An intelligent lock recognition method based on a time sequence diagram neural network

Abstract

The application discloses an intelligent lock recognition method based on a time sequence graph neural network, which comprises the following steps: collecting multi-source event data of the intelligent lock, uniformly formatting and embedding codes, and generating an event feature vector set; constructing a time-varying multi-relation graph, performing message transmission and feature aggregation, and obtaining a neighborhood aggregation result; constructing a double-frequency memory time sequence graph neural network, extracting long-term and short-term behavior features in slow-frequency and fast-frequency memory units; inputting the slow-frequency and fast-frequency memory states into a gate fusion layer, combining time and scene information to generate a full-time domain embedding vector; calculating identity matching, abnormal risk and forgery credibility, and outputting a comprehensive recognition result; performing real-time response and safety disposal according to the recognition result, and updating the network. The application utilizes the time sequence graph neural network and the double-frequency memory mechanism, realizes dynamic recognition of multiple elements of the intelligent lock, and has the advantages of strong self-learning, high recognition accuracy and high safety robustness.

Description

Technical Field

[0001] This invention relates to the field of security identification, and in particular to a smart lock identification method based on a time-series graph neural network. Background Technology

[0002] Existing smart lock recognition technologies primarily rely on static authentication, typically depending on single features such as fingerprints, passwords, facial recognition, or Bluetooth signals for matching and judgment. This approach suffers from significant shortcomings in security and stability. Some improvement solutions attempt to introduce data modeling based on time series or device behavior, achieving dynamic recognition through deep learning or temporal neural networks. However, these methods often only model single-dimensional input data, lacking a joint characterization of the complex interactions between multiple factors such as users, devices, methods, time, and space. This results in poor model adaptability to environmental changes and multi-user scenarios. Furthermore, existing solutions generally rely on centralized cloud processing, making it difficult to respond promptly to real-time events at the edge, leading to high recognition latency.

[0003] In the field of dynamic behavior modeling and anomaly recognition, existing research has proposed combining graph neural networks and long short-term memory networks to capture temporal dependencies or periodic features between nodes. However, these methods generally employ a single-frequency time update mechanism, which can only reflect a certain time scale of behavioral changes and cannot simultaneously take into account short-term fluctuations and long-term patterns. For high-frequency interactive devices such as smart locks, existing models struggle to balance rapidly changing unlocking events with long-term user habits, leading to frequent misjudgments in anomaly detection. Furthermore, the lack of a mechanism for real-time feedback at the edge and collaborative updating of the cloud-based behavioral baseline prevents the model from developing self-learning and cross-period evolution capabilities, limiting its reliability and robustness in dynamic environments with multiple scenarios and multiple users.

[0004] Therefore, how to provide a smart lock recognition method based on temporal graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a smart lock recognition method based on a temporal graph neural network. This invention utilizes a temporal graph neural network and a dual-frequency memory mechanism to achieve dynamic recognition of multiple elements of a smart lock, and has the advantages of strong self-learning, high recognition accuracy, and excellent security robustness.

[0006] A smart lock recognition method based on a temporal graph neural network according to an embodiment of the present invention includes the following steps:

[0007] Acquire multi-source event data during the operation of the smart lock, perform format unification and embedding encoding, and obtain a set of event feature vectors;

[0008] A time-varying multi-relationship graph structure is constructed based on the event feature vector set. User nodes, device nodes, mode nodes, time nodes and spatial nodes are associated. Graph message passing and neighborhood feature aggregation are performed to obtain the neighborhood aggregation result of the node at the current time.

[0009] A dual-frequency memory time-series graph neural network is constructed. The neighborhood aggregation result is input into the dual-frequency memory layer, and the memory state is updated in the slow-frequency memory unit and the fast-frequency memory unit respectively. The slow-frequency memory state and the fast-frequency memory state are output.

[0010] The slow-frequency memory state and the fast-frequency memory state are input into the gating fusion layer, and adaptive fusion is performed through the gating function based on the time context and scene information to generate a full-time domain node embedding vector.

[0011] The input nodes of the full-time domain are embedded into the output layer and mapped to low-dimensional node representation vectors. The identity matching result, the abnormal risk result and the forgery credibility result are calculated through the identification and anomaly judgment layer, and the comprehensive identification decision is output.

[0012] Based on comprehensive identification decisions, immediate responses and security measures are executed at the edge, and slow-frequency memory units and behavioral baselines are updated in the cloud to form an updated dual-frequency memory time-series neural network.

[0013] Optionally, the generation of the event feature vector set specifically includes:

[0014] Receive multi-source event data during the operation of the smart lock, and reorganize user identifier, device identifier, unlocking method, timestamp, spatial location and communication signal elements into a structured event element sequence according to a unified field order;

[0015] Discrete mapping and sparse coding are performed on the user identifier, device identifier, and unlocking method in the structured event element sequence to obtain the identity method coding result;

[0016] The timestamps in the structured event element sequence are decomposed into time, expanded into cycles, and mapped into phases. The timestamps are then transformed into time features containing information on intraday time periods, intraweek locations, rhythmic phases, and interval lengths, resulting in time coding results.

[0017] The spatial location in the structured event element sequence is analyzed by coordinates and quantized by region. The latitude and longitude and door location are reduced to the spatial characteristics of the floor, room, door type and the relationship with adjacent areas to obtain the spatial coding result.

[0018] Amplitude normalization, time alignment, and short sequence aggregation are performed on the communication signal elements in the structured event element sequence. Physical side statistics such as signal strength variation amplitude, stable interval length, rise and fall edge duration, pulse count, and jitter measurement are extracted to obtain the communication signal encoding results.

[0019] The identity encoding result, time encoding result, spatial encoding result and communication signal encoding result are concatenated and scale-aligned to form a single event vector, thus obtaining the event embedding result;

[0020] The format and storage layout of the event embedding results are standardized, the field order, numerical range and storage alignment are determined, and the results are organized into a set of event feature vectors based on the timestamp sequence.

[0021] Optionally, the generation of the neighborhood aggregation result specifically includes:

[0022] Based on the event feature vector set, an initial node list is established for user nodes, device nodes, mode nodes, time nodes, and space nodes, and a corresponding index is generated for each event;

[0023] Based on the initial node list, according to the co-occurrence relationship of elements of the same event, the interaction relationship between user nodes and device nodes, the selection relationship between user nodes and mode nodes, the occurrence relationship between device nodes and time nodes, the positioning relationship between device nodes and spatial nodes, and the time period relationship between mode nodes and time nodes are generated. The timestamp order, attempt number, success or failure flag and communication signal statistics of each relationship are recorded to obtain the relationship instance table.

[0024] The relation instance table is divided into pieces according to the timestamp and the continuous sliding time window to form a relation piece sequence organized by window number; within each window, relation instances are summarized according to the co-occurrence relation type of the elements to obtain a time-varying multi-relationship graph structure containing user nodes, device nodes, mode nodes, time nodes and space nodes.

[0025] Within each window, a relationship weight and a time-series attenuation coefficient are calculated for each type of element co-occurrence relationship. The relationship weight is determined based on the weighted result of the relationship occurrence frequency, communication signal stability, and time proximity. The time-series attenuation coefficient is calculated based on the time interval between the relationship and the end of the window. The two are multiplied to obtain the effective relationship weight and written into the relationship weight table.

[0026] Based on the relation weight table, an adjacency index and batch routing are established for the co-occurrence relationship of each type of element. User nodes, device nodes, mode nodes, time nodes and space nodes are divided into several subgraphs that can be processed in parallel, and relation adjacency index and batch routing table are generated.

[0027] Within each window, graph messaging and relationship aggregation are performed according to the batch routing table. The neighbor contributions of each type of element co-occurrence relationship are weighted and summarized based on the relationship weight table to obtain the temporary node representation and message set. The temporary node representation is the result of weighted aggregation of the neighbor node features of each type of element co-occurrence relationship.

[0028] The temporary representations of nodes are merged and stabilized according to time windows, and the neighborhood aggregation result of the node at the current time is generated by using the mean of the window and the weighting strategy of the recent window.

[0029] Optionally, the generation of the slow-frequency memory state and the fast-frequency memory state specifically includes:

[0030] A dual-frequency memory time-series neural network is constructed, comprising a dual-frequency memory layer, a gated fusion layer, a node embedding output layer, and a recognition and anomaly determination layer. The dual-frequency memory layer includes a slow-frequency memory unit, a fast-frequency memory unit, a trigger manager, a memory read / write interface, and a state cache.

[0031] At the edge, a node-state index is established based on the neighborhood aggregation result. The node-state index is used to record the storage location, access path and last update time of each node in the slow frequency memory unit and the fast frequency memory unit.

[0032] Based on the node-state index, the slow frequency memory unit template and behavior baseline of the previous cycle are loaded from the cloud, and the fast frequency memory unit snapshot retained during the last processing is read from the edge to generate the initial slow frequency memory state and the initial fast frequency memory state.

[0033] Align the neighborhood aggregation results with the event feature vector set by timestamp and index, extract the embedded fields corresponding to user identifier, device identifier, unlocking method, timestamp, spatial location and communication signal elements, and generate a memory update message sequence in chronological order;

[0034] Based on the initial fast-frequency memory state, the memory update message sequence is sequentially written into the fast-frequency memory unit. The manager is triggered to perform gating write, forgetting, and state advancement operations when each message is written. The gating write filters the effective feature dimensions of the current event, the forgetting operation reduces the weight of outdated or noisy features, and the state advancement writes the immediate impact of the latest event on the node's behavior into the short-term state cache. After each memory update message is written, an intermediate fast-frequency memory state is generated and stored in chronological order to form an intermediate fast-frequency memory state sequence. The fast-frequency memory unit takes the middle state of the last item in the intermediate fast-frequency memory state sequence as the current fast-frequency memory state, which represents the short-term immediate behavior response of the node at the current moment. Finally, the current fast-frequency memory state and the intermediate fast-frequency memory state set are output.

[0035] Based on the initial slow-frequency memory state, when the trigger manager detects the end of the time window or a continuous shift in the behavior baseline, it collects all memory update messages and the corresponding fast-frequency memory state sequence within the window, performs window aggregation and smoothing, calculates the mean of stable features and trend shift within the window, and performs template fine-tuning and state correction in combination with historical slow-frequency memory states. The slow-frequency memory unit writes the stable behavior patterns and slow-changing features in the window aggregation result into the long-term template through an adaptive gating mechanism, updates the center distribution of the slow-frequency memory according to the trend shift, and obtains the current slow-frequency memory state corresponding to the window number.

[0036] Align and deduplicate the current fast-frequency memory state and the current slow-frequency memory state according to the node-state index, write them to the state cache to form a pair of outputs of fast-frequency memory state and slow-frequency memory state, and generate a state update timestamp and trigger flag for each pair of outputs.

[0037] Optionally, the generation of the full-time-domain node embedding vector specifically includes:

[0038] Receive the slow-frequency memory state and the fast-frequency memory state, as well as the corresponding state update timestamp and trigger flag. Combine the neighborhood aggregation results with the time context and scene information provided by the time node, method node, and space node to establish a fusion batch index and obtain the fusion input set of the gated fusion layer.

[0039] Perform scale alignment, dimension alignment, and missing completion on the slow-frequency and fast-frequency memory states in the fusion input set to generate an aligned set of memory state pairs.

[0040] Based on the state update timestamp and trigger marker, the time period attribute corresponding to the time node, the unlocking method attribute corresponding to the method node, and the area attribute corresponding to the spatial node, the time fluctuation index, behavior consistency score and signal disturbance intensity are calculated for the set according to the aligned memory state, and a gating prior element table is generated.

[0041] In the gated fusion layer, based on the gated prior feature table and the aligned memory state pair set, a gated coefficient is generated for each node, and the gated coefficient sequence is output and a corresponding relationship is established with the fusion batch index.

[0042] The set is weighted and combined node by node according to the aligned memory state based on the gating coefficient sequence, and then stabilized and constrained, including range pruning, anomaly suppression and time smoothing, to suppress abnormal fluctuations and generate candidate full-time domain node embedding vectors.

[0043] The consistency of the candidate full-time domain node embedding vector with the relation weight table and the window number of the neighborhood aggregation result is checked. For nodes that do not meet the consistency, a slow frequency priority or fast frequency priority backoff strategy is triggered according to the gating coefficient sequence, and the full-time domain node embedding vector is output.

[0044] Optionally, the generation of the comprehensive identification decision specifically includes:

[0045] The full-time domain nodes are embedded into the vector input nodes and the output layer is embedded into the vector input nodes. Dimensionality reduction mapping, range normalization and numerical pruning are performed according to the preset field order to generate a set of node representation vectors corresponding to the event index.

[0046] Based on the event index, user node representation, device node representation, mode node representation, time node representation and spatial node representation are extracted from the node representation vector set and aligned with the corresponding records in the event feature vector set to form an event alignment vector group;

[0047] The event alignment vector group is input into the identification and anomaly determination layer, which consists of an identity matching branch, an anomaly risk branch, a forgery credibility branch, and a fusion decision module. In the identity matching branch, the identity matching score is calculated based on the event alignment vector group, and the identity matching result and threshold determination label are output.

[0048] In the abnormal risk branch, short-term fluctuation indicators, failure retry density and cross-regional switching rate are extracted based on the event alignment vector group and the event feature vector set. The abnormal risk score is calculated by combining the user node representation and the device node representation, and the abnormal risk result and risk level label are output.

[0049] In the forgery credibility branch, the signal disturbance intensity, stable interval length conversion value and rise and fall edge duration ratio are extracted based on the communication signal encoding results in the event alignment vector group and event feature vector set. The forgery credibility score is calculated by combining the method node representation and spatial node representation, and the forgery credibility result and suspicious type label are output.

[0050] The identity matching score, abnormal risk score, and forgery credibility score are input into the fusion decision module to generate a comprehensive identification score, which is formed by the weighted results of the three branches. A comprehensive identification decision is generated based on the comprehensive identification score and a preset threshold table.

[0051] Optionally, the generation of the updated dual-frequency memory time-series neural network specifically includes:

[0052] The system receives a comprehensive identification decision, a handling level, and an explanatory tag. At the edge, it executes an immediate response based on the handling level. If the comprehensive identification decision is to allow, it allows the unlocking action and records the execution time. If the comprehensive identification decision is to perform secondary verification, it triggers the user identity multi-factor verification module and suspends the unlocking response. If the comprehensive identification decision is to deny, it blocks the unlocking request and generates a security event log.

[0053] After an immediate response, the edge device records the operation feedback results, including whether the unlocking action was successful or failed, whether the verification was passed or rejected, the integrity of the communication signal and the response delay of the device. The feedback results are then linked with the comprehensive identification decision to form a handling record table.

[0054] Align the handling record table with the node representation vector set, extract the behavior pattern fields of user nodes, device nodes, method nodes, time nodes and space nodes to form a set of behavior feedback features;

[0055] At the edge, fast learning and memory calibration are performed based on the behavioral feedback feature set. When there is a deviation between the feedback result and the comprehensive recognition decision, the fast frequency memory unit update operation is triggered to adjust the short-term state cache corresponding to the node, update the short-term behavior pattern and signal response law, recalculate the fast frequency memory state of the node, and output the updated fast frequency memory state.

[0056] The updated set of fast-frequency memory states and behavioral feedback features generated at the edge is packaged by node index and time window number and uploaded to the cloud memory synchronization module. Based on the time window number, the historical slow-frequency memory unit template and behavioral baseline are matched, and the feedback offset and trend change are calculated.

[0057] In the slow frequency memory cell, template updates and behavioral baseline corrections are performed based on the feedback offset and trend change. The changing trend of the offset features is written into the slow frequency memory cell template, abnormal perturbation samples are deleted and the slow frequency memory state is updated to form the slow frequency memory cell correction template for the current window.

[0058] The slow-frequency memory unit correction template and the updated fast-frequency memory state at the edge are reintegrated into a synchronization packet, which is then sent to the edge to replace the old state cache and fast-frequency memory unit snapshot, forming an updated dual-frequency memory timing graph neural network for the next smart lock recognition operation.

[0059] The beneficial effects of this invention are:

[0060] This invention constructs a smart lock recognition method based on a temporal graph neural network, achieving unified modeling and dynamic fusion of multiple elements such as user, device, method, time, and space. This effectively overcomes the shortcomings of existing recognition technologies in terms of data isolation and context fragmentation. By introducing a time-varying multi-relationship graph structure, this invention can capture the interaction patterns and relationship weight changes between multiple entities within a sliding time window, enabling the recognition model to possess a global spatiotemporal relational understanding capability. Furthermore, a dual-frequency memory temporal graph neural network enables parallel updates of fast-frequency and slow-frequency memories, corresponding to short-term immediate behavior and long-term stable habits, respectively. This maintains both recognition sensitivity and stability in usage scenarios with frequent interactions and periodic changes, thereby significantly improving the accuracy of identity recognition and anomaly detection.

[0061] Furthermore, the gating fusion layer designed in this invention generates gating coefficients based on temporal context, scene information, and signal perturbation features, adaptively adjusting the fusion ratio of slow-frequency and fast-frequency memory states to ensure the model maintains robustness and generalization under different unlocking environments. Through a multi-branch joint decision-making mechanism in the identification and anomaly judgment layers, identity matching results, anomaly risk levels, and forgery credibility can be output simultaneously in a single identification, achieving multi-dimensional security judgment. Combining edge-end real-time response with a collaborative update mechanism of cloud-based slow-frequency memory templates, this invention forms a closed-loop self-learning system, enabling the model to automatically optimize behavioral templates and baseline distribution in continuous identification cycles, thereby achieving fast, accurate, and continuously evolving security identification effects in complex, multi-user, and cross-scenario smart lock environments. 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 flowchart of a smart lock recognition method based on a temporal graph neural network proposed in this invention;

[0064] Figure 2 This is a schematic diagram of a dual-frequency memory time-series graph neural network for a smart lock recognition method based on a time-series graph neural network proposed in this invention. Detailed Implementation

[0065] 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.

[0066] refer to Figures 1-2 A smart lock recognition method based on a temporal graph neural network includes the following steps:

[0067] Acquire multi-source event data during the operation of the smart lock, perform format unification and embedding encoding, and obtain a set of event feature vectors;

[0068] A time-varying multi-relationship graph structure is constructed based on the event feature vector set. User nodes, device nodes, mode nodes, time nodes and spatial nodes are associated. Graph message passing and neighborhood feature aggregation are performed to obtain the neighborhood aggregation result of the node at the current time.

[0069] A dual-frequency memory time-series graph neural network is constructed. The neighborhood aggregation result is input into the dual-frequency memory layer, and the memory state is updated in the slow-frequency memory unit and the fast-frequency memory unit respectively. The slow-frequency memory state and the fast-frequency memory state are output.

[0070] The slow-frequency memory state and the fast-frequency memory state are input into the gating fusion layer, and adaptive fusion is performed through the gating function based on the time context and scene information to generate a full-time domain node embedding vector.

[0071] The input nodes of the full-time domain are embedded into the output layer and mapped to low-dimensional node representation vectors. The identity matching result, the abnormal risk result and the forgery credibility result are calculated through the identification and anomaly judgment layer, and the comprehensive identification decision is output.

[0072] Based on comprehensive identification decisions, immediate responses and security measures are executed at the edge, and slow-frequency memory units and behavioral baselines are updated in the cloud to form an updated dual-frequency memory time-series neural network.

[0073] In this embodiment, the generation of the event feature vector set specifically includes:

[0074] Receive multi-source event data during the operation of the smart lock, and reorganize user identifier, device identifier, unlocking method, timestamp, spatial location and communication signal elements into a structured event element sequence according to a unified field order;

[0075] Discrete mapping and sparse coding are performed on the user identifier, device identifier, and unlocking method in the structured event element sequence to obtain the identity method coding result;

[0076] The timestamps in the structured event element sequence are decomposed into time, expanded into cycles, and mapped into phases. The timestamps are then transformed into time features containing information on intraday time periods, intraweek locations, rhythmic phases, and interval lengths, resulting in time coding results.

[0077] The spatial location in the structured event element sequence is analyzed by coordinates and quantized by region. The latitude and longitude and door location are reduced to the spatial characteristics of the floor, room, door type and the relationship with adjacent areas to obtain the spatial coding result.

[0078] Amplitude normalization, time alignment, and short sequence aggregation are performed on the communication signal elements in the structured event element sequence. Physical side statistics such as signal strength variation amplitude, stable interval length, rise and fall edge duration, pulse count, and jitter measurement are extracted to obtain the communication signal encoding results.

[0079] The identity encoding result, time encoding result, spatial encoding result and communication signal encoding result are concatenated and scale-aligned to form a single event vector, thus obtaining the event embedding result;

[0080] The format and storage layout of the event embedding results are standardized, the field order, numerical range and storage alignment are determined, and the results are organized into a set of event feature vectors based on the timestamp sequence.

[0081] In this embodiment, the generation of the neighborhood aggregation result specifically includes:

[0082] Based on the event feature vector set, an initial node list is established for user nodes, device nodes, mode nodes, time nodes, and space nodes, and a corresponding index is generated for each event;

[0083] Based on the initial node list, according to the co-occurrence relationship of elements of the same event, the interaction relationship between user nodes and device nodes, the selection relationship between user nodes and mode nodes, the occurrence relationship between device nodes and time nodes, the positioning relationship between device nodes and spatial nodes, and the time period relationship between mode nodes and time nodes are generated. The timestamp order, attempt number, success or failure flag and communication signal statistics of each relationship are recorded to obtain the relationship instance table.

[0084] The relation instance table is divided into pieces according to the timestamp and the continuous sliding time window to form a relation piece sequence organized by window number; within each window, relation instances are summarized according to the co-occurrence relation type of the elements to obtain a time-varying multi-relationship graph structure containing user nodes, device nodes, mode nodes, time nodes and space nodes.

[0085] Within each window, a relationship weight and a time-series attenuation coefficient are calculated for each type of element co-occurrence relationship. The relationship weight is determined based on the weighted result of the relationship occurrence frequency, communication signal stability, and time proximity. The time-series attenuation coefficient is calculated based on the time interval between the relationship and the end of the window. The two are multiplied to obtain the effective relationship weight and written into the relationship weight table.

[0086] Based on the relation weight table, an adjacency index and batch routing are established for the co-occurrence relationship of each type of element. User nodes, device nodes, mode nodes, time nodes and space nodes are divided into several subgraphs that can be processed in parallel, and relation adjacency index and batch routing table are generated.

[0087] Within each window, graph messaging and relationship aggregation are performed according to the batch routing table. The neighbor contributions of each type of element co-occurrence relationship are weighted and summarized based on the relationship weight table to obtain the temporary node representation and message set. The temporary node representation is the result of weighted aggregation of the neighbor node features of each type of element co-occurrence relationship.

[0088] The temporary representations of nodes are merged and stabilized according to time windows, and the neighborhood aggregation result of the node at the current time is generated by using the mean of the window and the weighting strategy of the recent window.

[0089] In this embodiment, the generation of the slow-frequency memory state and the fast-frequency memory state specifically includes:

[0090] A dual-frequency memory time-series neural network is constructed, comprising a dual-frequency memory layer, a gated fusion layer, a node embedding output layer, and a recognition and anomaly determination layer. The dual-frequency memory layer includes a slow-frequency memory unit, a fast-frequency memory unit, a trigger manager, a memory read / write interface, and a state cache.

[0091] At the edge, a node-state index is established based on the neighborhood aggregation result. The node-state index is used to record the storage location, access path and last update time of each node in the slow frequency memory unit and the fast frequency memory unit.

[0092] Based on the node-state index, the slow frequency memory unit template and behavior baseline of the previous cycle are loaded from the cloud, and the fast frequency memory unit snapshot retained during the last processing is read from the edge to generate the initial slow frequency memory state and the initial fast frequency memory state.

[0093] Align the neighborhood aggregation results with the event feature vector set by timestamp and index, extract the embedded fields corresponding to user identifier, device identifier, unlocking method, timestamp, spatial location and communication signal elements, and generate a memory update message sequence in chronological order;

[0094] Based on the initial fast-frequency memory state, the memory update message sequence is sequentially written into the fast-frequency memory unit. The manager is triggered to perform gating write, forgetting, and state advancement operations when each message is written. The gating write filters the effective feature dimensions of the current event, the forgetting operation reduces the weight of outdated or noisy features, and the state advancement writes the immediate impact of the latest event on the node's behavior into the short-term state cache. After each memory update message is written, an intermediate fast-frequency memory state is generated and stored in chronological order to form an intermediate fast-frequency memory state sequence. The fast-frequency memory unit takes the middle state of the last item in the intermediate fast-frequency memory state sequence as the current fast-frequency memory state, which represents the short-term immediate behavior response of the node at the current moment. Finally, the current fast-frequency memory state and the intermediate fast-frequency memory state set are output.

[0095] Based on the initial slow-frequency memory state, when the trigger manager detects the end of the time window or a continuous shift in the behavior baseline, it collects all memory update messages and the corresponding fast-frequency memory state sequence within the window, performs window aggregation and smoothing, calculates the mean of stable features and trend shift within the window, and performs template fine-tuning and state correction in combination with historical slow-frequency memory states. The slow-frequency memory unit writes the stable behavior patterns and slow-changing features in the window aggregation result into the long-term template through an adaptive gating mechanism, updates the center distribution of the slow-frequency memory according to the trend shift, and obtains the current slow-frequency memory state corresponding to the window number.

[0096] Align and deduplicate the current fast-frequency memory state and the current slow-frequency memory state according to the node-state index, write them to the state cache to form a pair of outputs of fast-frequency memory state and slow-frequency memory state, and generate a state update timestamp and trigger flag for each pair of outputs.

[0097] In this embodiment, the generation of the full-time-domain node embedding vector specifically includes:

[0098] Receive the slow-frequency memory state and the fast-frequency memory state, as well as the corresponding state update timestamp and trigger flag. Combine the neighborhood aggregation results with the time context and scene information provided by the time node, method node, and space node to establish a fusion batch index and obtain the fusion input set of the gated fusion layer.

[0099] Perform scale alignment, dimension alignment, and missing completion on the slow-frequency and fast-frequency memory states in the fusion input set to generate an aligned set of memory state pairs.

[0100] Based on the state update timestamp and trigger marker, the time period attribute corresponding to the time node, the unlocking method attribute corresponding to the method node, and the area attribute corresponding to the spatial node, the time fluctuation index, behavior consistency score and signal disturbance intensity are calculated for the set according to the aligned memory state, and a gating prior element table is generated.

[0101] The generation of the gated prior element table specifically includes:

[0102] The event time series of the same node within the current window is extracted from the neighborhood aggregation results and state update timestamps. The dispersion of the interval between adjacent events and the proportion of abnormal triggers are statistically analyzed to calculate the time fluctuation index. Using the slow-frequency memory state and behavioral baseline as references, and combining the unlocking methods, time periods, and regional attributes corresponding to the mode nodes, time nodes, and spatial nodes within the current window, the consistency between the current event set and the long-term template in terms of mode matching rate, time period matching rate, and regional matching rate is calculated to obtain the behavioral consistency score. Jitter metric, signal strength variation amplitude, and stable interval length are extracted from the communication signal encoding results. The jitter metric and variation amplitude are positively accumulated, and the stable interval length is calculated. The signal disturbance intensity is obtained by reverse conversion of the degree. After normalizing the time fluctuation index, behavior consistency score and signal disturbance intensity, a gating priority label and gating prior weight are generated according to the rules. When the time fluctuation index and signal disturbance intensity are higher than the corresponding index threshold and the behavior consistency score is lower than the corresponding index threshold, it is set to fast frequency priority and the gating prior weight is set to 0.9. When the time fluctuation index and signal disturbance intensity are lower than the corresponding index threshold and the behavior consistency score is higher than the corresponding index threshold, it is set to slow frequency priority and the low gating prior weight is set to 0.1. In other cases, the priority direction and gating prior weight are determined according to the weighted result of the three, forming a gating prior element table.

[0103] In the gated fusion layer, based on the gated prior feature table and the aligned memory state pair set, a gated coefficient is generated for each node, and the gated coefficient sequence is output and a corresponding relationship is established with the fusion batch index.

[0104] The set is weighted and combined node by node according to the aligned memory state based on the gating coefficient sequence, and then stabilized and constrained, including range pruning, anomaly suppression and time smoothing, to suppress abnormal fluctuations and generate candidate full-time domain node embedding vectors.

[0105] The consistency of the candidate full-time domain node embedding vector with the relation weight table and the window number of the neighborhood aggregation result is checked. For nodes that do not meet the consistency, a slow frequency priority or fast frequency priority backoff strategy is triggered according to the gating coefficient sequence, and the full-time domain node embedding vector is output.

[0106] In this embodiment, the generation of the comprehensive identification decision specifically includes:

[0107] The full-time domain nodes are embedded into the vector input nodes and the output layer is embedded into the vector input nodes. Dimensionality reduction mapping, range normalization and numerical pruning are performed according to the preset field order to generate a set of node representation vectors corresponding to the event index.

[0108] Based on the event index, user node representation, device node representation, mode node representation, time node representation and spatial node representation are extracted from the node representation vector set and aligned with the corresponding records in the event feature vector set to form an event alignment vector group;

[0109] The event alignment vector group is input into the identification and anomaly determination layer, which consists of an identity matching branch, an anomaly risk branch, a forgery credibility branch, and a fusion decision module. In the identity matching branch, the identity matching score is calculated based on the event alignment vector group, and the identity matching result and threshold determination label are output. The identity matching score is generated by a combination feature mapping of user node representation, device node representation, mode node representation, time node representation, and spatial node representation.

[0110] In the abnormal risk branch, short-term fluctuation indicators, failure retry density and cross-regional switching rate are extracted based on the event alignment vector group and the event feature vector set. The abnormal risk score is calculated by combining the user node representation and the device node representation, and the abnormal risk result and risk level label are output. The abnormal risk score reflects the behavioral stability and risk offset trend of the node within the time window.

[0111] In the forgery credibility branch, the signal disturbance intensity, stable interval length conversion value and rise and fall edge duration ratio are extracted based on the communication signal encoding results in the event alignment vector group and event feature vector set. The forgery credibility score is calculated by combining the mode node representation and spatial node representation, and the forgery credibility result and suspicious type mark are output. The forgery credibility score measures the consistency of physical signal and the naturalness of unlocking behavior.

[0112] The identity matching score, abnormal risk score, and forgery credibility score are input into the fusion decision module to generate a comprehensive identification score, which is formed by the weighted results of the three branches. A comprehensive identification decision is generated based on the comprehensive identification score and a preset threshold table. The comprehensive identification decision includes three handling levels: release, secondary verification, and rejection. The handling result carries branch descriptions and key feature summaries of the identity matching result, abnormal risk result, and forgery credibility result.

[0113] In this embodiment, the generation of the updated dual-frequency memory time-series neural network specifically includes:

[0114] The system receives a comprehensive identification decision, a handling level, and an explanatory tag. At the edge, it executes an immediate response based on the handling level. If the comprehensive identification decision is to allow, it allows the unlocking action and records the execution time. If the comprehensive identification decision is to perform secondary verification, it triggers the user identity multi-factor verification module and suspends the unlocking response. If the comprehensive identification decision is to deny, it blocks the unlocking request and generates a security event log.

[0115] After an immediate response, the edge device records the operation feedback results, including whether the unlocking action was successful or failed, whether the verification was passed or rejected, the integrity of the communication signal and the response delay of the device. The feedback results are then linked with the comprehensive identification decision to form a handling record table.

[0116] Align the handling record table with the node representation vector set, extract the behavior pattern fields of user nodes, device nodes, method nodes, time nodes and space nodes to form a set of behavior feedback features;

[0117] At the edge, fast learning and memory calibration are performed based on the behavioral feedback feature set. When there is a deviation between the feedback result and the comprehensive recognition decision, the fast frequency memory unit update operation is triggered to adjust the short-term state cache corresponding to the node, update the short-term behavior pattern and signal response law, recalculate the fast frequency memory state of the node, and output the updated fast frequency memory state.

[0118] The updated set of fast-frequency memory states and behavioral feedback features generated at the edge is packaged by node index and time window number and uploaded to the cloud memory synchronization module. Based on the time window number, the historical slow-frequency memory unit template and behavioral baseline are matched, and the feedback offset and trend change are calculated.

[0119] In the slow frequency memory cell, template updates and behavioral baseline corrections are performed based on the feedback offset and trend change. The changing trend of the offset features is written into the slow frequency memory cell template, abnormal perturbation samples are deleted and the slow frequency memory state is updated to form the slow frequency memory cell correction template for the current window.

[0120] The slow-frequency memory unit correction template and the updated fast-frequency memory state at the edge are reintegrated into a synchronization packet, which is then sent to the edge to replace the old state cache and fast-frequency memory unit snapshot, forming an updated dual-frequency memory timing graph neural network for the next smart lock recognition operation.

[0121] Example 1:

[0122] To verify the feasibility of this invention in practice, it was applied to a smart lock group management scenario in a large residential area in a coastal city. This residential area is equipped with networked smart locks of multiple brands and models, and has a complex user base, including residents' family members, as well as various access entities such as property management, cleaning, maintenance, visitors, and delivery personnel. Traditional lock identification often relies on single authentication or static records. In scenarios with frequent user switching, fluctuating communication signals, and significant differences in behavior over time, problems such as authorization misjudgment, delayed access, and erroneous alarms frequently occur. Especially in scenarios with frequent visitor access and cross-regional scheduling, traditional identification methods struggle to balance sensitivity and stability. This invention, in this scenario, integrates multi-source event data from the lock control terminal, mobile terminal, and property management platform, covering elements such as user identifier, device identifier, unlocking method, timestamp, spatial location, and communication signal. This data is then uniformly formatted and embedded at the edge to form a multi-dimensional event feature vector set.

[0123] In practical applications, the time-varying multi-relationship graph structure constructed by this invention unifies user, device, method, time, and spatial information into a graph model, dynamically capturing the contextual relationships of unlocking behavior. The system automatically identifies the interaction relationships between different nodes, such as the stable binding relationship between the owner and their own device, the time period patterns of property staff and multi-point access, and the communication characteristics of visitors and temporary authorized paths. In the graph neural network layer, parallel modeling of short-term access and long-term behavior is achieved through neighborhood aggregation and weight decay mechanisms. The dual-frequency memory time-series graph neural network runs continuously at the edge, with the fast-frequency memory unit capturing real-time behavioral changes, such as user unlocking method switching or signal disturbance anomalies; while the slow-frequency memory unit is periodically updated in the cloud, continuously learning the user's stable routine and behavioral baseline to form behavioral templates and trend distributions.

[0124] During continuous operation over several days, the gating fusion layer of this invention dynamically adjusts the fusion ratio of fast and slow frequency memories based on the temporal context and scene conditions, ensuring high recognition accuracy even under conditions of frequent nighttime entry and exit, device signal attenuation, or changes in user movement trajectories. The recognition and anomaly determination layer simultaneously outputs three types of results: identity matching, anomaly risk, and forgery credibility, enabling real-time responses at the edge: allowing users with normal identification to pass directly, triggering secondary verification for users whose behavior deviates but still has high credibility, and blocking and issuing alarms for detected signal forgery or cross-device anomalies. Subsequently, the cloud adjusts the slow frequency memory template based on the recognition and feedback results, allowing the model to gradually learn the personalized patterns of different users in terms of time, space, and mode, achieving continuous optimization.

[0125] By analyzing the temporal behavior data continuously collected from multiple smart locks within the same residential area, the dual-frequency memory temporal graph neural network constructed in this invention significantly reduces the cross-regional false recognition rate and the false judgment rate of spoofed signals, while improving the stability and latency tolerance of the recognition response. During overall operation, the smart lock recognition maintained stable output under high concurrency and complex signal environments, without any locking errors or recognition stalls, verifying the effectiveness and applicability of this method in multi-source data fusion, secure recognition, and self-learning updates.

[0126] Table 1. Performance comparison between the smart lock recognition method based on temporal graph neural network and traditional methods.

[0127]

[0128] As shown in Table 1, the dual-frequency memory temporal graph neural network method proposed in this invention significantly improves performance compared to traditional graph neural network methods in several key indicators. The recognition accuracy increased from 91.2% to 97.8%, mainly due to the joint modeling capability of the dual-frequency memory layer for long-term and short-term behaviors, enabling the model to simultaneously capture stable patterns and instantaneous changes. The anomaly detection recall increased from 84.7% to 93.5%, indicating that the mechanism of capturing sudden behaviors with fast-frequency memory units and correcting long-term templates with slow-frequency memory units effectively reduces the risk of missed detections. The forged signal recognition rate increased from 82.5% to 94.2%, attributed to the dynamic weight allocation strategy of the gating fusion layer between behavioral consistency and signal perturbation features. In terms of response performance, the average response time decreased from 245 milliseconds to 186 milliseconds, demonstrating the advantages of fast edge inference and the dual-frequency synchronization mechanism in low-latency processing. The cross-device false recognition rate decreased significantly from 4.6% to 1.2%, indicating that the model has higher discrimination accuracy in node feature alignment and context judgment. Furthermore, the long-term operational stability was improved from 72 hours to 168 hours, verifying the robustness and maintainability of the dual-frequency memory structure under continuous learning and cloud-edge collaborative updates. In summary, this invention significantly outperforms traditional methods in terms of recognition accuracy, anomaly detection, response efficiency, and long-term stability, demonstrating the combined advantages of model structure design and temporal feature fusion strategies.

[0129] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart lock recognition method based on a temporal graph neural network, characterized in that, Includes the following steps: Acquire multi-source event data during the operation of the smart lock, perform format unification and embedding encoding, and obtain a set of event feature vectors; A time-varying multi-relationship graph structure is constructed based on the event feature vector set. User nodes, device nodes, mode nodes, time nodes and spatial nodes are associated. Graph message passing and neighborhood feature aggregation are performed to obtain the neighborhood aggregation result of the node at the current time. A dual-frequency memory time-series graph neural network is constructed. The neighborhood aggregation result is input into the dual-frequency memory layer, and the memory state is updated in the slow-frequency memory unit and the fast-frequency memory unit respectively. The slow-frequency memory state and the fast-frequency memory state are output. The slow-frequency memory state and the fast-frequency memory state are input into the gating fusion layer, and adaptive fusion is performed through the gating function based on the time context and scene information to generate a full-time domain node embedding vector. The input nodes of the full-time domain are embedded into the output layer and mapped to low-dimensional node representation vectors. The identity matching result, the abnormal risk result and the forgery credibility result are calculated through the identification and anomaly judgment layer, and the comprehensive identification decision is output. Based on comprehensive identification decisions, immediate responses and security measures are executed at the edge, and slow-frequency memory units and behavioral baselines are updated in the cloud to form an updated dual-frequency memory time-series neural network.

2. The smart lock recognition method based on a time-series graph neural network according to claim 1, characterized in that, The generation of the event feature vector set specifically includes: Receive multi-source event data during the operation of the smart lock, and reorganize user identifier, device identifier, unlocking method, timestamp, spatial location and communication signal elements into a structured event element sequence according to a unified field order; Discrete mapping and sparse coding are performed on the user identifier, device identifier, and unlocking method in the structured event element sequence to obtain the identity method coding result; The timestamps in the structured event element sequence are decomposed into time, expanded into cycles, and mapped into phases. The timestamps are then transformed into time features containing information on intraday time periods, intraweek locations, rhythmic phases, and interval lengths, resulting in time coding results. The spatial location in the structured event element sequence is analyzed by coordinates and quantized by region. The latitude and longitude and door location are reduced to the spatial characteristics of the floor, room, door type and the relationship with adjacent areas to obtain the spatial coding result. Amplitude normalization, time alignment, and short sequence aggregation are performed on the communication signal elements in the structured event element sequence. Physical side statistics such as signal strength variation amplitude, stable interval length, rise and fall edge duration, pulse count, and jitter measurement are extracted to obtain the communication signal encoding results. The identity encoding result, time encoding result, spatial encoding result and communication signal encoding result are concatenated and scale-aligned to form a single event vector, thus obtaining the event embedding result; The format and storage layout of the event embedding results are standardized, the field order, numerical range and storage alignment are determined, and the results are organized into a set of event feature vectors based on the timestamp sequence.

3. The smart lock recognition method based on a time-series graph neural network according to claim 1, characterized in that, The generation of the neighborhood aggregation result specifically includes: Based on the event feature vector set, an initial node list is established for user nodes, device nodes, mode nodes, time nodes, and space nodes, and a corresponding index is generated for each event; Based on the initial node list, according to the co-occurrence relationship of elements of the same event, the interaction relationship between user nodes and device nodes, the selection relationship between user nodes and mode nodes, the occurrence relationship between device nodes and time nodes, the positioning relationship between device nodes and spatial nodes, and the time period relationship between mode nodes and time nodes are generated. The timestamp order, attempt number, success or failure flag and communication signal statistics of each relationship are recorded to obtain the relationship instance table. The relation instance table is divided into pieces according to the timestamp and the continuous sliding time window to form a relation piece sequence organized by window number; within each window, relation instances are summarized according to the co-occurrence relation type of the elements to obtain a time-varying multi-relationship graph structure containing user nodes, device nodes, mode nodes, time nodes and space nodes. Within each window, a relationship weight and a time-series attenuation coefficient are calculated for each type of element co-occurrence relationship. The relationship weight is determined based on the weighted result of the relationship occurrence frequency, communication signal stability, and time proximity. The time-series attenuation coefficient is calculated based on the time interval between the relationship and the end of the window. The two are multiplied to obtain the effective relationship weight and written into the relationship weight table. Based on the relation weight table, an adjacency index and batch routing are established for the co-occurrence relationship of each type of element. User nodes, device nodes, mode nodes, time nodes and space nodes are divided into several subgraphs that can be processed in parallel, and relation adjacency index and batch routing table are generated. Within each window, graph messaging and relationship aggregation are performed according to the batch routing table. The neighbor contributions of each type of element co-occurrence relationship are weighted and summarized based on the relationship weight table to obtain the temporary node representation and message set. The temporary node representation is the result of weighted aggregation of the neighbor node features of each type of element co-occurrence relationship. The temporary representations of nodes are merged and stabilized according to time windows, and the neighborhood aggregation result of the node at the current time is generated by using the mean of the window and the weighting strategy of the recent window.

4. The smart lock recognition method based on a time-series graph neural network according to claim 1, characterized in that, The generation of the slow-frequency memory state and the fast-frequency memory state specifically includes: A dual-frequency memory time-series neural network is constructed, comprising a dual-frequency memory layer, a gated fusion layer, a node embedding output layer, and a recognition and anomaly determination layer. The dual-frequency memory layer includes a slow-frequency memory unit, a fast-frequency memory unit, a trigger manager, a memory read / write interface, and a state cache. At the edge, a node-state index is established based on the neighborhood aggregation result. The node-state index is used to record the storage location, access path and last update time of each node in the slow frequency memory unit and the fast frequency memory unit. Based on the node-state index, the slow frequency memory unit template and behavior baseline of the previous cycle are loaded from the cloud, and the fast frequency memory unit snapshot retained during the last processing is read from the edge to generate the initial slow frequency memory state and the initial fast frequency memory state. Align the neighborhood aggregation results with the event feature vector set by timestamp and index, extract the embedded fields corresponding to user identifier, device identifier, unlocking method, timestamp, spatial location and communication signal elements, and generate a memory update message sequence in chronological order; Based on the initial fast-frequency memory state, the memory update message sequence is sequentially written into the fast-frequency memory unit. The manager is triggered to perform gating write, forgetting, and state advancement operations when each message is written. The gating write filters the effective feature dimensions of the current event, the forgetting operation reduces the weight of outdated or noisy features, and the state advancement writes the immediate impact of the latest event on the node's behavior into the short-term state cache. After each memory update message is written, an intermediate fast-frequency memory state is generated and stored in chronological order to form an intermediate fast-frequency memory state sequence. The fast-frequency memory unit takes the middle state of the last item in the intermediate fast-frequency memory state sequence as the current fast-frequency memory state, which represents the short-term immediate behavior response of the node at the current moment. Finally, the current fast-frequency memory state and the intermediate fast-frequency memory state set are output. Based on the initial slow-frequency memory state, when the trigger manager detects the end of the time window or a continuous shift in the behavior baseline, it collects all memory update messages and the corresponding fast-frequency memory state sequence within the window, performs window aggregation and smoothing, calculates the mean of stable features and trend shift within the window, and performs template fine-tuning and state correction in combination with historical slow-frequency memory states. The slow-frequency memory unit writes the stable behavior patterns and slow-changing features in the window aggregation result into the long-term template through an adaptive gating mechanism, updates the center distribution of the slow-frequency memory according to the trend shift, and obtains the current slow-frequency memory state corresponding to the window number. Align and deduplicate the current fast-frequency memory state and the current slow-frequency memory state according to the node-state index, write them to the state cache to form a pair of outputs of fast-frequency memory state and slow-frequency memory state, and generate a state update timestamp and trigger flag for each pair of outputs.

5. The smart lock recognition method based on a time-series graph neural network according to claim 1, characterized in that, The generation of the full-time-domain node embedding vector specifically includes: Receive the slow-frequency memory state and the fast-frequency memory state, as well as the corresponding state update timestamp and trigger flag. Combine the neighborhood aggregation results with the time context and scene information provided by the time node, method node, and space node to establish a fusion batch index and obtain the fusion input set of the gated fusion layer. Perform scale alignment, dimension alignment, and missing completion on the slow-frequency and fast-frequency memory states in the fusion input set to generate an aligned set of memory state pairs. Based on the state update timestamp and trigger marker, the time period attribute corresponding to the time node, the unlocking method attribute corresponding to the method node, and the area attribute corresponding to the spatial node, the time fluctuation index, behavior consistency score and signal disturbance intensity are calculated for the set according to the aligned memory state, and a gating prior element table is generated. In the gated fusion layer, based on the gated prior feature table and the aligned memory state pair set, a gated coefficient is generated for each node, and the gated coefficient sequence is output and a corresponding relationship is established with the fusion batch index. The set is weighted and combined node by node according to the aligned memory state based on the gating coefficient sequence, and then stabilized and constrained, including range pruning, anomaly suppression and time smoothing, to suppress abnormal fluctuations and generate candidate full-time domain node embedding vectors. The consistency of the candidate full-time domain node embedding vector with the relation weight table and the window number of the neighborhood aggregation result is checked. For nodes that do not meet the consistency, a slow frequency priority or fast frequency priority backoff strategy is triggered according to the gating coefficient sequence, and the full-time domain node embedding vector is output.

6. The smart lock recognition method based on a time-series graph neural network according to claim 1, characterized in that, The generation of the comprehensive identification decision specifically includes: The full-time domain nodes are embedded into the vector input nodes and the output layer is embedded into the vector input nodes. Dimensionality reduction mapping, range normalization and numerical pruning are performed according to the preset field order to generate a set of node representation vectors corresponding to the event index. Based on the event index, user node representation, device node representation, mode node representation, time node representation and spatial node representation are extracted from the node representation vector set and aligned with the corresponding records in the event feature vector set to form an event alignment vector group; The event alignment vector group is input into the identification and anomaly determination layer, which consists of an identity matching branch, an anomaly risk branch, a forgery credibility branch, and a fusion decision module. In the identity matching branch, the identity matching score is calculated based on the event alignment vector group, and the identity matching result and threshold determination label are output. In the abnormal risk branch, short-term fluctuation indicators, failure retry density and cross-regional switching rate are extracted based on the event alignment vector group and the event feature vector set. The abnormal risk score is calculated by combining the user node representation and the device node representation, and the abnormal risk result and risk level label are output. In the forgery credibility branch, the signal disturbance intensity, stable interval length conversion value and rise and fall edge duration ratio are extracted based on the communication signal encoding results in the event alignment vector group and event feature vector set. The forgery credibility score is calculated by combining the method node representation and spatial node representation, and the forgery credibility result and suspicious type label are output. The identity matching score, abnormal risk score, and forgery credibility score are input into the fusion decision module to generate a comprehensive identification score, which is formed by the weighted results of the three branches. A comprehensive identification decision is generated based on the comprehensive identification score and a preset threshold table.

7. The smart lock recognition method based on a time-series graph neural network according to claim 1, characterized in that, The generation of the updated dual-frequency memory time-series neural network specifically includes: The system receives a comprehensive identification decision, a handling level, and an explanatory tag. At the edge, it executes an immediate response based on the handling level. If the comprehensive identification decision is to allow, it allows the unlocking action and records the execution time. If the comprehensive identification decision is to perform secondary verification, it triggers the user identity multi-factor verification module and suspends the unlocking response. If the comprehensive identification decision is to deny, it blocks the unlocking request and generates a security event log. After an immediate response, the edge device records the operation feedback results, including whether the unlocking action was successful or failed, whether the verification was passed or rejected, the integrity of the communication signal and the response delay of the device. The feedback results are then linked with the comprehensive identification decision to form a handling record table. Align the handling record table with the node representation vector set, extract the behavior pattern fields of user nodes, device nodes, method nodes, time nodes and space nodes to form a set of behavior feedback features; At the edge, fast learning and memory calibration are performed based on the behavioral feedback feature set. When there is a deviation between the feedback result and the comprehensive recognition decision, the fast frequency memory unit update operation is triggered to adjust the short-term state cache corresponding to the node, update the short-term behavior pattern and signal response law, recalculate the fast frequency memory state of the node, and output the updated fast frequency memory state. The updated set of fast-frequency memory states and behavioral feedback features generated at the edge is packaged by node index and time window number and uploaded to the cloud memory synchronization module. Based on the time window number, the historical slow-frequency memory unit template and behavioral baseline are matched, and the feedback offset and trend change are calculated. In the slow frequency memory cell, template updates and behavioral baseline corrections are performed based on the feedback offset and trend change. The changing trend of the offset features is written into the slow frequency memory cell template, abnormal perturbation samples are deleted and the slow frequency memory state is updated to form the slow frequency memory cell correction template for the current window. The slow-frequency memory unit correction template and the updated fast-frequency memory state at the edge are reintegrated into a synchronization packet, which is then sent to the edge to replace the old state cache and fast-frequency memory unit snapshot, forming an updated dual-frequency memory timing graph neural network for the next smart lock recognition operation.

Images