Multi-dimensional access load evaluation method and system for memory particle lifetime prediction, electronic device, and storage medium

Abstract

The application discloses a kind of multi-dimensional access load evaluation methods and systems for memory grain life prediction, electronic equipment, storage medium, wherein the method comprises: obtaining the write access stream corresponding to target memory grain, online monitoring is carried out to write access stream, irreversible content summary calculation is carried out to write data under the premise that the plaintext content of write data is not output, and content form statistics are obtained;Space dimension load characteristics are extracted based on write address information, time dimension load characteristics are extracted based on the time sequence of write access stream, and operation dimension load characteristics are extracted based on write operation type or write granularity, to form load feature vector by fusion;Load feature vector is input into life prediction model, and the life consumption measure or residual life prediction result of target memory grain is output.Can solve the problem that traditional reliability evaluation is mainly based on simple load intensity or temperature index, and cannot provide fine basis for memory grain life prediction.

Description

Technical Field

[0001] This application relates to the field of computers, and more specifically, to a multi-dimensional access load assessment method and system for predicting memory chip lifetime, electronic devices, and storage media. Background Technology

[0002] Modern computing systems place increasingly higher demands on memory reliability, especially with the application of non-volatile memory, high-density DRAM, and emerging memory technologies. The durability of memory chips has become a key bottleneck for system stability. Traditional reliability assessments are often based on simple load intensity or temperature indicators, which fail to characterize the combined effects of multiple factors such as access frequency, data patterns, spatial distribution, and timing characteristics on the physical wear and tear of memory under real-world workloads. This makes it impossible to provide a refined basis for predicting the lifespan of memory chips. Summary of the Invention

[0003] The summary section introduces a series of simplified concepts, which will be further explained in detail in the detailed description section. The summary section of this invention is not intended to limit the key features and essential technical features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.

[0004] To address the problem that traditional reliability assessments, which often rely on simple load intensity or temperature indicators, struggle to characterize the synergistic impact of multiple factors such as access frequency, data patterns, spatial distribution, and timing characteristics on memory physical wear and tear under real-world workloads, and thus fail to provide a refined basis for memory chip lifetime prediction, this invention proposes a multi-dimensional access load assessment method for memory chip lifetime prediction. The method includes:

[0005] Obtain the write access stream corresponding to the target memory particle, monitor the write access stream online, and perform irreversible content digest calculation on the write data without outputting the plaintext content of the written data to obtain content morphology statistics. The write access stream includes write address information and write data. The content morphology statistics include at least one of the following to characterize the bit flip density, bit entropy, and repetitive block ratio of the written data.

[0006] Based on the write address information, spatial load features are extracted; based on the time series of the write access stream, temporal load features are extracted; based on the write operation type or write granularity, operation load features are extracted; and the temporal load features, spatial load features, operation load features, and content format statistics are fused together to form a load feature vector.

[0007] The load feature vector is input into the lifetime prediction model, and the lifetime consumption metric or remaining lifetime prediction result of the target memory particle is output.

[0008] Secondly, the present invention also proposes a multi-dimensional access load assessment system for predicting memory particle lifetime, the apparatus comprising:

[0009] The acquisition unit is used to acquire the write access stream corresponding to the target memory particle, monitor the write access stream online, and perform irreversible content digest calculation on the write data without outputting the plaintext content of the written data to obtain content morphology statistics. The write access stream includes write address information and write data. The content morphology statistics include at least one of the following to characterize the bit flip density, bit entropy, and repetitive block ratio of the written data.

[0010] The fusion unit is used to extract spatial load features based on the write address information, extract time load features based on the time series of the write access stream, extract operation load features based on the write operation type or write granularity, and fuse the time load features, spatial load features, operation load features and the content form statistics to form a load feature vector.

[0011] The prediction unit is used to input the load feature vector into the lifetime prediction model and output the lifetime consumption metric or remaining lifetime prediction result of the target memory particle.

[0012] Thirdly, an electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program stored in the memory to implement the steps of the multidimensional access load assessment method for memory particle lifetime prediction as described in any of the first aspects above.

[0013] Fourthly, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the multi-dimensional access load assessment method for memory particle lifetime prediction as described in any of the preceding claims of the first aspect.

[0014] In summary, the multi-dimensional access load assessment method for memory particle lifetime prediction proposed in this application obtains the write access flow corresponding to the target memory particle, monitors the write access flow online, and performs irreversible content digest calculation on the write data without outputting the plaintext content of the written data to obtain content morphology statistics. The write access flow includes write address information and write data, and the content morphology statistics include at least one of the following to characterize the bit flip density, bit entropy, and duplicate block ratio of the write data. Spatial dimension load features are extracted based on the write address information, temporal dimension load features are extracted based on the time series of the write access flow, and operational dimension load features are extracted based on the write operation type or write granularity. The temporal dimension load features, spatial dimension load features, operational dimension load features, and content morphology statistics are then fused with the content morphology statistics to form a load feature vector. The load feature vector is input into a lifetime prediction model, which outputs a lifetime consumption metric or remaining lifetime prediction result for the target memory particle. This improves the resolution and accuracy of lifetime prediction. By introducing content morphology statistics and fusing them with multi-dimensional load features, the model can distinguish workloads with the same write volume but significantly different wear levels, thereby reducing prediction bias. It enhances interpretability and transferability. Content morphology statistics provide an interpretive dimension closer to physical stress, while spatial and temporal features offer interpretive dimensions for wear distribution and superposition patterns, making the model output easier for engineers to understand and easier to transfer to different workload combinations. It also reduces privacy and operational overhead risks. Irreversible summarization avoids outputting plaintext content, and online sampling and window statistics control computational burden, enabling the solution to run on production systems for extended periods. Furthermore, it supports finer-grained lifetime management strategies. The output lifetime consumption metrics can be used for granular, block-level, or region-level management. Combined with strategies such as hotspot avoidance, write shaping, and load migration, it can delay premature local failures and improve overall lifetime. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0016] Figure 1 A schematic diagram of a multi-dimensional access load assessment method for predicting memory particle lifetime is provided in an embodiment of this application.

[0017] Figure 2 A schematic diagram of a multi-dimensional access load assessment system for predicting memory particle lifetime is provided in this application embodiment;

[0018] Figure 3This is a schematic diagram of an electronic device structure for multi-dimensional access load assessment based on memory chip lifetime prediction, provided as an embodiment of this application. Detailed Implementation

[0019] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The technical solutions of the embodiments of this application will now be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0020] Understandably, the lifespan degradation of memory chips essentially stems from the cumulative damage to materials and device structures under long-term electrical and thermal stress. For many storage media, damage depends not only on the number of writes or bytes written, but also strongly on the form of the written data, such as the density of bit flips, the randomness of data blocks, and the probability distribution of repeated writes of the same data. These content forms alter the frequency of cell programming state transitions, the statistical distribution of local charge injection and release, the execution path of write pulses and verification processes, and the write amplification and background cleanup behavior on the controller side, thereby creating stress accumulation of varying intensities at the physical layer.

[0021] To address the issue that traditional reliability assessments, which often rely on simple load intensity or temperature metrics, struggle to characterize the synergistic impact of multiple factors such as access frequency, data patterns, spatial distribution, and timing characteristics on memory physical wear under real-world workloads, and thus fail to provide a refined basis for memory chip lifetime prediction, this paper proposes a multi-dimensional access load assessment method for memory chip lifetime prediction. Please refer to [link / reference]. Figure 1 , Figure 1 This is a flowchart illustrating a multi-dimensional access load assessment method for predicting memory particle lifetime provided in an embodiment of this application. In some examples, the method further includes steps S110 to S130.

[0022] S110, obtain the write access stream corresponding to the target memory chip, monitor the write access stream online, perform irreversible content digest calculation on the write data without outputting the plaintext content of the written data, and obtain content form statistics. The write access stream includes write address information and write data. The content form statistics include at least one of the following to characterize the bit flip density, bit entropy, and repeating block ratio of the written data.

[0023] S120, extract spatial dimension load features based on the write address information, extract time dimension load features based on the time series of the write access stream, extract operation dimension load features based on write operation type or write granularity, and fuse the time dimension load features, spatial dimension load features, operation dimension load features with the content form statistics to form a load feature vector.

[0024] S130, input the load feature vector into the lifetime prediction model, and output the lifetime consumption metric or remaining lifetime prediction result of the target memory particle.

[0025] Understandably, without revealing the plaintext of the written data, the content form of the written data is transformed into an irreversible statistic, and together with the load characteristics of time, space and operation dimensions, they form a feature vector. This enables the prediction model to distinguish between load types with the same surface writing amount but significant differences in internal stress, and thus output a life consumption metric or remaining life that is closer to the real wear process.

[0026] For example, the system obtains the write access stream corresponding to the target memory particle at any location in the host-side driver layer, file system layer, memory controller-side command queue, or telemetry channel. The write access stream includes at least write address information and write data, and may further include fields such as write length, write granularity identifier, write request arrival timestamp, and queue depth. The lifespan degradation of memory chips is not only determined by the number of writes, but also strongly correlated with the statistically significant content form of the written data. Content form alters the frequency and magnitude of cell state transitions, the execution probability of programming verification paths, and the rate of local stress accumulation. Therefore, quantifiable indicators need to be extracted without revealing plaintext. Specifically, the system performs irreversible content digest calculations on the written data only within the volatile buffer. A windowed online processing method based on preset time windows or preset write byte windows is adopted. After dividing the written data into blocks of fixed size, irreversible mapping is performed on each data block to obtain a digest representation. This irreversible mapping can employ hash mapping, multi-hash bucket counting, low-dimensional signatures after random projection, or combinations thereof. The digest output is insufficient to recover the plaintext of the written data. Furthermore, the written data itself is discarded immediately after digest calculation to ensure that no plaintext is output or written to disk. When constructing content form statistics based on the digest representation, at least one of bit-flip density and bit entropy / repeating block ratio can be statistically analyzed within the window. Bit flip density characterizes the proportion of bit state changes during adjacent writes or same-address overwrites. It can be achieved by statistically analyzing the differences in signatures between adjacent blocks or by counting plaintext immediately after a single scan. Bit entropy characterizes the randomness of written data and can be estimated by the uniformity of hash bucket distribution. The proportion of duplicate blocks characterizes the proportion and concentration of the same digest value repeated within a window. For example, for high-randomness ciphertext writes, bit entropy is usually higher and signature distribution is more uniform. However, for services containing a large number of duplicate templates or zero pages, the proportion of duplicate blocks is usually significantly higher and may lead to an increase in local overwrites. Through the above irreversible digest and statistical construction, without introducing the risk of plaintext leakage, the influencing factors of the written content are transformed into quantifiable features that can be calculated online. This allows subsequent lifetime assessments to distinguish the differences in lifetime consumption caused by the same amount of writing but different state flip frequencies, randomness, and hotspot repetition, thereby reducing systematic misjudgments caused by relying solely on the number of writes or bandwidth.

[0027] For example, the system characterizes write access behavior from spatial, temporal, and operational dimensions, respectively, and integrates these with the aforementioned content form statistics to form a load feature vector. The stress accumulation of memory particles is related to the content form, the distribution of writes in the address space, the superposition of writes over time, and the form of write operations. Using only the content form or only the access count is insufficient to characterize true wear. Specifically, the spatial load feature is calculated based on write address information to characterize the distribution of wear within the particle. It can include at least the address hotspot index, hotspot persistence, and local rewrite ratio. The address hotspot index can be obtained by aggregating write addresses at the page or block granularity and calculating the degree of uneven distribution. Hotspot persistence can be obtained by statistically analyzing the probability of maintaining high write intensity in the same address interval within multiple consecutive windows. The local rewrite ratio can be obtained by statistically analyzing the percentage of times the same address interval is overwritten in a short period of time. The temporal load feature is calculated based on the time series of the write access flow to characterize whether the stress is sudden, sustained high intensity, or periodic. The system employs a combination of methods, including at least write burstability, distribution of sustained high-intensity segment lengths, and write rhythm stability. Write burstability can be obtained by the deviation of the peak write rate within a window from the mean, while distribution of sustained high-intensity segment lengths can be obtained by statistically analyzing the segment lengths where the write intensity exceeds a threshold in consecutive windows. Operational load features, calculated based on write operation type or write granularity, characterize differences in internal execution paths and potential write amplification risks. These features include at least write granularity distribution, alignment deviation features, and overwrite ratio. Alignment deviation features can be obtained by determining whether writes frequently cross the natural boundaries of the medium, leading to additional internal movement or reorganization. The system then normalizes these features and integrates them with content morphology statistics. In one implementation, feature concatenation forms a unified vector; in another, higher weights are assigned to certain features based on the level of content morphology statistics. For example, when bit flip density and bit entropy are both high, the sensitivity weights to burstability and hotspot persistence are increased to reflect the stronger destructive impact of highly random and frequently flipped content on lifespan in burst hotspot scenarios. For example, if two services have the same write volume, where service A has higher bit entropy and higher burstiness, forming a long-term hotspot for continuous writing, while service B has higher bit entropy but a more uniform spatial distribution and a stable write rhythm, then the fusion vector can explicitly distinguish the differences in stress superposition and wear distribution between the two. Therefore, by uniformly encoding the write content, location, and form into feature vectors that can be input into the model, the lifetime prediction model can learn the coupling relationships between multi-dimensional features. This improves interpretability and generalization stability when facing complex workloads and reduces the bias caused by relying on only a single-dimensional feature.

[0028] For example, the system inputs the load feature vector into the lifetime prediction model and outputs the lifetime consumption metric or remaining lifetime prediction result of the target memory particle. The lifetime prediction model can learn the mapping relationship between the load feature vector and the lifetime degradation characterization through historical data, thereby transforming the abstract access load into a quantifiable lifetime consumption estimate. The lifetime prediction model can be a statistical model or a machine learning model. When establishing training samples, sample input vectors are generated in units of preset time windows or preset write volume windows, and lifetime-related observations are used as supervision labels or regression targets. These observations can come from reliability telemetry indicators or maintenance records, such as the growth rate of error correction intensity over time, the upward trend of the number of retries, the rate of decrease in available blocks, changes in error occurrence density, or failure events triggered under specific threshold conditions. After the model training is completed, the system continuously generates feature vectors by window and calculates lifetime output during the online inference phase. The lifetime consumption metric can be expressed as the lifetime consumption rate per unit time, the cumulative lifetime consumption value, or the amount of health decline. The remaining lifetime prediction result can be expressed as the remaining time or remaining writeable amount estimate before the failure threshold. To ensure output stability, the system can perform smoothing or confidence assessment on continuous window outputs. For example, when the sampling rate decreases or some features are missing, the output confidence is reduced and abnormal spikes are suppressed. For example, under the same write bandwidth, if the model detects an increase in bit entropy accompanied by enhanced hotspot persistence and increased burstiness, the output lifetime attrition rate will increase, and a shorter remaining lifetime estimate will be given. However, if the bit entropy increases but its spatial distribution is uniform and the rhythm is stable, the increase in output will be relatively limited, thus avoiding the generalization of all high-random writes as high-risk. Therefore, by utilizing content morphological statistics and multidimensional load feature vectors, the sensitivity of lifetime prediction to content-related stress, hotspot wear, burst superposition stress, and differences in operation paths is significantly improved. This makes the prediction results closer to the actual wear process and can be used for granular lifetime management, risk warning, or resource scheduling decisions. At the same time, since the content summary is irreversible and does not output plaintext, it can run online for a long time while meeting privacy and security constraints.

[0029] In some examples, performing an irreversible content digest computation on the written data includes:

[0030] The written data is divided into blocks according to a preset block size;

[0031] Perform a random projection operation or a hash mapping operation on each data block to obtain a digest result, and calculate at least one of the bit flip density, the bit entropy and the proportion of repeated blocks based on the digest result, wherein the digest result irreversibly represents the statistical form of the written data and is insufficient to recover the plaintext content of the written data.

[0032] For example, although the plaintext content of the written data may involve privacy or commercially sensitive information, the statistically significant morphological characteristics of the written data, such as the density of bit state changes, the level of randomness, and the degree of block content repetition, significantly affect the accumulation of electrical stress in memory chips, the probability of triggering programming verification paths, and the distribution of local wear. Therefore, irreversible digest representation can be used to extract lifetime-related statistical information to avoid content leakage. In specific implementation, the preset block size can be determined based on the minimum programming unit of the medium or the system write granularity. For example, a block size consistent with the page size or an integer multiple thereof can be selected to ensure that the digest statistics and physical write behavior have a consistent spatial scale. At the same time, in high-throughput write scenarios, the data blocks after being divided can be sampled at a preset sampling rate to participate in digest calculation to control overhead. When performing hash mapping operations on each data block, an irreversible hash function can be used to output a fixed-length digest value, and multiple independent hashes can be further constructed to obtain a multi-bucket counting vector, so that the digest result only retains the statistical form and not the reversible information. The proportion of duplicate blocks can be obtained by statistically analyzing the proportion of data blocks with the same digest value in the same window and their concentration, thereby reflecting the... Within this window, there is a tendency for content reuse or hotspot overwriting. When performing random projection operations on each data block, the bit sequence or byte sequence of the data block can be regarded as a high-dimensional vector. A pre-generated and fixed random projection matrix or random symbol vector is used to project it and quantize it into a low-dimensional signature. The signature can be used to estimate bit entropy or randomness level. For example, when the projected signature shows a nearly uniform distribution in multiple dimensions, it can be determined that the data form is closer to random. Conversely, if the signature distribution is highly biased, it indicates that the data form is more structured. The calculation of bit flip density can be achieved by statistically analyzing the differences between digest results of adjacent write windows or adjacent overwriting. For example, in the scenario of repeated overwriting of the same address range in a short period of time, the intensity of bit state change can be approximately characterized by comparing the difference between the two digest signatures, thereby reflecting the frequency of state transition without saving plaintext. To ensure irreversibility and privacy, the system can only retain the digest result and the statistics obtained therefrom, and release the written data body immediately after the digest calculation is completed. At the same time, it avoids outputting the digest result together with auxiliary information that can be used to recover plaintext, so that the digest result is insufficient to recover the plaintext content of the written data.For example, when the data written by the business is a compressed ciphertext stream, its digest signature is statistically closer to random, the bit entropy estimate is higher, and the proportion of duplicate blocks is lower. Based on this, the prediction model can identify the load pattern that is more likely to cause frequent state transitions and higher programming stress. When the data written by the business contains a large number of templated fields or zero-fill pages, the probability of the digest value being repeated is higher, and the proportion of duplicate blocks is significantly increased. Combined with address hotspot characteristics, the wear risk of local areas being repeatedly overwritten can be further identified. Through the above irreversible content digest calculation, without the constraint of outputting plaintext written data, the form of written content is transformed into statistical features that can be monitored online and quantified. This allows the lifetime prediction model to distinguish the stress accumulation differences caused by the same amount of writing but different data forms. Furthermore, through mechanisms such as block division, sampling, and fixed-length digests, the computation and storage overhead is controlled within a range suitable for long-term online operation.

[0033] In some examples, the content format statistics also include a content stability index, which characterizes the stability of the data state distribution within a continuous write window. The method further includes:

[0034] The change in bit flip density or bit entropy within adjacent time windows is statistically analyzed to obtain the intensity of content change.

[0035] The content change intensity is normalized with the number of writes or the number of bytes written to obtain the content change rate per unit of write volume, which is used as the content stability index.

[0036] The input feature weights or lifespan consumption estimation coefficients of the lifespan prediction model are adaptively adjusted based on the content stability index.

[0037] For example, considering that the same average bit-flip density or bit entropy does not necessarily correspond to the same lifetime consumption, when the statistical form of the written content changes frequently over time, it often means that the write mode switches more frequently between different types of data. This may lead to storage units experiencing more complex state transition combinations, more frequent switching of compression and merging strategies on the controller side, and more uneven accumulation of local stress. Therefore, relying solely on the static content form statistics of a single window can easily underestimate or overestimate the destructiveness of certain loads. In specific implementation, the system segments the write access stream according to a preset window. The window can be a fixed time window, a fixed write byte window, or a composite window combining both. In each window, at least one of the bit-flip density or bit entropy is calculated. Then, the changes in the bit-flip density or bit entropy in adjacent time windows are statistically analyzed to obtain the content change intensity. The content change intensity can be calculated using the cumulative value of the absolute change, the mean square change, the quantile change, or the weighted change statistics. The weighted change statistics can be used to emphasize the impact of sudden changes on lifetime. For example, when the bit entropy of a certain window jumps from a low level to a high level, it can be given a higher weight to reflect the change in data form from structured to structured. The system addresses the increased programming stress caused by content switching to near-random content. Furthermore, it normalizes the intensity of content change with the number of writes or the number of bytes written, obtaining the content change rate per unit of write volume as the content stability index. This eliminates the scaling effect caused by differences in write volume across different windows, ensuring comparability between high-throughput and low-throughput windows. For example, in scenarios with fewer writes but drastic fluctuations in data form, the content change rate per unit of write volume will significantly increase, indicating unstable content form. At the model usage level, the system adaptively adjusts the input feature weights or lifetime consumption estimation coefficients of the lifetime prediction model based on the content stability index. In one implementation, when the content stability index exceeds a preset threshold, the input feature weights related to write randomness, bit flip intensity, and burstiness are increased, while the feature weights reflecting only the average write volume are decreased. This makes the model more focused on the rapid switching of content form, a factor more sensitive to lifetime. In another implementation, the system directly applies an adjustment term monotonically correlated with the content stability index to the lifetime consumption estimation coefficients, resulting in a higher lifetime consumption estimate for the same write volume under unstable content conditions.For example, a certain business writes a large amount of structured logs during the day, resulting in low bit entropy and gradual changes. However, at night, encrypted backups are written, causing a significant increase in bit entropy and frequent fluctuations between windows. If only the average bit entropy or average write volume is used, it may not be possible to accurately assess the risk of lifetime consumption. However, by introducing content stability indicators, the system can identify that the content change rate per unit write volume is higher during the nighttime phase, thereby adjusting the model weights or lifetime consumption coefficients to make the prediction results closer to the actual stress accumulation. Through the construction and application of the above-mentioned content stability indicators, it is possible not only to characterize the static form of written content within a single window, but also to characterize the form evolution and switching intensity across windows. This improves the sensitivity and robustness of lifetime prediction to phase-change content loads and multi-business mixed loads, and reduces the lifetime consumption estimation bias caused by feature staticization in scenarios with frequent content form switching.

[0038] In some examples, it also includes:

[0039] The content format statistics are subjected to interval binning or differential privacy noise injection processing to limit the identifiable contribution of a single data write to the statistical results.

[0040] The processed content morphology statistics are concatenated or weighted with the time-dimensional load features, spatial-dimensional load features, and operational-dimensional load features to form the load feature vector used as input to the lifetime prediction model.

[0041] Understandably, in load assessment for memory granular lifetime prediction, content format statistics such as bit flip density, bit entropy, duplicate block ratio, content stability indicators, and boundary content format change indicators are essentially statistical features extracted from written data. Although these features do not directly output plaintext, in some scenarios, if the features are continuously reported in high-precision numerical form or can be observed externally for a long time, they may still be used to infer business type, data structure changes, or whether a specific write event has occurred, forming an identifiable contribution to a single write, thus bringing privacy and compliance risks. Meanwhile, the goal of lifetime prediction models is not to identify single writes, but to identify the cumulative effect of long-term load patterns on wear. Therefore, in most scenarios, the model does not need high-precision continuous values; it only needs to be able to distinguish between low, medium, and high risk levels or maintain statistical ordering of feature expressions. Based on the above reasons, interval binning or differential privacy noise injection can be used to process content format statistics. Interval binning maps continuous values ​​to finite-level intervals, reducing the identifiability of small feature changes caused by a single write and enhancing resistance to reverse inference. Differential privacy noise injection, by superimposing random perturbations that satisfy the privacy budget onto the statistics, ensures that the impact of the presence or absence of any single write on the output distribution is strictly limited, thus theoretically restricting the identifiable contribution of a single write to the statistical results. Subsequently, the processed content morphology statistics are fused with temporal, spatial, and operational dimension features into a load feature vector, enabling the model to retain lifetime-related information while reducing the risk of exposing sensitive details.

[0042] For example, the system employs differential privacy noise injection to protect content format statistics. Specifically, the system determines the sensitivity range for each statistic, i.e., the maximum impact a single write event may have on that statistic, and selects a privacy budget parameter based on the required privacy strength. After obtaining the original statistics within a window, the system generates random noise based on the sensitivity range and privacy budget parameter, and superimposes the noise onto the original statistics to form the protected statistics. Necessary truncation and range constraints are then applied to ensure the output still falls within a physically reasonable range. To avoid uncontrollable impacts of noise on model training, the system can employ a consistent noise mechanism and simulate similar noise during the training phase, enabling the lifetime prediction model to learn a robust mapping relationship under noise conditions. Furthermore, the system can allocate different privacy budgets to different statistics; for example, assigning stricter protection strengths to more sensitive statistics that are more likely to be back-inferred, and relatively milder protection strengths to less sensitive statistics, thereby achieving a balance between privacy protection and prediction accuracy. The system concatenates the processed content morphology statistics, presented as discrete tags or protected values, with time-dimension, spatial-dimension, and operational-dimension load features to form a fixed-length load feature vector. To ensure that features of different dimensions can be input into the model, the system can normalize continuous features and perform one-hot or ordered encoding on discrete tags. If insufficient sampling within a window leads to missing content morphology statistics, the system can fill in missing markers and simultaneously generate missing mask features, enabling the model to identify the source of uncertainty for the feature, rather than mistaking missing features for low risk. In another embodiment, the system uses a weighted combination to construct the feature vector, with weights determined based on the reliability or confidence level of the features. For statistics injected with strong noise, the system can reduce their weights and increase the weights of associated stable features. For example, when bit entropy is strongly protected, the system can increase the contribution of repeated block proportion bucket tags and spatial hotspot features to maintain the ability to identify lifetime-sensitive patterns such as hotspot repeated overwrite. The weights can also be dynamically adjusted according to the external environment. For example, when the intensity of external interference is high or the queue is significantly congested, the explanatory weight for the long tail of delay can be reduced and the explanatory weight for the content form and spatial distribution can be increased to reduce the interference of congestion noise on the lifetime model.

[0043] In some examples, it also includes:

[0044] Within a preset sliding time window, the read access intensity and write access intensity are statistically analyzed, and the access behavior is classified into phase types based on the relative magnitude or trend of the read access intensity and the write access intensity. The phase types include read-dominant phase, write-dominant phase, and mixed phase.

[0045] A phase sequence composed of the phase types is generated in chronological order, and the phase duration distribution, phase switching frequency, and phase duty cycle are extracted as phase features.

[0046] The phase features, content morphology statistics, time-dimensional load features, spatial-dimensional load features, and operational-dimensional load features are used together to construct or update the load feature vector.

[0047] For example, considering that read and write accesses typically have different mechanisms and triggering paths, writes more directly correspond to programming stress and write amplification risks, while reads may correspond to read interference, increased error correction overhead, and exposure to related degradation. When read and write intensity exhibits phase dominance and frequent switching over time, simply using the average read / write ratio or single-window statistics can easily mask the additional damage contribution caused by phase switching. Therefore, it is necessary to classify access behavior into phases and model the phase sequence. In specific implementation, the system statistically analyzes read and write access intensity within a preset sliding time window. The intensity can be defined by the number of accesses, the number of bytes accessed, the access bandwidth, the time occupied by requesting services, or a combination thereof. The intensity can be weighted by writing granularity, concurrency, and queue depth to make it closer to the actual operational pressure borne by the granularity. Then, based on the relative magnitude or trend of read and write access intensity, the access behavior is classified into phase types, where write access... When the intensity of a read access is consistently higher than the read access intensity within the window and exceeds a preset percentage threshold, it is determined to be a write-dominant phase. When the intensity of a read access is consistently higher than the write access intensity within the window and exceeds a preset percentage threshold, it is determined to be a read-dominant phase. When the difference between the two is insufficient to meet the dominance determination condition or both are at a significant level, it is determined to be a mixed phase. In order to avoid frequent misjudgments of phase due to short-term jitter, a minimum duration window constraint or smoothing of the intensity sequence can be introduced for phase determination. After obtaining the phase type, the system generates a phase sequence composed of phase types in chronological order, and extracts the phase duration distribution, phase switching frequency, and phase duty cycle as phase features from the phase sequence. The phase duration distribution is used to reflect the stability of the write-dominant or read-dominant state, the phase switching frequency is used to reflect the density of the read-write dominance state transition, and the phase duty cycle is used to reflect the time proportion or write volume proportion of each phase type in a longer observation interval. For example, in scenarios where online query services dominated by random reads during the day and data archiving services dominated by high-throughput writes at night are mixed on the same granularity, the phase sequence will show long-term alternation between read-dominant and write-dominant phases. For real-time streaming services, the phase sequence may show characteristics of mixed phases with high duty cycles and high switching frequencies. Different phase structures often correspond to different stress superposition methods and lifetime consumption trajectories. Finally, the system combines the phase features with content form statistics, time-dimensional load features, spatial-dimensional load features, and operational-dimensional load features to form or update the load feature vector. This allows the lifetime prediction model to not only use content form and multi-dimensional load statistics to characterize the stress intensity and distribution within a single window, but also to use phase sequence features to characterize the read-write dominant structure and switching behavior across windows. This improves the sensitivity and robustness of lifetime consumption estimation under complex loads with significant read-write alternation, frequent phase switching, or multiple service superpositions, and reduces lifetime prediction bias caused by relying solely on the average read-write ratio.

[0048] In some examples, it also includes:

[0049] In the phase sequence, determine the phase boundary time points between adjacent phases, and set boundary observation windows before and after each phase boundary;

[0050] Within the boundary observation window, a boundary stress index is calculated, which includes at least one of the peak access intensity, the rate of increase of access intensity, and the access concurrency.

[0051] Within the boundary observation window, the content morphology statistics are simultaneously statistically analyzed to obtain boundary content morphology change indicators, which include bit flip density change and / or bit entropy change.

[0052] The boundary stress index and the boundary content morphology change index are used as components of the load feature vector.

[0053] For example, considering that the lifetime loss of many loads is not solely determined by the average intensity within each phase, phase switching is often accompanied by a surge in request arrival rate, increased concurrency, changes in write granularity or alignment, and abrupt changes in write data format. These changes can trigger higher electrical and thermal stress superposition in a short period of time and may trigger more complex internal execution paths, making it difficult to characterize the amplification effect of this switching transient on lifetime using only phase duration or phase duty cycle. Therefore, it is necessary to construct a dedicated observation window near the phase boundary and form boundary features. In specific implementation, the system locates the boundary time point between adjacent phases in the generated phase sequence. The boundary time point can be defined as the start time of the window where the phase type changes or the estimated time when the change occurs. Boundary observation windows are set before and after each phase boundary. The boundary observation windows can be determined by a fixed duration, a fixed number of visits, or a combination of both. At the same time, a minimum observation sample number constraint can be introduced to avoid statistical instability under low load conditions. Within the boundary observation window, the system calculates the boundary stress index, which includes at least the number of visits. The system queries at least one of the following: peak intensity, access intensity rise rate, and access concurrency. Peak access intensity can be defined as the maximum number of accesses or bytes accessed per unit time. The access intensity rise rate can be defined as the ratio of the difference in intensity between the window before and after the boundary to the time interval. Access concurrency can be defined as the number of concurrent incomplete requests, queue depth, or the number of simultaneously active threads. Furthermore, the concurrency can be weighted according to the difference in the impact of write and read operations on lifetime, making it more closely reflect the actual stress level of the granular components. Simultaneously, to incorporate transient content factors into the characterization, the system synchronously performs statistical analysis on content morphology within the boundary observation window to obtain a boundary content morphology change index. This index can include at least one of bit-flip density change and bit entropy change. The bit-flip density change can be obtained by the absolute value or weighted difference of the bit-flip density difference between the windows before and after the boundary, while the bit entropy change can be obtained by the difference in bit entropy between the windows before and after the boundary or its quantile difference, reflecting the sudden increase in programming stress when switching from structured content to near-random content or from low-flip mode to high-flip mode.For example, when the online query load, which is mainly read-based during the day, switches to the backup load, which is concentrated on writing at night, peak access intensity and increased concurrency often appear within the boundary observation window. At the same time, the bit entropy of the written data increases significantly and the change in bit flip density increases. Boundary features can reveal the high-risk transient wear corresponding to this switching moment. In contrast, during small-scale switching between read-dominant and mixed phases, the rate of increase in access intensity may be lower and the change in content form may be smaller, so the boundary features will show a lower risk. Finally, the system incorporates the boundary stress index and the boundary content form change index as components of the load feature vector, enabling the lifetime prediction model to explicitly learn the additional contribution of phase switching transients to lifetime consumption. This improves the sensitivity and interpretability of lifetime consumption measurement in scenarios with frequent read-write alternation, switching accompanied by sudden concurrency and drastic changes in content form, and reduces the problem of underestimation of lifetime assessment caused by relying solely on the average load within a phase.

[0054] In some examples, it also includes:

[0055] Within the boundary observation window, operational indicators reflecting the reliability status of the storage medium are collected, and the operational indicators include at least one of error correction strength, retry count, and long-tail index of access latency distribution;

[0056] The operation indication quantity is correlated with the boundary stress index to form a phase switching trigger signature, which is used to characterize the strength of the phase switching trigger for a specific failure mechanism.

[0057] Based on the phase switching trigger signature, the lifetime consumption estimation coefficients, feature weights, or output correction terms of the lifetime prediction model are adaptively adjusted.

[0058] For example, considering that although boundary stress indicators and boundary content morphology change indicators can characterize the load intensity and content mutation at the moment of phase switching, the same boundary stress does not necessarily lead to the same reliability degradation performance under different media, different aging stages, and different controller strategies. Therefore, it is necessary to introduce an operational indicator that can directly reflect the reliability status of the medium as bypass evidence, observe the coupling relationship between stress input and reliability response within the boundary observation window, and solidify this coupling relationship into a trigger signature for dynamic correction of lifetime prediction. In specific implementation, the system collects operational indicators within the boundary observation window. The operational indicators include at least one of error correction strength, retry count, and long-tail index of access delay distribution. The error correction strength can be defined as the mean, quantile, or histogram distribution characteristics of the number of error correction bits per unit time or unit access, or it can be defined as the frequency of error correction events. The retry count can be defined as read / write operations. The number of re-executions or the retry ratio due to verification failures or timing anomalies, and the long-tail index of access latency distribution can be defined as the deviation of high quantile latency from median latency, the occurrence rate of over-threshold latency events, or long-tail fitting parameters, used to reflect the impact of queue congestion and internal media recovery operations on external latency under boundary transients. Subsequently, the system correlates the operation indication quantity with the boundary stress index to form a phase switching trigger signature. The correlation calculation can be performed on the basis of time alignment, for example, calculating the correlation coefficient, mutual information, conditional probability gain, or quantile difference between the operation indication quantity and the peak access intensity, access intensity rise rate, and access concurrency within the boundary observation window. Furthermore, the content form change index can be introduced as a condition variable to distinguish between degradation response triggered by concurrent bursts and degradation response triggered by content form mutations, thereby obtaining a trigger signature used to characterize the strength of phase switching triggering a specific failure mechanism.For example, if the access intensity rate increases significantly within the boundary window of the switch from a read-dominant phase to a write-dominant phase, and the high quantile of the error correction intensity rises significantly while the retry ratio increases, accompanied by a significant increase in the latency tail, then the trigger signature can indicate that the switch has a strong contribution to the exposure of accumulated read disturbances, write programming stress, or triggering of internal recovery operations. Conversely, if the boundary stress index is high but the error correction intensity and the number of retryes do not increase synchronously, then the trigger signature can indicate that the switch has a weak triggering effect on failure mechanisms, avoiding excessive overestimation of lifetime consumption due to load surges alone. At the model adjustment level, the system adaptively adjusts the lifetime consumption estimation coefficients, feature weights, or output correction terms of the lifetime prediction model based on the phase switch trigger signature. In one implementation, when the trigger signature exceeds a preset trigger threshold, the system increases the correlation with boundary stress, content morphology changes, and phase. The system switches frequency-related feature weights and adds a penalty term that is monotonically correlated with the trigger signature amplitude to the lifetime consumption estimation coefficient. This allows the same load intensity to correspond to a higher lifetime consumption estimate under boundary conditions with significant reliability response. In another implementation, the system uses the trigger signature as a correction input to the model output, applying piecewise or quantile correction to the original lifetime prediction result. This enables the model to automatically enhance or weaken its sensitivity to boundary transients as the medium aging stage changes. Through the above-mentioned trigger signature construction and adaptive adjustment, the stress input of phase switching transients is correlated with the reliability response, forming an interpretable and quantifiable trigger strength metric. This effectively suppresses false alarms and false negatives under different media, different aging stages, and different control strategies, improving the robustness and consistency of lifetime consumption measurement and remaining lifetime prediction for phase-switching workloads.

[0059] In some cases, error correction coding in many storage systems can hide early aging behaviors. When memory chips experience initial degradation, it often manifests as an increase in the number of correctable errors, an increase in the number of correction bits, and a more frequent error correction event. However, read and write operations remain successful at the upper layers, and there is no direct correlation between the upper-layer access sequence and failure events. If lifetime prediction only uses load statistics such as write volume, access intensity, or address hotspots, it may misjudge the correctable but degraded stage as healthy, leading to overly optimistic lifetime prediction results. Therefore, it is necessary to transform the degradation signals reflected by the error correction process into quantifiable load-related features, incorporating the gradual changes in reliability status into the load feature vector, making the model more sensitive to early aging. Based on this, some examples also include collecting error correction-related operational information and constructing error correction strength features to eliminate the masking of aging behaviors by the error correction mechanism. The collection of error correction-related operational information and the construction of error correction strength features at least include:

[0060] Within a preset time window or preset write volume window, collect at least one of the following: the distribution of error correction bit widths corresponding to the target memory chip, the frequency of error correction events, or the comprehensive verification results after error correction, and generate an error correction intensity sequence.

[0061] By performing joint statistics on the error correction intensity sequence, the spatial dimension load feature, and the temporal dimension load feature, an error correction aging feature is obtained to characterize the cumulative trend of correctable errors.

[0062] The error correction aging features can be used as a component of the load feature vector or as an auxiliary input to the lifetime prediction model.

[0063] Understandably, the system collects error correction-related operational information of the target memory chip within a preset time window or preset write volume window. This information includes at least one or more of the following indicators: error correction bit depth, error correction event frequency, post-correction verification results, and count of uncorrectable events, forming an error correction intensity sequence. Subsequently, the system performs structured statistics on the error correction intensity sequence to form error correction aging features that can be input into the model: The error correction bit depth is aggregated by address range, page, block, or subarray granularity to obtain the error correction hotspot distribution, used to identify situations where local areas degrade faster. The error correction intensity growth rate, growth acceleration, or high quantile value rise is calculated between adjacent windows to characterize the degradation trend rather than just the current level. Error correction intensity is jointly statistically analyzed with spatial load characteristics, such as whether the write density and error correction bit depth increase synchronously within the same address range, to identify local aging caused by write hotspots. Finally, the system incorporates the error correction aging features into the load feature vector or as auxiliary input to the lifetime prediction model, enabling the lifetime model to distinguish between different health states—increasing error correction intensity and stable error correction intensity—under the same write load.

[0064] In some cases, the write volume observed on the host side may not equal the actual write activity experienced by the media. The controller may generate internal writes due to garbage collection, wear leveling, address remapping, metadata updates, log structure maintenance, etc., causing write amplification. Sometimes host writes are not large, but background maintenance writes are heavy, and lifetime consumption is severely underestimated. Therefore, it is necessary to infer internal write intensity using observable proxy variables to convert invisible internal writes into measurable metrics and correct biases in lifetime consumption estimation. Based on this, some examples also include estimating internal write intensity based on proxy variables of invisible internal writes to correct the underestimation of lifetime consumption caused by write amplification or background maintenance. The estimated internal write intensity includes at least:

[0065] Collect at least one of the following as proxy variables: write access latency distribution, background write bandwidth usage, write energy consumption, or media maintenance operation count corresponding to the target memory chip;

[0066] Based on the degree of deviation between the proxy variable and the number of bytes or writes written on the host side, calculate the internal write strength index or write amplification indicator;

[0067] The internal write strength index or the write amplification indicator can be used as a component of the load feature vector, or used to adjust the lifetime consumption estimation coefficient of the lifetime prediction model.

[0068] Understandably, the system collects proxy variables that are strongly correlated with internal writes but do not depend on plaintext data. These variables include at least one of the following: write latency distribution, background write bandwidth usage, write energy consumption, current or temperature rise changes, maintenance operation counts, and write queue congestion levels. Subsequently, the system constructs an internal write intensity metric to calculate whether the long tail of write latency significantly increases under the same host write count or number of writes. A significant increase in the long tail can be considered an indication of increased internal relocation or maintenance writes. The proportion of write bandwidth usage during background periods is statistically analyzed; if continuous background write usage occurs even when host writes are low, the internal write intensity is higher. If maintenance operation counts are available, a write amplification indicator is constructed using the ratio of maintenance operation frequency to host writes. Finally, the internal write intensity metric is used as a component of the load feature vector or to adjust the lifetime consumption estimation coefficients of the lifetime prediction model, so that the same host writes correspond to higher lifetime consumption when internal writes are heavier.

[0069] In some cases, the relationship between logical addresses and physical locations changes when the controller performs address remapping, bad block replacement, or wear leveling migration. If load assessment is based solely on hotspot statistics of logical addresses, it may miss the actual worn physical areas, leading to distortion of spatial dimension features. Therefore, it is necessary to establish an observation mechanism sensitive to mapping drift, estimate the degree of logical-to-physical drift, and use this drift information to correct spatial features, making hotspot identification more robust to remapping. Based on this, some examples also include anchor point detection and spatial correction to offset spatial hotspot drift caused by address remapping and bad block replacement. The anchor point detection and spatial correction include at least:

[0070] Select a preset number of anchor address sets, and periodically insert probe accesses into the anchor address sets to obtain anchor probe responses. The anchor probe responses include at least one of access latency, error correction strength, or number of retries.

[0071] The drift degree from logical address to physical location is estimated based on the changes in anchor point detection response between adjacent cycles, and drift correction parameters are generated.

[0072] The spatial dimension load features are corrected or reprojected based on the drift correction parameters, and the corrected spatial dimension load features are used to construct or update the load feature vector.

[0073] Understandably, the system selects a set of anchor addresses. This set can be sparse address points distributed across various logical regions, or it can be a set of pages or blocks selected through hierarchical sampling. The system periodically inserts probe accesses into the anchor address set. These probe accesses may include low-frequency reads, low-frequency writes, or read-only checksum accesses, and collects anchor probe responses. These responses include at least one of the following: access latency, error correction strength, and number of retries. Subsequently, the system statistically analyzes the changes in anchor probe responses between adjacent periods to estimate the degree of logical-to-physical drift. If a certain anchor address experiences a sudden change in latency distribution, error correction strength, or retry behavior within a short period, while the service load does not change accordingly, a mapping adjustment or a change in the physical location corresponding to that logical region may occur. This change is summarized into drift correction parameters, such as assigning drift confidence or drift strength scores to logical regions. Finally, the system uses drift correction parameters to correct or reproject the spatial dimension load features. For example, it reduces the direct interpretation weight of high drift regions when performing hotspot statistics, or maps hotspot statistics from the logical space to a stable representation of the potential physical space, and then inputs the corrected spatial features into the lifetime prediction model.

[0074] In some cases, memory chip lifespan is affected not only by electrical stress but also by temperature and temperature gradients. Access activities generate temperature rises, which diffuse spatially. High access activity in neighboring regions can raise the temperature of the target region and accelerate degradation. Characterizing thermal stress solely by the access intensity of the local region fails to reflect the contribution of neighboring heat sources, leading to lifespan prediction biases in layouts with significant thermal coupling. Therefore, it is necessary to construct an equivalent thermal stress feature that senses thermal diffusion, converting the intensity of neighboring accesses into an equivalent thermal load on the target region using a thermal diffusion kernel function. Based on this, some examples also include a thermally coupled-sensitized equivalent thermal stress feature construction to characterize the impact of temperature rise diffusion caused by neighboring region accesses on lifespan. This equivalent thermal stress feature construction includes at least:

[0075] The target memory particle is divided into multiple spatial units, and the access intensity time series of each spatial unit is statistically analyzed.

[0076] Based on the preset thermal diffusion kernel function, the access intensity time series of each spatial unit is calculated by proximity weighting to obtain the equivalent thermal stress time series of each spatial unit.

[0077] The statistics of the equivalent thermal stress time series are used as a component of the load feature vector.

[0078] Understandably, the system divides the target memory granules into multiple spatial units, which can be defined by subarrays, blocks, page groups, or other observable granularities. The system then calculates the access intensity time series for each spatial unit. Access intensity can be defined by the number of bytes written per unit time, the number of accesses, concurrency, or service time consumption. Subsequently, the system performs a neighbor-weighted calculation of the access intensity based on a preset thermal diffusion kernel function to obtain the equivalent thermal stress time series for each spatial unit. The thermal diffusion kernel function can decay according to spatial distance and introduce time lag, allowing high access from neighboring units within a shorter timeframe to contribute more to the target unit. For each target unit, the system weights and sums its own access intensity with the access intensity of neighboring units under the kernel function to obtain the equivalent thermal stress. Finally, the system extracts statistics from the equivalent thermal stress time series, such as mean, peak value, duration of sustained high thermal stress, and thermal stress burst rate, and inputs these as components of the load feature vector into the lifetime prediction model.

[0079] In some cases, considering that many workloads exhibit distinct read-write phases, the write-dominated phase directly applies programming stress, while the read-dominated phase may expose read disturbances, maintain degradation, or increase error correction overhead. The switching between read-write dominance states alters the type and superposition of stresses, and the average read-write ratio often masks this structural difference. Therefore, it is necessary to segment access behavior into phases based on read-write intensity, forming a phase sequence, and extracting features reflecting the phase structure so that lifetime prediction models can use phase structure information to explain the lifetime trajectories corresponding to different stress combinations. Based on this, some examples also include read-write phase identification and phase sequence feature extraction to characterize the impact of read-write dominance state switching on lifetime. The read-write phase identification and phase sequence feature extraction at least include:

[0080] Within a preset sliding time window, the read access intensity and write access intensity are statistically analyzed, and based on the relative magnitude or trend of the read access intensity and the write access intensity, the access behavior is divided into read-dominant phase, write-dominant phase, and mixed phase.

[0081] A phase sequence composed of the phase types is generated in chronological order, and the phase duration distribution, phase switching frequency, and phase duty cycle are extracted as phase features.

[0082] The phase feature is used as a component of the load feature vector.

[0083] Understandably, the system separately calculates read and write access intensity within a preset sliding time window. Intensity can be defined by access frequency, byte count, bandwidth, and service time usage, and can be weighted based on concurrency. The system classifies phase types based on the relative magnitude and trend of read and write intensity: a write-dominant phase is defined when write access intensity meets the dominance condition within the window; a read-dominant phase is defined when read access intensity meets the dominance condition; and a mixed phase is defined when both are high or the dominance relationship is unclear. To suppress noise jitter, a minimum phase duration constraint can be set to avoid frequent misjudgments of phase due to short-term fluctuations. Subsequently, the system generates a phase sequence in chronological order and extracts phase features such as phase duration distribution, phase switching frequency, and phase duty cycle, finally incorporating these phase features into the load feature vector.

[0084] In some cases, it's important to consider that idle periods do not equate to no lifetime loss. Certain media may experience hold degradation and accumulated threshold drift during idle periods, leading to increased readout error correction or retries after the idle period ends. If the model relies solely on access load, long idle periods may be misclassified as zero-stress phases, thus underestimating the contribution of hold-like degradation to lifetime. Therefore, it's necessary to identify idle periods and observe the reliability response after the idle period ends, incorporating idle-related degradation into the load feature vector or for output correction. Based on this, some examples also include idle period stress characterization to incorporate the risk of hold degradation during non-access periods into lifetime assessments. This idle period stress characterization includes at least:

[0085] Identify idle time periods written to the access stream and statistically analyze the distribution of idle segment lengths;

[0086] Within a preset observation window after the idle period ends, collect at least one of the following indicators: error correction intensity, number of retry attempts, or long-tail index of access delay distribution, to form an idle post-degradation response index.

[0087] The idle segment length distribution and the idle post-degradation response index can be used as components of the load feature vector, or used to adjust the output correction term of the lifetime prediction model.

[0088] Understandably, the system identifies idle periods in the write access stream. An idle period can be defined as a continuous duration without any write or read / write requests, and the duration threshold can be configured by the system. The system statistically analyzes the distribution of idle period lengths, the frequency of idle periods, and the distribution of idle periods at different temperature levels. Subsequently, after the idle period ends, the system collects operational indicators such as error correction intensity, retry count, and long-tail access latency metrics within a preset observation window to form post-idle degradation response metrics, such as the increase in error correction bits, the increase in retry ratio, and the increase in long-tail latency after the idle period ends. Finally, the system incorporates the idle period length distribution and post-idle degradation response metrics into the load feature vector, or applies a correction term to the lifetime prediction output when the idle period length exceeds a threshold and the degradation response is significant, so that the lifetime consumption estimate can reflect the risk of degradation.

[0089] In some cases, considering that in multi-tenant or multi-process environments, a single service can only see its own access, but the granular load represents the actual load of all services combined. Queue congestion, background maintenance writes, and temperature rise caused by other services can alter reliability status and lifetime consumption. If the model only uses the load of this single service, a systematic error in lifetime consumption will occur. Therefore, it is necessary to construct an external interference intensity fingerprint using runtime characteristics that do not depend on global visibility, incorporating the unobservable cumulative load impact into the model. Based on this, some examples also include the construction of an external interference intensity fingerprint to estimate the unobservable cumulative load impact in multi-tenant or multi-process scenarios. The construction of the external interference intensity fingerprint includes at least:

[0090] Collect at least one of the following characteristics corresponding to the target memory particle: queue waiting time distribution, service time jitter, or access latency autocorrelation feature, as an interference observation;

[0091] The external interference intensity index is calculated based on the interference observations, and the external interference intensity index is used as a latent variable to input the lifetime prediction model, or to correct the confidence level of the load feature vector.

[0092] When the external interference intensity index exceeds a preset threshold, the feature weights or lifetime consumption estimation coefficients of the lifetime prediction model are adjusted.

[0093] Understandably, the system collects interference observations related to interference intensity, including at least queue wait time distribution, service time jitter, and access latency autocorrelation characteristics. Queue wait time distribution reflects the congestion level of the equipment during that period, often caused by superimposed loads. Service time jitter reflects the instability of internal operation paths, potentially related to background maintenance or resource contention. Latency autocorrelation characteristics reflect the presence of periodic or batch processing interference. The system calculates external interference intensity indicators based on these observations. For example, it combines the high quantile rise in wait time, jitter intensity, and latency autocorrelation peak into an interference intensity score, using this score as a latent variable input to the lifetime prediction model or to correct the confidence level of the load feature vector. When the interference intensity exceeds a threshold, the system can increase the model's weighting of reliability indicators and boundary transient characteristics, reducing its reliance on single-service load statistics.

[0094] In some cases, considering that concurrent bursts and sharp increases in write intensity may trigger power integrity fluctuations, such as voltage drops, current surges, and sudden changes in local temperature rise, this can lead to deterioration of write boundary conditions, increased retries, and a heavier delay tail, which will accelerate device aging in the long run. This type of electrical stress-related degradation is not entirely determined by the average write amount, but is closely related to the burst rise rate, concurrency, and the length of the sustained high-intensity segment. Therefore, it is necessary to construct electrical stress candidate features and electrical stress response features, and combine them to form an electrical stress trigger index to characterize the contribution of power supply transients to lifetime loss. Based on this, some examples also include the construction of power supply transient-related electrical stress features to characterize the impact of power integrity fluctuations caused by concurrent bursts on lifetime. The electrical stress feature construction includes at least the following:

[0095] Within a preset time window, statistically analyze at least one of the following: the sudden rise rate of writing, the degree of writing concurrency, or the duration of continuous high-intensity segments, to form candidate features of electrical stress.

[0096] Within the preset time window, at least one of the following is collected: long-tail index of access delay distribution, number of retryes, or error correction intensity, to form electrical stress response characteristics;

[0097] An electric stress triggering index is constructed based on the electric stress candidate features and the electric stress response features, and the electric stress triggering index is used as a component of the load feature vector or to adjust the output correction term of the lifetime prediction model.

[0098] Understandably, the system statistically analyzes candidate electrical stress features within a preset time window, including at least the write burst rise rate, write concurrency, and duration of continuous high-intensity periods. Subsequently, within the same time window, it collects electrical stress response features, including at least the long-tail index of access latency distribution, the number of retries, and error correction intensity. Based on these candidate and response features, the system constructs an electrical stress trigger index. For example, when the burst rise rate and concurrency increase simultaneously with an increased retry ratio or a significantly heavier long-tail latency, the electrical stress trigger index takes a higher value; when the burst is strong but the response is not obvious, the trigger index is lower, to avoid over-judging high risk solely based on load bursts. Finally, the electrical stress trigger index is incorporated into the load feature vector or used to apply a correction term to the lifetime prediction output, making the model more accurate under electrical stress-sensitive loads.

[0099] Please see Figure 2 One embodiment of the multi-dimensional access load assessment system for predicting memory particle lifetime in this application includes:

[0100] The acquisition unit 21 is used to acquire the write access stream corresponding to the target memory particle, monitor the write access stream online, and perform irreversible content digest calculation on the write data without outputting the plaintext content of the written data to obtain content form statistics. The write access stream includes write address information and write data. The content form statistics include at least one of the following to characterize the bit flip density, bit entropy, and repeating block ratio of the written data.

[0101] The fusion unit 22 is used to extract spatial dimension load features based on the write address information, extract time dimension load features based on the time series of the write access stream, extract operation dimension load features based on the write operation type or write granularity, and fuse the time dimension load features, spatial dimension load features, operation dimension load features and the content form statistics to form a load feature vector.

[0102] The prediction unit 23 is used to input the load feature vector into the lifetime prediction model and output the lifetime consumption metric or remaining lifetime prediction result of the target memory particle.

[0103] like Figure 3 As shown, this application embodiment also provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored on the memory 310 and executable on the processor. When the processor 320 executes the computer program 311, it implements the steps of any of the above-described methods for multidimensional access load assessment for memory particle lifetime prediction.

[0104] Since the electronic device described in this embodiment is the device used to implement a multi-dimensional access load assessment system for predicting memory chip lifetime in the embodiments of this application, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the method described in the embodiments of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. Any device used by those skilled in the art to implement the method in the embodiments of this application is within the scope of protection of this application.

[0105] In practical implementation, when the computer program 311 is executed by the processor, it can achieve the following: Figure 1 Any of the corresponding implementation methods in the embodiments.

[0106] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

Claims

1. A method for multi-dimensional access load evaluation oriented to memory granule lifetime prediction, characterized in that, include: Obtain the write access stream corresponding to the target memory particle, monitor the write access stream online, and perform irreversible content digest calculation on the write data without outputting the plaintext content of the written data to obtain content morphology statistics. The write access stream includes write address information and write data. The content morphology statistics include at least one of the following to characterize the bit flip density, bit entropy, and repetitive block ratio of the written data. Based on the write address information, spatial load features are extracted; based on the time series of the write access flow, temporal load features are extracted; and based on the write operation type or write granularity, operational load features are extracted. These temporal, spatial, and operational load features are then fused with the content morphology statistics to form a load feature vector. The spatial load features characterize the distribution of wear within particles, including address hotspot index, hotspot persistence, and local rewrite ratio. The temporal load features characterize whether stress is superimposed in a sudden, sustained high-intensity, or periodic manner, including write burstiness, sustained high-intensity segment length distribution, and write rhythm stability. The operational load features characterize the differences in internal execution paths and potential write amplification risks, including write granularity distribution, alignment deviation features, and overwrite ratio. The load feature vector is input into the lifetime prediction model, and the lifetime consumption metric or remaining lifetime prediction result of the target memory particle is output.

2. The method of claim 1, wherein, The irreversible content digest calculation performed on the written data includes: The written data is divided into blocks according to a preset block size; Perform a random projection operation or a hash mapping operation on each data block to obtain a digest result, and calculate at least one of the bit flip density, the bit entropy and the proportion of repeated blocks based on the digest result, wherein the digest result irreversibly represents the statistical form of the written data and is insufficient to recover the plaintext content of the written data.

3. The method as described in claim 1, characterized in that, The content format statistics also include a content stability index, which characterizes the stability of the data state distribution within a continuous write window. The method further includes: The change in bit flip density or bit entropy within adjacent time windows is statistically analyzed to obtain the intensity of content change. The content change intensity is normalized with the number of writes or the number of bytes written to obtain the content change rate per unit of write volume, which is used as the content stability index. The input feature weights or lifespan consumption estimation coefficients of the lifespan prediction model are adaptively adjusted based on the content stability index.

4. The method as described in claim 1, characterized in that, Also includes: The content format statistics are subjected to interval binning or differential privacy noise injection processing to limit the identifiable contribution of a single data write to the statistical results. The processed content morphology statistics are concatenated or weighted with the time-dimensional load features, spatial-dimensional load features, and operational-dimensional load features to form the load feature vector used as input to the lifetime prediction model.

5. The method according to any one of claims 1 to 4, characterized in that, Also includes: Within a preset sliding time window, the read access intensity and write access intensity are statistically analyzed, and the access behavior is classified into phase types based on the relative magnitude or trend of the read access intensity and the write access intensity. The phase types include read-dominant phase, write-dominant phase, and mixed phase. A phase sequence composed of the phase types is generated in chronological order, and the phase duration distribution, phase switching frequency, and phase duty cycle are extracted as phase features. The phase features, content morphology statistics, time-dimensional load features, spatial-dimensional load features, and operational-dimensional load features are used together to construct or update the load feature vector.

6. The method as described in claim 5, characterized in that, Also includes: In the phase sequence, determine the phase boundary time points between adjacent phases, and set boundary observation windows before and after each phase boundary; Within the boundary observation window, a boundary stress index is calculated, which includes at least one of the peak access intensity, the rate of increase of access intensity, and the access concurrency. Within the boundary observation window, the content morphology statistics are simultaneously statistically analyzed to obtain boundary content morphology change indicators, which include bit flip density change and / or bit entropy change. The boundary stress index and the boundary content morphology change index are used as components of the load feature vector.

7. The method as described in claim 6, characterized in that, Also includes: Within the boundary observation window, operational indicators reflecting the reliability status of the storage medium are collected, and the operational indicators include at least one of error correction strength, retry count, and long-tail index of access latency distribution; The operation indication quantity is correlated with the boundary stress index to form a phase switching trigger signature, which is used to characterize the strength of the phase switching trigger for a specific failure mechanism. Based on the phase switching trigger signature, the lifetime consumption estimation coefficients, feature weights, or output correction terms of the lifetime prediction model are adaptively adjusted.

8. A multi-dimensional access load assessment system for predicting memory chip lifetime, characterized in that, include: The acquisition unit is used to acquire the write access stream corresponding to the target memory particle, monitor the write access stream online, and perform irreversible content digest calculation on the write data without outputting the plaintext content of the written data to obtain content morphology statistics. The write access stream includes write address information and write data. The content morphology statistics include at least one of the following to characterize the bit flip density, bit entropy, and repetitive block ratio of the written data. The fusion unit is used to extract spatial load features based on the write address information, extract time load features based on the time series of the write access stream, extract operation load features based on the write operation type or write granularity, and fuse the time load features, spatial load features, operation load features and the content form statistics to form a load feature vector. The prediction unit is used to input the load feature vector into the lifetime prediction model and output the lifetime consumption metric or remaining lifetime prediction result of the target memory chip. The spatial dimension load feature is used to characterize the distribution pattern of wear inside the chip, including address hotspot index, hotspot persistence and local rewrite ratio. The temporal dimension load feature is used to characterize whether the stress is superimposed in a burst, sustained high intensity or periodic manner, including write burst degree, sustained high intensity segment length distribution and write rhythm stability. The operational dimension load feature is used to characterize the differences in internal execution paths and potential write amplification risk, including write granularity distribution, alignment deviation characteristics and overwrite ratio.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program stored in the memory, implements the steps of the multidimensional access load assessment method for memory particle lifetime prediction as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the multi-dimensional access load assessment method for memory particle lifetime prediction as described in any one of claims 1-7.

Images