Smart home control system based on Internet of Things

Abstract

The invention belongs to the technical field of smart home, and particularly relates to a smart home control system based on Internet of Things, which comprises a data acquisition and processing module, an intelligent decision module, a privacy optimization module and an execution control module. The data acquisition and processing module is used for acquiring multi-source data and encoding the multi-source data into a local intention vector; the intelligent decision-making module comprises a similarity calculation unit, a first decision-making unit, a second decision-making unit and a third decision-making unit and is used for selecting an optimal control strategy under the three conditions; and the privacy optimization module adopts homomorphic encryption and differential privacy technologies to protect local data, and realizes collaborative updating of the encrypted knowledge base through secure multi-party calculation. According to the invention, the problems of limited edge computing capability, cloud training privacy leakage and poor fixed rule adaptability in the prior art are solved, and efficient and personalized smart home control on the premise of protecting user privacy is realized.

Description

Technical Field

[0001] This invention belongs to the field of smart home technology, specifically relating to a smart home control system based on the Internet of Things. Background Technology

[0002] Smart home systems have become an important part of modern life. Smart home control systems primarily utilize various sensors, actuators, controllers, and communication technologies to monitor and manage the home environment, providing a more comfortable, convenient, and energy-efficient living experience. Smart home systems can connect various smart devices within the home to meet the family's needs for safety, convenience, comfort, and energy efficiency.

[0003] However, traditional smart home systems often rely on preset fixed rules or static thresholds to control and manage smart devices, while users' lifestyles are actually dynamic and changing, such as adjustments to work hours, seasonal differences in work and rest, and the addition of new family members.

[0004] To further optimize the fit of control strategies, existing technologies commonly incorporate machine learning models and algorithms. By learning and analyzing long-term user behavior data, the control system can autonomously predict user needs and adjust control logic accordingly. However, the training and inference process of the learning model consumes significant computing resources, making the choice of deployment location a key bottleneck restricting the technology's widespread adoption. From a technical implementation perspective, deploying the learning model on local edge computing nodes (such as smart home gateways and smart central control devices) enables local data processing and rapid response. However, the hardware configuration of ordinary edge computing nodes is limited, and their computing power is insufficient to meet the training requirements of complex learning models. Configuring high-performance processors and storage devices for edge computing nodes would lead to a significant increase in hardware costs, exceeding the affordability of ordinary households and severely limiting the technology's widespread application. To address the shortcomings of edge node deployment, the mainstream solution in current technology is to adopt a cloud-based training model. This involves local edge nodes uploading collected user behavior data to a cloud server, where a large-scale computing cluster in the cloud completes the training process of the learning model. The trained model parameters are then distributed back to the local edge nodes, which execute model inference and adjust control strategies. While this model effectively solves the problems of insufficient computing power and high costs associated with edge nodes, it introduces a serious risk of user privacy data leakage.

[0005] Specifically, user behavior data constitutes highly sensitive personal privacy information. During its transmission from edge nodes to the cloud, it may be leaked due to unauthorized eavesdropping or attacks on network transmission links. Furthermore, during storage and processing on cloud servers, user privacy data may be illegally leaked and misused due to security vulnerabilities in the cloud platform, improper access control, or third-party data access. These privacy and security risks not only seriously infringe upon users' legitimate rights and interests but also become a core technological obstacle restricting the development of smart home control systems towards higher levels of intelligence. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides an Internet of Things (IoT)-based smart home control system. The objective of this invention can be achieved through the following technical solution: An IoT-based smart home control system, characterized in that: the system includes: The data acquisition and processing module is used to collect environmental data, smart device status data, and user direct interaction command data, and encode the collected data into local intent vectors. The intelligent decision-making module includes a similarity calculation unit with communication connection, a first decision unit, a second decision unit, and a third decision unit; The similarity calculation unit is used to calculate the similarity between the local intent feature vector and each encrypted feature in the preset encrypted intent feature library under homomorphic encryption, and compare the maximum similarity in the calculation result with the set similarity threshold interval, which includes a first threshold interval, a second threshold interval and a third threshold interval. The first decision unit outputs the encryption control strategy corresponding to the feature vector most similar to the local intent feature vector as the current first control strategy when the maximum similarity belongs to the third threshold interval. The second decision unit is used to input the local intent vector into the pre-trained local learning model for prediction and reasoning when the maximum similarity falls within the first threshold range, and to generate and output the second control strategy. The third decision unit is used to obtain the encryption control strategy corresponding to the most similar feature vector of the local intent feature vector as the first candidate strategy when the maximum similarity is within the second threshold interval, and simultaneously input the local intent feature vector into the local learning model to perform prediction and reasoning to obtain the control strategy as the second candidate strategy. The first candidate strategy and the second candidate strategy are respectively used to predict comfort, compare the comfort prediction results, and output the better candidate control strategy as the third control strategy. The control execution module uses a first control strategy, a second control strategy, or a third control strategy to intelligently control home appliances.

[0007] Preferably, the cryptographic intent feature library stores cryptographic knowledge units from multiple anonymous families, each cryptographic knowledge unit including a triplet, the triplet including: cryptographic intent feature vector, cryptographic control policy vector, and cryptographic context label.

[0008] Preferably, the third decision-making unit performs comfort predictions on the first candidate strategy and the second candidate strategy respectively according to a set prediction effect scoring function, wherein the calculation formula of the prediction effect scoring function is: In the formula, Let p be the score function for predicting the effect of the control policy p, where p is either the first or second candidate policy. These are the weighting coefficients, and their sum is 1. For comfort prediction function, Let be the energy efficiency function. This is the user compliance function.

[0009] Preferably, the comfort prediction function is calculated based on temperature, humidity, and light deviation, and the calculation formula is as follows: ; In the formula, These are the temperature deviation function, humidity deviation function, and light deviation function, respectively. This is a Gaussian penalty term.

[0010] Preferably, the formula for calculating the user compliance function is: ; In the formula, For strategy similarity, K is the total number of historical samples used for similarity calculation, and j is the index of the current historical record. This is a function for calculating the similarity between a candidate policy p and the j-th historical control policy. Similarity between them Acceptance of historical strategies; For the user's explicit preference matching degree, R is the total number of pre-defined display rules, and r is the index of the display rule. Preferably, the indicator function also includes a privacy optimization module that is communicatively connected to the intelligent decision-making module. The privacy optimization module includes a local privacy unit, a differential privacy addition unit, and a collaborative update unit that are communicatively connected. The privacy optimization module includes a local privacy unit for communication connections, a differential privacy addition unit, and a collaborative update unit: The local privacy unit is used to perform homomorphic encryption on the local intent feature vector, the second control policy, and the third control policy. The differential privacy-adding unit is used to add random noise to the encrypted intent feature vector and then encode it to generate an encrypted data frame. The collaborative update unit is used to update the encrypted intent feature library using encrypted data frames through a clustering algorithm based on secure multi-party computation.

[0011] Preferably, the encrypted data frame has the following frame format: the frame header includes the protocol version, time window identifier, and anonymized device identifier hash; the frame body includes triple-protected intent feature ciphertext, control policy ciphertext, and context ciphertext; and the frame tail is appended with an integrity check code based on elliptic curve digital signature.

[0012] Preferably, the collaborative update unit constructs a secure computing network based on a star topology, with multiple home nodes as participants, uploading their respective encrypted data frames to a semi-trusted aggregation server. In the encrypted state, a clustering algorithm based on secure multi-party computation is used to cluster multiple encrypted intent feature vectors, filter them, and then use the filtered data to update the network.

[0013] Preferably, the collaborative update unit further includes using an incremental strategy to implement the update process.

[0014] The beneficial effects of this invention are as follows: This invention effectively addresses the core contradictions of traditional smart home systems through a collaborative decision-making mechanism involving a first, second, and third decision-making unit. The first decision-making unit achieves millisecond-level matching based on a local encrypted feature library, avoiding real-time computation of complex models and mitigating the challenge of insufficient computing power on edge devices. The second decision-making unit handles novel scenarios using a lightweight deep learning model, overcoming the limitation of fixed-rule systems in adapting to dynamic user preferences. The third decision-making unit evaluates the execution effects of two control strategies using a prediction performance scoring function when the similarity falls within a second threshold range, selecting the optimal strategy from the two candidate strategies. This improves user comfort while enabling the system to respond quickly in common scenarios and adaptively learn in new ones, fundamentally eliminating reliance on preset rules or heavy cloud-based models, thus improving system response speed. Encryption and noise reduction ensure privacy protection and enhance the overall quality of home system control. Attached Figure Description

[0015] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0016] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a schematic diagram of the privacy optimization module of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0018] Please see Figures 1-2 This embodiment provides an Internet of Things (IoT) based smart home control system, the system including: The data acquisition and processing module is used to collect environmental data, smart device status data, and user direct interaction command data, and encode the collected data into local intent vectors. The data acquisition and processing module includes a data acquisition unit, a data preprocessing unit, and a data encoding unit. The data acquisition unit is used to collect environmental data, smart device status data, and user direct interaction command data based on various IoT sensors, device interfaces, and interactive terminals. The data preprocessing unit is used to preprocess the data collected by the data acquisition unit; The data encoding unit is used to extract and fuse classification features from the preprocessed data, and to perform temporal context modeling to generate local intent vectors. The specific functional implementation process of each unit includes: The data acquisition unit, based on various IoT sensors, device interfaces, and interactive terminals, achieves comprehensive and multimodal perception of the physical environment, device status, and user intent. Specific data acquisition methods include: Environmental data: Environmental data is used to assist the control system in perceiving environmental data and adjusting control strategies accordingly. This data includes: acquiring ambient temperature and humidity through digital temperature and humidity sensors deployed in key areas such as living rooms, bedrooms, and kitchens, typically communicating with the gateway via low-power wireless protocols (such as Zigbee, Z-Wave, or Bluetooth Mesh); collecting ambient light intensity through integrated smart lights, smart panels, or standalone light sensors to determine day / night conditions and direct sunlight, automatically controlling curtains and lights; acquiring PM2.5 / PM10 concentrations, total volatile organic compound concentrations, and carbon dioxide concentrations through air quality monitors installed indoors (such as living rooms and studies), with built-in gas and particulate matter sensors, reporting data via Wi-Fi or wirelessly; and acquiring ambient noise decibel values ​​and specific sound events (such as glass breaking sounds, smoke alarm beeping sounds, etc.) through sound sensors or microphone arrays. Furthermore, local edge computing nodes perform real-time edge AI analysis on the audio stream, extracting only characteristic events or decibel values ​​without uploading the original audio to protect privacy. Smart device status data reflects the working status of each execution unit within the home control system, facilitating coordinated control and status feedback. This status data includes: A) Switch and energy consumption status: For smart sockets or switches, the status and power information of connected devices are directly read. For large appliances such as smart air conditioners and smart washing machines, detailed status and energy consumption data are directly obtained via Wi-Fi, Bluetooth, or dedicated protocols (such as Matter); B) Device operating parameters: Status queries and subscriptions are performed through the open API interfaces or standard object models of each smart device. This periodically or through event monitoring, data such as air conditioner set temperature and operating mode, light brightness and color temperature, curtain opening / closing percentage, and TV content source and volume are obtained; C) Security device status: Security sensors are typically networked using low-power, high-reliability protocols (such as Zigbee and NB-IoT). Status changes are directly obtained through this network. Device status includes door and window magnetic sensor status (open / closed), smart door lock status (locked / unlocked, unlocking method), and camera working status (online / offline). User-direct interaction command data includes voice commands, touch and interface interactions, as well as preset scene and automated trigger data. Voice commands are collected through the voice input function of smart speakers, smart panels with microphones, or mobile apps. To protect privacy, a combination of local wake word detection and cloud semantic understanding or fully local edge AI voice recognition is used for collection. User GUI operations are recorded as structured control command logs through interactive terminals such as mobile applications, web interfaces, and smart touch screens. This includes button clicks, slider adjustments, and scene mode selections on the app; touch and interface interaction information from physical buttons or touch screen operations on smart wall panels; and user behaviors that are manually triggered by the user through the app, voice, or panel, as well as the execution logs of preset time / conditional automation rules. All of these are considered as indirect user preference commands.

[0019] The data preprocessing unit is used to preprocess the data collected by the data acquisition unit. Preprocessing includes outlier filtering and missing value handling, dimensional normalization, and timestamp alignment and buffer construction. Specifically, rule-based filters are applied to the sensor data stream, replacing missing values ​​caused by brief communication interruptions with the previous valid value or moving average, and filling them using linear interpolation or state preservation methods. Preferably, Hampel filters can be used for anomaly detection. Minimum-maximum normalization or Z-score standardization is used for dimensional normalization. The data preprocessing unit maintains a sliding time window buffer (e.g., the most recent 5 minutes), aligning all collected asynchronous data to a uniform sampling time point (e.g., once per second) according to their timestamps. For non-real-time changing state data (e.g., device on / off status), their values ​​remain unchanged before the state changes. After preprocessing, an arbitrary decision time t is obtained, and the system forms a multimodal dataset X(t). X(t) is a structured data object or table containing the normalized values ​​of all environmental, device, and interaction data at time t (or within a very short time window).

[0020] The data encoding unit is used to extract and fuse classification features from the preprocessed data, and to perform temporal context modeling to generate local intent vectors, including the following steps: S1: For different parts of the multimodal dataset, classify them according to environmental features, device status features, and user interaction features, and extract features for each, including: Environmental feature extraction: Aggregate sensor readings from different rooms (such as average temperature and maximum CO2 concentration) to generate a spatial feature vector characterizing the overall environmental conditions of the house. Calculate the slope, variance, and mean of temperature or light intensity indicators over a past time window to construct a trend vector to determine the changing trend of environmental data. For the extraction of device state features, the state of associated devices is encoded into a state vector, such as a lighting system and a temperature control system. Then, knowledge graphs or rules are used to extract features between devices. For user interaction features, the most recent voice commands and APP operations are converted into bag-of-words model vectors or pre-trained language model micro-embedding vectors, and the device types with high frequency of interaction in a short period of time are statistically analyzed based on the interaction pattern to form an interaction focus vector. S2: Fuse and concatenate the multi-source features, concatenating all the extracted feature vectors into a high-dimensional joint feature vector; S3: Since the joint feature vector may have high dimensionality or redundancy, a lightweight autoencoder or principal component analysis is used for dimensionality reduction. Preferably, a multilayer perceptron is used as the feature fusion encoder to achieve feature fusion and dimensionality reduction. The calculation formula includes: First fully connected layer: ; Second hidden layer: ; Output layer: ; In the formula, and are the output vectors of the first hidden layer and the second hidden layer, respectively. The latent vector of the output layer is a dense, low-dimensional vector. Through multilayer perceptron processing, the joint feature vector undergoes dimensionality reduction compression and multilayer fusion, preserving key information from the input features. It is a condensed encoding of the joint feature vector. For joint feature vectors, , , These are the weight matrices for the first fully connected layer, the second hidden layer, and the third output layer, respectively. , , These are the bias vectors for the first fully connected layer, the second hidden layer, and the third output layer, respectively. The linear unit activation function is modified by setting negative values ​​to 0 and retaining positive values, introducing a non-linear transformation to enable the model to fit more complex feature relationships.

[0021] S4: Perform context modeling, fusing the current latent space representation with the short-term historical context, and then using the latent vector at time t. Along with the latent vectors from the past few time points (such as the past 4 decision points), they are input into a gated recurrent unit or a causal convolutional network, and the output is a hidden state vector.

[0022] Preferably, when using GRU to compute the hidden state vector, the compressed features of the current time step are combined with the historical memory of the previous time step to generate a new hidden state containing current information and recent historical dynamic patterns, thereby capturing the temporal dependencies in the sequence data (such as decision points and time steps). In the single-step iterative computation of GRU, the formula for calculating the hidden vector at time t is: ; In the formula, Let be the hidden vector at time t. This is the hidden vector at time t-1, carrying historical information from the beginning of the sequence to time t-1. The GRU computation function intelligently controls the retention and updating of information through the "Update Gate" and "Reset Gate" mechanisms, solving the gradient vanishing problem of traditional recurrent neural networks (RNNs). It is better at capturing time dependencies. By modeling the temporal context through GRU, the model can not only utilize the features of the current moment, but also capture the dynamic evolution patterns in the time dimension (such as the temporal dependencies of user behavior and the continuous changes in system state), thereby improving the accuracy of downstream tasks (such as decision-making, prediction, and anomaly detection). S5: The current GRU hidden state and latent vector are linearly projected and normalized using the LayerNorm function to generate the final fixed-dimensional user intent vector.

[0023] Through the collaborative efforts of multiple computing units within the data acquisition and processing module, the complex raw data from the home IoT is intelligently translated into a machine-understandable, representational intent code that can be used for efficient decision-making. This provides precise input for subsequent intelligent decision-making modules, and each unit of the module is designed to be lightweight to ensure real-time operation on resource-constrained edge devices.

[0024] The intelligent decision-making module includes a similarity calculation unit for communication connections, a first decision-making unit, a second decision-making unit, and a third decision-making unit. The similarity calculation unit is used to perform similarity matching between the local intent feature vector and the encrypted features in the preset encrypted intent feature library, and compare the maximum similarity in the calculation result with the set similarity threshold interval, which includes a first threshold interval, a second threshold interval and a third threshold interval; The first decision unit outputs the encryption control strategy corresponding to the feature vector most similar to the local intent feature vector as the current first control strategy when the maximum similarity belongs to the third threshold interval. The second decision unit is used to input the local intent vector into the pre-trained local learning model for prediction and reasoning when the maximum similarity falls within the first threshold range, and to generate and output the second control strategy. The third decision unit is used to, when the maximum similarity falls within the second threshold range, obtain the encryption control policy corresponding to the most similar feature vector of the local intent feature vector as the first candidate policy, and simultaneously input the local intent feature vector into the local learning model for prediction and inference to obtain the control policy as the second candidate policy. Comfort prediction is then performed on the first and second candidate policies respectively, the comfort prediction results are compared, and the superior candidate control policy is output as the third control policy. The computation and execution of each unit includes the following processes: The similarity calculation unit compares the maximum similarity in the calculation results with a set maximum similarity threshold. The first decision-making unit adopts a privacy-preserving retrieval architecture based on homomorphic encryption. The encrypted intent feature library is stored in encrypted form in a protected memory area of ​​the edge device. A hierarchical index structure is used to optimize retrieval efficiency. The encrypted intent feature library adopts a hierarchical organization method based on scene clustering. The top layer is the scene category index, and each category contains multiple encrypted features of specific scenes. The encrypted intent feature library contains multiple encrypted knowledge units. The encrypted intent feature library stores encrypted knowledge units from multiple anonymous homes. Each encrypted knowledge unit includes a triplet. Each triplet contains an intent feature vector processed by homomorphic encryption and differential privacy, a corresponding encrypted control policy vector, and an encrypted context label describing the context in which the encrypted control policy is generated. The encrypted intent feature vector is a 128-dimensional feature vector encrypted by the CKKS homomorphic encryption scheme, which represents the abstract semantics of a specific home scenario. Encryption control strategy vector: The encoded representation of device control commands using Paillier partial homomorphic encryption; Encrypted context tags: Encrypted tags that assist in matching, including timestamps, season identifiers, and device availability status.

[0025] Upon receiving a plaintext geographic intent vector, the similarity calculation unit performs the following process: A1: Homomorphically encrypt the plaintext ground intent vector using the same public key PK as the encrypted intent feature library to obtain the encrypted vector. The calculation formula is: ; A2: To protect privacy, similarity matching is needed in the ciphertext space to calculate the similarity between the encrypted vector and each encrypted feature in the encrypted intent feature library. Since directly calculating traditional similarity (such as Euclidean distance and cosine similarity) of ciphertext is extremely complex, in practice, dedicated algorithms that support homomorphic comparison or approximate schemes such as comparative hash mapping are used. The similarity calculation formula under homomorphic encryption is as follows: For each encrypted historical feature in the feature library, its ciphertext similarity with the current encrypted feature is calculated, resulting in a set of similarity values ​​corresponding to all historical features. ; In the formula, Let i be the encrypted historical intent feature vector in the encrypted intent feature library. This is a similarity calculation function under homomorphic encryption, representing an algorithm that can directly calculate similarity in ciphertext (such as approximate algorithms that support homomorphic comparison, comparative hash mapping, etc.), and outputs the similarity value of two ciphertext features; Iterate through the calculated similarity values ​​and find the maximum similarity value. Compare the maximum similarity value with a set similarity threshold interval, which includes the first threshold interval. Second threshold interval and the third threshold interval The result is determined by whether the maximum similarity falls within the threshold range. The first decision unit obtains the output of the similarity calculation unit. When the maximum similarity falls into the third threshold range, it determines that the current situation is highly similar to the historical records of the historical records index and directly outputs the encryption control strategy corresponding to the feature vector i that is most similar to the local intent feature vector as the current first control strategy. Homomorphic encryption ensures that feature data remains in ciphertext during computation, preventing the leakage of original information; meanwhile, memory-based retrieval enables low-latency strategy matching, making it suitable for scenarios requiring real-time decision-making.

[0026] For the maximum similarity falling within the first threshold range At the same time, the second decision unit is used to synchronously input the local intent vector into the pre-trained local learning model for model prediction and inference, generating a personalized policy. The personalized policy is output as the second control policy. The local learning model here can be pre-trained using a 2-3 layer MLP (Multilayer Perceptron), MobileNet, ShuffleNet, or a small random forest or gradient boosting tree (GBDT). The training process includes: On the edge server side or the cloud side, historical data (or initial dataset) can be used to train model parameters in the cloud. Historical data can be anonymized aggregated data uploaded by edge devices or initial data generated by simulating scenarios to ensure data privacy. Supervised learning minimizes the error between predicted control commands and the true optimal commands, allowing the model to learn the mapping relationship between local intent vectors and control policies. The model parameters of the learning model are updated periodically through federated learning or by adding noise to the parameter updates using differential privacy algorithms. For example, with federated learning, the edge device calculates the parameter update gradient locally, uploads it to the cloud through encrypted aggregation (such as secure multi-party computation), and then distributes the new parameters after aggregation in the cloud. The original data is not leaked throughout the process, allowing the model to adapt to the dynamic changes in the edge scenario (such as the access of new devices and changes in user behavior habits), while avoiding the resource bottleneck of direct training on edge devices. To address the issue of insufficient computing power at the edge, during the model design phase, the number of network layers is reduced (e.g., MLP uses only 2 layers, CNN uses only 3 layers), the number of neurons per layer is reduced (e.g., ≤256 neurons per layer), and lightweight operators are used instead of traditional operators. Separable convolutions can be used instead of standard convolutions, and ReLU can be used instead of complex activation functions.

[0027] The third decision unit obtains the encryption control strategy corresponding to the feature vector most similar to the local intent feature vector as the first candidate strategy. Simultaneously, it inputs the local intent feature vector into the local learning model for prediction and inference, resulting in a control strategy as the second candidate strategy. It then performs comfort prediction on both the first and second candidate strategies, compares the comfort prediction results, and outputs the superior candidate control strategy as the third control strategy. Specifically, the third decision unit invokes the decision-making mechanisms of the first and second decision units to calculate the first and second candidate strategies, respectively. It then uses a preset prediction effect scoring function to perform comfort prediction on both the first and second candidate strategies. The calculation formula for the prediction effect scoring function is as follows: ; In the formula, Let p be the score function for predicting the effect of the control policy p, where p is either the first or second candidate policy. These are the weighting coefficients, and their sum is 1. This is a comfort prediction function, calculated based on the expected deviations from temperature, humidity, and light intensity. Given an energy efficiency function, predict the energy cost of implementing this control strategy. The user compliance function is based on the acceptance rate of historical similar strategies. The comfort prediction function, energy efficiency function, and user compliance function are all normalized values. The formula for calculating the comfort prediction function is: ; In the formula, These are the temperature deviation function, humidity deviation function, and light deviation function, respectively. The Gaussian penalty term is applied as an additional penalty when the predicted value deviates significantly from the ideal value. The formula for calculating the Gaussian penalty term is: ; In the formula, To penalize the weight value, For the standard deviation tolerances of various temperatures, humidity levels, and light intensity, The predicted value of k, Let k be the ideal value for the customer, where k is one of temperature, humidity, and light intensity. The calculation formula is: ; In the formula, These are temperature predictions based on thermodynamic models. The dynamic ideal temperature value that changes over time. The maximum acceptable deviation range is set at 5°C. and Calculation model reference The calculation is performed using the calculation model, which uses predicted values, ideal values, and acceptable deviation ranges. The formula for calculating the energy efficiency function is: ; In the formula, , These represent the actual total energy consumption and the predicted total energy consumption of the home system executing candidate strategy p, respectively. User compliance function The calculation formula is: ; In the formula, The similarity metric is denoted by K, ranging from 0 to 1. A higher value indicates a higher similarity. It quantifies the degree of fit between the current control policy and policies used in similar historical scenarios to assess policy compliance with past user behavior patterns. K represents the total number of historical data samples used for similarity calculation, and j is the index of the current historical data. This is a function for calculating the similarity between a candidate strategy p and the j-th historical record. Similarity between (historical control strategies) This is the acceptance tag for the historical strategy. It takes a value of 1 when the user accepts the strategy without intervention, and a value of 0 when the user actively withdraws the strategy. The user explicit preference matching degree is used to quantify the degree of fit between the current control strategy p and the user's preset explicit rules. It ranges from 0 to 1, with a higher value indicating a higher matching degree. Here, R is the total number of preset display rules, and r is the index of the display rule. This is an indicator function that determines whether strategy p violates the r-th rule. If it violates the rule, it returns 1; otherwise, it returns 0. By establishing a user compliance function calculation formula that combines user display preferences with policy display, the system transforms user-preset qualitative rules into quantitative matching degrees with values ​​ranging from 0 to 1. Through simple counting and normalization calculations, it can assess the degree to which control policies fit the user's explicit rules in real time and seamlessly integrate into a multi-objective optimization framework. From a business perspective, this ensures that machine decisions are anchored to the user's core needs and can pinpoint the reasons for violations, thereby improving decision transparency and user trust. At the same time, its extremely low computing power consumption is fully adapted to the real-time and privacy requirements of edge devices, demonstrating a closed-loop support for rule constraints → quantitative evaluation → compliant decision-making.

[0028] Compare the prediction performance scores of the first and second candidate control strategies, and select the one with the higher score as the output of the third control strategy; The third decision-making unit, when the maximum similarity falls within the second threshold range, uses a comfort prediction function and a quantitative mathematical model to comprehensively evaluate the expected improvement effect of a given control strategy P on the user's environmental comfort. The calculated comprehensive score intuitively reflects the degree to which strategy p adjusts the environment to the user's ideal state. The higher the value, the better the strategy is in meeting the user's personalized comfort needs. This determines whether to use the first or second candidate strategy, further optimizes the control process, and selects a control strategy that improves user comfort.

[0029] The control execution module uses a first control strategy, a second control strategy, or a third control strategy to intelligently control home appliances.

[0030] The privacy optimization module communicates and connects with the intelligent decision-making module. It is activated immediately after the control policy is generated and performs multi-layered protection on data involving user privacy, ensuring that sensitive information is always encrypted or obfuscated when participating in global knowledge sharing. The privacy optimization module includes a local privacy unit for communication connections, a differential privacy addition unit, and a collaborative update unit: The local privacy unit is used to perform homomorphic encryption on the local intent feature vector, the second control policy, and the third control policy. The differential privacy-adding unit is used to add random noise to the encrypted intent feature vector and then encode it to generate an encrypted data frame. The collaborative update unit is used to update the encrypted intent feature library using encrypted data frames through a clustering algorithm based on secure multi-party computation. The execution of each unit includes the following processes: The local privacy unit is used to perform homomorphic encryption processing on the local intent feature vector, the second control policy, and the third control policy using a homomorphic encryption algorithm when the control command is triggered. The CKKS fully homomorphic encryption scheme is used to encrypt the three data, allowing arithmetic operations to be performed directly in the ciphertext state without decryption. The encryption process uses a preset public key, which is periodically rotated and updated by the key management service. Specifically, each numerical component of the feature vector and the control policy vector is encrypted separately to generate the corresponding ciphertext component sequence. Since CKKS supports floating-point encryption, it can maintain the precision characteristics of the original data, making it easy for the teaching model to recognize. The differential privacy addition unit is used to add random noise that meets the differential privacy requirements to the encrypted intent feature vector, and then encode it to generate an encrypted data frame. Based on the ciphertext, differential privacy protection that satisfies strict mathematical definitions is further added to the encrypted intent features, according to a preset privacy budget parameter (usually...). The algorithm generates random noise vectors of appropriate strength from the Laplace distribution. For sensitive dimensions with high identifiability in the feature vector (such as dimensions that may reflect specific lifestyle habits), relatively strong noise is injected; for dimensions with strong generality, weak noise is injected. This differentiated processing protects privacy while preserving the statistical utility of the data to the greatest extent. By adding noise in ciphertext form to the encrypted feature vector through homomorphic addition, the original user data has been protected by both encryption and noise addition. Even if a single data packet is intercepted, the attacker will not be able to deduce any valid user information. The processed components are encapsulated into a standardized data frame structure. The frame header contains the protocol version, time window identifier, and anonymized device identifier hash; the frame body contains triple-protected intent feature ciphertext, control policy ciphertext, and context ciphertext; and the frame tail is appended with an integrity check code based on elliptic curve digital signature to prevent tampering during transmission. To optimize transmission efficiency, ciphertext packing technology is used to compress multiple numerical components into a single ciphertext polynomial, significantly reducing data volume. A typical encrypted data frame size is controlled within 5KB, ensuring efficient transmission even in IoT bandwidth-constrained environments. The collaborative update unit is used to update the encrypted intent feature library using encrypted data frames. The update process includes the following steps: By constructing a secure computing network based on a star topology, multiple home nodes participate by uploading their respective encrypted data frames to a semi-trusted aggregation server. In the encrypted state, a clustering algorithm based on secure multi-party computation is used to perform cluster analysis on the intent feature vectors after multiple rounds of encryption, and the most representative encrypted knowledge units in each cluster are selected to update the global encrypted feature library.

[0031] In encrypted form, the server can only see the encrypted intermediate results and cannot know the specific details of any family. The clustering algorithm employs an improved k-means variant adapted to homomorphic encryption environments. During algorithm initialization, the server randomly selects several points in the ciphertext space as initial cluster centers. These points are themselves encrypted, and the server is unaware of their actual values. Furthermore, during iteration, the distance to each cluster center needs to be calculated for each encrypted data point. Distance calculation is achieved through homomorphic operations: the server leverages the computability of encrypted vectors to calculate the square of the Euclidean distance without decryption, and the result remains encrypted. To determine the nearest cluster center, a secure multi-party comparison protocol is employed. This protocol allows multiple participants to collaboratively compare the size of two encrypted values ​​without revealing the values ​​themselves. Through multiple rounds of interaction, each data point is securely assigned to the nearest cluster, and the assignment results are secretly shared among the participants, ensuring that no single party is aware of the complete assignment. The selected representative knowledge units, after being standardized in format, are added to the global encryption feature library. The update process adopts an incremental strategy: new units first enter the "observation area" to verify their actual effectiveness over the following 1-2 weeks; after successful verification, they are then formally incorporated into the intent encryption feature library. After the update is complete, the new intent encryption signature library is distributed to each edge node in a differential privacy-preserving manner. After verifying the integrity and freshness of the library locally, the nodes replace the old version of the library files. The entire update cycle is typically once a month, executed automatically during system downtime.

[0032] The privacy optimization module constructs a complete technical closed loop of encrypted matching, secure learning, and privacy-protected updates through the collaborative work of multiple units. At the data generation end, the local privacy unit and the differential privacy addition unit use homomorphic encryption and differential privacy technologies to provide dual protection for sensitive information, ensuring that the original user data never leaves the local machine. In the knowledge aggregation stage, the collaborative update unit realizes clustering analysis and knowledge extraction in encrypted state through a secure multi-party computation protocol, enabling the global knowledge base to continuously evolve without touching plaintext data. This deeply integrates federated learning and edge computing, realizing privacy-secure distributed collaborative intelligence and providing a new technical paradigm for the smart home field.

[0033] The control system effectively addresses the core contradictions of traditional smart home systems through a collaborative decision-making mechanism involving a first, second, and third decision-making unit. The first decision-making unit achieves millisecond-level matching based on a local encrypted feature library, avoiding real-time computation of complex models and mitigating the challenge of insufficient computing power on edge devices. The second decision-making unit handles novel scenarios using a lightweight deep learning model, overcoming the limitation of fixed-rule systems in adapting to dynamic user preferences. Furthermore, the third decision-making unit evaluates the execution effects of two control strategies using a prediction performance scoring function when the similarity falls within a second threshold range, selecting the optimal candidate strategy. This improves user comfort while enabling the system to respond quickly in common scenarios and adaptively learn in new ones, fundamentally eliminating reliance on preset rules or heavy cloud-based models.

[0034] In terms of technical performance, the system balances response speed, personalization accuracy, and resource consumption; in terms of privacy and security, encryption and noise reduction achieve usability without visibility of user behavior data, eliminating the risk of data leakage in cloud training mode; low-cost edge deployment significantly lowers the barrier to system adoption, while continuously evolving collective intelligence ensures long-term improvement in user experience.

[0035] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An Internet of Things based smart home control system characterized by: The system includes: The data acquisition and processing module is used to collect environmental data, smart device status data, and user direct interaction command data, and encode the collected data into local intent vectors. The intelligent decision-making module includes a similarity calculation unit with communication connection, a first decision unit, a second decision unit, and a third decision unit; The similarity calculation unit is used to calculate the similarity between the local intent feature vector and each encrypted feature in the preset encrypted intent feature library under homomorphic encryption, and compare the maximum similarity in the calculation result with the set similarity threshold interval, which includes a first threshold interval, a second threshold interval and a third threshold interval. The first decision unit outputs the encryption control strategy corresponding to the feature vector most similar to the local intent feature vector as the current first control strategy when the maximum similarity belongs to the third threshold interval. The second decision unit is used to input the local intent vector into the pre-trained local learning model for prediction and inference when the maximum similarity falls within the first threshold range, and to generate and output the second control strategy. The third decision unit is used to obtain the encryption control strategy corresponding to the most similar feature vector of the local intent feature vector as the first candidate strategy when the maximum similarity is within the second threshold interval, and simultaneously input the local intent feature vector into the local learning model to perform prediction and reasoning to obtain the control strategy as the second candidate strategy. The first candidate strategy and the second candidate strategy are respectively used to predict comfort, compare the comfort prediction results, and output the better candidate control strategy as the third control strategy. The control execution module uses a first control strategy, a second control strategy, or a third control strategy to intelligently control home appliances.

2. The smart home control system based on the Internet of Things according to claim 1, characterized in that: The encrypted intent feature library stores encrypted knowledge units from multiple anonymous families. Each encrypted knowledge unit includes a triple, which includes: an encrypted intent feature vector, an encrypted control policy vector, and an encrypted context label.

3. The smart home control system based on the Internet of Things according to claim 2, characterized in that: The third decision-making unit performs comfort predictions on the first and second candidate strategies respectively based on a set prediction effect scoring function. The calculation formula for the prediction effect scoring function is as follows: In the formula, Let p be the score function for predicting the effect of the control policy p, where p is either the first or second candidate policy. These are the weighting coefficients, and their sum is 1. For comfort prediction function, Let be the energy efficiency function. This is the user compliance function.

4. A smart home control system based on the Internet of Things according to claim 3, characterized in that: The comfort prediction function is calculated based on temperature, humidity, and light deviation, and the calculation formula is as follows: ; In the formula, These are the temperature deviation function, humidity deviation function, and light deviation function, respectively. This is a Gaussian penalty term.

5. A smart home control system based on the Internet of Things according to claim 3, characterized in that: The formula for calculating the user compliance function is as follows: ； In the formula, For strategy similarity, K is the total number of historical samples used for similarity calculation, and j is the index of the current historical record. This is a function for calculating the similarity between a candidate policy p and the j-th historical control policy. Similarity between them Acceptance of historical strategies; For the user's explicit preference matching degree, R is the total number of pre-defined display rules, and r is the index of the display rule. This is an indicator function.

6. A smart home control system based on the Internet of Things according to claim 1, characterized in that: It also includes a privacy optimization module that is communicatively connected to the intelligent decision-making module. The privacy optimization module includes a local privacy unit, a differential privacy addition unit, and a collaborative update unit that are communicatively connected. The local privacy unit is used to perform homomorphic encryption on the local intent feature vector, the second control policy, and the third control policy. The differential privacy-adding unit is used to add random noise to the encrypted intent feature vector and then encode it to generate an encrypted data frame. The collaborative update unit is used to update the encrypted intent feature library using encrypted data frames through a clustering algorithm based on secure multi-party computation.

7. A smart home control system based on the Internet of Things according to claim 6, characterized in that: The encrypted data frame has the following frame format: the frame header contains the protocol version, time window identifier, and anonymized device identifier hash; the frame body contains triple-protected intent feature ciphertext, control policy ciphertext, and context ciphertext; and the frame tail is appended with an integrity check code based on elliptic curve digital signature.

8. A smart home control system based on the Internet of Things according to claim 6, characterized in that: The collaborative update unit constructs a secure computing network based on a star topology, with multiple home nodes as participants. Each node uploads its encrypted data frame to a semi-trusted aggregation server. In the encrypted state, a clustering algorithm based on secure multi-party computation is used to cluster multiple encrypted intent feature vectors, filter them, and then use the filtered data to update the network.

9. A smart home control system based on the Internet of Things according to claim 8, characterized in that: The collaborative update unit also includes using an incremental strategy to implement the update process.

Images