Smart home adaptive control system based on user habits

Abstract

The invention relates to the technical field of intelligent furniture control, in particular to an intelligent home adaptive control system based on user habits, which comprehensively depicts the user habits and outputs habit mode vectors through multi-modal feature extraction and fusion to provide a basis for accurate control. An environment context feature vector is generated by constructing a dynamic environment state map, so that a system can sense environment change; a dynamic equipment control strategy is generated by splicing habit and environment vectors through a neural network and a reinforcement learning algorithm, and user habits and environment factors are considered; a control signal is generated by analyzing a strategy mapping table, random delay is injected to avoid conflicts, and accurate execution and stable operation of equipment are ensured through verification and adjustment of a feedback loop; the performance is evaluated by collecting feedback data, habit changes are adapted, continuous optimization of the system is guaranteed, the time sequence and resource allocation are controlled by coordinating multiple devices, and the overall operation efficiency and the user experience are improved.

Description

Technical Field

[0001] This invention relates to the field of smart furniture control technology, and in particular to a smart home adaptive control system based on user habits. Background Technology

[0002] Smart home systems integrate various sensors and smart devices to provide users with a personalized, automated living experience. Existing technologies typically employ rule-based or single-modal machine learning methods to achieve adaptive control: the system uses a pre-set rule base or a behavioral model trained with single-dimensional data such as time or simple device status to infer user intent and execute device control. This mode can achieve basic automation in ideal scenarios where user habits are fixed and the environment is simple.

[0003] However, when faced with dynamic and complex real-world home environments, the limitations of existing methods become apparent: user behavior is inherently multimodal (e.g., the temporal regularity of operation sequences, the instantaneous semantics of voice commands, and the frequency domain characteristics of activity rhythms). Existing methods lack in-depth exploration of the intrinsic connections between multimodal behaviors, resulting in a one-sided and delayed identification of users' potential habitual intentions. The environmental information perceived by the system (e.g., temperature, light, sound) is interconnected and dynamically changing. Existing methods treat each environmental factor as an independent parameter, failing to construct its internal dynamic correlation map, leading to a superficial understanding of complex environmental states. These limitations prevent the system from achieving proactive and precise control when user habits subtly change or the environment undergoes abrupt changes. Its control strategies often exhibit mechanical responses or frequent false triggers, severely restricting the smoothness of the intelligent experience and user trust. Therefore, this invention proposes a smart home adaptive control system based on user habits. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the background art and to propose a smart home adaptive control system based on user habits.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A smart home adaptive control system based on user habits includes: a multimodal habit modeling module, an environmental context perception module, an adaptive control strategy generation module, a device control execution module, and a system feedback and optimization module; The multimodal habit modeling module is used to construct a multimodal deep behavior model and output a habit pattern vector, wherein the habit pattern vector includes at least habit strength, habit stability and habit triggering condition probability; The environment context awareness module is used to generate environment context feature vectors and construct a dynamic environment state map; The adaptive control strategy generation module is used to generate dynamic device control strategies based on habitual pattern vectors and environmental context feature vectors. The device control execution module is used to control the operating status of one or more smart home devices according to a dynamic device control strategy. The system feedback and optimization module is used to monitor user feedback signals and device status, and optimize the multimodal deep behavior model and dynamic device control strategy based on the feedback data.

[0006] Optionally, the process of constructing a multimodal deep behavioral model and outputting habit pattern vectors by the multimodal habit modeling module includes: S21. Collect user interaction data with smart home devices, perform data cleaning and standardization, and output the cleaned interaction data. S22. Perform multimodal feature extraction on the cleaned interactive data, and extract its temporal feature vector, frequency domain feature vector and semantic feature vector respectively; S23. Based on the extracted temporal feature vector, frequency domain feature vector and semantic feature vector, an attention mechanism is used to perform weighted fusion, and the fused features are input into the first neural network composed of a long short-term memory network and a Transformer encoder for deep behavior modeling to generate an implicit representation of user habits. S24. Based on implicit representation, dimension mapping and normalization are performed through a fully connected layer to output a habit pattern vector, where the vector dimension includes habit strength, habit stability and habit triggering condition probability.

[0007] Optionally, the weighted fusion process using an attention mechanism in S23 includes: S231. Input the temporal feature vector, frequency domain feature vector and semantic feature vector into the multi-head attention layer in parallel; S232. Based on the current time context, the multi-head attention layer calculates an attention score for each feature vector; S233. Based on the calculated attention score, the temporal feature vector, frequency domain feature vector and semantic feature vector are weighted and summed to generate the fused feature vector.

[0008] Optionally, the process by which the environment context-aware module generates environment context feature vectors and constructs a dynamic environment state map includes: S31. Collect environmental sensor data, perform data cleaning and standardization, and output the cleaned environmental sensor data. S32. Perform time series alignment and sliding window segmentation on the cleaned environmental sensor data to form environmental sensor data sequence blocks; S33. Based on the data sequence blocks of environmental sensors, their spatial and temporal features are extracted through a feature extraction network; wherein, the feature extraction network is composed of a one-dimensional convolutional neural network and a gated recurrent unit connected in series, the one-dimensional convolutional neural network is used to extract local spatial patterns, and the gated recurrent unit is used to capture long-term dependencies. S34. Based on the spatial and temporal features output by the feature extraction network, generate dynamic environmental context feature vectors and construct an environmental state map.

[0009] Optionally, the adaptive control strategy generation module generates a dynamic device control strategy based on habitual pattern vectors and environmental context feature vectors, including the following process: S41. Concatenate the habitual pattern vector with the environmental context feature vector to form a joint feature vector; S42. Input the joint feature vector into the pre-trained second neural network and output a preliminary set of device control actions; wherein, the second neural network is a deep feedforward neural network; S43. Based on the initial set of device control actions and combined with the real-time device status read from the device status register, the strategy is optimized and the action is selected through the reinforcement learning Actor-Critic algorithm to generate the final dynamic device control strategy. At the same time, a strategy mapping table containing device identifier, control parameters and execution time window is generated.

[0010] Optionally, the process by which the device control execution module controls the operating status of one or more smart home devices according to a dynamic device control strategy includes: S51. Parse the policy mapping table to obtain the control commands for the target device; S52. Based on the control commands of the target device, generate low-level control signals through the device communication protocol stack, and inject random delays when distributing the low-level control signals. S53. After the bottom-level control signal is issued, the device status is verified in real time through the device status feedback loop to see if the device status matches the expected control effect. If there is a deviation, the control intensity is dynamically adjusted through the soft actuator.

[0011] Optionally, the process of injecting random delays during the distribution of low-level control signals in S52 includes: S521. For each underlying control signal data frame to be distributed, determine the reference delay parameter according to its corresponding device type and instruction priority; S522. Based on the baseline delay parameter, independently sample random delay values ​​for each data frame from a preset probability distribution; S523. Append the sampled random delay value to the corresponding data frame, and put the data frame into the physical transmission queue according to the delayed time point.

[0012] Optionally, the system feedback and optimization module monitors user feedback signals and device status, and optimizes the multimodal deep behavior model and dynamic device control strategy based on the feedback data, including the following process: S61. Real-time collection of explicit user feedback data and implicit behavior correction data, combined with actual device status data obtained from the device control execution module, to calculate user satisfaction index and system performance index. S62. Construct a comprehensive system performance evaluation score based on user satisfaction indicators and system performance indicators; S63. When the overall system performance evaluation score is lower than the preset dynamic threshold, the optimization process is triggered. The parameters of the first neural network are updated through the gradient backpropagation algorithm, and the parameters of the second neural network are updated through the policy gradient method.

[0013] Optionally, the system feedback and optimization module also includes a habit drift detection submodule, which is used to identify long-term changing trends in user habits based on long-term collected habit pattern vectors, and adjust the network structure parameters of the first neural network through an online learning mechanism.

[0014] Optionally, the system also includes a cross-device collaborative management module, which coordinates the control timing and resource allocation among multiple devices based on habitual pattern vectors and environmental context feature vectors, and generates a global device collaborative strategy.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: By constructing a multimodal deep behavioral model, it comprehensively collects and processes user interaction data with smart home devices, extracts features from multiple dimensions including time series, frequency domain, and semantics, and performs deep modeling after fusion using an attention mechanism. This outputs a habit pattern vector containing habit strength, stability, and trigger condition probability, characterizing user habit features and providing a personalized and detailed basis for subsequent control strategy generation, aligning with users' daily behavior patterns. Furthermore, by collecting environmental sensor data, it constructs a dynamic environmental state map and generates environmental context feature vectors. It performs time series alignment and sliding window segmentation on the data, and utilizes a feature extraction network cascaded with a one-dimensional convolutional neural network and gated recurrent units to effectively extract spatial and temporal features, thereby enabling real-time perception of environmental changes and allowing control strategies to dynamically adjust according to the environment, improving the system's adaptability to different environments. By concatenating habitual pattern vectors with environmental context feature vectors, a preliminary set of device control actions is output through a pre-trained neural network. This set is then combined with real-time device status and optimized using reinforcement learning algorithms to generate dynamic device control strategies and strategy mapping tables. This approach considers both user habits and environmental and device status, achieving intelligent and dynamic control decisions. Control commands are obtained by parsing the strategy mapping table, generating underlying control signals and injecting random delays to avoid conflicts from simultaneous control of multiple devices. Feedback loops verify device status and dynamically adjust control strength to ensure accurate execution and stable operation. Real-time collection of feedback data evaluates performance, updates neural network parameters, identifies habit changes, and adjusts the network structure to ensure continuous system optimization. By coordinating the control timing and resource allocation of multiple devices, a global collaborative strategy is generated to avoid inter-device conflicts and improve overall operational efficiency and user experience. Attached Figure Description

[0016] Figure 1 This is a block diagram of a smart home adaptive control system based on user habits proposed in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Reference Figure 1 A smart home adaptive control system based on user habits, comprising a multimodal habit modeling module, an environmental context perception module, an adaptive control strategy generation module, a device control execution module, and a system feedback and optimization module; The multimodal habit modeling module is used to construct a multimodal deep behavioral model of user habits and output a habit pattern vector that represents the inherent laws of user behavior. The habit pattern vector includes at least habit strength, habit stability and habit triggering condition probability. The environmental context awareness module is used to generate environmental context feature vectors that represent the current environmental state and construct a dynamic environmental state map, which is used to represent the real-time correlation and weight between various environmental factors. The adaptive control strategy generation module is used to generate dynamic device control strategies based on habitual pattern vectors and environmental context feature vectors. The device control execution module is used to control the operating status of one or more smart home devices according to a dynamic device control strategy. The system feedback and optimization module is used to monitor user feedback signals and device status, and optimize the multimodal deep behavior model and dynamic device control strategy based on the feedback data.

[0019] It should be further explained that, in the specific implementation process, the multimodal habit modeling module constructs a multimodal deep behavioral model and outputs habit pattern vectors, including the following steps: S21. Collect user interaction data with smart home devices, perform data cleaning and standardization, and output the cleaned interaction data. The user interaction data with smart home devices includes device operation sequences, operation duration, operation frequency, voice command text, and user location trajectory. The multimodal habit modeling module also integrates a privacy protection sub-module, which is used to desensitize and homomorphically encrypt the original interaction data before data cleaning and standardization, and add noise conforming to the Laplace distribution through differential privacy technology to protect user identity information. S22. Perform multimodal feature extraction on the cleaned interactive data, and extract its temporal feature vector, frequency domain feature vector and semantic feature vector respectively; S23. Based on the extracted temporal feature vector, frequency domain feature vector and semantic feature vector, an attention mechanism is used to perform weighted fusion, and the fused features are input into the first neural network composed of a long short-term memory network and a Transformer encoder for deep behavior modeling to generate an implicit representation of user habits. In S23, the process of weighted fusion using an attention mechanism includes: S231. Input the temporal feature vector, frequency domain feature vector and semantic feature vector into the multi-head attention layer in parallel; S232. Based on the current time context, the multi-head attention layer calculates an attention score for each feature vector, where the attention score is used to reflect the importance of the feature modality to habit modeling in the current time context. S233. Based on the calculated attention score, the temporal feature vector, frequency domain feature vector and semantic feature vector are weighted and summed to generate the fused feature vector. S24. Based on implicit representation, dimension mapping and normalization are performed through a fully connected layer to output a habit pattern vector, where the vector dimension includes habit strength, habit stability and habit triggering condition probability. Specifically, time-series feature vector extraction includes: constructing an event sequence from user operation events (such as turning on the main living room light) and their precise timestamps; calculating the mean and variance of the time intervals between consecutive operations; and using Fast Fourier Transform to analyze the spectral characteristics of this event sequence within a 24-hour period to identify periodic patterns such as concentrated turning on bedroom devices at 7:00 AM and concentrated turning off living room devices at 10:00 PM. Frequency-domain feature vector extraction includes: treating the hourly operation event count sequence as a discrete signal; and using Hanning window-weighted Short-Time Fourier Transform to analyze its performance at different time periods (such as working hours). The spectrum energy distribution of the user's activity (during peak and leisure time periods) is analyzed, and its dominant frequency component and spectrum roll-off coefficient are extracted to quantify the rhythmic intensity and burst patterns of user activity. Semantic feature vector extraction includes: for unstructured voice commands from users (such as adjusting the ambient light to a warm yellow and to the darkest setting), a BERT model finely tuned with a large-scale Chinese smart home command corpus is used to generate context-related word vectors. These high-dimensional vectors are then mapped to fixed-length semantic feature vectors through a fully connected layer (semantic encoding layer) designed specifically for this task. The semantic feature vectors can understand the device object, operation attributes, and intensity modifiers in the command. A weighted fusion of attention is achieved using a gated multi-head attention mechanism. Extracted temporal, frequency, and semantic feature vectors are used as inputs. The gating mechanism uses the current time of day and day of the week as prior conditions, mapping them into dense vectors through a learnable embedding layer and labeling them as temporal gating vectors. These temporal gating vectors are used to modulate the calculation of attention weights. The temporal gating vectors are then multiplied by the temporal, frequency, and semantic feature vectors to generate an initial set of gating biases. These biases are added to the standard attention score (i.e., the dot product of the query and key) and normalized using the Softmax function to obtain a time-modulated attention weight distribution. For example, during the sleep period from 10 PM to 6 AM, the temporal gating vectors result in higher final attention weights for temporal and semantic feature vectors related to turning off lights and adjusting the air conditioner to a sleep temperature, while suppressing the weights of frequency feature vectors reflecting daytime activity. The fused feature vector obtained after weighted fusion is input into the first neural network. The first neural network adopts a collaborative modeling architecture: the Long Short-Term Memory network processes the weighted fused features in a sequential manner, captures the short-term progressive relationship of user habits (such as turning off the TV before turning off the lights), and outputs a hidden state sequence containing all time steps. The Transformer encoder performs temporal self-attention calculation on the hidden state sequence. The calculation method is as follows: the entire hidden state sequence is used as input, each hidden state in the sequence is used as a query, and attention operation is performed with all hidden states in the sequence (as keys and values) to capture the long-term dependency relationship between behaviors at different time points (such as identifying the fundamental difference in device operation combinations between weekend movie-watching habits and weekday office work habits). Finally, the final hidden state of the Long Short-Term Memory network is concatenated with the [CLS] label vector output by the Transformer encoder to form an implicit representation that comprehensively represents the user's potential behavioral intentions. The fully connected layer takes the generated implicit representation as input and is a three-layer neural network structure. The first layer maps the high-dimensional implicit representation to 128 dimensions and introduces non-linearity using the ReLU activation function. The second layer compresses it to 32 dimensions to extract more abstract core features of the habit. The third layer finally maps to three specific output nodes and outputs a final scalar using different normalization functions: the first output node uses the Sigmoid function to output the habit strength between 0 and 1, which is obtained by quantifying the frequency and consistency of the user's execution of the habit sequence in similar scene types throughout history; the second output... The node calculates the Euclidean distance variance in the feature space of the implicit representation corresponding to the habit in the most recent 30 occurrences, takes its reciprocal, and normalizes it using the Sigmoid function to output habit stability. The higher the value, the more stubborn and difficult to change the habit. The third output node uses the Softmax function to output a probability distribution vector. This vector represents the conditional probability of the habit being triggered under various preset scenario types (such as working from home, meeting guests, watching movies, and sleeping). This probability is calculated by the inner product of the implicit representation and the prototype vector of each scenario type (a fixed vector used to represent the essential features of each scenario).

[0020] It should be further explained that, in the specific implementation process, the environmental context awareness module generates environmental context feature vectors and constructs a dynamic environmental state map, including: S31. Collect environmental sensor data, perform data cleaning and standardization, and output the cleaned environmental sensor data; the environmental sensor data includes indoor temperature, humidity, light intensity, sound decibels, and human infrared signals. S32. Perform time series alignment and sliding window segmentation on the cleaned environmental sensor data to form environmental sensor data sequence blocks; S33. Based on the data sequence blocks of environmental sensors, their spatial and temporal features are extracted through a feature extraction network; wherein, the feature extraction network is composed of a one-dimensional convolutional neural network and a gated recurrent unit connected in series, the one-dimensional convolutional neural network is used to extract local spatial patterns, and the gated recurrent unit is used to capture long-term dependencies. S34. Based on the spatial and temporal features output by the feature extraction network, generate dynamic environmental context feature vectors and construct an environmental state map to characterize the correlation and weight between various environmental factors. Specifically, time series alignment includes: handling heterogeneous sensor data with varying sampling rates (from 1Hz to 0.1Hz) for temperature, humidity, light intensity, sound decibels, and human infrared signals; creating a globally equidistant time axis based on the sensor with the highest sampling rate (e.g., a light sensor, 1Hz); establishing a unified time coordinate sequence with fixed time intervals based on the highest data sampling frequency in the system, and aligning all other low-frequency sampled data to this time axis using an interpolation algorithm, thereby resolving the timestamp misalignment problem caused by different sampling rates in multi-source sensor data; and for sensor data with lower sampling rates (e.g., a temperature sensor, 0.1Hz),... The system uses a timestamp-based cubic spline interpolation algorithm to estimate smooth and continuous data values ​​at each point on the global time axis (0.1Hz), ensuring that all environmental variables have synchronized observations at any given time. A variable-length overlapping window mechanism is used for sliding window segmentation, with the window length adaptively adjusted according to the type of smart home device served: a short window of 5 minutes with a step size of 30 seconds is used for lighting control with high response requirements; a long window of 30 minutes with a step size of 3 minutes is used for air conditioning control with higher inertia. The aligned multi-dimensional sensor data matrix arranged in chronological order within each window constitutes the environmental sensor data sequence block. The feature extraction network takes the acquired environmental sensor data sequence blocks as input; a one-dimensional convolutional neural network performs convolutions in the one-dimensional time domain, with its kernel width set to cover multiple consecutive time steps (e.g., 5 time steps) to capture the local coupling and cooperative change patterns of different environmental variables within a short time window; for example, a specific convolutional kernel can learn to detect a typical local spatial pattern of indoor people preparing to rest, namely, a continuous decrease in illuminance within one minute accompanied by a decrease in sound decibels; the output of the one-dimensional convolutional neural network is fed into a gated recurrent unit, which adaptively memorizes or forgets historical information through its update and reset gates, thereby learning and capturing the macroscopic trends and long-term dependencies in the environmental data across the entire sliding window; for example, it learns to identify that from morning to noon, although there are small fluctuations in illuminance and temperature, the overall trend is a steady increase; the hidden state of the gated recurrent unit at the last time step is defined as a comprehensive environmental feature vector that integrates local spatial patterns and global temporal dependencies; The generation of environmental context feature vectors includes: inputting the comprehensive environmental feature vector output by the gated recurrent unit into the encoder part of the variational autoencoder; the encoder maps it into a low-dimensional latent vector that follows a Gaussian distribution, and the mean vector of this vector is used as a distributed and dense representation of the complex state of the environment, i.e., the environmental context feature vector; constructing an environmental state map based on a dynamic graph neural network method; the graph nodes in the environmental state map are predefined as five environmental factors: temperature, humidity, light, sound, and human infrared; the initial edge weights between graph nodes are determined by calculating the Pearson correlation coefficient matrix of each factor on historical data; the system periodically (e.g., every 5 minutes) inputs the current environmental context feature vector into a graph attention network, dynamically updates the feature embedding of each graph node, and recalculates the edge weights between graph nodes using the attention mechanism, thereby forming an environmental state map that reflects the dynamic causal and correlation relationships between environmental factors in real time; for example, when a person is detected at home, the edge weights between the human infrared graph node and the sound graph node in the environmental state map will be enhanced, while the weights with the temperature graph node may be weakened.

[0021] It should be further explained that, in the specific implementation process, the adaptive control strategy generation module generates a dynamic device control strategy based on habitual pattern vectors and environmental context feature vectors, including the following steps: S41. Concatenate the habitual pattern vector with the environmental context feature vector to form a joint feature vector; S42. Input the joint feature vector into the pre-trained second neural network and output a preliminary set of device control actions; wherein, the second neural network is a deep feedforward neural network; S43. Based on the initial set of device control actions and combined with the real-time device status read from the device status register, the strategy is optimized and the action is selected through the reinforcement learning Actor-Critic algorithm to generate the final dynamic device control strategy. At the same time, a strategy mapping table containing device identifier, control parameters and execution time window is generated. Specifically, the concatenation process of habitual pattern vectors and environmental context feature vectors is a gated cross-fusion mechanism. This mechanism takes the output habitual pattern vector and environmental context feature vector as input; projects the habitual pattern vector and environmental context feature vector through independent learnable linear layers; calculates the element-wise dot product of the two projected vectors and uses the sigmoid function to generate a gate signal vector between 0 and 1; this gate signal is used to control the information flow intensity of the habitual pattern vector and environmental context feature vector during fusion; for example, when the environmental context feature vector strongly indicates a sleep state, the gate signal amplifies the sleep-related dimension in the habitual pattern vector while suppressing the activity-related dimension; finally, the gated and modulated habitual pattern vector is concatenated with the environmental context feature vector to form a joint feature vector that can deeply reflect the user's potential intentions in a specific environment. The second neural network is a deep feedforward neural network with three hidden layers. Its input layer receives a joint feature vector, and each neuron in the output layer corresponds to a set of specific and normalized device control parameters (e.g., the brightness of the main living room light [0,1], the air conditioner set temperature [18,30] degrees Celsius, and the opening and closing degree of the smart curtains [0,1]). This network learns a robust initial mapping from the user-environment joint state space to the device control action space through supervised pre-training on a dataset consisting of historical best decision data. This pre-trained model serves as a high-quality starting point for reinforcement learning, effectively avoiding the generation of a large number of invalid or negative experiences in the initial exploration stage of reinforcement learning. The policy optimization process includes: the Actor network takes the joint feature vector as input and outputs the probability distribution of the initial set of device control actions; the Critic network receives the joint feature vector and the actions sampled by the Actor network, and outputs the Q-value that evaluates the long-term expected reward of the state-action relationship; through interaction with the environment (i.e., the real home space), the resulting state transitions and rewards (the reward function is composed of a weighted average of multiple objectives such as user satisfaction, energy consumption, and equipment wear) are stored in the experience replay pool; during training, mini-batch data is sampled from the pool, the Critic network is updated by minimizing the temporal difference error, and then the Actor network is updated along the direction of increasing the Q-value using the policy gradient method; this process... During the process, the system reads the device status register in real time. If an action conflict is found (e.g., the Actor network suggests turning on the air conditioner for heating, but the device register shows that the air conditioner is already in cooling mode and has not been running for 5 minutes), the probability of that action is set to zero and the device is resampled. Finally, the optimized set of actions, along with the control parameters of the corresponding device's globally unique identifier and a flexible execution time window calculated based on the probability of habitual trigger conditions (e.g., for habits with high trigger probability, execution is performed immediately within 0-2 seconds after the trigger condition is detected; for habits with low probability, execution is performed after a 2-5 second delay for observation and confirmation), are written into a structured strategy mapping table. This table is stored in the system's shared memory for the device control execution module to read in real time.

[0022] It should be further explained that, in the specific implementation process, the process by which the device control execution module controls the operating status of one or more smart home devices according to the dynamic device control strategy includes: S51. Parse the policy mapping table to obtain the control commands for the target device; S52. Based on the control commands of the target device, generate low-level control signals through the device communication protocol stack, and inject random delays when distributing the low-level control signals to avoid conflicts between multiple device commands. In S52, the process of injecting random delays during the distribution of low-level control signals includes: S521. For each underlying control signal data frame to be distributed, determine the reference delay parameter according to its corresponding device type and instruction priority; S522. Based on the baseline delay parameter, a random delay value is independently sampled for each data frame from the preset probability distribution; where the probability distribution refers to the exponential distribution. The exponential distribution is chosen because its memoryless performance most effectively simulates the arrival of random events. It can generate a random delay value with a mean of the average processing time of the instruction for each data frame, thereby maximally dispersing dense signals on the timeline and fundamentally avoiding collisions. S523. Append the sampled random delay value to the corresponding data frame, and put the data frame into the physical transmission queue according to the delayed time point. S53. After the bottom control signal is issued, the device status is verified in real time through the device status feedback loop to see if the device status matches the expected control effect. If there is a deviation, the control intensity is dynamically adjusted through the soft actuator. Specifically, the policy mapping table parsing process is a real-time task with priority scheduling and conflict re-detection; the device control execution module periodically (e.g., every 100 milliseconds) reads the policy mapping table; for each entry in the table, the system checks its execution time window, and if the current system time falls within that time window, the entry is marked as pending execution; the device identifier is extracted from the device identifier field of the pending execution entry, and based on the device identifier, the corresponding device driver abstraction layer is loaded from the device protocol library; the device protocol library is a database or knowledge base that stores various smart home device communication protocol rules and data formats, equivalent to a communication dictionary, providing the system with... The system provides syntax and vocabulary for communicating with devices of different brands and models (such as Zigbee lights, Wi-Fi air conditioners, and Bluetooth door locks). The device driver abstraction layer is a software middleware layer located between the core system and specific device protocols. It translates the unified control commands generated by the system into low-level control signals that the corresponding devices in the device protocol library can understand, thereby shielding the differences between different device protocols and realizing unified control of heterogeneous devices. The driver abstraction layer converts the standardized control parameters (such as brightness 0.75) in the policy mapping table into complete low-level command objects with device address, cluster ID, attribute ID, and specific values ​​required by the device's specific communication protocol (such as Zigbee's ZCL command or Wi-Fi's HTTP API call). Before the final delivery, the system queries the device status register again to perform a final conflict security check. The generation and distribution process aims to solve the signaling storm problem in smart home IoT. The device control execution module sends the parsed complete low-level instruction object to the multi-protocol adaptation layer, which is responsible for serializing the instruction object into raw data frames that can be transmitted over physical networks (such as Zigbee 3.0, Wi-Fi, Bluetooth Mesh). Before placing the data frame into the transmission queue, the system independently generates random, millisecond-level transmission delays for all instructions in the same batch, based on their device type and instruction priority, from an exponential distribution with the average instruction processing time as the expected value. This mechanism ensures that even if the policy mapping table requires multiple devices to act simultaneously, their physical control signals can be naturally dispersed in time, fundamentally avoiding packet collisions and network congestion on the wireless channel. The verification and adjustment process constitutes a rapid closed-loop control. After issuing the underlying control signal, the system immediately activates a feedback listener for the device. This listener obtains the actual status of the device by subscribing to the MQTT topic for device status updates or by actively polling the device gateway. After obtaining the actual status, it compares it with the expected status in the policy mapping table and calculates the relative error. If the error exceeds a preset silent threshold (e.g., brightness error > 5%), a soft actuator is triggered. The silent threshold is a fixed error tolerance used to determine whether the deviation between the actual operating status of the device and the expected control effect is significant enough to require immediate intervention. The control strength is dynamically adjusted, specifically reflected in the magnitude of the incremental adjustment command generated by the soft actuator. The soft actuator uses a proportional-integral controller to calculate a progressive adjustment command (e.g., a slight adjustment of +8% based on the current brightness) based on the current error and the accumulated error, and then issues it. This process repeats until the device status enters the expected tolerance range or reaches the maximum number of adjustments, thereby achieving smooth and seamless device control.

[0023] It should be further explained that, in the specific implementation process, the system feedback and optimization module monitors user feedback signals and device status, and optimizes the multimodal deep behavior model and dynamic device control strategy based on the feedback data. This process includes: S61. Real-time collection of explicit user feedback data and implicit behavior correction data, combined with actual device status data obtained from the device control execution module, to calculate user satisfaction index and system performance index. S62. Construct a comprehensive system performance evaluation score based on user satisfaction indicators and system performance indicators; S63. When the overall system performance evaluation score is lower than the preset dynamic threshold, the optimization process is triggered. The parameters of the first neural network are updated through the gradient backpropagation algorithm, and the parameters of the second neural network are updated through the policy gradient method. The dynamic threshold is a variable standard that is calculated in real time based on historical performance data (such as moving average and standard deviation) and is used to determine whether the overall system performance has dropped to the critical point that requires triggering the global model optimization process. Specifically, explicit user feedback data includes users' five-star ratings and text reviews of automated scenarios on the mobile app. The text reviews are quantified into sentiment scores using a pre-set sentiment analysis model. Implicit behavior correction data refers to users' actions of manually overriding system commands within a short period (e.g., within 2 minutes) after the system issues a control command. The intensity of these actions is determined by the speed and magnitude of the overriding operation. System performance indicators include device energy consumption per unit time, average command response time, and command failure rate due to device conflicts. Finally, the user satisfaction indicator is the weighted sum of explicit feedback scores and implicit behavior correction negative scores, and the system performance indicator is the normalized weighted product of all technical indicators. The overall system performance evaluation score is a scalar between 0 and 1, calculated from the user satisfaction index and the system performance index using a learnable weighted geometric mean. The formula is as follows: ; In the formula, This is a user satisfaction index, and its value range is [0,1]. The system performance index has a value range of [0,1]. α and β are weighted indices greater than zero, used to adjust the relative importance of user satisfaction and system performance in the comprehensive evaluation. The specific values ​​of α and β are determined by maximizing the correlation coefficient between the comprehensive system performance evaluation score and the expert rating sequence on the validation set. When the value of α is greater than the value of β, it indicates that the optimization strategy focuses more on improving user experience; conversely, when the value of β is greater than the value of α, it indicates that the optimization strategy focuses more on improving system energy efficiency and stability. A preset dynamic threshold is used, and it is dynamically adjusted based on the moving average and standard deviation of historical performance data to adapt to long-term drift in system performance. When the optimization process is triggered, the system samples batch data from recent interaction experience. For the first neural network, the interaction sequence and corresponding user satisfaction signal in the sampled data are used to calculate the gradient through the time series backpropagation algorithm, with the goal of minimizing the prediction error of user dissatisfaction, and the network weights are updated. For the second neural network (and its Actor-Critic structure), the state-action-reward sequence in the sampled data is used to update the Actor network through the policy gradient algorithm to maximize the expected cumulative reward, and the Critic network is updated by minimizing the time series difference error. This dual-path optimization mechanism ensures that the system can continuously learn from user feedback, while optimizing its habit recognition accuracy and control strategy effectiveness.

[0024] It should be further explained that, in the specific implementation process, the system feedback and optimization module also includes a habit drift detection submodule, which is used to identify the long-term changing trend of user habits based on the habit pattern vectors collected over a long period of time through cluster analysis, and to adjust the network structure parameters of the first neural network through an online learning mechanism;

[0025] Specifically, the habit drift detection submodule continuously stores the habit pattern vectors output by the multimodal habit modeling module and timestamps them, forming a time-sorted sequence of habit pattern vectors. Based on this long-term collected vector sequence, a K-means-based clustering analysis method is used: on a monthly cycle, all habit pattern vectors of the current month are merged with the vector set of the previous month, and cluster analysis is performed. The optimal number of clusters K is determined by the silhouette coefficient. By comparing the positional shift of cluster centers in adjacent cycles with the changes in the number of samples in each cluster, the long-term trend of user habits is identified. For example, the disappearance of an old late-night habit cluster and the formation of a new early-morning habit cluster are identified. When the average moving distance of the cluster centers exceeds the preset drift threshold, habit drift is determined to have occurred. The drift threshold is a preset distance threshold used to quantitatively determine whether user habits have undergone a sufficiently large substantial change. At this time, based on the determination result, the system triggers an online learning mechanism: fine-tuning the first neural network using recent (e.g., the last week) interaction data. If the performance improvement of the model on the validation set after fine-tuning is still lower than expected, a network structure search is initiated, dynamically adding an attention layer to the Transformer encoder in the first neural network to adapt to the increasingly complex long-term dependencies of habit patterns, thus completing the adjustment of the network structure parameters.

[0026] It should be further explained that, in the specific implementation process, the system also includes a cross-device collaborative management module, which is used to coordinate the control timing and resource allocation among multiple devices based on habitual pattern vectors and environmental context feature vectors, and generate a global device collaborative strategy to avoid device conflicts and resource competition. Specifically, the cross-device collaborative management module uses the habitual pattern vector output by the adaptive control strategy generation module and the environmental context feature vector as joint inputs. Internally, the cross-device collaborative management module maintains a device resource conflict graph, which defines potential conflict relationships between devices (e.g., a high-powered desk lamp and a high-definition projector cannot be turned on simultaneously due to visual conflict) and resource competition relationships (e.g., an air conditioner and an electric heater competing for high-power circuit resources). Based on the input joint feature vector, the cross-device collaborative management module predicts all device actions that may be triggered within a future time window (e.g., the next 5 minutes). The predicted device actions are then mapped onto the device resource conflict graph to detect whether conflicts exist. In the event of direct functional conflicts or indirect resource competition, if a conflict is detected, a priority-based time slot allocation algorithm is adopted based on the conflict detection results: a priority is assigned to each device action (the priority is jointly determined by habit strength, device importance, and the urgency of the user's real-time needs), and an execution time window is reserved for high-priority actions, while conflicting low-priority actions are delayed or canceled; finally, a global device coordination strategy is generated, which clearly defines the execution sequence and resource usage limits of all related devices, and is fed back as a constraint to the adaptive control strategy generation module to ensure that the final generated dynamic device control strategy will not cause device conflicts and resource competition.

[0027] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. The focus of each embodiment is on its differences from other embodiments. In particular, the apparatus embodiments are described simply because they are fundamentally based on the method embodiments; relevant details can be found in the descriptions of the method embodiments.

[0028] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0029] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0030] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0031] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0032] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0033] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart home adaptive control system based on user habits, characterized in that: It includes a multimodal habit modeling module, an environmental context awareness module, an adaptive control strategy generation module, an equipment control execution module, and a system feedback and optimization module; The multimodal habit modeling module is used to construct a multimodal deep behavior model and output a habit pattern vector, wherein the habit pattern vector includes at least habit strength, habit stability and habit triggering condition probability; The environment context awareness module is used to generate environment context feature vectors and construct a dynamic environment state map; The adaptive control strategy generation module is used to generate dynamic device control strategies based on habitual pattern vectors and environmental context feature vectors. The device control execution module is used to control the operating status of one or more smart home devices according to a dynamic device control strategy. The system feedback and optimization module is used to monitor user feedback signals and device status, and optimize the multimodal deep behavior model and dynamic device control strategy based on the feedback data.

2. The smart home adaptive control system based on user habits according to claim 1, characterized in that, The process by which the multimodal habit modeling module constructs a multimodal deep behavioral model and outputs habit pattern vectors includes: S21. Collect user interaction data with smart home devices, perform data cleaning and standardization, and output the cleaned interaction data. S22. Perform multimodal feature extraction on the cleaned interactive data, and extract its temporal feature vector, frequency domain feature vector and semantic feature vector respectively; S23. Based on the extracted temporal feature vector, frequency domain feature vector and semantic feature vector, an attention mechanism is used to perform weighted fusion, and the fused features are input into the first neural network composed of a long short-term memory network and a Transformer encoder for deep behavior modeling to generate an implicit representation of user habits. S24. Based on implicit representation, dimension mapping and normalization are performed through a fully connected layer to output a habit pattern vector, where the vector dimension includes habit strength, habit stability and habit triggering condition probability.

3. The smart home adaptive control system based on user habits according to claim 2, characterized in that, The process of weighted fusion using the attention mechanism in S23 includes: S231. Input the temporal feature vector, frequency domain feature vector and semantic feature vector into the multi-head attention layer in parallel; S232. Based on the current time context, the multi-head attention layer calculates an attention score for each feature vector; S233. Based on the calculated attention score, the temporal feature vector, frequency domain feature vector and semantic feature vector are weighted and summed to generate the fused feature vector.

4. The smart home adaptive control system based on user habits according to claim 1, characterized in that, The process by which the environmental context awareness module generates environmental context feature vectors and constructs a dynamic environmental state map includes: S31. Collect environmental sensor data, perform data cleaning and standardization, and output the cleaned environmental sensor data. S32. Perform time series alignment and sliding window segmentation on the cleaned environmental sensor data to form environmental sensor data sequence blocks; S33. Based on the data sequence blocks of environmental sensors, their spatial and temporal features are extracted through a feature extraction network; wherein, the feature extraction network is composed of a one-dimensional convolutional neural network and a gated recurrent unit connected in series, the one-dimensional convolutional neural network is used to extract local spatial patterns, and the gated recurrent unit is used to capture long-term dependencies. S34. Based on the spatial and temporal features output by the feature extraction network, generate dynamic environmental context feature vectors and construct an environmental state map.

5. The smart home adaptive control system based on user habits according to claim 4, characterized in that, The adaptive control strategy generation module generates dynamic device control strategies based on habitual pattern vectors and environmental context feature vectors. The process includes: S41. Concatenate the habitual pattern vector with the environmental context feature vector to form a joint feature vector; S42. Input the joint feature vector into the pre-trained second neural network and output a preliminary set of device control actions; wherein, the second neural network is a deep feedforward neural network; S43. Based on the initial set of device control actions and combined with the real-time device status read from the device status register, the strategy is optimized and the action is selected through the reinforcement learning Actor-Critic algorithm to generate the final dynamic device control strategy. At the same time, a strategy mapping table containing device identifier, control parameters and execution time window is generated.

6. The smart home adaptive control system based on user habits according to claim 5, characterized in that, The process by which the device control execution module controls the operating status of one or more smart home devices according to a dynamic device control strategy includes: S51. Parse the policy mapping table to obtain the control commands for the target device; S52. Based on the control commands of the target device, generate low-level control signals through the device communication protocol stack, and inject random delays when distributing the low-level control signals. S53. After the bottom-level control signal is issued, the device status is verified in real time through the device status feedback loop to see if the device status matches the expected control effect. If there is a deviation, the control intensity is dynamically adjusted through the soft actuator.

7. The smart home adaptive control system based on user habits according to claim 6, characterized in that, The process of injecting random delays during the distribution of low-level control signals in S52 includes: S521. For each underlying control signal data frame to be distributed, determine the reference delay parameter according to its corresponding device type and instruction priority; S522. Based on the baseline delay parameter, independently sample random delay values ​​for each data frame from a preset probability distribution; S523. Append the sampled random delay value to the corresponding data frame, and put the data frame into the physical transmission queue according to the delayed time point.

8. The smart home adaptive control system based on user habits according to claim 1, characterized in that, The system feedback and optimization module monitors user feedback signals and device status, and optimizes the multimodal deep behavior model and dynamic device control strategy based on the feedback data. This process includes: S61. Real-time collection of explicit user feedback data and implicit behavior correction data, combined with actual device status data obtained from the device control execution module, to calculate user satisfaction index and system performance index. S62. Construct a comprehensive system performance evaluation score based on user satisfaction indicators and system performance indicators; S63. When the overall system performance evaluation score is lower than the preset dynamic threshold, the optimization process is triggered. The parameters of the first neural network are updated through the gradient backpropagation algorithm, and the parameters of the second neural network are updated through the policy gradient method.

9. The smart home adaptive control system based on user habits according to claim 1, characterized in that, The system feedback and optimization module also includes a habit drift detection submodule, which is used to identify long-term trends in user habits based on long-term collected habit pattern vectors, and adjust the network structure parameters of the first neural network through an online learning mechanism.

10. A smart home adaptive control system based on user habits according to claim 1, characterized in that, The system also includes a cross-device collaborative management module, which coordinates the control timing and resource allocation among multiple devices based on habitual pattern vectors and environmental context feature vectors, and generates a global device collaborative strategy.

Images