storage device lifetime prediction method as described in the first aspect.
[0014] Thirdly, embodiments of this application provide a computer-readable storage medium storing an executable program, which is executed by a processor to implement the storage device lifetime prediction method as described in the first aspect.
[0015] This application provides a method for predicting the lifespan of a storage device, an electronic device, and a storage medium. This application achieves this by fusing internal structural images, sound data, electrical data, temperature data, and error monitoring data in the multimodal feature fusion layer of a multimodal lifespan prediction model to obtain a fused feature vector. This fused feature vector is then input into the lifespan prediction decision layer to obtain the predicted lifespan of the storage device. This method can calculate the predicted lifespan of the storage device based on multimodal data, thereby improving the accuracy of the prediction and meeting the practical application requirements for precise prediction of the lifespan of storage devices. Attached Figure Description
[0016] Figure 1 This is a flowchart illustrating the lifespan prediction method for storage devices provided in this application embodiment.
[0017] Figure 2 This is a schematic diagram of the structure of the multimodal lifetime prediction model provided in the embodiments of this application.
[0018] Figure 3 This is a schematic diagram of the structure of the multimodal feature fusion layer provided in the embodiments of this application.
[0019] Figure 4 This is a schematic diagram of the reliability weight calculation module provided in the embodiments of this application.
[0020] Figure 5 This is a schematic diagram of the storage device lifespan prediction device provided in the embodiments of this application.
[0021] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.
[0022] Figure 7 This is a structural block diagram of a computer-readable storage medium provided in an embodiment of this application. Detailed Implementation
[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0024] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "multiple" means two or more, unless otherwise explicitly specified.
[0025] This application provides a method for predicting the lifespan of a storage device, an electronic device, and a storage medium. By fusing internal structural images, sound data, electrical data, temperature data, and error monitoring data in the multimodal feature fusion layer of a multimodal lifespan prediction model, a fused feature vector is obtained. This fused feature vector is then input into the lifespan prediction decision layer to obtain the predicted lifespan of the storage device. The predicted lifespan of the storage device can be calculated based on multimodal data, thereby improving the accuracy of the prediction and meeting the need for accurate prediction of the lifespan of storage devices in practical applications.
[0026] The storage device involved in this application can be a hard disk drive (HDD), a solid-state drive (SSD), an optical disc, or a memory card, etc. The SSD can be a ball grid array (BGA) SSD. A BGA SSD may include flash memory chips, which may include NOR Flash and NAND Flash, etc. The following example uses a BGA SSD as the storage device, but this should not be considered a limitation of this application.
[0027] The storage device described in this application is a computer component; therefore, this application relates to the manufacture of computer components. The storage device described in this application can be used in databases and cloud computing centers; therefore, this application also relates to the fields of the Internet and cloud computing, big data services, specifically databases and cloud database services.
[0028] The storage device lifespan prediction method in this application is used to predict the lifespan of storage devices during preparation for computation or during idle periods. This allows staff to replace storage devices with backups in advance based on the predicted lifespan, avoiding data loss and business interruption due to sudden storage device failure. Essentially, it involves testing to detect or locate faulty hardware.
[0029] The lifespan prediction method for storage devices provided in this application will be described in detail below with reference to the accompanying drawings.
[0030] Please see Figure 1 , Figure 1 This is a flowchart illustrating the storage device lifetime prediction method provided in an embodiment of this application. Figure 1As shown, the method for predicting the lifespan of the storage device includes steps S100 to S500.
[0031] Step S100: Control the scanning device to scan the storage device and obtain an image of the internal structure of the storage device.
[0032] In some implementations, the internal structure image of the storage device is an X-ray image, ultrasound image, or magnetic resonance image, etc.
[0033] Preferably, the internal structure image of the storage device is an ultrasonic image. In this case, the scanning device is an ultrasonic imaging device, which is small in size and emits no radiation, making it convenient for scanning operations.
[0034] In some implementations, the internal structure image of the storage device may include multiple images obtained by continuous scanning of the storage device, or it may be a single image.
[0035] Step S200: Control the sound monitoring device to monitor the sound of the storage device and obtain the sound data when the storage device is working.
[0036] When the storage device is a hard disk drive (HDD), the sound it makes during operation is related to the head's seek operation or the disk's rotation speed. HDDs typically produce a regular sound during normal operation; abnormal noises can indicate head wear, disk scratches, or bearing failure, which will affect the HDD's lifespan. When the storage device is an SSD, abnormal electrical noises can indicate capacitor or controller failure. Therefore, the lifespan of a storage device can be predicted based on its operating sound data.
[0037] Step S300: Acquire electrical data, temperature data, and error monitoring data of the storage device.
[0038] The electrical data includes voltage and current curves during storage device operation. The temperature data includes the operating temperature curve of the storage device. The error monitoring data includes error records such as hard drive stop failure records, write error records, erase error records, read error records, and data verification error records.
[0039] Optionally, the temperature data may also include an ambient temperature profile of the storage device during operation.
[0040] Step S400: Input the internal structure image, sound data, electrical data, temperature data and error monitoring data into the trained multimodal lifespan prediction model.
[0041] Step S500: In the multimodal lifespan prediction model, multimodal feature fusion is performed on internal structure images, sound data, electrical data, temperature data and error monitoring data to obtain a fused feature vector. The fused feature vector is then input into the lifespan prediction decision layer to obtain the predicted lifespan of the storage device.
[0042] Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of the multimodal lifetime prediction model provided in the embodiments of this application. Figure 2 As shown, in some embodiments, the multimodal lifetime prediction model 1 includes a multimodal data preprocessing layer 10, a multimodal feature fusion layer 20, a lifetime prediction decision layer 30, and a calibration layer 40.
[0043] In some implementations, the multimodal data preprocessing layer 10 is used to preprocess internal structure images, sound data, electrical data, temperature data and error monitoring data, extract feature vectors corresponding to multiple data, and input the feature vectors corresponding to multiple data into the multimodal feature fusion layer 20.
[0044] In some implementations, the multimodal data preprocessing layer includes an image preprocessing module and an image feature extraction module. The image preprocessing module is used to perform image processing operations such as enhancement and denoising on the internal structure image, while the image feature extraction module is used to extract the feature vector corresponding to the processed image.
[0045] In some implementations, the image feature extraction module is a convolutional neural network model. The convolutional neural network model can extract feature vectors corresponding to the processed image through convolution operations, pooling operations, and fully connected layers.
[0046] As described above, the internal structure image of the storage device can be obtained from multiple images obtained by continuously scanning the storage device. In this case, the image feature extraction module can include a trunk branch module, a detail branch module, and a result vector module. The trunk branch module is used to extract two-dimensional feature vectors from the multiple internal structure images, and the detail branch module is used to extract three-dimensional feature vectors based on the multiple internal structure images. The result vector module is used to concatenate the two-dimensional and three-dimensional feature vectors to obtain the feature vector corresponding to the internal structure image.
[0047] In some implementations, the result vector module is also used to reduce the dimensionality of the vector obtained by concatenating the two-dimensional feature vector and the three-dimensional feature vector to obtain the feature vector corresponding to the internal structure image.
[0048] In some implementations, when the internal structure image is a single image, the feature extraction module may include only the trunk branch module and exclude the detail branch module. In this case, the trunk branch module is used to extract the two-dimensional feature vector of the single internal structure image.
[0049] In some implementations, the backbone module includes multiple convolutional layers and fully connected layers connected in sequence. The backbone module is used to progressively extract two-dimensional features from the internal structure image across multiple convolutional layers to obtain a feature map. These two-dimensional features may include the morphological features of chip solder joints and the appearance features of capacitors, etc. The fully connected layers are used to convert the feature maps output by the convolutional layers into two-dimensional feature vectors.
[0050] In some implementations, the detail branch module can be a 3D image neural network model. The detail branch module includes multiple sequentially connected 3D convolutional layers, multiple 3D pooling layers, and multiple fully connected layers. The 3D convolutional layers and multiple 3D pooling layers are used to extract 3D geometric features based on multiple internal structure images. The 3D geometric features may include 3D geometric features between stacked storage unit layers.
[0051] In some implementations, the multimodal data preprocessing layer further includes an audio data preprocessing module and an audio feature extraction module. The audio data preprocessing module performs spectral analysis on the audio data to obtain spectral data. The audio feature extraction module extracts audio signals whose frequencies fall within any one of multiple preset frequency bands from the audio data based on the spectral data, thereby obtaining audio signals corresponding to multiple preset frequency bands, and extracts the feature vector of each audio signal to obtain the feature vector corresponding to the audio data.
[0052] Optionally, multiple preset frequency bands include normal operating sound frequency bands and various types of abnormal operating sound frequency bands. These various types of abnormal operating sound frequency bands may include overload operating sound frequency bands and short-circuit operating sound frequency bands, etc. When the storage device is a mechanical hard drive, the various types of abnormal operating sound frequency bands may also include operating sound frequency bands for head wear, platter scratches, and bearing failure, etc.
[0053] In some implementations, the multimodal data preprocessing layer further includes an electrical data preprocessing module. This module extracts feature vectors from the voltage and current curves in the electrical data and concatenates these feature vectors into a corresponding feature vector for the electrical data.
[0054] In some implementations, the electrical data preprocessing module is a temporal recurrent neural network (TRRN) model. A TRRN model is a type of recurrent neural network specifically designed for processing time-series data. By introducing recurrent connections between hidden layers, the TRRN model allows information to be transferred between different time steps. This design enables the TRRN model to capture dependencies and patterns in the time series.
[0055] Optionally, the eigenvectors of the voltage and current curves include periodic fluctuation eigenvectors and abnormal fluctuation eigenvectors.
[0056] Optionally, in the time recurrent neural network model, a sliding window of preset length is used to extract the periodic fluctuation feature vectors of the pressure curve and the current curve.
[0057] Optionally, the preset length is the median of the erase / write cycles of the storage device.
[0058] For example, a temporal recurrent neural network model can be a long short-term memory network model.
[0059] In some implementations, the multimodal data preprocessing layer further includes a temperature data preprocessing module, which is used to calculate the self-heating temperature curve of the storage device based on the ambient temperature curve and the operating temperature curve in the temperature data, and extract the feature vector corresponding to the temperature data based on the self-heating temperature curve.
[0060] In some implementations, in the temperature data preprocessing module, for each time point, the temperature value of the operating temperature curve at that time point is subtracted from the temperature value of the ambient temperature curve at the same time point, and the difference is the self-heating temperature of the storage device at that time point, thereby obtaining the self-heating temperature curve of the storage device.
[0061] In some implementations, the multimodal data preprocessing layer further includes an error monitoring data preprocessing module. This module calculates a feature vector corresponding to the error monitoring data based on the type and recording time of the error records. The feature vector representing the error monitoring data indicates the frequency of various errors occurring in the storage device during different usage periods.
[0062] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of the multimodal feature fusion layer provided in an embodiment of this application. For example... Figure 3 As shown, in some embodiments, the multimodal feature fusion layer 20 includes a reliability weight calculation module 21, a multimodal feature alignment module 22, an error type association module 23, and a fusion operation module 24.
[0063] The reliability weight calculation module 21 calculates the weights corresponding to the internal structure image, sound data, electrical data, temperature data, and error monitoring data. The multimodal feature alignment module 22 maps multiple input feature vectors to the semantic space to obtain multiple aligned feature vectors. The error type association module 23 associates the multiple aligned feature vectors with error types to obtain multiple associated feature vectors. The fusion operation module 24 performs multimodal feature fusion based on the weights corresponding to the multiple data output by the reliability weight calculation module 21 and the multiple associated feature vectors output by the error type association module 23 to obtain a fused feature vector.
[0064] Please see Figure 4 , Figure 4 This is a schematic diagram of the reliability weight calculation module provided in an embodiment of this application. Figure 4 As shown, in some embodiments, the reliability weight calculation module 21 includes a spatiotemporal consistency calculation unit 211, a priori weight determination unit 212, and a weight calculation function unit 213.
[0065] The spatiotemporal consistency calculation unit 211 calculates the spatiotemporal consistency index corresponding to multiple data points. The prior weight determination unit 212 determines the prior weights corresponding to multiple data points based on various physical models. The weight calculation function unit 213 calculates the weight corresponding to each data point using a preset weight calculation function, based on the spatiotemporal consistency index and prior weights corresponding to multiple data points. The weight calculation function includes pre-trained adjustment parameters. The multiple data points include internal structure images, sound data, electrical data, temperature data, and error monitoring data.
[0066] In some implementations, for two-dimensional data, the spatiotemporal consistency index can be used to reflect the similarity of features at adjacent time points in the time dimension. For three-dimensional data, the spatiotemporal consistency index can be used to reflect the similarity of features in different regions in the spatial dimension. The higher the index value, the better the spatiotemporal consistency of the modality data, and the higher the reliability of the data.
[0067] Alternatively, electrical data, sound data, temperature data, and error monitoring data can be categorized as two-dimensional data.
[0068] Alternatively, when the internal structure image comprises multiple images, the internal structure image can be classified as three-dimensional data.
[0069] In some implementations, the spatiotemporal consistency calculation unit is used to divide the data into multiple data segments according to a preset time period when the data is two-dimensional data, extract the feature vector of each data segment and calculate the similarity of the feature vectors of two adjacent data segments in each group to obtain a time similarity sequence, and calculate the spatiotemporal consistency index of the data based on the time similarity sequence.
[0070] Optionally, the cosine similarity of the feature vectors of two adjacent data segments in each group is calculated to obtain a time similarity sequence including multiple cosine similarities, and the average value of the cosine similarity in the time similarity sequence is used as the spatiotemporal consistency index of the data.
[0071] Optionally, when the internal structure image includes multiple images, the spatiotemporal consistency calculation unit is also used to obtain the internal structure model of the storage device based on the multiple internal structure images. The internal structure model of the storage device is three-dimensional data. When the data is three-dimensional, the spatiotemporal consistency calculation unit is used to divide the data into multiple spatial units according to a preset spatial unit size, extract the feature vector of each spatial unit, calculate the similarity of the feature vectors of each pair of adjacent spatial units, obtain a spatial similarity sequence, and calculate the spatiotemporal consistency index of the data based on the spatial similarity sequence.
[0072] Optionally, the cosine similarity of the feature vectors of two adjacent spatial units in each group is calculated to form a spatial similarity sequence including multiple cosine similarities, and the average value of the cosine similarity in the spatial similarity sequence is used as the spatiotemporal consistency index of the data.
[0073] In some implementations, the prior weight determination unit incorporates multiple physical models, which may include models pre-determined based on physical principles to describe the relationship between memory cell failure and temperature and time, and models to describe the relationship between solder joint fatigue and temperature, etc.
[0074] In some implementations, the prior weight determination unit is used to determine the usage scenario type of the storage device based on at least one of multiple data points, and to determine the prior weights corresponding to the multiple data points according to the physical model corresponding to the usage scenario type. For example, when it is determined from temperature data that the operating temperature of the storage device is higher than a preset temperature, the usage scenario type of the storage device is determined to be a high-temperature scenario, and the model corresponding to the high-temperature scenario, which describes the relationship between storage cell failure and temperature and time, is selected to determine the prior weights corresponding to the multiple data points. In a high-temperature scenario, the prior weights corresponding to the temperature data are higher than the prior weights corresponding to the general scenario.
[0075] In some implementations, the weight calculation function is: ,in, Indicates the first The weights corresponding to each data point Indicates the adjustment parameter. Indicates the first The spatiotemporal consistency index corresponding to each data point Indicators representing historical average spatiotemporal consistency. Indicates the first Prior weights of each data point.
[0076] Optionally, and It can be obtained by training a multimodal lifetime prediction model in advance.
[0077] In some implementations, the adjustment parameter controls the weight pair. and The sensitivity to differences between them. The larger the value of the adjustment parameter, the steeper the function curve, meaning that when... and When the differences are slight, the weights will change significantly.
[0078] In some implementations, the multimodal feature alignment module is used to map the feature vectors corresponding to multiple data points to the semantic space and obtain multiple aligned feature vectors, each with the same spatial dimension, for subsequent calculation.
[0079] In some implementations, due to the different sampling rates of multiple data points, there may be a misalignment issue in the time dimension. In this case, the multimodal feature alignment module is also used to map the feature vectors corresponding to multiple data points to the semantic space, and then use a dynamic time warping algorithm to elastically match the time series of the data, so that the two feature vectors can be aligned on the time axis through scaling and translation, thereby achieving alignment of each feature vector in the time dimension.
[0080] In some implementations, the error type association module 23 includes an error type feature library and a graph neural network unit. The error type feature library includes error types, corresponding feature vectors, associated data types, and physical models. The graph neural network unit is used to construct a multimodal feature interaction graph based on multiple aligned feature vectors, using feature vectors from the error type feature library as edge weights. A node in the modal feature interaction graph represents an aligned feature vector.
[0081] In some implementations, error types include memory cell failure, open circuit, short circuit, solder joint fatigue, and exceeding a safe temperature threshold.
[0082] For example, the data types associated with storage cell failure include error monitoring data. The data types associated with open circuits include electrical data, error monitoring data, and internal structure images. The data types associated with short circuits include electrical data, error monitoring data, and internal structure images. The data types associated with solder joint fatigue include electrical data and internal structure images. The data types associated with exceeding a safe temperature threshold include temperature data and sound data, because the operating noise of the storage device tends to be higher when the safe temperature threshold is exceeded.
[0083] Optionally, the graph neural network unit is also used to aggregate information from adjacent nodes in the multimodal feature interaction graph using graph convolutional layers and update the aligned feature vector represented by each node to obtain multiple associated feature vectors. These associated feature vectors contain constraint information about error types. In this way, the fusion of multimodal features can be initially achieved. Furthermore, the associated feature vectors, containing constraint information about error types, can be calculated based on prior knowledge, thus ensuring that the output of the multimodal lifetime prediction model conforms to physical laws.
[0084] Optionally, there are multiple graph convolutional layers, which are connected sequentially. Each graph convolutional layer updates the aligned feature vector represented by each node by aggregating information from neighboring nodes. In some implementations, the fusion operation module is used to perform multimodal feature fusion based on the weights corresponding to multiple data points output by the reliability weight calculation module and the multiple associated feature vectors output by the error type association module, to obtain a fused feature vector.
[0085] Optionally, the fusion operation module is used to perform a weighted summation of the associated feature vectors corresponding to each data point based on the weights corresponding to each data point, thereby obtaining a fused feature vector. In this way, the fusion of multimodal features can be ultimately achieved.
[0086] In some implementations, the lifetime prediction decision layer is used to calculate the initial predicted lifetime of the storage device based on the fused feature vectors.
[0087] Optionally, the lifetime prediction decision layer includes a main task regression module, which includes multiple fully connected layers and activation functions connected in sequence. The first fully connected layer of the main task regression module is used to input the fused feature vector and perform calculations, and the activation function of the main task regression module is used to determine the initial predicted lifetime of the storage device based on the output of the last fully connected layer.
[0088] In some implementations, the lifetime prediction decision layer is also used to input the fused feature vector into the fully connected layer, input the vector output by the fully connected layer into the activation layer, and obtain the probability of each error type occurring in the storage device from the output of the activation layer.
[0089] Optionally, the lifetime prediction decision layer is a multi-task learning architecture, which also includes an auxiliary task module. The auxiliary task module includes multiple fully connected layers and activation layers connected in sequence. The first fully connected layer of the auxiliary task module is used to input the fused feature vector and perform calculations. The activation layer of the auxiliary task module is used to determine the probability of each error type occurring in the storage device based on the output of the last fully connected layer.
[0090] In some implementations, when training a multimodal lifetime prediction model, the loss function used by the main task regression module of the lifetime prediction decision layer is the mean square error loss function or the root mean square error loss function, etc.
[0091] In some implementations, the cross-entropy loss function is used as the auxiliary task module of the lifetime prediction decision layer when training the multimodal lifetime prediction model. The cross-entropy loss function is used to evaluate the difference between the probability distribution of the probability of each error type occurring in the storage device during the current training and the true distribution.
[0092] Optionally, the fully connected layer includes multiple groups of neurons, each group of neurons corresponding to an error type.
[0093] As described above, in some implementations, error types include memory cell failure, open circuit, short circuit, solder joint fatigue, and exceeding safe temperature thresholds. This allows for the interpretation of the multimodal lifetime prediction model's calculations, enabling users to understand which errors are affecting the lifespan of the storage device.
[0094] In some implementations, the calibration layer is used to calibrate the initial predicted lifespan of the storage device to obtain the final predicted lifespan.
[0095] In some implementations, the calibration layer is also used to obtain the device type, usage scenario type, and usage duration of the storage device; to obtain the corresponding calibration coefficients from the calibration error information based on the device type, usage scenario type, and usage duration; and to calibrate the initial predicted lifespan of the storage device based on the calibration coefficients to obtain the final predicted lifespan. This approach can further improve the accuracy of the prediction.
[0096] Optionally, log data from the storage device can be obtained, and the device type and usage duration of the storage device can be determined based on the log data. For example, it can be determined that the storage device is an SSD and has been used for 1000 hours.
[0097] Optionally, the use case type can be determined by the prior weight determination unit based on at least one of multiple data, or it can be obtained according to the received user instructions.
[0098] In some implementations, the calibration error information is a calibration error table that represents a multidimensional mapping relationship between device type, usage scenario type, usage duration, and calibration coefficient, determined based on training data.
[0099] Optionally, in the calibration error table, a combination of device type, usage scenario type, and duration interval of used time corresponds to a calibration coefficient.
[0100] In some implementations, the training data is divided into three dimensions according to device type, usage scenario type, and usage duration, and a calibration error table is pre-trained based on the partial training data corresponding to the combination of duration intervals of each device type, usage scenario type, and usage duration.
[0101] Optionally, the duration interval can be preset or obtained through training.
[0102] For example, when the storage device is an SSD, the usage scenario type is a general scenario, and the usage time is 2000 hours, the usage time range can be determined to be 1500 hours to 2500 hours according to the calibration error table, and the calibration coefficient corresponding to the combination of the current device type, usage scenario type and usage time range is 0.1.
[0103] In some implementations, the initial calibration formula for the predicted service life is: ,in, This indicates the final predicted service life. This indicates the initial predicted service life. Indicates the calibration coefficient. It represents the base of the natural logarithm.
[0104] in, It can be any value. The values can be obtained through training.
[0105] In some implementations, when training a multimodal lifetime prediction model, the calibration layer uses a loss function such as mean squared error loss function, logarithmic loss function, or mean absolute error loss function.
[0106] In some implementations, the method further includes step S600.
[0107] Step S600: Control the 3D scanning device to scan the storage device and obtain the 3D point cloud data of the storage device.
[0108] 3D point cloud data can be classified as 3D data.
[0109] At this point, step S400 includes: inputting the internal structure image, sound data, electrical data, temperature data, error monitoring data, and three-dimensional point cloud data into the trained multimodal lifespan prediction model.
[0110] The processing of multiple data in the multimodal lifetime prediction model is described in step S500.
[0111] In summary, the storage device lifetime prediction method provided in this application has the following advantages: 1. By fusing internal structural images, sound data, electrical data, temperature data, and error monitoring data in the multimodal feature fusion layer of the multimodal lifetime prediction model, a fused feature vector is obtained. This fused feature vector is then input into the lifetime prediction decision layer to obtain the predicted lifetime of the storage device. The predicted lifetime of the storage device can be calculated based on multimodal data, thereby improving the accuracy of the prediction and meeting the need for accurate prediction of the lifetime of storage devices in practical applications.
[0112] 2. The graph neural network unit is also used to aggregate information of adjacent nodes in the multimodal feature interaction graph using graph convolutional layers and update the aligned feature vector represented by each node to obtain multiple associated feature vectors, which can initially achieve the fusion of multimodal features. Furthermore, it can calculate associated feature vectors based on prior knowledge. These associated feature vectors contain constraint information about error types, thus ensuring that the output of the multimodal lifespan prediction model conforms to physical laws.
[0113] 3. The lifetime prediction decision layer is also used to input the fused feature vector into the fully connected layer, and input the vector output by the fully connected layer into the activation layer to obtain the probability of each type of error occurring in the storage device output by the activation layer. This can explain the calculation results of the multimodal lifetime prediction model, so that users know which errors affect the lifetime of the storage device.
[0114] Please see Figure 5 , Figure 5 This is a schematic diagram of the storage device lifetime prediction apparatus provided in an embodiment of this application. Figure 5 As shown, the storage device lifespan prediction device 300 includes an acquisition module 310 and a processing module 320.
[0115] In some embodiments, the acquisition module 310 is used to control the scanning device to scan the storage device to obtain an image of the internal structure of the storage device; control the sound monitoring device to perform sound monitoring on the storage device to obtain sound data when the storage device is working; and acquire electrical data, temperature data and error monitoring data of the storage device. The electrical data includes the voltage curve and current curve when the storage device is working, the temperature data includes the operating temperature curve when the storage device is working, and the error monitoring data includes hard disk stop failure records, write error records, erase error records, read error records and data verification error records.
[0116] In some implementations, the processing module 320 is used to input internal structure images, sound data, electrical data, temperature data, and error monitoring data into a pre-trained multimodal lifespan prediction model; in the multimodal feature fusion layer of the multimodal lifespan prediction model, multimodal feature fusion is performed on the internal structure images, sound data, electrical data, temperature data, and error monitoring data to obtain a fused feature vector, and the fused feature vector is input into the lifespan prediction decision layer to obtain the predicted lifespan of the storage device. The multimodal lifespan prediction model includes a multimodal data preprocessing layer, a multimodal feature fusion layer, a lifespan prediction decision layer, and a calibration layer. The multimodal data preprocessing layer is used to preprocess the internal structure images, sound data, electrical data, temperature data, and error monitoring data, and extract feature vectors corresponding to multiple data points. The multimodal feature fusion layer includes a reliability weight calculation module. The system comprises a block, a multimodal feature alignment module, an error type association module, and a fusion operation module. The reliability weight calculation module calculates the weights corresponding to internal structure images, sound data, electrical data, temperature data, and error monitoring data. The multimodal feature alignment module maps multiple input feature vectors to the semantic space to obtain multiple aligned feature vectors. The error type association module associates multiple aligned feature vectors with error types to obtain multiple associated feature vectors. The fusion operation module performs multimodal feature fusion based on the weights corresponding to multiple data points output by the reliability weight calculation module and the multiple associated feature vectors output by the error type association module to obtain a fused feature vector. The lifetime prediction decision layer calculates the initial predicted lifetime of the storage device based on the fused feature vector. The calibration layer calibrates the initial predicted lifetime of the storage device to obtain the final predicted lifetime.
[0117] Please see Figure 6 , Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. For example... Figure 6 As shown, the electronic device 400 includes: one or more processors 410 and a memory 420. Figure 6 Take a processor 410 as an example.
[0118] In some implementations, the processor 410 and the memory 420 may be connected via a bus or other means. Figure 6 Taking the example of a connection between China and Israel via a bus.
[0119] In some embodiments, the processor 410 is configured to control a scanning device to scan the storage device to obtain an image of the internal structure of the storage device; control a sound monitoring device to monitor the sound of the storage device to obtain sound data during operation; acquire electrical data, temperature data, and error monitoring data of the storage device, wherein the electrical data includes voltage and current curves during operation, the temperature data includes operating temperature curves during operation, and the error monitoring data includes hard disk stop failure records, write error records, erase error records, read error records, and data verification error records; input the internal structure image, sound data, electrical data, temperature data, and error monitoring data into a pre-trained multimodal lifetime prediction model; perform multimodal feature fusion on the internal structure image, sound data, electrical data, temperature data, and error monitoring data in the multimodal feature fusion layer of the multimodal lifetime prediction model to obtain a fused feature vector, and input the fused feature vector into the lifetime prediction decision layer to obtain the predicted lifetime of the storage device. The multimodal lifetime prediction model includes a multimodal data preprocessing layer, a multimodal feature fusion layer, and a lifetime prediction layer. The system comprises a decision layer and a calibration layer; a multimodal data preprocessing layer for preprocessing internal structure images, sound data, electrical data, temperature data, and error monitoring data, and extracting feature vectors corresponding to multiple data points; a multimodal feature fusion layer including a reliability weight calculation module, a multimodal feature alignment module, an error type association module, and a fusion operation module; the reliability weight calculation module for calculating the weights corresponding to internal structure images, sound data, electrical data, temperature data, and error monitoring data; the multimodal feature alignment module for mapping multiple input feature vectors to a semantic space and obtaining multiple aligned feature vectors; the error type association module for associating multiple aligned feature vectors with error types and obtaining multiple associated feature vectors; and the fusion operation module for performing multimodal feature fusion based on the weights corresponding to multiple data points output by the reliability weight calculation module and the multiple associated feature vectors output by the error type association module, resulting in a fused feature vector; a lifetime prediction decision layer for calculating the initial predicted lifetime of the storage device based on the fused feature vector; and a calibration layer for calibrating the initial predicted lifetime of the storage device to obtain the final predicted lifetime.
[0120] In some embodiments, memory 420 serves as a non-volatile computer-readable storage medium, used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules for the storage device lifetime prediction method in the embodiments of this application. Processor 410 executes various functional applications and data processing of electronic device 400 by running the non-volatile software programs, instructions, and modules stored in memory 420, thereby implementing the storage device lifetime prediction method of the above-described method embodiments.
[0121] In some embodiments, memory 420 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of electronic device 400, etc. Furthermore, memory 420 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, memory 420 may optionally include memory remotely located relative to processor 410, and this remote memory may be connected to the controller via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.
[0122] In some implementations, one or more modules are stored in memory 420, and when executed by one or more processors 410, they perform the storage device lifetime prediction method in any of the above method embodiments, for example, performing the above-described... Figure 1 The method steps S100 to S500.
[0123] Please refer to Figure 7 , Figure 7 This is a structural block diagram of a computer-readable storage medium provided in an embodiment of this application. The computer-readable storage medium 500 stores program code 510, which can be called by a processor to execute the storage device lifetime prediction method described in the above method embodiments.
[0124] The computer-readable storage medium 500 may be an electronic storage device such as flash memory, electrically erasable programmable read-only memory (EEPROM), hard disk, or read-only memory (ROM). Optionally, the computer-readable storage medium includes a non-volatile computer-readable medium. The computer-readable storage medium 500 has storage space for program code that performs any of the method steps of the above-described method for predicting the lifetime of the storage device. This program code can be read from or written to one or more computer program products. The program code may, for example, be compressed in a suitable form.
[0125] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for predicting the lifespan of a storage device.
[0126] In summary, this application provides a method for predicting the lifespan of a storage device, an electronic device, and a storage medium. The method for predicting the lifespan of a storage device includes: controlling a scanning device to scan the storage device to obtain an image of its internal structure; controlling a sound monitoring device to monitor the sound of the storage device to obtain sound data during operation; acquiring electrical data, temperature data, and error monitoring data of the storage device, wherein the electrical data includes voltage and current curves during operation, the temperature data includes operating temperature curves during operation, and the error monitoring data includes hard disk stop failure records, write error records, erase error records, read error records, and data verification error records; inputting the internal structure image, sound data, electrical data, temperature data, and error monitoring data into a pre-trained multimodal lifespan prediction model; fusing the internal structure image, sound data, electrical data, temperature data, and error monitoring data in the multimodal feature fusion layer of the multimodal lifespan prediction model to obtain a fused feature vector, and inputting the fused feature vector into a lifespan prediction decision layer to obtain the predicted lifespan of the storage device. The multimodal lifetime prediction model comprises a multimodal data preprocessing layer, a multimodal feature fusion layer, a lifetime prediction decision layer, and a calibration layer. The multimodal data preprocessing layer preprocesses internal structure images, sound data, electrical data, temperature data, and error monitoring data, extracting feature vectors corresponding to multiple data points. The multimodal feature fusion layer includes a reliability weight calculation module, a multimodal feature alignment module, an error type association module, and a fusion operation module. The reliability weight calculation module calculates the weights corresponding to the internal structure images, sound data, electrical data, temperature data, and error monitoring data. The multimodal feature alignment module maps multiple input feature vectors to a semantic space to obtain multiple aligned feature vectors. The error type association module associates multiple aligned feature vectors with error types to obtain multiple associated feature vectors. The fusion operation module performs multimodal feature fusion based on the weights of the multiple data points output by the reliability weight calculation module and the multiple associated feature vectors output by the error type association module, obtaining a fused feature vector. The lifetime prediction decision layer calculates the initial predicted lifetime of the storage device based on the fused feature vector. The calibration layer calibrates the initial predicted lifetime of the storage device to obtain the final predicted lifetime. This application integrates internal structural images, sound data, electrical data, temperature data, and error monitoring data in the multimodal feature fusion layer of a multimodal lifespan prediction model to obtain a fused feature vector. This fused feature vector is then input into the lifespan prediction decision layer to obtain the predicted lifespan of the storage device. This method can calculate the predicted lifespan of the storage device based on multimodal data, thereby improving the accuracy of the prediction and meeting the need for accurate prediction of the lifespan of storage devices in practical applications.
[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for predicting the lifespan of a storage device, characterized in that, include: The scanning device is controlled to scan the storage device to obtain an image of the internal structure of the storage device; The sound monitoring device is controlled to monitor the sound of the storage device, and the sound data of the storage device during operation is obtained. Acquire electrical data, temperature data, and error monitoring data of the storage device. The electrical data includes voltage and current curves when the storage device is operating. The temperature data includes the operating temperature curve when the storage device is operating. The error monitoring data includes hard disk stop failure records, write error records, erase error records, read error records, and data verification error records. The internal structure image, the sound data, the electrical data, the temperature data, and the error monitoring data are input into the trained multimodal lifespan prediction model; In the multimodal lifespan prediction model, multimodal feature fusion is performed on the internal structure image, the sound data, the electrical data, the temperature data, and the error monitoring data to obtain a fused feature vector. The fused feature vector is then input into the lifespan prediction decision layer to obtain the predicted lifespan of the storage device. The multimodal lifespan prediction model includes a multimodal data preprocessing layer, a multimodal feature fusion layer, a lifespan prediction decision layer, and a calibration layer. The multimodal data preprocessing layer is used to preprocess the internal structure image, the sound data, the electrical data, the temperature data, and the error monitoring data, and extract feature vectors corresponding to multiple data. The multimodal feature fusion layer includes a reliability weight calculation module, a multimodal feature alignment module, an error type association module, and a fusion operation module. The reliability weight calculation module is used to calculate the weights corresponding to the internal structure image, the sound data, the electrical data, the temperature data, and the error monitoring data. The multimodal feature alignment module is used to map multiple input feature vectors to a semantic space and obtain multiple aligned feature vectors. The error type association module is used to associate the multiple aligned feature vectors with error types to obtain multiple associated feature vectors. The fusion operation module is used to perform multimodal feature fusion based on the weights corresponding to the multiple data output by the reliability weight calculation module and the multiple associated feature vectors output by the error type association module to obtain the fused feature vector. The lifetime prediction decision layer is used to calculate the initial predicted lifetime of the storage device based on the fused feature vector; The calibration layer is used to calibrate the initial predicted lifespan of the storage device to obtain the final predicted lifespan.
2. The method for predicting the lifespan of a storage device according to claim 1, characterized in that, The reliability weight calculation module includes a spatiotemporal consistency calculation unit, a prior weight determination unit, and a weight calculation function unit. The spatiotemporal consistency calculation unit is used to calculate the spatiotemporal consistency index corresponding to multiple data points. The prior weight determination unit is used to determine the prior weights corresponding to multiple data points based on multiple physical models. The weight calculation function unit is used to calculate the weight corresponding to each data point based on the spatiotemporal consistency index and prior weights corresponding to multiple data points using a preset weight calculation function. The weight calculation function includes pre-trained adjustment parameters.
3. The method for predicting the lifespan of a storage device according to claim 2, characterized in that, The spatiotemporal consistency calculation unit is used to divide the data into multiple data segments according to a preset time period when the data is two-dimensional data, extract the feature vector of each data segment and calculate the similarity of the feature vectors of two adjacent data segments in each group to obtain a time similarity sequence, and calculate the spatiotemporal consistency index of the data based on the time similarity sequence. The spatiotemporal consistency calculation unit is used to divide the data into multiple spatial units according to a preset spatial unit size when the data is three-dimensional data, extract the feature vector of each spatial unit and calculate the similarity of the feature vectors of two adjacent spatial units in each group to obtain a spatial similarity sequence, and calculate the spatiotemporal consistency index of the data based on the spatial similarity sequence.
4. The method for predicting the lifespan of a storage device according to claim 1, characterized in that, The error type association module includes an error type feature library and a graph neural network unit. The error type feature library includes error types, corresponding feature vectors, associated data types, and physical models. The graph neural network unit is used to construct a multimodal feature interaction graph based on the multiple aligned feature vectors, using the feature vectors in the error type feature library as edge weights. A node in the multimodal feature interaction graph is one of the aligned feature vectors. The graph neural network unit is also used to aggregate the information of adjacent nodes in the multimodal feature interaction graph using graph convolutional layers and update the alignment feature vector represented by each node to obtain multiple associated feature vectors, wherein the associated feature vectors contain constraint information of error type.
5. The method for predicting the lifespan of a storage device according to claim 1, characterized in that, The lifetime prediction decision layer is also used to input the fused feature vector into the fully connected layer, and input the vector output by the fully connected layer into the activation layer to obtain the probability of each error type occurring in the storage device output by the activation layer.
6. The method for predicting the lifespan of a storage device according to claim 1, characterized in that, The calibration layer is also used to obtain the device type, usage scenario type and usage duration of the storage device, obtain the corresponding calibration coefficient from the calibration error information based on the device type, usage scenario type and usage duration, and calibrate the initial predicted lifespan of the storage device based on the calibration coefficient to obtain the final predicted lifespan.
7. The method for predicting the lifespan of a storage device according to claim 6, characterized in that, The calibration error information is a calibration error table that represents a multidimensional mapping relationship between device type, usage scenario type, usage duration, and calibration coefficient, determined based on training data. In the calibration error table, a combination of device type, usage scenario type, and usage duration interval corresponds to a calibration coefficient.
8. The method for predicting the lifespan of a storage device according to claim 1, characterized in that, The method further includes: The 3D scanning device is controlled to scan the storage device to obtain the 3D point cloud data of the storage device; The step of inputting the internal structure image, the sound data, the electrical data, the temperature data, and the error monitoring data into the pre-trained multimodal lifespan prediction model includes: The internal structure image, the sound data, the electrical data, the temperature data, the error monitoring data, and the three-dimensional point cloud data are input into the pre-trained multimodal lifespan prediction model.
9. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the storage device lifetime prediction method as described in any one of claims 1 to 8.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an executable program, which is executed by a processor to implement the storage device lifetime prediction method as described in any one of claims 1 to 8.