Smart home anti-theft system based on smart mobile terminal

Abstract

The invention relates to the technical field of smart home and mobile terminals, and discloses a smart home anti-theft system based on a smart mobile terminal, which comprises a dynamic behavior learning module, a multi-modal sensor collaborative network, a dual-channel verification unit and a cross-platform emergency linkage engine. According to the smart home anti-theft system based on the smart mobile terminal, a false alarm scene is effectively eliminated and the false alarm rate is reduced through a multi-modal sensor collaborative network and a dual-channel verification unit, and a personalized security policy is dynamically generated by adopting a mixed learning algorithm and an edge-cloud collaborative computing architecture, so that the privacy data of a user is protected; through the cross-platform emergency linkage engine, hierarchical alarm and multi-dimensional response are supported, the intrusion event processing efficiency is improved, it is ensured that an evidence chain cannot be tampered, judicial evidence obtaining standards are met, and the problem that evidence of a traditional system is prone to being lost or tampered is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of smart homes and mobile terminals, and in particular to a smart home anti-theft system based on a smart mobile terminal. Background Art

[0002] Smart home security system has many important impacts. It uses advanced technologies such as artificial intelligence, big data analysis, cloud computing, etc., and through intelligent monitoring equipment and sensors, it can monitor and identify various security threats in real time, such as intrusion, fire, gas leakage, high-altitude throwing, etc., and can automatically trigger alarms and emergency notifications, so that users can take timely actions to protect themselves and their property. The smart home security system maintains the residential function of traditional houses, gets rid of the passive mode, and is active and intelligent. It provides a full range of information exchange functions, optimizes people's lifestyles and living environments, and realizes home appliance control, lighting control, indoor and outdoor remote control, curtain intelligent control, anti-theft alarm, timing Control and mobile phone APP remote control and other functions. The system is highly customizable and flexible. It can provide all-round security protection for home and corporate users according to user needs and actual conditions. Users can view and manage their security equipment anytime and anywhere through mobile phones or computers, which improves safety and convenience. In addition, the smart home anti-theft system not only solves the practical problems of users, but also has certain social significance. It creates a safer and more stable environment for society, enabling people to enjoy a more secure and comfortable living and working environment. At the same time, the system can also help the police and relevant departments improve the efficiency of emergency response and crime investigation, and contribute to the maintenance of social order.

[0003] However, traditional home anti-theft systems usually rely on a single sensor, infrared or camera, and have problems such as high false alarm rate, lack of adaptive ability, and single alarm mode. They use fixed thresholds to trigger alarms and cannot dynamically adjust the sensitivity according to user behavior, resulting in frequent false alarms. In addition, existing systems mostly use local sound and light alarms, lack the ability to link with smart home devices and community security systems, and have low emergency response efficiency. Therefore, there is an urgent need for a smart home anti-theft system based on smart mobile terminals. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent home anti-theft system based on an intelligent mobile terminal to solve the problems raised in the above background technology.

[0005] To solve the above technical problems, the present invention provides the following technical solutions: a smart home anti-theft system based on a smart mobile terminal, comprising: a dynamic behavior learning module: which uses a hybrid learning algorithm that combines supervised learning with reinforcement learning to generate personalized security strategies by analyzing the user's work and rest time, regional activity trajectory, and device operation history, and updates the trigger conditions in real time; Multimodal sensor collaborative network: This includes cameras, millimeter-wave radars, sound sensors, and temperature and humidity sensors. The millimeter-wave radars are used to penetrate obstructions and detect human movement, while the temperature and humidity sensors assist in identifying environmental anomalies. Data from each sensor is dynamically integrated through an adaptive weight allocation mechanism, with weights adjusted in real time based on environmental complexity and historical false alarm rates. Dual-channel verification unit: After the infrared sensor is triggered, the millimeter-wave radar's micro-motion detection and the camera's AI skeleton recognition are simultaneously activated. An alarm is triggered only when both are determined to be human activity. The skeleton recognition algorithm is embedded in an adversarial neural network to eliminate interference from wearing disguises. Cross-platform emergency linkage engine: Deeply integrated with smart home devices, community security systems, and third-party emergency service platforms, it automatically performs differentiated actions based on the level of intrusion: Level 1 alarm: close doors and windows and activate local sound and light alarm; Level 2 alert: Push encrypted video clips and environmental data snapshots to the user terminal; Level 3 alert: Upload the complete chain of evidence to the cloud through the blockchain evidence storage interface and directly call the police terminal device.

[0006] In the hybrid learning algorithm of the dynamic behavior learning module, the reinforcement learning part adopts a multi-objective optimization strategy to minimize the false alarm rate and maximize the response speed, and adapts to different home scenarios through transfer learning, retaining the localization of user privacy data during migration.

[0007] In the adaptive weight allocation mechanism, the weight of the millimeter-wave radar is automatically increased to above 0.6 at night or when the curtains are closed. The weight of the temperature and humidity sensor is dynamically increased when the environment changes suddenly. The weight adjustment formula is: Among them, S i is the current sensor signal strength, E 历史 is the number of false alarms of the sensor in the past 24 hours, and α and β are the environmental complexity adjustment coefficients.

[0008] The adversarial neural network of the dual-channel verification unit includes a generator and a discriminator. The generator simulates the thermal image of a human body wearing a thick camouflage coat and a mask. The discriminator distinguishes between real intrusion and camouflaged interference through multi-scale feature extraction. A federated learning framework is used for training, and each user terminal only uploads the model gradient to protect privacy.

[0009] In the third-level alarm of the cross-platform emergency linkage engine, the blockchain evidence storage interface adopts sharding storage technology to divide the video data into multiple segments and encrypt and store them in different nodes. When storing evidence, a unique hash value is generated and bound to the user's biometric features to ensure that the chain of evidence cannot be tampered with.

[0010] The system deploys an edge-cloud collaborative computing architecture, where: Millimeter-wave radar signal processing and bone recognition algorithms run on edge computing nodes, with response latency less than 50ms. The user behavior learning model and blockchain evidence storage interface are deployed in a private cloud, and data is desensitized using differential privacy technology.

[0011] The temperature and humidity sensor is linked to the sound sensor. When the characteristic sound pattern of broken glass and a sudden change in temperature and humidity are detected, a level 3 alarm is directly triggered, skipping the dual-channel verification process.

[0012] The system supports the virtual security fence function. Users can draw a dynamic protection area through mobile terminals. When the overlap between the intruder's trajectory and the fence exceeds 80%, the drone tracking module is automatically activated and high-definition video is transmitted back to the security platform in real time.

[0013] Compared with the prior art, the present invention has the following beneficial effects: First, the present invention uses a multimodal sensor collaborative network including millimeter-wave radar, camera, sound sensor and adaptive weight distribution mechanism, combined with a dual-channel verification unit millimeter-wave micro-motion detection + AI skeleton recognition, to effectively eliminate false alarm scenarios such as pet activities, light and shadow interference, and wearing disguises. For example, millimeter-wave radar can penetrate curtains to detect human micro-movements, and AI skeleton recognition is embedded in an adversarial neural network to resist camouflage attacks. Compared with the traditional single infrared sensor solution, the false alarm rate is reduced and improved.

[0014] Second, the present invention adopts a hybrid learning algorithm, combining supervised learning with reinforcement learning to dynamically generate personalized security strategies, adapting to different home scenarios through transfer learning, and using a federated learning framework to protect user privacy data and only upload model gradients. In addition, the edge-cloud collaborative computing architecture localizes sensitive data such as user work and rest patterns, and only desensitized information is stored in the cloud, solving the cumbersome configuration problem caused by the traditional system's reliance on fixed thresholds, while avoiding the risk of user behavior data leakage.

[0015] Third, the cross-platform emergency linkage engine of the present invention supports hierarchical alarms and multi-dimensional responses. The first-level alarm links smart home devices, including closing doors and windows and activating sound and light alarms. The second-level alarm pushes encrypted video clips to the user terminal. The third-level alarm uploads a complete chain of evidence through blockchain sharding evidence storage technology, combining biometric features including voiceprint binding hash values ​​to ensure that evidence cannot be tampered with. Compared with the single local alarm or manual reporting mode in the existing technology, the present invention improves the efficiency of intrusion incident processing, and the evidence chain meets the judicial evidence collection standards, solving the defect that evidence in traditional systems is easily lost or tampered with. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the structure of the present invention; Figure 2It is a structural schematic diagram of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. Example

[0018] like Figure 1 As shown, the present invention provides a technical solution: a smart home anti-theft system based on a smart mobile terminal, including: a dynamic behavior learning module: using a hybrid learning algorithm that combines supervised learning with reinforcement learning, generates personalized security strategies by analyzing the user's work and rest time, regional activity trajectory, and device operation history, and updates the trigger conditions in real time; Multimodal sensor collaborative network: This includes cameras, millimeter-wave radars, sound sensors, and temperature and humidity sensors. The millimeter-wave radars are used to penetrate obstructions and detect human movement, while the temperature and humidity sensors assist in identifying environmental anomalies. Data from each sensor is dynamically integrated through an adaptive weight allocation mechanism, with weights adjusted in real time based on environmental complexity and historical false alarm rates. Dual-channel verification unit: After the infrared sensor is triggered, the millimeter-wave radar's micro-motion detection and the camera's AI skeleton recognition are simultaneously activated. An alarm is triggered only when both are determined to be human activity. The skeleton recognition algorithm is embedded in an adversarial neural network to eliminate interference from wearing disguises. Cross-platform emergency linkage engine: Deeply integrated with smart home devices, community security systems, and third-party emergency service platforms, it automatically performs differentiated actions based on the level of intrusion: Level 1 alarm: close doors and windows and activate local sound and light alarm; Level 2 alert: Push encrypted video clips and environmental data snapshots to the user terminal; Level 3 alert: Upload the complete chain of evidence to the cloud through the blockchain evidence storage interface and directly call the police terminal device.

[0019] The dynamic behavior learning module analyzes the user's work and rest time, regional activity trajectory and equipment operation history data through a hybrid learning algorithm, combining supervised learning and reinforcement learning. Supervised learning establishes an initial security model based on labeled data including the user's normal activity pattern. Reinforcement learning dynamically adjusts the trigger conditions through environmental feedback including false alarm events, generates personalized security strategies and updates threshold parameters in real time. For example, it automatically reduces the sensitivity of infrared sensors according to the user's nighttime sleep habits to reduce false alarms; the multimodal sensor collaborative network integrates cameras, millimeter-wave radars, sound sensors and temperature and humidity sensors. The millimeter-wave radar emits high-frequency electromagnetic waves to penetrate curtains or furniture obstructions, and detects human micro-motion characteristics including breathing or limb movements through the Doppler effect. The temperature and humidity sensor monitors environmental mutations including sudden changes in temperature and humidity caused by broken windows or gas leaks. The data of each sensor is input into the adaptive weight distribution mechanism. The weight calculation formula is: is the current sensor signal strength, E 历史 is the number of false alarms of the sensor in the past 24 hours, α and β are the environmental complexity adjustment coefficients, which realize dynamic data fusion. The dual-channel verification unit simultaneously starts the millimeter-wave radar micro-motion detection and camera AI skeleton recognition after the infrared sensor is triggered. The millimeter-wave radar confirms the presence of the human body through phase change analysis. After the camera captures the picture, the adversarial neural network performs skeleton recognition. The generator simulates the thermal interference of wearing camouflage including thick coats, masks, etc. The discriminator extracts bone joint features through multi-scale convolutional layers. The alarm is triggered only when both are judged to be real human activities, excluding interference from pets or light and shadow; the cross-platform emergency linkage engine performs graded operations according to the severity of the intrusion: the first-level alarm is detected through the smart home The interface closes the electric doors and windows and activates the high-decibel sound and light alarm. The second-level alarm pushes the encrypted video clip AES-256 encryption and environmental data snapshots including temperature, humidity, and sound spectrum to the user's mobile terminal. The third-level alarm calls the blockchain evidence storage interface to divide the video into encrypted fragments and store them in distributed nodes, generate a unique hash value and bind it to the user's voiceprint, and directly call the police terminal equipment through the VoIP protocol to ensure the integrity of the chain of evidence and judicial traceability. In the hybrid learning algorithm of the dynamic behavior learning module, the reinforcement learning part adopts a multi-objective optimization strategy, while minimizing the false alarm rate and maximizing the response speed, and adapts to different family scenarios through transfer learning, retaining the local processing of user privacy data during migration.

[0020] In the hybrid learning algorithm of the dynamic behavior learning module, the reinforcement learning part adopts a multi-objective optimization strategy, and dynamically balances the system performance by designing a dual-objective reward function false alarm rate penalty and a response speed gain. The false alarm rate penalty is calculated based on the frequency of historical false alarm events, including the proportion of false alarms in the past week, and the response speed gain is inversely proportional to the time delay from alarm triggering to action execution. The algorithm iteratively optimizes the policy network parameters through a deep deterministic policy gradient framework, so that the system can automatically find the Pareto optimal solution of false alarm rate and response speed in a complex environment; when transfer learning is adapted to different home scenarios, the feature space mapping technology is used to train the source home scenario. The basic model parameters are frozen, and only the local network layer of the target scene is fine-tuned. During the migration process, user privacy data is always stored in the local edge device. Noise is injected into the model through differential privacy technology to ensure that the original data cannot be restored reversibly. At the same time, the federated learning architecture is used to aggregate multi-user model update gradients in the cloud and only transmit gradients instead of original data, realizing the unification of cross-scenario knowledge sharing and personalized adaptation. For example, when migrating the window breaking features of a high-rise home to a low-rise model, the privacy of local sensor data is retained, and only the feature extraction layer weights are shared. Ultimately, the system can quickly adapt to the new environment without retraining the entire model, taking into account both efficiency and privacy security.

[0021] In the adaptive weight allocation mechanism, the weight of the millimeter-wave radar is automatically increased to above 0.6 at night or when the curtains are closed. The weight of the temperature and humidity sensor is dynamically increased when the environment changes suddenly. The weight adjustment formula is: Among them, S i is the current sensor signal strength, E 历史 is the number of false alarms of the sensor in the past 24 hours, and α and β are the environmental complexity adjustment coefficients.

[0022] The adaptive weight allocation mechanism significantly improves the system's adaptability to environmental changes and detection reliability by dynamically adjusting the weight of each sensor in data fusion. Calculate the weight of each sensor in real time, where S i Reflects the current signal strength, including the signal strength of the millimeter-wave radar detecting the slight movement of the human body, E 历史 Statistics on the number of false alarms of sensors in the past 24 hours include false alarm records of infrared sensors due to pet activities. α and β are dynamically adjusted according to the complexity of the environment. For example, at night or when the curtains are closed, α is increased to 0.7 to enhance the weight of the millimeter-wave radar. When the temperature and humidity change suddenly, β is increased to enhance the weight of the temperature and humidity sensor. This mechanism uses the dual constraints of signal strength and historical false alarm rate to give priority to the most reliable sensor data in the current environment for decision-making. For example, when the infrared sensor has frequent false alarms recently, E 历史 The value is high and the current signal is weak S iWhen the accuracy is low, its weight is greatly reduced, and it relies on millimeter-wave radar and camera data, thereby effectively suppressing false alarms. At the same time, in complex and changeable scenarios (such as strong light interference causing camera failure), the mechanism maintains detection continuity by increasing the weight of millimeter-wave radar and sound sensors, ensuring the robustness of the system in obstructed, camouflaged or extreme environments, and ultimately achieving multi-sensor collaborative optimization and minimization of false alarm rates.

[0023] The adversarial neural network of the dual-channel verification unit includes a generator and a discriminator. The generator simulates the heat map of a human body wearing a thick camouflage coat and a mask. The discriminator distinguishes between real intrusion and camouflaged interference through multi-scale feature extraction. A federated learning framework is used for training, and each user terminal only uploads the model gradient to protect privacy.

[0024] The adversarial neural network in the dual-channel verification unit improves the system's anti-camouflage capability through the collaborative training of the generator and the discriminator: the generator simulates the heat map of a human body wearing camouflage such as a thick coat or a mask, and generates realistic interference samples through adversarial training, forcing the discriminator to learn more refined feature differentiation capabilities; the discriminator uses a multi-scale feature extraction network containing different levels of convolution kernels and attention mechanisms, and analyzes images at multiple levels from local details to global forms such as the proportion of human body contours, identifying the differences between real human activities and camouflage interference, such as abnormal temperature distribution in heat maps, discontinuous heat radiation caused by camouflage clothing, or unnatural skeletal joint movement trajectories and masks that restrict facial micro-movements for judgment; during the training process, the federal The learning framework ensures privacy and security. Each user terminal uses its own sensor data, including intrusion event videos captured by the camera, to train the generator and discriminator locally. Only the non-raw model gradient data is encrypted and uploaded to the cloud server for aggregation and optimization. The cloud then sends the updated global model to each terminal. This mechanism avoids the direct transmission of raw data containing user sensitive information. At the same time, differential privacy technology is used to add random noise to the gradient to further prevent privacy leakage through gradient reverse inference. This design enables the system to leverage multi-user data to jointly improve the generalization ability of the adversarial neural network without sharing any specific user's real-scene data, ultimately achieving the dual goals of high-precision anti-masquerade detection and privacy protection.

[0025] In the third-level alarm of the cross-platform emergency linkage engine, the blockchain evidence interface uses sharding storage technology to divide the video data into multiple segments and encrypt and store them in different nodes. When storing evidence, a unique hash value is generated and bound to the user's biometric features to ensure that the chain of evidence cannot be tampered with.

[0026] The blockchain evidence storage interface adopts shard storage technology, which divides the video data of the intrusion event into multiple fragments according to the timestamp or spatial area. Each fragment is encrypted by the AES-256 algorithm and then distributedly stored in multiple independent nodes of the blockchain network. The shard storage strategy prevents a single node from holding complete data and prevents malicious attacks or tampering. When storing evidence, the system generates a unique hash value for each encrypted fragment and links all fragment hash values ​​in sequence into a Merkle tree structure, and finally generates a global root hash. At the same time, the root hash is bound to the user's biometric features. The voiceprint extracts key frequency band parameters through MFCC and converts them into non-reversible hash values, which are written into the blockchain smart contract together with the root hash. This binding mechanism ensures that even if the number of fragments is large, the global root hash value is not large. If the data is illegally obtained, it is impossible to decrypt or reconstruct the complete video. It is necessary to simultaneously crack the multi-node encrypted shards and match the user's biometric features. Any tampering with the content of the shards will cause the Merkle tree hash chain to break, triggering the blockchain consensus mechanism to automatically mark the abnormal node and start the data repair process. In addition, the hash value bound to the biometric is authenticated through zero-knowledge proof technology. Users can extract the matching hash value through voiceprint verification without directly exposing the original biometric data, further protecting privacy. This design gives the evidence chain of the third-level alarm judicial credibility, meets data integrity, non-repudiation and privacy compliance requirements, and completely solves the defects of traditional security systems such as easy tampering of evidence, centralized storage and weak identity authentication.

[0027] The system deploys an edge-cloud collaborative computing architecture, where: Millimeter-wave radar signal processing and bone recognition algorithms run on edge computing nodes, with response latency less than 50ms. The user behavior learning model and blockchain evidence storage interface are deployed in a private cloud, and data is desensitized using differential privacy technology.

[0028] The edge-cloud collaborative computing architecture achieves the dual goals of efficient real-time processing and privacy and security protection through division of labor and cooperation: millimeter-wave radar signal processing and bone recognition algorithms are deployed on edge computing nodes to directly process raw sensor data, and use the local computing power of edge devices to complete signal analysis and bone joint feature extraction, compressing the response delay to less than 50ms to ensure the real-time performance of intrusion detection; at the same time, the user behavior learning model and blockchain evidence interface are deployed on the private cloud. The cloud uses a high-performance GPU cluster to train the global model of the hybrid learning algorithm and uses differential privacy technology to desensitize the uploaded data - Laplace noise is injected into the user behavior data to blur sensitive information and prevent data reverse restoration; when the blockchain is stored, the cloud receives the encrypted metadata transmitted by the edge node. Data, the video fragments are encrypted with AES-256 and then stored in distributed nodes, and a unique hash value is generated and bound to the user's voiceprint to ensure the judicial compliance of the evidence storage process; the edge nodes only transmit desensitized features to the cloud to avoid exposure of the original video or radar signal, while the cloud aggregates the gradients of multiple edge node models through the federated learning framework to optimize the global model to achieve knowledge sharing and privacy protection; this architecture uses resources in a hierarchical manner, with the edge focusing on low-latency tasks and the cloud handling complex computing and storage, reducing bandwidth pressure and supporting dynamic expansion, while meeting the requirements of privacy regulations such as GDPR, solving the latency bottleneck of traditional centralized cloud computing and the computing power limitations of edge devices, and achieving an integrated balance of millisecond-level response, data full life cycle protection and cross-regional collaboration capabilities in security scenarios.

[0029] The temperature and humidity sensors are linked to the sound sensors. When the characteristic sound pattern of broken glass is detected and the temperature and humidity change suddenly, a level 3 alarm is directly triggered, skipping the dual-channel verification process.

[0030] The linkage mechanism between the temperature and humidity sensor and the sound sensor achieves rapid response to emergencies through multi-dimensional environmental anomaly detection. When the sound sensor captures the characteristic soundprint of glass breaking in the frequency band of 5kHz to 8kHz, the frequency domain energy peak is extracted through Fourier transform, and combined with the time domain short-time energy mutation analysis, the interference of similar frequencies such as metal impact is eliminated. At the same time, the temperature and humidity sensor detects a sudden change in temperature and humidity, such as a temperature drop of more than 2°C and a humidity increase of more than 10%. It is determined to be indoor and outdoor air exchange caused by broken windows. The system will determine it as an extreme intrusion event and immediately skip the conventional dual-channel verification process of millimeter-wave radar and camera verification, and directly trigger a level 3 alarm. At this time, the system executes the highest priority operation - through the block The chain sharding evidence storage interface encrypts and stores real-time video in distributed nodes, generates a unique hash value and binds it to the user's voiceprint, and simultaneously activates the drone tracking module to lock the intruder's location and send back high-definition images, while directly calling the police terminal equipment to ensure that the response delay is less than 10 seconds; this linkage logic is based on the characteristic coupling of physical destruction events and the high temporal and spatial correlation between the sound of glass breaking and sudden changes in temperature and humidity. While ensuring high confidence, it sacrifices some redundant verification in exchange for millisecond-level response in critical scenarios, solving the problem of delays caused by lengthy verification processes in traditional security systems during violent intrusions. At the same time, the integrity of the evidence chain is maintained by binding sharded evidence to biometrics, avoiding the risk of judicial credibility caused by skipping verification.

[0031] The system supports the virtual security fence function. Users can draw a dynamic protection area through mobile terminals. When the overlap between the intruder's trajectory and the fence exceeds 80%, the drone tracking module is automatically activated and high-definition video is transmitted back to the security platform in real time.

[0032] Users draw a dynamic protection zone using their mobile devices. The system converts the polygonal coordinates drawn by the user into a virtual electronic fence based on a geofencing algorithm and monitors intruders' movements in real time. When an intruder is detected by infrared sensors, cameras, or millimeter-wave radar, the system uses integrated positioning technology to track their movement in real time and calculate the overlap between their trajectory and the virtual fence. Once a threshold is triggered, the drone tracking module is automatically activated. Using RTK high-precision positioning and obstacle avoidance algorithms, the drone quickly takes off and flies along the pre-set path to the target area. Its onboard 4K camera and thermal imager capture high-definition footage of the intruder in real time, which is transmitted back to the security platform via a 5G private network with low latency. Simultaneously, the system initiates a multi-level coordinated response: triggering audible and visual alarms to deter intruders, displaying the drone tracking footage on split-screen on the user's device and the community security center, and encrypting and storing the tracking video via a blockchain evidence storage interface, generating a spatiotemporal hash chain as legal evidence. This feature, by combining dynamic fencing with intelligent tracking, overcomes the blind spots of traditional fixed surveillance, enabling proactive defense and precise evidence collection. This feature is particularly suitable for large-scale outdoor scenarios, providing early warning and locking onto the target before the intruder approaches the core area, upgrading the security response from passive alarm to active interception.

[0033] Working principle: The system receives user configuration instructions through smart mobile terminals and interacts in real time. The multi-sensor module includes cameras, infrared sensors, sound sensors, millimeter-wave radars and temperature and humidity sensors to continuously collect environmental data. The millimeter-wave radar uses high-frequency electromagnetic waves to penetrate obstructions to detect human micro-motion characteristics. The temperature and humidity sensor monitors environmental mutations, including sudden changes in temperature and humidity caused by broken windows. All sensor data are input into the data processing module. The hybrid learning algorithm combines supervised learning with reinforcement learning to analyze the user's historical behavior patterns, including work and rest time and activity trajectory, dynamically generate personalized security strategies and optimize trigger conditions in real time; when the infrared sensor detects an abnormality, the multi-source data fusion unit activates the adaptive weight allocation mechanism, and dynamically adjusts the weight of each sensor according to the complexity of the environment, including night or curtain closed status and historical false alarm rate, including increasing the weight of the millimeter-wave radar to above 0.6, and synchronously triggering the dual-channel verification process: the millimeter-wave radar performs micro-motion detection, and the camera captures the picture through the AI ​​skeleton recognition algorithm embedded in the adversarial neural network. The algorithm simulates the heat map of a camouflaged human body through the generator, and the discriminator distinguishes between real intrusion and interference based on multi-scale feature extraction. Only when both are judged as human activity, the alarm execution module is activated. Activate the graded response mechanism - if it is a minor anomaly including a short trigger, push an encrypted notification to the user terminal; if it is determined to be a serious intrusion, the cross-platform emergency linkage engine automatically executes three-level operations: the first-level alarm closes the doors and windows and activates the sound and light alarm, the second-level alarm uploads the encrypted video clip to the user terminal, and the third-level alarm uses the blockchain sharding technology to divide the video into encrypted clips and store them in distributed nodes, generate a unique hash value and bind it to the user's voiceprint, and directly call the police terminal at the same time; the system adopts an edge-cloud collaborative architecture, and the millimeter wave signal processing and bone recognition algorithm run on the edge node. The response delay is less than 50ms, and the user's line After the data is desensitized with differential privacy, it is uploaded to the private cloud for training the model. The federated learning framework ensures that each terminal only shares the model gradient to protect privacy. In addition, the virtual security fence function allows users to draw dynamic protection areas through mobile terminals. When the overlap between the intruder's trajectory and the fence exceeds 80%, the drone tracking module automatically starts and transmits high-definition images back to the security platform to achieve active defense. During the whole process, the temperature and humidity sensors and sound sensors are linked to monitor the 5kHz to 8kHz frequency band of the glass shattering characteristic voiceprint. If an abnormality is detected, the dual-channel verification is skipped and the third-level alarm is directly triggered to ensure zero-delay response in extreme situations.

[0034] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and alterations may be made to the embodiments without departing from the principles and spirit thereof, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A smart home anti-theft system based on a smart mobile terminal, characterized in that: include: Dynamic Behavior Learning Module: This module uses a hybrid learning algorithm that combines supervised learning with reinforcement learning. By analyzing user schedules, regional activity trajectories, and device operation history, it generates personalized security strategies and updates trigger conditions in real time. Multimodal sensor collaborative network: This includes cameras, millimeter-wave radars, sound sensors, and temperature and humidity sensors. The millimeter-wave radars are used to penetrate obstructions and detect human movement, while the temperature and humidity sensors assist in identifying environmental anomalies. Data from each sensor is dynamically integrated through an adaptive weight allocation mechanism, with weights adjusted in real time based on environmental complexity and historical false alarm rates. Dual-channel verification unit: After the infrared sensor is triggered, the millimeter-wave radar's micro-motion detection and the camera's AI skeleton recognition are simultaneously activated. An alarm is triggered only when both are determined to be human activity. The skeleton recognition algorithm is embedded in an adversarial neural network to eliminate interference from wearing disguises. Cross-platform emergency linkage engine: Deeply integrated with smart home devices, community security systems, and third-party emergency service platforms, it automatically performs differentiated actions based on the level of intrusion: Level 1 alarm: close doors and windows and activate local sound and light alarm; Level 2 alert: Push encrypted video clips and environmental data snapshots to the user terminal; Level 3 alert: Upload the complete chain of evidence to the cloud through the blockchain evidence storage interface and directly call the police terminal device.

2. The smart home anti-theft system based on a smart mobile terminal according to claim 1, characterized in that: In the hybrid learning algorithm of the dynamic behavior learning module, the reinforcement learning part adopts a multi-objective optimization strategy to minimize the false alarm rate and maximize the response speed, and adapts to different home scenarios through transfer learning, retaining the localization of user privacy data during migration.

3. The smart home anti-theft system based on a smart mobile terminal according to claim 1, characterized in that: In the adaptive weight allocation mechanism, the weight of the millimeter-wave radar is automatically increased to above 0.6 at night or when the curtains are closed. The weight of the temperature and humidity sensor is dynamically increased when the environment changes suddenly. The weight adjustment formula is: Among them, S i is the current sensor signal strength, E 历史 is the number of false alarms of the sensor in the past 24 hours, and α and β are the environmental complexity adjustment coefficients.

4. The smart home anti-theft system based on a smart mobile terminal according to claim 1, characterized in that: The adversarial neural network of the dual-channel verification unit includes a generator and a discriminator. The generator simulates the thermal image of a human body wearing a thick camouflage coat and a mask. The discriminator distinguishes between real intrusion and camouflaged interference through multi-scale feature extraction. A federated learning framework is used for training, and each user terminal only uploads the model gradient to protect privacy.

5. The smart home anti-theft system based on a smart mobile terminal according to claim 1, characterized in that: In the third-level alarm of the cross-platform emergency linkage engine, the blockchain evidence storage interface adopts sharding storage technology to divide the video data into multiple segments and encrypt and store them in different nodes. When storing evidence, a unique hash value is generated and bound to the user's biometric features to ensure that the chain of evidence cannot be tampered with.

6. The smart home anti-theft system based on a smart mobile terminal according to claim 1, characterized in that: The system deploys an edge-cloud collaborative computing architecture, where: Millimeter-wave radar signal processing and bone recognition algorithms run on edge computing nodes, with response latency less than 50ms. The user behavior learning model and blockchain evidence storage interface are deployed in a private cloud, and data is desensitized using differential privacy technology.

7. The smart home anti-theft system based on a smart mobile terminal according to claim 1, characterized in that: The temperature and humidity sensor is linked to the sound sensor. When the characteristic sound pattern of broken glass and a sudden change in temperature and humidity are detected, a level 3 alarm is directly triggered, skipping the dual-channel verification process.

8. The smart home anti-theft system based on a smart mobile terminal according to claim 1, characterized in that: The system supports the virtual security fence function. Users can draw a dynamic protection area through mobile terminals. When the overlap between the intruder's trajectory and the fence exceeds 80%, the drone tracking module is automatically activated and high-definition video is transmitted back to the security platform in real time.

Images