A deep learning-based storage device failure prediction method

Abstract

This invention discloses a deep learning-based method for predicting storage device failures, comprising: collecting operational monitoring data of the storage device; preprocessing to construct a time-series health representation sequence; performing adaptive window construction to extract the modeling sequence; establishing a topology graph and aggregating the node-level health representation sequences; constructing an improved SCINet model to obtain a prediction sequence; calculating the prediction residual sequence and generating a standardized residual sequence; outputting a failure risk judgment result based on the sequential probability ratio test; generating alarm information and triggering coordinated operation and maintenance actions. This invention, by combining the improved SCINet model and the sequential probability ratio test, achieves online prediction and early warning-based coordinated handling of storage device failure risks.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method for predicting storage device failures based on deep learning. Background Technology

[0002] In data centers and enterprise applications, existing storage devices typically rely on operational monitoring data for health management and fault handling. Common practices include collecting SMART health metrics, media error and correction statistics, interface and link error statistics, temperature and power supply status metrics, and IO performance and load statistics. Alarms and maintenance are then implemented based on threshold rules, empirical models, or shallow learning models. However, current operational monitoring data is characterized by multi-source heterogeneity, high noise, numerous missing data points, inconsistent sampling frequencies, and asynchronous timestamps. Log events are discrete and sudden, and statistical definitions are inconsistent, making it difficult to align data and express features on a unified timeline. This hinders the formation of a stable, time-series health representation that can accurately characterize the degradation and evolution process. Threshold rules are susceptible to noise and missing data, leading to false positives or false negatives. Shallow models rely on manual features and struggle to capture long-term dependencies and cross-metric coupling relationships, making it difficult to maintain stable predictive performance in complex loads and multi-device environments.

[0003] Storage device failure samples are inherently scarce and diverse, with a significant class imbalance between normal and pre-failure samples. Furthermore, different models, firmware versions, and workloads lead to significant variations in data distribution, making it easy for models to experience decreased generalization ability when deployed across different scenarios. Fault labeling typically relies on replacement records, maintenance records, or post-failure confirmation at the time of failure, making it difficult to precisely define the boundaries of the pre-failure stage. This results in unstable supervised learning objectives and further exacerbates the difficulty of model training. Existing alarm methods, primarily based on fixed windows and fixed thresholds, often lack a sequential decision structure for online streaming data. Alarm triggering relies on single observations or short-term statistics, making it difficult to achieve evidence-based early warning while maintaining a controllable false alarm rate. This often leads to alarm jitter, frequent triggering, and difficulty in reproducing the alarms. Moreover, in actual operation, device status gradually drifts with aging, making it difficult to adapt fixed thresholds and fixed distribution assumptions over the long term. There is a lack of a linkage mechanism that directly maps risk judgment results to data migration, load adjustment, path switching, isolation processing, or replacement plans.

[0004] Therefore, how to provide a deep learning-based method for predicting storage device failures is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method for predicting storage device failures. This invention comprehensively utilizes techniques such as unified temporal health characterization construction of operational monitoring data, link topology modeling, improved SCINet model structure innovation, and online decision-making via sequential probability ratio testing. It details the entire process from storage device operational monitoring data acquisition and preprocessing, event-consistency-driven adaptive window construction, topology graph construction and node-level health characterization generation, prediction sequence generation based on recursive variable split-aggregation and frequency-time dual-domain cross-routing combined with graph topology awareness, prediction residual standardization, to failure risk judgment and operation-maintenance linkage triggering via dual-channel symmetry testing and mean square composite statistical chain. This invention introduces core innovations in its model and decision structure, including recursive variable split-aggregation, frequency-time dual-domain cross-routing, graph topology awareness, dual-channel symmetry testing, and mean square composite statistical chain, enabling online rolling prediction of failure risks and low false alarm rates under multi-source heterogeneous noise data conditions. Compared with existing technologies, this invention has advantages such as strong health characterization consistency, good adaptability across models and loads, large early warning lead time, low false alarm rate, and ease of engineering-based linkage deployment.

[0006] A storage device fault prediction method based on deep learning according to an embodiment of the present invention includes:

[0007] Collect operational monitoring data from storage devices, preprocess the operational monitoring data, and construct a time-series health characterization sequence;

[0008] A variational entropy-fluctuation coupling-driven multi-scale adaptive window partitioning is performed on the time-series health representation sequence, and the most relevant window is selected to extract the modeling sequence through a dynamic memory pool voting mechanism.

[0009] A topology map is established based on the link connections of storage devices, and node-level health characterization sequences are obtained by aggregating nodes.

[0010] An improved SCINet model is constructed. Based on recursive variable split-aggregation, dynamic multi-split recursive convolution interaction and aggregation restoration are performed on the modeling sequence. Frequency-time dual-domain cross routing is introduced to perform time-domain convolution modeling and frequency-domain transformation modeling in parallel. Graph topology awareness is used to perform neighborhood propagation on the node-level health representation sequence to obtain topology embedding and thus obtain the prediction sequence.

[0011] The predicted residual sequence is calculated based on the predicted sequence and the actual observed sequence, and the predicted residual sequence is then used to generate a standardized residual sequence according to the standardization rules.

[0012] Based on the sequential probability ratio test, online decision is made on the standardized residual sequence. A dual-channel symmetric test is used to construct and update two symmetric sequential log-likelihood cumulative statistics for positive and negative drift respectively. The sequential cumulative statistics are constructed and updated for the squared terms of the standardized residual sequence through the mean square composite statistical chain, and the failure risk decision result is output.

[0013] Based on the fault risk assessment results, alarm information is generated, triggering joint operation and maintenance actions related to storage devices.

[0014] Optionally, the operational monitoring data includes SMART health indicators, media error and error correction statistics, interface and link error statistics, temperature and power supply status indicators, and IO performance and load statistics.

[0015] Optionally, the construction of the time-series health representation sequence includes:

[0016] Collect operational monitoring data and establish a unified time axis according to time granularity. Map SMART health indicators, media error and error correction statistics, interface and link error statistics, temperature and power supply status indicators, IO performance and load statistics to the corresponding time windows of the unified time axis to form the original observation vector sequence of each time window.

[0017] The original observation vector sequence is preprocessed, missing values ​​are filled according to time windows and a missing mask vector is generated, count data is aggregated according to time windows and an event count vector is generated, and numerical indicators are normalized to obtain normalized indicator vectors. The normalization process adopts the median-absolute deviation median normalization rule. The normalized indicator value is equal to the difference between the indicator value and the median of the indicator in the historical window, and then divided by the median absolute deviation of the indicator in the historical window. The median absolute deviation is the median of the absolute values ​​of the difference between the indicator value and the indicator median in the historical window.

[0018] The normalized index vector, event count vector, and missing mask vector are concatenated by channel to obtain a time-series health representation sequence corresponding to a unified time axis.

[0019] Optionally, the extraction of the modeling sequence includes:

[0020] Three types of windows—short, medium, and long—are slid simultaneously on the time axis, with a step size of a single time step. For each candidate window, variational entropy and fluctuation coupling degree are calculated. Variational entropy is obtained by statistically analyzing the frequency of occurrence of each normalized index within the discrete quantile interval and applying the information entropy formula. Fluctuation coupling degree is obtained by comparing the absolute sum of the differences between index vectors in adjacent time steps.

[0021] A score is generated for each candidate window. The score is a linear combination of variational entropy and fluctuation coupling with weights. Windows that are higher than the overall average level of candidate windows plus one standard deviation are marked as high-relevance windows. Adjacent high-relevance windows are merged into the same window segment, with the start and end index of the window segment as the boundary.

[0022] All window segments are written into a fixed-capacity dynamic memory pool in the order of generation and maintained in a first-in-first-out manner. If the current time step is in an interval where the number of window segments in the memory pool exceeds a preset threshold, the number of times the window segment containing the current time step appears is counted and the one with the most occurrences is selected as the valid window. The time-series health representation fragments corresponding to the valid window are extracted as the modeling sequence, and the window start index, end index and window length are recorded to form a window state identifier.

[0023] Optionally, the node-level health representation sequence obtained by node aggregation includes:

[0024] Obtain the link connection relationship data of the storage device, and represent the link connection relationship data as a set of nodes and a set of edges. The set of nodes includes controller nodes, storage medium nodes and link port nodes, and the set of edges includes physical connection edges or logical communication edges between nodes. Generate an adjacency matrix based on the set of edges.

[0025] According to the node mapping rules, the time-series health representation sequence is mapped to each node in the node set. Within each time window, the indicator channels mapped to the same node are aggregated to obtain the node feature vector, forming the node feature vector sequence of each time window.

[0026] Align the node feature vector sequences of each node according to the time window index to obtain the node-level health representation sequence.

[0027] Optionally, obtaining the predicted sequence includes:

[0028] An improved SCINet model is constructed, which includes a recursive variable split-aggregation module, a frequency-time dual-domain cross routing module, and a graph topology-aware module.

[0029] The modeling sequence is extracted using a recursive variable split-aggregation module. In each recursive layer, the number of splits and the set of split indices are determined based on the split control vector. The modeling sequence is divided into subsequences according to the set of split indices. One-dimensional convolution is performed on each subsequence to obtain subsequence features. Interactive operations are performed on the features of each subsequence to obtain interactive features. Aggregation is performed according to the set of split indices to restore the output sequence of the current recursive layer. After recursing to the preset number of recursive layers, the temporal feature sequence is obtained.

[0030] The modeling sequence is modeled in parallel across two domains using a frequency-time dual-domain cross-routing module. The time-domain branch performs convolutional interaction processing on the time-domain feature sequence to obtain the time-domain output feature sequence. The frequency-domain branch performs discrete cosine transform on the modeling sequence to obtain the frequency-domain sequence. Splitting, convolution, interaction and aggregation restoration processing are then performed to obtain the frequency-domain output feature sequence. After each layer of processing is completed, channel swapping is performed to obtain the fused feature sequence.

[0031] The topology embedding sequence is obtained by performing neighborhood propagation on the node-level health representation sequence through the graph topology perception module. The neighborhood propagation is based on the adjacency matrix. The node feature vector of each node and the feature vector of the neighboring nodes are linearly transformed and then summed to obtain the propagation result. The propagation result is then mapped nonlinearly to obtain the topology embedding sequence.

[0032] The topological embedding sequence is aligned with the fused feature sequence according to the time window index, and a fusion operation is performed to obtain the joint feature sequence. The joint feature sequence is then mapped to the output to obtain the prediction sequence.

[0033] Optionally, generating a standardized residual sequence from the predicted residual sequence according to a standardization rule includes:

[0034] Within each time window of a unified time axis, the predicted sequence is obtained, and the actual observation sequence corresponding to the same time window is obtained. The predicted sequence and the actual observation sequence are aligned according to the index channel and the prediction step size.

[0035] Calculate the predicted residual sequence. The residual value of any time window, any step size, and any index channel in the predicted residual sequence is equal to the actual observed value of the index channel at the current time window step size minus the predicted value of the index channel at the current time window step size.

[0036] The predicted residual sequence is standardized to generate a standardized residual sequence. The standardized value of any residual term in the standardized residual sequence is equal to the residual term minus the mean of the corresponding residual term in the historical window, and then divided by the standard deviation of the corresponding residual term in the historical window.

[0037] Optionally, the output fault risk judgment result includes:

[0038] The input observation sequence for the sequential probability ratio test is established based on the standardized residual sequence. The current standardized residual value is obtained as the current observation value in each time window, and the initial values ​​of the first cumulative statistic, the second cumulative statistic and the mean square cumulative statistic are initialized to zero.

[0039] Two symmetric sequential log-likelihood cumulative statistics are constructed and updated for positive and negative drift respectively using a dual-channel symmetric test. The first cumulative statistic for positive drift and the second cumulative statistic for negative drift are updated in each time window according to the log-likelihood increment. The natural logarithm of the ratio of the probability density of the current observation under the negative drift assumption to the probability density under the no drift assumption is added to the second cumulative statistic of the previous time window. The first cumulative statistic is compared with the first upper threshold and the first lower threshold respectively, and the second cumulative statistic is compared with the second upper threshold and the second lower threshold respectively.

[0040] The mean square cumulative statistic is updated by the squared terms of the observed sequence through a mean squared composite statistical chain. The mean square cumulative statistic is updated in each time window by adding the square of the current observation to the mean square cumulative statistic of the previous time window and comparing the mean square cumulative statistic with the mean square upper threshold.

[0041] Based on the threshold comparison results, the fault risk judgment result is output. When the first cumulative statistic is not less than the first upper threshold, a positive drift alarm is output. When the second cumulative statistic is not less than the second upper threshold, a negative drift alarm is output. When the mean square cumulative statistic is not less than the mean square upper threshold, a fluctuation expansion alarm is output. When the first cumulative statistic is not greater than the first lower threshold, the second cumulative statistic is not greater than the second lower threshold, and the mean square cumulative statistic is less than the mean square upper threshold, a normal judgment is output.

[0042] Optionally, generating alarm information based on the fault risk judgment result includes:

[0043] Receive the fault risk judgment result and generate alarm information. The alarm information includes storage device identifier, alarm time window index, alarm type, trigger statistics identifier and trigger threshold identifier. The alarm types include positive drift alarm, negative drift alarm and fluctuation expansion alarm.

[0044] The set of operation and maintenance linkage actions is determined based on the alarm type and linkage mapping table. The linkage mapping table defines the correspondence between alarm type and operation and maintenance linkage action. The operation and maintenance linkage action includes data migration, load adjustment, path switching, isolation processing and replacement plan. Execution parameters are generated for the set of operation and maintenance linkage actions. The execution parameters include the target object of the action, the start time of the action and the duration of the action.

[0045] Output alarm information and trigger the execution of a set of operation and maintenance linkage actions. Record the execution results to generate linkage records. The linkage records include storage device identifier, alarm type, set of operation and maintenance linkage actions, execution timestamp and execution status.

[0046] The beneficial effects of this invention are:

[0047] This invention forms a consistent time-series health representation sequence by uniformly aligning the storage device operation monitoring data with a time axis, constructing a missing mask, vectorizing event counts, and performing robust normalization. It also constructs a topology graph and node-level health representation sequence by combining link connection relationships. Compared to methods relying on a single indicator or a fixed window, this invention can stably extract degradation-related information under conditions of high noise, numerous missing data, and asynchronous multi-source heterogeneous data. Furthermore, the improved SCINet model introduces recursive variable split-aggregation, frequency-time dual-domain cross-routing, and graph topology awareness to achieve joint modeling of degradation trends, periodic weak signals, and topology-related faults. This enhances prediction accuracy and generalization ability across different models and load scenarios, and strengthens the ability to capture early degradation features.

[0048] This invention constructs and standardizes the prediction residuals based on the predicted sequence and the actual observed sequence, and then uses a sequential probability ratio test for online decision-making. It simultaneously covers positive and negative drift through a dual-channel symmetric test and covers fluctuation-induced degradation through a mean square composite statistical chain. Compared with fixed threshold alarms, this invention can reduce false alarms and false negatives and increase the early warning lead time. This invention maps the fault risk judgment results into alarm information and triggers coordinated operation and maintenance actions such as data migration, load adjustment, path switching, isolation processing, or plan replacement, forming a closed-loop handling process that can be engineered and implemented, thereby improving the availability and operation and maintenance efficiency of the storage platform. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0050] Figure 1 This is a flowchart of a deep learning-based storage device fault prediction method proposed in this invention;

[0051] Figure 2 This is a block diagram of the improved SCINet model for a deep learning-based storage device fault prediction method proposed in this invention.

[0052] Figure 3 This is a functional diagram of the sequential probability ratio test in a deep learning-based storage device fault prediction method proposed in this invention. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0054] refer to Figure 1 , Figure 2 and Figure 3A deep learning-based method for predicting storage device failures includes:

[0055] Collect operational monitoring data from storage devices, preprocess the operational monitoring data, and construct a time-series health characterization sequence;

[0056] A variational entropy-fluctuation coupling-driven multi-scale adaptive window partitioning is performed on the time-series health representation sequence, and the most relevant window is selected to extract the modeling sequence through a dynamic memory pool voting mechanism.

[0057] A topology map is established based on the link connections of storage devices, and node-level health characterization sequences are obtained by aggregating nodes.

[0058] An improved SCINet model is constructed. Based on recursive variable split-aggregation, dynamic multi-split recursive convolution interaction and aggregation restoration are performed on the modeling sequence. Frequency-time dual-domain cross routing is introduced to perform time-domain convolution modeling and frequency-domain transformation modeling in parallel. Graph topology awareness is used to perform neighborhood propagation on the node-level health representation sequence to obtain topology embedding and thus obtain the prediction sequence.

[0059] The predicted residual sequence is calculated based on the predicted sequence and the actual observed sequence, and the predicted residual sequence is then used to generate a standardized residual sequence according to the standardization rules.

[0060] Based on the sequential probability ratio test, online decision is made on the standardized residual sequence. A dual-channel symmetric test is used to construct and update two symmetric sequential log-likelihood cumulative statistics for positive and negative drift respectively. The sequential cumulative statistics are constructed and updated for the squared terms of the standardized residual sequence through the mean square composite statistical chain, and the failure risk decision result is output.

[0061] Based on the fault risk assessment results, alarm information is generated, triggering joint operation and maintenance actions related to storage devices.

[0062] In this embodiment, the operation monitoring data includes SMART health indicators, media error and error correction statistics, interface and link error statistics, temperature and power supply status indicators, and IO performance and load statistics.

[0063] In this embodiment, constructing the time-series health representation sequence includes:

[0064] Collect operational monitoring data and establish a unified time axis according to time granularity. Map SMART health indicators, media error and error correction statistics, interface and link error statistics, temperature and power supply status indicators, IO performance and load statistics to the corresponding time windows of the unified time axis to form the original observation vector sequence of each time window.

[0065] The original observation vector sequence is preprocessed. Missing values ​​are filled according to time windows to generate a missing mask vector. Count data is aggregated according to time windows to generate an event count vector. Numerical indicators are normalized to obtain a normalized indicator vector. The normalization process uses the median-median absolute deviation normalization rule. The normalized indicator value is equal to the indicator value minus the median of the indicator within the historical window, divided by the median absolute deviation of the indicator within the historical window. The median absolute deviation is the median of the absolute values ​​of the differences between the indicator value and the indicator median within the historical window. Specifically, the aggregation of count data according to time windows to generate the event count vector is as follows:

[0066] The unified timeline is divided into continuous time windows of fixed duration. Within each time window, all count records falling into the time window are categorized according to event type. For each event type, the number of times the type record appears in the current time window is counted. When the same event type corresponds to multiple sources or multiple channels within the time window, the counts of each source or channel are added together to obtain the total number of times the event type appears in the current time window. The total number of times the event type appears is compared with the upper threshold, and the smaller value is taken as the final count of the current event type within the time window. The upper threshold is set to one hundred. The final counts of all event types within the time window are arranged in the fixed number order of the event types to obtain the event count vector corresponding to the time window.

[0067] The normalized index vector, event count vector, and missing mask vector are concatenated by channel to obtain a time-series health representation sequence corresponding to a unified time axis.

[0068] In this embodiment, the extraction of the modeling sequence includes:

[0069] Three types of windows—short, medium, and long—are simultaneously slid along the time axis, with a step size of a single time step. For each candidate window, variational entropy and fluctuation coupling degree are calculated. Variational entropy is obtained by statistically analyzing the frequency of occurrence of each normalized index within the discrete quantile interval and applying the information entropy formula. Fluctuation coupling degree is obtained by comparing the absolute sum of the differences between index vectors at adjacent time steps. Where:

[0070] The three types of windows—short, medium, and long—are slid simultaneously on the timeline. Specifically, a coverage length is set for each type of window. At any given time, the end positions of the three windows are aligned with the current time step. As the time step moves forward, the three windows are simultaneously shifted forward by one time step. At each time step, three candidate windows with different spans and ending points at the current time point are generated and marked with start and end indices until the end of the timeline.

[0071] The variational entropy and fluctuation coupling degree are calculated as follows: For each candidate window, the normalized index value is first divided into discrete quantile intervals according to the quantile boundaries. After counting the frequency of each quantile interval, the variational entropy is obtained according to the information entropy formula. The frequency of each quantile interval is multiplied by the negative of the product of the logarithm with base 2 and summed. The adjacent time steps within the window are traversed in time order. The absolute sum of the differences between the index vectors of each pair of adjacent time steps is calculated and accumulated. The accumulated value is the fluctuation coupling degree of the window.

[0072] A score is generated for each candidate window. The score is a linear combination of variational entropy and fluctuation coupling weights. Windows with a score higher than the overall average level plus one standard deviation of the candidate windows are marked as highly relevant windows. Adjacent highly relevant windows are merged into the same window segment, with the start and end indices of the window segment as the boundaries. The score is a linear combination of variational entropy and fluctuation coupling weights, specifically:

[0073] For each candidate window, the variational entropy and fluctuation coupling degree are calculated. The window score is obtained by multiplying the variational entropy by 0.6 and the fluctuation coupling degree by 0.4 according to the linear combination formula. The arithmetic mean and standard deviation of the scores of all candidate windows are calculated. The average value plus the single standard deviation is used as the threshold. Window scores higher than the threshold are marked as highly relevant windows.

[0074] All window segments are written into a fixed-capacity dynamic memory pool in the order of generation and maintained in a first-in-first-out manner. If the current time step is in an interval where the number of window segments in the memory pool exceeds a preset threshold, the number of occurrences of the window segment containing the current time step is counted and the one with the most occurrences is selected as the valid window. The time-series health representation fragments corresponding to the valid window are extracted as the modeling sequence. The window start index, end index and window length are recorded to form a window status identifier. The fixed capacity of the dynamic memory pool is set to 60 and the preset threshold interval is set to ±15.

[0075] In this embodiment, the step of obtaining the node-level health representation sequence by node aggregation includes:

[0076] Obtain the link connection relationship data of the storage device, and represent the link connection relationship data as a set of nodes and a set of edges. The node set includes controller nodes, storage medium nodes, and link port nodes, and the edge set includes physical connection edges or logical communication edges between nodes. Generate an adjacency matrix based on the edge set. Specifically, the generation of the adjacency matrix based on the edge set is as follows:

[0077] Assign a unique node number to each node in the node set. Determine the row and column order of the adjacency matrix according to the node numbers in ascending order. Create a matrix with the number of rows and columns equal to the number of nodes. Initialize all elements of the matrix to zero. Traverse each edge in the edge set and read the node numbers i and j of the two endpoints of the edge. When the edge is a physical connection edge or a logical communication edge, set the element in the i-th row and j-th column of the matrix to one and the element in the j-th row and i-th column to one. When there are multiple edges with the same pair of nodes in the edge set, keep the corresponding elements of the matrix as one. To avoid invalid connections, connections with a count of less than five for logical communication edges within the time window are not written into the matrix. After traversal, the adjacency matrix is ​​obtained.

[0078] The time-series health representation sequence is mapped to each node in the node set according to the node mapping rules. Within each time window, the indicator channels mapped to the same node are aggregated to obtain the node feature vector, forming the node feature vector sequence of each time window. Specifically, the time-series health representation sequence is mapped to each node in the node set according to the node mapping rules.

[0079] A mapping table is established between indicator channels and nodes. Each indicator channel is uniquely assigned to one of the following nodes: controller node, storage medium node, or link port node. For each time window, the time-series health representation vector of the time window is read, and the values ​​of each indicator channel are written into the input set of the assigned node according to the mapping table. When the same node corresponds to multiple indicator channels in the same time window, the arithmetic mean is taken for numerical indicators, the sum is taken for count indicators, and the maximum bitwise value is taken for the missing mask to obtain the node feature vector of the node in the time window. When the missing mask of an indicator channel in the time window is 1, the indicator channel does not participate in the arithmetic mean and summation calculation. When the number of valid indicator channels participating in aggregation in a node in the time window is less than three, the node feature vector of the node in the time window is set to zero and marked as invalid.

[0080] Align the node feature vector sequences of each node according to the time window index to obtain the node-level health representation sequence.

[0081] In this embodiment, obtaining the predicted sequence includes:

[0082] An improved SCINet model is constructed, comprising a recursive variable split-aggregation module, a frequency-time dual-domain cross-routing module, and a graph topology-aware module. Specifically, the improved SCINet model is constructed as follows:

[0083] The recursive variable split-aggregation module is embedded into the fixed odd-even sampling split position of the original SCINet. The fixed binary sampling rule is replaced by a dynamic split control vector. Each recursive layer performs multi-split recursive convolution interaction on the modeling sequence according to the split number and split index set and completes aggregation and restoration. The frequency-time dual-domain cross routing module is connected to the feature backbone of the recursive interaction output. The time domain branch and the frequency domain branch are introduced to model in parallel. The frequency domain branch performs a split convolution interaction structure that is isomorphic to the time domain after discrete cosine transform. Channel exchange is implemented at the interaction output position of each layer to realize dual-domain feature routing. After the dual-domain routing output, the graph topology awareness module is connected. The neighborhood propagation driven by the link topology adjacency matrix is ​​used to replace the topologyless fusion method. Node-level topology embedding is generated and aligned with the dual-domain fusion features for fusion. The output mapping layer replaces the original single path output and maps the fused joint features to the prediction sequence.

[0084] The improved SCINet model is trained by taking the modeling sequence and the corresponding actual observation sequence as training samples. The prediction error is used as the optimization target. The prediction error is calculated by the difference between the predicted sequence and the actual observation sequence at each prediction step and each index channel. The loss value is formed by summing the squares of the differences. During the training process, the model parameters are updated through backpropagation. The split control vector generation unit, subsequence one-dimensional convolution unit and interaction operation unit in the recursive variable split-aggregation module are continuously optimized. The time-domain branch convolution interaction unit, frequency-domain branch discrete cosine transform post-processing unit and channel exchange unit in the frequency-time dual-domain cross routing module are optimized. The neighborhood propagation linear transformation unit and nonlinear mapping unit in the graph topology sensing module are optimized. The parameters of the fusion operation and output mapping layer are optimized simultaneously to gradually reduce the error between the predicted sequence and the actual observation sequence output by the improved SCINet model.

[0085] Temporal features are extracted from the modeling sequence using a recursive variable split-aggregation module. At each recursive layer, the number of splits and the set of split indices are determined based on the split control vector. The modeling sequence is then divided into subsequences according to the set of split indices. One-dimensional convolution is performed on each subsequence to obtain subsequence features. Interactive operations are performed on the subsequence features to obtain interactive features. Aggregation is then performed according to the set of split indices to restore the output sequence of the current recursive layer. After recursing to a preset number of recursive layers, the temporal feature sequence is obtained, where:

[0086] In each recursive layer, the split number and split index set are determined based on the split control vector. Specifically, the input sequence of the recursive layer is first subjected to one-dimensional convolution and fully connected mapping to obtain the split control vector. The split control vector contains control components corresponding to three candidate split numbers, which are set as two, three, and four. The split control vector is normalized to obtain the probability values ​​of the three components. The candidate split number with the largest probability value is selected as the split number. After determining the split number, the time step index of the input sequence of the recursive layer is counted starting from zero. The time step index is modulo the split number and classified according to the remainder. The index with a remainder of zero is assigned to the first subsequence, the index with a remainder of one is assigned to the second subsequence, and so on, to obtain the split index set.

[0087] Interaction operations are performed on the features of each subsequence to obtain interaction features. Specifically, for each subsequence feature, a gate vector is first obtained by one-dimensional convolution. The gate vector is then subjected to element-wise nonlinear mapping to obtain interaction coefficients with values ​​between zero and one. For the i-th subsequence feature, the remaining subsequence features are aligned in time and summed element-wise to obtain complementary features. The complementary features are then multiplied element-wise by the interaction coefficients and added element-wise to the i-th subsequence feature to obtain the interaction features of the i-th subsequence. The interaction coefficients are generated using the convolution kernels shared by the same recursive layer. When the interaction coefficients are less than 0.1, they are truncated to 0.1 to avoid interaction failure.

[0088] The current recursive layer output sequence is obtained by performing aggregation and restoration according to the split index set. After recursing to the preset number of recursive layers, the temporal feature sequence is obtained. Specifically, an output placeholder is established for each time step of the input sequence of the recursive layer. The interaction features of each subsequence are written back to the corresponding time step position of the output placeholder according to the corresponding split index set, so that the sequence written back is consistent with the input sequence of the recursive layer in time order, thus obtaining the current recursive layer output sequence. The current recursive layer output sequence is used as the input of the next recursive layer to repeatedly perform split control vector generation, split number determination, split index set construction, subsequence convolution and interaction, aggregation and restoration, until the number of recursive layers reaches the preset number of recursive layers. The preset number of recursive layers is set to three. When the number of recursive layers reaches three, the final recursive layer output sequence is output as the temporal feature sequence.

[0089] The modeling sequence is modeled in parallel across two domains using a frequency-time dual-domain cross-routing module. The time-domain branch performs convolutional interaction processing on the time-domain feature sequence to obtain the time-domain output feature sequence. The frequency-domain branch performs discrete cosine transform on the modeling sequence to obtain the frequency-domain sequence, and then performs splitting, convolution, interaction, and aggregation restoration processing to obtain the frequency-domain output feature sequence. After each layer of processing, channel swapping is performed to obtain the fused feature sequence, where:

[0090] The temporal output feature sequence is obtained by performing convolutional interaction processing on the temporal feature sequence using a temporal branch, specifically as follows:

[0091] The temporal feature sequence is input into the convolutional interaction unit of the temporal branch according to the channel dimension. One-dimensional convolution is applied to the temporal feature sequence in each layer to obtain convolutional features. The convolutional features are divided into sub-sequence features according to the split index set. Interaction operation is performed on each sub-sequence feature according to the interaction rules of the same layer to obtain interaction features. The interaction features are aggregated and restored to the temporal layer output sequence according to the split index set. The temporal layer output sequence is used as the input of the next layer to repeat the convolution, split, interaction and aggregation restoration process until the number of layers reaches the preset number three, and the temporal output feature sequence is obtained.

[0092] The interaction rule is as follows: the i-th subsequence feature is first generated by one-dimensional convolution to form a gated vector and then the interaction coefficient is obtained by nonlinear mapping. The remaining subsequence features other than the i-th are aligned in time and then summed element by element to obtain complementary features. The complementary features are multiplied element by element by the interaction coefficient and added element by element to the i-th subsequence feature to obtain the i-th interaction feature.

[0093] The frequency domain branch is used to perform discrete cosine transform on the modeling sequence to obtain the frequency domain sequence. Specifically, the discrete cosine transform is performed on the modeling sequence for each channel within each time window, mapping the time series with a length equal to the number of sample points in the time window to a frequency domain coefficient sequence of the same length. In the discrete cosine transform, each frequency domain coefficient is obtained by multiplying the original sequence value with the corresponding cosine basis function value point by point for each time point from the start to the end of the time index, summing the results, and multiplying by the normalization coefficient corresponding to the frequency index. Only the coefficients of the low-frequency part are retained in the obtained frequency domain coefficient sequence, and the coefficients of the high-frequency part are set to zero to suppress high-frequency noise. The retention ratio of the low-frequency part is set to one-half and rounded down. The frequency domain coefficient sequence after retaining the low-frequency coefficients is used as the frequency domain sequence input to the frequency domain branch, and the frequency domain output feature sequence is obtained by splitting, convolution, interaction and aggregation restoration.

[0094] The topological embedding sequence is obtained by performing neighborhood propagation on the node-level health representation sequence through the graph topology perception module. Neighborhood propagation is based on the adjacency matrix, where the node feature vector of each node and the feature vectors of its neighboring nodes are linearly transformed and then summed to obtain the propagation result. The propagation result is then nonlinearly mapped to obtain the topological embedding sequence. Specifically, the process of obtaining the topological embedding sequence through neighborhood propagation of the node-level health representation sequence is as follows:

[0095] For each time window, a graph propagation calculation is performed. The node feature vector of each node is linearly transformed to obtain the node's transformation vector. The feature vectors of each neighboring node connected to the node in the adjacency matrix are linearly transformed to obtain the neighborhood transformation vector. The node's transformation vector and all neighborhood transformation vectors are summed element by element and divided by the larger of the number of the node's neighboring nodes and one to obtain the propagation result of the node in the time window.

[0096] The topological embedding vector of a node in a time window is obtained by applying a nonlinear mapping to each element of the propagation result. The nonlinear mapping adopts the rule of setting elements less than zero to zero. When the number of neighboring nodes is greater than eight, only the eight neighboring nodes with the highest connection strength are selected to participate in the summation calculation. The topological embedding vectors of all nodes in each time window are arranged in chronological order to obtain the topological embedding sequence.

[0097] The topological embedding sequence is aligned with the fused feature sequence according to the time window index, and a fusion operation is performed to obtain a joint feature sequence. The joint feature sequence is then mapped to the output to obtain a prediction sequence, which is specifically as follows:

[0098] Within each time window, the fused feature sequence and the topological embedding sequence are aligned by the time window index. The fused feature vector and the topological embedding vector of the same time window are concatenated in the channel dimension to form a joint feature vector. The joint feature vectors are then arranged into a joint feature sequence according to the time window order. An output mapping layer is applied to the joint feature sequence to map the feature dimension to the prediction target dimension. The output mapping layer is implemented using one-dimensional convolution with a kernel length of one and the number of kernels set to the product of the prediction target dimension and the prediction stride. The output of the output mapping layer is rearranged into a prediction sequence according to the prediction stride and the prediction target dimension. To avoid abnormal output amplitude, each predicted value in the prediction sequence is compared with an amplitude threshold, and the smaller absolute value amplitude is taken. The amplitude threshold is set to six.

[0099] In this embodiment, generating a standardized residual sequence from the predicted residual sequence according to the standardization rules includes:

[0100] Within each time window of a unified time axis, the predicted sequence is obtained, and the actual observation sequence corresponding to the same time window is obtained. The predicted sequence and the actual observation sequence are aligned according to the index channel and the prediction step size.

[0101] Calculate the predicted residual sequence. The residual value of any time window, any step size, and any index channel in the predicted residual sequence is equal to the actual observed value of the index channel at the current time window step size minus the predicted value of the index channel at the current time window step size.

[0102] The predicted residual sequence is standardized to generate a standardized residual sequence. The standardized value of any residual term in the standardized residual sequence is equal to the residual term minus the mean of the corresponding residual term in the historical window, and then divided by the standard deviation of the corresponding residual term in the historical window.

[0103] In this embodiment, the output of the fault risk judgment result includes:

[0104] The input observation sequence for the sequential probability ratio test is established based on the standardized residual sequence. The current standardized residual value is obtained as the current observation value in each time window, and the initial values ​​of the first cumulative statistic, the second cumulative statistic and the mean square cumulative statistic are initialized to zero.

[0105] A dual-channel symmetric test is used to construct and update two symmetric sequential log-likelihood cumulative statistics for both positive and negative drift. The first cumulative statistic for positive drift and the second cumulative statistic for negative drift are updated in each time window according to the log-likelihood increment. The natural logarithm of the ratio of the probability density of the current observation under the negative drift assumption to the probability density under the no-drift assumption is added to the second cumulative statistic of the previous time window. The first cumulative statistic is compared with the first upper threshold and the first lower threshold, and the second cumulative statistic is compared with the second upper threshold and the second lower threshold, respectively.

[0106] The first cumulative statistic for positive drift and the second cumulative statistic for negative drift are updated in each time window according to the log-likelihood increment, specifically:

[0107] At each time window, the current observation value is obtained, and the probability density of the current observation value under the no-drift assumption, the positive-drift assumption, and the negative-drift assumption is calculated respectively. The probability density under the positive-drift assumption is divided by the probability density under the no-drift assumption to obtain the positive likelihood ratio. The natural logarithm of the positive likelihood ratio is taken to obtain the positive log-likelihood increment. The positive log-likelihood increment is added to the first cumulative statistic of the previous time window to obtain the first cumulative statistic of the current time window.

[0108] The negative likelihood ratio is obtained by dividing the probability density under the negative drift assumption by the probability density under the no drift assumption. The negative log-likelihood increment is obtained by taking the natural log of the negative likelihood ratio. The negative log-likelihood increment is added to the second cumulative statistic of the previous time window to obtain the second cumulative statistic of the current time window.

[0109] The first cumulative statistic is compared with the first upper threshold and the first lower threshold, and the second cumulative statistic is compared with the second upper threshold and the second lower threshold, respectively. Specifically, after updating the first and second cumulative statistics in each time window, the first cumulative statistic is compared with the first upper threshold. When the first cumulative statistic is greater than or equal to the first upper threshold, a positive drift trigger flag is output. The first cumulative statistic is compared with the first lower threshold. When the first cumulative statistic is less than or equal to the first lower threshold, a positive drift release flag is output. The second cumulative statistic is compared with the second upper threshold and the second lower threshold in the same way. When the second cumulative statistic is greater than or equal to the second upper threshold, a negative drift trigger flag is output. When the second cumulative statistic is less than or equal to the second lower threshold, a negative drift release flag is output. The first upper threshold is set to 4.5, the first lower threshold is set to -4.5, the second upper threshold is set to 4.5, and the second lower threshold is set to -4.5.

[0110] The mean squared cumulative statistic is updated by using a mean squared composite statistical chain to update the squared terms of the observed sequence. The update method for each time window is to add the square of the current observation to the mean squared cumulative statistic of the previous time window, and then compare the mean squared cumulative statistic with an upper mean square threshold. Specifically, the comparison with the upper mean square threshold is as follows:

[0111] The mean square cumulative statistic of the current time window is compared with the mean square upper threshold. When the mean square cumulative statistic is greater than or equal to the mean square upper threshold, a fluctuation expansion trigger flag is output and the trigger time window index is recorded. When the mean square cumulative statistic is less than the mean square upper threshold, a fluctuation expansion not triggered flag is output and the update continues in the next time window. In order to make the threshold consistent with the time window length, the mean square upper threshold is set to the preset cumulative window length multiplied by the threshold constant. The cumulative window length is set to thirty time windows, the threshold constant is set to four, and the mean square upper threshold is equal to one hundred and twenty.

[0112] Based on the threshold comparison results, the fault risk judgment result is output. When the first cumulative statistic is not less than the first upper threshold, a positive drift alarm is output. When the second cumulative statistic is not less than the second upper threshold, a negative drift alarm is output. When the mean square cumulative statistic is not less than the mean square upper threshold, a fluctuation expansion alarm is output. When the first cumulative statistic is not greater than the first lower threshold, the second cumulative statistic is not greater than the second lower threshold, and the mean square cumulative statistic is less than the mean square upper threshold, a normal judgment is output.

[0113] In this embodiment, generating alarm information based on the fault risk judgment result includes:

[0114] Receive the fault risk judgment result and generate alarm information. The alarm information includes storage device identifier, alarm time window index, alarm type, trigger statistics identifier and trigger threshold identifier. The alarm types include positive drift alarm, negative drift alarm and fluctuation expansion alarm.

[0115] The set of operation and maintenance linkage actions is determined based on the alarm type and linkage mapping table. The linkage mapping table defines the correspondence between alarm type and operation and maintenance linkage action. The operation and maintenance linkage action includes data migration, load adjustment, path switching, isolation processing and replacement plan. Execution parameters are generated for the set of operation and maintenance linkage actions. The execution parameters include the target object of the action, the start time of the action and the duration of the action.

[0116] Output alarm information and trigger the execution of a set of operation and maintenance linkage actions. Record the execution results to generate linkage records. The linkage records include storage device identifier, alarm type, set of operation and maintenance linkage actions, execution timestamp and execution status.

[0117] Example 1: To verify the feasibility of this invention in practice, it was applied to a storage cluster operation and maintenance scenario for online business. The operation and maintenance system continuously accesses the operation monitoring data of each storage device. The data comes from device self-test and health statistics, media error and error correction statistics, interface and link error statistics, temperature and power supply status indicators, and IO performance and load statistics. In actual operation, the data suffers from high noise spikes, missing data, and asynchrony. Some devices experience intermittent reporting interruptions, leading to continuous data loss. Some indicators are refreshed hourly while IO statistics are refreshed minutely. Log events are sudden and dense under high load, and the count distribution has a long tail, making it difficult for traditional fixed window features to stably represent the degradation process. The proportion of real fault samples is low, and the distribution varies significantly across models and loads, resulting in high lag and false alarm rates in early warnings based on fixed thresholds or shallow models, making it difficult to meet the operation and maintenance requirements for lead time, stability, and executable linkage.

[0118] When applying this invention, the monitoring data is first aligned to a uniform time granularity and missing data filling, missing mask generation, aggregation of count-type events into event count vectors by time windows, and robust normalization are completed to obtain a time-series health representation sequence. Based on the changes in log event counts and the fluctuation intensity of numerical indicators, an adaptive window driven by event consistency is constructed, and a modeling sequence is output with window status indicators. At the same time, a topology graph is constructed according to the device link connection relationship, and node-level health representation sequences are obtained by aggregating controller nodes, storage medium nodes, and link port nodes. The modeling sequence enters the improved SCINet model to complete the prediction. Recursive variable split-aggregation is used for dynamic multi-split recursive interaction. Frequency-time dual-domain cross routing is used for parallel modeling in the time and frequency domains and channel switching. Graph topology awareness is used to perform neighborhood propagation on the node-level health representation sequence to obtain topological embedding and align and fuse it with dual-domain fusion features to output the prediction sequence. After the prediction sequence and the actual observation sequence form a prediction residual and are standardized to the observation sequence, they enter the sequential probability ratio test to complete the online decision. The dual-channel symmetric test accumulates log-likelihood evidence for positive and negative drift respectively. The mean square composite statistical chain performs square cumulative decision on fluctuation expansion. Finally, alarm information is generated and the linkage actions in the data migration, load adjustment, path switching, isolation processing or replacement plan are triggered.

[0119] In continuous operation evaluation, the method of this invention can maintain stable output under conditions of multi-source heterogeneity, asynchrony, and the presence of missing and spike noise. This indicates that the adaptive window construction driven by the consistency of time-series health representation sequences and events can effectively reduce the interference of missing and sudden events on modeling. The improved SCINet, through dynamic multi-splitting recursive interaction, parallel modeling in the time and frequency domains and channel switching, as well as link topology neighborhood propagation fusion, makes the model more sensitive to periodic weak signals and link-related fault propagation, thus maintaining good prediction consistency and stability even when deployed across different models and loads. In the online decision-making stage, the dual-channel symmetric test can simultaneously cover both risk uplink and risk downlink drift patterns, and the mean square composite statistical chain can supplement the decision on fluctuation expansion type degradation. After alarm output, alarm information can directly drive the linkage actions of data migration, load adjustment, path switching or isolation processing, forming a closed-loop process from prediction to handling, which helps to reduce the risk of business interruption caused by sudden failures and improve operation and maintenance efficiency.

[0120] Table 1. Comparison of Comprehensive Indicators of Storage Device Failure Prediction Methods

[0121] Comparison Indicators / Methods Threshold rules ARIMA LSTM TCN Original SCINet Method of the present invention Warning lead time (h) 8.4 14.7 19.6 23.1 26.5 33.2 False alarm rate (%) 6.8 5.1 4.6 4.0 3.7 2.3 Missed Reporting Rate (%) 21.6 17.8 14.9 13.5 11.8 7.9 AUC 0.71 0.76 0.81 0.83 0.85 0.90 F1 0.52 0.57 0.63 0.66 0.68 0.74 Alarm jitter (times / day) 1.9 1.4 1.1 0.9 0.8 0.3 Inference delay (ms) 2 6 18 14 16 21 Migration decline (%) 18.5 16.2 13.8 12.9 11.6 6.7

[0122] As shown in Table 1, in terms of early warning capability and overall discrimination performance, the method of this invention has an early warning lead time of 33.2 hours, which is higher than the original SCINet's 26.5 hours, ARIMA's 14.7 hours, and the threshold rule's 8.4 hours, demonstrating a larger early intervention window. At the same time, the method of this invention achieves AUC and F1 scores of 0.90 and 0.74 respectively, both higher than the original SCINet's 0.85 / 0.68 and better than LSTM's 0.81 / 0.63, indicating that the discrimination capability and precision-recall balance are more stable under different thresholds.

[0123] From the perspective of false alarm and false negative control, the false alarm rate and false negative rate of the method of the present invention are 2.3% and 7.9% respectively, which are significantly lower than the threshold rule's 6.8% / 21.6% and better than LSTM's 4.6% / 14.9%. The data shows that the present invention reduces invalid alarms while reducing the risk of missing real faults, and the overall risk judgment is more reliable.

[0124] From the perspective of engineering implementation metrics, the alarm jitter of the method of this invention is 0.3 times / day, which is lower than the original SCINet's 0.8 times / day, and the alarm stability is more suitable for triggering linkage processing; the inference latency is 21ms, which is higher than the threshold rule's 2ms and ARIMA's 6ms, but is still within the range of usable online inference, and is on the same order of magnitude as TCN's 14ms and the original SCINet's 16ms; in terms of migration degradation metrics, the present invention is 6.7%, which is lower than the threshold rule's 18.5%, reflecting better control of platform costs after linkage actions are triggered.

[0125] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting storage device failures based on deep learning, characterized in that, include: Collect operational monitoring data from storage devices, preprocess the operational monitoring data, and construct a time-series health characterization sequence; A variational entropy-fluctuation coupling-driven multi-scale adaptive window partitioning is performed on the time-series health representation sequence, and the most relevant window is selected to extract the modeling sequence through a dynamic memory pool voting mechanism. A topology map is established based on the link connections of storage devices, and node-level health characterization sequences are obtained by aggregating nodes. An improved SCINet model is constructed. Based on recursive variable split-aggregation, dynamic multi-split recursive convolution interaction and aggregation restoration are performed on the modeling sequence. Frequency-time dual-domain cross routing is introduced to perform time-domain convolution modeling and frequency-domain transformation modeling in parallel. Graph topology awareness is used to perform neighborhood propagation on the node-level health representation sequence to obtain topology embedding and thus obtain the prediction sequence. The predicted residual sequence is calculated based on the predicted sequence and the actual observed sequence, and the predicted residual sequence is then used to generate a standardized residual sequence according to the standardization rules. Based on the sequential probability ratio test, online decision is made on the standardized residual sequence. A dual-channel symmetric test is used to construct and update two symmetric sequential log-likelihood cumulative statistics for positive and negative drift respectively. The sequential cumulative statistics are constructed and updated for the squared terms of the standardized residual sequence through the mean square composite statistical chain, and the failure risk decision result is output. Based on the fault risk assessment results, alarm information is generated, triggering joint operation and maintenance actions related to storage devices.

2. The method for predicting storage device faults based on deep learning according to claim 1, characterized in that, The operational monitoring data includes SMART health indicators, media error and error correction statistics, interface and link error statistics, temperature and power supply status indicators, and IO performance and load statistics.

3. The storage device fault prediction method based on deep learning according to claim 1, characterized in that, The construction of the time-series health representation sequence includes: Collect operational monitoring data and establish a unified time axis according to time granularity. Map SMART health indicators, media error and error correction statistics, interface and link error statistics, temperature and power supply status indicators, IO performance and load statistics to the corresponding time windows of the unified time axis to form the original observation vector sequence of each time window. The original observation vector sequence is preprocessed, missing values ​​are filled according to time windows and a missing mask vector is generated, count data is aggregated according to time windows and an event count vector is generated, and numerical indicators are normalized to obtain normalized indicator vectors. The normalization process adopts the median-absolute deviation median normalization rule. The normalized indicator value is equal to the difference between the indicator value and the median of the indicator in the historical window, and then divided by the median absolute deviation of the indicator in the historical window. The median absolute deviation is the median of the absolute values ​​of the difference between the indicator value and the indicator median in the historical window. The normalized index vector, event count vector, and missing mask vector are concatenated by channel to obtain a time-series health representation sequence corresponding to a unified time axis.

4. The method for predicting storage device faults based on deep learning according to claim 1, characterized in that, The extracted modeling sequence includes: Three types of windows—short, medium, and long—are slid simultaneously on the time axis, with a step size of a single time step. For each candidate window, variational entropy and fluctuation coupling degree are calculated. Variational entropy is obtained by statistically analyzing the frequency of occurrence of each normalized index within the discrete quantile interval and applying the information entropy formula. Fluctuation coupling degree is obtained by comparing the absolute sum of the differences between index vectors in adjacent time steps. A score is generated for each candidate window. The score is a linear combination of variational entropy and fluctuation coupling with weights. Windows that are higher than the overall average level of candidate windows plus one standard deviation are marked as high-relevance windows. Adjacent high-relevance windows are merged into the same window segment, with the start and end index of the window segment as the boundary. All window segments are written into a fixed-capacity dynamic memory pool in the order of generation and maintained in a first-in-first-out manner. If the current time step is in an interval where the number of window segments in the memory pool exceeds a preset threshold, the number of times the window segment containing the current time step appears is counted and the one with the most occurrences is selected as the valid window. The time-series health representation fragments corresponding to the valid window are extracted as the modeling sequence, and the window start index, end index and window length are recorded to form a window state identifier.

5. The method for predicting storage device faults based on deep learning according to claim 1, characterized in that, The node-level health representation sequence obtained by node aggregation includes: Obtain the link connection relationship data of the storage device, and represent the link connection relationship data as a set of nodes and a set of edges. The set of nodes includes controller nodes, storage medium nodes and link port nodes, and the set of edges includes physical connection edges or logical communication edges between nodes. Generate an adjacency matrix based on the set of edges. According to the node mapping rules, the time-series health representation sequence is mapped to each node in the node set. Within each time window, the indicator channels mapped to the same node are aggregated to obtain the node feature vector, forming the node feature vector sequence of each time window. Align the node feature vector sequences of each node according to the time window index to obtain the node-level health representation sequence.

6. The storage device fault prediction method based on deep learning according to claim 1, characterized in that, The obtained predicted sequence includes: An improved SCINet model is constructed, which includes a recursive variable split-aggregation module, a frequency-time dual-domain cross routing module, and a graph topology-aware module. The modeling sequence is extracted using a recursive variable split-aggregation module. In each recursive layer, the number of splits and the set of split indices are determined based on the split control vector. The modeling sequence is divided into subsequences according to the set of split indices. One-dimensional convolution is performed on each subsequence to obtain subsequence features. Interactive operations are performed on the features of each subsequence to obtain interactive features. Aggregation is performed according to the set of split indices to restore the output sequence of the current recursive layer. After recursing to the preset number of recursive layers, the temporal feature sequence is obtained. The modeling sequence is modeled in parallel across two domains using a frequency-time dual-domain cross-routing module. The time-domain branch performs convolutional interaction processing on the time-domain feature sequence to obtain the time-domain output feature sequence. The frequency-domain branch performs discrete cosine transform on the modeling sequence to obtain the frequency-domain sequence. Splitting, convolution, interaction and aggregation restoration processing are then performed to obtain the frequency-domain output feature sequence. After each layer of processing is completed, channel swapping is performed to obtain the fused feature sequence. The topology embedding sequence is obtained by performing neighborhood propagation on the node-level health representation sequence through the graph topology perception module. The neighborhood propagation is based on the adjacency matrix. The node feature vector of each node and the feature vector of the neighboring nodes are linearly transformed and then summed to obtain the propagation result. The propagation result is then mapped nonlinearly to obtain the topology embedding sequence. The topological embedding sequence is aligned with the fused feature sequence according to the time window index, and a fusion operation is performed to obtain the joint feature sequence. The joint feature sequence is then mapped to the output to obtain the prediction sequence.

7. The method for predicting storage device faults based on deep learning according to claim 1, characterized in that, The step of generating a standardized residual sequence from the predicted residual sequence according to the standardization rules includes: Within each time window of a unified time axis, the predicted sequence is obtained, and the actual observation sequence corresponding to the same time window is obtained. The predicted sequence and the actual observation sequence are aligned according to the index channel and the prediction step size. Calculate the predicted residual sequence. The residual value of any time window, any step size, and any index channel in the predicted residual sequence is equal to the actual observed value of the index channel at the current time window step size minus the predicted value of the index channel at the current time window step size. The predicted residual sequence is standardized to generate a standardized residual sequence. The standardized value of any residual term in the standardized residual sequence is equal to the residual term minus the mean of the corresponding residual term in the historical window, and then divided by the standard deviation of the corresponding residual term in the historical window.

8. The method for predicting storage device faults based on deep learning according to claim 1, characterized in that, The output fault risk judgment result includes: The input observation sequence for the sequential probability ratio test is established based on the standardized residual sequence. The current standardized residual value is obtained as the current observation value in each time window, and the initial values ​​of the first cumulative statistic, the second cumulative statistic and the mean square cumulative statistic are initialized to zero. Two symmetric sequential log-likelihood cumulative statistics are constructed and updated for positive and negative drift respectively using a dual-channel symmetric test. The first cumulative statistic for positive drift and the second cumulative statistic for negative drift are updated in each time window according to the log-likelihood increment. The natural logarithm of the ratio of the probability density of the current observation under the negative drift assumption to the probability density under the no drift assumption is added to the second cumulative statistic of the previous time window. The first cumulative statistic is compared with the first upper threshold and the first lower threshold respectively, and the second cumulative statistic is compared with the second upper threshold and the second lower threshold respectively. The mean square cumulative statistic is updated by the squared terms of the observed sequence through a mean squared composite statistical chain. The mean square cumulative statistic is updated in each time window by adding the square of the current observation to the mean square cumulative statistic of the previous time window and comparing the mean square cumulative statistic with the mean square upper threshold. Based on the threshold comparison results, the fault risk judgment result is output. When the first cumulative statistic is not less than the first upper threshold, a positive drift alarm is output. When the second cumulative statistic is not less than the second upper threshold, a negative drift alarm is output. When the mean square cumulative statistic is not less than the mean square upper threshold, a fluctuation expansion alarm is output. When the first cumulative statistic is not greater than the first lower threshold, the second cumulative statistic is not greater than the second lower threshold, and the mean square cumulative statistic is less than the mean square upper threshold, a normal judgment is output.

9. The method for predicting storage device faults based on deep learning according to claim 1, characterized in that, The generation of alarm information based on the fault risk judgment result includes: Receive the fault risk judgment result and generate alarm information. The alarm information includes storage device identifier, alarm time window index, alarm type, trigger statistics identifier and trigger threshold identifier. The alarm types include positive drift alarm, negative drift alarm and fluctuation expansion alarm. The set of operation and maintenance linkage actions is determined based on the alarm type and linkage mapping table. The linkage mapping table defines the correspondence between alarm type and operation and maintenance linkage action. The operation and maintenance linkage action includes data migration, load adjustment, path switching, isolation processing and replacement plan. Execution parameters are generated for the set of operation and maintenance linkage actions. The execution parameters include the target object of the action, the start time of the action and the duration of the action. Output alarm information and trigger the execution of a set of operation and maintenance linkage actions. Record the execution results to generate linkage records. The linkage records include storage device identifier, alarm type, set of operation and maintenance linkage actions, execution timestamp and execution status.

Images