CN-121482897-B - Intelligent lock identification method based on time sequence diagram neural network

Abstract

The invention discloses an intelligent lock identification method based on a time sequence diagram neural network, which comprises the following steps of collecting multisource event data of an intelligent lock, unifying formats and embedding codes to generate an event feature vector set, constructing a time-varying multiple relation diagram, executing message transmission and feature aggregation to obtain a neighborhood aggregation result, constructing a double-frequency memory time sequence diagram neural network, extracting long-term and short-term behavior features from a slow-frequency and fast-frequency memory unit, inputting the slow-frequency and fast-frequency memory states into a gating fusion layer, combining time and scene information to generate a full-time-domain embedded vector, calculating identity matching, abnormal risk and falsification credibility, outputting a comprehensive identification result, executing instant response and safety treatment according to the identification result, and carrying out network updating. The invention utilizes the time sequence diagram neural network and the double-frequency memory mechanism to realize the dynamic identification of multiple elements of the intelligent lock, and has the advantages of strong self-learning, high identification accuracy and excellent safety and robustness.

Inventors

• DAI ZHIBO
• ZHAO JINGCHAO
• ZHANG LEI

Assignees

• 北京国金源富科技有限公司

Dates

Publication Date

20260508

Application Date

20251127

Claims (7)

1. 1. The intelligent lock identification method based on the time sequence diagram neural network is characterized by comprising the following steps of: the method comprises the steps of obtaining multi-source event data in the running process of the intelligent lock, and carrying out format unification and embedded coding to obtain an event feature vector set; Constructing a time-varying multi-relation graph structure according to the event feature vector set, establishing association among user nodes, equipment nodes, mode nodes, time nodes and space nodes, and executing graph message transfer and neighborhood feature aggregation to obtain a neighborhood aggregation result of the nodes at the current moment; constructing a dual-frequency memory timing diagram neural network, inputting the neighborhood aggregation result into a dual-frequency memory layer, updating the memory state in a slow frequency memory unit and a fast frequency memory unit respectively, and outputting the slow frequency memory state and the fast frequency memory state; inputting the slow frequency memory state and the fast frequency memory state into a gating fusion layer, and carrying out self-adaptive fusion through a gating function according to the time context and scene information to generate a full-time domain node embedded vector; Embedding the full-time domain node embedded vector input node into the output layer, mapping the full-time domain node embedded vector input node into a low-dimensional node characterization vector, calculating an identity matching result, an abnormal risk result and a falsification credibility result through the identification and abnormality judgment layer, and outputting a comprehensive identification decision; And performing instant response and safety treatment on the edge end according to the comprehensive recognition decision, and updating the slow frequency memory unit and the behavior base line at the cloud end to form an updated dual-frequency memory time sequence diagram neural network.
2. 2. The intelligent lock identification method based on the time sequence diagram neural network according to claim 1, wherein the generating of the event feature vector set specifically comprises: receiving multi-source event data in the intelligent lock operation process, and reorganizing a user identifier, a device identifier, an unlocking mode, a time stamp, a space position and a communication signal element into a structured event element sequence according to a unified field sequence; Performing discrete mapping and sparse coding on a user identifier, a device identifier and an 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 time stamp in the structured event element sequence, and converting the time stamp into a time characteristic containing intra-day time period, intra-week position, rhythm phase and interval length information to obtain a time coding result; carrying out coordinate analysis and region quantization on the spatial positions in the structured event element sequence, and classifying longitude and latitude and door points into spatial features of the relationship between floors, rooms, door categories and adjacent regions to obtain a spatial coding result; Performing amplitude normalization, time alignment and short sequence aggregation on communication signal elements in the structured event element sequence, extracting physical side statistics of signal intensity variation amplitude, stable interval length, rising and falling edge duration, pulse count and jitter measurement, and obtaining a communication signal coding result; splicing and aligning the identity mode coding result, the time coding result, the space coding result and the communication signal coding result to form a single event vector, and obtaining an event embedding result; and executing format unification and storage layout standardization on the event embedding result, determining a field sequence, a numerical range and a storage alignment mode, and sequentially sorting the field sequence, the numerical range and the storage alignment mode into an event feature vector set according to the time stamp.
3. 3. The intelligent lock identification method based on the time sequence diagram neural network according to claim 1, wherein the generation of the neighborhood aggregation result specifically comprises: respectively establishing initial node lists of user nodes, equipment nodes, mode nodes, time nodes and space nodes according to the event feature vector set, and generating corresponding indexes for each event; Based on the initial node list, generating an interaction relation between a user node and a device node, a selection relation between the user node and a mode node, an occurrence relation between the device node and a time node, a positioning relation between the device node and a space node and a time period relation between the mode node and the time node according to element co-occurrence relation of the same event, and recording a time stamp sequence, an attempt sequence number, success or failure marks and communication signal statistics of each relation to obtain a relation example table; The time stamp is used as an order, and the relation instance table is sliced according to a continuous sliding time window to form a relation slicing sequence organized according to window numbers; respectively summarizing the relation examples according to the element co-occurrence relation types in each window to obtain a time-varying multi-relation graph structure comprising user nodes, equipment nodes, mode nodes, time nodes and space nodes; Calculating a relation weight and a time sequence attenuation coefficient for each type of element co-occurrence relation in each window, wherein the relation weight is determined according to the relation occurrence frequency and the weighting result of the communication signal stability and the time proximity, the time sequence attenuation coefficient is calculated according to the relation and the time interval of the window tail end, and the relation weight and the time sequence attenuation coefficient are multiplied to obtain an effective relation weight and written into a relation weight table; establishing an adjacency index and a batch route for the co-occurrence relation of each type of elements based on the relation weight table, dividing user nodes, equipment nodes, mode nodes, time nodes and space nodes into a plurality of sub-graphs which can be processed in parallel, and generating a relation adjacency index and a batch route table; performing graph message transfer and relationship aggregation in each window according to a batch routing table, and carrying out weighted summarization on neighbor contributions of the co-occurrence relationships of each type of elements based on a relationship weight table to obtain a node temporary representation and a message set, wherein the node temporary representation is a weighted aggregation result of neighbor node characteristics of the co-occurrence relationships of various types of elements; And merging and stabilizing the temporary node representation according to a time window, and generating a neighborhood aggregation result of the node at the current moment by adopting a weighted strategy of an average value in the window and a recent window.
4. 4. The intelligent lock identification method based on the time sequence diagram neural network according to claim 1, wherein the generation of the slow frequency memory state and the fast frequency memory state specifically comprises the following steps: the method comprises the steps of constructing a dual-frequency memory time sequence diagram neural network, wherein the dual-frequency memory layer comprises a slow-frequency memory unit, a fast-frequency memory unit, a trigger manager, a memory read-write interface and a state buffer; establishing node-state indexes at the edge end according to the neighborhood aggregation result, wherein the node-state indexes are used for recording storage positions, access paths 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, loading a slow frequency memory unit template and a behavior baseline of the previous period from the cloud, and reading a fast frequency memory unit snapshot reserved in the last processing from the edge end to generate an initial slow frequency memory state and an initial fast frequency memory state; Aligning the neighborhood aggregation result with the event feature vector set according to the time stamp and the index, extracting the embedded fields corresponding to the user identifier, the equipment identifier, the unlocking mode, the time stamp, the spatial position and the communication signal element, and generating a memory update message sequence according to the time sequence; Based on the initial fast frequency memory state, writing a memory update message sequence into a fast frequency memory unit in sequence, triggering a manager to execute gating writing, forgetting and state pushing operation when each message is written, wherein the gating writing screens the effective characteristic dimension of a current event, the forgetting operation carries out weight attenuation on outdated or noise characteristics, the state pushing writes the instant influence of the latest event on node behaviors into a short-term state cache, each memory update message is written to generate an intermediate fast frequency memory state, the intermediate fast frequency memory state is stored according to time sequence to form an intermediate fast frequency memory state sequence, the fast frequency memory unit takes the intermediate state of the last strip of the intermediate fast frequency memory state sequence as the current fast frequency memory state, represents the short-term instant behavior response of the node at the current moment, and finally outputs the current fast frequency memory state and an intermediate fast frequency memory state set; Based on the initial slow frequency memory state, when a trigger manager detects that a time window is over or a behavior base line continuously deviates, collecting all memory update messages and corresponding fast frequency memory state sequences in the window, carrying out intra-window aggregation and smoothing treatment on the memory update messages, calculating stable characteristic mean values and trend deviation values in the window, carrying out template fine adjustment and state correction by combining the historical slow frequency memory states, writing stable behavior patterns and slow variation characteristics in a window aggregation result into a long-term template by a slow frequency memory unit through a self-adaptive gating mechanism, and updating the central distribution of the slow frequency memory according to the trend deviation values to obtain the current slow frequency memory state corresponding to a 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, the writing state is cached to form paired output of the fast frequency memory state and the slow frequency memory state, and a state update time stamp and a trigger mark are generated for each pair of output.
5. 5. The intelligent lock identification method based on the time sequence diagram neural network according to claim 1, wherein the generation of the full time domain node embedded vector specifically comprises: Receiving a slow frequency memory state, a fast frequency memory state, a corresponding state update time stamp and a trigger mark, and establishing a fusion batch index by combining a neighborhood aggregation result with time context and scene information provided by a time node, a mode node and a space node to obtain a fusion input set of a gating fusion layer; Performing scale alignment, dimension alignment and missing complement on the slow frequency memory state and the fast frequency memory state in the fusion input set to generate an aligned memory state pair set; Based on the state update time stamp and the trigger mark, the time period attribute corresponding to the time node, the unlocking mode attribute corresponding to the mode node and the region attribute corresponding to the space node, calculating a time fluctuation index, a behavior consistency score and a signal disturbance intensity according to the aligned memory state pair set, and generating a gating prior element table; generating gating coefficients for each node according to the gating prior element table and the aligned memory state pair set at a gating fusion layer, outputting a gating coefficient sequence and establishing a corresponding relation with a fusion batch index; The method comprises the steps of carrying out weighted combination on a set node by node according to a memory state aligned by a gating coefficient sequence, and carrying out stabilization and constraint processing, wherein the processes comprise range cutting, abnormal suppression and time smoothing, and abnormal fluctuation suppression, so as to generate candidate full-time domain node embedded vectors; and checking consistency of the candidate full-time domain node embedded vector, a relation weight table and window numbers of the neighborhood aggregation result, triggering a rollback strategy of slow frequency priority or fast frequency priority according to a gating coefficient sequence for nodes which do not meet consistency, and outputting the full-time domain node embedded vector.
6. 6. The intelligent lock identification method based on the time sequence diagram neural network according to claim 1, wherein the generating of the comprehensive identification decision specifically comprises: embedding all-time domain nodes into a vector input node embedded output layer, performing dimension reduction mapping, range normalization and numerical clipping according to a preset field sequence, and generating a node characterization vector set corresponding to an event index; Extracting user node characterization, equipment node characterization, mode node characterization, time node characterization and space node characterization from the node characterization vector set according to the event index, and aligning with corresponding records in the event feature vector set to form an event alignment vector group; inputting the event alignment vector group into an identification and abnormality judgment layer, wherein the identification and abnormality judgment layer consists of an identity matching branch, an abnormality risk branch, a fake credibility branch and a fusion decision module, and in the identity matching branch, an identity matching score is calculated on the basis of an event pair Ji Xiangliang group, and an identity matching result and a threshold judgment mark are output; At the abnormal risk branch, short-term fluctuation indexes, failure retry density and cross-region switching rate are extracted based on an event alignment vector set and an event feature vector set, the abnormal risk score is calculated by combining user node characterization and equipment node characterization, and an abnormal risk result and a risk grade mark are output; In the fake credibility branch, extracting signal disturbance intensity, a stable interval length conversion value and a rising edge duration ratio based on communication signal coding results in an event alignment vector group and an event feature vector set, calculating fake credibility score by combining mode node characterization and space node characterization, and outputting a fake credibility result and suspicious type marks; The identity matching score, the abnormal risk score and the counterfeit credibility score are input into a fusion decision module to generate a comprehensive identification score, the comprehensive identification score is formed by weighted results output by three branches, and a comprehensive identification decision is generated according to the comprehensive identification score and a preset threshold value table.
7. 7. The intelligent lock identification method based on the time sequence diagram neural network according to claim 1, wherein the generation of the updated dual-frequency memory time sequence diagram neural network specifically comprises: Receiving a comprehensive identification decision, a treatment grade and an explanatory mark, executing instant response at an edge end according to the treatment grade, allowing unlocking action to be executed and recording execution time if the comprehensive identification decision is released, triggering a user identity multi-factor verification module and suspending 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 refused; Recording operation feedback results by the edge end after instant response, wherein the operation feedback results comprise success or failure of unlocking action, passing or rejecting verification, integrity of communication signals and response delay of the equipment end, and associating the feedback results with comprehensive recognition decisions to form a treatment record table; aligning the treatment record table with the node characterization vector set, and extracting behavior mode fields of user nodes, equipment nodes, mode nodes, time nodes and space nodes to form a behavior feedback feature set; Performing fast learning and memory calibration at the edge according to the behavior feedback feature set, triggering a fast frequency memory unit to update when a feedback result deviates from a comprehensive recognition decision, adjusting a short-term state buffer corresponding to the node, updating a short-term behavior mode and a signal response rule, recalculating a fast frequency memory state of the node, and outputting the updated fast frequency memory state; The updated fast frequency memory state and behavior feedback feature set generated by the edge end are packaged according to the node index and the time window number and then uploaded to a cloud memory synchronization module, a historical slow frequency memory unit template and a behavior base line are matched according to the time window number, and feedback offset and trend change degree are calculated; Performing template updating and behavior baseline correction according to the feedback offset and the trend change degree in the slow frequency memory unit, writing the change trend of the offset characteristic into the slow frequency memory unit template, deleting an abnormal disturbance sample and updating the slow frequency memory state to form a slow frequency memory unit correction template of the current window; And re-integrating the slow frequency memory unit correction template and the updated fast frequency memory state of the edge end into a synchronous packet, and transmitting the synchronous packet to the edge end to replace the old state buffer and the fast frequency memory unit snapshot to form an updated double-frequency memory time sequence diagram neural network for the next intelligent lock identification operation.

Description

Intelligent lock identification method based on time sequence diagram neural network Technical Field The invention relates to the field of security and protection identification, in particular to an intelligent lock identification method based on a time sequence diagram neural network. Background The existing intelligent lock identification technology mainly uses static identity authentication, usually relies on fingerprint, password, face or Bluetooth signal and other single characteristics to carry out matching judgment, and has obvious defects in terms of safety and stability. Partial improvement schemes attempt to introduce data modeling based on time sequences or device behaviors, and dynamic identification is realized through deep learning or a time sequence neural network, but the methods are often only used for modeling input data with single dimension, and lack of joint description of complex interaction relations among multiple elements such as users, devices, modes, time and space, and the like, so that the adaptability of the model to environmental changes and multi-user scenes is poor. In addition, the existing scheme generally depends on cloud centralized processing, real-time events at the edge end are difficult to respond in time, and recognition delay is high. In the field of dynamic behavior modeling and anomaly identification, a combination mode of a graph neural network and a long-term and short-term memory network has been proposed to capture time dependency or periodic characteristics among nodes, but the methods generally adopt a single-frequency time update mechanism, only a certain time scale of behavior change can be reflected, and short-term fluctuation and long-term regularity cannot be simultaneously considered. For high-frequency interaction equipment such as intelligent locks, the existing model is difficult to balance between rapidly-changing unlocking events and long-term user habits, so that abnormal detection is easy to misjudge. In addition, due to the lack of an edge-end instant feedback and cloud behavior baseline collaborative updating mechanism, the model cannot form self-learning and cross-period evolution capability, and the reliability and the robustness of the model in a multi-scene and multi-user dynamic environment are limited. Therefore, how to provide an intelligent lock identification method based on a neural network of a timing diagram is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide an intelligent lock identification method based on a time sequence diagram neural network, which utilizes the time sequence diagram neural network and a double-frequency memory mechanism to realize the dynamic identification of multiple elements of an intelligent lock and has the advantages of strong self-learning, high identification accuracy and excellent safety robustness. According to the embodiment of the invention, the intelligent lock identification method based on the time sequence diagram neural network comprises the following steps of: the method comprises the steps of obtaining multi-source event data in the running process of the intelligent lock, and carrying out format unification and embedded coding to obtain an event feature vector set; Constructing a time-varying multi-relation graph structure according to the event feature vector set, establishing association among user nodes, equipment nodes, mode nodes, time nodes and space nodes, and executing graph message transfer and neighborhood feature aggregation to obtain a neighborhood aggregation result of the nodes at the current moment; constructing a dual-frequency memory timing diagram neural network, inputting the neighborhood aggregation result into a dual-frequency memory layer, updating the memory state in a slow frequency memory unit and a fast frequency memory unit respectively, and outputting the slow frequency memory state and the fast frequency memory state; inputting the slow frequency memory state and the fast frequency memory state into a gating fusion layer, and carrying out self-adaptive fusion through a gating function according to the time context and scene information to generate a full-time domain node embedded vector; Embedding the full-time domain node embedded vector input node into the output layer, mapping the full-time domain node embedded vector input node into a low-dimensional node characterization vector, calculating an identity matching result, an abnormal risk result and a falsification credibility result through the identification and abnormality judgment layer, and outputting a comprehensive identification decision; And performing instant response and safety treatment on the edge end according to the comprehensive recognition decision, and updating the slow frequency memory unit and the behavior base line at the cloud end to form an updated dual-frequency memory time sequence diagram neural network. Optionally,