Remote internet-of-things control system and control method for household equipment

Abstract

The invention provides a remote internet-of-things control system and control method for home equipment, and the method comprises the steps: obtaining the operation data, environment data and user operation data of each piece of home equipment in a home system through a data obtaining module, and carrying out the edge calculation of the data; and online learning is carried out according to the processed data by adopting a deep reinforcement learning method through an instruction generation module, and an equipment control instruction set is constructed. The instruction verification module verifies the feasibility of each instruction in the equipment control instruction set through the preset virtual twin model based on the historical fault data and the user preference data corresponding to the home system, determines the initial control instruction set, and avoids the problem of equipment misoperation. The network state is monitored in real time through the network self-adaptive transmission module, a transmission path is dynamically adjusted according to factors such as network delay and packet loss rate, and it is ensured that a control instruction can be timely and accurately transmitted to home equipment. The instruction transmission delay or loss caused by network problems is avoided, and the user experience is improved.

Description

Technical Field

[0001] The present invention relates to the field of smart home technology, and in particular to a remote IoT control system and a control method for home appliances. Background Art

[0002] With the rapid development of the Internet of Things (IoT) technology, remote IoT control systems for home appliances are playing a vital role in improving the convenience and intelligence of life. These systems allow users to monitor and control home appliances in real time via remote devices such as mobile terminals, enabling intelligent management of lighting, air conditioners, and appliances, creating a convenient and comfortable home environment. However, existing remote IoT control systems for home appliances suffer from significant technical deficiencies in practical applications. For one thing, these systems lack effective verification mechanisms when receiving control commands from users or the cloud, directly transmitting them to home appliances. This can lead to unverified commands causing malfunctions due to data errors, formatting errors, or malicious tampering. For example, an erroneous temperature control command could cause an air conditioner to operate at high load for extended periods, shortening the lifespan of the appliance and potentially posing a safety hazard. Maliciously tampered power control commands could even cause electrical appliances to short-circuit and catch fire, seriously threatening the safety of users and property.

[0003] Furthermore, the transmission of control commands is highly dependent on the network environment, and network fluctuations such as signal attenuation, latency, and packet loss are common, severely impacting the real-time nature of command transmission. For example, if a user remotely sends an emergency command to close a gas valve, network latency prevents the command from being delivered in time, potentially missing the optimal time to act and potentially causing a gas leak or other safety incident. Furthermore, when adjusting smart lighting, network fluctuations can cause delayed updates to lighting status, leading to a disconnect between user actions and device responses, significantly reducing the user experience. Summary of the Invention

[0004] The invention provides a remote IoT control system and control method for home appliances, which solves the technical problems that the existing remote IoT control systems for home appliances directly issue unverified control instructions, which may cause the equipment to operate incorrectly, and the control instructions are easily affected by network fluctuations, affecting real-time performance and resulting in a poor user experience.

[0005] The present invention provides a remote IoT control system for household appliances, comprising: a data acquisition module, an instruction generation module, an instruction verification module, and a network adaptive transmission module;

[0006] The data acquisition module is used to obtain the operating data, environmental data and user operation data of each home appliance in the home system and perform edge computing on the data to generate target operating data, target environmental data and target user operation data corresponding to the home appliance;

[0007] The instruction generation module is used to use a deep reinforcement learning method to perform online learning based on all the target operation data, all the target environment data, and all the target user operation data to construct a device control instruction set;

[0008] The instruction verification module is used to verify the feasibility of each instruction in the device control instruction set based on the historical fault data and user preference data corresponding to the home system through a preset virtual twin model, and determine the initial control instruction set;

[0009] The network adaptive transmission module is used to dynamically select the transmission path corresponding to the initial control instruction set according to network delay and packet loss rate, and generate target control instructions.

[0010] Optionally, the data acquisition module includes:

[0011] A data collection module is used to collect operating data, environmental data and user operation data of each home appliance in the home system, and generate multiple operating data, multiple environmental data and multiple user operation data;

[0012] an association data generating module, configured to determine association data between the household devices according to the device type, device location, and device function corresponding to each of the operation data;

[0013] a data correction module, configured to use a machine learning method to perform anomaly detection and data correction on the operating data, the environmental data, and the user operation data, respectively, to generate initial operating data, initial environmental data, and initial user operation data corresponding to the home appliance;

[0014] A feature extraction module is used to extract time series features of the initial operation data, the initial environment data and the initial user operation data using an LSTM network based on the associated data to generate target operation data, target environment data and target user operation data corresponding to the home appliance.

[0015] Optionally, the data correction module performs the following steps:

[0016] Performing timestamp alignment and data normalization on the operation data, the environment data, and the user operation data respectively to generate normalized data corresponding to the home appliance;

[0017] Performing data anomaly detection on the normalized data using a preset time series anomaly detection model to generate an abnormal location mask for operation and environment data and an abnormal location mask for user operation;

[0018] Correcting the normalized data corresponding to the short-term anomaly mask in the abnormal position mask of the operation and environmental data using a sliding window difference method to generate initial corrected data;

[0019] Using a generative adversarial network to correct the initial correction data corresponding to the long-term anomaly mask in the anomaly position mask of the operation and environment data to generate intermediate correction data;

[0020] Using a preset user behavior model to correct the intermediate correction data corresponding to the user operation abnormal position mask to generate target correction data;

[0021] The target correction data is denormalized to generate initial operation data, initial environment data and initial user operation data corresponding to the home appliance.

[0022] Optionally, the execution steps of the feature extraction module include:

[0023] splicing the initial operation data, the initial environment data, the initial user operation data, and the associated data according to feature dimensions to obtain fused data;

[0024] Using a sliding window technique to divide the fused data into windows according to preset division data, generating a plurality of window data blocks;

[0025] Input the window data modules into the preset LSTM network for feature extraction to generate feature vectors;

[0026] All the feature vectors are linearly projected to generate target operation data, target environment data and target user operation data corresponding to the home appliance.

[0027] Optionally, the instruction generation module includes:

[0028] An environment modeling module, configured to construct an environment model corresponding to the home system using all of the target operation data, all of the target environment data, and all of the target user operation data, and to generate a plurality of local environment models and a global environment model;

[0029] A user preference modeling module is used to learn the user operation mode and user preference corresponding to the target user operation data through a deep learning model to obtain a user preference model;

[0030] An environmental prediction module, configured to perform environmental prediction on the target environmental data using a time series analysis method to generate environmental prediction data;

[0031] A deep reinforcement learning decision module is used to determine the device control instructions corresponding to each of the home appliances based on the local environment model, the global environment model, the user preference model and the environmental prediction data, and generate a device control instruction set using a deep reinforcement learning algorithm.

[0032] Optionally, the execution steps of the deep reinforcement learning decision module are:

[0033] Performing data splicing on the local environment state data corresponding to the local environment model, the global environment state data corresponding to the global environment model, and the environment prediction data to generate a comprehensive state vector;

[0034] Calculating the attention weights of each part of the comprehensive state vector using an attention network to generate attention weight data;

[0035] Performing weighted summation on the integrated state vector using the attention weight data to generate a fusion state vector;

[0036] splicing the user operation vector predicted by the user preference model with the fusion state vector to generate a target state vector;

[0037] A multi-agent deep deterministic policy gradient algorithm is used to calculate the control instructions corresponding to each of the household devices in the target state vector to generate a device control instruction set.

[0038] Optionally, the instruction verification module includes:

[0039] A multi-dimensional validity verification module is used to verify the feasibility of each instruction in the device control instruction set from the three dimensions of safety, functionality, and energy efficiency, and generate a feasibility score vector;

[0040] a dynamic risk assessment module, configured to perform a dynamic risk assessment on each instruction in the device control instruction set based on the feasibility score vector, historical fault data corresponding to the home system, and user preference data, and generate risk score data;

[0041] An instruction screening module is used to select instructions in the device control instruction set whose risk score data, feasibility score vector, and user preference weight corresponding to the home device meet a preset threshold to generate an initial control instruction set.

[0042] Optionally, the execution steps of the risk dynamic assessment module are:

[0043] Using the historical fault data corresponding to the household appliances to train a random forest model to generate a target risk assessment model;

[0044] Calculating the failure probability of each instruction in the device control instruction set by using the target risk assessment model to generate multiple failure probabilities;

[0045] Dynamically determine the weight data corresponding to each data in the feasibility score vector using the hierarchical analysis method according to the user preference data;

[0046] The weight data, the failure probability and the feasibility score vector are substituted into a preset risk score formula to calculate and obtain risk score data.

[0047] Optionally, the execution steps of the network adaptive transmission module are:

[0048] The network probe is used to obtain the delay rate, packet loss rate, available bandwidth and device signal strength of each protocol channel in the home system in real time to generate real-time network status data;

[0049] Predicting the network status of the home system using the real-time network status data and the historical network status data to generate network status prediction data;

[0050] Prioritizing the initial control instruction set according to the user preference data and the instruction types corresponding to the initial control instruction set to generate a classified instruction set;

[0051] Substituting the real-time network status data, the network status prediction data, the classification instruction set, and the network topology data corresponding to the home system into a preset comprehensive evaluation index formula to calculate a comprehensive evaluation index;

[0052] Selecting a transmission path corresponding to the initial control instruction set according to the comprehensive evaluation index to generate a target transmission path;

[0053] The target transmission path and the initial instruction corresponding to the target transmission path are used to construct a target control instruction.

[0054] The present invention also provides a remote IoT control method for household appliances, comprising:

[0055] Obtain the operating data, environmental data, and user operation data of each home appliance in the home system and perform edge computing on the data to generate target operating data, target environmental data, and target user operation data corresponding to the home appliance;

[0056] Using a deep reinforcement learning method to conduct online learning based on all of the target operation data, all of the target environment data, and all of the target user operation data to construct a device control instruction set;

[0057] Based on the historical fault data and user preference data corresponding to the home system, the feasibility of each instruction in the device control instruction set is verified through a preset virtual twin model to determine the initial control instruction set;

[0058] The transmission path corresponding to the initial control instruction set is dynamically selected according to the network delay and packet loss rate to generate the target control instruction.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] The present invention provides a remote IoT control system and control method for household appliances. The data acquisition module acquires the operating data, environmental data and user operation data of each household appliance in the household system, and performs edge computing on the data. The original data can be effectively screened, analyzed and processed to generate target operating data, target environmental data and target user operation data. The instruction generation module adopts a deep reinforcement learning method to conduct online learning based on the processed data to construct a device control instruction set. Deep reinforcement learning can continuously optimize the control instruction set to make it more in line with the actual operating conditions of household appliances and user needs. Through online learning, the system can adjust the control instructions in real time according to changes in the home environment and device status, thereby improving the pertinence and effectiveness of the instructions, further reducing the risk of device misoperation, and making the control of household appliances more intelligent and precise, thereby improving the user experience.

[0061] The command verification module uses a preset virtual twin model to verify the feasibility of each instruction in the device control instruction set based on historical fault data and user preference data corresponding to the home system, thereby determining the initial control instruction set. The virtual twin model simulates and verifies the control instructions before actual execution. Incorporating historical fault data can prevent the recurrence of instructions that have previously caused device failures or malfunctions. Incorporating user preference data can also align control instructions with user habits and expectations, effectively avoiding the potential for device malfunctions caused by directly issuing unverified control instructions. The network adaptive transmission module monitors network status in real time and dynamically adjusts transmission paths based on factors such as network latency and packet loss rate, ensuring that control instructions are transmitted to home devices in a timely and accurate manner. This effectively mitigates the impact of network fluctuations on control instruction transmission, improving the real-time and reliability of control instruction transmission, and avoiding delays or loss of instruction transmission due to network issues, thereby enhancing the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0063] Figure 1 This is a schematic structural diagram of a remote IoT control system for household appliances according to the present invention;

[0064] Figure 2 This is a structural diagram of a data acquisition module in a remote IoT control system for household appliances of the present invention;

[0065] Figure 3 This is a schematic structural diagram of an instruction generation module in a remote IoT control system for household appliances of the present invention;

[0066] Figure 4 This is a schematic structural diagram of a command verification module in a remote IoT control system for household appliances of the present invention;

[0067] Figure 5 This is a flowchart of the steps of a remote IoT control method for household appliances of the present invention. DETAILED DESCRIPTION

[0068] In order to better understand the technical content of the present invention, a specific embodiment is provided below, and the present invention is further described in conjunction with the accompanying drawings.

[0069] See also Figures 1 to 4 , the present invention provides a remote IoT control system for household appliances, comprising: a data acquisition module, an instruction generation module, an instruction verification module and a network adaptive transmission module;

[0070] The data acquisition module is used to obtain the operating data, environmental data, and user operation data of each home appliance in the home system and perform edge computing on the data to generate the target operating data, target environmental data, and target user operation data corresponding to the home appliance;

[0071] An instruction generation module is used to construct a device control instruction set by performing online learning based on all target operation data, all target environment data, and all target user operation data using a deep reinforcement learning method;

[0072] The instruction verification module is used to verify the feasibility of each instruction in the device control instruction set based on the historical fault data and user preference data corresponding to the home system through a preset virtual twin model, and determine the initial control instruction set;

[0073] The network adaptive transmission module is used to dynamically select the transmission path corresponding to the initial control instruction set according to the network delay and packet loss rate, and generate the target control instruction.

[0074] In the embodiment of the present invention, Figure 1As shown in the figure, the data acquisition module, instruction generation module, instruction verification module, and network adaptive transmission module are connected in sequence. The data acquisition module collects real-time operating data (such as air conditioner power consumption and refrigerator temperature), environmental data (such as indoor humidity and light intensity), and user operation data (such as mobile app control records and voice commands) from each home appliance. Using an LSTM (Long Short-Term Memory) network, it performs edge computing to generate high-precision target operating data, target environmental data, and target user operation data. The instruction generation module uses deep reinforcement learning to generate a set of device control instructions online based on all the target operating data, all target environmental data, and all target user operation data output by the data acquisition module. This enables dynamic learning and constructing intelligent instructions, replacing traditional fixed rules. The instruction verification module uses a virtual twin model to simulate the operation of the home system. Combining historical fault data and user preference data, it verifies the feasibility of instructions and selects a safe and efficient initial control instruction set. The network adaptive transmission module monitors network parameters such as network latency and packet loss rate in real time, dynamically selects the optimal transmission path based on instruction priority, generates target control instructions, and reliably sends them to home appliances. The invention solves the technical problems that the remote IoT control system of existing home appliances directly issues unverified control instructions, which may cause the equipment to operate incorrectly, and the control instructions are easily affected by network fluctuations, affecting real-time performance and leading to a poor user experience.

[0075] Preferably, the data acquisition module includes:

[0076] A data collection module is used to collect operating data, environmental data and user operation data of each home appliance in the home system, and generate multiple operating data, multiple environmental data and multiple user operation data;

[0077] A correlation data generation module is used to determine correlation data between home devices according to the device type, device location and device function corresponding to each operation data;

[0078] A data correction module is used to use machine learning methods to detect anomalies in the operating data, environmental data, and user operation data and perform data correction to generate initial operating data, initial environmental data, and initial user operation data corresponding to the home device;

[0079] The feature extraction module is used to extract time series features of the initial operation data, initial environment data and initial user operation data based on the associated data using the LSTM network to generate target operation data, target environment data and target user operation data corresponding to the home appliance.

[0080] In this embodiment of the present invention, a data acquisition module collects operational data, environmental data, and user operation data from various home appliances. Operational data includes device power, temperature, and speed; environmental data includes indoor temperature, humidity, and light intensity; and user operation data includes user switch commands and setting parameters.

[0081] The associated data generation module builds associations between devices based on device type (e.g., refrigeration equipment, lighting equipment), location (e.g., living room, bedroom), and function (e.g., temperature adjustment, on / off control) (e.g., "bedroom air conditioner linked to bedside lamp," "kitchen smoke sensor linked to gas valve"), generating device association data (e.g., association strength matrix, collaborative rule base). Traditional systems process individual device data in isolation (e.g., controlling the air conditioner based solely on the living room temperature) and ignore device collaboration (e.g., "leaving windows closed when the air conditioner is on, resulting in energy waste"). This module uses associated data (e.g., "window open status → air conditioner energy-saving mode triggered") to provide a global perspective for subsequent command generation, reducing ineffective operations (e.g., avoiding "forced cooling with windows open"). Association rules are constructed for complex scenarios such as "home mode" and "away mode" (e.g., "automatically turning off lights and lowering the refrigerator temperature to energy-saving mode when the user leaves home"), upgrading the system from single-device control to multi-device collaboration and improving the user experience.

[0082] The data correction module uses machine learning methods (such as time series anomaly detection models, generative adversarial networks (GANs), and user behavior models) to detect and correct anomalies in the raw data. This removes or repairs abnormal data caused by sensor failures, transmission errors, and user misoperation, generating clean initial operation / environmental / user operation data. The feature extraction module uses an LSTM network to extract time series features from the corrected initial data based on device association data, mining the temporal dependencies in the data to generate target operation data, target environment data, and target user operation data containing time series features for each home device.

[0083] Preferably, the execution steps of the data correction module are:

[0084] Perform timestamp alignment and data normalization on the operating data, environmental data, and user operation data to generate normalized data corresponding to home devices;

[0085] Perform data anomaly detection on normalized data using a preset time series anomaly detection model to generate operation and environment data anomaly location masks and user operation anomaly location masks;

[0086] The sliding window difference method is used to correct the normalized data corresponding to the short-term anomaly mask in the abnormal position mask of the operation and environmental data to generate the initial corrected data;

[0087] A generative adversarial network is used to correct the initial correction data corresponding to the long-term anomaly mask in the anomaly position mask of the operation and environment data to generate intermediate correction data;

[0088] The preset user behavior model is used to correct the intermediate correction data corresponding to the user operation abnormal position mask to generate target correction data;

[0089] The target correction data is denormalized to generate initial operation data, initial environment data and initial user operation data corresponding to the home device.

[0090] In an embodiment of the present invention, by aligning the timestamps of operating data, environmental data, and user operation data, the data is calibrated with a unified time base to ensure that the various types of data can be accurately associated at the same time. At the same time, due to the different physical dimensions of each data (such as temperature in ° C and device power in W), directly inputting the machine learning model will lead to model training deviation. Therefore, a data normalization method is used to uniformly scale the data to a specific interval (such as the 0-1 interval or a standard normal distribution interval with a mean of 0 and a standard deviation of 1) to eliminate the impact of dimensional differences on data features. After this processing, normalized data corresponding to home appliances is generated, providing standardized and unified data input for subsequent anomaly detection and feature extraction.

[0091] A preset time series anomaly detection model is used to analyze the normalized data at each time point. This model can be built based on an LSTM autoencoder. For example, the LSTM autoencoder learns from normal data to establish a normal time series pattern. When the input data does not conform to this pattern, the model outputs an anomaly signal.

[0092] After detection is complete, an abnormal location mask for operating and environmental data and an abnormal location mask for user operations are generated. The abnormal location mask for operating and environmental data marks the time points when abnormal device operating and environmental parameters occur; the abnormal location mask for user operations identifies abnormal behavior in user operation data (such as operating the device at unusual times or frequent, meaningless operation commands). These masks provide clear location information for subsequent targeted data corrections.

[0093] For the short-term anomalies marked in the abnormal position mask of the operation and environmental data (usually refers to abnormal data that lasts for 1-5 sampling cycles), this step uses the sliding window difference method for correction. The specific implementation method is: with the abnormal data point as the center, select n (n can be set according to actual conditions, such as n = 5) normal data points before and after to form a sliding window, calculate the mean or median of the data in the window, and fill the abnormal data position with this value. This method is suitable for repairing short-term accidental noise such as current spikes at the moment of equipment startup and instantaneous false alarms caused by slight vibrations of sensors. Its advantage is that it uses the time correlation of adjacent normal data to quickly and efficiently repair abnormal data, with low computational complexity, and can meet the needs of real-time processing of home appliance data, and generate initial correction data after correction.

[0094] For long-term anomalies in the anomaly position mask of operation and environmental data (data anomalies that last for more than 5 sampling cycles, such as continuous data drift caused by sensor aging), the sliding window difference method is difficult to effectively repair. This step uses a generative adversarial network (GAN) for processing. GAN consists of a generator and a discriminator. During the training process, the generator learns the distribution law of normal data and generates samples close to the real data; the discriminator distinguishes between the samples output by the generator and the real data. Through the adversarial training of the two, the generator can eventually output reasonable data that conforms to the distribution of real data. The data sequence generated by the generator is replaced with the data in the long-term anomaly area to obtain intermediate corrected data. Compared with traditional interpolation methods, GAN can better capture the complex distribution characteristics and change trends of data, effectively avoid equipment misjudgment caused by long-term data distortion, and significantly improve the accuracy and reliability of data repair.

[0095] Taking into account that there may be invalid operation instructions in the user operation data due to accidental touches on the APP, system failures, etc., this step uses a preset user behavior model to correct the intermediate correction data corresponding to the user operation abnormal position mask. The preset user behavior model is based on the user's historical operation data and is trained through a machine learning algorithm (such as a decision tree, a deep learning network). It can learn the user's operation patterns and habits. For example, if a user uses a coffee machine frequently from 7:00 to 9:00 on weekdays, but has no operation records during the same period on weekends, when the system detects operation data on the coffee machine in the early morning of a certain weekend, the user behavior model can determine that the operation may be an erroneous operation based on historical data, and use normal operation data from the same period in history to replace it, and finally generate target correction data. This step ensures that the user operation data can truly reflect the user's intentions and provide a reliable basis for the system to learn user preferences and generate personalized control instructions.

[0096] The target correction data obtained through the above correction steps has its original dimensions and numerical range changed during the normalization process. In order to restore the data to its original physical meaning and meet the requirements of the subsequent feature extraction module and local decision generation module, a denormalization operation is required. Specifically, according to the scaling parameters recorded during normalization (including minimum value, maximum value, mean value, standard deviation, etc.), the target correction data is inversely transformed and restored to its original dimensions, thereby generating the initial operation data, initial environmental data, and initial user operation data corresponding to the home appliance.

[0097] Preferably, the execution steps of the feature extraction module include:

[0098] The initial operation data, initial environment data, initial user operation data and associated data are spliced according to feature dimensions to obtain fused data;

[0099] The sliding window technology is used to divide the fused data into windows according to the preset division data to generate multiple window data blocks;

[0100] Input the window data modules into the preset LSTM network for feature extraction and generate feature vectors;

[0101] All feature vectors are linearly projected to generate target operation data, target environment data, and target user operation data corresponding to the home devices.

[0102] In this embodiment of the present invention, initial operating data, initial environmental data, initial user operation data, and associated data are spliced together according to feature dimensions. Initial operating data includes real-time operating parameters for household appliances, such as air conditioner operating current and temperature settings; initial environmental data covers environmental indicators such as indoor temperature, humidity, and light intensity; initial user operation data records control commands issued by users via apps, voice, and other means; and associated data defines the relationships between household appliances based on type, location, and function. By splicing together according to feature dimensions, previously scattered and independent data is integrated into fused data containing multiple aspects of information.

[0103] Using sliding window technology, the fused data is windowed according to preset partitioning data. These preset partitioning data include two key parameters: window size and sliding step size. The window size determines the amount of data contained in each data block, while the sliding step size controls the distance the window moves each time. For example, the window size can be set to 100 time series data points, and the sliding step size can be set to 10 data points. This means that the window moves 10 data points each time, generating a new window data block. Through sliding window partitioning, the continuous fused data is divided into multiple window data blocks with overlapping portions. This processing method can capture the changing characteristics of data over different time spans, preserving the local characteristics of the data while taking into account the long-term trends of the data.

[0104] The generated multiple window data blocks are fed into a pre-set LSTM network for feature extraction. During training, the LSTM network learns the time series patterns within the window data blocks and extracts features that reflect the inherent regularities of the data. After each window data block is processed by the LSTM network, a corresponding feature vector is output.

[0105] Linear projection is performed on all feature vectors to generate target operating data, target environment data, and target user operation data for the home devices. Linear projection reduces or transforms the feature vectors using a linear transformation matrix, mapping them to a dimensional space suitable for subsequent processing while retaining key features.

[0106] Preferably, the instruction generation module includes:

[0107] An environmental modeling module is used to construct an environmental model corresponding to the home system using all target operation data, all target environment data, and all target user operation data, generating multiple local environment models and a global environment model;

[0108] The user preference modeling module is used to learn the user operation patterns and user preferences corresponding to the target user operation data through a deep learning model to obtain a user preference model;

[0109] An environmental prediction module is used to perform environmental prediction on target environmental data using a time series analysis method to generate environmental prediction data;

[0110] The deep reinforcement learning decision module is used to determine the device control instructions corresponding to each home appliance based on the local environment model, the global environment model, the user preference model and the environmental prediction data, and generate the device control instruction set using the deep reinforcement learning algorithm.

[0111] In an embodiment of the present invention, the environmental modeling module constructs an environmental model of the home system based on the target operating data, target environmental data, and target user operation data output by the data correction module. Among them, the target operating data covers the real-time operating parameters of home appliances, the target environmental data includes environmental indicators such as indoor and outdoor temperature and humidity, light intensity, and the target user operation data records the user's control instructions and operating habits for the equipment. This module generates multiple local environmental models and global environmental models through hierarchical modeling. Among them, for a single room or specific area (such as a living room or bedroom), a refined local environmental model is constructed based on the equipment operating data, environmental data, and user operation data in the area. Starting from the overall level of the home system, all local environmental models and the associated data between each area are integrated to construct a global environmental model, which can present the overall operating status of the home system.

[0112] The user preference modeling module uses deep learning models (such as recurrent neural networks) to conduct in-depth analysis of target user operation data and learn users' operation patterns and preferences. Target user operation data includes control commands issued by users through various methods such as mobile phone apps, voice assistants, physical buttons, as well as information such as operation time and frequency. By training on a large amount of historical operation data, the deep learning model can capture the underlying patterns and personalized needs of user operations. Through continuous learning and optimization, a user preference model is ultimately obtained. This model converts users' personalized needs into quantifiable parameters, providing a key basis for generating control commands that meet users' expectations, solving the problem that traditional systems cannot meet users' personalized needs.

[0113] The environmental prediction module uses time series analysis methods (such as the autoregressive integrated moving average model) to analyze and predict target environmental data. This target environmental data includes time-varying environmental indicators such as indoor temperature and humidity, light intensity, and air quality, all of which exhibit distinct time series characteristics. Time series analysis methods predict future environmental conditions by mining trends, seasonality, and cyclical patterns in historical data. The resulting environmental prediction data can proactively reflect environmental trends, making the control system proactive and avoiding untimely control issues caused by delayed response to environmental changes.

[0114] The deep reinforcement learning decision module integrates local and global environmental models, user preference models, and environmental prediction data, and uses deep reinforcement learning algorithms (such as the multi-agent deep deterministic policy gradient algorithm) to generate device control instruction sets. This module views the home system's operating environment as a dynamic decision space, with each home device acting as an independent agent.

[0115] Preferably, the execution steps of the deep reinforcement learning decision module are:

[0116] The local environment state data corresponding to the local environment model, the global environment state data corresponding to the global environment model and the environment prediction data are concatenated to generate a comprehensive state vector;

[0117] Use the attention network to calculate the attention weights of each part of the comprehensive state vector and generate attention weight data;

[0118] Use attention weight data to perform weighted summation on the comprehensive state vector to generate a fusion state vector;

[0119] The user operation vector predicted by the user preference model is concatenated with the fusion state vector to generate the target state vector;

[0120] A multi-agent deep deterministic policy gradient algorithm is used to calculate the control instructions corresponding to each home appliance in the target state vector and generate a device control instruction set.

[0121] In this embodiment of the present invention, local environmental status data corresponding to the local environmental model, global environmental status data corresponding to the global environmental model, and environmental prediction data are spliced together. Local environmental status data reflects the operating parameters and environmental indicators of devices in a single room or area; global environmental status data provides an overview of the entire home system's operation, including energy consumption distribution throughout the house and the linkage between devices in each area; and environmental prediction data provides a prediction of future environmental trends. Through data splicing, environmental information of different dimensions and scales is integrated into a comprehensive state vector that encompasses the current and future states of the home system.

[0122] The attention network is used to analyze the comprehensive state vector and calculate the attention weight of each part of the data. The corresponding expression of the attention network is:

[0123]

[0124] Among them, α i is the comprehensive state vector s concat The attention weight of the i-th part in , which reflects the importance of this part in the fusion process; is the comprehensive state vector s concat The i-th component in s (such as local environment state, global environment state or environment prediction data, etc.); w is a learnable weight vector used to adjust the importance of the input; V is a learnable matrix that concat Perform linear transformation to extract features; b is a learnable bias term that increases the model's fitting ability; m is the comprehensive state vector s concat The total number of components (for example, local environmental state, global environmental state, and environmental prediction data, a total of m=3 parts).

[0125] Each element in the integrated state vector is multiplied by its corresponding attention weight, and all weighted elements are summed to generate a fused state vector. An attention network is used to calculate the weights of each component of the integrated state vector, highlighting the impact of key environmental factors on decision-making (e.g., air conditioning operating status and indoor temperature have higher weights during high summer temperatures). This attention weight data is then generated and weighted summed over the integrated state vector to produce the fused state vector. This step optimizes the integrated state vector, highlighting the role of key environmental information and ensuring that the fused state vector better reflects the core environmental characteristics required for the current decision. The user action vector predicted by the user preference model is concatenated with the fused state vector to form a target state vector that reflects the user's personalized needs. Based on the target state vector, a multi-agent deep deterministic policy gradient algorithm, through trial and error and learning, determines the optimal control instructions for each home device, while meeting safety, functionality, and energy efficiency constraints, and generates a device control instruction set.

[0126] Preferably, the instruction verification module includes:

[0127] A multi-dimensional validity verification module is used to verify the feasibility of each instruction in the device control instruction set from the three dimensions of safety, functionality, and energy efficiency, and generate a feasibility score vector;

[0128] A dynamic risk assessment module is used to dynamically assess the risk of each instruction in the device control instruction set based on the feasibility score vector, historical fault data corresponding to the home system, and user preference data, and generate risk score data;

[0129] The instruction screening module is used to select instructions in the device control instruction set whose risk score data, feasibility score vector and user preference weight corresponding to the home device meet the preset threshold to generate an initial control instruction set.

[0130] In an embodiment of the present invention, the multi-dimensional validity verification module is intended to perform comprehensive and systematic feasibility verification on each instruction in the device control instruction set from three key dimensions: safety, functionality, and energy efficiency, and generate a feasibility scoring vector.

[0131] Safety Verification: Checks whether the instruction exceeds the device safety threshold (such as electric water heater temperature ≤ 80°C, socket current ≤ 10A). The corresponding scoring formula is: Among them, F A is the total number of parameters involved in security verification in the instruction; a i is the i-th parameter in the instruction (such as the temperature instruction value of the electric water heater, the current instruction value of the socket, etc.); Q is the indicator function, when a i Falling in the safe range [L i , U i ], Q=1; otherwise, Q=0; [L i , U i ] is the safety interval of the i-th parameter (such as the safety interval of the electric water heater temperature [L i , U i ]=[-∞,80℃],the socket current safety range [L i , U i ]=[-∞,10A].

[0132] Functional verification: Use a virtual twin model to simulate instruction execution and determine whether the expected goal is achieved (e.g., room temperature change ≤ ±2°C within 10 minutes): Among them, S sim (t+τ) is the device state after the instruction execution time τ is simulated by the virtual twin model; S target is the expected target state of instruction execution; S max is the possible maximum value of the device state; S min The minimum possible value for the device status.

[0133] Energy efficiency verification: Calculate the expected energy consumption after the instruction is executed (such as power consumption in air conditioning cooling mode): E consume =P(a t )·τ, Among them, P(a t ) is the same as instruction a t Corresponding device power function; E consume For instruction a t The estimated energy consumption after execution; τ is the instruction a t Length of execution time; E base For instruction a t Baseline energy consumption after execution (for comparative evaluation of energy efficiency).

[0134] The dynamic risk assessment module dynamically assesses the risk of each instruction in the device control instruction set based on the feasibility score vector output by the multi-dimensional validity verification module, combined with historical failure data and user preference data for the home system. This generates risk scores. Utilizing historical failure data: By collecting and analyzing past failure records of home devices, the module trains machine learning algorithms such as random forest models to construct a target risk assessment model. This model predicts the probability of each instruction causing a failure during actual execution based on historical failure patterns. For example, if a certain brand of water heater has historically caused circuit failures due to frequent on and off cycles, the model will assign a higher probability of failure to the instruction set when frequent on and off commands for that water heater appear in the control instruction set. User preference data integration: Taking into account the differences in user habits and preferences for home devices, this module uses the analytic hierarchy process to dynamically determine the weighting of each dimension in the feasibility score vector based on user preference data. For example, for users who prioritize energy conservation, energy efficiency will be given a higher weight; for users who prioritize safe device operation, safety will be given a higher weighting. Risk Score Calculation: Substitute the aforementioned weighted data, failure probability, and feasibility score vector into a pre-set risk scoring formula to generate a comprehensive risk score for each command. This score comprehensively reflects the potential risk level of the command given the current home system environment and user preferences, providing a quantitative basis for command screening.

[0135] The instruction screening module screens instructions within the device control instruction set based on the risk score data and feasibility score vector generated by the dynamic risk assessment module, as well as the user preference weights corresponding to the home devices. Specifically, this module pre-sets a series of thresholds, including a risk score threshold, a feasibility score threshold, and a user preference weight threshold. An instruction is selected to generate the initial control instruction set only if its risk score falls below the pre-set risk threshold, the scores in the feasibility score vector meet the functional requirement thresholds for the corresponding dimensions, and the user preference weights are met.

[0136] Preferably, the execution steps of the risk dynamic assessment module are:

[0137] Use historical fault data corresponding to home appliances to train a random forest model to generate a target risk assessment model;

[0138] The target risk assessment model is used to calculate the failure probability of each instruction in the device control instruction set and generate multiple failure probabilities;

[0139] According to the user preference data, the hierarchical analysis method is used to dynamically determine the weight data corresponding to each data in the feasibility score vector;

[0140] Substitute the weight data, failure probability and feasibility score vector into the preset risk score formula to calculate the risk score data.

[0141] In this embodiment of the present invention, a random forest model is trained using historical fault data corresponding to household appliances to generate a target risk assessment model. This historical fault data includes multi-dimensional information such as the time the fault occurred, the fault type, the device state before the fault, and related environmental parameters. For example, for a smart refrigerator, historical fault data might include the operating time of the compressor at the time of the fault, abnormal cooling temperature records, and voltage fluctuations at the time.

[0142] As an ensemble learning algorithm, the random forest model effectively handles high-dimensional data and complex relationships by constructing multiple decision trees and synthesizing their results. During training, the model automatically learns the correlation between features in historical failure data and the occurrence of failures. For example, it was found that the probability of failure increases significantly when a refrigerator compressor operates continuously for more than a certain period of time and the ambient temperature exceeds a certain threshold. After multiple rounds of training and optimization, the resulting target risk assessment model is capable of predicting the probability of equipment failure based on input data, providing a core tool for subsequent instruction risk assessment.

[0143] Each instruction in the device control instruction set is input into the target risk assessment model. The model calculates the corresponding failure probability for each instruction, generating multiple failure probability data. During the calculation process, the model analyzes the device type, operation method, and current home system status of the instruction, combined with the learned failure patterns.

[0144] For example, if an instruction is to frequently turn a certain brand of water heater on and off, the target risk assessment model will retrieve historical failure data for that model of water heater that resulted from frequent on and off cycles. Based on this data, it calculates the probability of the instruction causing a failure. Each instruction is assigned a failure probability value, which quantifies the likelihood of equipment failure after the instruction is executed, providing a key basis for comprehensively assessing instruction risk.

[0145] Based on user preference data, the Analytic Hierarchy Process (AHP) is used to dynamically determine the weights corresponding to each data point in the feasibility score vector. User preference data includes the user's emphasis on safety, functionality, and energy efficiency of home devices. For example, some users prioritize device safety and are willing to sacrifice energy efficiency for safety; others prioritize energy conservation and have relatively low requirements for device functionality.

[0146] The Analytic Hierarchy Process (AHP) constructs a hierarchical model to break down user preferences into different levels of factors. For example, it categorizes home appliance evaluation factors into safety, functionality, and energy efficiency. It then compares each factor pairwise to determine its relative importance and calculates the weight of each factor within the feasibility score vector. For example, analysis may determine that for a particular user, safety has a weight of 0.5, functionality has a weight of 0.3, and energy efficiency has a weight of 0.2. These weights are dynamically adjusted as user preferences evolve, ensuring that risk assessments fully reflect individual user needs.

[0147] The determined weights, calculated failure probabilities, and the feasibility score vector output by the multi-dimensional validity verification module are substituted into a pre-set risk scoring formula to calculate the final risk score. This formula comprehensively considers the feasibility of the instruction, the failure probability, and the user's preference weight, integrating these factors into a quantified risk score through specific mathematical operations.

[0148] The default risk scoring formula is: Risk Score = Safety Score × Safety Weight × (1 + Failure Probability) + Functionality Score × Functionality Weight × (1 + Failure Probability) + Energy Efficiency Score × Energy Efficiency Weight × (1 + Failure Probability). Substituting the specific weight data, failure probability, and each score in the feasibility score vector into the formula, the corresponding risk score data for each instruction can be obtained.

[0149] Preferably, the execution steps of the network adaptive transmission module are:

[0150] Through network probes, the delay rate, packet loss rate, available bandwidth, and device signal strength of each protocol channel in the home system are obtained in real time to generate real-time network status data;

[0151] Use real-time network status data and historical network status data to predict the network status of the home system and generate network status prediction data;

[0152] Prioritizing the initial control instruction set according to the user preference data and the instruction type corresponding to the initial control instruction set to generate a classified instruction set;

[0153] Substituting the real-time network status data, network status prediction data, classification instruction set, and network topology data corresponding to the home system into a preset comprehensive evaluation index formula to calculate a comprehensive evaluation index;

[0154] Selecting a transmission path corresponding to the initial control instruction set according to the comprehensive evaluation index to generate a target transmission path;

[0155] A target transmission path and an initial instruction corresponding to the target transmission path are used to construct a target control instruction.

[0156] In an embodiment of the present invention, network probes are deployed in the smart home network to collect key network parameters of each protocol channel (such as Wi-Fi, Zigbee, and Bluetooth) in the home system in real time, including delay rate, packet loss rate, available bandwidth, and device signal strength, to generate real-time network status data. Taking the home Wi-Fi network as an example, when multiple devices are playing high-definition videos online at the same time, the network probe can immediately monitor the increase in Wi-Fi channel delay rate and the decrease in available bandwidth; if a smart device is far away from the router, its signal strength data will be significantly weakened. These real-time data provide the most realistic network status basis for subsequent network status analysis and path selection, avoiding errors in transmission path selection due to the use of lagging network information.

[0157] By combining real-time and historical network status data, and applying techniques such as time series analysis and machine learning prediction algorithms, the system can predict the home system's network status for a period of time in the future, generating network status forecasts. For example, based on historical data showing home network congestion due to multiple devices connecting to the network between 8:00 PM and 10:00 PM each night, combined with current real-time network status, it predicts that Wi-Fi network latency may rise to 200ms and packet loss to 8% during that time period. This forward-looking network status prediction enables the system to plan transmission paths in advance, avoiding temporary command transmission failures due to network deterioration and enhancing the system's ability to cope with network fluctuations.

[0158] Based on user preference data and the corresponding instruction types of the initial control instruction set, the initial control instruction set is prioritized to generate a classified instruction set. Emergency instructions, such as powering off devices triggered by fire alarms or closing valves in the event of a gas leak, are related to home safety and are given the highest priority. High-frequency operation instructions, such as daily light on / off and air conditioning temperature adjustment, affect the user's immediate experience and are given the next highest priority. Non-real-time instructions, such as device firmware upgrades, have lower timeliness requirements and are given the lowest priority. This prioritization ensures that critical instructions are transmitted first when network resources are limited, ensuring the timeliness and safety of home device control.

[0159] Substitute real-time network status data, network status prediction data, classified instruction sets, and network topology data corresponding to the home system (describing the device connection relationships and paths in the network) into the preset comprehensive evaluation index formula. This formula comprehensively considers factors such as network parameters (delay, packet loss, bandwidth), instruction priority, and network topology. The preset comprehensive evaluation index formula is:

[0160]

[0161] Among them, Score is a comprehensive evaluation index; μ is the first weight coefficient; β is the second weight coefficient; γ is the third weight coefficient; δ is the fourth weight coefficient; Dalay represents the network transmission delay, which refers to the time (in milliseconds) required for data to travel from the sender to the receiver on the selected transmission path. The smaller the delay, The larger the value, the higher the positive contribution to the Score, reflecting the advantage of low-latency paths for transmission. LossRate represents the network packet loss rate, that is, the ratio of the number of data packets lost during transmission to the total number of data packets sent (such as 0.1 represents a 10% packet loss rate). 1-LossRate ensures that the lower the packet loss rate, the greater the contribution of this part to the Score, reflecting the pursuit of complete data transmission. Bandwidth refers to the available bandwidth of the transmission path, which measures the amount of data that can be transmitted per unit time (units such as Mbps). The larger the bandwidth, the stronger the positive impact on the Score, reflecting the improvement of data transmission efficiency on high-bandwidth paths. Priority indicates the priority of the instruction (such as urgent instructions have high priority and ordinary instructions have low priority). The higher the priority, the greater the contribution to the Score, making it easier for high-priority instructions to match high-quality transmission paths.

[0162] Based on the calculated comprehensive evaluation indicators, the optimal path is selected from all possible transmission path options to generate the target transmission path. For example, if the comprehensive evaluation indicators show that under the current network conditions, the comprehensive evaluation indicators for transmitting emergency commands through the Zigbee channel are optimal (low latency, low packet loss rate, and meeting command priority requirements), the Zigbee channel is determined as the target transmission path for the emergency command. Even in complex multi-protocol mixed network environments, this step can quickly identify the path most suitable for transmitting each command, ensuring that the command is efficiently delivered to the device.

[0163] The target control command is constructed using the determined target transmission path and the initial command corresponding to that path. The initial command is encapsulated and adapted according to the protocol format of the target transmission path. For example, if the target transmission path is a Bluetooth channel, the initial command is converted into a data format recognizable by the Bluetooth communication protocol and then sent to the target home device via that path. This ensures the accurate and timely transmission of control commands in complex network environments, effectively improving the stability and user experience of remote IoT control of home devices.

[0164] Preferably, Figure 5 As shown, the present invention also provides a remote IoT control method for household appliances, comprising:

[0165] Step 101: Obtain the operating data, environmental data, and user operation data of each home appliance in the home system and perform edge computing on the data to generate target operating data, target environmental data, and target user operation data corresponding to the home appliance;

[0166] Step 102: Using a deep reinforcement learning method to perform online learning based on all target operation data, all target environment data, and all target user operation data, to construct a device control instruction set;

[0167] Step 103: Based on the historical fault data and user preference data corresponding to the home system, the feasibility of each instruction in the device control instruction set is verified through a preset virtual twin model to determine the initial control instruction set;

[0168] Step 104: Dynamically select a transmission path corresponding to the initial control instruction set according to the network delay and packet loss rate, and generate a target control instruction.

[0169] In the embodiment of the present invention, the operation data, environmental data and user operation data of each home appliance in the home system are obtained. Edge computing is performed on these data to explore the patterns and dependencies of the data in the time dimension. Calculation is performed on the edge side, and there is no need to upload a large amount of raw data to the cloud, which greatly reduces the data transmission delay and the dependence on the cloud server, and improves the real-time performance of data processing. Through this process, the raw data is converted into target operation data, target environmental data and target user operation data.

[0170] Using deep reinforcement learning, online learning is performed based on the target operating data, target environment data, and target user operation data generated above. The deep reinforcement learning algorithm learns from the target data the complex relationships between device operation, the environment, and user operations, continuously adjusting the control strategy. Through continuous online learning, a device control instruction set is constructed. This instruction set is derived from in-depth analysis and learning of various data. Compared to traditional fixed-rule instructions, this instruction set is more intelligent and adaptable, better meeting the dynamic changes in the home environment and the personalized needs of users.

[0171] Based on the historical fault data and user preference data corresponding to the home system, the feasibility of each instruction in the device control instruction set is verified through a preset virtual twin model. The virtual twin model is a digital mapping of the home system, which can simulate the process and effect of instruction execution in a virtual environment. After the instruction set is input into the virtual twin model, the instructions are evaluated from multiple dimensions such as safety, functionality, and energy efficiency. Through this verification process, risky or unfeasible instructions are eliminated, and the initial control instruction set is determined to ensure that the instructions finally issued will not cause damage to the equipment, while meeting user needs, avoiding equipment misoperation due to unverified instructions, and ensuring the safe and stable operation of home appliances.

[0172] The transmission path corresponding to the initial control instruction set is dynamically selected based on network latency and packet loss rate to generate the target control instructions. In actual home network environments, network conditions are complex and changeable. Problems such as Wi-Fi congestion and signal interference can cause network latency and packet loss, affecting the timeliness and accuracy of instruction transmission. By monitoring parameters such as network latency and packet loss rate in real time, combined with the type and priority of the instructions, the optimal path is selected from multiple transmission paths. After determining the transmission path, the initial control instructions are encapsulated according to the protocol and format of the path to generate the target control instructions. This ensures that the instructions can be transmitted quickly and accurately to the target home devices, enabling remote control of the devices. This effectively addresses the impact of network fluctuations on instruction transmission and improves the real-time and smoothness of users' remote control of home devices.

[0173] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A remote IoT control system for household appliances, characterized in that: include: Data acquisition module, instruction generation module, instruction verification module and network adaptive transmission module; The data acquisition module is used to obtain the operating data, environmental data and user operation data of each home appliance in the home system and perform edge computing on the data to generate target operating data, target environmental data and target user operation data corresponding to the home appliance; The instruction generation module is used to use a deep reinforcement learning method to perform online learning based on all the target operation data, all the target environment data, and all the target user operation data to construct a device control instruction set; The instruction verification module is used to verify the feasibility of each instruction in the device control instruction set based on the historical fault data and user preference data corresponding to the home system through a preset virtual twin model, and determine the initial control instruction set; The network adaptive transmission module is used to dynamically select the transmission path corresponding to the initial control instruction set according to network delay and packet loss rate, and generate target control instructions.

2. A remote IoT control system for household appliances according to claim 1, characterized in that: The data acquisition module includes: A data collection module is used to collect operating data, environmental data and user operation data of each home appliance in the home system, and generate multiple operating data, multiple environmental data and multiple user operation data; an association data generating module, configured to determine association data between the household devices according to the device type, device location, and device function corresponding to each of the operation data; a data correction module, configured to use a machine learning method to perform anomaly detection and data correction on the operating data, the environmental data, and the user operation data, respectively, to generate initial operating data, initial environmental data, and initial user operation data corresponding to the home appliance; A feature extraction module is used to extract time series features of the initial operation data, the initial environment data and the initial user operation data using an LSTM network based on the associated data to generate target operation data, target environment data and target user operation data corresponding to the home appliance.

3. A remote IoT control system for household appliances according to claim 2, characterized in that: The execution steps of the data correction module are: Performing timestamp alignment and data normalization on the operation data, the environment data, and the user operation data respectively to generate normalized data corresponding to the home appliance; Performing data anomaly detection on the normalized data using a preset time series anomaly detection model to generate an abnormal location mask for operation and environment data and an abnormal location mask for user operation; Correcting the normalized data corresponding to the short-term anomaly mask in the abnormal position mask of the operation and environmental data using a sliding window difference method to generate initial corrected data; Using a generative adversarial network to correct the initial correction data corresponding to the long-term anomaly mask in the anomaly position mask of the operation and environment data to generate intermediate correction data; Using a preset user behavior model to correct the intermediate correction data corresponding to the user operation abnormal position mask to generate target correction data; The target correction data is denormalized to generate initial operation data, initial environment data and initial user operation data corresponding to the home appliance.

4. A remote IoT control system for household appliances according to claim 2, characterized in that: The execution steps of the feature extraction module include: splicing the initial operation data, the initial environment data, the initial user operation data, and the associated data according to feature dimensions to obtain fused data; Using a sliding window technique to divide the fused data into windows according to preset division data, generating a plurality of window data blocks; Input the window data modules into the preset LSTM network for feature extraction to generate feature vectors; All the feature vectors are linearly projected to generate target operation data, target environment data and target user operation data corresponding to the home appliance.

5. The remote IoT control system for household appliances according to claim 1, characterized in that: The instruction generation module includes: An environment modeling module, configured to construct an environment model corresponding to the home system using all of the target operation data, all of the target environment data, and all of the target user operation data, and to generate a plurality of local environment models and a global environment model; A user preference modeling module is used to learn the user operation mode and user preference corresponding to the target user operation data through a deep learning model to obtain a user preference model; An environmental prediction module, configured to perform environmental prediction on the target environmental data using a time series analysis method to generate environmental prediction data; A deep reinforcement learning decision module is used to determine the device control instructions corresponding to each of the home appliances based on the local environment model, the global environment model, the user preference model and the environmental prediction data, and generate a device control instruction set using a deep reinforcement learning algorithm.

6. A remote IoT control system for household appliances according to claim 5, characterized in that: The execution steps of the deep reinforcement learning decision module are: Performing data splicing on the local environment state data corresponding to the local environment model, the global environment state data corresponding to the global environment model, and the environment prediction data to generate a comprehensive state vector; Calculating the attention weights of each part of the comprehensive state vector using an attention network to generate attention weight data; Performing weighted summation on the integrated state vector using the attention weight data to generate a fusion state vector; splicing the user operation vector predicted by the user preference model with the fusion state vector to generate a target state vector; A multi-agent deep deterministic policy gradient algorithm is used to calculate the control instructions corresponding to each of the household devices in the target state vector to generate a device control instruction set.

7. The remote IoT control system for household appliances according to claim 1, characterized in that: The instruction verification module includes: A multi-dimensional validity verification module is used to verify the feasibility of each instruction in the device control instruction set from the three dimensions of safety, functionality, and energy efficiency, and generate a feasibility score vector; a dynamic risk assessment module, configured to perform a dynamic risk assessment on each instruction in the device control instruction set based on the feasibility score vector, historical fault data corresponding to the home system, and user preference data, and generate risk score data; An instruction screening module is used to select instructions in the device control instruction set whose risk score data, feasibility score vector, and user preference weight corresponding to the home device meet a preset threshold to generate an initial control instruction set.

8. A remote IoT control system for household appliances according to claim 7, characterized in that: The execution steps of the risk dynamic assessment module are: Using the historical fault data corresponding to the household appliances to train a random forest model to generate a target risk assessment model; Calculating the failure probability of each instruction in the device control instruction set by using the target risk assessment model to generate multiple failure probabilities; Dynamically determine the weight data corresponding to each data in the feasibility score vector using the hierarchical analysis method according to the user preference data; The weight data, the failure probability and the feasibility score vector are substituted into a preset risk score formula to calculate and obtain risk score data.

9. A remote IoT control system for household appliances according to claim 1 or 7, characterized in that: The execution steps of the network adaptive transmission module are: The network probe is used to obtain the delay rate, packet loss rate, available bandwidth and device signal strength of each protocol channel in the home system in real time to generate real-time network status data; Predicting the network status of the home system using the real-time network status data and the historical network status data to generate network status prediction data; Prioritizing the initial control instruction set according to the user preference data and the instruction types corresponding to the initial control instruction set to generate a classified instruction set; Substituting the real-time network status data, the network status prediction data, the classification instruction set, and the network topology data corresponding to the home system into a preset comprehensive evaluation index formula to calculate a comprehensive evaluation index; Selecting a transmission path corresponding to the initial control instruction set according to the comprehensive evaluation index to generate a target transmission path; The target transmission path and the initial instruction corresponding to the target transmission path are used to construct a target control instruction.

10. A remote IoT control method for household appliances, characterized in that: include: Obtain the operating data, environmental data, and user operation data of each home appliance in the home system and perform edge computing on the data to generate target operating data, target environmental data, and target user operation data corresponding to the home appliance; Using a deep reinforcement learning method to conduct online learning based on all of the target operation data, all of the target environment data, and all of the target user operation data to construct a device control instruction set; Based on the historical fault data and user preference data corresponding to the home system, the feasibility of each instruction in the device control instruction set is verified through a preset virtual twin model to determine the initial control instruction set; The transmission path corresponding to the initial control instruction set is dynamically selected according to the network delay and packet loss rate to generate the target control instruction.

Images