Machine learning based inorganic media archival assessment method

Abstract

The application discloses a machine learning-based inorganic medium archiving evaluation method, and belongs to the field of long-term archiving storage of inorganic media and machine learning risk evaluation. In order to solve the problem that the environment state, reading frequency and medium aging performance of the inorganic medium storage unit are difficult to be uniformly evaluated, the application collects environment, reading and aging detection data, extracts multi-modal vectors and uses the multi-modal vectors as at least part of the node features of a dynamic medium relationship graph, establishes the dynamic medium relationship graph, combines a medium aging prototype memory bank, a graph Transformer and a conformal calibration layer to output a unified archiving health score, a reading risk score, a page missing probability interval and a confidence identifier, and forms an archiving disposal suggestion according to the reading risk level, the reading risk score, the page missing probability interval and the confidence identifier, so that the technical effect of identifying potential reading risks in advance when the file state is still normal is achieved.

Description

Technical Field

[0001] This invention relates to the fields of long-term archiving storage of inorganic media and risk assessment using machine learning, and particularly to a machine learning-based method for archiving assessment of inorganic media. Background Technology

[0002] As the retention periods for cold data, archival data, and strategic backup data continue to lengthen, inorganic media storage units are being used to hold critical data that needs to remain recoverable for years or even longer. Existing archiving management systems typically rely on file directories, checksums, write times, and manual inspection records as management criteria. These systems can record whether files exist, whether checksums have been completed, and whether explicit read / write anomalies have occurred. However, they lack a unified model for factors that media are exposed to during long-term storage, such as temperature and humidity, light, particulate matter, vibration, cabinet opening and closing, differences in reading devices, and changes in reading intervals.

[0003] In real-world long-term storage scenarios, media from the same batch may be distributed across different cabinets, and media of different material types and write times may experience similar environmental exposures or reading device deviations. Traditional methods often save environmental records, read logs, and media surface images or spectral detection results separately, lacking a mechanism to uniformly transform environmental exposure, read degradation, and apparent aging into comparable characteristics. This results in potential read risks only being discovered after file verification fails, the number of rereads increases abnormally, or page faults are recovered.

[0004] Furthermore, existing risk assessment methods typically rely on fixed thresholds or human experience, which are insufficient in distinguishing between different aging stages such as normal aging, moisture exposure, thermal cycling exposure, low-frequency reading instability, and equipment deviation. When outputting page loss probability or reading risk, they also lack probability ranges and confidence indicators that are calibrated by combining material type, storage period, reading frequency, and environmental fluctuation level, making it difficult to provide verifiable handling basis for rereading, spot checks, migration, or redundant verification.

[0005] Therefore, there is a need for an inorganic media archiving evaluation method that can address the shortcomings of the existing technologies. Summary of the Invention

[0006] One objective of this invention is to propose a machine learning-based method for evaluating inorganic media archiving. Addressing the challenges of unifying the assessment of environmental conditions, read frequency, and media aging behavior during long-term storage of cold data, archival data, and strategic backup data in existing technologies, and the difficulty in timely detection of potential read risks, this invention proposes a technical solution that collects environmental condition data, read behavior data, and media aging detection data from inorganic media storage units. It extracts environmental exposure vectors, read degradation vectors, and media apparent aging vectors, establishes a dynamic media relationship graph with multimodal fusion vectors as at least some node features, and combines a media aging prototype memory, a graph Transformer, and a conformal calibration layer to output archiving risks. This invention has the technical effect of identifying read risks in advance when the file status is still normal, outputting page fault probability intervals and confidence indicators, and forming archiving disposal suggestions.

[0007] This invention provides a machine learning-based method for evaluating inorganic media archiving, comprising: S1, collecting environmental state data, read behavior data, and media aging detection data of inorganic media storage units, aligning them according to media identification and sampling time to form a multimodal sample sequence; S2, inputting the multimodal sample sequence into a trained self-supervised temporal encoder, read degradation encoder, and image encoder or spectral encoder to obtain an environmental exposure vector, a read degradation vector, and a media apparent aging vector, and fusing them into a multimodal fusion vector; S3, establishing a dynamic media relationship graph according to media batch, storage cabinet location, material type, write time, window master read device, and environmental exposure similarity, with node features including the multimodal fusion vector, and using relationship type and time... Interval and exposure similarity are used as edge features to obtain graph samples; S4, the multimodal fusion vector is matched with the prototype units in the media aging prototype memory bank for similarity matching, and aging stage vectors are generated according to the matching similarity and prototype reliability weights. The aging stage vectors are then incorporated into the node features of the graph samples; S5, the incorporated graph samples are input into the trained graph Transformer, which outputs a unified archive health score, read risk level and expected page fault probability. The conformal calibration layer generates read risk score, page fault probability interval and confidence label, and generates archive disposal suggestions based on read risk level, read risk score, page fault probability interval and confidence label. The conformal calibration layer is selected according to material type, storage period, read frequency and environmental fluctuation level.

[0008] Optionally, S1 includes: Environmental condition data includes temperature, relative humidity, light intensity, particulate matter concentration, vibration intensity, and storage cabinet opening and closing events; The read behavior data includes read time, read device identifier, read duration, number of rereads, error correction count, read rate, and verification failure records; The media aging test data includes media surface images, reflectance spectra, transmission spectra, coating integrity test values, and edge defect test values; The data are merged into time windows according to media identifier, batch number and sampling time, and the uncollected items are recorded using a missing measurement mask to obtain the multimodal sample sequence.

[0009] Optionally, S2 includes: The environmental state data is divided into preset fixed-length windows in chronological order, and the environmental exposure vector is obtained by a self-supervised temporal encoder that reconstructs the occlusion segment and predicts the next window. The number of rereads, error correction counts, reading rate, and verification failure records in the reading behavior data are input into the reading degradation encoder to obtain the reading degradation vector; After the image or spectrum in the medium aging detection data is processed by convolution feature extraction and spectral peak normalization, it is input into the corresponding encoder to obtain the medium apparent aging vector. The three types of vectors are applied with normalized weights determined by the modality availability mask and then concatenated to obtain the multimodal fusion vector; Furthermore, the self-supervised temporal encoder, read degradation encoder, image encoder, spectral encoder, and graph Transformer are trained as follows: using historical inorganic media storage unit samples as training samples, the self-supervised temporal encoder is trained using temporal segment occlusion reconstruction loss and next window prediction loss; the read degradation encoder is trained using read degradation labels or degradation pseudo-labels constructed from reread counts and error correction counts; the image encoder and spectral encoder are trained using batch consistency loss of media aging detection samples; and the scoring head and probability head of the graph Transformer are trained using labeled recovery failure records, page fault records, and verification failure records.

[0010] Optionally, S3 includes: Each inorganic media storage unit is used as a node; The reading device adopts a window master reading device, which is the reading device with the longest reading duration within the current sampling time window. If the reading duration is the same, the most recent reading device is used. When two nodes have the same media batch, the same storage cabinet, the same material type, the write time difference falls within the preset time difference range, the same window master reading device, or the environmental exposure similarity is not less than the preset similarity threshold, an edge is established between the corresponding nodes. The environmental exposure similarity is calculated from the cosine similarity of the environmental exposure vectors of the two nodes; The relationship type, time interval, and exposure similarity of the edges are encoded as the edge features, and the node features and edge features are updated according to the sampling time of the newly added detection data to obtain the dynamic medium relationship graph.

[0011] Optionally, S4 includes: Set normal aging, moisture exposure, thermal cycling exposure, reading interval instability, and equipment deviation as prototype categories respectively; The prototype of read interval instability is updated by samples whose read interval is greater than the 75th percentile of the historical read interval in the same batch and whose number of rereads is greater than the 75th percentile of the historical number of rereads in the same batch; The aging stage vector is generated based on the matching similarity and prototype reliability weight. Specifically, this includes generating a matching similarity distribution by using the cosine similarity between the multimodal fusion vector and the center vector of each prototype unit, and multiplying the matching similarity distribution by the prototype reliability weight determined by the number of prototype samples and the most recent update time to obtain the aging stage vector. The aging stage vector generated based on the matching similarity and prototype reliability weights is used as the node additional feature input to the subsequent graph Transformer; Furthermore, each prototype unit in the media aging prototype memory includes a prototype category, a center vector, a covariance matrix, a number of samples, a recent update time, and a reliability weight. When the reading results after re-examination or migration confirm the sample category, the corresponding center vector and covariance matrix are updated using the multimodal fusion vector of the confirmed sample according to the preset learning rate; The reliability weight is obtained by multiplying the sample size normalization factor by the time decay factor obtained by looking up the table according to the most recent update time; When the highest similarity value between the confirmed sample and all prototype units does not reach the new prototype threshold determined by the 10th percentile of the highest similarity of historical confirmed samples, the confirmed sample is written into the candidate prototype unit. The candidate prototype unit is converted into a prototype unit after the number of samples reaches the preset number threshold.

[0012] Optionally, S5 includes: The Graph Transformer performs multi-head attention calculation and neighborhood aggregation on the dynamic medium relationship graph, including node feature projection, edge feature bias, and node aggregation, to obtain the target node representation. By representing the target node with the input score header and probability header, we can obtain the unified archive health score, read risk level and expected page fault probability, respectively. Based on material type, storage period, reading frequency and environmental fluctuation level, a target residual set is selected from the calibration residual set. The target residual set is used to perform conformal calibration on the expected page fault probability, and the reading risk score, page fault probability range and confidence flag indicating whether the target coverage is met are output. The archiving disposal recommendations are generated based on the read risk level, read risk score, page fault probability range, and confidence indicator. Furthermore, the calibration residual set is constructed as follows: Historical calibration sample groups were divided according to material type, preset storage period range, preset reading frequency range, and environmental fluctuation level. For each historical calibration sample, calculate the non-compliance score between the predicted page fault probability and the actual page fault mark; The quantile of the non-compliance score under the target coverage is taken as the calibration residual; Add and subtract the calibration residual corresponding to the sample group to which the current media storage unit belongs from the expected page fault probability, and set the interval with the lower bound less than 0 to 0 and the interval with the upper bound greater than 1 to 1 to obtain the page fault probability interval; When the number of samples in the sample group to which the current media storage unit belongs is not less than the preset minimum number of samples, it is considered that the target coverage guarantee is met and the confidence flag is a valid calibration flag; otherwise, it is considered that the target coverage guarantee is not met and the confidence flag is a calibration flag to be verified. Furthermore, archiving and disposal recommendations are generated in the following manner: When the read risk score is not less than the migration risk threshold, the upper bound of the page fault probability interval is not less than the migration probability threshold, or the confidence flag is a flag to be reviewed and calibrated, generate data migration suggestions and redundancy check suggestions. When the upper bound of the page fault probability interval is less than the migration probability threshold and not less than the repeat probability threshold, a repeat suggestion is generated. When the upper bound of the page missing probability interval is less than the repeat probability threshold and the unified archive health score is not greater than the sampling health threshold, a sampling suggestion is generated. When the upper bound of the page fault probability interval is less than the repeat probability threshold and the unified archive health score is greater than the random inspection health threshold, the recommended re-inspection cycle is determined according to the mapping table between the upper bound of the page fault probability interval and the preset re-inspection cycle, and the recommended re-inspection cycle is written into the media archive record.

[0013] The beneficial effects of this invention are: 1. By aligning environmental status data, reading behavior data, and media aging detection data according to media identification and sampling time, and extracting environmental exposure vectors, reading degradation vectors, and media apparent aging vectors using self-supervised temporal encoders, reading degradation encoders, and image encoders or spectral encoders respectively, the originally scattered environmental records, reading logs, and detection results can be converted into multimodal fusion vectors under the same media node, thereby improving the consistency and integrity of the long-term archived state characterization of inorganic media.

[0014] 2. By establishing a dynamic media relationship graph based on media batch, storage cabinet location, material type, write time, window master read device, and environmental exposure similarity, and by using multimodal fusion vectors as at least part of the node features and relationship type, time interval, and exposure similarity as edge features input to the Transformer graph, it is possible to simultaneously utilize the correlation information between media of the same batch, same cabinet location, same material, and similar exposure when evaluating a single media storage unit. This allows the unified archive health score, read risk level, and expected page fault probability to better reflect the group risk and proximity risk in long-term storage.

[0015] 3. By setting up a media aging prototype memory library after multimodal fusion and setting up a conformal calibration layer at the risk output end, the current media can be matched with prototype units such as normal aging, moisture exposure, thermal cycling exposure, read interval instability, and equipment deviation to generate aging stage vectors. At the same time, the calibration residual set is selected according to material type, storage period, read frequency, and environmental fluctuation level, and the read risk score, page missing probability interval, and confidence mark are output. Based on the read risk level, read risk score, page missing probability interval, and confidence mark, archiving disposal suggestions are generated, thus providing a verifiable disposal basis for rereading, spot checking, migration, or redundancy verification. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart shows a machine learning-based evaluation method for inorganic media archiving. Figure 2 A flowchart for establishing a dynamic media relationship diagram in step S3 of this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-2The inorganic media archiving assessment method based on machine learning includes: S1, collecting environmental state data, read behavior data, and media aging detection data of inorganic media storage units, aligning them according to media identification and sampling time to form a multimodal sample sequence; S2, inputting the multimodal sample sequence into a trained self-supervised temporal encoder, read degradation encoder, and image encoder or spectral encoder to obtain environmental exposure vector, read degradation vector, and media apparent aging vector, and fusing them into a multimodal fusion vector; S3, establishing a dynamic media relationship graph according to media batch, storage cabinet location, material type, write time, window master read device, and environmental exposure similarity, with node features including the multimodal fusion vector, and using relationship type and time interval as parameters. S4. Using exposure similarity as edge features, graph samples are obtained; S5. The multimodal fusion vector is matched with the prototype units in the media aging prototype memory bank for similarity. An aging stage vector is generated based on the matching similarity and prototype reliability weight. The aging stage vector is then incorporated into the node features of the graph samples; S6. The incorporated graph samples are input into the trained graph Transformer, which outputs a unified archive health score, read risk level, and expected page fault probability. The conformal calibration layer generates a read risk score, page fault probability interval, and confidence indicator. Based on the read risk level, read risk score, page fault probability interval, and confidence indicator, archive disposal suggestions are generated. The conformal calibration layer is selected according to material type, storage period, read frequency, and environmental fluctuation level.

[0019] In this specific embodiment, S1 includes: The data acquisition gateway of the archiving evaluation system uses inorganic media storage units as the processing object, reads the media identifier, batch number, material type, storage cabinet location, write time, and the identifier of the most recent reading device. In this specific embodiment, the sampling period is set to 24 hours and the time window length is set to 7 days. The environmental acquisition records include temperature, relative humidity, light intensity, particulate matter concentration, vibration intensity, and storage cabinet opening and closing events. The reading log records include reading time, reading device identifier, reading duration, number of rereads, error correction count, reading rate, and verification failure records. The aging test records include media surface images, reflectance spectrum, transmission spectrum, coating integrity test values, and edge defect test values. The system writes these raw records into a sampling buffer table with the media identifier and sampling time as the joint primary key, for use in the sample sequence construction before S2. The sampling buffer table stores the data source, acquisition device, physical unit, sampling time, record checksum, and cabinet position status for each original record. Environmental status data comes from warehouse sensors and cabinet door event controllers. Reading behavior data comes from archived reading device logs and verification program logs. Media aging detection data comes from surface imaging equipment, reflectance spectrometer, transmission spectrometer, coating integrity detection device, and edge defect detection device. The acquisition gateway unifies the time of different devices to the time of the archive server according to the device clock calibration table and writes the calibration offset as a field into the original record. The system performs alignment processing on the sampling buffer table under the same time reference. First, environmental status data is placed into the corresponding 7-day window according to the acquisition time. Then, read behavior data is merged into the window with the closest time and a time difference of no more than 12 hours according to the read time. Media aging test data is merged into the same window according to the test completion time. The merging rule is as follows: ,in Indicates the first The set of original records within the current window for each media storage unit. Indicates media identification. Representing records Sampling time, Indicates the time at the center of the window. In this specific implementation, the time is set to 12 hours and recorded in the acquisition parameter table. Records that do not fall into the window are entered into the candidate queue of the next window. After merging time windows, the alignment module generates environmental sub-records, reading sub-records, and aging sub-records. The environmental sub-records store the average temperature, average relative humidity, cumulative light intensity, 95th percentile of particulate matter concentration, root mean square of vibration intensity, and number of opening and closing events. The reading sub-records store the cumulative reading duration, total number of rereads, total error correction count, median reading rate, and number of verification failures. The aging sub-records store the surface image file index, reflectance spectrum sequence, transmission spectrum sequence, coating integrity detection value, and edge defect detection value. The module also writes the field names, physical units, window statistical rules, and data sources into the sample field list. The cleaning module performs linear normalization on temperature, relative humidity, light intensity, particulate matter concentration, vibration intensity, read rate, and spectral intensity according to the upper and lower limits in the pre-deployment calibration table. The normalization rule is as follows: ,in Medium In the field Window statistics on and These represent the fields in the data acquisition parameter table. The effective lower limit and effective upper limit, For dimensionless normalized results, fields that exceed the effective physical range are simultaneously written to the exception field table while retaining the original measurement value; The system uses a missing test mask to record uncollected items and retain the source of the missing test. The mask calculation rule is as follows: Medium In the field There are valid sampled values ​​above. Representation field If a test is missing in the current window or removed from the abnormal field table, and if the surface image is missing but the reflectance or transmission spectrum exists, the aging sub-record retains the spectrum field and sets the image field mask to zero. If the read log in the same window is empty, the read duration, reread count, error correction count, read rate, and verification failure record are all identified by the test missing mask instead of being replaced by a zero value. The system writes the test missing mask into the sample sequence for S2 to determine the modal availability weight. The data quality module performs deduplication on duplicate records within the same media identifier, batch number, and sampling time window, retaining the record with the latest acquisition completion time and consistent checksum. When the time difference between the environmental sensor record and the cabinet opening / closing event exceeds the synchronization upper limit in the acquisition parameter table, the system marks the window as a window to be reviewed and stops writing to the model training set, but still saves it as a sample with missing test mask during inference. A record in the acquisition parameter table includes field name, physical unit, effective lower limit, effective upper limit, normalized denominator, synchronization upper limit, and version number. This table is constructed from the empty cabinet sensor records, historical reading logs, and aging tester calibration results for 30 consecutive days before deployment, and is updated after the testing equipment is replaced. The system ultimately encapsulates the windowed results of each media storage unit into a multimodal sample sequence. Sequence elements include media identifier, batch number, material type, storage cabinet location, write time, sampling time, environmental sub-vector, read sub-vector, aging test index, missing test mask, and abnormal field flags. The sequence is written to the multimodal sample library in ascending order of sampling time, and the most recent sample is added to the database. The sample tensor composed of windows is used as the input object of S2. When the number of valid windows is less than 16, placeholder windows with full missing test masks are added to the front end and the actual number of windows is retained.

[0020] In this specific embodiment, S2 includes: The feature encoding server reads the multimodal sample sequence written by S1 and organizes the environment sub-vectors of the most recent 16 time windows of each media storage unit into... The environmental time matrix will organize the read subvectors into... The reading behavior matrix scales the surface image to... The pixel and reflectance and transmission spectra are resampled into 512 band points. The missing measurement mask is input synchronously with the corresponding mode. All numerical fields continue to use the normalized range of the S1 acquisition parameter table to ensure that the input scale of the same medium is consistent during the training and inference stages. The input cleaning thread performs window order checks and outlier replacements on the environmental time matrix. If the same field is removed from the outlier field table in two consecutive windows, the missing test mask is retained and the normalized value of the field is filled into the field mean of the training set. During inference, the median of the same batch of media with the same material in the most recent 30 days is filled in. The replacement source, replacement value and field mask are written into the coded input record. The reading behavior matrix is ​​grouped by reading device identifier and then summarized into the media window to avoid losing device deviation information when the same medium is read by multiple reading devices on the same day. The self-supervised temporal encoder employs a two-layer bidirectional gated recurrent network and a one-layer temporal attention pooling structure. During training, 20% of continuous environment fragments are randomly occluded from historical inorganic media storage unit samples, and the environment sub-vectors for the next window are predicted. The training loss is a weighted sum of the occluded fragment reconstruction loss and the next window prediction loss, with the loss function set as follows: ,in This represents the set of time segments that are occluded. Indicates the first An environment window vector, Represents the reconstructed vector. The training parameter is set to 0.4. The training set is made from historical media samples that have completed periodic inspections and is divided into training subsets and validation subsets by batch. The encoder version is frozen when the validation loss changes no more than 0.001 for 5 consecutive rounds. During inference, the self-supervised temporal encoder inputs the environment-time matrix into the frozen version of the model. The hidden states, jointly trained by occlusion reconstruction and next-window prediction, are then processed by temporal attention pooling to output a 128-dimensional environment exposure vector. The pooling rule is as follows: ,in Medium No. The hidden state of a window. Calculated by the time attention layer and renormalized after handling zeros in the missing test window. This represents the environment exposure vector. If there are fewer than four valid environment windows, the system will... Write the low-confidence environment vector state and retain the mask for subsequent fusion weight reduction; The degenerate encoder reads reread counts, error correction counts, read rate, and verification failure records. The cumulative read duration is used as a normalized baseline field and input into a three-layer multilayer perceptron. Training labels are manually verified degenerate labels. Degenerate pseudo-labels for unlabeled samples are constructed solely from reread counts and error correction counts, following the rules... ,in This represents the number of rereads normalized to the historical maximum number of rereads for the same batch. This indicates the error correction count, which is normalized to the historical maximum error correction count for the same batch. Both values ​​are cropped to the range of 0 to 1. The larger the value, the stronger the reading degradation. The reading rate and verification failure records are used as input features of the encoder to participate in the degradation vector inference but do not enter the construction of the degradation pseudo-label. The reading degrade encoder inference outputs a 128-dimensional reading degrade vector and a degrade trust flag. The degrade trust flag is jointly determined by the number of valid windows in the reading log, reading device consistency, and pseudo-label distribution stability, according to the following rules: ,in This indicates the mean of the valid mask for the read modality. This indicates the consistency score of the read device. The degradation pseudo-label stability score is represented by three items, all of which are calculated after being normalized to the range of 0 to 1 from the historical read logs of the same batch. When the value is not less than 0.5, the degradation vector is read using the current window. When the value is less than 0.5 and there is an output from the previous confidence window, the current low-confidence output is not used; instead, the read degradation vector from the previous confidence window is adopted, and the read mode fusion weights are reduced. When the value is less than 0.5 and there is no previous trusted window output, the degenerate vector is set to the missing test mask state; the security state of the above backoff branch is recorded as follows: or Each of these is written into the security status field of the fusion vector table, so that subsequent S3 to S5 can identify that the read degraded vector comes from the previous trusted window or the missing test protection branch. The image encoder uses a convolutional residual network to extract scratches, spots, and edge defects from the medium surface image and outputs a 128-dimensional image aging vector. Image preprocessing includes dividing the gray value by 255 to obtain the range of 0 to 1, cropping the edge region according to the medium edge positioning box, and removing low contrast images using the lower limit of contrast in the image quality threshold table. The spectral encoder first performs peak normalization with the maximum peak intensity of each spectrum as the denominator and cropped to the range of 0 to 1. Then, it extracts the absorption peak shift, peak width variation, and transmission attenuation morphology through a one-dimensional convolutional network and outputs a 128-dimensional spectral aging vector. The image encoder and spectral encoder are trained using batch consistency loss of media aging detection samples, with the loss function set as follows: ,in This indicates a sample pair from the same batch that has the same aging confirmation label. This refers to sample pairs from the same batch but with different confirmation labels. Represents an image or spectral encoding vector. The training parameter was set to 1.0 in the training parameter table. The training samples came from the re-examination confirmation results, the post-transfer reading results, and the manually labeled apparent aging samples. When the same window contains both image and spectrum data, the media apparent aging vector is obtained by weighting the image aging vector and the spectral aging vector according to detection quality weights. When only one type of detection exists, the corresponding encoder output is used and the mask of the other mode is set to zero. The detection quality weights are weighted according to... Calculation, where A mask representing the availability of an image or spectral mode. Indicates the test quality score. The dimensionless zero constant 0.001 in the mass weight table is used to write the weight results into the modal weight cache. If the denominator is only composed of... If the contribution is positive, the aging vector is set to zero and a flag to be re-inspected is output. The multimodal fusion module applies normalized weights determined by a modal availability mask to the environmental exposure vector, read degradation vector, and media apparent aging vector. The fusion weights are: ,in For modal availability mask, For modal-based reliability, the three basic reliability categories—environment, read, and apparent aging—are calibrated to 0.30, 0.30, and 0.40 on the pre-deployment validation set. When set to 0.001, the system concatenates the three weighted vectors into a 384-dimensional multimodal fusion vector in the order of environment, reading, and apparent aging. After splicing, the fusion module adds modal weights, low confidence flags, and encoder version numbers to form a fusion vector table record. The record fields include media identifier, sampling time, environmental exposure vector, reading degradation vector, media apparent aging vector, three types of modal weights, fusion vector, encoder version, training data time range, and abnormal input summary. When all modal availability masks are zero, the system does not generate a fusion vector and returns a supplementary sampling task identifier to S1. The training scheduling module registers the self-supervised temporal encoder, read-down degenerate encoder, image encoder, spectral encoder, and the graph Transformer used by S5 in the same model version table. The table fields include model name, input field set, output dimension, training sample time range, loss function, validation metric, freeze time, and rollback version. When the validation metric of any encoder is lower than the activation threshold in the version table, the system continues to use the previous frozen version for inference and marks the current multimodal fusion vector as the inference result of the old version. The fusion vector table output by S2 serves as the source of node features for establishing the dynamic media relationship graph in S3, and as the input vector for matching the media aging prototype memory bank in S4. The system saves the reverse index of the sample sequence in S1 in each fusion vector record, so that subsequent reading risk calculation can be traced back to the original window of environmental state data, reading behavior data and media aging detection data.

[0021] In this specific embodiment, S3 includes: The graph construction module reads the fusion vector table output by S2 and builds a node table with each inorganic media storage unit as a node. The node fields include media identifier, batch number, storage cabinet location, material type, write time, window master reading device identifier, sampling time, multimodal fusion vector, environmental exposure vector, modal weight, missing test mask, and node update time. The node primary key consists of media identifier and sampling time. When a new sampling time arrives for the same medium, the node features are updated with the multimodal fusion vector as the node's main feature, and the previous window node snapshot is retained. The graph construction module first performs candidate source cleaning on the node table, deleting node records with empty media identifiers, missing batch numbers, and those that cannot be filled back from the archive directory. When the material type is missing, it is filled back from the media entry table according to the media identifier. When the storage cabinet is missing, the node is retained, but the generation of edges related to the same cabinet is prohibited. The cleaned node records are written into the graph construction input table and the reasons for deletion and the source of filling are retained for subsequent graph sample verification. Environmental exposure similarity is calculated from the environmental exposure vectors of two nodes, according to the following rules: ,in and Representing nodes respectively and nodes Environmental exposure vector, To construct a threshold table for the graph, a zero-prevention constant of 0.000001 is used. The range is written after being cropped. When any node has a low-confidence environment vector state, the candidate edge can still be retained by batch, cabinet location, material, write time or read device conditions, but the environment exposure similarity field is marked as low confidence. When the same medium is read by multiple reading devices within the same sampling time window, the graph construction module does not directly concatenate all device identifiers into the node relationship judgment field. Instead, it stores the device set, reading duration of each device, most recent reading time, and number of verification failures in the reading device details table, and determines the main reading device identifier of the window in the order of longest reading duration and most recent reading time when the durations are the same. The edge generation conditions, node table fields, and subsequent device deviation prototypes all use the main reading device identifier of this window. The device set is only stored as an explanation field and a device deviation verification field, thereby ensuring that the relationship type of the same reading device has a unique and reproducible judgment object. The edge generator writes one edge record for each retained candidate node pair. The relation type field is encoded according to the trigger conditions as multi-hot encoding based on same batch, same cabinet location, same material, recent write time, same reading device, and similar environmental exposure. The time interval field is the absolute value of the number of days between two write times, normalized to a maximum of 180 days. The exposure similarity field uses... If multiple triggering conditions are met simultaneously, all of them are written into the multi-hot encoding of the relation type without creating edges repeatedly. The edge record also includes the edge update time, threshold table version, and candidate generation reason. The dynamic update module maintains the graph samples according to the sampling time of newly added detection data, and the node update rules are as follows: ,in This represents the multimodal fusion vector output by S2 and used as the main feature of the node. Additional structured attribute encoding representing material type, storage period, reading frequency, and environmental fluctuation level; edge feature update rule is as follows. ,in Represents relation type encoding, Indicates the normalized time interval. The environmental exposure similarity is represented, and the updated node main features, additional structured attributes, and edge features are written to the dynamic graph cache; The graph construction threshold table stores the time difference interval, similarity threshold, zero constant protection, low confidence similarity processing method and version number. Its calibration source is the environmental exposure vector of media in the same batch and cabinet in history, the environmental exposure vector of media randomly paired in different cabinets, and confirmed samples exposed to moisture or thermal cycling. Each record in the threshold table includes material type, cabinet area, similarity threshold, validity date and sample quantity. When the graph construction module does not hit the material type, the full material threshold is used and a threshold miss flag is written in the edge record. When a node's newly detected data lacks the fusion vector output by S2, the system retains the node features from the previous window and marks the node status as waiting for encoding, with a retention time not exceeding two sampling periods. When the candidate node pair set is empty, the graph construction module still outputs a single-node graph sample containing only the target node and self-loop edges, with the self-loop edge relationship type set to its own historical continuation and the exposure similarity set to 1, to avoid the S5 graph Transformer being unable to reason due to empty neighborhoods. The edge table update also executes expired edge pruning and conflict edge retention rules. If two nodes no longer meet any candidate conditions in the last two sampling periods, the edge is marked as expired and removed from the current graph sample. If the same node pair has both a relationship with the same reading device and a low-confidence environment exposure similarity relationship, the relationship with the same reading device is retained and the low-confidence environment exposure similarity is used as an additional edge field instead of as the reason for edge creation. The pruning results are written to the edge change log for S5 to explain the source of the reading risk level. The graph construction input table, graph construction threshold table, node table, edge table, and edge change log are associated with the same graph version number. When new detection data arrives, only the neighborhood subgraphs exposed in the same batch, cabinet, material, or similar environment as the target node are recalculated. Unaffected nodes and edges use the previous version. The system records the incremental update range, the number of recalculated nodes, the number of recalculated edges, and the update time in the graph sample metadata so that the dynamic graph state at that time can be restored during subsequent graph Transformer training and replay. The dynamic graph cache also maintains a neighborhood statistics summary for each target node. The summary fields include the number of neighbors in the same batch, the number of neighbors in the same cabinet, the number of neighbors with the same material, the number of neighbors with similar environmental exposure, the average exposure similarity, the maximum time interval, and the number of low-confidence edges. These summaries do not replace the node feature matrix and edge feature matrix, but are written into the graph sample metadata for subsequent verification of the graph construction results. When new detection data causes changes in node material type, cabinet location, or window main reading device, the graph construction module first closes the candidate edges triggered by the old attributes, then regenerates the candidate edges according to the new attributes, and writes the old edge closing time, the new edge effective time, and the source of attribute change into the edge change log to ensure that the relationship type, time interval, window main reading device, and exposure similarity in the dynamic graph are consistent with the current media archive record. The output of the graph sample includes a node table, an edge table, a node feature matrix, an edge feature matrix, and a target node index. The target node index is used to specify the media storage unit for which the reading risk needs to be assessed. The graph sample is saved to the dynamic media relationship graph library with the batch number, sampling date, and graph version number as keys. The graph version number increases with the sampling time of newly added detection data. Old version graph samples are retained for graph Transformer training playback. The dynamic graph cache simultaneously passes the multimodal fusion vector of the target node and the graph sample identifier to S4, so that S4 can append the aging stage vector to the same node features and pass the reverse indexes of the node table and edge table to S5, so as to trace the source of its neighborhood relationship for reading risk level, page fault probability range and archiving disposal suggestions.

[0022] In this specific embodiment, S4 includes: The prototype matching module reads the target node multimodal fusion vector and graph sample identifier transmitted by S3, and calls the medium aging prototype memory. In this specific embodiment, the memory contains five types of prototype units: normal aging, moisture exposure, thermal cycling exposure, reading interval instability, and equipment deviation. Each prototype unit stores the prototype category, center vector, covariance matrix, number of samples, last update time, reliability weight, material type coverage, and prototype version number. The center vector has the same dimension as the 384-dimensional multimodal fusion vector output by S2. The covariance matrix is ​​used for subsequent candidate prototype outlier elimination and reliability maintenance, but does not replace cosine similarity matching. The system initializes each prototype category based on historically confirmed samples. Normal aging prototypes are samples whose storage years have increased but whose verification failure records are zero and whose apparent aging test has not triggered anomalies. Moisture exposure prototypes are samples whose average relative humidity or particulate matter concentration has continuously exceeded the corresponding material calibration threshold and whose moisture traces have been confirmed by re-inspection. Thermal cycling exposure prototypes are samples whose temperature fluctuation exceeds the thermal cycling threshold and whose spectral peak position has undergone a stable shift. Equipment deviation prototypes are samples whose reread count and error correction count are significantly higher than those of other devices for the same reading device. The prototype for unstable read interval is updated by samples whose read interval is greater than the 75th percentile of the historical read interval in the same batch and whose number of rereads is greater than the 75th percentile of the historical number of rereads in the same batch. The read interval is calculated as the difference in the number of days between the current read time and the previous read time. The historical percentile is constructed from the historical read logs of the same batch and written into the prototype threshold table. If the number of historical samples in the same batch is less than 30, the historical percentile of the same material type is used and the prototype update record is marked as the cross-batch threshold source. Before prototype matching, the input check thread reads the modal weights and low confidence flags of the target node fusion vector. When the fusion vector is missing, the dimension is not equal to 384, or all modal weights are zero, the system stops prototype matching and writes the graph sample identifier into the supplementary encoding queue. When there is only a single modal low confidence flag, the system retains the target node and records the low confidence source in the matching similarity distribution for reference when calculating the S5 confidence flag. The matching module performs cosine similarity calculation between the target node fusion vector and each prototype center vector, according to the following rules: ,in This represents the multimodal fusion vector of the target node. Indicates the first class prototype unit center vector, The zero-prevention constant in the prototype parameter table is 0.000001. The larger the value, the closer the target node is to the corresponding aging category. Prototype matching items that are below the effective lower limit of similarity still retain their values, but are reduced in weight when the prototype reliable weights are multiplied. The prototype reliability weight is obtained by multiplying the sample size normalization factor by the time decay factor, and the calculation rule is as follows: ,in Indicates the first Number of prototype unit samples Set it to 30 in the prototype parameter table. This represents the time decay factor obtained by looking up the table based on the most recent update time, ranging from 0.5 to 1, when the most recent update time is more than 180 days from the current time. Take 0.5, write the reliable weight back to the prototype memory and save it along with the prototype version number; The aging stage vector is obtained by multiplying the matching similarity distribution by the reliable weight, according to the following rule: ,in Indicates the target node is at the . The aging phase component on the class prototype, The sum of the five components is 0.000001, and is normalized to 1 when a valid prototype exists. If the reliability weights of all prototypes are less than 0.2, the system will set the aging stage vector to a uniform distribution and write the prototype low reliability flag to allow S5 to reduce the confidence of the model output. When the sample category is confirmed by the re-examination or migration results, the prototype maintenance thread updates the corresponding center vector and covariance matrix according to a preset learning rate. The center vector update rule is as follows: The covariance matrix is ​​updated exponentially by the outer product of the confirmed sample fusion vector and the updated center vector. Set as follows in the prototype parameter table After the update, the sample count is incremented by 1 and the latest update time is refreshed. If it is confirmed that there is a record of failure to recover after migration, the record is synchronously written to the S5 graph Transformer training queue. The system establishes transition rules for new category candidates, first calculating the highest similarity between the confirmed sample and all prototype units. ,in The cosine similarity between the dimensionless multimodal fusion vector and the dimensionless prototype center vector. It is still a dimensionless similarity value, if Below the new prototype threshold Then the confirmed sample is written into the candidate prototype unit. The prototype units are determined by the 10th percentile of the highest similarity of historical confirmed samples and stored in the prototype threshold table. When the number of candidate prototype units reaches the preset threshold of 20 and their internal average similarity is not less than 0.75, they are converted into formal prototype units. Otherwise, they are retained as candidates to be reviewed and do not participate in the S4 matching output. The prototype parameter table stores the prototype category, effective lower limit of similarity, zero protection constant, sample number normalization constant, time decay factor table, learning rate upper limit, new prototype threshold and candidate conversion number threshold. The table records are from historical confirmed samples, re-inspection records, migration read results and manual review records. The version update is triggered when the number of newly confirmed samples reaches 100 or the recovery failure rate distribution of any material type drifts. Before the candidate prototype unit is officially adopted, it saves the candidate sample fusion vector, confirmed category source, re-examination time, migration result and average similarity within the sample. The maintenance thread calculates the covariance trace within the candidate unit every week and removes outlier candidate samples. The outlier determination threshold is determined by the 90th percentile of the covariance trace of the official prototype. The removed samples are written into the list to be manually reviewed. The candidate samples that are not removed continue to accumulate to the preset number threshold. The prototype memory also stores the category sorting map. The five fixed prototypes are output in the order of normal aging, moisture exposure, thermal cycling exposure, reading interval instability and equipment deviation. The newly added formal prototypes are placed after the fixed prototypes and carry the candidate source identifier. When S5 reads the aging stage vector, it also reads the category sorting map to avoid the misalignment of node attached features caused by the change of prototype order during the training and inference stages of the graph Transformer. The prototype matching module appends the aging stage vector to the tail of the target node features in the S3 dynamic graph cache and generates the merged graph sample. The node features use the multimodal fusion vector as the main feature and the structured attribute encoding as the additional attribute input. The aging stage vector is merged as the node additional feature generated by S4. At the same time, the prototype category order, reliability weight vector, low reliability flag and prototype version number are written. If the target node already has the previous version aging stage vector, the vector corresponding to the current sampling time is used to overwrite the current additional feature of the node and the previous version is stored in the historical node snapshot. The merged graph sample identifier, aging stage vector, matching similarity distribution, reliable weight vector, and prototype version number are passed to S5's graph Transformer inference service. S5 can then use five types of prototype knowledge—normal aging, moisture exposure, thermal cycling exposure, read interval instability, and device deviation—when outputting a unified archive health score, read risk level, and predicted page fault probability, while maintaining a clear source of node-attached features.

[0023] In this specific embodiment, S5 includes: A page fault refers to a data block or sector that cannot be recovered after three rereads and error corrections during the reading process of inorganic media. The page fault flag is generated by the page fault record field in the read log and takes a value of 0 or 1. The recovery failure record refers to the event record when the file cannot be fully recovered, the verification fails, or the page fault flag is 1 during the re-examination or migration verification. Both the training service and the calibration layer extract supervision flags from the read log, re-examination log and migration verification log according to this definition. The Graph Transformer inference service reads the merged graph sample from the S4 output. The graph sample contains node feature matrices, edge feature matrices, target node indices, prototype reliable weight vectors, and missing test masks. The service first performs linear projection on the node features to obtain... The model is represented by a dimensional node, and edge bias mapping is performed on the edge features composed of relation type, time interval and exposure similarity. The edge bias results are written into the bias table of each attention head, and the model version number is consistent with the frozen version registered in S2. The Graph Transformer performs multi-head attention computation on edge feature biases at each layer, with the attention rule being: ,in and Represents a node and neighboring nodes The upper layer representation, and Indicates the first The projection matrix of each attention head. Indicates a single-head dimension. Represents edge features The bias value obtained by mapping, the attention score of severely missing nodes is reduced by the missing test mask penalty term, and the neighborhood aggregation output is written into the target node representation through residual connection and feedforward network; The scoring header reads the target node representation and outputs a unified archive health score; the probability header outputs the estimated page fault probability; and the risk level is output by the risk classification header, calculated according to the following rules. , , ,in This represents the target node representation. The range is 0 to 1, with higher values ​​indicating better archive health. The value ranges from 0 to 1, with higher values ​​indicating a higher expected probability of page faults. This indicates the risk level to be read. The scoring head, probability head, and classification head are all trained using labeled recovery failure records, page fault records, and verification failure records. The training service constructs Transformer training samples using historical inorganic media storage cell samples. Supervision information includes recovery failure labels, page fault markers, verification failure records, and read risk level labels. The training loss is... ,in For the regression loss of health scores, The binary cross-entropy loss is used for page fault probability. To read the risk level cross-entropy loss, the weights , and The training parameters are set to 0.3, 0.4, and 0.3. This model version is not enabled when the calibration error of the expected page fault probability on the validation set exceeds 0.05. The conformal calibration layer includes a calibration residual selection module, an interval generation module, and a disposal rule module. The calibration residual selection module selects the target residual set according to material type, storage period, reading frequency, and environmental fluctuation level. The historical calibration sample group is divided by the same material type, preset storage period interval, preset reading frequency interval, and environmental fluctuation level. The environmental fluctuation level is obtained by weighting the normalized variance of temperature, relative humidity, and vibration intensity within the most recent 16 windows. The calibration grouping table fields include material type, storage period interval, reading frequency interval, environmental fluctuation level, sample number, target coverage, residual quantile, minimum sample number, and version number. This table is constructed from historical samples that have recently completed re-inspection or migration verification and is updated monthly. For each historical calibration sample, the calibration residual selection module calculates the non-compliance score between the predicted page fault probability and the actual page fault mark. ,in Representing historical samples The expected probability of a page fault. Indicates the actual page fault marker; at the target coverage The quantile of the non-compliance score is taken as the calibration residual, according to the following rule: ,in Describe the target residual set. This indicates the number of samples in the set. The target coverage is set to 0.90 in the calibration parameter table. Write the residual quantile value field into the calibration grouping table and make it available for the current sample group to which the medium belongs; The conformal calibration layer's interval generation module performs interval calibration on the estimated page fault probability of the current medium. The page fault probability interval is... ,in Indicates the lower bound of the interval. Indicates the upper bound of the interval. The risk value is read from the target residual set of the current medium, setting values ​​less than 0 at the lower bound of the interval to 0 and values ​​greater than 1 at the upper bound of the interval to 1. Generate, where , and All are risk components in the range of 0 to 1. The larger the value, the higher the reading risk. The scoring weights are determined by the calibration parameter table on the validation set and sum to 1. The results are written into the risk output table. The conformal calibration layer's handling rules module reads the risk output table, risk level, and confidence flag, and generates archiving handling suggestions according to the threshold table. When the read risk value is not less than the migration risk threshold, the upper bound of the page fault probability interval is not less than the migration probability threshold, or the confidence flag is a calibration flag to be reviewed, data migration suggestions and redundancy verification suggestions are generated. When the upper bound of the page fault probability interval is less than the migration probability threshold but not less than the reread probability threshold, a reread suggestion is generated. When the upper bound of the page fault probability interval is less than the reread probability threshold and the unified archiving health score is not greater than the sampling health threshold, a sampling suggestion is generated. When the upper bound of the page fault probability interval is less than the reread probability threshold and the unified archiving health score is greater than the sampling health threshold, the recommended reread cycle is determined according to the mapping table between the upper bound of the page fault probability interval and the reread cycle. The conformal calibration layer's handling rules module reads the risk output table, risk level, and confidence flag. Based on the read risk level, read risk score, page fault probability interval, and confidence flag, it generates archiving handling suggestions according to the threshold table. When the read risk score is not less than the migration risk threshold, the upper bound of the page fault probability interval is not less than the migration probability threshold, or the confidence flag is a calibration flag to be reviewed, it generates data migration suggestions and redundancy verification suggestions. When the upper bound of the page fault probability interval is less than the migration probability threshold but not less than the reread probability threshold, it generates a reread suggestion. When the upper bound of the page fault probability interval is less than the reread probability threshold and the unified archiving health score is not greater than the sampling health threshold, it generates a sampling suggestion. When the upper bound of the page fault probability interval is less than the reread probability threshold and the unified archiving health score is greater than the sampling health threshold, it determines the recommended reread cycle according to the mapping table between the upper bound of the page fault probability interval and the reread cycle. The priority order for handling recommendations is as follows: migration and redundancy check are the highest priority, reread is the second priority, sampling inspection is the third priority, and writing during the reread inspection cycle is the lowest priority. When multiple conditions are triggered on the same medium in the same evaluation cycle, only the highest priority recommendation is output and the remaining triggering conditions are saved as additional reasons. If the read risk level is the highest and the upper bound of the page fault probability interval is lower than the reread probability threshold, then migration and redundancy check recommendations are still output, and the reread requirement is written as an additional reason and threshold conflict reason into the review field. The reread recommendation is not output as the main recommendation separately. After generating the master recommendation, the disposal rules module synchronously writes it to the archived disposal record. The record fields include media identifier, assessment period, read risk level, read risk score, estimated page fault probability, lower bound of page fault probability interval, upper bound of page fault probability interval, confidence index, master recommendation, additional reasons, threshold version, and calibration group version. The read risk score is directly obtained from the runtime output of the interval generation module according to the risk formula. The threshold table is only used to determine the migration, rereading, sampling or re-inspection cycle branches, and does not recalculate or replace the risk score, so that the treatment suggestions can be traced from the graph Transformer output, the shape-preserving calibration results and the rule branches one by one; The above output links constitute the final evaluation process of the inorganic media archiving evaluation method based on machine learning. The machine learning model is responsible for generating health scores, risk levels, and page fault probabilities from multimodal fusion vectors, dynamic media relationship graphs, and aging stage vectors. The calibration residual selection module and interval generation module in the conformal calibration layer are responsible for probability intervals and confidence indicators. The disposal rule module in the conformal calibration layer converts the model output, conformal calibration residuals, and threshold tables into executable archiving disposal suggestions, without replacing the model inference itself.

[0024] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0025] This invention connects environmental conditions, read behavior, and media aging detection results into a continuous archive health assessment chain through multimodal sample sequences, dynamic media relationship diagrams, and graph Transformers. This allows the read risk of inorganic media storage units to no longer rely solely on file status or single verification results, but to make a comprehensive judgment by combining long-term environmental exposure, read degradation trends, apparent aging performance, and the correlation between similar media. This more directly solves the problem that critical data may be distorted or have missing pages when recovered years later, but which is difficult to detect in time in the early stages.

[0026] The media aging prototype memory and conformal calibration layer do not change the basic object of archiving assessment, but rather improve the structure to address the differences in aging stages and the insufficient reliability of probability outputs in the long-term storage of inorganic media. The aging stage vector enables the graph Transformer to utilize prototype knowledge such as normal aging, moisture exposure, thermal cycling exposure, read interval instability, and equipment deviation. The conformal calibration layer converts the read risk output into page fault probability intervals and confidence indicators, and generates archiving disposal suggestions based on the read risk level, read risk score, page fault probability interval, and confidence indicator through the disposal rule module. This makes the suggestions for rereading, sampling, migration, and redundancy verification more suitable for early warning and review disposal in long-term archiving management.

Claims

1. A machine learning-based method for evaluating inorganic media archiving, characterized in that, include: S1. Collect environmental status data, reading behavior data and media aging test data of inorganic media storage units, align them according to media identification and sampling time, and form a multimodal sample sequence; S2. Input the multimodal sample sequence into the self-supervised temporal encoder, readout degradation encoder, and image encoder or spectral encoder obtained by training to obtain the environmental exposure vector, readout degradation vector and media apparent aging vector, and fuse them into a multimodal fusion vector. S3. Establish a dynamic media relationship graph based on media batch, storage cabinet location, material type, write time, window main read device, and environmental exposure similarity. The node features include multimodal fusion vectors, and the graph samples are obtained with relationship type, time interval, and exposure similarity as edge features. S4. Perform similarity matching between the multimodal fusion vector and the prototype units in the media aging prototype memory. Generate an aging stage vector based on the matching similarity and prototype reliability weights. Incorporate the aging stage vector into the node features of the graph samples. Set normal aging, moisture exposure, thermal cycling exposure, read interval instability, and equipment deviation as prototype categories respectively. Among them, the read interval instability prototype is updated by samples whose read interval is greater than the 75th percentile of the historical read interval in the same batch and whose reread count is greater than the 75th percentile of the historical reread count in the same batch. Generate the aging stage vector based on the matching similarity and prototype reliability weights. Specifically, this includes generating a matching similarity distribution by using the cosine similarity between the multimodal fusion vector and the center vector of each prototype unit, and multiplying the matching similarity distribution by the prototype reliability weights determined by the number of prototype samples and the most recent update time to obtain the aging stage vector. The aging stage vector generated by the matching similarity and prototype reliability weight is used as the node's additional feature input to the subsequent graph Transformer; each prototype unit in the media aging prototype memory includes a prototype category, center vector, covariance matrix, number of samples, most recent update time, and reliability weight; when the reading result after re-examination or migration confirms the sample category, the corresponding center vector and covariance matrix are updated using the multimodal fusion vector of the confirmed sample according to a preset learning rate; the reliability weight is obtained by multiplying the sample number normalization factor with the time decay factor obtained by looking up the table according to the most recent update time; when the highest similarity value between the confirmed sample and all prototype units does not reach the new prototype threshold determined by the 10th percentile of the highest similarity of historical confirmed samples, the confirmed sample is written into a candidate prototype unit, and the candidate prototype unit is converted into a prototype unit after the number of samples reaches a preset number threshold; S5. Input the merged graph samples into the trained graph Transformer; The graph Transformer performs multi-head attention calculation and neighborhood aggregation on the dynamic medium relationship graph to obtain the target node representation; By representing the target node with the input score header and probability header, we can obtain the unified archive health score, read risk level and expected page fault probability, respectively. The conformal calibration layer generates a read risk score, a page missing probability range, and a confidence level. Based on the read risk level, read risk score, page missing probability range, and confidence level, it generates archiving and disposal recommendations. The conformal calibration layer is selected according to material type, storage period, read frequency, and environmental fluctuation level.

2. The machine learning-based inorganic media archiving evaluation method according to claim 1, characterized in that, S1 includes: Environmental condition data includes temperature, relative humidity, light intensity, particulate matter concentration, vibration intensity, and storage cabinet opening and closing events; The reading behavior data includes reading time, reading device identification, reading duration, number of rereads, error correction count, reading rate, and verification failure records; the media aging detection data includes media surface images, reflectance spectra, transmission spectra, coating integrity detection values, and edge defect detection values; the data are merged into time windows according to media identification, batch number, and sampling time, and uncollected items are recorded using a missing test mask to obtain the multimodal sample sequence.

3. The machine learning-based inorganic media archiving evaluation method according to claim 1, characterized in that, S2 include: The environmental state data is divided into preset fixed-length windows in chronological order. The environmental exposure vector is obtained by reconstructing the occlusion segment and predicting the next window using a self-supervised temporal encoder. The number of rereads, error correction counts, reading rate, and verification failure records in the reading behavior data are input into the reading degradation encoder to obtain the reading degradation vector. The image or spectrum in the medium aging detection data is processed by convolution feature extraction and spectral peak normalization, and then input into the corresponding encoder to obtain the medium apparent aging vector; after applying normalized weights determined by modal availability mask to the three types of vectors respectively, they are concatenated to obtain the multimodal fusion vector.

4. The machine learning-based inorganic media archiving evaluation method according to claim 1, characterized in that, S3 includes: using each inorganic media storage unit as a node; the window master reading device is the reading device with the longest reading duration within the current sampling time window, and the most recent reading device is used when the reading durations are the same; when two nodes have the same media batch, the same storage cabinet, the same material type, the writing time difference falls within a preset time difference interval, the window master reading device is the same, or the environmental exposure similarity is not less than a preset similarity threshold, an edge is established between the corresponding nodes; the environmental exposure similarity is calculated by the cosine similarity of the environmental exposure vectors of the two nodes; the relationship type, time interval, and exposure similarity of the edge are encoded as the edge features, and the node features and edge features are updated according to the sampling time of the newly added detection data to obtain the dynamic media relationship graph.

5. The machine learning-based inorganic media archiving evaluation method according to claim 1, characterized in that, S5 includes: selecting a target residual set from the calibration residual set based on material type, storage period, reading frequency, and environmental fluctuation level; performing conformal calibration on the expected page defect probability using the target residual set; outputting a reading risk score, a page defect probability range, and a confidence flag indicating whether the target coverage is met; and generating the archiving disposal recommendation based on the reading risk level, reading risk score, page defect probability range, and confidence flag.

6. The machine learning-based inorganic media archiving evaluation method according to claim 3, characterized in that, The self-supervised temporal encoder, read degradation encoder, image encoder, spectral encoder, and graph Transformer are trained as follows: using historical inorganic media storage unit samples as training samples, the self-supervised temporal encoder is trained using temporal segment occlusion reconstruction loss and next window prediction loss; the read degradation encoder is trained using read degradation labels or degradation pseudo-labels constructed from reread counts and error correction counts; the image encoder and spectral encoder are trained using batch consistency loss of media aging detection samples; and the scoring head and probability head of the graph Transformer are trained using labeled recovery failure records, page fault records, and verification failure records.

7. The machine learning-based inorganic media archiving evaluation method according to claim 5, characterized in that, The calibration residual set is constructed as follows: Historical calibration sample groups were divided according to material type, preset storage period range, preset reading frequency range, and environmental fluctuation level; the non-compliance score between the expected page miss probability and the actual page miss mark was calculated for each historical calibration sample. The quantile of the non-compliance score under the target coverage is taken as the calibration residual; the calibration residual corresponding to the sample group to which the current media storage unit belongs is added and subtracted from the expected page fault probability, and the lower bound of the interval is set to 0 and the upper bound of the interval is set to 1 to obtain the page fault probability interval; When the number of samples in the sample group to which the current media storage unit belongs is not less than the preset minimum number of samples, it is considered that the target coverage guarantee is met and the confidence flag is a valid calibration flag; otherwise, it is considered that the target coverage guarantee is not met and the confidence flag is a calibration flag to be verified.

8. The machine learning-based inorganic media archiving evaluation method according to claim 7, characterized in that, Archiving and disposal recommendations are generated in the following way: When the read risk score is not less than the migration risk threshold, the upper bound of the page fault probability interval is not less than the migration probability threshold, or the confidence flag is a flag to be reviewed and calibrated, generate data migration suggestions and redundancy check suggestions. When the upper bound of the page fault probability interval is less than the migration probability threshold and not less than the repeat probability threshold, a repeat recommendation is generated; when the upper bound of the page fault probability interval is less than the repeat probability threshold and the unified archive health score is not greater than the sampling health threshold, a sampling recommendation is generated; when the upper bound of the page fault probability interval is less than the repeat probability threshold and the unified archive health score is greater than the sampling health threshold, a recommended re-inspection cycle is determined according to the mapping table between the upper bound of the page fault probability interval and the preset re-inspection cycle, and the recommended re-inspection cycle is written into the media archive record.

Images