An automated testing method and system for SSD products
By constructing a multi-source state window and a timing anomaly detection model, and generating recovery control codes, automated detection and hardware-level recovery in SSD testing are integrated, solving the closed-loop problem of anomaly identification and recovery, and improving the automation and efficiency of testing.
Patent Information
- Application Number
- CN202611592876.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-15
- Publication Date
- 2026-11-10
Smart Images

Figure CN123068210A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hard disk testing technology, and in particular to an automated testing method and system suitable for SSD products. Background Technology
[0002] Currently, in the research and development and mass production testing of solid-state drives (SSDs), automated testing technology has been widely used in read / write performance verification, functional traversal, and stress testing. Existing testing solutions can typically issue basic commands to the SSD, control power-on and power-off, and record some exception logs.
[0003] However, in actual testing, especially when the SSD under test is used as the system disk of the test host, existing technologies still have significant shortcomings. On the one hand, anomaly identification methods are relatively simple, often relying solely on explicit faults such as lost heartbeats or system blue screens as the basis for judgment, making it difficult to effectively capture implicit degradation states such as I / O latency deterioration and gradual deterioration of health parameters, leading to delayed anomaly detection or misjudgment. On the other hand, anomaly handling and physical recovery processes are disconnected. When a serious anomaly is detected, manual intervention is usually required to disassemble the disk, transfer it to a dedicated card-opening platform for I / O shorting and firmware reconstruction, and then redeploy the test environment. This not only results in long test interruption times and low efficiency, but also, due to the lack of a unified anomaly context passing mechanism, the recovered tests often cannot accurately continue to the effective checkpoints before the anomaly occurred, causing the repeated execution of completed test tasks. In addition, the existing anomaly detection model lacks a deterministic control mapping between the hardware recovery actions, resulting in the intelligent detection results not being able to directly drive the underlying hardware to complete the automatic recovery closed loop.
[0004] Therefore, there is an urgent need for an automated testing solution that can deeply integrate intelligent anomaly identification with hardware-level automatic recovery. Summary of the Invention
[0005] To address the technical problem of the lack of closed-loop linkage between anomaly identification and physical recovery in existing SSD system disk testing technologies, this invention provides an automated testing method and system suitable for SSD products.
[0006] The technical solutions provided by the embodiments of the present invention are as follows:
[0007] First aspect
[0008] An automated testing method for SSD products provided in this embodiment of the invention includes:
[0009] S1: Obtain SSD automated testing tasks and configure the timing anomaly detection model and device mapping relationship;
[0010] S2: Collect the multi-source operating status of the SSD under test and construct a status window;
[0011] S3: Based on the aforementioned time-series anomaly detection model and preset hard anomaly rules, generate recovery control codes and anomaly contexts;
[0012] S4: In response to the recovery control code, save the test checkpoint and switch to the backup storage device;
[0013] S5: Based on the device mapping relationship, perform low-level hardware reset and firmware reconstruction on the tested SSD;
[0014] S6: Rebuild the test environment and generate a recovery execution point based on the exception context to continue testing.
[0015] Second aspect
[0016] An automated testing system for SSD products provided in this embodiment of the invention includes:
[0017] processor;
[0018] A memory storing computer-readable instructions, which, when executed by the processor, implement the automated testing method for SSD products as described in the first aspect.
[0019] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:
[0020] In this embodiment of the invention, by constructing a multi-source status window that includes heartbeat, system status, health parameters, I / O performance, and log semantics, and combining it with a fixed time-series anomaly detection model and hard anomaly rules, comprehensive and accurate perception of sudden SSD failures and gradual degradation is achieved. By mapping anomaly detection results to explicit recovery control codes and establishing a task mapping relationship between the tested SSD, backup storage device, and hardware control components, the anomaly identification results can directly drive the underlying hardware to perform in-situ I / O shorting, firmware reconstruction, and automatic system environment recovery, completely solving the technical problems of relying on manual transitions and fragmented recovery processes in traditional testing. Furthermore, by saving the anomaly context and performing secondary verification using a model with the same parameters after recovery, not only is the test progress accurately continued, but the validity verification of the recovery results is also ensured, significantly improving the automation, testing efficiency, and reliability of SSD system disk testing. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 This is a flowchart illustrating an automated testing method for SSD products provided in an embodiment of the present invention.
[0023] Figure 2 This is a schematic diagram of a joint anomaly determination and recovery control code generation process provided in an embodiment of the present invention.
[0024] Figure 3 This is a schematic diagram of a spatiotemporal gating fusion mechanism provided in an embodiment of the present invention.
[0025] Figure 4 This is a schematic diagram of the sample distribution for a joint criterion of correlation difference and reconstruction error provided in an embodiment of the present invention. Detailed Implementation
[0026] The technical solution of the present invention will now be described with reference to the accompanying drawings.
[0027] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.
[0028] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.
[0029] Example 1
[0030] Reference manual attached Figure 1 The diagram illustrates a flowchart of an automated testing method for SSD products provided by an embodiment of the present invention.
[0031] This invention provides an automated testing method for SSD products. This method can be implemented by an automated testing system for SSD products, which can be a terminal or a server. The processing flow of the automated testing method for SSD products may include the following steps:
[0032] Step S1 involves acquiring the SSD automated testing task and configuring the timing anomaly detection model's mapping relationship with the device. Specifically, after receiving the test task, the server workstation not only parses the basic test parameters but, more importantly, establishes a unique task mapping relationship between the SSD under test, the test host, the backup storage device, and hardware control components (such as relay groups and controllers). This mapping relationship defines the physical control channel, enabling subsequent software instructions to accurately address specific hardware execution units. Simultaneously, timing anomaly detection model parameters matching the SSD model under test are loaded and kept fixed throughout the entire testing cycle, ensuring consistency in anomaly judgment criteria and avoiding misjudgments caused by online data drift. This step lays a definite topological foundation and a unified judgment benchmark for subsequent automated closed-loop testing.
[0033] Step S2 involves collecting multi-source operational status data of the SSD under test and constructing a status window. During test execution, the system synchronously collects multi-dimensional data, including but not limited to heartbeat status, operating system status, health parameters, input / output status, I / O latency, log semantic features, and key event time intervals. This data is not isolated but aligned with a unified timestamp and combined into a continuous status window. By constructing this status window containing rich temporal information, the system can simultaneously perceive instantaneous hard faults and gradual performance degradation trends, providing comprehensive and structured data input for subsequent intelligent analysis. This solves the problem that traditional single indicators cannot accurately characterize the complex operational status of SSDs.
[0034] Step S3 generates recovery control codes and anomaly contexts based on the timing anomaly detection model and preset hard anomaly rules. This is the key link connecting software diagnostics and hardware actions. The system inputs the status window into the timing anomaly detection model for inference and performs a comprehensive judgment based on preset hard anomaly rules (such as disk failure, inability to enumerate, etc.). Unlike traditional methods that only output alarm logs, this step directly generates recovery control codes with clear control semantics (such as test continuation, controlled recovery, immediate recovery, or isolation), and simultaneously records the anomaly context containing information such as test progress and anomaly type. This design allows the anomaly identification results to directly drive subsequent physical recovery actions, eliminating the need for manual judgment and intervention, and achieving a deterministic binding between diagnosis and execution.
[0035] Step S4: In response to the recovery control code, save the test checkpoint and switch to the backup storage device. When the recovery control code indicates that recovery is required, the system first pauses the current test and saves the test checkpoint and exception context before the anomaly occurred, ensuring that no on-site information is lost. Subsequently, the system automatically powers down the SSD under test and switches the boot path of the test host to the backup storage device. This process is completed without the SSD under test leaving the test station, using the backup storage device to quickly take over the system operating environment, providing a stable and independent operating platform for subsequent low-level repair operations on the SSD under test, completely replacing the traditional manual disk disassembly and transfer process.
[0036] Step S5 involves performing a low-level hardware reset and firmware reconstruction on the SSD under test based on the device mapping relationship. In the system environment maintained by the backup storage device, the system automatically controls the hardware execution unit to perform low-level operations on the SSD under test according to the device mapping relationship established in step S1. For example, by coordinating timing such as shorting control pins and controlled power-on, the SSD under test is brought into a specific engineering mode (such as Boot drive state), and then the mass production tool is pushed to perform firmware reconstruction. This step transforms the original card opening and repair operation, which relied on manual experience, into a standardized automatic control sequence, achieving a firmware-level reset of the SSD under test at the original test station, significantly improving recovery efficiency and operational standardization.
[0037] Step S6: Rebuild the test environment and generate a recovery execution point based on the exception context to continue testing. After firmware reconstruction, the system automatically removes the boot path of the backup storage device, restores the SSD under test as the boot path of the system disk, and redeploys the target operating system and test tools to it. Furthermore, the system reads the previously saved exception context, extracts test checkpoints from it, and generates a recovery execution point, enabling the test to precisely resume from the most recent valid position before the exception occurred, rather than starting from scratch. This not only avoids the duplication of completed test tasks but also ensures the continuity and integrity of test results through dual recovery of the environment and schedule.
[0038] Through steps S1 to S6 above, this embodiment constructs a complete automated closed loop of "detection-decision-execution-recovery-continued testing". This closed-loop mechanism directly transforms the intelligent anomaly identification results into automatic recovery actions at the physical layer, and achieves lossless transmission of test states through the anomaly context, effectively solving the technical problems of fragmented anomaly handling, high dependence on manual intervention, and loss of test breakpoints after recovery in the prior art.
[0039] Example 2
[0040] Based on Embodiment 1, this embodiment further details the specific implementation of steps S1 and S2. As an optional implementation, step S1 specifically includes sub-steps S101 and S102.
[0041] In sub-step S101, a task mapping relationship is established between the SSD under test, the test host, the backup storage device, the hardware control components, and the server workstation based on the automated test task. Specifically, this task mapping relationship is not merely a logical association, but also a digital definition of the physical control channels. For example, after parsing the test task, the server workstation, according to the preset device topology configuration, binds the unique serial number of the SSD under test to a specific interface of the test host, the boot sequence of the backup storage device, and a specific pin channel of the hardware control components (such as relay groups or GPIO controllers). This binding relationship clarifies the issues of "who controls whom" and "through which path" in the subsequent recovery process. For example, it determines which relay is responsible for cutting off the power supply to the SSD under test, which signal is responsible for triggering the I / O short circuit of the SSD under test, and the BIOS boot item identifier corresponding to the backup storage device. By establishing this precise task mapping relationship, the recovery control code generated by the upper-layer software can be accurately translated into executable instructions of the lower-layer hardware, providing a definite physical topology basis for the entire automated closed loop and avoiding recovery failures caused by chaotic device connections or addressing errors.
[0042] In sub-step S102, fixed model parameters matching the SSD under test are loaded. These model parameters remain fixed during testing. Specifically, the "fixed model parameters" refer to a set of parameters that have been determined during offline training and are no longer updated during online inference. For example, this parameter set includes at least the network weight matrix of the temporal anomaly detection model, the anomaly judgment threshold, the standardized parameters of the input features (such as mean and standard deviation), the sliding window length setting, and the spectral gating parameters. In this embodiment, it is strictly prohibited to use real-time data generated by the SSD under test during the current testing process to fine-tune or retrain the above parameters online. The technical effect of this design is to help maintain the consistency of the anomaly judgment scale. If the model parameters are allowed to change dynamically with the data during testing, it may cause the model to gradually adapt to a certain gradual performance degradation, misjudging it as a "normal" state, or making it impossible to reproduce the judgment result before the anomaly occurred during the recovery verification stage due to parameter drift. By solidifying parameters, the system ensures that the same "ruler" is used for measurement, whether in the anomaly detection phase or the secondary verification phase after recovery. This effectively defends against the risk of data distribution drift caused by online learning and improves the credibility of test results.
[0043] As an optional implementation, step S2 specifically includes sub-step S201 and sub-step S202.
[0044] In sub-step S201, heartbeat status, operating system status, health parameters, input / output status, I / O latency status, log semantic features, and time interval features between key events are collected synchronously to construct a running status vector:
[0045]
[0046] in, This represents the running state vector at time t. This represents the heartbeat state at time t. This indicates the operating system state at time t. Represents the health parameters at time t. This represents the input and output states at time t. Indicates the I / O delay state at time t. This represents the semantic features of the log at time t. This represents the time interval characteristics between key events at time t. Specifically, these seven components constitute a comprehensive, multi-dimensional characterization of the operational health of the tested SSD. Among them, It not only includes the presence of a heartbeat, but also the jitter of the heartbeat cycle, which can be used to directly reflect the stability of the communication link between the testing tool and the SSD under test. It covers blue screen codes, boot phase flags, or kernel panic messages from the operating system, used to identify system-level crashes. This includes underlying physical metrics such as wear level and number of remapped sectors found in SMART information or NVMe Health Log. The system counted read / write errors, command timeouts, and device enumeration status to reflect the interaction quality at the protocol layer. It records performance metrics such as average latency and tail latency (e.g., P99 / P99.9) to capture performance degradation that is not visible to the naked eye. It is a semantic vector extracted from serial port logs and system event logs after natural language processing or keyword matching, which can transform unstructured text alarms into computable numerical features. This characterizes timing patterns such as "the time difference between the last power failure and the current enumeration" and "the time interval between two consecutive error logs," features that are crucial for identifying intermittent faults. By fusing the aforementioned heterogeneous data, this embodiment solves the problem that traditional technologies relying solely on a single heartbeat or error count cannot fully perceive the complex operating state of an SSD, enabling subsequent anomaly detection models to simultaneously capture both sudden hard faults and progressive soft failures.
[0047] In sub-step S202, the most recently acquired running state vectors are combined into a state window:
[0048]
[0049] in, Let W represent the state window up to time t, W represent the number of sampling points included in the sliding window, and T represent the matrix transpose. Specifically, the sliding window mechanism transforms discrete instantaneous sampling points into a two-dimensional matrix input with a temporal structure. The value of W determines the span of historical information observed by the model. For example, when the sampling period is 1 second, if W is set to 60, the state window covers the running trajectory of the most recent minute. This structural design allows the temporal anomaly detection model to not only analyze the current state but also capture the trend and dependencies of the state's evolution over time. For example, a single increase in I / O latency might be a normal load fluctuation, but if it shows a continuously increasing monotonic trend within W sampling points, or is accompanied by a specific log semantic sequence, it highly likely indicates an anomaly such as garbage collection blocking or cache exhaustion within the firmware. Therefore, the state window... The construction of this system is key to upgrading static snapshots to dynamic time series analysis, providing the necessary data foundation for accurately distinguishing between normal fluctuations and real anomalies.
[0050] Example 3
[0051] Building upon Examples 1 and 2, this example further details the specific algorithm implementation for generating the recovery control code based on the timing anomaly detection model in step S3. (See attached specification.) Figure 2 The diagram illustrates a joint anomaly determination and recovery control code generation process provided by an embodiment of the present invention.
[0052] As an optional implementation, step S3 specifically includes sub-steps S301 to S307.
[0053] In sub-step S301, different state features are projected onto the same latent space through an input mapping layer, forming an input temporal feature sequence representing the current SSD operating state. Specifically, since the state window constructed in Example 2 contains seven types of heterogeneous data, including heartbeat, operating system state, health parameters, I / O latency, and log semantics, their numerical ranges and physical dimensions differ greatly. This step maps these heterogeneous features onto the same high-dimensional latent space through a learnable embedding layer, enabling the subsequent neural network to process multi-source information within a unified metric space. This process is not a simple numerical concatenation, but rather achieves semantic alignment through matrix operations within the computer, laying a data foundation for subsequently capturing cross-modal anomaly correlations.
[0054] In sub-step S302, an anomaly attention mechanism is employed to model the correlation differences between the input temporal feature sequences. This mechanism aims to quantify the degree of deviation between the current SSD operating state and the normal temporal pattern.
[0055] First, the temporal feature sequence output from layer (l-1) is input into the corresponding query mapping, key mapping, value mapping, and scaling parameter mapping of layer l. A linear transformation is then performed on the temporal feature sequence using the corresponding parameter weight matrices to obtain the query tensor, key tensor, value tensor, and scaling parameters of layer l. The temporal feature sequence of layer (l-1) represents the input features of the current encoding layer. The query parameter weight matrix, key parameter weight matrix, and value parameter weight matrix of layer l are used to generate the corresponding query tensor, key tensor, and value tensor, respectively. The scaling parameter weight matrix of layer l is used to generate the corresponding scaling parameters based on the current input features.
[0056] The scaling parameter is a trainable parameter used to adjust the bandwidth of the Gaussian prior correlation distribution, thereby controlling the rate at which the prior correlation between different time points decays over time. Since different anomaly patterns in SSD automated testing have different timescale characteristics—for example, I / O instantaneous timeouts typically manifest as short-duration, concentrated local anomalies, while NAND wear or long-term performance degradation typically manifests as longer-duration, wide-range trend anomalies—training the scaling parameter enables the model to adaptively adjust the prior correlation range based on the current state characteristics to match anomaly patterns at different timescales, thus improving the model's sensitivity to detecting different types of SSD anomalies.
[0057] The l-th layer corresponds to the l-th coding layer in the multi-layer stacked encoder. Each coding layer processes the temporal features of the input layer by layer. In each coding layer, an anomaly attention mechanism, spectral energy filtering mechanism, and spatiotemporal gating fusion mechanism are combined to jointly extract the temporal dependencies, frequency domain variation features, and correlations between different state features in the SSD operating state.
[0058] Furthermore, a prior association matrix for layer l is constructed based on the scaling parameter and the distance between different time positions within the state window. Specifically, the relative distance between each time position within the state window is used as the basis for prior association calculation, and a Gaussian distribution with the scaling parameter corresponding to the current time position as the bandwidth parameter is used to determine the prior association strength between different time positions. When the distance between two time positions is small, the corresponding prior association strength is high. As the distance between two time positions increases, the corresponding prior association strength gradually decreases. By adjusting the scaling parameter to regulate the decay rate of association strength with time distance, a prior association pattern that can adapt to different abnormal time scales is formed.
[0059] After obtaining the initial prior association strength between each time position, the initial prior association strength is scaled. Preferably, it is normalized according to the row direction of the prior association matrix so that the sum of the prior association values corresponding to the same time position is 1, thus obtaining the prior association matrix of the l-th layer.
[0060] The prior correlation matrix describes the inherent correlation between different time positions within a state window based on time distance, under a preset normal or ideal stable operating state. The time position index within the state window identifies the relative position of each sampling moment within the current state window, and the state window length represents the number of consecutive sampling time points currently participating in the time-series correlation analysis.
[0061] After obtaining the query tensor and key tensor, the actual association strength between different time positions within the state window is calculated based on the matching relationship between the query tensor and the key tensor. Specifically, correlation calculations are performed on the query tensor and key tensor to obtain the feature similarity between different time positions, and the correlation calculation results are scaled according to the model feature dimension to suppress excessively large association values caused by the increase in feature dimension.
[0062] Subsequently, the scaled correlation results are normalized using Softmax to normalize the actual correlation weights corresponding to each time position, thereby obtaining the sequence correlation matrix of the l-th layer.
[0063] The sequence correlation matrix is used to characterize the actual dependencies formed between different time points based on current feature content in the currently collected multi-source SSD operating status data. Unlike the prior correlation matrix, which is mainly determined by time distance, the sequence correlation matrix is dynamically determined by the current actual operating data, and therefore can reflect atypical time dependencies that occur under abnormal conditions.
[0064] Furthermore, the value tensor is weighted and aggregated using the sequence correlation matrix, so that each time position is fused with the corresponding state features according to its actual correlation weight with other time positions, thus obtaining the correlation output of the l-th layer.
[0065] The associated output is used to characterize the contextual features formed by the current SSD running state under the actual temporal dependencies, and serves as one of the inputs for subsequent feature processing in the current coding layer.
[0066] Therefore, by constructing a priori correlation matrix reflecting normal timing distance patterns and a sequential correlation matrix reflecting actual SSD operational data dependencies, the differences between the two can be further compared. When the SSD is in a normal or relatively stable state, the actual sequential correlation usually shows high consistency with the priori correlation based on the normal timing pattern. When the SSD experiences instantaneous I / O timeouts, disk drops, abnormal retries, performance degradation caused by NAND wear, or other abnormal states, the dependencies between actual time positions may deviate from the normal timing correlation pattern, thus increasing the correlation difference between the sequential correlation matrix and the priori correlation matrix.
[0067] Therefore, correlation differences can be used as anomaly features to characterize the degree to which the current operating state of the SSD deviates from the normal timing pattern. In combination with correlation output, these features can be further used in subsequent encoding, timing state reconstruction, and time point anomaly score calculation to improve the ability to identify abnormal states at different time scales during automated testing.
[0068] After obtaining the prior correlation matrix and sequence correlation matrix for each coding layer, KL divergence is used to measure the distribution difference between the two correlation matrices in each layer. The difference results of each layer are then averaged to obtain the multi-layer average correlation difference. The multi-layer average correlation difference is used to comprehensively characterize the degree of deviation of the current SSD actual temporal correlation pattern from the normal temporal correlation pattern.
[0069] When the SSD is in a normal or relatively stable state, the sequence correlation matrix and the prior correlation matrix have high consistency and small correlation differences. When anomalies such as transient I / O timeouts, disk failures, abnormal retries, heartbeat disorders, NAND wear, or continuous performance degradation occur, the actual timing dependencies change, causing the sequence correlation matrix to deviate from the prior correlation matrix, thus increasing the correlation differences.
[0070] By synthesizing the correlation differences among multiple coding layers, both local instantaneous anomalies and gradual anomalies over longer time scales can be characterized simultaneously, and the likelihood of a single coding layer being affected by local fluctuations can be reduced.
[0071] Specifically, correlation discrepancies do not merely determine whether a single operating parameter exceeds a preset threshold, but rather whether the temporal correlation structure between multiple consecutive time points has changed. Therefore, latent faults where individual parameters remain within the normal range but the timing logic is abnormal can also be identified. For example, if the SSD heartbeat continues and the interval between individual heartbeats does not exceed the timeout threshold, but the fluctuation pattern of the intervals between consecutive heartbeats has deviated from the normal cycle pattern, correlation discrepancies can reflect this abnormal timing relationship.
[0072] Therefore, the anomaly attention mechanism obtains the correlation difference that characterizes the degree of SSD temporal correlation anomaly by comparing the prior temporal correlation relationship under normal conditions with the sequence correlation relationship formed by the current actual operating data. The correlation difference is then combined with the subsequent temporal state reconstruction results to calculate the time point anomaly score, providing a basis for anomaly evolution degree calculation and differentiated recovery decision-making.
[0073] In sub-step S303, a spatiotemporal gating fusion mechanism is used to fuse the input temporal feature sequence to obtain spatiotemporal fused features. (See attached instruction manual.) Figure 3 The diagram illustrates a spatiotemporal gating fusion mechanism provided by an embodiment of the present invention.
[0074] Specifically, the input temporal feature sequence is input into the first convolutional processing branch and the second convolutional processing branch respectively. Parallel feature extraction is performed on the input temporal features through the first convolutional layer and the second convolutional layer with different receptive fields to obtain local state change information at different time scales.
[0075] The first and second convolutional layers employ different receptive fields, allowing the two convolutional processing paths to focus on extracting temporal local features within different ranges. Convolutional processing with a smaller receptive field enhances the perception of local abrupt changes within a short timeframe, while convolutional processing with a larger receptive field extracts state change features spanning a longer local timeframe, thus forming a multi-scale local feature extraction structure.
[0076] In the first processing branch, the features obtained by the first convolutional layer for the input temporal feature sequence are interacted element-wise with the features obtained by the second convolutional layer for the input temporal feature sequence, and the interaction result is processed by GELU nonlinear activation to obtain the output of the first branch.
[0077] The first branch output is mainly used to preserve local state information with strong common responses among convolutional features of different receptive fields, enabling the model to highlight abnormal changes occurring within the current time range. For transient anomalies such as disk failure, command timeout, sudden increase in I / O latency, abnormal retries, and short-term heartbeat interruptions in SSD automated testing scenarios, the first branch can enhance the response of local mutation features through the interaction between multi-scale convolutional features.
[0078] In the second processing branch, the features obtained from the first and second convolutional layers are fused element-wise according to the crossover method corresponding to the first processing branch, and the fused result is subjected to GELU nonlinear activation to obtain the output of the second branch.
[0079] The output of the second branch is further input into the convolutional gated recurrent unit, which updates the hidden state corresponding to the current time step by combining the input features of the current time step with the hidden state saved in the previous time step.
[0080] The hidden state of the previous time step is used to store the SSD operating state information accumulated from multiple previous time points, while the hidden state of the current time step, based on the historical state information, further integrates the local features corresponding to the current time step. Through the gating state update process, selective control is exercised over the historical information that needs to be retained, the historical information that needs to be decayed, and the current state information that needs to be introduced, thereby establishing long-term dependencies between consecutive time points.
[0081] Therefore, the convolutional gated recurrent unit can continuously memorize the state evolution process of SSD over a long test period. For example, it can model the progressive degradation process such as NAND media wear, gradual increase in the number of bad blocks, slow decrease in health parameters, continuous increase in average I / O latency, and gradual increase in the number of error retries. This can make up for the problem that simple convolution processing mainly focuses on local time range and is difficult to fully describe the long-term state change trend.
[0082] Furthermore, the output of the first branch is fused with the hidden state corresponding to the current time step, so that the current local features and the historical state evolution features form a joint representation in the same feature space.
[0083] The first branch output primarily carries local mutation information extracted by multi-scale convolution within the current time range, while the hidden state at the current time step primarily carries historical state change information accumulated by convolutional gated recurrent units. By fusing the two, both current local anomaly features and long-term operational trend features can be preserved simultaneously.
[0084] The fused features are further processed by a third convolutional layer. The third convolutional layer completes channel adjustment, feature recombination, and local information integration to obtain the final spatiotemporal fused features.
[0085] The spatiotemporal fusion feature contains information in two dimensions: on the one hand, it contains information on the local operating state changes of the tested SSD at the current moment and in the adjacent time range; on the other hand, it contains information on the historical evolution of the tested SSD at multiple consecutive time points.
[0086] Therefore, when an SSD experiences short-term anomalies such as disk failure, command timeout, transient I / O jitter, or sudden errors, the multi-scale convolutional branch can quickly extract the corresponding local mutation features. When an SSD experiences long-term degradation such as NAND wear, continuously declining health, gradual deterioration of read / write performance, or continuously increasing error rate, the convolutional gated loop branch can continuously record the state change process using historical hidden states.
[0087] By combining the two types of features, the spatiotemporal gating fusion mechanism can both sense sudden anomalies occurring near the current time point and retain the memory of historical operating states and long-term degradation trajectories, thus forming an SSD operating state representation that takes into account both short-term anomaly response capability and long-term state evolution perception capability.
[0088] Furthermore, the spatiotemporal fusion features can be used as input features for subsequent multi-layer stacked encoders, and encoded together with the temporal correlation features extracted by the anomaly attention mechanism and other state features to enhance the model's comprehensive representation of the SSD's full lifecycle operating state, transient failures, and progressive degradation process.
[0089] In sub-step S304, a multi-layer stacked encoder is employed, integrating anomaly attention and spatiotemporal gating fusion mechanisms at each layer. The spatiotemporal fusion features extracted from each layer are then combined to obtain the encoder output. This deep stacked structure allows the model to abstract higher-order anomaly semantic features from the original signal layer by layer.
[0090] In sub-step S305, the decoder performs decoding reconstruction based on the encoder output to obtain the decoder output. The reconstruction process essentially involves the model attempting to "reconstruct" the current input. If the input contains anomalies, the model will be unable to accurately reconstruct it, resulting in a significant deviation.
[0091] In sub-step S306, the time-point anomaly score is calculated based on the reconstruction error and correlation difference between the encoder output and the decoder output:
[0092]
[0093]
[0094] in, Indicates the score of anomalies at a given time point. This represents the activation function. Indicates the correlation difference. Indicates the reconstruction error. Denotes the Frobenius norm. Indicates encoder output, Indicates the decoder output. This represents the weighting coefficient of the reconstruction error. This formula embodies the idea of dual verification: only when a point in time exhibits both anomalies in the association pattern (…) Large), and also shows difficulties in reconstruction ( Only when the anomaly score is large will it be assigned a high anomaly score. This joint criterion effectively avoids false alarms caused by a single indicator due to environmental noise or normal load switching, ensuring a high confidence level in anomaly identification.
[0095] Reference manual attached Figure 4 This diagram illustrates the sample distribution of a joint criterion for correlation difference and reconstruction error provided by an embodiment of the present invention. With the normalized correlation difference as the horizontal axis and the reconstruction error as the vertical axis, normal state samples are mainly distributed in regions where both indicators are low. Non-substantial anomalies such as test load switching and transient I / O jitter may result in an increase in one indicator while the other remains at a low level. Abnormal samples that truly correspond to continuous degradation, abnormal retries, or changes in equipment state structure are more likely to enter regions where both correlation difference and reconstruction error are high. The time-point anomaly score is not a simple binary choice between correlation difference and reconstruction error, but rather a joint weighting of reconstruction error using correlation difference. Therefore, when normal load switching only causes instantaneous reconstruction difficulties, but the temporal correlation still basically conforms to the normal prior, it is unlikely to be assigned an excessively high anomaly score; conversely, if, although a single operating parameter does not exceed the traditional fixed threshold at a certain time period, the dependency structure between consecutive time points has changed, and state reconstruction also becomes difficult, it can enter the high-confidence anomaly region.
[0096] In sub-step S307, the system classifies the SSD status based on the time-point anomaly score and preset hard anomaly rules, generating recovery control codes and anomaly contexts. Specifically, the system compares consecutive anomaly scores with preset thresholds and, combined with deterministic hard anomaly rules such as disk failure and inability to enumerate, classifies the SSD status into four levels: normal, degraded, severely abnormal, and non-automatically recoverable candidate. Based on this, the system outputs corresponding test continue, controlled recovery, immediate recovery, or isolation control codes. The anomaly context records the correlation information between the anomaly and the current operating status of the tested SSD, test conditions, anomaly judgment criteria, and subsequent recovery operations. This allows step S4 and subsequent recovery steps to execute targeted recovery control based on the original test state at the time of the anomaly, and to restore the test progress and detection conditions before the anomaly after recovery. This step transforms the aforementioned complex mathematical calculations into explicit device control commands, completing the decision-making closed loop from data perception to physical execution.
[0097] Example 4
[0098] Building upon Example 3, this example further details the specific execution process for anomaly classification, hardware-level automatic recovery, and closed-loop verification. This series of steps constitutes a complete automated execution chain from software diagnosis to physical repair and then to effectiveness confirmation.
[0099] As an optional implementation, step S307 specifically includes sub-steps S3071 and S3072. In sub-step S3071, the state of the SSD under test is divided into normal state, degraded state, severe abnormal state, and non-automatically recoverable candidate state. Specifically, this four-level state classification mechanism is the core basis for automated decision-making. Among them, the determination of the degraded state not only requires the abnormal score at a certain time point to exceed the warning threshold, but also requires the satisfaction of preset continuous conditions, such as exceeding the standard for N consecutive sampling points or exceeding the standard a certain number of times within a specified time. This design can effectively filter out isolated noise spikes caused by instantaneous load fluctuations or electromagnetic interference, preventing the system from overreacting to non-substantial anomalies. The triggering of severe abnormal states is more sensitive, including both the abnormal score exceeding a higher severe abnormal threshold and directly hitting preset hard anomaly rules, such as device failure, inability to enumerate, or controller error. The introduction of hard anomaly rules compensates for the perception lag that may exist in the pure data-driven model when dealing with sudden physical failures, ensuring the system's safety baseline. The "candidate status for non-automatic recovery" is used to identify situations where hardware damage is irreversible or the cumulative number of recovery attempts has reached its limit, providing a basis for subsequent isolation and treatment.
[0100] In sub-step S3072, test continuation control codes, controlled recovery control codes, immediate recovery control codes, and isolation control codes are generated and mapped respectively. Specifically, this mapping relationship transforms abstract state hierarchies into machine-executable deterministic instructions. For example, when a degraded state is identified, a controlled recovery control code is generated, instructing the system to pause and prepare for recovery in an orderly manner while ensuring data security. When a severe abnormal state is identified, an immediate recovery control code is generated, triggering the highest priority emergency takeover process. This one-to-one mapping mechanism eliminates the ambiguity of relying on manual interpretation of alarm logs to decide on operations in traditional solutions, enabling anomaly identification results to directly drive the transition of the underlying hardware state machine, significantly improving the response speed and decision consistency of the test system.
[0101] As an optional implementation, step S4 specifically includes sub-steps S401 and S402. In sub-step S401, in response to the recovery control code being a controlled recovery control code or an immediate recovery control code, new requests to the SSD under test are paused, and the test checkpoint and exception context before the exception occurred are saved. Specifically, saving the test checkpoint is a crucial prerequisite for implementing subsequent "breakpoint continuation testing". This checkpoint not only records the execution progress of the current test case but also includes the state of key variables in memory, the count of completed I / O operations, and temporary data not written to persistent storage. Simultaneously, the exception context encapsulates complete context information triggered during recovery, including the exception type, associated difference values, reconstruction error distribution, and the device topology state at that time. This information is serialized and stored in non-volatile media, ensuring that even after the SSD under test is completely powered off or its firmware is reset, the system can still accurately trace back to the logical location before the failure, avoiding the time wasted by full retesting.
[0102] In sub-step S402, the tested SSD is powered off, and the boot path of the backup storage device is switched. Once the backup environment is detected as ready, this is output to the subsequent recovery process. Specifically, this step achieves seamless hot-switching of the operating environment. Since the tested SSD is already in an abnormal state, continuing to use it as the system disk may cause the operating system to crash or the file system to become corrupted; therefore, its power supply must be cut off first to protect the situation. Subsequently, the boot order of the test host is modified through hardware control components, causing it to boot from the preset backup storage device. The backup storage device hosts an independent micro-operating system and a recovery agent, which can maintain the basic computing power and network communication capabilities of the test host without relying on the tested SSD. Only after the system confirms that the backup environment's service heartbeat is normal, the network is connected, and the recovery agent is ready will it proceed to the next stage, thus providing a stable and reliable execution platform for subsequent low-level repair operations.
[0103] As an optional implementation, step S5 specifically includes sub-steps S501 to S503. This is the core step in achieving in-situ hardware repair. In sub-step S501, the control pin shorting channel is closed, and the SSD under test is powered on under controlled power while the shorting is maintained, causing it to enter the boot disk state. Specifically, this operation is an automated simulation of the traditional manual card opening fixture function. For most SSD controller chips, shorting a specific GPIO pin is the physical trigger condition that forces the chip to skip the normal firmware loading process and directly enter the underlying Boot ROM engineering mode. The system precisely controls the closing timing of the shorting path through relays or electronic switches to ensure that the shorting signal is stably established at the moment of power-on, thereby enabling the SSD under test to be recognized as a device awaiting mass production rather than a regular storage disk. This in-situ operation method completely eliminates the cumbersome process of disassembling the SSD, transferring it to a dedicated card opening station, and reinstalling it, significantly shortening the physical recovery cycle.
[0104] In sub-step S502, the pin shorting channel is disconnected after the preset shorting hold condition is met or the Boot enumeration signal is detected. Specifically, the timing of the shorting signal release is crucial. Releasing it too early may cause the SSD to exit engineering mode, while releasing it too late may interfere with subsequent mass production command interactions. The system monitors device enumeration events on the USB or PCIe bus in real time. Once it confirms that the Boot ROM has been successfully loaded and recognized by the host, it immediately disconnects the shorting circuit, or automatically releases it after a preset safe hold period. This feedback-based dynamic timing control ensures the reliability of engineering mode switching.
[0105] In sub-step S503, after the standby storage device enumerates the SSD under test, it pushes the mass production tool to perform firmware reconstruction. Specifically, the recovery agent in the standby storage device pulls the matching mass production toolkit and firmware image from the server based on the SSD model and firmware version information recorded in the previously saved exception context, and issues a reconstruction command to the SSD under test in boot state through a standard interface. This process covers low-level operations such as bad block table reconstruction, firmware area erasure, and parameter calibration, restoring the SSD to its initial, normally functioning state. Since the entire process is completed in a closed loop within the test host, no manual intervention is required, truly achieving automated on-site revival of the failed disk.
[0106] As an optional implementation, step S6 specifically includes sub-steps S601 to S603. In sub-step S601, the boot path of the backup storage device is revoked, and the system disk boot path of the SSD under test is restored. Specifically, after successful firmware reconstruction, the system again switches the boot order via hardware control components, returning control to the newly repaired SSD under test. In sub-step S602, the target operating system and testing tools are deployed to the SSD under test. Since mass production operations erase all user data on the SSD, it is necessary to reinstall the required operating system environment and testing tools such as Ater via network deployment services or local cache images to ensure that the software stack is consistent with the test task requirements. In sub-step S603, a recovery execution point is generated based on the test checkpoints in the exception context, and the test continues from the recovery execution point. Specifically, the system reads the checkpoint information saved in step S401, parses out the test case ID, loop iteration, and data offset at the time of interruption, constructs the recovery execution point accordingly, and injects it into the test framework. This allows the testing process to be precisely resumed from the breakpoint, rather than mechanically starting from scratch, greatly improving the efficiency of testing resource utilization.
[0107] As an optional implementation, steps S7 to S9 are included after step S6 to form a closed-loop verification mechanism for the recovery effect. In step S7, the multi-source operating status is re-acquired within a preset observation period to construct a post-recovery status window. Specifically, the test is not immediately considered passed after the recovery operation is completed. Instead, a short observation period (e.g., 5 minutes or a specific number of I / O operations) is set, during which seven types of status data of the same dimensions as in Embodiment 2 are continuously collected to form a new status window. This step aims to capture potential hidden instability factors in the early stages of recovery, such as temporary performance jitter or residual errors caused by incomplete firmware initialization.
[0108] In step S8, a time-series anomaly detection model with identical parameters to that before the anomaly occurred is invoked to perform anomaly detection on the recovered state window. Specifically, it is emphasized here that "identical parameters" is the logical cornerstone for verifying effectiveness. The system must load a model instance with the same version, weights, standardized parameters, and thresholds as when the recovery was triggered for inference. If a different model is used or the parameters are updated, it becomes impossible to distinguish whether the "disappearance of the anomaly" is due to the successful repair of the SSD or a drift in the judgment criteria. Only by using the same "ruler" for verification can the true effect of the recovery action be objectively evaluated.
[0109] As an optional implementation, step S8 specifically includes sub-steps S801 and S806. In sub-step S801, after completing the underlying recovery of the SSD under test and the reconstruction of the test environment, the system continuously collects the multi-source operating status of the SSD under test within a preset recovery observation period, and calls a time-series anomaly detection model with the same model structure, model weights, standardized parameters, and anomaly judgment parameters as before the anomaly occurred to perform anomaly detection on the recovered state window, thereby obtaining the time-point anomaly score corresponding to each detection time point within the recovery observation period. The average time-point anomaly score within the recovery observation period is calculated, and the anomaly state recovery parameters are calculated based on the average time-point anomaly score after recovery.
[0110]
[0111] in, Indicates parameters for abnormal state recovery. This represents the average abnormality score within the recovery observation period. This represents the normalization parameter for outlier scores. This represents the truncation function, limiting the result to between 0 and 1. When the average anomaly score after recovery is low, it indicates that the tested SSD deviated less from the normal timing pattern during the recovery observation period, and $R_A$ is larger. When the average anomaly score after recovery approaches or reaches the severe anomaly threshold, it indicates that the recovered SSD still exhibits significant anomaly characteristics. The corresponding decrease occurs. Therefore, the abnormal state recovery parameter can convert the "abnormality level" output by the model into a "recovery level" that facilitates subsequent comprehensive evaluation. In sub-step S802, the baseline I / O latency corresponding to the normal testing phase of the tested SSD is obtained, and the average I / O latency within the recovery observation period is obtained. The I / O latency recovery parameter is then calculated:
[0112]
[0113] in, Indicates I / O latency recovery parameters. This represents the average I / O delay measured during the recovery observation period. This represents the baseline I / O latency corresponding to the normal testing phase. Indicated for preventing A hyperparameter that is 0 or too small will result in division by zero. Before the tested SSD malfunctions, the system can select a preset time range of I / O latency statistics as the baseline I / O latency from its normal testing phase. The baseline I / O latency can be the average latency, or a sliding average from the stable phase or a preset baseline value can be used depending on the specific test task. After recovery, the system re-statistically calculates the I / O latency within the recovery observation period and compares the recovered average I / O latency with the baseline I / O latency of the normal phase. The smaller the relative deviation between the two, the closer the recovered I / O performance is to the normal state before the malfunction, and the closer the I / O latency recovery parameter is to 1. The larger the deviation between the two, the more it indicates that although the SSD can be reinstalled and perform I / O operations, its performance state has not been fully recovered, and the I / O latency recovery parameter is correspondingly reduced. In sub-step S803, the heartbeat continuity parameter is calculated based on the number of heartbeats that should be received within the recovery observation period and the number of valid heartbeats actually received:
[0114]
[0115] in, This represents the continuous parameters of the heartbeat. This indicates the actual number of valid heartbeats received during the recovery observation period. This represents the theoretical number of heartbeats that should be received within the recovery observation period according to the preset heartbeat sending cycle. After the tested SSD recovers and re-enters the testing phase, the test host or test tool continuously monitors the device's heartbeat according to the preset cycle. The system counts the valid heartbeats received within the recovery observation period and compares them with the theoretical number of heartbeats that should be received to determine whether the communication link and operating status after recovery have continuous stability. When no heartbeat loss or interruption occurs within the recovery observation period, the actual number of valid heartbeats is basically consistent with the theoretical number that should be received, and the heartbeat continuity parameter is close to 1. When there are intermittent heartbeat absences, communication timeouts, or temporary system disconnection after recovery, the actual number of valid heartbeats decreases, and the heartbeat continuity parameter decreases accordingly. In sub-step S804, the number of repeated abnormal events with the same abnormal identifier as the abnormal context before recovery within the recovery observation period is counted, and the abnormal non-recurrence parameter is calculated:
[0116]
[0117] in, Indicates parameters that do not recur due to abnormality. This indicates the number of duplicate exception events with the same exception identifier as the exception context before recovery within the recovery observation period. This indicates the preset allowable number of recurring exception events. When an exception occurs, the system records the corresponding exception identifier in the exception context. The exception identifier can be determined based on the exception type, error code, log semantic category, device status code, or a combination of multiple exception features. For example, for a command timeout exception, the corresponding command type, timeout error code, and relevant log features can be combined to form the exception identifier. For a disk loss exception, the device enumeration loss event and corresponding bus status can be used as the exception identifier. During the recovery observation period, the system continuously checks whether events with the same exception identifier as before recovery reappear and counts the number of times they occur. When no recurrence of the same type of exception occurs, When the number of recurring abnormal events is 0, the non-recurrence parameter approaches 1. As the number of recurring similar abnormal events increases, the non-recurrence parameter gradually decreases. When the number of recurring abnormal events reaches or exceeds the preset allowable number, the non-recurrence parameter decreases to 0. In sub-step S805, the hard abnormality recovery parameter is set according to the hard abnormality rule hit status within the recovery observation period:
[0118]
[0119] in, This represents the hard anomaly recovery parameter. Hard anomaly rules are used to detect abnormal states with deterministic fault meanings that should not be judged solely based on statistical anomaly scores, including but not limited to device disk failure, inability to complete Boot enumeration, controller errors, firmware integrity detection failure, system blue screen, or kernel crash. When no preset hard anomaly rule is hit during the entire recovery observation period, the hard anomaly recovery parameter is set to 1. Once any hard anomaly rule is hit during the recovery observation period, the hard anomaly recovery parameter is set to 0. In sub-step S806, the recovery stability R is calculated by weighted summation based on the abnormal state recovery parameter, I / O latency recovery parameter, heartbeat continuity parameter, anomaly non-recurrence parameter, and hard anomaly recovery parameter. All weight coefficients are non-negative and can be predetermined based on different SSD models, automated test task types, and fault risk levels. During the execution of the current automated test task, the weight coefficients remain fixed to ensure consistency of evaluation scales between different recovery stages. By weighted and integrated from the above multiple dimensions, this embodiment jointly evaluates the recovery results from five dimensions: abnormal state, performance state, communication continuity, abnormal recurrence, and deterministic hard faults. This further expands the traditional single recovery action into a quantifiable and verifiable recovery effect evaluation process, providing a unified judgment basis for stable recovery, recovery under observation, and recovery failure classification in subsequent step S9, and forming a closed-loop automated testing mechanism of "anomaly detection - automatic recovery - recovery verification - continued testing or recovery again".
[0120] In step S9, if the detection results continuously meet the preset stability conditions, the recovery is confirmed as successful, and the remaining tests continue. Specifically, the stability conditions typically require that the anomaly score remains below the warning threshold throughout the observation period, no hard anomaly rules are hit, and key indicators such as heart rate and I / O latency return to the normal baseline. Only when these conditions are continuously met does the system finally mark the recovery as successful, remove the observation status, and allow the testing process to continue. If an anomaly is detected again during the observation period, a new round of recovery or direct isolation is triggered based on the anomaly level. This rigorous secondary verification mechanism eliminates the phenomenon of "false recovery," ensuring the credibility of the final test report and the consistency of product quality.
[0121] As an optional implementation, step S9 specifically includes sub-steps. When If the preset hard anomaly rule is not hit for Q consecutive detection cycles, the SSD under test is determined to be in a stable recovery state, the recovery is confirmed to be successful, and the remaining automated test tasks continue to be executed from the recovery execution point. If the preset hard anomaly rule is not met, the SSD under test will be identified as being in a recovery observation state, the recovery observation period will be extended, and the I / O test load will be reduced according to a preset ratio, while the sampling frequency of multi-source operation will be increased. The recovery stability will be recalculated after the extended observation period ends. If the recalculated recovery stability reaches... If the recovery is successful, proceed with the remaining tests. If the recalculated recovery stability is still within [a certain range], then [the recovery is considered successful]. Within this context, the decision is made based on the number of previous recovery attempts to either perform a controlled recovery again or move the tested SSD into isolation. When Alternatively, if a severe hard anomaly rule is hit again within the recovery observation period, the SSD under test will be identified as a recovery failure. If the current number of automatic recovery attempts is less than the preset recovery attempt limit, the anomaly evolution degree is recalculated based on the recovered status window, and a controlled recovery control code or an immediate recovery control code is generated based on the recalculated anomaly evolution degree. The process then returns to S4 to execute the recovery again. If the current number of automatic recovery attempts reaches the preset recovery attempt limit, an isolation control code is generated, the current automated test task of the SSD under test is terminated, and the anomaly context, previous anomaly evolution degrees, previous recovery stability, and recovery operation records are sent to the server workstation. This represents the first recovery stability threshold. This represents the second recovery stability threshold, and satisfies: This embodiment evaluates recovery results by grading stability and forms a multi-level closed-loop control mechanism based on the number of automatic recovery attempts. When the recovery effect is good, testing continues. When the recovery effect is critical, testing pressure is reduced and observation time is extended. When recovery fails but automatic recovery conditions are still met, the anomaly evolution degree is recalculated and a matching recovery level is selected. When the number of recovery attempts reaches the limit or there is an unrecoverable risk, isolation measures are implemented.
[0122] Example 5
[0123] This invention also provides an automated testing system for SSD products, applied to the aforementioned automated testing method for SSD products, comprising:
[0124] processor.
[0125] The memory stores computer-readable instructions, which, when executed by a processor, implement an automated testing method for SSD products as described in the method embodiments.
[0126] The automated testing system for SSD products provided by this invention can execute the above-mentioned automated testing methods for SSD products and achieve the same or similar technical effects. To avoid duplication, this invention will not elaborate further.
[0127] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
[0128] The following points need to be explained:
[0129] (1) The accompanying drawings of the embodiments of the present invention only involve the structures involved in the embodiments of the present invention. Other structures can refer to the general design.
[0130] (2) For clarity, the thickness of layers or regions is enlarged or reduced in the drawings used to describe embodiments of the invention, i.e., these drawings are not drawn to scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being “above” or “below” another element, the element may be “directly” located “above” or “below” the other element or there may be intermediate elements.
[0131] (3) Where there is no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.
[0132] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. An automated testing method for SSD products, characterized in that, include: S1: Obtain SSD automated testing tasks and configure the timing anomaly detection model and device mapping relationship; S2: Collect the multi-source operating status of the SSD under test and construct a status window; S3: Based on the aforementioned time-series anomaly detection model and preset hard anomaly rules, generate recovery control codes and anomaly contexts; S4: In response to the recovery control code, save the test checkpoint and switch to the backup storage device; S5: Based on the device mapping relationship, perform low-level hardware reset and firmware reconstruction on the tested SSD; S6: Rebuild the test environment and generate a recovery execution point based on the exception context to continue testing.
2. The automated testing method for SSD products according to claim 1, characterized in that, S1 specifically includes: S101: Based on the automated test task, establish a task mapping relationship between the SSD under test, the test host, the backup storage device, the hardware control component and the server workstation; S102: Load the fixed model parameters that match the SSD under test, and the model parameters remain fixed during the test.
3. The automated testing method for SSD products according to claim 1, characterized in that, S2 specifically includes: S201: Synchronously collect heartbeat status, operating system status, health parameters, input / output status, I / O latency status, log semantic features, and time interval features between key events to construct a running status vector; S202: Combine the most recent running state vectors collected continuously into the state window.
4. The automated testing method for SSD products according to claim 1, characterized in that, S3 specifically includes: S301: By projecting different state features onto the same latent space through the input mapping layer, an input temporal feature sequence representing the current SSD operating state is formed; S302: Employ an anomaly attention mechanism to model the correlation differences of the input time-series feature sequences in order to quantify the degree of deviation between the current SSD operating state and the normal time-series mode; S303: A spatiotemporal gating fusion mechanism is used to fuse the input time-series feature sequences to obtain spatiotemporal fused features; S304: A multi-layer stacked encoder is used, which integrates the anomaly attention mechanism and the spatiotemporal gating fusion mechanism in each layer, and integrates the spatiotemporal fusion features extracted from each layer to obtain the encoder output; S305: The decoder performs decoding and reconstruction based on the encoder output to obtain the decoder output; S306: Calculate the time point anomaly score based on the reconstruction error between the encoder output and the decoder output and the correlation difference; S307: Based on the time point anomaly score and the preset hard anomaly rule, perform state classification and generate the recovery control code and anomaly context.
5. The automated testing method for SSD products according to claim 4, characterized in that, Specifically, S307 includes: S3071: The state of the SSD under test is divided into normal state, degraded state, severely abnormal state and candidate state that cannot be automatically recovered; S3072: Map and generate test continuation control code, controlled recovery control code, immediate recovery control code and isolation control code respectively; The degradation state is determined based on the time point anomaly score exceeding the warning threshold and meeting the preset persistence condition, while the severe anomaly state is determined based on the time point anomaly score exceeding the severe anomaly threshold or hitting the preset hard anomaly rule.
6. The automated testing method for SSD products according to claim 5, characterized in that, S4 specifically includes: S401: In response to the recovery control code being a controlled recovery control code or an immediate recovery control code, pause the issuance of new requests to the SSD under test, and save the test checkpoint before the anomaly occurred and the anomaly context. S402: Control the SSD under test to power off and switch the boot path of the backup storage device. After detecting that the backup environment is ready, output to the subsequent recovery process.
7. The automated testing method for SSD products according to claim 1, characterized in that, S5 specifically includes: S501: The control pin shorting channel is closed, and the tested SSD is powered on in a controlled manner while the shorting is maintained, so that it enters the boot state. S502: After the preset short-circuit holding condition is met or the Boot enumeration signal is detected, the pin short-circuit channel is disconnected; S503: After the backup storage device enumerates the tested SSD, push the mass production tool to perform firmware reconstruction.
8. The automated testing method for SSD products according to claim 1, characterized in that, S6 specifically includes: S601: Cancel the boot path of the backup storage device and restore the system disk boot path of the tested SSD; S602: Deploy the target operating system and testing tools onto the SSD under test; S603: Generate the recovery execution point based on the test checkpoint in the exception context, and continue executing the test from the recovery execution point.
9. The automated testing method for SSD products according to claim 1, characterized in that, Following S6, it also includes: S7: Re-acquire the multi-source operating status within the preset observation period and construct the restored status window; S8: Call the time-series anomaly detection model with the same parameters as before the anomaly occurred, and perform anomaly detection on the recovered state window; S9: When the test results continuously meet the preset stability conditions, confirm successful recovery and continue the remaining tests.
10. An automated testing system for SSD products, characterized in that, include: processor; A memory storing computer-readable instructions, which, when executed by the processor, implement an automated testing method for SSD products as described in any one of claims 1 to 9.
