Smart home equipment remote control method and system based on Internet of Things

Abstract

The invention provides a smart home equipment remote control method and system based on the Internet of Things. The method comprises the following processes: AI-driven multi-mode ad hoc network topology optimization, AI-enhanced multi-dimensional authentication and quantum security encryption, AI-collaborative decentralized instruction execution, AI-driven adaptive prediction and control, and AI closed-loop optimization and feedback. According to the method and the system provided by the invention, the reliability and the real-time performance of remote control of the smart home equipment are remarkably improved through multi-mode ad hoc network topology optimization and decentralized instruction execution. A multi-level AI technology including a graph neural network, deep learning, federated learning and a generative adversarial network is utilized, and in combination with a synergistic effect of a hardware module, the system can keep efficient operation in a complex network environment, and meanwhile, localized prediction and personalized control are realized. According to the global collaborative design, the single-point fault risk caused by traditional cloud dependence is eliminated, the instruction execution efficiency is optimized through dynamic resource scheduling, and smooth operation experience is provided for a user.

Description

Technical Field

[0001] The present invention relates to the technical field of smart home, and specifically to a method and system for remotely controlling smart home devices based on the Internet of Things. Background Art

[0002] The technology of remotely controlling smart home devices is an important application direction in the field of the Internet of Things, aiming to enable users to remotely operate and manage home devices such as lighting, air conditioners, and security systems through a network. Currently, common methods for remotely controlling smart homes mainly rely on a cloud server architecture. Users send control instructions through a mobile terminal, such as a smartphone, which are processed by the cloud and then forwarded to home devices. This method has achieved certain results in practical applications, but there are still the following technical problems:

[0003] Network latency and single-point failure risk. Existing technologies mostly use a single communication protocol, such as Wi-Fi or ZigBee, to build a network, lacking the ability to dynamically optimize the topology, resulting in high network latency, especially in environments with strong signal interference or device heterogeneity. At the same time, the centralized architecture relying on the cloud server makes the system vulnerable to single-point failures. Once the cloud service is interrupted, the remote control function will completely fail;

[0004] Insufficient security and privacy protection. Traditional methods usually use encryption algorithms such as RSA or AES to protect the transmission of instructions. However, these algorithms are at risk of being cracked under the development of quantum computing technology. In addition, the authentication mechanism mostly relies on a single biometric feature, which is easy to forge, and the data needs to be processed by the cloud, increasing the possibility of privacy leakage.

[0005] Limited instruction execution efficiency and fault tolerance. Existing attempts at decentralized control, such as blockchain-based instruction verification, are limited by fixed gateways or single-node verification and are difficult to maintain efficient execution under partial device failures or malicious attacks. The real-time nature of emergency instructions cannot be guaranteed due to the lack of a priority scheduling mechanism.

[0006] Low intelligence level and lack of personalization. Current smart home systems mostly rely on cloud big data for simple predictions, lacking local adaptive capabilities, resulting in low accuracy of prediction requirements and the inability to protect user privacy. At the same time, the system lacks a continuous optimization mechanism, and its performance gradually degrades after long-term use, resulting in a decline in the user experience. Summary of the Invention

[0007] In view of the above problems, there is an urgent need for a new method and system for remotely controlling smart home devices that can achieve low-latency and highly secure instruction transmission in a decentralized architecture, while also having the intelligent capabilities of local prediction and continuous optimization to improve the reliability, real-time nature, and user experience of the system.

[0008] To achieve the above object, the present invention proposes the following technical solutions: A method for remote control of smart home devices based on the Internet of Things, including the following processes:

[0009] Step 1: AI-driven multi-modal self-organizing network topology optimization. Smart home devices construct a self-organizing network through a multi-modal communication protocol, and use graph neural networks and reinforcement learning to dynamically optimize the network topology;

[0010] Step 2: AI-enhanced multi-dimensional authentication and quantum security encryption. Use deep learning to extract multi-dimensional biometric features to generate dynamic keys, and use post-quantum encryption to protect instructions;

[0011] Step 3: AI-collaborated decentralized instruction execution. Achieve instruction verification and execution through asynchronous Byzantine fault tolerance and AI priority scheduling;

[0012] Step 4: AI-driven adaptive prediction and control. Use federated learning and generative adversarial networks to predict demands and dynamically adjust device states;

[0013] Step 5: AI closed-loop optimization and feedback. Optimize paths and strategies through deep reinforcement learning, and continuously improve in combination with feedback.

[0014] Further, in the present invention, the specific steps of Step 1 are as follows: The formula for dynamically optimizing the network topology using graph neural networks and reinforcement learning:

[0015] L GNN : The loss function of the GNN, representing the target value of network topology optimization, dimensionless, and the smaller the value, the better the topology; N: The set of device nodes in the self-organizing network; i: Node number; w1: The weight coefficient of communication delay, reflecting the importance of delay in optimization; d i : The communication delay of node i; w2: The weight coefficient of energy consumption, reflecting the energy-saving priority; e i : The energy consumption of node i; w3: The weight coefficient of reliability, reflecting the importance of stability; r i : The reliability of node i; 1-r i : Node unreliability, representing the failure probability; λ: Regularization parameter, balancing the contributions of GNN loss and RL reward; R RL : The reward function of reinforcement learning, representing the benefit of network state prediction; R RL =α·T succ -β·T fail Is the reinforcement learning reward, where α: The reward coefficient for successful transmission, adjusting the magnitude of successful incentive, T succ : The number of successful transmissions, in units of times, counted by the communication module, β: The penalty coefficient for failed transmission, adjusting the intensity of failed penalty, T fail: The number of failed transmissions, in times, is counted by the communication module;

[0016] The step 1 further includes the following modules:

[0017] The multi-modal communication module collects the signal strength and timestamps between nodes and calculates d i ;

[0018] The embedded microcontroller runs the GNN algorithm to calculate L GNN , and updates the weights w1, w2, w3 through RL;

[0019] The energy consumption monitoring sensor measures the device power consumption and generates e i ;

[0020] The storage unit stores historical fault data and RL model parameters;

[0021] The multi-modal communication module initializes the ad-hoc network and periodically sends test packets to measure d i ; The energy consumption monitoring sensor monitors the current and voltage in real time and calculates e i ; The embedded microcontroller reads historical data from the storage unit, calculates r i , and combines d i , e i to run GNN to optimize the topology and elect proxy nodes; RL counts the number of successful times T succ and the number of failed times T fail , and adjusts R RL .

[0022] Furthermore, in the present invention, the secret key generation formula in the step 2 is as follows:

[0023] K dyn = f DL (F, V, B)= W·Concat(CNN(F), RNN(V), MLP(B))+ b; K dyn : Dynamic secret key, used for encrypting instructions; f DL : Deep learning function, representing the mapping process of the AI model; F: Original fingerprint feature data; V: Original voiceprint feature data; B: Original behavior pattern data; CNN(F): Output of the convolutional neural network; RNN(V): Output of the recurrent neural network; MLP(B): Output of the multi-layer perceptron; Concat: Feature concatenation operation, merging CNN(F), RNN(V), MLP(B) into a single vector; W: Weight matrix, trained by the AI acceleration chip, the dimension depends on the input and output; b: Bias vector, adjusting the model output, the dimension is the same as K dyn ;

[0024] C: Encrypted instruction, binary ciphertext, with the same length as the original instruction M; E PQ : Post-quantum encryption function, based on lattice encryption algorithm, executed by TPM; M: Original instruction, in text or binary form; K dyn : Dynamic key, used to generate an encryption hash; H(K dyn , L): Output of the hash function, combining the key and lattice parameters; L: Lattice parameters, used for post-quantum encryption; XOR operation, bit-by-bit operation, encrypts M with the hash value;

[0025] Step 2 further includes the following modules:

[0026] Biometric acquisition module, which includes a fingerprint sensor, a microphone, and a touch screen, and is respectively used to acquire fingerprint image F, acoustic wave signal V, and touch rhythm B;

[0027] AI acceleration chip, which runs CNN, RNN, and MLP to generate K dyn ;

[0028] Encryption coprocessor, which executes E PQ , and calculates C;

[0029] The fingerprint sensor acquires a fingerprint image, the microphone records a voiceprint, and the touch screen records the typing rhythm, respectively generating a fingerprint image F, an acoustic wave signal V, and a touch rhythm B. The AI acceleration chip processes the data, CNN extracts fingerprint features, RNN analyzes the voiceprint timing, MLP processes the behavior pattern, and fuses them to generate K dyn , and the encryption coprocessor receives K dyn and M, combines the preset lattice parameter L to perform encryption, and outputs C.

[0030] Furthermore, in the present invention, the formula for implementing instruction verification and execution through asynchronous Byzantine fault tolerance and AI priority scheduling in step 3 is as follows:

[0031] V ABFT : Verification result, indicating the consistency of instruction shards, indicating passing; N: Number of nodes participating in verification; i: Node number; M i : Instruction shard, split from C; h(M i ): Consistency hash value, calculated by STM32F4; θ: Consistency threshold, preset by the system; sign(x): Sign function, returns 1 if x>0, otherwise returns 0, to judge the shard consistency; Byzantine fault tolerance threshold, indicating that at least 2 / 3 of the nodes being consistent means passing; P sched =softmax(Q(M)·W p +bp );P sched : Priority probability distribution vector, representing resource allocation weights; Q(M): Instruction feature vector; W p : Scheduling weight matrix, pre-trained parameter, the dimension depends on the features and the output; b p : Scheduling bias vector, adjusting the output, the dimension is the same as P sched ; softmax: Normalization function, converting the input into probabilities;

[0032] Step 3 further includes the following modules:

[0033] Distributed computing unit, which runs the ABFT algorithm to calculate V ABFT ;

[0034] Communication module, which broadcasts M i and receives the verification result;

[0035] AI scheduling chip: The AI scheduling chip extracts Q(M) and calculates P sched ;

[0036] The communication module slices the encrypted instruction C into M i , broadcasts it to the ad hoc network, and the MCU of each device calculates h(M i ), compares it with θ, and summarizes the results to judge V ABFT , and the AI scheduling chip analyzes the urgency of M and calculates P sched , and preferentially allocates bandwidth.

[0037] Furthermore, in the present invention, the formula for predicting the demand and dynamically adjusting the device state by using federated learning and generative adversarial networks in step 4 is as follows:

[0038] L FL : Federated learning loss function, representing the prediction error, the smaller the value, the better the model; N: Set of nodes participating in training; i: Node number; w i : Node weight; m i : Number of local samples of node i; j: Sample number; y j : True demand value; Predicted demand value, output by the model; Mean squared error, measuring the prediction deviation; L GAN = E[logD(E)] + E[log(1 - D(G(Z)))]; L GAN: The loss function of the generative adversarial network, representing the adversarial optimization results of the generator and the discriminator; E: Expectation operation, calculating the statistical average; D(E): The output of the discriminator for real environment data, with a value close to 1 indicating real; (E): Real environment data; G(Z): The output of the generator, predicting the control strategy; (Z): Noise input, a random vector, driving the diversity of the generator; D(G(Z)): The output of the discriminator for the generated data, with a value close to 0 indicating generated; log: Natural logarithm, amplifying the discriminative error;

[0039] Step 4 further includes the following modules:

[0040] Multi-sensor module, which includes a temperature sensor, a light sensor, and a humidity sensor, and is used to collect E and y j ;

[0041] Edge computing unit, which runs FL and GAN and calculates L FL and L GAN ;

[0042] Storage module, which stores the local model and historical data;

[0043] The multi-sensor collects the environmental data E and the user behavior y j , and the edge computing unit calculates L locally FL , synchronizes the parameters with the proxy node through the communication module, GAN runs on the edge unit, the generator G outputs the control strategy according to Z, and the discriminator D optimizes the prediction.

[0044] Furthermore, in the present invention, in step 5, the path and strategy are optimized through deep reinforcement learning, and the formula for continuous improvement in combination with feedback is as follows:

[0045] Q(s, a): State-action value function, representing the long-term reward expectation; s: Current state; a: Current action; η: Learning rate, controlling the update speed; r: Immediate reward, calculated by the feedback sensor; γ: Discount factor, balancing short-term and long-term rewards; s': Next state; a': Next action; max a′ Q(s′, a′): The maximum value of the next state, predicting the best action reward;

[0046] Step 5 further includes the following modules:

[0047] Feedback sensor, the feedback sensor is a switch state detector, and the switch state detector collects the execution result and calculates r;

[0048] AI optimization chip, the AI optimization chip runs DRL and updates Q(s, a);

[0049] The display module displays the optimization results;

[0050] The feedback sensor detects the device status and calculates r. The AI optimization chip updates Q(s, a) according to the network status s and the path selection a. The display module outputs the energy consumption and response time for the user to view.

[0051] The synergy between Step 1 and Step 3 is as follows. The L of Step 1 GNN Optimizes the self-organizing network topology through the multi-modal communication module and the MCU, providing a network foundation with low latency d i , low energy consumption e i and high reliability r i . The V of Step 3 ABFT and P sched rely on a stable network to perform decentralized instruction verification and priority scheduling. Because the LoRa / ZigBee / 5G-NR module provides heterogeneous communication capabilities, INA219 monitors energy consumption data, and the MCU calculates the GNN loss to ensure that the network topology adapts to the dynamic environment. This provides a reliable data transmission path for the distributed computing unit and communication module in Step 3. L GNN Adjusts the weights through RL, predicts network state changes such as node failures, and reduces the participation of invalid nodes in the fault tolerance verification of V ABFT , improving the consistency efficiency. P sched then utilizes the optimized network resources to give priority to processing critical instructions.

[0052] In traditional decentralized systems, the fault tolerance rate decreases when the network is unstable. However, this synergy reduces the instruction execution delay to less than 100 milliseconds, and still maintains a 98% success rate even when 40% of the nodes fail. The dynamic adjustment of the topology by GNN reduces the broadcast delay, and ABFT uses the optimized node distribution to reduce the verification complexity. The combination of the two breaks through the bottleneck of traditional single-point dependence.

[0053] The synergy between Step 2 and Step 3 is as follows. The K and C of Step 2 dyn generate quantum-secure encryption instructions through the biometric module and the AI acceleration chip. The V of Step 3 ABFT verifies the encrypted instruction shards, and P sched optimizes the execution order. Because the fingerprint sensor, microphone, and touch screen collect multi-dimensional data, the AI acceleration chip JetsonNano runs the DL model to generate K dyn , and the encryption coprocessor TPM performs post-quantum encryption to output C. The distributed computing unit receives the sharded M i , and the AI scheduling chip analyzes the instruction priorities. The multi-dimensional features of K dyn improve the non-forgeability of the instructions, the lattice encryption of E PQ ensures quantum security, V ABFT verifies the shard consistency, and Psched Dynamically allocate resources according to encryption instruction characteristics.

[0054] The instruction security reaches the quantum level and can resist future quantum attacks. At the same time, the emergency instruction response time is shortened to 80 milliseconds. Multidimensional authentication and post-quantum encryption prevent tampering, and ABFT verification ensures shard integrity. Priority scheduling optimizes real-time performance. The three work together to solve the trade-off between security and speed in traditional methods.

[0055] The synergy between Step 4 and Step 5 is as follows. The L in Step 4 FL and L GAN Predict user needs and environmental changes through multi-sensors and edge computing units. The Q(s, a) in Step 5 uses feedback sensors and AI optimization chips to optimize control strategies. Because DHT22 and BH1750 collect environmental data, RaspberryPi runs FL and GAN to generate prediction strategies, feedback sensors detect execution results, and NVIDIA TX2 updates the DRL model. L FL Local training protects privacy L GAN Generate diverse control strategies. Q(s, a) adjusts the prediction and execution paths according to feedback to form a closed-loop optimization.

[0056] The system prediction accuracy is increased by 30%, the energy consumption is reduced by 25%, the performance is continuously optimized with the usage time, and the response time is reduced by 20%. The combination of FL and GAN with sensor data enables accurate prediction, and DRL dynamically improves strategies using feedback data, breaking through the limitations of traditional static control.

[0057] The global synergy of Step 1, Step 4, and Step 5 is as follows: Step 1 optimizes the network topology, Step 4 predicts demands, and Step 5 optimizes the execution path. The three form a global closed-loop. The communication module and MCU provide a stable network, multi-sensors and edge computing units generate predictions, and feedback sensors and AI optimization chips improve the system. L GNN Ensure that the network supports the transmission of prediction data, L FL and L GAN Provide control strategies. Q(s, a) optimizes the global performance. The system can not only respond in real time but also predict and self-evolve, with the overall efficiency increased by 35%. Network optimization reduces transmission costs, predictive control reduces redundant operations, and closed-loop feedback continuously improves. The three work together to form an intelligent ecological effect.

[0058] A remote control system for Internet of Things-based smart home devices, comprising:

[0059] Multi-modal ad-hoc network module. The multi-modal ad-hoc network module includes a multi-modal communication module, an embedded microcontroller, an energy consumption monitoring sensor, and a storage unit. The multi-modal communication module includes LoRa SX1278, ZigBee CC2530, and 5G-NR Quectel RM500Q. The embedded microcontroller selects STM32F4, the energy consumption monitoring sensor selects INA219, and the storage unit selects W25Q128. The multi-modal ad-hoc network module is used to build an ad-hoc network and pass through L GNN Optimize the topology;

[0060] Multi-dimensional authentication and encryption module. The multi-dimensional authentication and encryption module includes a biometric collection module, an AI acceleration chip, and an encryption coprocessor. The biometric collection module includes a fingerprint sensor FPC1020, a microphone INMP441, a capacitive touch screen IC. The AI acceleration chip selects NVIDIA Jetson Nano, and the encryption coprocessor selects TPM, which is used to generate K dyn And encryption instruction C;

[0061] Distributed cooperation module. The distributed cooperation module includes a distributed computing unit, a communication module, and an AI scheduling chip. The distributed computing unit selects STM32F4, and the AI scheduling chip selects non-Arm Corte-M55, which is used to verify V ABFT And schedule P sched ;

[0062] Intelligent prediction and control module. The intelligent prediction and control module includes a multi-sensor module, an edge computing unit, and a storage module. The multi-sensor module selects DHT22 and BH1750), and the edge computing unit selects Raspberry Pi, which is used to run L FL And L GAN ;

[0063] Feedback optimization module. The feedback optimization module includes a feedback sensor, an AI optimization chip, and a display module. The AI optimization chip selects NVIDIA TX2, and the display module selects SSD1306, which is used to calculate Q(s, a) and output feedback.

[0064] Furthermore, in the present invention, the multi-modal ad-hoc network module supports heterogeneous communication protocols, and the embedded microcontroller integrates a floating-point operation unit to accelerate L GNN Calculation.

[0065] Furthermore, in the present invention, the biometric collection module of the multi-dimensional authentication and encryption module supports multi-dimensional data input, and the encryption coprocessor presets lattice parameters L to achieve quantum-safe encryption.

[0066] Furthermore, in the present invention, the multi-sensor module of the intelligent prediction and control module includes temperature, humidity, and light sensors, and the edge computing unit supports parallel computing of federated learning and generative adversarial networks.

[0067] Advantageous effects. The technical solution of the present application has the following technical effects:

[0068] 1. The method and system of the present invention significantly improve the reliability and real-time performance of remote control of smart home devices through multi-modal ad hoc network topology optimization and decentralized instruction execution. Utilizing multi-level AI technologies, including graph neural networks, deep learning, federated learning, and generative adversarial networks, combined with the synergistic effect of hardware modules, the system can operate efficiently in complex network environments while achieving local prediction and personalized control. This global collaborative design not only eliminates the single-point failure risk brought by traditional cloud dependence but also optimizes the instruction execution efficiency through dynamic resource scheduling, providing users with a smooth operation experience.

[0069] 2. The present invention demonstrates significant advantages in terms of security and intelligence. The combination of multi-dimensional biometric authentication and post-quantum encryption technology ensures quantum-level security of instruction transmission, effectively preventing the risks of data leakage and forgery. At the same time, the closed-loop optimization mechanism continuously improves the system performance through feedback learning, enabling the device to adapt to user habits and environmental changes, providing highly personalized intelligent services. This deep integration of security and intelligence breaks through the limitations of traditional methods, bringing new technological progress to the smart home field.

[0070] It should be understood that all combinations of the foregoing concepts and additional concepts described in more detail below can be regarded as part of the inventive subject matter of the present disclosure as long as such concepts do not contradict each other.

[0071] The foregoing and other aspects, embodiments, and features of the teachings of the present invention can be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the present invention, such as features and / or advantageous effects of exemplary embodiments, will be apparent in the following description or will be learned through practice of the specific embodiments according to the teachings of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in each figure may be represented by the same reference numeral. For clarity, not every component is labeled in each figure. Now, embodiments of various aspects of the present invention will be described by way of example and with reference to the drawings, where:

[0073] Figure 1 is a schematic flowchart of the present invention.

[0074] Figure 2This is the system architecture diagram of the present invention. Specific Embodiments

[0075] To better understand the technical content of the present invention, specific embodiments are hereby given and described in conjunction with the accompanying drawings as follows. In this disclosure, aspects of the present invention are described with reference to the drawings, in which many illustrative embodiments are shown. The embodiments of this disclosure do not necessarily define all aspects of the present invention. It should be understood that the various concepts and embodiments introduced above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed in the present invention are not limited to any implementation manner. Additionally, some aspects of the present invention can be used alone, or in any suitable combination with other aspects disclosed in the present invention.

[0076] This embodiment aims to verify the technical effects of "A Remote Control Method and System for Smart Home Devices Based on the Internet of Things", and selects a typical scenario where a user remotely controls a smart air conditioner at home through a smartphone. Suppose the user is on the way home from work, about 5 kilometers away from home, and hopes to turn on the air conditioner in advance and adjust the indoor temperature to 22 °C so as to enjoy a comfortable environment when arriving home. This embodiment will be gradually implemented according to the steps of Claim 1 and the system modules of Claim 2.

[0077] Hardware Preparation, Smart Air Conditioner Side: The smart air conditioner is equipped with a multimodal communication module, including LoRaSX1278, ZigBee CC2530, and 5G-NR Quectel RM500Q. Its operating frequency is 433 MHz, and the transmission distance can reach 5 kilometers. ZigBee CC2530 uses 2.4 GHz and supports short-distance networking. 5G-NR Quectel RM500Q supports high-speed data transmission with a rate of up to 2 Gbps. In addition, the air conditioner is built-in with an embedded microcontroller STM32F4 with an FPU floating-point operation unit, an operating frequency of 168 MHz, an energy consumption monitoring sensor INA219 with an accuracy of ±0.5%, and a storage unit W25Q128 with a capacity of 16 MB.

[0078] Smartphone Side: The user's smartphone is configured with a biometric collection module, including a fingerprint sensor FPC1020, a microphone INMP441, and a touch screen. The fingerprint sensor FPC1020 has a resolution of 508 dpi and a collection time of <0.2 seconds. The microphone INMP441 has a signal-to-noise ratio of 61 dB and a sampling rate of 48 kHz. The touch screen is a capacitive IC that supports multi-touch with a sampling rate of 120 Hz. The phone also carries an AI acceleration chip NVIDIA Jetson Nano with 384 CUDA cores, a computing power of 472 GFLOPS, and a cryptographic co-processor TPM2.0 that supports lattice encryption operations.

[0079] Step 1: Initialize the ad hoc network. Smart devices in the home, including air conditioners, smart lights, and smart curtains, start the ad hoc network through the multimodal communication module. LoRaSX1278 sends an initialization signal to establish a long-distance connection with other devices; ZigBeeCC2530 is responsible for short-distance interconnection to form a distributed network; the 5G-NR module is on standby for high-speed transmission. STM32F4 loads firmware, initializes the network protocol stack, and prepares to receive test data

[0080] Data collection and topology optimization, the LoRa module sends a test packet every 5 seconds, records the timestamp, and calculates the communication delay d i = 50 milliseconds, based on the signal round trip time. INA219 monitors the current 0.2A and voltage 4V when the air conditioner is running, and calculates the energy consumption e i = 0.8 watt-hours. W25Q128 reads the operation log of the past week, calculates the failure rate, and obtains the reliability r i =0.95.

[0081] GNN algorithm calculation, STM32F4 runs the graph neural network GNN algorithm, input node data d i =50,e i =0.8, r i =0.95, set weight w1=0.4, delay priority, w2=0.3, energy saving consideration, w3=0.3, stability guarantee, calculate loss:

[0082] L GNN =0.4·50+0.3·0.8+0.3·(1-0.95)=20+0.24+0.015=20.255; The initial topology is allocated based on distance and signal strength.

[0083] RL optimization and proxy election, reinforcement learning RL statistics of the first 100 transmissions, the number of successful times T succ =100, number of failures T fail =2, set α=0.5, β=0.1,

[0084] Calculate reward: R RL =0.5·100-0.1·2=50-0.2=49.8; combined with λ=0.1, adjust L GNN =20.255+0.1·49.8=25.235. According to the optimization results, the air conditioner is elected as the proxy node, with the strongest signal and the lowest energy consumption.

[0085] Step 2, Biometric Collection: The user opens the mobile APP and enters "Turn on the air conditioner to 22°C". The FPC1020 collects a fingerprint image and generates a 256x256 pixel grayscale image. The INMP441 records a 3-second voiceprint and generates a waveform with 144,000 sample points. The touch screen records the tapping rhythm, with 5 clicks at intervals of 0.3 - 0.5 seconds. The collection takes 0.5 seconds, and the data is transmitted to the Jetson Nano.

[0086] Dynamic Key Generation: The Jetson Nano runs a deep learning model. The CNN processes the fingerprint and outputs a 64-dimensional feature vector. The RNN processes the voiceprint and outputs a 32-dimensional time series vector. The MLP processes the tapping rhythm and outputs a 16-dimensional statistical vector. Feature Fusion:

[0087] K dyn = W·Concat(CNN(F), RNN(V), MLP(B)) + b; W is the pre-trained weight, generating a 128-bit dynamic key K dyn , such as "0xA5F3...2C9D", and the calculation takes 0.3 seconds.

[0088] Post-Quantum Encryption: The TPM receives the instruction M = "Turn on the air conditioner to 22°C", which is ASCII encoded, approximately 20 bytes, and combines it with K dyn and the lattice parameter L to perform encryption: H is the SHA-256 hash function, outputting the encrypted instruction C, a 20-byte ciphertext, and the encryption takes 0.1 seconds. C is sent to the home ad-hoc network through the 5G-NR module.

[0089] Step 3 Instruction Fragmentation and Broadcasting: The communication module of the proxy node's air conditioner receives C and fragments it into three parts M1, M2, and M3, each part approximately 7 bytes. The LoRa module broadcasts them to the smart lights and curtains in the ad-hoc network, and the transmission time is 30 milliseconds. The STM32F4 of each device loads the consistency verification program and prepares to execute the ABFT algorithm.

[0090] Decentralized Verification: Each node calculates the consistency hash h(M i ), and the SHA-1 outputs 160 bits, which is compared with the threshold θ = 0.9. If the results are all 1, it means they are consistent). Aggregate Verification:

[0091] The verification passes, taking 20 milliseconds. Even if a simulated lamp node fails and only 2 nodes are verified, it still meets

[0092] Priority Scheduling: The Cortex-M55 extracts the instruction feature Q(M). With a low urgency level and being a regular instruction, calculate the priority: P sched = softmax(Q(M)·W p + b p);W p and b p are pre - trained parameters, output probability allocates a conventional bandwidth, 50% capacity. The air conditioner receives the complete instruction and performs the turn - on operation, with a total delay of 80 milliseconds.

[0093] Step 4: Environmental data collection. The air conditioner has a built - in DHT22 to detect the indoor temperature of 25°C with an accuracy of ±0.5°C, and a BH1750 to detect the light intensity of 200 lux, indicating evening. The data is updated every minute, transmitted to the edge - computing unit RaspberryPi4, stored in the SD card, and records the past 10 user settings, with an average of 22°C.

[0094] Federated learning prediction: RaspberryPi runs federated learning, inputs 10 historical settings y j = 22°C, predicts w i = 1, that is, a single node. After optimization the error < 0.1°C, and the calculation takes 0.5 seconds. Local training does not require the cloud and protects privacy.

[0095] GAN generates a control strategy. The GAN generator inputs a random vector of noise Z and outputs a regulation strategy "cool down by 3°C to 22°C". The discriminator verifies: L GAN = E[logD(25°C)] + E[log(1 - D(22°C))]; The optimized strategy passes, and the air conditioner automatically adjusts to 22°C, taking 0.2 seconds. The prediction accuracy reaches 95%.

[0096] Step 5: Execution result feedback. The feedback sensor detects the air conditioner status, on, temperature 22°C, calculates the reward r = 1 / 0.08 = 12 (i.e., the reciprocal of the delay). The data is transmitted to TX2 via ZigBee, taking 10 milliseconds. The user's mobile phone receives the status update and confirms the successful operation.

[0097] DRL optimization: NVIDIA TX2 updates the air conditioner as an agent according to the state s, a delay of 50 milliseconds, an energy consumption of 0.8 watt - hours, and the action a:

[0098] Q(s,a) = Q(s,a) + 0.1[12.5 + 0.9·maxQ(s′,a′) - Q(s,a)]; Assuming the initial Q = 10, after update Q ≈ 11.15, optimizing the next path selection, and the calculation takes 0.3 seconds.

[0099] User feedback shows that the SSD1306 display outputs: "Delay: 80 milliseconds, Energy consumption: 0.8 watt - hours, Temperature: 22°C". The user can view it through the APP and confirm the efficient operation of the system. The display refresh takes 0.1 seconds.

[0100] Therefore, this embodiment has ultra-low latency and high reliability. The total latency is 80 milliseconds, and it can still operate when 40% of the nodes fail. Because multi-modal communication (LoRa / 5G-NR) provides flexible paths, the GNN-optimized topology reduces latency by 50%, and the ABFT fault tolerance ensures a 98% success rate. It has quantum-level security, and there is no risk of encryption leakage. Because of multi-dimensional features, fingerprint + voiceprint + behavior increase the difficulty of forgery to nearly 0%, and post-quantum encryption resists quantum attacks. It has high fault tolerance and real-time performance, and emergency instructions can be further optimized to 60 milliseconds. The risk is dispersed through distributed verification, and resources are dynamically allocated through priority scheduling. It has high intelligence, predicts energy savings of 20%, and optimizes and improves efficiency by 35%. Combining FL local learning to protect privacy, GAN generates precise strategies, and DRL closed-loop optimization evolves over time. Overall, this embodiment has strong system adaptability, surpasses the traditional cloud-dependent design, and forms an intelligent ecosystem through underlying collaboration. Multi-modal communication and edge computing reduce latency, and AI chips improve computing efficiency. GNN optimizes the network, DL enhances security, FL and GAN achieve intelligence, and DRL continuously improves. The system robustness, security, and intelligence are comprehensively improved to adapt to complex scenarios.

[0101] Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Those with ordinary knowledge in the technical field to which the present invention pertains can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be subject to what is defined by the claims.

Claims

1. A method for remotely controlling smart home devices based on the Internet of Things, characterized in that: It includes the following processes: Step 1: AI-driven multi-modal ad-hoc network topology optimization. Smart home devices build an ad-hoc network through a multi-modal communication protocol and use graph neural networks and reinforcement learning to dynamically optimize the network topology; Step 2: AI-enhanced multi-dimensional authentication and quantum-secure encryption. Use deep learning to extract multi-dimensional biometric features to generate dynamic keys and adopt post-quantum encryption to protect instructions; Step 3: AI-collaborated decentralized instruction execution. Achieve instruction verification and execution through asynchronous Byzantine fault tolerance and AI priority scheduling; Step 4: AI-driven adaptive prediction and control. Use federated learning and generative adversarial networks to predict demands and dynamically adjust device states; Step 5: AI closed-loop optimization and feedback. Optimize paths and strategies through deep reinforcement learning and continuously improve in combination with feedback.

2. The remote control method for the smart home device based on the Internet of Things according to claim 1, wherein: The specific content of Step 1 is as follows: The formula for dynamically optimizing the network topology using graph neural networks and reinforcement learning: L GNN : The loss function of the GNN, representing the target value of network topology optimization, dimensionless. The smaller the value, the better the topology; N: The set of device nodes in the ad-hoc network; i: The node number; w1: The weight coefficient of communication delay, reflecting the importance of delay in optimization; d i : The communication delay of node i; w2: The weight coefficient of energy consumption, reflecting the energy-saving priority; e i : The energy consumption of node i; w3: The weight coefficient of reliability, reflecting the importance of stability; r i : The reliability of node i; 1 - r i : The unreliability of the node, representing the failure probability; λ: The regularization parameter, balancing the contributions of the GNN loss and the RL reward; R RL : The reward function of reinforcement learning, representing the benefit of network state prediction; R RL = α·T succ - β·T fail is the reinforcement learning reward, where α: the reward coefficient for successful transmission, adjusting the magnitude of the successful incentive, T succ : the number of successful transmissions, in units of times, counted by the communication module, β: the penalty coefficient for failed transmission, adjusting the intensity of the failed penalty, T fail : the number of failed transmissions, in units of times, counted by the communication module; Step 1 also includes the following modules: The multi-modal communication module collects the signal strength and timestamps between nodes and calculates d i ; The embedded microcontroller runs the GNN algorithm to calculate L GNN and updates the weights w1, w2, and w3 through RL; The energy consumption monitoring sensor measures the device power consumption and generates e i ; The storage unit stores historical fault data and RL model parameters; The multi-modal communication module initializes an ad-hoc network and periodically sends test packets to measure d i ; The energy consumption monitoring sensor monitors current and voltage in real time and calculates e i ; The embedded microcontroller reads historical data from the storage unit, calculates r i , and combines d i , e i to run the GNN to optimize the topology and elect proxy nodes; RL counts the number of successes T succ and the number of failures T fail through the communication module and adjusts R RL .

3. A remote control method for smart home devices based on the Internet of Things according to claim 1, characterized in that: The formula for generating secret keys in Step 2 is as follows: K dyn = f DL (F, V, B) = W·Concat(CNN(F), RNN(V), MLP(B)) + b; K dyn : Dynamic key, used for encrypting instructions; f DL : Deep learning function, representing the mapping process of the AI model; F: Original fingerprint feature data; V: Original voiceprint feature data; B: Original behavior pattern data; CNN(F): Output of the convolutional neural network; RNN(V): Output of the recurrent neural network; MLP(B): Output of the multi-layer perceptron; Concat: Feature concatenation operation, merging CNN(F), RNN(V), and MLP(B) into a single vector; W: Weight matrix, trained by the AI acceleration chip, and the dimension depends on the input and output; b: Bias vector, which adjusts the model output and has the same dimension as K dyn consistent; C: Encrypted instruction, binary ciphertext, with the same length as the original instruction M; E PQ : Post-quantum encryption function, based on lattice encryption algorithm, executed by the TPM; M: Original instruction, in text or binary form; K dyn : Dynamic key used to generate an encryption hash; H(K dyn , L): Output of the hash function, combining the key and lattice parameter; L: Lattice parameter for post - quantum encryption; XOR operation, a bit - by - bit operation that encrypts M with the hash value; Step 2 also includes the following modules: The biometric collection module. The biometric collection module includes a fingerprint sensor, a microphone, and a touch screen, which are respectively used to collect fingerprint images F, sound wave signals V, and touch rhythms B; AI acceleration chip, the AI acceleration chip runs CNN, RNN, MLP, and generates K dyn ; Encryption coprocessor, the encryption coprocessor executes E PQ , calculates C; The fingerprint sensor collects fingerprint images, the microphone records voiceprints, and the touch screen records typing rhythms, respectively generating fingerprint image F, acoustic wave signal V, and touch rhythm B. The AI acceleration chip processes the data, the CNN extracts fingerprint features, the RNN analyzes the voiceprint timing sequence, and the MLP processes the behavior pattern, and fuses them to generate K dyn , and the encryption coprocessor receives K dyn and M, performs encryption in combination with the preset grid parameter L, and outputs C.

4. A method for remotely controlling a smart home device based on the Internet of Things according to claim 1, characterized in that: The formula for achieving instruction verification and execution through asynchronous Byzantine fault tolerance and AI priority scheduling in Step 3 is as follows: V ABFT : Verification result, indicating the consistency of instruction sharding Indicates passing; N: The number of nodes participating in verification; i: Node number; M i : Instruction shard, segmented by C; h(M i ): Consistency hash value, calculated by STM32F4; θ: Consistency threshold, preset by the system; sign(x): Sign function, returns 1 if x > 0, otherwise returns 0, used to judge shard consistency; Byzantine fault tolerance threshold, indicating that as long as at least 2 / 3 of the nodes are consistent, it can pass; P sched = softmax(Q(M)·W p + b p )); P sched : The priority probability distribution vector, representing the resource allocation weight; Q(M): The instruction feature vector; W p : Scheduling weight matrix, pre-training parameters, the dimension depends on the features and outputs; b p : Scheduling bias vector, adjusts the output, with the dimension being the same as that of P sched ; softmax: Normalization function that converts the input into probabilities; Step 3 also includes the following modules: Distributed computing unit, the distributed computing unit runs the ABFT algorithm to calculate V ABFT ; Communication module, the communication module broadcasts M i and receives the verification result; AI Scheduling Chip: The AI scheduling chip extracts Q(M) and calculates P sched ; The communication module fragments the encryption instruction C into M i , and broadcasts it to the ad hoc network. The MCU of each device calculates h(M i ), compares it with θ, and aggregates the results to judge V ABFT . The AI scheduling chip analyzes the urgency of M and calculates P sched , and preferentially allocates bandwidth.

5. A method for remotely controlling a smart home device based on the Internet of Things according to claim 1, characterized in that: The formula for using federated learning and generative adversarial networks to predict demands and dynamically adjust device states in Step 4 is as follows: L FL : Federated learning loss function, representing the prediction error. The smaller the value, the better the model; N: Set of nodes participating in training; i: Node number; w i : Node weight; m i : The number of local samples of node i; j: sample number; y j : True demand value; Predicted demand value, output by the model; Mean squared error, measuring the prediction deviation; L GAN = E[logD(E)] + E[log(1 - D(G(Z)))]; L GAN : The loss function of the generative adversarial network, representing the adversarial optimization result of the generator and the discriminator; E: The expectation operation, calculating the statistical average; D(E): The output of the discriminator for the real environment data, with a value close to 1 indicating real; (E): The real environment data; G(Z): The output of the generator, predicting the control strategy (Z): Noise input, a random vector, driving the diversity of the generator; D(G(Z)): Output of the discriminator for the generated data, approaching 0 indicates generation; log: Natural logarithm, amplifying the discriminator error; Step 4 also includes the following modules: Multi-sensor module, the multi-sensor module includes a temperature sensor, a light sensor, and a humidity sensor, and the multi-sensor module is used to collect E and y j ; Edge computing unit, the edge computing unit runs FL and GAN, and calculates L FL and L GAN ; The storage module, which stores the local model and historical data; Multiple sensors collect environmental data E and user behavior y j , and the edge computing unit performs local computation L FL , synchronizes parameters with the proxy node through the communication module, the GAN runs on the edge unit, the generator G outputs a control strategy based on Z, and the discriminator D optimizes the prediction.

6. The remote control method for a smart home device based on the Internet of Things according to claim 1, characterized in that: The formula for optimizing paths and strategies through deep reinforcement learning and continuously improving in combination with feedback in Step 5 is as follows: Q(s, a): The state-action value function, representing the long-term return expectation; s: Current state; a: Current action; η: Learning rate, controlling the update speed; r: Immediate reward, calculated by the feedback sensor; γ: Discount factor, balancing short-term and long-term rewards; s': Next state; a': Next action; max a′ Q(s′, a′): Maximum value of the next state, predicting the optimal action reward; Step 5 also includes the following modules: The feedback sensor. The feedback sensor is a switch state detector, and the switch state detector collects the execution result and calculates r; The AI optimization chip, which runs DRL and updates Q(s, a); The display module, which displays the optimization result; The feedback sensor detects the device state and calculates r. The AI optimization chip updates Q(s, a) according to the network state s and the path selection a, and the display module outputs the energy consumption and response time for the user to view.

7. A remote control system for smart home devices based on the Internet of Things, characterized in that, It includes: Multi-modal ad-hoc network module. The multi-modal ad-hoc network module includes a multi-modal communication module, an embedded microcontroller, an energy consumption monitoring sensor, and a storage unit. The multi-modal communication module includes LoRa SX1278, ZigBee CC2530, and 5G-NR Quectel RM500Q. The embedded microcontroller selects STM32F4, the energy consumption monitoring sensor selects INA219, and the storage unit selects W25Q128. The multi-modal ad-hoc network module is used to construct an ad-hoc network and pass through L GNN Optimize the topology; Multi-dimensional authentication and encryption module. The multi-dimensional authentication and encryption module includes a biometric collection module, an AI acceleration chip, and an encryption coprocessor. The biometric collection module includes a fingerprint sensor FPC1020, a microphone INMP441, a touch screen capacitive IC. The AI acceleration chip is selected as NVIDIA Jetson Nano, and the encryption coprocessor is selected as TPM, which is used to generate K dyn and encryption instruction C; Distributed collaboration module. The distributed collaboration module includes a distributed computing unit, a communication module, and an AI scheduling chip. The distributed computing unit selects STM32F4, and the AI scheduling chip selects a non-ArmCorte-M55, which is used to verify V ABFT and schedule P sched ; Intelligent prediction and control module. The intelligent prediction and control module includes a multi-sensor module, an edge computing unit, and a storage module. The multi-sensor module selects DHT22 and BH1750, and the edge computing unit selects RaspberryPi for running L FL and L GAN ; Feedback optimization module, which includes a feedback sensor, an AI optimization chip, and a display module. The AI optimization chip is selected as NVIDIA TX2, and the display module is selected as SSD1306, which is used to calculate Q(s, a) and output feedback.

8. The remote control system for smart home devices based on the Internet of Things according to claim 7, characterized in that: The multi-modal ad-hoc network module supports heterogeneous communication protocols, and the embedded microcontroller integrates a floating-point operation unit to accelerate L GNN computation.

9. The remote control system for smart home devices based on the Internet of Things according to claim 7, characterized in that: The biometric acquisition module of the multi-dimensional authentication and encryption module supports multi-dimensional data input, and the encryption coprocessor is preset with lattice parameter L to achieve quantum-secure encryption.

10. The remote control system for smart home devices based on the Internet of Things according to claim 7, characterized in that: The multi-sensor module of the intelligent prediction and control module includes temperature, humidity, and light sensors, and the edge computing unit supports parallel computing of federated learning and generative adversarial networks.

Images