Smart home system based on One NET platform

Abstract

The invention relates to the technical field of smart home and Internet of Things, and discloses a smart home system based on an One NET platform, and the system comprises a data collection module which is used for collecting original data related to family activities, and generating a multi-dimensional time sequence feature vector representing a current family state according to the original data; the data transmission module is used for uploading the multi-dimensional time sequence feature vector to a cloud processing module deployed on a One NET platform; and the cloud processing module is used for receiving the multi-dimensional time sequence feature vector and analyzing the multi-dimensional time sequence feature vector based on an activity scene recognition model with a dynamic updating mechanism so as to determine a target activity scene of the current family. Through the dynamically updated scene recognition model, the multi-objective optimization algorithm and the enhanced security mechanism, the optimal intelligent control capable of accurately adapting to user behaviors, actively performing health and safety intervention and balancing multiple objectives is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart home and Internet of Things, in particular to a smart home system based on One NET platform. BACKGROUND

[0002] With the development of social economy and the continuous improvement of people's demand for living quality, the demand for intelligent, personalized and healthy living environment of modern families is increasing. Especially in the context of the diversification of family structure, how to provide a safe and comfortable living space for the elderly, Alzheimer's patients and other special groups, which can be passively monitored, has become an important social issue.

[0003] The existing smart home technology has realized the effective networking and basic control of home devices. By applying various sensors and wireless communication technologies, users can remotely control the lights, curtains, air conditioners and other appliances in their homes through mobile terminals and other devices, and can preset multiple scene modes according to personal preferences, such as home mode or cinema mode. These technical solutions have realized the linkage of multiple devices to a certain extent, providing users with convenient automation experience through simple trigger rules, and providing a feasible technical path for basic home energy management and security monitoring.

[0004] However, the existing technology has obvious shortcomings in the depth of intelligence, security and the breadth of function; first, its scene recognition ability depends on static rules or models, which cannot dynamically adapt to changes in long-term behavior patterns of users, resulting in a decrease in decision accuracy over time. Secondly, its data transmission security mechanism is relatively single, lacking dynamic key update and active detection capability of abnormal traffic, and there are security risks. In addition, the functions of the existing technology are mainly limited to environmental control and basic security, generally lacking non-contact normal monitoring capability of physiological indicators of users, and it is difficult to provide active health and safety intervention for special groups such as Alzheimer's patients. SUMMARY

[0005] In view of the shortcomings of the prior art, the present application provides a smart home system based on One NET platform, which solves the problem that the existing smart home system generally uses a static scene recognition model that cannot adapt to changes in user behavior, the data transmission security mechanism is relatively single, and lacks dynamic key update and active defense capability of abnormal traffic.

[0006] To achieve the above purpose, the present application realizes the following technical solutions:

[0007] The present application provides a smart home system based on One NET platform, which comprises:

[0008] a data collection module configured to collect raw data related to household activities and generate a multi-dimensional time-series feature vector representing a current household state based on the raw data;

[0009] a data transmission module configured to upload the multi-dimensional time-series feature vector to a cloud processing module deployed on a One NET platform;

[0010] a cloud processing module configured to receive the multi-dimensional time-series feature vector and analyze the multi-dimensional time-series feature vector based on an activity scenario recognition model with a dynamic updating mechanism to determine a target activity scenario of the current household, and then generate and match an optimal composite energy and health strategy from a scenario-strategy mapping library based on the target activity scenario using a multi-objective optimization algorithm;

[0011] a strategy execution module configured to receive control instructions contained in the composite energy and health strategy and drive home appliance devices to execute the control instructions through a device linkage rule engine.

[0012] In a specific embodiment, the data collection module includes:

[0013] a multi-source perception unit configured to obtain sensors of the raw data, the sensors including environment state sensors and physiological indicator sensors;

[0014] the environment state sensors include passive infrared sensors, door magnetic sensors, sound sensors, and alternating current sensors;

[0015] the physiological indicator sensors are configured to non-contact monitor a user's heart rate, breathing rate, body temperature, blood pressure, or blood glucose using a millimeter wave radar;

[0016] a feature generation unit configured to process the raw data to generate the multi-dimensional time-series feature vector.

[0017] Further, the feature generation unit is specifically configured to:

[0018] quantize a state signal of the passive infrared sensor into a binary feature value representing the presence of a human body;

[0019] quantize a state signal of the door magnetic sensor into a binary feature value representing the opening and closing of a door or window;

[0020] normalize an analog output signal of the sound sensor to obtain an environmental sound decibel feature value;

[0021] normalize an output signal of the alternating current sensor to obtain an appliance operating current feature value;

[0022] Filtering and calibrating the output signal of the physiological index sensor to obtain a standardized physiological index feature value;

[0023] Splicing all the feature values above and attaching a time stamp to construct the multi-dimensional time sequence feature vector.

[0024] In one specific embodiment, the data transmission module comprises:

[0025] A wireless communication unit for establishing a network connection between the data transmission module and the cloud processing module;

[0026] A protocol packaging unit for packaging the multi-dimensional time sequence feature vector in the format of the MQTT protocol and transmitting it through the wireless communication unit, wherein the One NET platform provides the infrastructure and device identity authentication service for the communication in the format of the MQTT protocol.

[0027] Further, the protocol packaging unit is specifically used for:

[0028] Using the device identity certificate pre-installed in the data transmission module to complete the two-way authentication with the cloud processing module through the One NET platform and establish an initial encrypted communication channel;

[0029] The two-way authentication includes a dynamic key update mechanism and an abnormal traffic detection algorithm, and the session key of the encrypted communication channel is dynamically negotiated and updated at a preset time period;

[0030] The abnormal traffic detection algorithm is used to analyze the traffic mode of data transmission in real time, and when an attack behavior deviating from the normal mode is detected, the communication is automatically cut off and an alarm is sent to the user or system administrator.

[0031] In one specific embodiment, the cloud processing module comprises:

[0032] A scenario recognition unit for receiving the multi-dimensional time sequence feature vector and calculating the target activity scenario according to the activity scenario recognition model with a dynamic update mechanism;

[0033] A strategy generation unit for querying in the scenario-strategy mapping library by taking the target activity scenario and real-time physiological index data as retrieval conditions and generating and obtaining the optimal composite strategy corresponding thereto through the multi-objective optimization algorithm.

[0034] Further, the activity scenario recognition model with a dynamic update mechanism specifically determines the target activity scenario by the following ways:

[0035] The activity scenario recognition model is built-in a user behavior prediction model based on a long short-term memory network, which learns a historical multi-dimensional time sequence feature vector to predict a user activity scenario in a next time window;

[0036] The real-time input multi-dimensional time sequence feature vector V current is compared with the feature vector V predicted predicted by the prediction model, and a deviation Δ = ||V current -V predicted ||2 is calculated;

[0037] When the deviation Δ exceeds a preset threshold ∈, it indicates that the user behavior pattern has changed, and the system triggers a weight self-adaptive adjustment mechanism, adopts a gradient descent algorithm to optimize the weight vector W i in the scenario recognition model, and the update rule is represented as:

[0038]

[0039] In the formula, W i (t+1) is the updated weight; η is a learning rate; is a gradient of the loss function J with respect to the weight vector W i (t); W i is the weight vector in the scenario recognition model;

[0040] In identifying a current scenario, a dot product between the current feature vector V current and all preset activity scenario corresponding dynamically adjusted weight vectors W i is calculated to obtain a matching score S of each scenario, and a scenario with the highest score S max is determined as a target activity scenario.

[0041] In one specific embodiment, the policy execution module comprises:

[0042] An instruction analysis unit, configured to receive and analyze a control instruction issued from the cloud processing module;

[0043] A device linkage rule engine, configured to receive a plurality of sensor state signals from the data acquisition module as evidence, and perform probability inference according to a preset conditional probability table to calculate a posterior probability of one or more preset user activity scenarios, and automatically generate and deliver a linkage control instruction corresponding to the user activity scenario when the posterior probability exceeds a preset confidence threshold;

[0044] A device driving unit, connected with the instruction analysis unit and the device linkage rule engine, and containing a relay or switch circuit, for performing physical on-off control on the running state of the home appliance device according to the direct instruction of the instruction analysis unit or the instruction inferred and generated by the device linkage rule engine.

[0045] Further, the physical on-off control specifically includes:

[0046] According to the analysis result of the instruction analysis unit, the contact of the relay or switch circuit is controlled to be closed to turn on the circuit for supplying power to the home appliance device;

[0047] The contact of the relay or switch circuit is controlled to be opened to turn off the circuit for supplying power to the home appliance device.

[0048] In a specific embodiment, the smart home system is also applied to the anti-wandering behavior intervention scene of patients with dementia, and specifically includes:

[0049] In the data acquisition module, the trajectory heat map of the indoor activities of the user is constructed by using the millimeter wave radar and the sensor deployed in the key area;

[0050] In the cloud processing module, the trajectory heat map and the activity time sequence are analyzed by using a behavior pattern learning model based on a long short-term memory network to learn the daily behavior pattern of the patient;

[0051] When the behavior pattern learning model identifies an abnormal behavior deviating from the daily pattern, the system determines that this is a potential wandering target scene;

[0052] The strategy execution module receives an intervention strategy for the potential wandering target scene and automatically performs a linkage operation, and the linkage operation includes encrypting the electronic lock of the door to prevent going out, adjusting the indoor light to a soft color temperature and playing soothing music familiar to the patient to guide him / her back to the bedroom, and simultaneously sending an emergency alarm to the mobile terminal of the preset family member or caregiver.

[0053] The present application provides a smart home system based on One NET platform, which has the following beneficial effects:

[0054] 1、The application improves the self-adaptive ability and recognition accuracy of the system to user behavior by introducing an active scene recognition model with a dynamic updating mechanism. The active scene recognition model is built-in with a user behavior prediction model based on a long short-term memory network, and the model weight is adaptively adjusted in combination with a gradient descent algorithm. When the actual behavior pattern of the user changes over time, the system can predict by learning historical data, and automatically optimize its internal parameters when the prediction deviates from the actual occurrence, thereby continuously maintaining high-precision scene recognition, avoiding the technical defects of the system using a static model leading to decision failure due to user habit change.

[0055] 2、The application enhances the security and reliability of data in the transmission process by constructing a multi-level security protection mechanism. On the basis of completing the two-way authentication by using the device identity certificate, the system integrates a dynamic key updating mechanism, which reduces the risk of static key cracking by periodically negotiating and replacing the communication session key. At the same time, the deployed abnormal traffic detection algorithm can analyze the data flow pattern in real time, actively identify and block potential network attack behavior, and ensure the integrity and confidentiality of sensitive data such as user multi-dimensional time sequence feature vectors during uploading to the cloud processing module.

[0056] 3、The application deeply integrates physiological index monitoring and environmental control, realizes the functional expansion from single device management to active health and safety intervention. The system collects user health data through physiological index sensors such as millimeter wave radar, and generates composite energy and health strategies. In particular, in the application scenario of preventing Alzheimer's patients from getting lost, the system uses a long short-term memory network to analyze the user's activity trajectory heat map to learn their daily behavior patterns, and when it identifies abnormal behaviors such as walking to the door at night, it can automatically perform a series of intervention operations such as encrypting the door lock, adjusting the environment, and sending an alarm to the caregiver, providing an effective safety monitoring technical solution for special groups. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 The system architecture diagram of the embodiment of the application;

[0058] Figure 2 The data acquisition module framework diagram of the embodiment of the application;

[0059] Figure 3 The data transmission module framework diagram of the embodiment of the application;

[0060] Figure 4 The cloud processing module framework diagram of the embodiment of the application;

[0061] Figure 5 The strategy execution module framework diagram of the embodiment of the application.

[0062] Explanation of reference numerals:

[0063] 100, data acquisition module; 200, data transmission module; 300, cloud processing module; 400, policy execution module. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0065] Reference Figure 1 , Figure 1 is a system architecture diagram according to an embodiment of the present application. The embodiment of the present application provides an intelligent home system based on One NET platform, which comprises a data acquisition module 100, a data transmission module 200, a cloud processing module 300 and a policy execution module 400.

[0066] The data acquisition module 100 is used for collecting original data related to user activities and physiological states in a home environment, and generating a multi-dimensional time sequence feature vector after processing the original data.

[0067] The data transmission module 200 communicates data with the data acquisition module 100, and is used for uploading the multi-dimensional time sequence feature vector to the cloud processing module 300.

[0068] The cloud processing module 300 is deployed on the One NET platform and connected with the data transmission module 200 through a network. The cloud processing module 300 is used for receiving the multi-dimensional time sequence feature vector, analyzing to determine a target activity scenario based on an activity scenario recognition model with a dynamic updating mechanism, and generating a composite energy and health policy by using a multi-objective optimization algorithm.

[0069] The policy execution module 400 communicates with the cloud processing module 300, and is physically or signal connected with home appliances in the home environment. The policy execution module 400 is used for receiving and executing control instructions issued by the cloud processing module 300 to drive the home appliances to run.

[0070] The overall working process of the intelligent home system in the embodiment of the present application is as follows:

[0071] Firstly, the data acquisition module 100 deployed in the user's home continuously collects environmental state data and physiological index data of the user through the multi-source perception unit integrated in it. The feature generation unit in it performs quantization, normalization and filtering and other preprocessing operations on the collected various original data, and finally constructs a multi-dimensional time sequence feature vector containing a time stamp.

[0072] Subsequently, the data transmission module 200 obtains the multi-dimensional time sequence feature vector from the data acquisition module 100. The protocol encapsulation unit in the data transmission module 200 encapsulates the multi-dimensional time sequence feature vector according to the MQTTS protocol format, and through the wireless communication unit, establishes an encrypted communication channel with the cloud processing module 300 on the basis of the infrastructure and device identity authentication service provided by the One NET platform, and transmits the encapsulated data packet.

[0073] Next, in combination with the description of Figure 4 , the cloud processing module 300 deployed on the One NET platform receives the encrypted data. The scene recognition unit in the cloud processing module 300 uses an activity scene recognition model with a built-in long short-term memory network (LSTM) to analyze the received time sequence feature vector sequence, and combines its dynamic weight update mechanism to calculate the target activity scene of the current home environment. Next, the strategy generation unit takes the target activity scene and real-time physiological index data as input, runs a multi-objective optimization algorithm (such as NSGA-II), and generates and matches an optimal composite energy and health strategy, i.e., the optimal composite strategy, in the preset scene-strategy mapping library, which can balance energy consumption, comfort and device life.

[0074] Finally, the cloud processing module 300 issues the specific control instructions parsed from the optimal composite strategy to the strategy execution module 400. After the strategy execution module 400 receives the instructions, the device linkage rule engine inside it reasons according to the preset Bayesian network rules to determine whether to trigger the coordinated action of other devices. Finally, the instructions generated by the instruction parsing unit and the device linkage rule engine are physically controlled by the device driving unit to control the power supply circuit of the household appliance, completing the closed-loop regulation and control of the home environment.

[0075] Referring to Figure 2 , Figure 2 is a functional block diagram of a data acquisition module according to an embodiment of the present application. In specific embodiments of the present application, the functions of the data acquisition module 100 are implemented through its internal physical and logical implementation, mainly including a multi-source perception unit and a feature generation unit.

[0076] The multi-source perception unit is used to directly interact with the physical environment of the home, and is internally integrated with multiple sensors to obtain raw data. The environmental state sensors included in the multi-source perception unit can be specifically configured as: one or more passive infrared (PIR) sensors, which are deployed at the corners of walls or ceilings in corridors, living rooms and other areas where people mainly move, and are used to output high-low level signals representing human movement by detecting changes in infrared radiation emitted by the human body; door magnetic sensors, which are respectively installed on the doors and windows of the entrance door and key rooms, and are composed of a magnet part and a switch part, and are used to output on-off signals representing the opening or closing state of the doors and windows; a sound sensor, which is a microphone and a signal amplification circuit, and is deployed in core areas such as living rooms, and is used to collect intensity information of environmental sound, and its output is an analog voltage signal proportional to the sound decibel value; an alternating current sensor, which can use a non-intrusive current transformer, and is clamped on the total fire line of a household meter or the power supply line of a specific high-power device, and is used to monitor the running power consumption of electrical appliances, and its output is a signal proportional to the current size.

[0077] The multi-source perception unit also includes physiological indicator sensors. In one specific embodiment, the physiological indicator sensor is a millimeter wave radar module, which is deployed facing the user's bed or a sofa where the user usually sits. The millimeter wave radar module transmits millimeter wave electromagnetic waves of a specific frequency (for example, 60 GHz) and receives the echo signals reflected by the human body, analyzes the tiny displacement of the chest caused by the user's heartbeat and breathing through the Doppler effect, and thus calculates the raw data stream of physiological indicators such as the user's heart rate and breathing frequency without contacting the user's body.

[0078] The feature generation unit is electrically connected with the multi-source perception unit, and is used to receive and process the raw signals output by the various types of sensors. The feature generation unit performs a series of determined data processing steps to generate structured multi-dimensional time sequence feature vectors. First, for the discrete state signals output by the passive infrared sensor and the door magnetic sensor, the feature generation unit directly performs binary quantization, for example, mapping the state of human movement or door / window opening to binary value 1, and mapping the state of no activity or door / window closing to 0.

[0079] For the continuous analog signals output by the sound sensor and the alternating current sensor, the feature generation unit first digitizes them through an analog-to-digital converter (ADC), and then performs normalization processing to eliminate differences in dimension and numerical range. Normalization can use the min-max normalization method, and its calculation formula is:

[0080]

[0081] In the formula, X norm is the normalized feature value; X is the instantaneous value collected by the sensor; X minand X max are respectively the minimum and maximum values of the normal operating range of the sensor determined during the calibration process.

[0082] For the physiological indicator raw data stream output by the millimeter wave radar, the feature generation unit performs filtering and calibration operations on it due to its containing environmental noise and body motion artifacts. The Kalman filter algorithm can be implemented inside the feature generation unit to process the continuous heart rate and respiratory rate data sequences to filter out random noise and obtain a smooth estimated value closer to the true value. The processed data is calibrated to international standard units, such as the unit of heart rate is beats per minute (bpm).

[0083] Finally, the feature generation unit splices and combines all the processed feature values in a predefined fixed order. At the end of the combined vector, a high-precision timestamp provided by the internal Real-Time Clock (RTC) module is attached to form a complete multi-dimensional time series feature vector. The multi-dimensional time series feature vector is then output to the data transmission module 200.

[0084] Referring to Figure 3 , Figure 3 is a functional block diagram of the data transmission module 200 according to an embodiment of the present application. In a specific embodiment of the present application, the data transmission module 200 is used to establish a secure and reliable data link between the data acquisition module 100 and the cloud processing module 300. The data transmission module 200 mainly includes a wireless communication unit and a protocol encapsulation unit.

[0085] The wireless communication unit can be a Narrow Band Internet of Things (NB-IoT) communication module in hardware. The wireless communication unit is used to search for and connect to a local wireless network to establish a physical connection with the Internet, providing network layer and transport layer support for the upper layer data transmission.

[0086] The protocol encapsulation unit interacts with the protocol stack of the wireless communication unit to perform data formatting and encapsulation, encryption, and security authentication. The protocol encapsulation unit first serializes the multi-dimensional time series feature vector received from the data acquisition module 100, converting it into a general data exchange format such as a JSON (JavaScript Object Notation) string.

[0087] Subsequently, the protocol encapsulation unit constructs an MQTT publish packet in the standard format of MQTT (Message Queuing Telemetry Transport) protocol, taking the JSON string as the message payload. The packet specifies the target topic, through which the cloud processing module 300 receives data. The entire communication process is based on the TLS (Transport Layer Security) protocol, constituting an MQTTTS communication. The One NET platform acts as an MQTTTS broker in this process, providing an access point for devices and routing and distributing messages.

[0088] To ensure communication security, the protocol encapsulation unit performs the following specific security operations when establishing a communication link and transmitting data. When initially connected, the data transmission module 200 performs two-way authentication with the One NET platform using the device identity credentials (including product ID, device ID, and device key) pre-stored in its internal non-volatile memory. During the TLS handshake process, the data transmission module 200 presents its identity credentials to the platform and signs key connection parameters using the device key. After the platform verifies that the signature is correct, it confirms the legality of the device identity, thereby establishing an initial encrypted communication channel.

[0089] After the initial encrypted channel is established, the system also performs a dynamic key update mechanism. The protocol encapsulation unit and the cloud processing module 300 negotiate and update the session key through an application layer protocol at a pre-set time period (e.g., 600 seconds). This process can be based on the Diffie-Hellman key exchange algorithm. Both parties exchange their public keys in the encrypted channel and calculate a new shared secret based on the other party's public key and their own private key. Then, through a key derivation function (KDF), a new symmetric session key is generated for subsequent data transmission encryption, and the original session key is invalidated.

[0090] In addition, the cloud processing module 300 is deployed with an abnormal traffic detection model based on the Isolation Forest algorithm at the data receiving end. The abnormal traffic detection model takes the transmission frequency, packet size, connection mode, etc. of each device data stream as input features, and analyzes the traffic mode of data transmission in real time. The Isolation Forest algorithm detects outliers by constructing a set of randomized binary trees. Due to its sparsity, abnormal data points will be quickly isolated in the tree. When the calculated anomaly score of a certain device data stream exceeds the preset threshold, the system determines that the device has attack behavior. At this time, the system automatically calls the device management interface of the One NET platform, immediately cuts off the MQTTS connection with the device, and sends alarm information to the terminal of the system administrator.

[0091] Referring to Figure 4 , Figure 4 is a functional block diagram of the cloud processing module 300 according to an embodiment of the present application. In the specific embodiment of the present application, the cloud processing module 300 is deployed in the One NET platform, and the cloud processing module 300 mainly includes a scenario recognition unit and a policy generation unit.

[0092] The scenario recognition unit is used to receive the multi-dimensional time sequence feature vector sequence uploaded from the data transmission module 200, and perform dynamic scenario recognition. The scenario recognition unit internally has an active scenario recognition model with a dynamic updating mechanism. The active scenario recognition model includes a Long Short-Term Memory Network (LSTM) for user behavior prediction, and a mechanism for adaptive adjustment of weights. The LSTM network processes the input historical feature vector sequence through its unique input gate, forget gate and output gate structure, and predicts the feature vector corresponding to the user activity state in the next time window to generate a predicted feature vector.

[0093] In runtime, the scenario recognition unit compares the current feature vector V current output by the LSTM network with the predicted feature vector V predicted , and calculates the deviation Δ between the two, which can be obtained by calculating the Euclidean distance between the two vectors: Δ = ||V current -V predicted ||2.

[0094] When the calculated deviation Δ exceeds a preset threshold ∈, the system determines that the current behavior pattern of the user deviates from the historical learned pattern. This condition will trigger the weight adaptive adjustment mechanism. The mechanism uses the gradient descent algorithm to optimize and update the weight vector used for final scenario classification. The mathematical expression of its update rule is:

[0095]

[0096] wherein W i (t+1) is the updated weight; η is the learning rate; is the gradient of the loss function J with respect to the weight vector W i (t); W i is the weight vector in the scenario recognition model.

[0097] After the above-mentioned weight adjustment is completed or in the case that the bias does not exceed the threshold, the scenario recognition unit obtains a matching score S current for each scenario by calculating the dot product between the current feature vector V i and the weight vector W i corresponding to all preset activity scenarios (for example: sleep, leaving home, cooking, dining, etc.). The scenario with the highest score is determined as the target activity scenario of the current family and is transmitted as output to the strategy generation unit.

[0098] The strategy generation unit receives the target activity scenario determined by the scenario recognition unit and the real-time physiological indicator data extracted from the multi-dimensional time-series feature vector. The strategy generation unit uses NSGA-II (Non-dominated Sorting Genetic Algorithm II) as a multi-objective optimization algorithm to generate and match an optimal composite energy and health strategy from the scenario-strategy mapping library. The algorithm is configured to simultaneously optimize the following three conflicting objective functions:

[0099] An energy consumption cost function f1(P), the goal of which is to minimize, is defined as the sum of the instantaneous power of all controlled home appliances, i.e. where P j is the power of the jth device.

[0100] A user comfort function f2(C), the goal of which is to maximize, is defined as a fit score of a comprehensive environmental parameter and user preference setting. For example, f2(C) = 1 / (1+w T |T-T set |+w H |H-H set |), where T and H are the current temperature value and comfortable humidity, T set and Hset are the user's preset comfortable temperature value and comfortable humidity value, and w T and w H are the corresponding weight coefficients. The determination of the weight coefficients w T and w H includes a preset value method, which presets a set of standardized weight values for different geographical regions or seasons according to a large amount of statistical data, for example, setting w H higher than wT The user-defined method allows users to explicitly set w according to their personal preferences through the system's user interface. T and w H The relative size. Adaptive learning refers to the system recording user manual interventions in devices (such as adjusting air conditioner temperature or turning humidifier on and off). When the system detects that the user frequently adjusts the temperature but rarely intervenes in humidity, it will gradually increase w through a learning rate factor. T The value of is reduced accordingly, and w is decreased accordingly. H The value of w is used to automatically adapt the comfort model to the user's actual preferences. For ease of calculation and comparison, the weighting coefficients can be normalized to satisfy w. T +w H =1, but this is not a mandatory constraint. The absolute value of the weight coefficient can also be used to adjust the overall priority of the comfort function in the entire multi-objective optimization problem.

[0101] The equipment lifespan function f3(L) aims to minimize equipment wear and tear. It can be defined as a function related to the start-stop frequency of critical equipment (e.g., an air conditioning compressor) per unit time, i.e., minimizing... Where N switch,j Let α be the number of times device j starts and stops per unit time. j This is the loss weighting coefficient for the equipment. The loss weighting coefficient α j It is determined based on the specific equipment type, and the determination method includes: based on manufacturer data, which refers to the determination of α based on technical specifications such as Mean Time Between Failures (MTBF) or rated start-stop cycles provided by the equipment manufacturer. j Assign an initial baseline value. Based on equipment type and cost, this refers to setting α... j This is related to the replacement cost of the equipment and the physical wear and tear of the start-stop mechanism. For example, for air conditioning equipment that includes a compressor and mechanical relays, its α... AC The value was set much higher than the α value of an LED luminaire controlled by a solid-state relay. LED The value is used to reflect that the loss per start-stop cycle of the former is much greater than that of the latter.

[0102] The strategy generation unit iteratively evolves a set of solutions located at the Pareto front by executing standard operations of the NSGA-II algorithm, such as fast non-dominated sorting, crowding calculation, simulated binary crossover, and polynomial mutation. Each solution in this set corresponds to a specific device control strategy that achieves different balances among the three objectives mentioned above. Based on a preset meta-strategy (e.g., energy saving priority or comfort priority), the system selects a final composite strategy from this set of solutions, parses the strategy into a set of specific control instructions, and sends them to the strategy execution module.

[0103] Referring to Figure 5 , Figure 5 is a functional block diagram of a policy execution module 400 according to an embodiment of the present application. In specific embodiments of the present application, the policy execution module 400 is a terminal execution mechanism of a smart home system, which is used to convert the decision generated by the cloud into physical control of home appliances. The policy execution module 400 mainly includes an instruction analysis unit, a device linkage rule engine, and a device driver unit.

[0104] The instruction analysis unit communicates with the cloud processing module 300 through wireless or wired means. The instruction analysis unit is used to receive the control instruction issued by the cloud in the JSON format. After receiving, the instruction analysis unit parses the JSON string, extracts the fields including the target device ID, the action to be performed, and the specific parameter value (such as the brightness value, the temperature value), and converts them into the standardized control command format of the module, and then distributes them to the device driver unit.

[0105] The device linkage rule engine has its internal logic based on the Bayesian network model. The Bayesian network defines the sensor state, the user activity, and the device state in the home environment as nodes in the network, and the directed edges between the nodes represent the causal dependence relationship between them. Each node is attached with a conditional probability table (CPT), which is pre-set according to historical data statistics or expert knowledge. For example, for the hidden node of inferring cooking activity, its conditional probability table can be defined as P.

[0106] When the device linkage rule engine receives new evidence from the data acquisition module 100 or the internal bus of the system, such as receiving two evidence inputs of “kitchen door magnet state = open” and “kitchen infrared state = active”, it will use the belief propagation algorithm to perform probability inference in the Bayesian network, and calculate the posterior probability of the cooking activity node being “true”. If the calculated probability value exceeds a preset confidence threshold (for example, 0.9), the engine determines that the user is performing a cooking activity, and automatically generates a new control instruction with high confidence, such as triggering the exhaust hood to open and linking to adjust the gas stove to the preset fire power. The generated instruction is also sent to the device driver unit.

[0107] The device driving unit is the final interface for performing physical control. It is connected with the instruction analysis unit and the device linkage rule engine, and receives control instructions from the two. The device driving unit is internally integrated with multiple driving circuits, such as a relay circuit for controlling an alternating-current high-power electrical appliance or a TRIAC (Triode for Alternating Current) circuit for light dimming control.

[0108] When the device driving unit receives a control instruction, it first selects the corresponding driving circuit according to the device ID in the control instruction. Then, the circuit is operated according to the action in the control instruction. If the instruction is to turn on the power supply of the device, the device driving unit applies a driving voltage to the corresponding relay coil, so that the normally open contact of the relay is closed, thereby turning on the circuit for supplying power to the home appliance. If the control instruction is to turn off the power supply of the device, the device driving unit removes the voltage applied to the relay coil, so that the contact is disconnected under the action of the spring, thereby cutting off the circuit for supplying power to the home appliance, and completing the physical control of the running state of the device.

[0109] The smart home system of the embodiment of the present application can be applied to the scene of intervening in the specific behavior of patients with dementia.

[0110] In this application embodiment, the primary function of the system is to learn the user's daily behavior pattern. The data acquisition module 100 uses multiple millimeter wave radar sensors deployed in the user's bedroom, living room, and key path area leading to the main door to continuously collect the user's position coordinate data in the indoor space. At the same time, the pressure sensor deployed under the mattress provides binary state data of whether the user is in bed. The cloud processing module 300 receives these coordinate and state data streams containing timestamps. To construct the trajectory heat map of the user's indoor activities, the cloud processing module 300 first divides the indoor plane space into a two-dimensional grid. In each preset time window (for example, 5 minutes), the system counts the residence time of the user in each grid element, thereby generating an MxN-dimensional heat map matrix representing the activity distribution in the time window.

[0111] The behavior pattern learning model based on the LSTM (Long Short-Term Memory) network in the cloud processing module 300 takes a series of trajectory heat map matrices generated in a continuous time window (for example, 12, representing the past one hour) as an input sequence, and trains and learns the normal activity pattern of the user in different time periods (such as night) for several weeks. Through this process, the model establishes a baseline behavior pattern, for example, learns that the normal activity trajectory of the user from 2:00 to 4:00 in the morning is limited to a short path from the bedroom grid element to the bathroom grid element and back.

[0112] After the system enters the continuous monitoring stage, when the abnormal behavior deviating from the daily mode is identified, the system will trigger the intervention mechanism. In a specific running scenario, at 3:00 in the morning, the pressure sensor of the data acquisition module 100 reports a user getting-out-of-bed signal, and then the millimeter wave radar sensor array detects that the user's moving track continuously moves from the bedroom area to the door area. The cloud processing module 300 generates a track heat map sequence in real time, which presents a path completely inconsistent with the learned normal activity baseline at night.

[0113] At this time, the behavior pattern learning model produces a huge calculation deviation between the current input track sequence and the normal track sequence predicted by the model based on historical data when making a prediction. The deviation value exceeds the preset threshold. The scene recognition unit determines the current scene as a potential lost target scene accordingly. The strategy generation unit retrieves and matches the preset intervention strategy for this high-priority scene and generates a set of linked and determined control instruction set.

[0114] Finally, the strategy execution module 400 receives the instruction set and immediately executes a series of parallel intervention operations. The instruction analysis unit analyzes the instructions and executes through the device driving unit: sends encrypted instructions to the smart electronic lock deployed at the door, forces it to enter the locked state, and temporarily disables the regular unlocking function on the inside; sends instructions to the smart lighting devices in the living room and corridor area to adjust the color temperature to 2700K soft warm color temperature and set the brightness to 30%, and sends instructions to the lighting devices in the bedroom to increase the brightness to 70%; sends instructions to the smart sound box in the living room to play a pre-stored piece of soothing music familiar to the patient at a volume of 40 decibels. At the same time, the cloud processing module 300 pushes an emergency alert information to the mobile terminal application of the preset family member or caregiver through the communication service interface of the One NET platform, which contains the event type, occurrence time and current location of the user.

[0115] The principles and implementation manners of the present application are described in the specific embodiments, and the above embodiment descriptions are only used to help understand the method and core idea of the present application; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manners and application ranges will be changed, and the above description of the present application should not be understood as a limitation.

Claims

1. A smart home system based on the One NET platform, characterized in that, The system includes: The data acquisition module is used to collect raw data related to family activities and generate a multi-dimensional time-series feature vector representing the current family status based on the raw data. The data transmission module is used to upload the multidimensional time-series feature vector to the cloud processing module deployed on the One NET platform; The cloud processing module is used to receive the multi-dimensional time-series feature vector, and analyze the multi-dimensional time-series feature vector based on the activity scenario recognition model with a dynamic update mechanism to determine the target activity scenario of the current household. Then, based on the target activity scenario, the optimal composite energy and health strategy is generated and matched from the scenario-strategy mapping library using a multi-objective optimization algorithm. The strategy execution module is used to receive the control instructions contained in the composite energy and health strategy, and drive the home appliances to execute the control instructions through the device linkage rule engine.

2. The smart home system based on the One NET platform according to claim 1, characterized in that, The data acquisition module includes: A multi-source sensing unit is a sensor used to acquire the raw data, the sensor including an environmental state sensor and a physiological indicator sensor; The environmental status sensors include passive infrared sensors, door magnetic sensors, sound sensors, and AC current sensors; The physiological indicator sensor is a millimeter-wave radar used for non-contact monitoring of a user's heart rate, respiratory rate, body temperature, blood pressure, or blood sugar. The feature generation unit is used to process the original data to generate the multidimensional time-series feature vector.

3. A smart home system based on the One NET platform according to claim 2, characterized in that, The feature generation unit is specifically used for: The state signal of the passive infrared sensor is quantified as a binary feature value representing the presence of the human body; The state signal of the door magnetic sensor is quantized into binary feature values ​​representing the opening and closing of doors and windows; The analog output signal of the sound sensor is normalized to obtain the ambient sound decibel characteristic value. The output signal of the AC current sensor is normalized to obtain the characteristic value of the operating current of the electrical appliance; The output signal of the physiological index sensor is filtered and calibrated to obtain standardized physiological index characteristic values; All the above feature values ​​are concatenated and a timestamp is added to construct the multidimensional temporal feature vector.

4. A smart home system based on the One NET platform according to claim 1, characterized in that, The data transmission module includes: A wireless communication unit is used to establish a network connection between the data transmission module and the cloud processing module; The protocol encapsulation unit is used to encapsulate the multi-dimensional time-series feature vector according to the MQTTS protocol format and transmit it encrypted through the wireless communication unit. The One NET platform provides infrastructure and device authentication services for communication in the MQTTS protocol format.

5. A smart home system based on the One NET platform according to claim 4, characterized in that, The protocol encapsulation unit is specifically used for; Using the device identity credentials pre-installed in the data transmission module, two-way authentication is completed through the One NET platform and the cloud processing module to establish an initial encrypted communication channel; The two-way authentication includes a dynamic key update mechanism and an abnormal traffic detection algorithm. The session key of the encrypted communication channel is dynamically negotiated and updated according to a preset time period. The abnormal traffic detection algorithm is used to analyze the traffic patterns of data transmission in real time. When it detects an attack behavior that deviates from the normal pattern, it automatically cuts off communication and sends an alert to the user or system administrator.

6. A smart home system based on the One NET platform according to claim 1, characterized in that, The cloud processing module includes: The scenario recognition unit is used to receive the multi-dimensional temporal feature vector and calculate the target activity scenario based on the activity scenario recognition model with a dynamic update mechanism. The strategy generation unit is used to query the scenario-strategy mapping library using the target activity scenario and real-time physiological index data as retrieval conditions, and to generate and obtain the corresponding optimal composite strategy through the multi-objective optimization algorithm.

7. A smart home system based on the One NET platform according to claim 6, characterized in that, The activity scenario recognition model with a dynamic update mechanism determines the target activity scenario in the following ways: The activity scenario recognition model incorporates a user behavior prediction model based on a long short-term memory network. This user behavior prediction model learns historical multi-dimensional temporal feature vectors to predict user activity scenarios within the next time window. The real-time input multidimensional time-series feature vector V current The feature vector V predicted by the user behavior prediction model predicted Compare and calculate the deviation Δ=||V current -V predicted ||2; When the deviation Δ exceeds the preset threshold ∈, it indicates a change in user behavior pattern. The system then triggers an adaptive weight adjustment mechanism, employing a gradient descent algorithm to adjust the weight vector W in the scene recognition model. i The optimization is performed, and its update rules are expressed as follows: In the formula, W i (t+1) represents the updated weight; η is the learning rate; For the loss function J on the weight vector W i The gradient of (t); W i This refers to the weight vector in the context recognition model; When identifying the current context, calculate the current feature vector V. current The dynamically adjusted weight vector W corresponding to all preset activity scenarios i The dot product between the two scenarios yields the matching score for each scenario. And the highest scorer S max The scenario was identified as the target activity scenario.

8. A smart home system based on the One NET platform according to claim 1, characterized in that, The strategy execution module includes: The instruction parsing unit is used to receive and parse control instructions issued from the cloud processing module; The device linkage rule engine is built based on a Bayesian network. The device linkage rule engine receives multiple sensor status signals from the data acquisition module as evidence and performs probabilistic reasoning based on a preset conditional probability table to calculate the posterior probability of one or more preset user activity scenarios. When the posterior probability exceeds a preset confidence threshold, the linkage control command corresponding to the user activity scenario is automatically generated and transmitted. The device driving unit is connected to the instruction parsing unit and the device linkage rule engine, and includes a relay or switch circuit, for physically controlling the power supply to the home appliance based on the direct instructions from the instruction parsing unit or the instructions inferred by the device linkage rule engine.

9. A smart home system based on the One NET platform according to claim 8, characterized in that, The physical power on / off control specifically includes: Based on the parsing result of the instruction parsing unit, the contacts of the relay or switch circuit are controlled to close, so as to connect the circuit that supplies power to the home appliance. The contacts of the control relay or switch circuit are opened to cut off the circuit supplying power to the household appliance.

10. A smart home system based on the One NET platform according to claim 2, characterized in that, The smart home system is also applied to intervention scenarios to prevent dementia patients from wandering away, specifically including: In the data acquisition module, the millimeter-wave radar and the sensors deployed in key areas are used to construct a heat map of the user's indoor activity trajectory; In the cloud processing module, a behavior pattern learning model based on long short-term memory network is used to analyze the trajectory heatmap and activity time series in order to learn the patient's daily behavior patterns. When the behavior pattern learning model identifies abnormal behavior that deviates from the daily pattern, it determines it as a potential missing target scenario; The strategy execution module receives the intervention strategy for the potential lost target scenario and automatically executes the linkage operation, which includes encrypting the electronic lock of the door to prevent the person from going out, adjusting the indoor lights to a soft color temperature and playing soothing music familiar to the patient to guide them back to the bedroom, and sending an emergency alarm to the mobile terminal of a preset family member or caregiver.

Images