Intelligent lock identification method based on time sequence diagram neural network

Abstract

The invention discloses an intelligent lock recognition method based on a time sequence diagram neural network, and the method comprises the following steps: collecting multi-source event data of an intelligent lock, unifying the format, embedding codes, and generating an event feature vector set; constructing a time-varying multi-relation graph, and executing message passing and feature aggregation to obtain a neighborhood aggregation result; constructing a dual-frequency memory time sequence diagram neural network, and extracting long-term and short-term behavior characteristics in slow-frequency and fast-frequency memory units; inputting the slow frequency memory state and the fast frequency memory state into a gating fusion layer, and combining time and scene information to generate a full-time-domain embedded vector; calculating identity matching, abnormal risk and forgery credibility, and outputting a comprehensive identification result; and executing instant response and security processing according to an identification result, and carrying out network updating. According to the method, the sequence diagram neural network and the double-frequency memory mechanism are utilized, multi-element dynamic recognition of the intelligent lock is achieved, and the method has the advantages of being high in self-learning, high in recognition accuracy and excellent in safety robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of security and protection identification, and in particular to a smart lock identification method based on a time series graph neural network. BACKGROUND

[0002] Existing smart lock identification technologies are mainly based on static identity authentication, usually relying on single features such as fingerprints, passwords, faces, or Bluetooth signals for matching and judgment, which has obvious shortcomings in security and stability. Some improved solutions attempt to introduce data modeling based on time series or device behavior, and realize dynamic identification through deep learning or time series neural networks, but such methods usually only model single-dimensional input data, lack joint characterization of complex interaction relationships between users, devices, methods, time, and space, and thus have poor adaptability to environmental changes and multi-user scenarios. In addition, existing solutions generally rely on centralized cloud processing, which is difficult to respond to real-time events on the edge, and has high recognition delay.

[0003] In the field of dynamic behavior modeling and anomaly identification, existing research has proposed a combination of graph neural networks and long short-term memory networks to capture the time-dependent relationship between nodes or periodic features, but these methods generally use a single frequency time update mechanism, which can only reflect a certain time scale of behavior changes and cannot simultaneously consider short-term fluctuations and long-term regularities. For high-frequency interaction devices such as smart locks, existing models are difficult to balance between rapidly changing unlocking events and long-term user habits, leading to false positives in anomaly detection. In addition, the lack of edge instant feedback and cloud behavior baseline collaborative update mechanism makes the model unable to form 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 identification method based on a time series graph neural network is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a smart lock identification method based on a time series graph neural network. The present application uses a time series graph neural network and a dual-frequency memory mechanism to realize multi-element dynamic identification of a smart lock, and has the advantages of strong self-learning, high identification accuracy, and high security robustness.

[0006] According to the smart lock identification method based on the time series graph neural network of the present application, the method comprises the following steps: Obtain multi-source event data in the running process of the smart lock, perform format unification and embedding coding, and obtain an event feature vector set; A time-varying multi-relation graph structure is constructed according to the event feature vector set, user nodes, device nodes, mode nodes, time nodes and space nodes are associated, graph message passing and neighborhood feature aggregation are performed, and neighborhood aggregation results of the nodes at the current time are obtained; A dual-frequency memory time sequence graph neural network is constructed, the neighborhood aggregation results are input into the dual-frequency memory layer, the memory state is updated in the slow-frequency memory unit and the fast-frequency memory unit respectively, and 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 gate fusion layer, and the gate function is used for adaptive fusion according to the time context and the scene information, and a full-time domain node embedding vector is generated; The full-time domain node embedding vector is input into the node embedding output layer, mapped into a low-dimensional node representation vector, and through the identification and anomaly judgment layer, the identity matching result, the abnormal risk result and the forgery credibility result are calculated, and the comprehensive identification decision is output; According to the comprehensive identification decision, instant response and safety disposal are performed at the edge, and the slow-frequency memory unit and the behavior baseline are updated in the cloud to form an updated dual-frequency memory time sequence graph neural network.

[0007] Optionally, the generation of the event feature vector set specifically includes: Receiving multi-source event data in the operation process of the intelligent lock, reorganizing user identification, device identification, unlocking mode, timestamp, space position 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 identification, device identification and unlocking mode in the structured event element sequence to obtain an identity mode coding result; Time decomposition, period expansion and phase mapping are performed on the timestamp in the structured event element sequence, and the timestamp is converted into time features containing intra-day time period, intra-week position, rhythm phase and interval length information to obtain a time coding result; Coordinate analysis and area quantization are performed on the space position in the structured event element sequence, and the latitude and longitude and door position are reduced to the space features of floor, room, door type and adjacent area relationship to obtain a space coding result; Amplitude normalization, time alignment and short sequence aggregation are performed on the communication signal elements in the structured event element sequence, and physical side statistics such as signal strength change amplitude, stable interval length, rising and falling edge length, pulse count and jitter metric are extracted to obtain a communication signal coding result; The identity mode coding result, the time coding result, the space coding result and the communication signal coding result are spliced and scaled to form a single event vector, and an event embedding result is obtained; The event embedding result is subjected to format unification and storage layout normalization to determine field order, numerical range and storage alignment manner, and is sorted into an event feature vector set according to timestamps.

[0008] Optionally, the generation of the neighborhood aggregation result specifically includes: An initial node list of user nodes, device nodes, method nodes, time nodes and space nodes is established according to the event feature vector set, and an index corresponding to each event is generated; Based on the initial node list, an interaction relationship between user nodes and device nodes, a selection relationship between user nodes and method nodes, an occurrence relationship between device nodes and time nodes, a positioning relationship between device nodes and space nodes, and a time period relationship between method nodes and time nodes are generated according to the co-occurrence relationship of elements of the same event, the timestamp order, the attempt number, the success or failure mark and the communication signal statistics of each relationship are recorded, and a relationship instance table is obtained; The relationship instance table is divided into fragments according to a continuous sliding time window in timestamp order, forming a relationship fragment sequence organized by window number; within each window, the relationship instances are summarized according to the co-occurrence relationship type, and a time-varying multi-relationship graph structure containing user nodes, device nodes, method nodes, time nodes and space nodes is obtained; In each window, the relationship weight and the time sequence decay coefficient are calculated for each type of co-occurrence relationship, the relationship weight is determined according to the weighted results of relationship occurrence frequency, communication signal stability and time proximity, the time sequence decay coefficient is calculated according to the time interval between the relationship and the end of the window, the effective relationship weight is obtained by multiplying the two, and written into the relationship weight table; Based on the relationship weight table, an adjacency index and a batch routing are established for each type of co-occurrence relationship, and user nodes, device nodes, method nodes, time nodes and space nodes are divided into several subgraphs that can be processed in parallel, and a relationship adjacency index and batch routing table is generated; In each window, the graph message passing and relationship aggregation are performed according to the batch routing table, the neighbor contribution of each type of co-occurrence relationship is weighted and summarized based on the relationship weight table, and a node temporary representation and a message set are obtained, wherein the node temporary representation is the weighted aggregation result of the neighbor node features of each type of co-occurrence relationship; The node temporary representation is merged and stabilized according to the time window, and the node neighborhood aggregation result at the current time is generated by using the window mean value and the recent window weighting strategy.

[0009] Optionally, the generation of the slow frequency memory state and the fast frequency memory state specifically includes: The dual-frequency memory time series diagram neural network is constructed, including a dual-frequency memory layer, a gate fusion layer, a node embedding output layer, and an identification and anomaly judgment 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 end, a node-state index is established according to the neighborhood aggregation result, which 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 the behavior baseline of the last period are loaded from the cloud, and the fast-frequency memory unit snapshot reserved when the last processing is read from the edge end, to generate an initial slow-frequency memory state and an initial fast-frequency memory state; The neighborhood aggregation result and the event feature vector set are aligned by timestamp and index, the embedding fields corresponding to user identification, device identification, unlocking method, timestamp, spatial position and communication signal elements are extracted, and a memory update message sequence is generated in time sequence; Based on the initial fast-frequency memory state, the memory update message sequence is written into the fast-frequency memory unit in turn, and the trigger manager performs gate writing, forgetting and state advancing operations when each message is written. The gate writing filters the valid feature dimensions of the current event, the forgetting operation attenuates the weights of outdated or noise features, and the state advancing writes the immediate influence of the latest event on node behavior into the short-term state cache. After each memory update message is written, an intermediate fast-frequency memory state is generated, and the intermediate fast-frequency memory state sequence is formed by storing the intermediate fast-frequency memory states in time sequence. The fast-frequency memory unit takes the last intermediate state of the intermediate fast-frequency memory state sequence as the current fast-frequency memory state, representing the short-term immediate behavior response of the node at the current time. 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 the sustained deviation of the behavior baseline, all memory update messages and corresponding fast-frequency memory state sequences in the window are collected, and window aggregation and smoothing processing is performed on them. The stable feature mean and trend deviation in the window are calculated, the template fine-tuning and state correction are performed in combination with the historical slow-frequency memory state, the slow-frequency memory unit writes the stable behavior pattern and slow-changing feature in the window aggregation result into the long-term template through the adaptive gate mechanism, and the center distribution of the slow-frequency memory is updated according to the trend deviation. The current slow-frequency memory state corresponding to the window number is obtained. The current fast-frequency memory state and the current slow-frequency memory state are aligned and de-duplicated according to the node-state index, written into the state cache to form the paired output of the fast-frequency memory state and the slow-frequency memory state, and the state update timestamp and the trigger mark are generated for each pair of output.

[0010] Optionally, the generation of the full-time domain node embedding vector specifically includes: The received slow frequency memory state and fast frequency memory state, the corresponding state update timestamp and trigger mark, the neighborhood aggregation result, the time node, the mode node and the space node, the time context and scene information provided by the neighborhood aggregation result, the time node, the mode node and the space node, the fusion batch index established by the time context and scene information, and the fusion input set of the gating fusion layer are obtained; The slow frequency memory state and the fast frequency memory state in the fusion input set are subjected to scale alignment, dimension alignment and missing completion to generate a set of aligned memory state pairs; Based on the state update timestamp and the trigger mark, the time period attribute corresponding to the time node, the unlocking mode attribute corresponding to the mode node and the area attribute corresponding to the space node, the time fluctuation index, the behavior consistency score and the signal disturbance intensity are calculated from the set of aligned memory state pairs to generate a gating prior element table; In the gating fusion layer, the gating prior element table and the set of aligned memory state pairs are used to generate a gating coefficient for each node, and the gating coefficient sequence is output and the corresponding relationship with the fusion batch index is established; The set of aligned memory state pairs is weighted and combined node by node according to the gating coefficient sequence, and is subjected to stabilization and constraint processing, including range clipping, abnormality suppression and time smoothing, to suppress abnormal fluctuations and generate a candidate full-time domain node embedding vector; The candidate full-time domain node embedding vector, the relationship weight table and the window number of the neighborhood aggregation result are subjected to consistency checking, and the nodes that do not meet the consistency are subjected to a slow frequency priority or fast frequency priority fallback strategy triggered by the gating coefficient sequence, and a full-time domain node embedding vector is output.

[0011] Optionally, the generation of the comprehensive identification decision specifically includes: The full-time domain node embedding vector is input into the node embedding output layer, and dimension reduction mapping, range normalization and numerical clipping are performed in the preset field order to generate a set of node representation vectors corresponding to the event index; The user node representation, the device node representation, the mode node representation, the time node representation and the space node representation are extracted from the set of node representation vectors according to the event index, and are 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 abnormality judgment layer, and the identification and abnormality judgment layer is composed of an identity matching branch, an abnormality risk branch, a fake credibility branch and a fusion decision module. In the identity matching branch, an identity matching score is calculated based on the event alignment vector group, and an identity matching result and a threshold judgment mark are output; In the abnormality risk branch, short-term fluctuation indicators, failure retry density and cross-region switching rate are extracted based on the event alignment vector group and the event feature vector set, and an abnormality risk score is calculated based on the user node representation and the device node representation, and an abnormality risk result and a risk level mark are output; In the forgery credibility branch, based on the event alignment vector set and the communication signal encoding result in the event feature vector set, the signal disturbance intensity, the stable interval length conversion value and the rising and falling edge time length ratio are extracted, the forgery credibility score is calculated combined with the mode node representation and the space node representation, and the forgery credibility result and the suspicious type label are output; The identity matching score, the abnormal risk score and the forgery credibility score are input into a fusion decision module to generate a comprehensive identification score, which is formed by the weighted results output by the three branches, and a comprehensive identification decision is generated according to the comprehensive identification score and a preset threshold table.

[0012] Optionally, the generation of the updated dual-frequency memory time sequence graph neural network specifically includes: The comprehensive identification decision, the treatment level and the explanatory label are received, and an instant response is performed according to the treatment level at the edge side; if the comprehensive identification decision is release, an unlocking action is allowed to be performed and the execution time is recorded; if the comprehensive identification decision is secondary verification, a user identity multi-factor verification module is triggered and the unlocking response is suspended; and if the comprehensive identification decision is rejection, the unlocking request is blocked and a security event log is generated; After the instant response, an operation feedback result is recorded by the edge side, including unlocking action success or failure, verification pass or rejection, communication signal integrity and device side response delay, the feedback result is associated with the comprehensive identification decision to form a treatment record table; The treatment record table is aligned with the node representation vector set, the behavior mode fields of the user node, the device node, the mode node, the time node and the space node are extracted to form a behavior feedback feature set; The behavior feedback feature set is used to perform fast learning and memory calibration at the edge side, when there is a deviation between the feedback result and the comprehensive identification decision, the fast frequency memory unit is triggered to update the operation, the short-term state cache corresponding to the node is adjusted, the short-term behavior mode and the signal response rule are updated, the fast frequency memory state is recalculated, and the updated fast frequency memory state is output; The updated fast frequency memory state generated by the edge side and the behavior feedback feature set are packaged according to the node index and the time window number, and then uploaded to the cloud end memory synchronization module, the historical slow frequency memory unit template and the behavior baseline are matched according to the time window number, the feedback offset and the trend change degree are calculated; In the slow frequency memory unit, the template update and the behavior baseline correction are performed according to the feedback offset and the trend change degree, the change trend of the offset feature is written into the slow frequency memory unit template, the abnormal disturbance sample is deleted and the slow frequency memory state is updated, and the slow frequency memory unit correction template of the current window is formed; The slow frequency memory unit correction template and the updated fast frequency memory state of the edge side are re-integrated into a synchronization package, which is sent to the edge side to replace the old state cache and the fast frequency memory unit snapshot, and an updated dual-frequency memory time sequence graph neural network is formed, which is used for the next intelligent lock identification operation.

[0013] The beneficial effects of the present application are: The present application realizes unified modeling and dynamic fusion of multiple factors such as users, devices, methods, time and space by constructing an intelligent lock recognition method based on a time series graph neural network, effectively overcoming the deficiencies of existing recognition technologies in data isolation and context fragmentation. By introducing a time-varying multi-relation graph structure, the present application can capture the interaction patterns and relationship weight changes between multiple entities within a sliding time window, enabling the recognition model to have global spatio-temporal correlation understanding capability. On this basis, the dual-frequency memory time series graph neural network realizes parallel updating of fast-frequency memory and slow-frequency memory, corresponding to short-term immediate behavior and long-term stable habits respectively, and can maintain recognition sensitivity and stability in frequent interaction and periodic change scenarios, thereby significantly improving the accuracy of identity recognition and anomaly detection.

[0014] In addition, the designed gating fusion layer generates gating coefficients according to time context, scene information and signal disturbance characteristics, and adaptively adjusts the fusion ratio of slow-frequency and fast-frequency memory states to ensure the robustness and generalization of the model in different unlocking environments. Through the multi-branch joint decision mechanism of the recognition and anomaly judgment layer, the identity matching result, anomaly risk level and forgery credibility can be output simultaneously in a single recognition, realizing multi-dimensional security judgment. Combined with the collaborative updating mechanism of edge end instant response and cloud slow-frequency memory template, the present application forms a closed-loop self-learning system, enabling the model to automatically optimize the behavior template and baseline distribution in continuous recognition cycles, thereby realizing fast, accurate and continuously evolving security recognition effect in complex, multi-user and cross-scene intelligent lock environments. BRIEF DESCRIPTION OF DRAWINGS

[0015] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application and explain the principles of the present application, and do not constitute a limitation of the present application. In the drawings:

[0016] Fig. 1 A flowchart of an intelligent lock recognition method based on a time series graph neural network according to the present application; Fig. 2 A dual-frequency memory time series graph neural network diagram of an intelligent lock recognition method based on a time series graph neural network according to the present application. DETAILED DESCRIPTION

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

[0018] REFERENCE Figs. 1-2A smart lock recognition method based on a time series graph neural network, comprising the following steps: Obtain multi-source event data in the operation process of the smart lock, perform format unification and embedding coding, and obtain an event feature vector set; According to the event feature vector set, a time-varying multi-relation graph structure is constructed, the user node, the device node, the mode node, the time node and the space node are associated, the graph message passing and the neighborhood feature aggregation are performed, and the neighborhood aggregation result of the node at the current time is obtained; A dual-frequency memory time series graph neural network is constructed, the neighborhood aggregation result is input into the dual-frequency memory layer, the memory state is updated in the slow-frequency memory unit and the fast-frequency memory unit respectively, and 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 gate fusion layer, and the gate function is used for adaptive fusion according to the time context and the scene information, and a full-time domain node embedding vector is generated; The full-time domain node embedding vector is input into the node embedding output layer, mapped into a low-dimensional node representation vector, and through the recognition and abnormality judgment layer, the identity matching result, the abnormal risk result and the forgery credibility result are calculated, and the comprehensive recognition decision is output; According to the comprehensive recognition decision, instant response and safety disposal are performed on the edge side, and the slow-frequency memory unit and the behavior baseline are updated on the cloud side, forming an updated dual-frequency memory time series graph neural network.

[0019] In this embodiment, the generation of the event feature vector set specifically includes: Receive multi-source event data in the operation process of the smart lock, reorganize the user identifier, the device identifier, the unlocking mode, the timestamp, the space position and the 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, the device identifier and the unlocking mode in the structured event element sequence to obtain an identity mode coding result; Time decomposition, period expansion and phase mapping are performed on the timestamp in the structured event element sequence, and the timestamp is converted into a time feature containing intra-day time period, intra-week position, rhythm phase and interval length information to obtain a time coding result; Coordinate analysis and area quantization are performed on the space position in the structured event element sequence, and the latitude and longitude and the door position are reduced to the space features of floor, room, door type and adjacent area relationship to obtain a space coding result; Amplitude normalization, time alignment and short sequence aggregation are performed on the communication signal elements in the structured event element sequence, and the physical side statistics of signal strength change amplitude, stable interval length, rising and falling edge time, pulse count and jitter metric are extracted to obtain a communication signal coding result; The identity mode encoding result, the time encoding result, the space encoding result, and the communication signal encoding result are spliced and scale-aligned to form a single event vector, and an event embedding result is obtained; Format unification and storage layout normalization are performed on the event embedding result to determine the field order, value range, and storage alignment mode, and the event feature vector set is sorted according to the timestamp.

[0020] In this embodiment, the generation of the neighborhood aggregation result specifically includes: According to the event feature vector set, initial node lists of user nodes, device nodes, mode nodes, time nodes, and space nodes are respectively established, and corresponding indexes are generated for each event; Based on the initial node list, according to the co-occurrence relationship of the elements of the same event, the interaction relationship between the user nodes and the device nodes, the selection relationship between the user nodes and the mode nodes, the occurrence relationship between the device nodes and the time nodes, the positioning relationship between the device nodes and the space nodes, and the time period relationship between the mode nodes and the time nodes are generated, and the timestamp order, the attempt serial number, the success or failure mark, and the communication signal statistics of each relationship are recorded, and a relationship instance table is obtained; According to the timestamp, the relationship instance table is divided into fragments according to the continuous sliding time window, and a relationship fragment sequence organized according to the window number is formed; in each window, the relationship instances are respectively summarized according to the co-occurrence relationship type, and a time-varying multi-relationship graph structure containing user nodes, device nodes, mode nodes, time nodes, and space nodes is obtained; In each window, the relationship weight and the time sequence decay coefficient are calculated for each type of co-occurrence relationship, the relationship weight is determined according to the weighted results of the relationship occurrence frequency, the communication signal stability, and the time proximity, and the time sequence decay coefficient is calculated according to the time interval between the relationship and the end of the window, the effective relationship weight is obtained by multiplying the two, and written into the relationship weight table; Based on the relationship weight table, an adjacency index and a batch routing are established for each type of co-occurrence relationship, and the user nodes, device nodes, mode nodes, time nodes, and space nodes are divided into several subgraphs that can be processed in parallel, and a relationship adjacency index and a batch routing table are generated; In each window, the graph message passing and relationship aggregation are performed according to the batch routing table, the neighbor contribution of each type of co-occurrence relationship is weighted and summarized based on the relationship weight table, and a node temporary representation and a message set are obtained, wherein the node temporary representation is the weighted aggregation result of the neighbor node features of each type of co-occurrence relationship; The node temporary representation is merged and stabilized according to the time window, and the neighborhood aggregation result of the node at the current time is generated by using the window mean value and the recent window weighting strategy.

[0021] In this embodiment, the generation of the slow frequency memory state and the fast frequency memory state specifically includes: The dual-frequency memory time series graph neural network is constructed, including a dual-frequency memory layer, a gate fusion layer, a node embedding output layer, and an identification and anomaly judgment 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 end, a node-state index is established according to the neighborhood aggregation result, which 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 the behavior baseline of the last period are loaded from the cloud, and the fast-frequency memory unit snapshot retained during the last processing is read from the edge end, to generate an initial slow-frequency memory state and an initial fast-frequency memory state. The neighborhood aggregation result and the event feature vector set are aligned according to the timestamp and index, the embedding fields corresponding to the user identification, device identification, unlocking method, timestamp, spatial position, and communication signal elements are extracted, and a memory update message sequence is generated in chronological order. Based on the initial fast-frequency memory state, the memory update message sequence is written into the fast-frequency memory unit in turn, and the trigger manager performs gate writing, forgetting, and state advancing operations when each message is written. The gate writing filters the valid feature dimensions of the current event, the forgetting operation attenuates the weights of outdated or noise features, and the state advancing writes the immediate influence of the latest event on node behavior into the short-term state cache. After each memory update message is written, an intermediate fast-frequency memory state is generated, and the intermediate fast-frequency memory state sequence is formed by storing the intermediate states in chronological order. The fast-frequency memory unit takes the last intermediate state of the intermediate fast-frequency memory state sequence as the current fast-frequency memory state, representing the short-term immediate behavior response of the node at the current time. 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 sustained deviation in the behavior baseline, all memory update messages and corresponding fast-frequency memory state sequences within the window are collected, aggregated and smoothed within the window, and the stable feature mean and trend deviation within the window are calculated. Combined with the historical slow-frequency memory state, template fine-tuning and state correction are performed. The slow-frequency memory unit writes the stable behavior pattern and slow-changing features in the window aggregation result into the long-term template through an adaptive gate mechanism, updates the center distribution of the slow-frequency memory according to the trend deviation, and obtains the current slow-frequency memory state corresponding to the window number. The current fast-frequency memory state and the current slow-frequency memory state are aligned and de-duplicated according to the node-state index, written into the state cache to form the paired output of the fast-frequency memory state and the slow-frequency memory state, and a state update timestamp and a trigger marker are generated for each pair of output.

[0022] In this embodiment, the generation of the full-time domain node embedding vector specifically includes: Receiving slow frequency memory state and fast frequency memory state, and corresponding state update timestamp and trigger mark, combining neighborhood aggregation result and time node, mode node, space node to provide time context and scene information, establishing fusion batch index, getting fusion input set of gating fusion layer; Performing scale alignment, dimension alignment and missing completion on slow frequency memory state and fast frequency memory state in fusion input set, generating aligned memory state pair set; Based on state update timestamp and trigger mark, time period attribute corresponding to time node, unlocking mode attribute corresponding to mode node, and area attribute corresponding to space node, calculating time fluctuation index, behavior consistency score and signal disturbance intensity according to aligned memory state pair set, generating gating prior element table; The generation of the gating prior element table specifically includes: Extracting event time sequence of the same node in the current window from the neighborhood aggregation result and the state update timestamp, and calculating the dispersion of adjacent event intervals and the abnormal trigger proportion to obtain the time fluctuation index. With the slow frequency memory state and the behavior baseline as the reference, combining the unlocking mode, time period and area attributes of the mode node, time node and space node in the current window, the consistency degree of the current event set and the long-term template in the mode matching rate, time period matching rate and area matching rate is calculated to obtain the behavior consistency score. The jitter metric, signal strength change amplitude and stable interval length are extracted from the communication signal coding result. The jitter metric and the change amplitude are positively accumulated, and the stable interval length is inversely converted to obtain the signal disturbance intensity. After normalizing the time fluctuation index, the behavior consistency score and the signal disturbance intensity, the gating priority mark and the gating prior weight are generated according to the rules. When the time fluctuation index and the signal disturbance intensity are higher than the corresponding index threshold and the behavior consistency score is lower than the corresponding index threshold, the fast frequency priority is set and the gating prior weight is set to 0.9. When the time fluctuation index and the signal disturbance intensity are lower than the corresponding index threshold and the behavior consistency score is higher than the corresponding index threshold, the slow frequency priority is set and the low gating prior weight is set to 0.1. In other cases, the priority direction and the gating prior weight are determined according to the weighted results of the three, forming the gating prior element table; In the gating fusion layer, according to the gating prior element table and the aligned memory state pair set, a gating coefficient is generated for each node, and the gating coefficient sequence is output and corresponding relationship with the fusion batch index is established; According to the gating coefficient sequence, the aligned memory state pair set is weighted and combined node by node, and is subjected to stabilization and constraint processing, including range clipping, abnormal suppression and time smoothing, to suppress abnormal fluctuations, and generate candidate full-time domain node embedding vector; The candidate full-time domain node embedding vector is checked for consistency with the relationship weight table and the window number of the neighborhood aggregation result, and a slow frequency priority or fast frequency priority fallback strategy is triggered according to a gating coefficient sequence for a node that does not meet the consistency, and a full-time domain node embedding vector is output.

[0023] In the embodiment, the generation of the comprehensive identification decision specifically includes: The full-time domain node embedding vector is input into a node embedding output layer, dimension reduction mapping, range normalization and numerical clipping are performed in a preset field order, and a node representation vector set corresponding to an event index is generated; According to the event index, user node representations, device node representations, mode node representations, time node representations and space node representations are extracted from the node representation vector set and aligned with corresponding records in the event feature vector set to form an event alignment vector group; The event alignment vector group is input into an identification and anomaly judgment layer, the identification and anomaly judgment layer is composed of an identity matching branch, an anomaly risk branch, a forgery credibility branch and a fusion decision module, in the identity matching branch, an identity matching score is calculated based on the event alignment vector group, and an identity matching result and a threshold judgment mark are output, the identity matching score is generated by combined feature mapping of the user node representation, the device node representation, the mode node representation, the time node representation and the space node representation; In the anomaly risk branch, short-term fluctuation indicators, failure retry density and cross-region switching rate are extracted based on the event alignment vector group and the event feature vector set, and the anomaly risk score is calculated by combining the user node representation and the device node representation, and the anomaly risk result and the risk level mark are output, the anomaly risk score reflects the behavior stability and risk deviation trend of the node in the time window; In the forgery credibility branch, signal disturbance intensity, stable interval length conversion value and rising and falling edge duration ratio are extracted based on the communication signal encoding results in the event alignment vector group and the event feature vector set, and the forgery credibility score is calculated by combining the mode node representation and the space node representation, and the forgery credibility result and the suspicious type mark are output, the forgery credibility score measures the physical signal consistency and the naturalness of the unlocking behavior; The identity matching score, the anomaly risk score and the 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, and a comprehensive identification decision is generated according to the comprehensive identification score and a preset threshold table, the comprehensive identification decision includes three disposal levels of release, secondary verification and rejection, and the branch explanation and key feature abstract of the identity matching result, the anomaly risk result and the forgery credibility result are carried in the disposal result.

[0024] In the embodiment, the generation of the updated dual-frequency memory time sequence graph neural network specifically includes: Receiving the comprehensive identification decision, the treatment level and the explanatory label, executing instant response according to the treatment level at the edge, allowing to execute the unlocking action and recording the execution time if the comprehensive identification decision is release, triggering the user identity multi-factor verification module and suspending the unlocking response if the comprehensive identification decision is secondary verification, blocking the unlocking request and generating a security event log if the comprehensive identification decision is rejection; Recording operation feedback results at the edge after instant response, including unlocking action success or failure, verification pass or rejection, communication signal integrity and device end response delay, associating the feedback results with the comprehensive identification decision to form a treatment record table; Aligning the treatment record table with the node representation vector set, extracting the behavior pattern fields of the user node, device node, method node, time node and space node to form a behavior feedback feature set; Performing fast learning and memory calibration at the edge according to the behavior feedback feature set, triggering the fast frequency memory unit update operation when there is a deviation between the feedback result and the comprehensive identification decision, adjusting the short-term state cache corresponding to the node, updating the short-term behavior pattern and signal response rule, recalculating the fast frequency memory state of the node, and outputting the updated fast frequency memory state; Packing the updated fast frequency memory state generated by the edge and the behavior feedback feature set according to the node index and time window number, uploading to the cloud memory synchronization module, matching the historical slow frequency memory unit template and behavior baseline according to the time window number, calculating the feedback offset and trend change degree; Performing template update and behavior baseline correction in the slow frequency memory unit according to the feedback offset and trend change degree, writing the change trend of the offset feature into the slow frequency memory unit template, deleting abnormal disturbance samples and updating the slow frequency memory state to form a slow frequency memory unit correction template for the current window; Reintegrating the slow frequency memory unit correction template and the updated fast frequency memory state of the edge into a synchronization package, replacing the old state cache and fast frequency memory unit snapshot at the edge to form an updated dual-frequency memory time sequence graph neural network for the next intelligent lock identification operation.

[0025] Embodiment 1: In order to verify the feasibility of the application in implementation, the application is applied to the intelligent door lock group management scene of a large residential area in a certain coastal city. The residential area is equipped with multiple brands and multiple models of networked intelligent locks, and the user group is complex, including homeowners, property, cleaning, maintenance, visitors and express delivery and other types of access subjects. In the past, door lock identification has been dependent on single identity verification or static records, and in the scene of frequent switching of multiple users, communication signal fluctuations, and significant differences in time period behavior, problems such as authorized misjudgment, delayed release and false alarms often occur, especially in the frequent access of visitors and cross-regional scheduling scenes, the traditional identification method is difficult to balance sensitivity and stability. The application accesses multi-source event data from the door lock control end, the mobile end and the property platform in this scene, covering user identification, device identification, unlocking method, timestamp, spatial location and communication signal, and unified formatting and embedded coding are carried out at the edge to form a multi-dimensional event feature vector set.

[0026] In actual application process, through the time-varying multi-relationship graph structure constructed by the application, the user, device, method, time and space information are unified into the graph model, and the context association of unlocking behavior is dynamically captured. The system automatically identifies the interaction between different nodes, such as the stable binding relationship between the owner and the self device, the time period mode of the property personnel and the multi-point access, the communication characteristics of the visitor and the temporary authorization path, etc. In the graph neural network layer, through the neighborhood aggregation and weight decay mechanism, the parallel modeling of short-term access and long-term behavior is realized. The dual-frequency memory time series graph neural network runs continuously at the edge, the fast-frequency memory unit captures the real-time behavior change in real time, such as the user's unlocking method switching or signal disturbance anomaly; the slow-frequency memory unit is periodically updated in the cloud, continuously learning the stable work and behavior baseline of the user, forming the behavior template and trend distribution.

[0027] In the continuous operation of multiple days, the gating fusion layer of the application dynamically adjusts the fusion ratio of fast and slow frequency memories according to the time context and scene conditions, so that the model still maintains high recognition accuracy in the case of frequent access at night, device signal attenuation or user moving track change. The recognition and anomaly judgment layer simultaneously outputs three types of results of identity matching, abnormal risk and forgery credibility, and completes the immediate response at the edge: directly releasing for the user with normal recognition, triggering secondary verification for the user with behavior deviation but still with high credibility, and executing blocking and alarm for the case of signal forgery or cross-device anomaly. Subsequently, the cloud adjusts the slow-frequency memory template according to the recognition and feedback results, so that the model gradually learns the individualized mode of different users in time period, space and method, and realizes continuous optimization.

[0028] Through analysis of the time sequence behavior data continuously collected by multiple intelligent locks in the same residential area, the dual-frequency memory time sequence graph neural network constructed by the application significantly reduces the cross-region misrecognition rate and the false signal misjudgment rate, and improves the stability and delay tolerance of the recognition response. During the overall operation, the intelligent lock recognition remains stable output in a high-concurrency and complex signal environment, and there is no locking error or recognition stagnation, verifying the effectiveness and applicability of the method in multi-source data fusion, safe recognition and self-learning update.

[0029] Table 1 Performance comparison of intelligent lock recognition method based on time sequence graph neural network and traditional method

[0030] As can be seen from Table 1, the dual-frequency memory time sequence graph neural network method proposed by the application shows significant improvement over the traditional graph neural network method in multiple key indicators. The recognition accuracy is improved 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 behavior, which enables the model to capture both stable patterns and immediate changes. The abnormality detection recall rate is improved from 84.7% to 93.5%, indicating that the mechanism of capturing sudden behavior through the fast-frequency memory unit and correcting the long-term template through the slow-frequency memory unit effectively reduces the risk of missed detection. The false signal recognition rate is improved from 82.5% to 94.2%, thanks to the dynamic weight distribution strategy of the gating fusion layer between behavior consistency and signal disturbance features. In terms of response performance, the average response time is reduced from 245 milliseconds to 186 milliseconds, reflecting the advantages of edge fast reasoning and dual-frequency synchronization mechanism in low-latency processing. The cross-device misrecognition rate is significantly reduced from 4.6% to 1.2%, indicating that the model has higher discrimination accuracy in node feature alignment and context determination. In addition, the long-time running stability is improved from 72 hours to 168 hours, verifying the robustness and maintainability of the dual-frequency memory structure in continuous learning and cloud-edge collaborative update. In summary, the application is significantly superior to the traditional method in recognition accuracy, abnormality detection, response efficiency and long-term stability, etc., reflecting the comprehensive advantages of the model structure design and time sequence feature fusion strategy.

[0031] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can make equivalent substitutions or changes to the technical solution and inventive concept of the application within the technical scope disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A smart lock recognition method based on a time sequence graph neural network, characterized in that, The method comprises the following steps: Obtain multi-source event data in the operation process of the smart lock, unify the format and embed the code to obtain an event feature vector set; According to the event feature vector set, a time-varying multi-relation graph structure is constructed, user nodes, device nodes, mode nodes, time nodes and space nodes are associated, graph message passing and neighborhood feature aggregation are performed, and neighborhood aggregation results of the nodes at the current time are obtained; A dual-frequency memory time sequence graph neural network is constructed, the neighborhood aggregation results are input into the dual-frequency memory layer, the memory state is updated in the slow-frequency memory unit and the fast-frequency memory unit respectively, and 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 a gate fusion layer, and the time context and the scene information are adaptively fused through a gate function to generate a full-time domain node embedding vector; The full-time domain node embedding vector is input into a node embedding output layer, mapped into a low-dimensional node representation vector, and passed through an identification and anomaly judgment layer to calculate an identity matching result, an abnormal risk result and a forgery credibility result, and output a comprehensive identification decision; According to the comprehensive identification decision, instant response and safety disposal are performed on the edge side, and the slow-frequency memory unit and the behavior baseline are updated on the cloud side to form an updated dual-frequency memory time sequence graph neural network.

2. The intelligent lock recognition method based on a time sequence graph neural network according to claim 1, characterized in that, The generation of the event feature vector set specifically comprises: Receiving multi-source event data in the operation process of the smart lock, reorganizing user identification, device identification, unlocking mode, timestamp, spatial position and communication signal elements into a structured event element sequence according to a unified field order; Performing discrete mapping and sparse coding on the user identification, device identification and unlocking mode in the structured event element sequence to obtain an identity mode coding result; Performing time decomposition, period expansion and phase mapping on the timestamp in the structured event element sequence to convert the timestamp into a time feature containing intra-day time period, intra-week position, rhythm phase and interval length information to obtain a time coding result; Performing coordinate analysis and area quantization on the spatial position in the structured event element sequence to reduce the latitude and longitude and door point position into the spatial features of floor, room, door type and adjacent area relationship to obtain a spatial coding result; Performing amplitude normalization, time alignment and short sequence aggregation on the communication signal elements in the structured event element sequence to extract the physical side statistics of signal strength change amplitude, stable interval length, rising and falling edge time, pulse count and jitter metric to obtain a communication signal coding result; Splicing and scaling the identity mode coding result, the time coding result, the spatial coding result and the communication signal coding result to form a single event vector to obtain an event embedding result; Performing format unification and storage layout standardization on the event embedding result to determine the field order, value range and storage alignment method, and arranging according to the timestamp to obtain an event feature vector set.

3. The intelligent lock recognition method based on a time sequence graph neural network according to claim 1, characterized in that, The generation of the neighborhood aggregation result specifically comprises: According to the event feature vector set, initial node lists of user nodes, device nodes, mode nodes, time nodes and space nodes are respectively established, and an index corresponding to each event is generated. Based on the initial node list, the interaction relationship between the user node and the device node, the selection relationship between the user node and the method node, the occurrence relationship between the device node and the time node, the positioning relationship between the device node and the space node, and the time period relationship between the method node and the time node are generated according to the co-occurrence relationship of the elements of the same event, and the timestamp order, the attempt number, the success or failure mark, and the communication signal statistics of each relationship are recorded to obtain a relationship instance table; The relationship instance table is divided into fragments according to the continuous sliding time window in the order of the timestamp to form a relationship fragment sequence organized by window number; the relationship instances are summarized according to the co-occurrence relationship type in each window to obtain a time-varying multi-relationship graph structure containing user nodes, device nodes, method nodes, time nodes, and space nodes; In each window, the relationship weight and the time sequence decay coefficient are calculated for each type of co-occurrence relationship element, the relationship weight is determined according to the weighted results of the relationship occurrence frequency, the communication signal stability, and the time proximity, and the time sequence decay coefficient is calculated according to the time interval between the relationship and the end of the window, and the effective relationship weight is obtained by multiplying the two to be written into the relationship weight table; Based on the relationship weight table, an adjacency index and a batch routing are established for each type of co-occurrence relationship element, and the user nodes, device nodes, method nodes, time nodes, and space nodes are divided into several subgraphs that can be processed in parallel to generate a relationship adjacency index and a batch routing table; In each window, the graph message passing and relationship aggregation are performed according to the batch routing table, the neighbor contribution of each type of co-occurrence relationship element is weighted and summarized based on the relationship weight table, and the node temporary representation and the message set are obtained, wherein the node temporary representation is the weighted aggregation result of the neighbor node features of each type of co-occurrence relationship element; The node temporary representation is merged and stabilized according to the time window, and the neighbor aggregation result of the node at the current time is generated by using the window mean value and the recent window weighting strategy.

4. The intelligent lock recognition method based on a time sequence 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 sequence graph neural network is constructed, including a dual-frequency memory layer, a gate fusion layer, a node embedding output layer, and an identification and anomaly judgment 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; A node-state index is established at the edge end according to the neighbor 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 the behavior baseline of the last period are loaded from the cloud, and the fast frequency memory unit snapshot retained during the last processing is read from the edge end to generate an initial slow frequency memory state and an initial fast frequency memory state; The neighbor aggregation result and the event feature vector set are aligned according to the timestamp and the index, the embedding fields corresponding to the user identification, the device identification, the unlocking method, the timestamp, the space position, and the communication signal element are extracted, and a memory update message sequence is generated in chronological order; Based on the initial fast-frequency memory state, a sequence of memory update messages is written into the fast-frequency memory unit in turn, triggering the manager to perform gated writing, forgetting and state advancing operations when each message is written, the gated writing filters the valid feature dimensions of the current event, the forgetting operation attenuates the weights of outdated or noise features, and the state advancing writes the immediate influence of the latest event on node behavior into the short-term state cache, after each memory update message is written, an intermediate fast-frequency memory state is generated, and the intermediate fast-frequency memory state sequence is formed by storing the intermediate fast-frequency memory states in chronological order, the fast-frequency memory unit takes the intermediate state of the last intermediate fast-frequency memory state in the sequence as the current fast-frequency memory state, representing the short-term immediate behavior response of the node at the current time, and finally outputs the current fast-frequency memory state and the set of intermediate fast-frequency memory states; Based on the initial slow-frequency memory state, when the trigger manager detects the end of the time window or the sustained deviation of the behavior baseline, all memory update messages and corresponding fast-frequency memory state sequences within the window are collected, aggregated and smoothed within the window, the stable feature mean and trend deviation within the window are calculated, the template fine-tuning and state correction are performed in combination with the historical slow-frequency memory state, and the stable behavior pattern and slow-changing feature in the window aggregation result are written into the long-term template through the adaptive gating mechanism, and the center distribution of the slow-frequency memory is updated according to the trend deviation, obtaining the current slow-frequency memory state corresponding to the window number; The current fast-frequency memory state and the current slow-frequency memory state are aligned and de-duplicated according to the node-state index, written into the state cache to form the paired output of the fast-frequency memory state and the slow-frequency memory state, and a state update timestamp and a trigger marker are generated for each pair of output.

5. The intelligent lock recognition method based on a time sequence graph neural network according to claim 1, characterized in that, The generation of the full-time-domain node embedding vector specifically includes: Receiving slow-frequency memory states and fast-frequency memory states and corresponding state update timestamps and trigger markers, combining the neighborhood aggregation result and the time node, the mode node and the space node to provide temporal context and scene information, establishing a fusion batch index to obtain a fusion input set of the gating fusion layer; Performing scale alignment, dimension alignment and missing completion on the slow-frequency memory states and fast-frequency memory states in the fusion input set to generate a set of aligned memory state pairs; Based on the state update timestamp and the trigger marker, the time period attribute corresponding to the time node, the unlocking mode attribute corresponding to the mode node, and the area attribute corresponding to the space node, the time fluctuation index, the behavior consistency score and the signal disturbance intensity are calculated according to the set of aligned memory state pairs to generate a table of gated prior elements; In the gating fusion layer, a gating coefficient is generated for each node according to the table of gated prior elements and the set of aligned memory state pairs, and the gating coefficient sequence is output and corresponding relationship with the fusion batch index is established; The set of aligned memory state pairs is weighted and combined node by node according to the gating coefficient sequence, and is subjected to stabilization and constraint processing, including range clipping, abnormality suppression and time smoothing, to suppress abnormal fluctuations and generate a candidate full-time-domain node embedding vector; The candidate full-time domain node embedding vector is checked for consistency with the relationship weight table and the window number of the neighborhood aggregation result, and a slow frequency priority or fast frequency priority fallback strategy is triggered according to a gating coefficient sequence for a node that does not meet the consistency, and a full-time domain node embedding vector is output.

6. The intelligent lock recognition method based on a time sequence graph neural network according to claim 1, characterized in that, The generation of the comprehensive identification decision specifically includes: The full-time domain node embedding vector is input into a node embedding output layer, dimension reduction mapping, range normalization and numerical clipping are performed in a preset field order, and a node representation vector set corresponding to the event index is generated; According to the event index, user node representations, device node representations, mode node representations, time node representations and space node representations are extracted from the node representation vector set, and are aligned with corresponding records in the event feature vector set to form an event alignment vector group; The event alignment vector group is input into an identification and anomaly judgment layer, and the identification and anomaly judgment layer is composed of an identity matching branch, an abnormal risk branch, a forgery credibility branch and a fusion decision module. In the identity matching branch, an identity matching score is calculated based on the event alignment vector group, and an identity matching result and a threshold judgment mark are output; In the abnormal risk branch, short-term fluctuation indicators, failure retry density and cross-region switching rate are extracted based on the event alignment vector group and the event feature vector set, and the abnormal risk score is calculated combined with the user node representation and the device node representation, and the abnormal risk result and the risk level mark are output; In the forgery credibility branch, signal disturbance intensity, stable interval length conversion value and rising and falling edge time length ratio are extracted based on the communication signal coding result in the event alignment vector group and the event feature vector set, and the forgery credibility score is calculated combined with the mode node representation and the space node representation, and the forgery credibility result and the suspicious type mark are output; The identity matching score, the abnormal risk score and the forgery credibility score are input into the fusion decision module to generate a comprehensive identification score, which is formed by the weighted results output by the three branches, and a comprehensive identification decision is generated according to the comprehensive identification score and a preset threshold table.

7. The intelligent lock recognition method based on a time sequence graph neural network according to claim 1, characterized in that, The generation of the updated dual-frequency memory time sequence graph neural network specifically includes: The comprehensive identification decision, the treatment level and the explanatory mark are received, and an immediate response is performed according to the treatment level at the edge, if the comprehensive identification decision is release, the unlocking action is allowed to be performed and the execution time is recorded, if the comprehensive identification decision is secondary verification, the user identity multi-factor verification module is triggered and the unlocking response is suspended, and if the comprehensive identification decision is rejection, the unlocking request is blocked and a security event log is generated; After the immediate response, the operation feedback result is recorded by the edge, including the success or failure of the unlocking action, the passing or rejection of the verification, the communication signal integrity and the device end response delay, the feedback result is associated with the comprehensive identification decision to form a treatment record table; The treatment record table and the node representation vector set are aligned, the behavior mode fields of the user node, the device node, the mode node, the time node and the space node are extracted, and a behavior feedback feature set is formed; Performing fast learning and memory calibration according to the behavior feedback feature set at the edge end, triggering the fast frequency memory unit update operation when the feedback result deviates from the comprehensive recognition decision, adjusting the short-term state cache corresponding to the node, updating the short-term behavior mode and signal response law, recalculating the fast frequency memory state of the node, and outputting the updated fast frequency memory state; Packing the updated fast frequency memory state generated at the edge end and the behavior feedback feature set according to the node index and the time window number, uploading them to the cloud memory synchronization module, matching the historical slow frequency memory unit template and the behavior baseline according to the time window number, calculating the feedback offset and the trend change degree; Performing template update and behavior baseline correction in the slow frequency memory unit according to the feedback offset and the trend change degree, writing the change trend of the offset feature into the slow frequency memory unit template, deleting abnormal disturbance samples and updating the slow frequency memory state, and forming the slow frequency memory unit correction template of the current window; Reintegrating the slow frequency memory unit correction template and the updated fast frequency memory state at the edge end into a synchronization package, issuing it to the edge end to replace the old state cache and fast frequency memory unit snapshot, and forming the updated dual-frequency memory time sequence graph neural network for the next intelligent lock recognition operation.

Images