A security intelligent early warning system for an archive storage environment

Abstract

This invention discloses an intelligent early warning system for the security of archival storage environments, comprising: an environmental monitoring signal acquisition module for providing raw data support; an environmental monitoring signal preprocessing module for preprocessing and generating standard basic signals; an environmental risk feature generation module for extracting composite risk features and constructing multidimensional feature sequences; a non-uniform patch partitioning module for dynamically adjusting patch parameters using a sensitivity algorithm to achieve accurate sampling of sudden risk; a regional risk embedding module for integrating regional coding to achieve spatial semantic enhancement of risk features; a damage accumulation judgment module for capturing long-term and short-term risks and accumulating historical damage contributions through coupled modulation and temporal coding; and a graded early warning alarm output module for executing non-negative incremental updates and monotonically non-decreasing constraints to output graded alarm commands. This invention achieves accurate capture of archival environmental risks and monotonically cumulative prediction of irreversible damage, improving the physical reliability and timeliness of early warning.

Description

Technical Field

[0001] This invention relates to the field of archive storage security technology, and in particular to an intelligent early warning system for archive storage environment security. Background Technology

[0002] Current archival storage environment security management mainly relies on single environmental monitoring devices to collect physical parameters such as temperature, humidity, and light, triggering alarms through fixed thresholds. Conventional environmental monitoring systems mostly use equal-interval sampling and fixed alarm rules, lacking the ability to respond sensitively to rapid changes in environmental anomalies or long-term cumulative damage, making it difficult to achieve refined risk assessment of archival materials. Existing systems cannot fully consider the differences in archival material types, storage areas, and equipment coverage, resulting in the inability to effectively distinguish the risk characteristics of different areas under the same environmental parameters, leading to low accuracy and reliability of early warning information. Single-channel data processing lacks the ability to fuse multi-source information, making it impossible to dynamically analyze the interaction of multiple factors such as temperature and humidity coupling, light and heat superposition, smoke and gas correlation, and pest activity, making it difficult to capture the complex impact of environmental changes on archival damage.

[0003] Existing technologies for environmental data time-series analysis mostly employ fixed-length time slices or simple sliding window processing methods, which are insufficient for simultaneous monitoring of sudden environmental events and continuous trends. This results in high-risk changes not being reflected in alarm results in a timely manner. Traditional cumulative damage assessment relies on single-point thresholds or simple superposition calculations, lacking a mechanism for the continuous accumulation and retention of historical damage information, and thus failing to accurately assess the long-term damage status of records. Existing systems generate alarm results with uniform threshold outputs, lacking hierarchical strategies, and are unable to distribute warning information of different risk levels to audible and visual alarm devices, management terminals, mobile receivers, access control systems, environmental control equipment, and fire alarm linkage interfaces to achieve refined management.

[0004] Existing technologies for monitoring and early warning of archival environment security have shortcomings in multi-source data fusion, dynamic time-series analysis, cumulative damage modeling, and tiered alarm output, failing to meet the high-precision, dynamic, and regionally differentiated security early warning requirements of archival management. Therefore, how to provide an intelligent early warning system for archival storage environment security is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent early warning system for the security of archival storage environments. This invention employs non-uniform risk perception patch partitioning, multi-source channel coupled attention, Transformer temporal coding, and irreversible cumulative memory technology to dynamically analyze and model the composite risk feature sequence of the archival environment and accumulate risks. It has the advantages of accurate early warning, clear risk classification, and sensitivity to sudden and long-term environmental changes.

[0006] An intelligent early warning system for the security of an archive storage environment according to an embodiment of the present invention includes:

[0007] The environmental monitoring signal acquisition module is used to acquire multi-source monitoring signals of the archived environment, add acquisition time identifiers, sensor numbers and storage area codes, and generate raw environmental monitoring signals.

[0008] The environmental monitoring signal preprocessing module is used to preprocess the raw environmental monitoring signals to generate archival environmental monitoring base signals.

[0009] The environmental risk feature generation module is used to extract composite risk features of the archive environment based on basic signals from archive environmental monitoring and generate a sequence of archive environmental risk features.

[0010] The non-uniform patch partitioning module is used to input the archival environment risk feature sequence into the improved archival environment risk perception patchTST model. Based on the rate of change, the amount of continuous deviation, the magnitude of mutation and the risk sensitivity level, the module calculates the patch length, the overlap ratio and the sliding step size through the sensitivity mapping algorithm to generate a non-uniform risk perception environment patch sequence.

[0011] The regional risk embedding module is used to construct the archive regional risk code and embed the non-uniform risk perception environment Patch sequence to generate the regional enhanced Patch feature sequence.

[0012] The damage accumulation determination module is used to perform cross-parameter coupling modulation, long-period trend encoding, short-time perturbation encoding and historical damage contribution accumulation retention on the region enhancement patch feature sequence to generate an archive damage accumulation feature sequence.

[0013] The graded early warning and alarm output module is used to perform non-negative incremental updates and time monotonically non-decreasing constraints on the cumulative characteristic sequence of archive damage, generate early warning results for archive environmental safety, and send graded alarm commands.

[0014] Optionally, the preprocessing process in the environmental monitoring signal preprocessing module includes: sorting the original environmental monitoring signals according to the acquisition time identifier, completing missing values, removing outliers and normalizing dimensions, binding storage area codes, and generating archived environmental monitoring basic signals.

[0015] Optionally, the process of extracting composite environmental risk features from the archives in the environmental risk feature generation module includes:

[0016] Read the basic environmental monitoring signals from the archives according to the storage area code and the collection time identifier, and calculate the product of temperature change and humidity change to generate temperature and humidity coupling risk characteristics;

[0017] The product of the cumulative illumination and the temperature change within a preset sliding window period is used to generate a photothermal superposition risk characteristic.

[0018] The cross-correlation coefficient between changes in smoke concentration and changes in volatile gas concentration is calculated using a cross-correlation function to generate smoke gas associated risk characteristics.

[0019] The number of pest triggers and the interval between adjacent triggers within a preset sliding window period are statistically analyzed to generate pest activity risk characteristics.

[0020] The frequency of access control opening and closing within a preset sliding window period and the temperature and humidity fluctuations after the access control opening and closing are used to generate access control disturbance risk characteristics.

[0021] The differences in humidity and temperature before and after the operation of the environmental control equipment are calculated to generate the risk characteristics of equipment control failure.

[0022] Sort and concatenate the data according to the collection time identifier to generate an archive environmental risk characteristic sequence.

[0023] Optionally, the structure of the improved archival environment risk perception PatchTST model includes:

[0024] The risk feature input layer is used to receive the risk feature sequence of the archive environment, which is arranged as the model input sequence according to the storage area code and the collection time identifier.

[0025] The non-uniform risk perception patch partitioning layer is used to calculate the patch length, overlap ratio and sliding step size based on the rate of change, continuous deviation, mutation magnitude and risk sensitivity level in the model input sequence, and to perform non-uniform temporal sampling on the model input sequence to generate a non-uniform risk perception environment patch sequence.

[0026] The regional risk coding embedding layer is used to perform a fusion process on the archive regional risk coding and the non-uniform risk perception environment Patch sequence, including at least one of vector concatenation, position superposition and linear mapping, to generate a regional enhanced Patch feature sequence.

[0027] A multi-source channel coupled attention layer is used to perform cross-parameter coupled modulation on different risk feature channels in the region enhancement patch feature sequence to generate a multi-source coupled patch feature sequence.

[0028] The Transformer temporal coding layer is used to capture the continuous patch dependencies and abrupt patch response relationships in the multi-source coupled patch feature sequence through a multi-head attention mechanism, and generate temporal coding features.

[0029] An irreversible damage accumulation memory layer is used to perform cumulative writing and retention updates on historical damage contributions in temporal coding features through gated recurrent units and residual connection compensation, generating an archive damage accumulation feature sequence.

[0030] The archive damage risk prediction head is used to perform non-negative incremental updates on the cumulative feature sequence of archive damage, constraining the cumulative damage risk value to exhibit a monotonically non-decreasing characteristic on the time axis, and generating archive damage risk prediction results.

[0031] Optionally, the non-uniform risk perception patch partitioning layer includes the following operations:

[0032] The rate of change, sustained deviation, and magnitude of abrupt change in the archival environmental risk characteristic sequence are calculated based on the collection time identifier. The preset risk level table is read according to the archival material type and storage area code to determine the risk sensitivity level.

[0033] The rate of change, the amount of continuous deviation, the magnitude of the mutation, and the risk sensitivity level are input into the sensitivity mapping algorithm, and the judgment is performed in the order of the first sampling segment, the second sampling segment, and the third sampling segment.

[0034] When at least one of the following conditions is met—the rate of change exceeds the first rate of change threshold, the amount of continuous deviation exceeds the first amount of deviation threshold, the magnitude of mutation exceeds the first magnitude of mutation threshold, or the risk sensitivity level reaches the first risk level—it is marked as the first sampling segment, and the first patch length, first overlap ratio, and first sliding step size are configured.

[0035] When a region is not marked as the first sampling segment, and at least one of the following conditions is met: the rate of change exceeds the second rate of change threshold, the amount of continuous deviation exceeds the second deviation threshold, the magnitude of mutation exceeds the second magnitude threshold, or the risk sensitivity level reaches the second risk level, it is marked as the second sampling segment, and the second patch length, the second overlap ratio, and the second sliding step size are configured.

[0036] When it is not marked as the first sampling segment or the second sampling segment, it is marked as the third sampling segment, and the third patch length, third overlap ratio and third sliding step size are configured.

[0037] The first sampling segment, the second sampling segment, and the third sampling segment are sliced ​​in the time domain according to the corresponding Patch length, overlap ratio, and sliding step size, respectively, to generate the first Patch sequence, the second Patch sequence, and the third Patch sequence. The length of the first Patch is less than the length of the second Patch, the length of the second Patch is less than the length of the third Patch, the first overlap ratio is greater than the second overlap ratio, the second overlap ratio is greater than the third overlap ratio, the first sliding step size is less than the second sliding step size, and the second sliding step size is less than the third sliding step size.

[0038] The first, second, and third patch sequences are merged according to the acquisition time identifier to generate a non-uniform risk perception environment patch sequence.

[0039] Optionally, the regional risk coding embedding layer includes the following operations:

[0040] Generate archive area risk codes based on the type of archive materials and the code of the storage area;

[0041] The archive area risk code and the non-uniform risk perception environment Patch sequence are fused together using at least one of the following methods: vector concatenation, position overlay and linear mapping, to generate an area-enhanced Patch feature sequence.

[0042] The document type is combined with the spatial characteristics of the storage area to mark the Patch location, and each Patch contains spatial risk information for the corresponding storage area.

[0043] Optionally, the multi-source channel coupled attention layer includes the following operations:

[0044] The received region is enhanced with Patch feature sequences and classified according to risk feature channels;

[0045] Calculate the correlation coefficient matrix between each risk characteristic channel to generate a cross-parameter correlation matrix;

[0046] Based on the cross-parameter correlation matrix, weighted modulation is performed on each risk feature channel, and the relevant channel features are linearly combined according to the coupling weight to generate coupling enhancement features. These features are then arranged in order of acquisition time to form a multi-source coupling patch feature sequence.

[0047] Optionally, the irreversible damage accumulation memory layer includes the following operations:

[0048] Receive the timing coding features output by the Transformer timing coding layer according to the acquisition time identifier;

[0049] The temporal coding features of the current patch are added one by one to the archive damage cumulative feature vector corresponding to the previous acquisition time identifier to generate the candidate cumulative feature vector of the current patch.

[0050] Perform a non-negative incremental update on the candidate cumulative feature vector of the current patch, with the cumulative increment restricted to a non-negative value. Add the cumulative increment of the current patch to the cumulative feature value of the archive damage corresponding to the previous acquisition time identifier to generate the archive damage cumulative feature of the current patch.

[0051] Based on the collection time identifier, the cumulative archival damage features of each patch are combined to generate a cumulative archival damage feature sequence.

[0052] Optionally, the process of generating archive environment safety early warning results in the hierarchical early warning alarm output module includes:

[0053] Receive the archive damage cumulative feature sequence according to the collection time identifier;

[0054] Subtract the feature vector corresponding to the current acquisition time marker from the feature vector of the previous time step element by element, set negative values ​​to zero, and generate a non-negative increment;

[0055] The non-negative increment is added element by element to the feature vector of the previous time step to generate the current cumulative damage risk value; the cumulative value is checked in chronological order, and if it is less than the previous time step value, it is replaced with the previous time step value to form a monotonically non-decreasing sequence.

[0056] Based on the processed cumulative values, an early warning result for the archive environment security is generated;

[0057] The warning results are converted into tiered alarm commands and sent to audible and visual alarm devices, management terminals, mobile receivers, access control controllers, environmental control equipment, and fire alarm linkage interfaces.

[0058] The beneficial effects of this invention are:

[0059] (1) By using non-uniform risk perception patch partitioning and sensitivity mapping algorithm, dynamic sampling of composite risk feature sequence of archive environment is realized. The system can perform high-density analysis for high-risk time periods and reduce sampling frequency for low-risk time periods to improve the timeliness and accuracy of early warning.

[0060] (2) By using multi-source channel coupled attention layer, regional risk coding embedding and Transformer temporal coding, coupling modeling and temporal coding are performed on different risk feature channels and storage areas to realize cross-parameter and cross-regional risk correlation analysis, so that sudden events and long-term cumulative environmental changes can be accurately identified and predicted.

[0061] (3) An irreversible damage accumulation memory layer and a hierarchical early warning alarm output module are adopted to perform non-negative incremental updates and time monotonic non-decreasing constraints on the accumulated damage risk, generate hierarchical early warning results and distribution instructions, and realize continuous, hierarchical and operable intelligent early warning management of the archive storage environment security. Attached Figure Description

[0062] 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:

[0063] Figure 1 This is a module connection diagram of an intelligent early warning system for the security of archive storage environment proposed in this invention;

[0064] Figure 2 This is a schematic diagram of the multi-source coupled attention, temporal coding, and early warning output proposed in this invention;

[0065] Figure 3This is a schematic diagram of the non-uniform patch partitioning and regional risk coding embedding proposed in this invention. Detailed Implementation

[0066] 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.

[0067] refer to Figures 1-3 An intelligent early warning system for the security of archive storage environments, comprising:

[0068] The environmental monitoring signal acquisition module is used to collect temperature, humidity, light, smoke, dust, volatile gas, pest monitoring, vibration, access control opening / closing signals, and environmental control equipment operation signals from the archive storage environment. It adds acquisition time stamps, sensor numbers, and storage area codes to generate raw environmental monitoring signals. The environmental monitoring signal preprocessing module performs time synchronization, anomaly removal, missing data completion, unit normalization, and area code binding on the raw environmental monitoring signals to generate basic environmental monitoring signals for the archives. The environmental risk feature generation module extracts temperature, humidity, and wind characteristics from the basic environmental monitoring signals for the archives. The system generates a sequence of archival environmental risk characteristics, including risk features related to light and heat superposition, smoke and gas correlation, pest activity, access control disturbance, and equipment adjustment failure. A non-uniform patch partitioning module is used to input this sequence into the non-uniform risk perception patch partitioning layer of the improved archival environmental risk perception patchTST model. Based on the rate of change, sustained deviation, abrupt change amplitude, and risk sensitivity level, a preset sensitivity mapping algorithm calculates the patch length, overlap ratio, and sliding step size. Non-uniform temporal sampling is then performed on the archival environmental risk characteristic sequence to generate a non-uniform risk... The system comprises: a perception environment patch sequence; a regional risk embedding module, used to construct archive area risk codes based on sensor locations, archive shelf area codes, cabinet level codes, wall proximity markers, ventilation zone codes, and equipment coverage markers, embedding these archive area risk codes into the non-uniform risk perception environment patch sequence to generate a regional enhanced patch feature sequence; and a damage accumulation determination module, used to input the regional enhanced patch feature sequence into a multi-source channel coupled attention layer, a Transformer temporal coding layer, and an irreversible damage accumulation memory layer based on the archive degradation kinetic weight initial matrix, performing cross-parameter coupled modulation and long-period trend... The system retains potential coding, short-term disturbance coding, and historical damage contribution accumulation to generate a cumulative feature sequence of archival damage. A tiered early warning alarm output module inputs this cumulative feature sequence into an archival damage risk prediction head with physical monotonicity constraints. It performs non-negative incremental updates on the cumulative damage risk value, constraining the cumulative damage risk value to exhibit a monotonic, non-decreasing characteristic on the time axis. This generates an archival environmental safety early warning result containing risk type, risk area, risk level, and disposal recommendations. Based on the archival environmental safety early warning result, it generates tiered alarm commands and sends them to audible and visual alarm devices, management terminals, mobile receivers, access control controllers, environmental control equipment, and fire alarm linkage interfaces.

[0069] In this embodiment, the multi-source monitoring signals of the archive environment in the environmental monitoring signal acquisition module are represented by a unified field structure. Each signal record includes at least a signal type field, an acquisition value field, an acquisition time identifier field, a sensor number field, a storage area code field, and a device status field. Among them, the signal type field is used to identify the category of temperature, humidity, light, smoke, dust, volatile gases, pests, vibration, access control opening / closing, or environmental control equipment operation; the acquisition value field is used to record the corresponding sensor output value or device status value; the acquisition time identifier field is used to record the signal acquisition time; the sensor number field is used to identify the signal source; the storage area code field is used to identify the archive storage area, archive shelf area, or cabinet area corresponding to the signal; and the device status field is used to record the operating status of the corresponding environmental control equipment at the time of signal acquisition.

[0070] In this embodiment, the preprocessing process in the environmental monitoring signal preprocessing module includes: sorting the original environmental monitoring signals according to the acquisition time identifier, completing missing values, removing outliers and normalizing dimensions, binding storage area codes, and generating archived environmental monitoring basic signals.

[0071] In this embodiment, the process of extracting composite environmental risk features from the archives in the environmental risk feature generation module includes:

[0072] Read the basic environmental monitoring signals from the archives according to the storage area code and the collection time identifier, and calculate the product of temperature change and humidity change to generate temperature and humidity coupling risk characteristics;

[0073] Temperature change is the difference between the temperature value corresponding to the current acquisition time identifier and the temperature value corresponding to the previous acquisition time identifier under the same storage area code; humidity change is the difference between the humidity value corresponding to the current acquisition time identifier and the humidity value corresponding to the previous acquisition time identifier under the same storage area code; when the product of temperature change and humidity change is positive, the product value is retained as a unidirectional fluctuation feature; when the product is negative, the absolute value of the product and the direction of change identifier are recorded to form a temperature and humidity coupling risk feature;

[0074] The cumulative light intensity value and the temperature change within the preset sliding window period are calculated to generate the photothermal superposition risk characteristic; the cumulative light intensity value is the sum of the products of the light intensity value corresponding to each collection time marker and the adjacent collection time interval within the preset sliding window period.

[0075] The cross-correlation coefficient between the change in smoke concentration and the change in volatile gas concentration is calculated using the cross-correlation function to generate a smoke gas associated risk feature. Within a preset lag step range, the cross-correlation coefficient between the smoke concentration change sequence and the volatile gas concentration change sequence is calculated respectively, and the maximum value of the cross-correlation coefficient is selected as the smoke gas associated risk feature.

[0076] The pest activity risk characteristics are generated by statistically analyzing the number of pest triggers and the interval between adjacent triggers within a preset sliding window period. The pest activity risk characteristics consist of the number of pest triggers, the average interval between adjacent triggers, and the minimum interval between adjacent triggers within the preset sliding window period.

[0077] The access control system generates access control disturbance risk characteristics by statistically analyzing the frequency of access control opening and closing within a preset sliding window period and the temperature and humidity fluctuations after access control opening and closing; the temperature and humidity fluctuations after access control opening and closing are the differences between the maximum and minimum temperature values ​​and the differences between the maximum and minimum humidity values ​​within a preset disturbance observation window after the access control opening and closing time is marked.

[0078] The differences in humidity and temperature before and after the operation of the environmental control equipment are calculated to generate the risk characteristics of equipment control failure. The differences in humidity and temperature before and after the operation of the environmental control equipment are the differences between the average value of the preset baseline window before the equipment start time and the average value of the preset response window after the equipment start time.

[0079] The risk characteristics of temperature and humidity coupling, light and heat superposition, smoke and gas correlation, pest activity, access control disturbance, and equipment adjustment failure are sorted and spliced ​​according to the collection time identifier to generate an archive environmental risk characteristic sequence.

[0080] In this embodiment, the structure of the improved archival environment risk perception PatchTST model includes:

[0081] The risk feature input layer receives the sequence of archival environmental risk features, arranged by storage area code and collection time identifier as the model input sequence. The rate of change in the model input sequence refers to the increase or decrease ratio between the composite risk feature values ​​of the archival environment corresponding to adjacent collection time identifiers under the same storage area code. The continuous deviation refers to the cumulative deviation of the composite risk feature values ​​of the archival environment from the preset safety benchmark value in multiple consecutive collection time identifiers. The abrupt change amplitude refers to the instantaneous and rapid change amplitude of the composite risk feature values ​​of the archival environment between adjacent collection time identifiers. The risk sensitivity level refers to the priority weight assigned to different risk features based on the type of archival material, storage area characteristics, and historical environmental risk records, which is used to guide the adjustment of sampling density and sliding step size in non-uniform patch division.

[0082] The non-uniform risk perception patch partitioning layer is used to calculate the patch length, overlap ratio and sliding step size based on the rate of change, continuous deviation, mutation magnitude and risk sensitivity level in the model input sequence, and to perform non-uniform temporal sampling on the model input sequence to generate a non-uniform risk perception environment patch sequence.

[0083] The regional risk coding embedding layer is used to perform a fusion process on the archive regional risk coding and the non-uniform risk perception environment Patch sequence, including at least one of vector concatenation, position superposition and linear mapping, to generate a regional enhanced Patch feature sequence.

[0084] A multi-source channel coupled attention layer is used to perform cross-parameter coupled modulation on different risk feature channels in the region enhancement patch feature sequence to generate a multi-source coupled patch feature sequence.

[0085] The Transformer temporal coding layer is used to capture the continuous patch dependencies and abrupt patch response relationships in the multi-source coupled patch feature sequence through a multi-head attention mechanism, and generate temporal coding features.

[0086] An irreversible damage accumulation memory layer is used to perform cumulative writing and retention updates on historical damage contributions in temporal coding features through gated recurrent units and residual connection compensation, generating an archive damage accumulation feature sequence.

[0087] The archive damage risk prediction head is used to perform non-negative incremental updates on the cumulative feature sequence of archive damage, constraining the cumulative damage risk value to exhibit a monotonically non-decreasing characteristic on the time axis, and generating archive damage risk prediction results.

[0088] In this embodiment, the non-uniform risk perception patch partitioning layer includes the following operations:

[0089] The rate of change, sustained deviation, and magnitude of abrupt change in the archival environmental risk characteristic sequence are calculated based on the collection time identifier. The preset risk level table is read according to the archival material type and storage area code to determine the risk sensitivity level.

[0090] A preset risk level table is used to establish the correspondence between archival material types, storage area codes, and risk sensitivity levels. The preset risk level table includes at least the following fields: an archival material type field, used to record paper archives, photographic archives, film archives, magnetic media archives, and electronic media archives; a storage area code field, used to record archive shelf areas, cabinet levels, wall-adjacent areas, ventilation zones, and areas covered by environmental control equipment; and a risk sensitivity level field, used to assign a first risk level, a second risk level, or a third risk level based on the environmental sensitivity corresponding to the archival material type and storage area code.

[0091] The rate of change, the amount of continuous deviation, the magnitude of the mutation, and the risk sensitivity level are input into the sensitivity mapping algorithm, and the judgment is performed in the order of the first sampling segment, the second sampling segment, and the third sampling segment.

[0092] When at least one of the following conditions is met—the rate of change exceeds the first rate of change threshold, the amount of continuous deviation exceeds the first amount of deviation threshold, the magnitude of mutation exceeds the first magnitude of mutation threshold, or the risk sensitivity level reaches the first risk level—it is marked as the first sampling segment, and the first patch length, first overlap ratio, and first sliding step size are configured.

[0093] When a region is not marked as the first sampling segment, and at least one of the following conditions is met: the rate of change exceeds the second rate of change threshold, the amount of continuous deviation exceeds the second deviation threshold, the magnitude of mutation exceeds the second magnitude threshold, or the risk sensitivity level reaches the second risk level, it is marked as the second sampling segment, and the second patch length, the second overlap ratio, and the second sliding step size are configured.

[0094] When it is not marked as the first sampling segment or the second sampling segment, it is marked as the third sampling segment, and the third patch length, third overlap ratio and third sliding step size are configured.

[0095] Time-domain slicing is performed on the first, second, and third sampling segments according to their corresponding patch length, overlap ratio, and sliding step size, generating a first patch sequence, a second patch sequence, and a third patch sequence, respectively. The first patch length is shorter than the second patch length, the second patch length is shorter than the third patch length, the first overlap ratio is greater than the second overlap ratio, the second overlap ratio is greater than the third overlap ratio, the first sliding step size is shorter than the second sliding step size, and the second sliding step size is shorter than the third sliding step size. Furthermore, the first rate of change threshold is greater than the second rate of change threshold, the first deviation threshold is greater than the second deviation threshold, the first mutation amplitude threshold is greater than the second mutation amplitude threshold, and the risk sensitivity corresponding to the first risk level is higher than that of the second risk level, and the risk sensitivity corresponding to the second risk level is higher than that of the third risk level.

[0096] The first, second, and third patch sequences are merged according to the acquisition time identifier to generate a non-uniform risk perception environment patch sequence.

[0097] In one embodiment, the sensitivity mapping algorithm is used to divide the collection period into a first sampling segment, a second sampling segment, and a third sampling segment based on the rate of change, sustained deviation, abrupt change magnitude, and risk sensitivity level of the archival environmental risk characteristic sequence, and to determine the patch length, overlap ratio, and sliding step size, respectively. The algorithm maps the sampling segment attributes by looking up a table or using a piecewise function based on the comparison results between the input features and preset thresholds, achieving high-density sampling in high-risk time periods and low-density sampling in low-risk time periods, as detailed in Table 1 below:

[0098] Table 1: Data Collection Time Period Table

[0099]

[0100] For example, when the humidity risk characteristic change rate Δ=22% in a certain storage area, the sensitivity mapping algorithm marks the corresponding time period as the first sampling segment, configuring a patch length of 16, an overlap ratio of 50%, and a sliding step size of 8; when the humidity risk characteristic change rate Δ=3% in another storage area, the continuous deviation is 2%, the mutation amplitude does not exceed the threshold, and the risk sensitivity level is the third risk level, the algorithm marks the corresponding time period as the third sampling segment, configuring a patch length of 64, an overlap ratio of 0%, and a sliding step size of 64.

[0101] In this embodiment, the regional risk coding embedding layer includes the following operations:

[0102] Generate archive area risk codes based on the type of archive materials and the code of the storage area;

[0103] The archive area risk code and the non-uniform risk perception environment Patch sequence are fused together using at least one of the following methods: vector concatenation, position overlay and linear mapping, to generate an area-enhanced Patch feature sequence.

[0104] The document type is combined with the spatial characteristics of the storage area to mark the Patch location, and each Patch contains spatial risk information for the corresponding storage area.

[0105] In this embodiment, vector concatenation refers to directly connecting the risk coding vector of the archive area with the feature vector of the corresponding Patch in the dimension to form an extended feature vector;

[0106] Location overlay refers to adding a spatial location vector of the archive region risk code to each Patch feature vector, and embedding regional spatial information into time series features by adding elements one by one;

[0107] Linear mapping refers to projecting the archive region risk encoding vector and the Patch feature vector onto the same dimension through a trainable linear transformation, generating a fused region-enhanced Patch feature vector, ensuring that the input dimension of the attention layer is consistent.

[0108] At least one of the three fusion methods should be used to ensure that the Patch feature sequence contains spatial risk information and to achieve differentiated representation of the same type of risk in different storage areas.

[0109] In this embodiment, the multi-source channel coupled attention layer includes the following operations:

[0110] The receiving area is enhanced with Patch feature sequences and classified according to risk feature channels. Risk feature channels refer to different risk category features in the composite risk feature sequence of the archive environment, including temperature and humidity coupling risk features, light and heat superposition risk features, smoke and gas association risk features, pest activity risk features, access control disturbance risk features, and equipment adjustment failure risk features. Each category of feature is independently input as a channel in the Patch sequence.

[0111] Calculate the correlation coefficient matrix between each risk feature channel to generate a cross-parameter correlation matrix; for each type of risk feature channel vector in each Patch, arrange them according to the collection time identifier, calculate the Pearson correlation coefficient between each pair of channels, and form a channel-to-channel correlation coefficient matrix; each element in the matrix represents the degree of linear correlation between the corresponding two risk feature channels in the current Patch, quantifying the coupling strength between channels.

[0112] Based on the cross-parameter correlation matrix, weighted modulation is performed on each risk feature channel. Specifically, for each risk feature channel, the row vector corresponding to the channel in the cross-parameter correlation matrix is ​​used as the weight to perform a weighted linear combination on the feature values ​​of other channels to generate the coupling enhancement feature of that channel. The relevant channel features are linearly combined according to the coupling weight to generate the coupling enhancement feature, and arranged in the order of the acquisition time identifier to form a multi-source coupling patch feature sequence.

[0113] In this embodiment, in the Transformer temporal coding layer, a multi-source coupled patch feature sequence is input and arranged according to the acquisition time identifier to form a temporal patch sequence;

[0114] Continuous Patch Dependency Capture: Through a multi-head attention mechanism, the feature vector of each Patch in the Patch sequence is used as the query, key, and value. The temporal dependencies between adjacent Patches in the sequence are globally weighted and summed to generate a continuous dependency encoding for each Patch, reflecting the temporal continuity information within the sequence.

[0115] Mutation Patch Response Relationship Capture: In different attention heads of multi-head attention, the patch position encoding gain is used to enhance the features of mutation patches, so that patches that undergo sudden feature changes in the sequence are given higher weights, thereby capturing the response of sudden environmental risk signals to the current patch.

[0116] The sequential dependency encoding and mutation response encoding are concatenated or weighted and fused at the Patch level to form the temporal encoding features of each Patch.

[0117] In this embodiment, the irreversible damage accumulation memory layer includes the following operations:

[0118] Receive the timing coding features output by the Transformer timing coding layer according to the acquisition time identifier;

[0119] The temporal coding features of the current patch are added one by one to the archive damage cumulative feature vector corresponding to the previous acquisition time identifier to generate the candidate cumulative feature vector of the current patch.

[0120] Among them, the archive damage accumulation feature corresponding to the previous collection time identifier refers to the damage state feature accumulated by the irreversible damage accumulation memory layer based on the temporal coding feature of the previous Patch before processing the current Patch; the archive damage accumulation feature is used to record the historical damage contribution formed by continuous humidity deviation, cumulative light exposure, pest activity, abnormal smoke and gas, access control disturbance and equipment adjustment failure during the previous collection time period, as the input for the current Patch's cumulative writing and retention update;

[0121] Perform a non-negative incremental update on the candidate cumulative feature vector of the current patch, with the cumulative increment restricted to a non-negative value. Add the cumulative increment of the current patch to the cumulative feature value of the archive damage corresponding to the previous acquisition time identifier to generate the archive damage cumulative feature of the current patch.

[0122] Based on the collection time identifier, the cumulative archival damage features of each patch are combined to generate a cumulative archival damage feature sequence.

[0123] In this embodiment, the cumulative feature sequence of archive damage is input into the archive damage risk prediction header, and the cumulative features of each patch are processed in the order of collection time identifier.

[0124] Perform non-negative incremental update by subtracting the cumulative feature value corresponding to the previous acquisition time marker from the cumulative feature vector of the current patch's archive damage element by element to obtain the damage increment of the current patch, and setting all negative values ​​to zero to obtain the non-negative increment.

[0125] The cumulative damage risk value of the current patch is obtained by adding the non-negative increment to the cumulative feature value of the archive damage corresponding to the previous acquisition time marker element by element.

[0126] The cumulative damage risk value of the constraint exhibits a monotonically non-decreasing characteristic on the time axis. For each patch's cumulative damage risk value obtained in chronological order, check whether it is less than the previous patch's cumulative value. If it is less, replace it with the previous patch's value to ensure that each cumulative value in the time series is not less than the previous time step.

[0127] The processed cumulative damage risk values ​​are output in the order of collection time to form the final archive damage risk prediction result.

[0128] In this embodiment, the process of generating archive environment safety early warning results in the graded early warning alarm output module includes:

[0129] Receive the archive damage cumulative feature sequence according to the collection time identifier;

[0130] Subtract the feature vector corresponding to the current acquisition time marker from the feature vector of the previous time step element by element, set negative values ​​to zero, and generate a non-negative increment;

[0131] The non-negative increment is added element by element to the feature vector of the previous time step to generate the current cumulative damage risk value; the cumulative value is checked in chronological order, and if it is less than the previous time step value, it is replaced with the previous time step value to form a monotonically non-decreasing sequence.

[0132] Based on the processed cumulative values, an early warning result for the archive environment security is generated;

[0133] Specifically, risk classification determination: compare the cumulative damage risk value corresponding to the current patch with the preset risk threshold table;

[0134] If the cumulative value is greater than or equal to the high-risk threshold, it is marked as high-risk;

[0135] If the cumulative value is greater than or equal to the medium-risk threshold and less than the high-risk threshold, it is marked as medium-risk.

[0136] If the cumulative value is less than the medium-risk threshold, it is marked as low-risk;

[0137] Region association: Map the storage region code corresponding to the cumulative damage risk value of each patch to the archive storage region to generate risk level information classified by region;

[0138] Time stamp binding: The collection time stamp of each patch is bound to the corresponding risk level to form a time series risk report;

[0139] Results integration: Combine the regional risk levels and time series information of all patches to form an early warning result for archival environmental safety, including risk level, risk area and collection time identifier.

[0140] The warning results are converted into tiered alarm commands and sent to audible and visual alarm devices, management terminals, mobile receivers, access control controllers, environmental control equipment, and fire alarm linkage interfaces.

[0141] Specifically, the operations for converting to a tiered alarm command include the following:

[0142] Risk level mapping: For each patch or storage area, the archive environment security warning results are mapped to their risk level (high risk, medium risk, low risk) to the corresponding preset alarm level;

[0143] High risk is mapped to a red alarm command, medium risk to a yellow alarm command, and low risk to a green alarm command.

[0144] Command format generation: Generates a command format that the target device can recognize based on the target device type (audio-visual alarm device, management terminal, mobile receiver, access control controller, environmental control equipment, fire linkage interface);

[0145] The command content includes: target device identifier, corresponding patch or storage area code, risk level identifier, alarm color or level, and trigger time;

[0146] Device distribution: The generated hierarchical alarm commands are sent to each target device according to the device identifier to achieve information synchronization;

[0147] Each instruction is guaranteed to be consistent with the collection time identifier and area code of the corresponding patch, ensuring that the alarm corresponds to the actual risk.

[0148] Example 1: To verify the feasibility of the present invention in practice, the present invention was applied to the environmental security management of a large archive. The archive contains paper archives and microfilms of different materials. The environment is complex and has problems such as temperature and humidity fluctuations, light interference, microbial pest activity, and failure of air conditioning and ventilation equipment. Traditional fixed threshold alarms cannot identify potential risks in a timely manner, nor can they provide graded early warnings for different storage areas.

[0149] By applying the intelligent early warning system of this invention, the archive storage environment can be comprehensively monitored and dynamically analyzed, enabling accurate prediction of environmental risks and hierarchical alarm management.

[0150] In the archive, the environmental monitoring signal acquisition module first collects multi-source signals such as temperature, humidity, light intensity, smoke concentration, dust concentration, volatile gas concentration, pest monitoring, vibration, access control opening and closing, and the operating status of environmental control equipment. Each signal is marked with an acquisition time identifier, sensor number, and storage area code to form complete raw environmental monitoring data. After the acquisition is completed, the environmental monitoring signal preprocessing module performs data time synchronization, anomaly removal, missing value completion, unit normalization, and area code binding to generate standardized basic environmental monitoring signals for the archive, ensuring the accuracy and consistency of the analysis.

[0151] The environmental risk feature generation module extracts risks such as temperature and humidity coupling risk, light and heat superposition risk, smoke and gas correlation risk, pest activity risk, access control disturbance risk, and equipment adjustment failure risk from the processed base signal. By accumulating and calculating the rate of change, continuous deviation, and abrupt change amplitude, it forms an archival environmental risk feature sequence that can reflect the potential impact of environmental fluctuations on archives. The non-uniform risk perception patch segmentation module divides the risk feature sequence into first, second, and third sampling segments, and configures the patch length, overlap ratio, and sliding step size according to a preset risk level table and sensitivity mapping algorithm to achieve dense analysis of high-risk segments and sparse sampling of low-risk segments. The segmented patch sequence is further input into the regional risk coding embedding module, which embeds storage area code, cabinet level code, wall proximity identification, ventilation zoning, and equipment coverage information into the patch features to generate a regional enhanced patch feature sequence, achieving effective fusion of environmental spatial information and risk features.

[0152] The region-enhanced patch feature sequence enters the multi-source channel coupled attention layer. By calculating the correlation coefficient matrix of different risk feature channels and performing weighted modulation, a multi-source coupled patch feature sequence is formed, ensuring that the interaction effects between different environmental parameters can be fully captured. The Transformer temporal coding layer performs long-period trend coding and short-time perturbation coding on the multi-source coupled patch feature sequence. Through a multi-head attention mechanism, it captures the continuous patch dependency relationship and the abrupt patch response relationship respectively, generating temporal coded features. The irreversible damage accumulation memory layer accumulates and retains the historical damage contribution in the temporal coded features. Through non-negative incremental updates, it ensures that the accumulated value exhibits a monotonically non-decreasing feature on the time axis, forming an archive damage accumulation feature sequence.

[0153] The archive damage risk prediction head generates archive environmental safety early warning results using cumulative feature sequences. It then generates alarm commands corresponding to high, medium, and low risk levels through a tiered early warning alarm output module, sending these commands to audible and visual alarm devices, management terminals, mobile receivers, access control controllers, environmental control equipment, and fire linkage interfaces to achieve real-time early warning and tiered management. In practical applications, the system triggers a red alarm when temperature and humidity fluctuations reach ±3℃ and ±10%RH in high-risk areas, a yellow alarm when fluctuations reach ±2℃ and ±5%RH in medium-risk areas, and a green alarm when fluctuations reach ±1℃ and ±2%RH in low-risk areas. When smoke concentration suddenly increases by more than 5ppm and volatile gas concentration exceeds 15ppm, the high-risk alarm trigger rate reaches 98%, the medium-risk alarm trigger rate is 91%, and the false alarm rate in low-risk areas is less than 3%, verifying the system's high sensitivity and accuracy to abnormal environmental changes. Compared to equal-length patch analysis, the non-uniform patch division strategy shortens the detection time for high-risk sections by approximately 40%, improving the efficiency of environmental risk identification. After applying this invention, the overall environmental security management accuracy and response speed of the archives are improved. It can implement differentiated and hierarchical early warning for archives in different regions and with different material types, providing scientific, efficient and intelligent security for long-term archive preservation; specific data are shown in Table 2 below.

[0154] Table 2: Comparison of Archival Environmental Risk Monitoring and Early Warning

[0155]

[0156] Table 2 illustrates the implementation effect of the archival environmental risk monitoring and graded early warning system. By comparing environmental data from different storage areas and material types, the advantages and practical application value of this invention can be intuitively analyzed. The five storage areas listed in the table cover two common archival material types: paper archives and microfilm. Environmental parameters such as temperature and humidity fluctuations, smoke concentration, and volatile gas concentration are key risk indicators used to assess the safety of the archival storage environment. Analysis of temperature and humidity fluctuations reveals that high-risk areas have relatively large temperature and humidity fluctuations. For example, area A has temperature and humidity fluctuations of ±3℃ and ±10%RH, while area E has ±3℃ and ±8%. These environmental conditions accelerate the acidification and embrittlement process of paper archives and also increase the long-term degradation risk of microfilm. Medium-risk areas have relatively moderate temperature and humidity fluctuations. For example, areas B and D have temperature and humidity fluctuations between ±2℃ and ±5~6%RH. Low-risk areas have the smallest fluctuations. For example, area C has temperature and humidity fluctuations of ±1℃ and ±2%RH. This indicates that the system can automatically identify different risk levels and classify and manage them according to environmental changes.

[0157] Smoke concentration and volatile gas concentration are important indicators of air quality and potential fire or chemical damage. The table shows that smoke concentrations in high-risk areas reach 5-6 ppm and volatile gas concentrations are 15-16 ppm; smoke concentrations in medium-risk areas are 3-4 ppm and volatile gas concentrations are 12-14 ppm; and the indicators for low-risk areas are lower than those for medium- and high-risk areas. This demonstrates that this invention can integrate multi-source environmental signals, capture abnormal factors in the air that may damage archives, and determine the risk level through cumulative calculation and sensitivity analysis.

[0158] The risk level column intuitively reflects the system's classification capability. High-risk areas are marked with a red alert, medium-risk areas with a yellow alert, and low-risk areas with a green alert, which is highly consistent with changes in environmental data. Alarm trigger rate data further verifies the system's high reliability and sensitivity. The alarm trigger rate reaches 96-98% in high-risk areas, 90-91% in medium-risk areas, and the false alarm rate in low-risk areas is less than 3%, indicating that the system can avoid false alarms in low-risk areas while possessing rapid response capabilities in high-risk environments.

[0159] Overall, this table fully demonstrates the technical advantages of this invention: the system can achieve fusion analysis of multi-source environmental data, dynamically identify different risk areas, and perform non-uniform patch sampling and cumulative risk calculation, realizing a comprehensive assessment of environmental fluctuations, emergencies, and long-term cumulative damage. Through hierarchical alarm output, this invention can distribute risk information in real time to audible and visual alarm devices, management terminals, mobile receivers, access control controllers, environmental control equipment, and fire linkage interfaces, achieving intelligent and refined archival environmental security management and providing a guarantee for the long-term preservation of archives.

[0160] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent early warning system for the security of archive storage environments, characterized in that, include: The environmental monitoring signal acquisition module is used to acquire multi-source monitoring signals of the archived environment, add acquisition time identifiers, sensor numbers and storage area codes, and generate raw environmental monitoring signals. The environmental monitoring signal preprocessing module is used to preprocess the raw environmental monitoring signals to generate archival environmental monitoring base signals. The environmental risk feature generation module is used to extract composite risk features of the archive environment based on basic signals from archive environmental monitoring and generate a sequence of archive environmental risk features. The non-uniform patch partitioning module is used to input the archival environment risk feature sequence into the improved archival environment risk perception patchTST model. Based on the rate of change, the amount of continuous deviation, the magnitude of mutation and the risk sensitivity level, the module calculates the patch length, the overlap ratio and the sliding step size through the sensitivity mapping algorithm to generate a non-uniform risk perception environment patch sequence. The regional risk embedding module is used to construct the archive regional risk code and embed the non-uniform risk perception environment Patch sequence to generate the regional enhanced Patch feature sequence. The damage accumulation determination module is used to perform cross-parameter coupling modulation, long-period trend encoding, short-time perturbation encoding and historical damage contribution accumulation retention on the region enhancement patch feature sequence to generate an archive damage accumulation feature sequence. The graded early warning and alarm output module is used to perform non-negative incremental updates and time monotonically non-decreasing constraints on the cumulative characteristic sequence of archive damage, generate early warning results for archive environmental safety, and send graded alarm commands.

2. The intelligent early warning system for the security of archive storage environment according to claim 1, characterized in that, The preprocessing process in the environmental monitoring signal preprocessing module includes: sorting the original environmental monitoring signals according to the acquisition time identifier, completing missing values, removing outliers and normalizing dimensions, binding storage area codes, and generating archived environmental monitoring basic signals.

3. The intelligent early warning system for the security of archive storage environment according to claim 1, characterized in that, The process of extracting composite environmental risk features from archives in the environmental risk feature generation module includes: Read the basic environmental monitoring signals from the archives according to the storage area code and the collection time identifier, and calculate the product of temperature change and humidity change to generate temperature and humidity coupling risk characteristics; The product of the cumulative illumination and the temperature change within a preset sliding window period is used to generate a photothermal superposition risk characteristic. The cross-correlation coefficient between changes in smoke concentration and changes in volatile gas concentration is calculated using a cross-correlation function to generate smoke gas associated risk characteristics. The number of pest triggers and the interval between adjacent triggers within a preset sliding window period are statistically analyzed to generate pest activity risk characteristics. The frequency of access control opening and closing within a preset sliding window period and the temperature and humidity fluctuations after the access control opening and closing are used to generate access control disturbance risk characteristics. The differences in humidity and temperature before and after the operation of the environmental control equipment are calculated to generate the risk characteristics of equipment control failure. Sort and concatenate the data according to the collection time identifier to generate an archive environmental risk characteristic sequence.

4. The intelligent early warning system for the security of archive storage environment according to claim 1, characterized in that, The structure of the improved archival environment risk perception PatchTST model includes: The risk feature input layer is used to receive the risk feature sequence of the archive environment, which is arranged as the model input sequence according to the storage area code and the collection time identifier. The non-uniform risk perception patch partitioning layer is used to calculate the patch length, overlap ratio and sliding step size based on the rate of change, continuous deviation, mutation magnitude and risk sensitivity level in the model input sequence, and to perform non-uniform temporal sampling on the model input sequence to generate a non-uniform risk perception environment patch sequence. The regional risk coding embedding layer is used to perform a fusion process on the archive regional risk coding and the non-uniform risk perception environment Patch sequence, including at least one of vector concatenation, position superposition and linear mapping, to generate a regional enhanced Patch feature sequence. A multi-source channel coupled attention layer is used to perform cross-parameter coupled modulation on different risk feature channels in the region enhancement patch feature sequence to generate a multi-source coupled patch feature sequence. The Transformer temporal coding layer is used to capture the continuous patch dependencies and abrupt patch response relationships in the multi-source coupled patch feature sequence through a multi-head attention mechanism, and generate temporal coding features. An irreversible damage accumulation memory layer is used to perform cumulative writing and retention updates on historical damage contributions in temporal coding features through gated recurrent units and residual connection compensation, generating an archive damage accumulation feature sequence. The archive damage risk prediction head is used to perform non-negative incremental updates on the cumulative feature sequence of archive damage, constraining the cumulative damage risk value to exhibit a monotonically non-decreasing characteristic on the time axis, and generating archive damage risk prediction results.

5. The intelligent early warning system for the security of archive storage environment according to claim 4, characterized in that, The non-uniform risk perception patch partitioning layer includes the following operations: The rate of change, sustained deviation, and magnitude of abrupt change in the archival environmental risk characteristic sequence are calculated based on the collection time identifier. The preset risk level table is read according to the archival material type and storage area code to determine the risk sensitivity level. The rate of change, the amount of continuous deviation, the magnitude of the mutation, and the risk sensitivity level are input into the sensitivity mapping algorithm, and the judgment is performed in the order of the first sampling segment, the second sampling segment, and the third sampling segment. When at least one of the following conditions is met—the rate of change exceeds the first rate of change threshold, the amount of continuous deviation exceeds the first amount of deviation threshold, the magnitude of mutation exceeds the first magnitude of mutation threshold, or the risk sensitivity level reaches the first risk level—it is marked as the first sampling segment, and the first patch length, first overlap ratio, and first sliding step size are configured. When a region is not marked as the first sampling segment, and at least one of the following conditions is met: the rate of change exceeds the second rate of change threshold, the amount of continuous deviation exceeds the second deviation threshold, the magnitude of mutation exceeds the second magnitude threshold, or the risk sensitivity level reaches the second risk level, it is marked as the second sampling segment, and the second patch length, the second overlap ratio, and the second sliding step size are configured. When it is not marked as the first sampling segment or the second sampling segment, it is marked as the third sampling segment, and the third patch length, third overlap ratio and third sliding step size are configured. The first sampling segment, the second sampling segment, and the third sampling segment are sliced ​​in the time domain according to the corresponding Patch length, overlap ratio, and sliding step size, respectively, to generate the first Patch sequence, the second Patch sequence, and the third Patch sequence. The length of the first Patch is less than the length of the second Patch, the length of the second Patch is less than the length of the third Patch, the first overlap ratio is greater than the second overlap ratio, the second overlap ratio is greater than the third overlap ratio, the first sliding step size is less than the second sliding step size, and the second sliding step size is less than the third sliding step size. The first, second, and third patch sequences are merged according to the acquisition time identifier to generate a non-uniform risk perception environment patch sequence.

6. The intelligent early warning system for archive storage environment security according to claim 4, characterized in that, The regional risk coding embedding layer includes the following operations: Generate archive area risk codes based on the type of archive materials and the code of the storage area; The archive area risk code and the non-uniform risk perception environment Patch sequence are fused together using at least one of the following methods: vector concatenation, position overlay and linear mapping, to generate an area-enhanced Patch feature sequence. The document type is combined with the spatial characteristics of the storage area to mark the Patch location, and each Patch contains spatial risk information for the corresponding storage area.

7. The intelligent early warning system for the security of archive storage environment according to claim 4, characterized in that, The multi-source channel coupled attention layer includes the following operations: The received region is enhanced with Patch feature sequences and classified according to risk feature channels; Calculate the correlation coefficient matrix between each risk characteristic channel to generate a cross-parameter correlation matrix; Based on the cross-parameter correlation matrix, weighted modulation is performed on each risk feature channel, and the relevant channel features are linearly combined according to the coupling weight to generate coupling enhancement features. These features are then arranged in order of acquisition time to form a multi-source coupling patch feature sequence.

8. The intelligent early warning system for the security of archive storage environment according to claim 4, characterized in that, The irreversible damage accumulation memory layer includes the following operations: Receive the timing coding features output by the Transformer timing coding layer according to the acquisition time identifier; The temporal coding features of the current patch are added one by one to the archive damage cumulative feature vector corresponding to the previous acquisition time identifier to generate the candidate cumulative feature vector of the current patch. Perform a non-negative incremental update on the candidate cumulative feature vector of the current patch, with the cumulative increment restricted to a non-negative value. Add the cumulative increment of the current patch to the cumulative feature value of the archive damage corresponding to the previous acquisition time identifier to generate the archive damage cumulative feature of the current patch. Based on the collection time identifier, the cumulative archival damage features of each patch are combined to generate a cumulative archival damage feature sequence.

9. The intelligent early warning system for the security of archive storage environment according to claim 1, characterized in that, The process of generating archive environment safety early warning results in the hierarchical early warning alarm output module includes: Receive the archive damage cumulative feature sequence according to the collection time identifier; Subtract the feature vector corresponding to the current acquisition time marker from the feature vector of the previous time step element by element, set negative values ​​to zero, and generate a non-negative increment; The non-negative increment is added element by element to the feature vector of the previous time step to generate the current cumulative damage risk value; the cumulative value is checked in chronological order, and if it is less than the previous time step value, it is replaced with the previous time step value to form a monotonically non-decreasing sequence. Based on the processed cumulative values, an early warning result for the archive environment security is generated; The warning results are converted into tiered alarm commands and sent to audible and visual alarm devices, management terminals, mobile receivers, access control controllers, environmental control equipment, and fire alarm linkage interfaces.

Images