Intelligent central control equipment, environment self-adaptive equipment control method and device, and medium

Abstract

The invention relates to an intelligent central control device, an environment adaptive device control method and device, and a medium. The method comprises the steps of obtaining environment measured data provided by a plurality of online devices which are located in the same environment space and provide different modal data sources in a private network; constructing a single-source state vector corresponding to each data source, and fusing each single-source state vector into a multi-mode state vector representing the overall environment state; inputting the multi-modal state vector into a preset parameter generation model for reasoning to obtain control parameter vectors corresponding to at least part of the online devices in the plurality of online devices; and according to the control parameter vector, constructing equipment control instructions of part of online equipment to control the corresponding online equipment, so that the adjusted overall environment state can optimize the environment evaluation index and the operation energy consumption index. The intelligent equipment can be adaptively controlled according to the environment state of the intelligent home, so that the user experience is improved.

Description

Technical Field

[0001] This application relates to the field of intelligent control, and in particular to a control method, device, and medium for intelligent central control equipment and environmental adaptive equipment. Background Technology

[0002] In the field of smart homes, scene control technology aims to provide users with a convenient and comfortable living experience by coordinating the interaction of multiple devices. Currently, the mainstream smart home scene control methods mainly rely on preset rule mechanisms. This method typically requires users or installers to pre-set the correspondence between specific environmental conditions and device actions, such as automatically turning on lights when indoor light levels fall below a certain fixed value, or setting a series of device operations to be performed at specific times. This control logic based on fixed rules constitutes the technological foundation of traditional smart home automation systems.

[0003] However, since the core of the aforementioned traditional methods lies in executing pre-set static rules, they have significant limitations in practical applications. First, the control strategy is fixed and cannot adapt to dynamic changes in environmental conditions. For example, controlling lights solely based on preset light thresholds fails to comprehensively consider diverse environmental information such as the current time, the presence of the user, and their specific activities. This often results in control outcomes that do not match the user's actual needs, failing to achieve true scene intelligence.

[0004] Furthermore, because the rules are pre-defined, they lack the ability to learn and self-optimize from user interactions. Simply mechanically executing predetermined rules without considering user satisfaction with the control effect—for example, frequent manual adjustments to device parameters by users effectively negate the pre-defined rules—fail to recognize and utilize this feedback to improve subsequent decisions. This keeps the control strategy stuck at the initial configuration level, making it difficult to meet users' personalized and dynamically changing needs.

[0005] Therefore, a core problem inherent in existing technologies is that their static, rule-based control paradigms are unable to adapt to dynamically changing environments or optimize control strategies through interactive experience, resulting in limited intelligence and poor user experience. Summary of the Invention

[0006] The primary objective of this application is to solve at least one of the above-mentioned problems by providing a control method, device, and medium for intelligent central control equipment and environmental adaptive equipment.

[0007] To achieve the various objectives of this application, the following technical solution is adopted: An environment-adaptive device control method provided for one of the purposes of this application includes the following steps: Acquire environmental measurement data from multiple online devices in the same environmental space within a private network that provide different modal data sources; A single-source state vector is constructed for each data source, and the single-source state vectors are merged into a multimodal state vector that represents the overall environmental state. The multimodal state vector contains semantic information for inferring the user's activity state. The multimodal state vector is input into a preset parameter generation model for inference to obtain control parameter vectors corresponding to at least some of the online devices among the multiple online devices. The parameter generation model is trained in advance by multi-objective reinforcement learning using a reward function constructed with environmental assessment indicators and operating energy consumption indicators. Based on the control parameter vector, device control instructions are constructed for the online devices to control the corresponding online devices, so that the adjusted overall environmental state can optimize the environmental evaluation indicators and operating energy consumption indicators.

[0008] An environment-adaptive device control apparatus, proposed to meet one of the objectives of this application, comprises: The data acquisition module is configured to acquire environmental measurement data provided by multiple online devices in the same environmental space within a private network that provide different modal data sources. The vector construction module is configured to construct a single-source state vector for each data source, and then merge the single-source state vectors into a multimodal state vector that represents the overall environmental state. The multimodal state vector contains semantic information for inferring the user's activity state. The reasoning generation module is configured to input the multimodal state vector into a preset parameter generation model for reasoning to obtain control parameter vectors corresponding to at least some of the online devices among the multiple online devices. The parameter generation model is trained in advance by multi-objective reinforcement learning using a reward function constructed with environmental evaluation indicators and operating energy consumption indicators. The equipment control module is configured to construct equipment control instructions for some online devices based on the control parameter vector, and control the corresponding online devices so that the adjusted overall environmental state can optimize the environmental evaluation indicators and operating energy consumption indicators.

[0009] On another front, an intelligent central control device provided for one of the purposes of this application includes a controller, the controller including a processor and a memory, the processor calling and running a computer program in the memory to perform the steps of the environmental adaptive device control method.

[0010] In another aspect, a computer-readable storage medium is provided to suit another purpose of this application, which stores, in the form of computer-readable instructions, a computer program implemented according to the described environment-adaptive device control method, which, when invoked by a computer, executes the steps included in the corresponding method.

[0011] Compared to traditional technologies, this application acquires and integrates measured multimodal environmental data within a private network to construct a multimodal state vector containing semantic information. Then, it utilizes a parameter generation model trained through multi-objective reinforcement learning for reasoning and decision-making, ultimately generating adaptive control commands. This enables comprehensive and dynamic perception and understanding of the environmental state, rather than relying on pre-set fixed rules. The reinforcement learning-based parameter generation model optimizes the control strategy based on a reward function constructed from environmental evaluation indicators and operational energy consumption indicators. This allows for dynamic adaptation to environmental changes and user needs, significantly improving the intelligence and accuracy of control and continuously enhancing the user experience over long-term use. Attached Figure Description

[0012] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a typical embodiment of the environmental adaptive device control method of this application; Figure 2 This is a schematic block diagram of the environmental adaptive device control device of this application; Figure 3 This is a schematic diagram of the structure of a computer device used in this application. Detailed Implementation

[0013] The environmental adaptive device control method claimed in this application can be deployed in a smart home private network environment centered on a smart home hub, home gateway, or edge server. This private network constructs a localized communication and control domain, which is widely connected to a wide variety of smart devices and sensing devices, collectively forming the infrastructure for the system to perceive the environment and perform control.

[0014] Specifically, these devices include, but are not limited to, temperature and humidity sensors, light sensors, air quality sensors, and noise sensors for collecting physical environmental data; visual acquisition devices such as infrared cameras and depth cameras for sensing user presence, identity, location, and behavior; and audio acquisition devices such as microphone arrays; as well as various controlled execution devices, such as smart lighting devices, smart air conditioners, fresh air systems, humidifiers, smart lamps, smart curtains, smart speakers, smart TVs, smart sockets, and smart door locks.

[0015] In the aforementioned network environment, the intelligent central control device, which performs core processing functions, is the key carrier for implementing the solution. Physically, this device can manifest as a smart home hub, a high-performance gateway, or a dedicated server, responsible for running the computer program implemented according to the method of this application. Functionally, the intelligent central control device maintains connectivity with all access devices through multiple communication protocols supported within the private network, continuously aggregating environmental measurement data from sensing devices and current operating status data from execution devices, thereby constructing a comprehensive and real-time perception of the overall environmental state. Simultaneously, the intelligent central control device incorporates a pre-trained parameter generation model, which can automatically generate optimized control parameters for one or more execution devices based on the fused multimodal environmental state information.

[0016] It should be noted that the environmental measurement data in this application is a comprehensive concept, with diverse sources and forms. It includes physical quantities directly measured by sensors, such as temperature, humidity, and light intensity, as well as semantic information extracted from the raw data collected by devices through intelligent analysis, such as user activity types identified by cameras or specific commands parsed from audio streams. Various devices play different but collaborative roles in the entire smart home system implemented by the intelligent central control device. Sensing devices are primarily responsible for providing input information describing the environment and user status, while execution devices, as the ultimate control targets, receive commands issued by the intelligent central control device, thereby actually changing the environmental state. This extensive device access and collaborative operation mechanism enables this application to adapt to complex and ever-changing home environments, achieving accurate contextual awareness and adaptive intelligent control.

[0017] Having clarified the aforementioned network architecture, device roles, and basic concepts, the following detailed explanation of the implementation details of the technical solution claimed in this application is provided through several specific embodiments.

[0018] Please see Figure 1 In some embodiments, the environment adaptive device control method of this application can be implemented as an application program running in the processor of an intelligent central control device. The method includes: Step S3100: Obtain environmental measurement data provided by multiple online devices in the same environmental space within the private network that provide different modal data sources; The intelligent central control device initiates its data acquisition function, targeting all online devices within the private network that are pre-labeled as belonging to the same physical or logical environmental space, and begins the process of acquiring environmental measurement data. This environmental space can be a specific room, an open living room area, or an entire residential unit consisting of multiple rooms; its scope can be defined by the geographical location information of the devices or by the user's logical division. Online devices refer to intelligent devices that have successfully connected to the private network, possess communication capabilities, and are currently in normal working order, including both controlled intelligent devices and intelligent devices acting as environmental sensors.

[0019] Intelligent central control devices acquire environmental measurement data from various online devices via one or more communication protocols supported by a private network, such as Wi-Fi, Zigbee, Bluetooth Mesh, or MQTT, establishing connections with each online device and requesting data. These data sources constitute different modalities due to their different physical characteristics and information types. In one embodiment, data acquisition can be achieved by subscribing to device event push messages, proactively reporting changes in device status or sensor readings. In another embodiment, the intelligent central control device periodically sends polling commands to the target device, requesting it to return the current measurement data.

[0020] The environmental measurement data is broad in scope, and its specific form depends on the functionality of the online devices providing the data. For example, temperature and humidity sensors may acquire numerical values ​​representing the physical properties of the environment, smart cameras may acquire pre-processed image sequences or raw video streams, while smart microphones may acquire audio streams or identified audio event features. These data collectively constitute a raw, low-level description of the current environmental state. Upon receiving this heterogeneous data, the intelligent central control device can perform preliminary verification and caching, providing a reliable data foundation for subsequent vector construction and fusion steps.

[0021] Before being used to construct state vectors, the acquired environmental measurement data can undergo a series of preprocessing operations, including but not limited to data cleaning to remove outliers or invalid readings, and timestamp alignment of data from devices with different sampling frequencies to ensure that all data correspond to environmental snapshots within the same time window. Furthermore, for numerical data, normalization or standardization can be performed to eliminate the influence of units, facilitating subsequent model processing.

[0022] Step S3200: Construct a single-source state vector for each data source, and merge the single-source state vectors into a multimodal state vector that represents the overall environmental state. The multimodal state vector contains semantic information for inferring the user's activity state. After acquiring environmental measurement data, the intelligent central control device further transforms the heterogeneous environmental measurement data from different data sources into a unified, semantically rich, multimodal state vector. To this end, feature extraction is first performed on each data source to construct a single-source state vector. For example, numerical data such as temperature, humidity, and light intensity acquired from physical sensors can be converted into numerical feature vectors with the same scale through standardization. For image sequences acquired from visual acquisition devices, in one embodiment, a pre-trained convolutional neural network model can be used to extract feature maps from the images, which are then aggregated into feature vectors representing the visual content. For audio streams, audio feature vectors can be constructed by extracting acoustic features such as Mel-frequency cepstral coefficients or by utilizing event codes output by an audio event detection model.

[0023] Subsequently, the intelligent central control device fuses the multiple single-source state vectors constructed separately. The purpose of fusion is to integrate information from different modalities to form a more comprehensive and accurate representation of the overall environmental state. In one embodiment, a simple vector concatenation operation can be used to connect the individual single-source state vectors into a long vector. In another, more preferred embodiment, an attention-based fusion model is used. This model can dynamically calculate the importance weights of different modal features in a specific environmental context and perform weighted summation or more complex interactive calculations on them, thereby highlighting key information and suppressing noise interference. Through this fusion, the final multimodal state vector not only contains low-level physical environment parameters but also contains high-level semantic information for inferring the user's activity state. For example, by combining the user's location, posture, and audio commands, it is possible to infer that the user is currently in an activity state such as watching a movie, eating, or resting.

[0024] It should be noted that the feature extraction method used to construct the single-source state vector and the specific architecture of the fusion model can be selected and designed according to the accuracy and efficiency requirements of the actual application. For example, on edge devices with limited computing resources, a lightweight neural network model or feature extractor can be chosen. On the server side, a deeper network model can be used to obtain richer feature representations. The final generated multimodal state vector will serve as the direct input for subsequent parameter generation models to perform inference and decision-making.

[0025] Step S3300: Input the multimodal state vector into a preset parameter generation model for inference to obtain control parameter vectors corresponding to at least some of the online devices among the multiple online devices. The parameter generation model is trained in advance by multi-objective reinforcement learning using a reward function constructed with environmental evaluation indicators and operating energy consumption indicators. The parameter generation model used in this application is essentially a decision function trained through multi-objective reinforcement learning, which internally encapsulates a complex mapping relationship between historical environmental states and optimal equipment control parameters. During the training phase, the model uses a reward function jointly constructed from environmental evaluation indicators and operational energy consumption indicators as the optimization objective. Through interactive learning with a large amount of historical data, it can automatically output an adapted control strategy based on the current comprehensive environmental state.

[0026] The parameter generation model can be implemented in various forms. In one embodiment, the model can be a deep Q-network, whose output layer corresponds to the Q-values ​​of all possible control actions, selecting the action with the highest Q-value as the current policy. In another embodiment, the model can be a policy gradient network, such as a model trained by a proximal policy optimization algorithm, which directly outputs the probability distribution of control parameters and determines the final action to be executed through sampling. The model inference process is the process of mapping a high-dimensional multimodal state vector into a simplified control parameter vector, where each dimension of the control parameter vector corresponds to a specific parameter setting of one or more controlled online devices, such as the brightness and color temperature values ​​of smart lights, or the target temperature and fan speed level of an air conditioner.

[0027] The training of the parameter generation model relies on a pre-built training dataset and reward function. The training process teaches the model to balance two potentially conflicting objectives: pursuing environmental comfort (quantified by environmental evaluation metrics) and reducing energy consumption (quantified by operational energy consumption metrics), seeking a Pareto optimal solution. This allows the model to dynamically generate optimal control parameters for the current situation, rather than relying on pre-defined fixed rules when facing complex and changing environments.

[0028] The parameter generation model takes a multimodal state vector as input and, after inference, generates a control parameter vector. This vector is not directly an executable instruction of the device, but an abstract, normalized set of parameters. After subsequent transformation processing, it can be further used to generate device control instructions.

[0029] Step S3400: Construct device control instructions for the online devices based on the control parameter vector to control the corresponding online devices, so that the adjusted overall environmental state can optimize the environmental evaluation indicators and operating energy consumption indicators.

[0030] The intelligent central control device further converts the control parameter vector obtained from the parameter generation model into specific executable control commands, and sends them to the corresponding online devices in the private network for execution. The dimensions in the control parameter vector are related to the device to be controlled and its adjustable parameters, such as normalized values ​​representing the percentage of light brightness, the set temperature of the air conditioner, and the volume level of the audio system. The intelligent central control device needs to convert these normalized parameter values ​​into a command format that can be recognized and executed by the specific device, based on a pre-established or dynamically negotiated device control protocol.

[0031] In one embodiment, the conversion process relies on a unified device capability model maintained by device adaptive middleware. This model stores the set of control commands supported by each registered device and parameter mapping rules. For example, for the same brightness parameter, the device capability model knows that device A supports linear adjustment from 0-100%, while device B may only support three levels: "low," "medium," and "high." In this case, the middleware will map the continuous brightness percentage to the nearest discrete level based on the capability description of device B. The converted control commands are sent to the target device via network protocols supported by the private network, such as MQTT topic publishing, CoAP requests, or proprietary TCP commands.

[0032] Control commands can be issued synchronously or asynchronously. In one embodiment, the intelligent central control device sends commands to each target device sequentially and waits for confirmation before sending the next command, ensuring the reliability and sequence of control. In another embodiment, to improve response speed, the intelligent central control device can send commands to all target devices in parallel and monitor the execution status feedback of each device. After a device executes a command, the change in its operating state directly affects the physical environment, thereby altering the state of the surrounding space.

[0033] By sending control commands to control the operation of equipment based on control parameter vectors, the adjusted overall environmental state can evolve towards optimizing environmental evaluation indicators and operational energy consumption indicators. For example, when the parameter generation model infers that a user is engaged in movie-watching activities through multimodal state vectors, the model may output parameter vectors to reduce the brightness of the main lights, turn on the backlight strips, and adjust the air conditioner to silent mode. After the corresponding online devices execute the corresponding device control commands, the ambient light, temperature, and noise levels change. These changes are continuously monitored by sensors and fed back as new environmental measurement data, forming a closed-loop control loop. The optimization of environmental evaluation indicators is reflected in the adjusted environment better meeting the inferred user activity needs, such as the immersive experience required for movie watching. The optimization of operational energy consumption indicators is reflected in avoiding unnecessary energy consumption through precise control while meeting comfort requirements, such as automatically reducing lighting or turning off air conditioning in unoccupied areas.

[0034] The effectiveness of the entire control process relies on the strategy learned during the training of the parameter generation model. This strategy comprehensively considers both immediate effects and long-term benefits, enabling the cumulative effect of a series of control actions to ultimately achieve the optimal balance between environmental comfort and energy efficiency. The intelligent central control device continuously executes this control cycle to achieve continuous adaptive optimization of the home environment.

[0035] This application constructs an intelligent decision-making system based on multimodal perception and multi-objective reinforcement learning through the above embodiments. Compared with traditional smart home control technology that relies on fixed preset rules, it has achieved significant technological progress. Its advantages are mainly reflected in the following aspects.

[0036] First, this application achieves a comprehensive and deep fusion of environmental state perception and understanding, overcoming the limitations of traditional technologies that rely on only a single or limited data source. It acquires environmental measurement data of different modalities from various online devices widely distributed within a private network; these data sources are complementary in terms of physical attributes and information types. By constructing a single-source state vector and fusing it using advanced algorithms such as attention mechanisms, a multimodal state vector containing high-level semantic information is generated. Based on this, accurate inferences about user activity states and other contextual information can be extracted from raw, low-level sensor readings, providing a much richer and more accurate environmental state representation for subsequent intelligent decision-making than traditional solutions, thus laying the foundation for truly user-intent-aligned control.

[0037] Secondly, the control strategy generation process of this application possesses high adaptability and continuous optimization capabilities. The parameter generation model, trained through multi-objective reinforcement learning, can dynamically generate the optimal control parameter vector based on the current complex environmental state. During the training phase, this model has learned to balance multiple objectives such as environmental comfort and energy efficiency; therefore, its decisions are not based on static rules but can adapt to dynamic environmental changes and potential user needs. This solves the fundamental shortcomings of traditional technologies, such as fixed strategies, inability to handle unforeseen scenarios, or failure to optimize long-term performance. Essentially, it possesses the ability to learn and improve its behavior from interactions with the environment.

[0038] Furthermore, this application significantly improves overall efficiency and user experience by forming an automated closed loop from perception and decision-making to execution. By translating control parameter vectors into specific device commands and executing them, the physical environment is altered. The changed environmental state is then continuously monitored by the system and fed back as new input. This closed-loop mechanism allows the system to verify its control effectiveness and provides data support for further model optimization. The entire process requires no manual user intervention in configuring scene rules, achieving full automation from environmental perception to device control. This not only lowers the user's barrier to entry and maintenance costs but also improves environmental comfort and optimizes energy consumption through precise and timely control.

[0039] Based on any embodiment of the method in this application, environmental measurement data provided by multiple online devices in the same environmental space within a private network and providing different modal data sources are obtained, including: Step S3110: Obtain physical environment measured data from the physical sensor device deployed in the environmental space as the first environmental measured data. The physical environment measured data includes any multiple of temperature, humidity, light intensity and noise decibel value. Intelligent central control equipment acquires measured physical environmental data from physical sensor devices deployed in the environment, serving as one of the fundamental inputs for constructing an environmental state model. Here, physical sensor devices specifically refer to environmental sensing devices capable of directly measuring environmental physical parameters, including but not limited to temperature sensors, humidity sensors, light sensors, noise sensors, air quality sensors (such as PM2.5 and CO2 sensors), and atmospheric pressure sensors. These devices continuously or periodically monitor the basic physical properties of the environment and report the measured values ​​via a private network.

[0040] The acquired environmental measurement data are quantified numerical values ​​that directly reflect the objective physical state of the environment. Examples include Celsius values ​​from temperature sensors, percentage relative humidity from humidity sensors, lux illuminance values ​​from light sensors, and decibel values ​​from noise sensors. These data typically have clearly defined physical units and dimensions. After receiving this raw data, the intelligent central control device can perform necessary preprocessing, such as unit conversion or filtering outliers that clearly exceed reasonable ranges, to ensure the validity and consistency of the data.

[0041] These measured physical environment data constitute a low-level, objective description of the environment and are important bases for judging environmental comfort, assessing energy efficiency, and inferring macro-level user activities (such as indirectly determining whether air conditioning needs to be run based on ambient temperature).

[0042] Step S3120: Obtain an image sequence from the visual acquisition device deployed in the environmental space, and perform real-time analysis on the image sequence to obtain visual perception measurement data as the second environmental measurement data. The visual perception measurement data includes any multiple of the following: user identity, number of people in the room, user location, and user posture. The intelligent central control device acquires image sequences from visual acquisition devices deployed in the environment and analyzes these image sequences in real time to extract visual perception measurement data. Visual acquisition devices include, but are not limited to, intelligent devices capable of capturing visual information such as wide-angle cameras, infrared thermal imagers, and depth-sensing cameras. These devices continuously acquire images at a certain frame rate, forming image sequences, and transmit the image data stream to the intelligent central control device via a private network.

[0043] After receiving the image sequence, the intelligent central control device calls its built-in computer vision analysis module to process the image data. In one embodiment, the processing employs deep learning-based object detection and recognition algorithms, such as YOLO or Faster R-CNN models, to identify the user in the image and determine their location. In another embodiment, human pose estimation algorithms, such as OpenPose, can be used to analyze the user's skeletal key points and then infer the user's posture, such as standing, sitting, lying down, or walking. For user identification, one embodiment can be to use facial recognition technology to match detected facial features with a pre-registered user feature database.

[0044] Through the above analysis, the intelligent central control device ultimately obtains structured visual perception measurement data. This data specifically includes information items such as user identity, number of people in the room, user location, and user posture. User identity information can be represented as a unique user identifier. The number of people in the room is an integer value, directly reflecting the number of people in the current environment. User location information can be represented as coordinates relative to a spatial origin, or simply as the area where the user is located, such as the living room sofa area or next to the dining table. User posture can be represented using predefined posture category labels.

[0045] These visual perception measurement data provide direct evidence about user presence and behavior, serving as a crucial source of high-level semantic information for inferring user intentions and understanding environmental context. For example, recognizing multiple people sitting around a table in a seated posture can help infer that a dining activity is in progress.

[0046] Step S3130: Obtain an audio stream from the audio acquisition device deployed in the environmental space, and perform real-time analysis on the audio stream to obtain audio perception test data as third environmental test data. The audio perception test data includes any multiple items from speech text, environmental noise features, and audio of specific events. The intelligent central control device acquires audio streams from audio acquisition devices deployed in the environment and analyzes these streams in real time to extract measured audio perception data. Audio acquisition devices include, but are not limited to, intelligent devices capable of capturing ambient sound, such as single microphones and microphone arrays. These devices continuously collect audio signals from the environment, forming an audio stream, and transmit the audio data to the intelligent central control device for processing via a private network.

[0047] After receiving the audio stream, the intelligent central control device invokes its built-in audio signal processing and analysis module. In one embodiment, the processing first preprocesses the original audio, including noise reduction, echo cancellation, and audio enhancement, to improve the accuracy of subsequent analysis. Subsequently, the processed audio is analyzed to extract specific information. For acquiring voice information, one embodiment employs automatic speech recognition technology to convert the speech content in the audio into text information, i.e., speech-text. For extracting environmental noise features, another embodiment can characterize the overall noise level and spectral characteristics of the environment by calculating the frequency domain features of the audio signal, such as Mel-frequency cepstral coefficients, or by analyzing the amplitude statistical characteristics of the signal. For detecting audio of specific events, yet another embodiment can apply an audio event detection model. This model, after training, can identify audio events with specific semantics, such as the sound of breaking glass, a doorbell, a baby crying, or the sound of flowing water, and assign corresponding type labels to the identified events.

[0048] Through the above analysis, the intelligent central control device obtains structured audio perception measurement data. This data specifically includes information items such as speech text, environmental noise characteristics, and audio of specific events. Speech text represents the recognized voice commands or dialogue content in string form. Environmental noise characteristics can be represented as a series of numerical feature vectors or a comprehensive decibel level. Audio of specific events can be represented by event type labels and their timestamps.

[0049] These audio perception measurement data provide crucial information about environmental acoustic scenarios and user voice interactions, serving as a valuable data source for understanding user intent and environmental events. For example, recognizing the voice text "dim the lights" can directly serve as explicit input for user control intent. Detecting high ambient noise characteristics may indicate a noisy environment, requiring adjustment of audio equipment volume or activation of noise reduction mode. Detecting specific event audio, such as a doorbell sound, can trigger corresponding security or welcome scenarios.

[0050] Step S3140: Obtain the current operating status data of the online device to be controlled in the private network as the fourth environmental measurement data. The operating status data includes any number of items among switch status, working mode and operating parameters.

[0051] Intelligent central control equipment acquires current operational status data from online devices to be controlled within a private network, serving as a crucial component in constructing a complete environmental status representation. Here, "online devices to be controlled" broadly refers to intelligent devices that act as both environmental control execution terminals and provide their own status feedback, such as intelligent air conditioners, intelligent lighting devices, intelligent curtains, and intelligent speakers. These devices play a dual role in the environment: they are both the end point for executing control commands and the information source reflecting the current baseline operational status of the environment.

[0052] The intelligent central control device acquires operational status data in ways including, but not limited to, two main embodiments. In one embodiment, the intelligent central control device proactively sends a status query request to the target online device to be controlled, and the device returns a snapshot of its current operational status upon receiving the request. In another embodiment, when the online device to be controlled changes its own status, it proactively pushes a status update message to the intelligent central control device, for example, by publishing it to a specific topic via the MQTT protocol. The intelligent central control device receives status changes in real time by subscribing to these topics.

[0053] The acquired operational status data provides a detailed description of the device's real-time working condition. The on / off status indicates whether the device is currently in an on or off mode. The operating mode defines the device's specific operating strategy, such as the cooling, heating, or dehumidifying mode of an air conditioner, or the reading, relaxation, or night light mode of a lamp. Operating parameters provide a more detailed description of the device's specific settings in the current mode, encompassing a wide range of parameters. For example, for a smart air conditioner, operating parameters include the set target temperature, current fan speed, and swing switch status; for smart lamps, operating parameters include the current brightness level, color temperature, and color saturation; and for smart curtains, operating parameters include the current percentage of opening or the specific location.

[0054] These operational status data, serving as the fourth set of environmental measurement data, provide environmental baseline information. This baseline information, along with data from physical sensors, visual acquisition devices, and audio acquisition equipment, constitutes a multi-dimensional description of the environment. For example, understanding the current set temperature and operating mode of the air conditioner, combined with the actual room temperature read from the temperature sensor, allows for a more accurate assessment of the current temperature control effectiveness and the scope for further adjustments. Obtaining the current brightness and color temperature of the lighting fixtures helps determine whether fine-tuning or scene switching is necessary. Therefore, operational status data extends the system's understanding of the environmental state from purely external observation to mastering the state of internal execution units. This provides crucial information for generating coordinated, precise control strategies that conform to the current equipment capacity constraints, preventing conflicts or ineffective operations between control commands and the current equipment state.

[0055] This application, through the above embodiments, achieves systematic collection and integration of four heterogeneous data sources: physical environment, visual perception, audio perception, and device operating status. This results in significant technical advantages in the comprehensiveness of data acquisition, the depth of semantic understanding, and the accuracy of control decisions. Specifically, this embodiment not only constructs an environmental state representation rich in high-level semantic information through multimodal data fusion, providing a richer context for intelligent decision-making than a single data source, but more importantly, by incorporating the operating status of the controlled device itself into the scope of measured environmental data, it can grasp the dynamic baseline of the environment in real time. The resulting control strategy not only responds to changes in the external environment but also seamlessly integrates with the current operating status of the device, effectively avoiding control conflicts and ineffective operations, improving control coordination and system efficiency, and ultimately achieving a more thorough perception of the environment, more intelligent decision-making, and more precise control.

[0056] Based on any embodiment of the method in this application, a single-source state vector is constructed for each data source, and the single-source state vectors are fused into a multimodal state vector representing the overall environmental state, including: Step S3210: Based on the first environmental measured data, construct a low-level feature vector representing the physical environment through standardization processing; After acquiring the environmental measurement data, the intelligent central control device needs to transform this heterogeneous data into a unified, machine-understandable state representation. First, a low-level feature vector is constructed based on the initial environmental measurement data. This initial environmental measurement data, namely numerical data such as temperature, humidity, and light intensity from physical sensor devices, typically has different dimensions and numerical ranges. In one embodiment, these raw data are converted into standard scores with zero mean and unit variance through standardization. In another embodiment, the min-max normalization method is used to linearly scale the data to a fixed interval between zero and one. This standardization process eliminates the influence of dimensions, allowing the values ​​of different physical quantities to be compared and calculated on the same scale, thereby constructing a low-level feature vector characterizing the physical environment. This vector directly reflects the objective physical conditions of the environment.

[0057] Step S3220: Based on the measured data of the second environment and the measured data of the third environment, extract high-level semantic feature vectors representing user activities through a deep learning model; Subsequently, high-level semantic feature vectors are constructed based on the measured data from the second and third environments. The measured data from the second environment includes visual information such as user identity, location, and posture, while the measured data from the third environment includes audio information such as speech text and audio of specific events. These data already possess a certain semantic level, requiring more complex models to extract their deeper features.

[0058] In one embodiment, for visual data, a pre-trained convolutional neural network model, such as ResNet or VGG, is used to extract spatial features of the image. This is then combined with a recurrent neural network, such as LSTM, to process the temporal dynamics of the image sequence, ultimately converging into a visual semantic vector representing the user's activity intent. For audio data, one embodiment utilizes a pre-trained speech recognition model to extract speech embedding features, which are then combined with the output encoding of an audio event detection model to jointly construct an audio semantic vector. The visual and audio semantic vectors can be further fused through a fully connected layer to generate a comprehensive high-level semantic feature vector representing user activity. This vector embodies an understanding of user behavior and intent.

[0059] Step S3230: Using an attention-based fusion model, the low-level feature vector, the high-level semantic feature vector, and the device state vector constructed from the fourth environment's measured data are weighted and fused to generate the multimodal state vector.

[0060] Finally, an attention-based fusion model is employed to weightedly fuse the aforementioned low-level feature vectors, high-level semantic feature vectors, and device state vectors constructed from measured data in the fourth environment. The device state vectors reflect the current operating baseline of the controlled devices, such as the air conditioner's setting mode or the current brightness of the lights. The core of the attention mechanism lies in dynamically calculating the importance weights of different feature vectors within the current environmental context.

[0061] In one embodiment, the fusion model first concatenates or adds the three feature vectors to obtain an initial fused representation. This representation is then used as the query vector, and each feature vector is used as a key vector for computation. A weight score for each feature vector is obtained through dot product attention or additive attention mechanisms. Feature vectors with higher weight scores will dominate the final fusion result. The multimodal state vector generated after weighted fusion is a comprehensive and context-aware representation of the environment state, encompassing low-level physical signals, high-level semantic understanding, and the current operating state of the system.

[0062] This application achieves significant technical advantages in data representation through the above embodiments. These embodiments transform raw, heterogeneous multimodal data into a unified, semantically rich, and context-aware multimodal state vector through feature extraction and attention-based intelligent fusion. This approach overcomes the limitations of simple data stacking or fixed-weight fusion, dynamically highlighting key information in different contexts. This provides a far more accurate and richer description of the environmental state for subsequent parameter generation models than single-modal or simple fusion methods.

[0063] Based on any embodiment of the method in this application, the method further includes the following steps for training the parameter generation model: Step S5100: Obtain a training dataset, which includes multiple training samples constructed based on historical data generated from the daily operation of the private network. Each training sample includes a historical multimodal state vector, a historical control parameter vector of the online device executed in the historical state, and a new multimodal state vector after the control is executed. The intelligent central control device is responsible for building the training dataset used to train the parameter generation model. The quality and representativeness of this dataset directly determine the performance of the final model. The construction of the training dataset relies on historical data continuously accumulated by the private network during its daily operation. This historical data is a record of the real interactions between the smart home system and the environment, containing rich information on states, actions, and state transitions.

[0064] The training dataset contains a large number of training samples. Each training sample is a complete state-action-new state triple extracted from historical data. Specifically, a training sample contains three core elements. The first element is the historical multimodal state vector, representing the complete representation of the overall environmental state by the system after multimodal fusion at a certain historical moment, i.e., the input state upon which the model makes decisions. The second element is the historical control parameter vector actually executed by the system in that historical state, representing the decision or action made by the model (or the control logic at that time) in the face of the historical multimodal state vector. The third element is the new multimodal state vector collected by the system after executing the above control, reflecting the change in the environmental state after executing the control action, i.e., the result of the state transition.

[0065] The process of constructing training samples may include segmenting and aligning historical data streams. In one embodiment, the intelligent central control device takes a snapshot of the system state at fixed time intervals, such as every minute or whenever the environmental state changes significantly, recording the current multimodal state vector and the control parameter vector to be executed. After a short time delay, ensuring that the control action has taken effect and the environmental state has stabilized, the multimodal state vector is collected again as the new state. By associating consecutive historical snapshots in chronological order, a large number of training samples can be constructed. In another embodiment, the construction of training samples may be triggered by specific events, such as intensive sampling only around the time points when significant user interaction is detected or when environmental parameters fluctuate drastically, in order to focus on learning strategies for key decision points.

[0066] These training samples collectively depict the dynamic patterns of environmental state changes resulting from specific control actions performed in historical operations, providing learning material for deep reinforcement learning algorithms. This allows them to deduce the mapping relationship between states, actions, and rewards, thereby training a parameter generation model capable of intelligent decision-making.

[0067] Step S5200: Calculate the corresponding instant reward value for each training sample according to the predefined reward function. The reward function takes environmental evaluation indicators and operating energy consumption indicators as input parameters. The environmental evaluation indicators are determined based on the implicit user feedback information obtained after the execution of the corresponding historical control parameter vector. The operating energy consumption indicators are determined based on the change in energy consumption data before and after execution. Furthermore, the intelligent central control device can calculate a quantified, instantaneous reward value for each training sample in the training dataset based on a predefined reward function. This calculation process is a crucial step in transforming abstract control effects into numerical signals that the model can understand and optimize. The reward function itself is a predefined mathematical formula or calculation rule, designed to provide a clear direction for the model's learning.

[0068] The reward function uses two core indicators as input parameters: environmental evaluation indicators and operational energy consumption indicators. The environmental evaluation indicators are quantified based on implicit user feedback information obtained after the execution of historical control parameter vectors. In one embodiment, feedback is inferred by analyzing user behavior data over a period of time after the control action is executed. For example, if a user manually increases the air conditioning temperature immediately after control execution, this is considered negative implicit feedback, indicating that the previous control decision failed to meet the user's comfort needs. Conversely, if the user does not intervene manually for a long period after control execution, it can be interpreted as positive implicit feedback, indicating that the control effect meets the user's expectations. In another embodiment, implicit user feedback information may also include whether the user canceled the scene mode triggered by the control, or whether the user's continuous stay in the environmental space increased due to environmental improvement. These behavioral signals are quantified into specific scores for the environmental evaluation indicators according to a set of predefined mapping rules; for example, positive feedback corresponds to positive scores, and negative feedback corresponds to negative scores.

[0069] The determination of operational energy consumption indicators depends on the change in energy consumption data before and after the execution of control actions. In one embodiment, the intelligent central control device obtains the total energy consumption reading within a specific time window before and after control execution from the smart meter or the controlled device itself, and calculates the difference as the energy consumption change. In another embodiment, for a single device, the energy consumption change can be calculated by directly querying the product of its power and operating time. Subsequently, according to predefined energy efficiency assessment rules, this energy consumption change is converted into an operational energy consumption indicator score. For example, the rules may stipulate that, under the premise of meeting comfort requirements, the greater the reduction in energy consumption, the higher the score; while unnecessary increases in energy consumption will result in a deduction of points.

[0070] Finally, the reward function takes the environmental assessment score and the operational energy consumption score as inputs, and calculates a comprehensive immediate reward value through a weighted summation formula or other combinations. This immediate reward value, as a scalar signal, encapsulates a comprehensive evaluation of the historical control actions recorded for the training sample. It indicates to the reinforcement learning model whether executing the historical control action is good or bad in the long run under a specific historical state, and the degree of good or bad, thus providing the most direct guidance signal for the optimization of the model's parameters.

[0071] Step S5300: Construct a deep reinforcement learning network as the parameter generation model, wherein the state space of the parameter generation model is defined according to the historical multimodal state vector, and the action space is defined according to the historical control parameter vector; The intelligent central control device constructs a deep reinforcement learning network as the specific implementation architecture for the parameter generation model. The parameter generation model is the core component for achieving intelligent decision-making in this application; its essence is a function approximator capable of learning the optimal mapping from environmental states to control actions. Deep reinforcement learning networks are chosen due to their powerful representation learning capabilities and consist of an input layer, hidden layers, and an output layer. The hidden layers can contain multiple fully connected layers, convolutional layers, or recurrent neural network layers. The specific architecture can be designed according to the complexity of the task and the requirements for real-time performance.

[0072] The state space of the parameter generation model is defined based on the historical multimodal state vector. This means that the dimension of the network's input layer needs to match the dimension of the historical multimodal state vector. For example, if the historical multimodal state vector is an N-dimensional vector containing physical environment features, user activity semantic features, and device state features, then the input layer of the deep reinforcement learning network needs to be designed with N neurons to receive and process this state vector. The accurate definition of the state space ensures that the model can perceive and understand complete information about the environment.

[0073] The action space of the parameter generation model is defined based on the historical control parameter vector. The action space represents all possible outputs of the model. The output layer dimension of the network is consistent with the dimension of the historical control parameter vector. For example, if the historical control parameter vector is a vector containing K control parameters such as light brightness, air conditioner set temperature, and curtain opening degree, then the output layer of the deep reinforcement learning network is typically designed with K neurons, each neuron outputting a continuous or discrete control parameter value. In one embodiment, for continuous control tasks, the output layer can directly generate real values ​​of the control parameters. In another embodiment, for discrete action selection, the output layer can output the probability distribution of each selectable action, or use a deep Q-network architecture to output the Q-value of each discrete action. The accurate definition of the action space ensures that the model can output specific, executable control commands.

[0074] By anchoring the state space and action space to historical multimodal state vectors and historical control parameter vectors, respectively, a direct correspondence is established between the input and output of the deep reinforcement learning network and the actual operating data of the system. This design enables the trained model to receive the current multimodal state vector as input and directly output the corresponding control parameter vector, achieving end-to-end intelligent decision-making. The constructed deep reinforcement learning network provides parameterized policy functions to be optimized for subsequent iterative training.

[0075] Step S5400: Using the training data consisting of the training samples and the corresponding immediate reward values, the deep reinforcement learning network is iteratively trained through an experience replay mechanism and a target network update strategy to maximize the long-term cumulative reward consisting of the immediate reward values, and the network parameters are optimized until the model converges.

[0076] The intelligent central control device initiates an iterative training process for a deep reinforcement learning network. This process utilizes training data consisting of training samples and their corresponding immediate reward values, and optimizes the network parameters through a specific training algorithm. The fundamental goal of training is to enable the parameter generation model to learn a policy that maximizes the expected value of the long-term cumulative reward after executing a series of control actions. The long-term cumulative reward is typically defined as the discounted sum of the immediate reward values ​​at future time points, and its calculation needs to consider the impact of the current action on the future.

[0077] The training process can utilize an experience replay mechanism. The intelligent central control device stores a large amount of experience data generated during training or extracted from historical data—namely, tuples of states, actions, rewards, and new states—in a fixed-size circular buffer called the experience replay buffer. In one embodiment, a small batch of experience data is randomly sampled from this buffer during each training iteration, instead of directly using the most recent experience. This method breaks the temporal correlation between data, making the sampled data more consistent with the assumption of independent and identically distributed data, thereby greatly improving training stability and data utilization efficiency. In another embodiment, priority-based experience replay can be used, assigning different sampling probabilities to experience data based on its learning value. For example, experiences with larger temporal difference errors have a higher probability of being sampled, thus accelerating the learning process.

[0078] The training process also relies on a target network update strategy to stabilize learning. A widely used strategy is to establish a target network with the exact same structure as the main network being trained, but whose parameter updates lag behind the main network. In one embodiment, the parameters of the main network are periodically copied to the target network, for example, once after a certain number of training iterations. When calculating the target Q-value, this relatively fixed target network is used instead of the continuously changing main network. This method reduces fluctuations in the target value and prevents divergence or oscillations during training, which is crucial for the success of algorithms such as deep Q-learning. In another embodiment, a soft update strategy can be used, where only a small portion of the main network parameters are updated to the target network after each training iteration, allowing the target network parameters to slowly track the main network and achieve a smoother stabilization effect.

[0079] Based on sampled empirical data and the target network, the intelligent central control device performs specific network parameter updates. In one embodiment, for value-based algorithms such as deep Q-networks, the main network parameters are updated by minimizing the temporal difference error loss between the predicted Q-value and the target Q-value. The target Q-value is calculated by the target network and combined with the immediate reward value. In another embodiment, for policy-based algorithms such as proximal policy optimization, the network parameters are updated by calculating the policy gradient and combining it with the advantage function estimation, so that the model is more inclined to choose actions that bring higher long-term cumulative rewards. Optimization algorithms typically employ stochastic gradient descent or its variants, such as the Adam optimizer, to iteratively adjust network weights and biases.

[0080] The training process continues, with the intelligent central control device constantly optimizing network parameters using new empirical data until the model performance converges. Convergence can be determined by whether, on an independent validation set, model performance metrics such as average cumulative reward no longer show significant improvement, or by reaching the preset maximum number of training epochs. Ultimately, a well-trained parameter generation model is obtained that can replace the older version, thereby improving capabilities. This model can output a control parameter vector with high probability of maximizing long-term cumulative reward based on the input multimodal state vector.

[0081] This application, through the above embodiments, constructs an autonomous learning framework based on real historical interaction data, guided by multi-objective optimization, and possessing a stable and efficient training mechanism. This embodiment not only automatically constructs high-quality training samples using massive historical data generated by the smart home system itself, ensuring a high degree of consistency between the learning materials and real-world scenarios, but also transforms user experience objectives, which are difficult to optimize directly, into computable optimization signals through implicit user feedback and energy efficiency objectives via a pre-defined reward function. This enables the model to autonomously learn Pareto optimal strategies that balance comfort and energy consumption in complex dynamic environments. Furthermore, by integrating experience replay and the target network, the instability problem of reinforcement learning training is effectively overcome, ensuring the reliability and efficiency of model convergence. Ultimately, this complete training process enables the parameter generation model to continuously learn and self-optimize from historical experience, continuously improving its decision-making intelligence without relying on manually preset rules or labeled data. This achieves a qualitative leap from static automation to dynamic adaptation, enabling the provision of highly personalized and energy-efficient smart home services.

[0082] Based on any embodiment of the method in this application, according to a predefined reward function, the immediate reward value corresponding to each training sample is calculated. Before the reward function takes environmental assessment indicators and operational energy consumption indicators as input parameters, it includes: Step S4110: For the historical control parameter vector in the training sample and its new multimodal state vector after execution, determine the user's behavior data within a preset time period after the historical control is executed. The intelligent central control device uses historical control parameter vectors from training samples and their new multimodal state vectors after execution to determine user behavior data within a preset time period following the historical control execution. This preset time period is a configurable parameter, long enough to allow the effects of the control action to be fully manifested and the user's reaction observed, such as five or ten minutes. The acquisition of behavioral data relies on the system's existing perception capabilities. In one embodiment, by querying the user's position changes and posture sequences recorded by the visual acquisition device within this time period, it can be determined whether the user has left the environmental space or performed any specific actions. In another embodiment, by analyzing the audio stream recorded by the audio acquisition device, it can be identified whether the user has issued voice commands or exhibited emotional sounds. Simultaneously, manual adjustments to device parameters by the user via mobile applications or smart panels, recorded in the system logs, are also important sources of behavioral data. These behavioral data collectively constitute the original basis for analyzing the user's reaction to historical control actions.

[0083] Step S4120: Based on the behavioral data, identify implicit user feedback information, which includes at least one of the following: the user's manual adjustment of the parameters of the controlled online device, whether the user cancels the historical control, and the duration of the user's continuous stay in the environmental space. Based on the identified behavioral data, the intelligent central control device further identifies implicit user feedback. Implicit user feedback is the user's satisfaction or preference for the control effect expressed through their behavior rather than direct ratings. In one embodiment, it identifies manual adjustments made by the user to the parameters of the controlled online device. For example, if the system automatically sets the air conditioner to 25 degrees Celsius, and the user quickly adjusts it manually to 26 degrees Celsius, this action is considered strong negative implicit feedback, indicating that the automatically set temperature is lower than the user's expectations. In another embodiment, it identifies whether the user cancels a historical control. For example, if the system activates "movie mode," but the user immediately cancels the mode via voice or an app, this explicitly indicates dissatisfaction with the control scenario. Yet another embodiment identifies the duration of the user's stay in the environmental space. For example, if the system automatically activates a comfortable environment based on the perceived entry of a user, and the user does not leave for an extended period, this can be interpreted as positive implicit feedback, indicating that the environmental settings meet the user's wishes and make them feel comfortable. This feedback is indirect, but more natural and frequent than explicit ratings, and truly reflects the user experience.

[0084] Step S4130: According to the predefined mapping rules, the identified implicit user feedback information is quantified into scores of environmental evaluation indicators, where positive feedback corresponds to positive scores and negative feedback corresponds to negative scores.

[0085] After identifying implicit user feedback, the intelligent central control device continues to quantify the identified implicit user feedback into environmental evaluation index scores according to predefined mapping rules. The mapping rules are a set of explicit logic that maps specific behaviors to standardized scores. In one embodiment, negative feedback, such as when a user manually adjusts device parameters or cancels a scene, can be assigned a fixed negative score, such as -1. For positive feedback, such as when a user's dwell time exceeds a certain threshold, a positive score can be assigned, such as +0.5. In another more complex embodiment, the mapping rules can consider the intensity of the feedback. For example, the larger the manual adjustment, the larger the absolute value of the assigned negative score; the longer the dwell time, the higher the assigned positive score. The quantified environmental evaluation index score is a standardized numerical value, allowing implicit feedback of different natures and intensities to be uniformly measured and calculated in the reward function, providing clear and calculable signals for model optimization.

[0086] This application achieves key technical advantages in the evaluation mechanism of model training through the above embodiments. This embodiment creatively transforms user experience, which is difficult to quantify directly, into calculable environmental evaluation indicators by analyzing implicit feedback information generated by natural user behavior and based on explicit rules. This shifts the optimization objective of the reinforcement learning model from simply pursuing physical parameter optimization to focusing on genuine user satisfaction, effectively aligning system decisions with human subjective feelings. By using user behaviors such as canceling control, manual adjustment, and dwell time as training signals, it can automatically learn and adapt to the personalized preferences of different users, thereby significantly improving the personalization level and user acceptance of control strategies. Ultimately, this drives the evolution of smart home systems from mechanically executing commands to truly understanding and meeting user needs.

[0087] Based on any embodiment of the method in this application, according to a predefined reward function, the immediate reward value corresponding to each training sample is calculated. Before the reward function takes environmental assessment indicators and operational energy consumption indicators as input parameters, it includes: Step S4210: Before and after execution of the historical control parameter vector in the training sample, obtain energy consumption monitoring data from the smart meter or controlled online device in the private network. The intelligent central control device, based on the historical moments corresponding to the training samples, can obtain energy consumption monitoring data within a specific time window before and after the execution of historical control parameter vectors from the energy consumption monitoring unit in the private network. Energy consumption monitoring data is a quantitative reading reflecting energy consumption. In one embodiment, the data source can be a smart meter deployed on the circuit bus, which provides the total energy consumption reading of the entire environment or a specific circuit within the time period before and after control execution. In another embodiment, data can be directly obtained from the controlled online device itself. Many intelligent devices, such as smart air conditioners and smart lights, have the function of monitoring and reporting their own real-time power or cumulative energy consumption. The acquired energy consumption monitoring data needs to have a precise timestamp to accurately align with the time point of control execution.

[0088] Step S4220: Based on the energy consumption monitoring data, calculate the change in energy consumption after the historical control parameter vector is executed relative to before the execution, and use it as a basic energy consumption indicator; Based on the acquired energy consumption monitoring data, the change in energy consumption after the execution of historical control parameter vectors relative to before execution can be calculated, serving as a basic energy consumption indicator. Calculating the change in energy consumption requires defining a reasonable time window for comparison. In one embodiment, the average energy consumption over a fixed period before control execution can be selected as a benchmark and compared with the average energy consumption over the same period after control execution, calculating the difference. In another embodiment, the transient characteristics of equipment operation can be considered; for example, for an air conditioner, power consumption increases significantly after the compressor starts, allowing for comparison of energy consumption differences within a complete operating cycle before and after startup. The change in energy consumption is a signed numerical value, with a positive value indicating an increase in energy consumption and a negative value indicating a decrease in energy consumption. This basic energy consumption indicator intuitively reflects the direct impact of a single control action on energy consumption.

[0089] Step S4230: Based on the predefined energy efficiency assessment rules and the ratio of the energy consumption change to the baseline energy consumption value, generate the final operating energy consumption index, wherein the baseline energy consumption value is the rated power or historical average energy consumption of the controlled online equipment.

[0090] Furthermore, based on predefined energy efficiency assessment rules, the basic energy consumption indicators are further transformed into final operational energy consumption indicators. The energy efficiency assessment rules aim to evaluate absolute energy consumption changes within a reasonable context, making the assessment results more fair and comparable. These rules generate indicators based on the ratio of energy consumption change to a baseline energy consumption value. In one embodiment, the baseline energy consumption value can be the rated power of the controlled online equipment, typically provided by the equipment manufacturer and stored in the equipment capacity model. In this case, the ratio of energy consumption change to rated power reflects the proportion of energy consumption change caused by the current control action relative to the equipment's maximum capacity. In another preferred embodiment, the baseline energy consumption value can be the historical average energy consumption of the equipment under similar operating conditions, such as the average energy consumption during the same period over the past week at similar ambient temperatures. This approach eliminates the influence of factors such as seasons and usage habits, more accurately assessing the energy efficiency improvement or loss resulting from control actions. Based on the ratio, it is mapped to a score for the operational energy consumption indicator using a predefined function or lookup table method; for example, the smaller the ratio (i.e., the greater the energy consumption reduction and the larger the proportion relative to the baseline value), the higher the score.

[0091] This application, through the above embodiments, achieves a precise quantitative technical advantage in the energy efficiency optimization dimension of model training. This embodiment acquires accurate energy consumption monitoring data from smart meters or the devices themselves and calculates the relative change relative to the device's rated capacity or historical baseline level. This allows the evaluation results of operating energy consumption indicators to overcome the limitations of absolute values, fairly measuring the control energy efficiency of different devices under different operating conditions. This ratio-based standardized evaluation method provides clear and consistent energy efficiency optimization signals for the reinforcement learning model, guiding the model to focus on the relative efficiency of energy use while pursuing user comfort. This systematically learns and generates control strategies that maximize energy savings while meeting demand, ultimately achieving an effective unification of environmental adaptive control and sustainable energy-saving goals.

[0092] Based on any embodiment of the method in this application, before obtaining environmental measurement data provided by multiple online devices in the same environmental space and providing different modal data sources in a private network, the method includes: Step S2100: Periodically send broadcast discovery messages to the private network and listen for device announcement messages sent by online devices in the private network to identify newly connected online devices; The intelligent central control device can execute a device discovery process to identify newly added online devices in a private network. This process can combine active probing and passive monitoring mechanisms. The intelligent central control device periodically sends broadcast discovery messages to the private network, following common device discovery protocols such as mDNS or SSDP, to query whether new devices exist within the network. Simultaneously, the intelligent central control device continuously listens for device announcement messages spontaneously emitted by online devices within the network; many intelligent devices actively broadcast their presence information after starting up or connecting to the network. By receiving and parsing these announcement messages, the intelligent central control device can promptly detect the access of new devices. In one embodiment, the discovery cycle can be set to once every few minutes to achieve a balance between timeliness and network load. After identifying a newly accessed online device, the system records its initial network identification information, such as its IP address or MAC address.

[0093] Step S2200: For newly identified online devices, send a standardized capability query command to them and receive and parse the response returned by the device to obtain its device capability description information. The device capability description information includes the device's unique identifier, device type, supported control command set, and parameter range of each command. For newly identified online devices, this step is responsible for capability negotiation. The intelligent central control device sends a standardized capability query command to the device. This command is a structured data request with a predefined format and semantics, based on IoT standards such as OCF or Matter. Upon receiving the query command, the device returns a response message containing its capability description information. The intelligent central control device receives and parses this response to extract the structured device capability description information. This information includes at least a unique device identifier to uniquely identify the device in the system; the device type for functional classification, such as distinguishing between lighting devices, environmental sensors, or entertainment devices; and the set of control commands supported by the device and the parameter range of each command. For example, for smart lighting, the command set might include switching, adjusting brightness, and adjusting color temperature, while the brightness parameter range might be an integer from 0 to 100.

[0094] Step S2300: Based on the device capability description information, create a corresponding drive interface and state mapping table for the device in the preset unified device capability model to achieve standardized encapsulation of its device control commands and state feedback. The intelligent central control device creates corresponding software interfaces for new devices within a pre-defined unified device capability model based on the acquired device capability description information. The unified device capability model is an abstraction layer defined within the system, using a unified, standardized data structure and interface to describe the functions of all devices, thus shielding the heterogeneity of underlying devices. Specifically, it creates a corresponding driver interface and state mapping table for the new device. The driver interface is a set of standardized functions or methods used to communicate with the device. The state mapping table defines how to map device-specific state information to standardized attributes in the unified model, and how to convert standardized control commands into specific instructions that the device can understand. For example, regardless of the brand of air conditioner and the specific instructions used to adjust the temperature, they can all be controlled with a unified command through their driver interface and state mapping table, with the driver interface responsible for the conversion to specific instructions. This process achieves standardized encapsulation of device control commands and state feedback.

[0095] Step S2400: Register the newly accessed online device that has completed standardized encapsulation to the list of available devices to complete device registration.

[0096] After standardized encapsulation, the intelligent central control device registers the newly connected online device to the available device list. The available device list can be a dynamic database maintained by the system, recording all currently discoverable, communicative, and functional devices, along with their standardized capability descriptions. The registration process includes writing the device's unique identifier, type, and the driver interface pointer created for it into the list. Once registration is successful, the new device is officially included in the system's management scope. Its sensor data can be acquired in step S3100 as environmental measurement data, and it can also become a controlled online device in step S3400. At this point, the device adaptive access process is complete, achieving plug-and-play functionality for new devices.

[0097] This application, through the above embodiments, achieves fundamental technical advantages in system scalability and compatibility. This embodiment establishes an automated plug-and-play mechanism for devices through standardized device discovery, capability querying, and unified abstraction and encapsulation processes. This enables new smart devices to be automatically identified, understood, and integrated by the system after accessing the private network, without the need for cumbersome manual intervention in driver installation or protocol configuration. This not only significantly reduces system deployment and maintenance costs, but more importantly, it architecturally solves the long-standing problem of device heterogeneity in the smart home field. It provides a stable and scalable hardware foundation for multimodal data acquisition and collaborative control, ensuring that the environmental adaptive control method can operate stably and reliably in a dynamic environment where the device ecosystem is continuously evolving.

[0098] Please see Figure 2An environment-adaptive device control apparatus provided to meet one of the purposes of this application is a functional embodiment of the environment-adaptive device control method of this application. The apparatus includes a data acquisition module 3100, a vector construction module 3200, an inference generation module 3300, and a device control module 3400. The data acquisition module 3100 is configured to acquire environmental measurement data provided by multiple online devices in the same environmental space within a private network, each providing different modal data sources. The vector construction module 3200 is configured to construct single-source state vectors for each data source and fuse these single-source state vectors into a multimodal state vector representing the overall environmental state. The multimodal state vector contains semantic information for inferring the user's activity state; the inference generation module 3300 is configured to input the multimodal state vector into a preset parameter generation model for inference to obtain control parameter vectors corresponding to at least some of the multiple online devices. The parameter generation model is trained in advance using a reward function constructed with environmental evaluation indicators and operating energy consumption indicators through multi-objective reinforcement learning; the device control module 3400 is configured to construct device control instructions for the partial online devices based on the control parameter vectors to control the corresponding online devices, so that the adjusted overall environmental state can optimize the environmental evaluation indicators and operating energy consumption indicators.

[0099] Based on any embodiment of the device in this application, the data acquisition module 3100 includes: a first acquisition module, configured to acquire measured physical environment data from physical sensor devices deployed in the environmental space as first environmental measured data, the measured physical environment data including any of the following: temperature, humidity, light intensity, and noise decibel value; a second acquisition module, configured to acquire image sequences from visual acquisition devices deployed in the environmental space, and perform real-time analysis on the image sequences to obtain visual perception measured data as second environmental measured data, the visual perception measured data including any of the following: user identity, number of people in the room, user location, and user posture; a third acquisition module, configured to acquire audio streams from audio acquisition devices deployed in the environmental space, and perform real-time analysis on the audio streams to obtain audio perception measured data as third environmental measured data, the audio perception measured data including any of the following: speech text, environmental noise characteristics, and audio of specific events; and a fourth acquisition module, configured to acquire current operating status data from online devices to be controlled in the private network as fourth environmental measured data, the operating status data including any of the following: on / off status, working mode, and operating parameters.

[0100] Based on any embodiment of the device in this application, the vector construction module 3200 includes: a low-level representation module, configured to construct a low-level feature vector representing the physical environment based on the first environmental measured data through standardization processing; a high-level representation module, configured to extract a high-level semantic feature vector representing user activities based on the second environmental measured data and the third environmental measured data through a deep learning model; and a vector fusion module, configured to use an attention-based fusion model to weightedly fuse the low-level feature vector, the high-level semantic feature vector, and the device state vector constructed from the fourth environmental measured data to generate the multimodal state vector.

[0101] Based on any embodiment of the device in this application, the device further includes: a data retrieval module, configured to acquire a training dataset, wherein the training dataset includes multiple training samples constructed based on historical data generated from the daily operation of the private network, wherein each training sample includes a historical multimodal state vector, a historical control parameter vector of the online device executing in the historical state, and a new multimodal state vector after the control is executed; and a function calculation module, configured to calculate the corresponding instantaneous reward value for each training sample according to a predefined reward function, wherein the reward function takes environmental evaluation indicators and operating energy consumption indicators as input parameters, and the environmental evaluation indicators are calculated based on the corresponding historical control parameter vector after execution. The obtained implicit user feedback information is determined, and the running energy consumption index is determined based on the change in energy consumption data before and after execution; the model construction module is configured to construct a deep reinforcement learning network as the parameter generation model, wherein the state space of the parameter generation model is defined according to the historical multimodal state vector, and the action space is defined according to the historical control parameter vector; the model training module is configured to use training data composed of the training samples and corresponding immediate reward values, and iteratively train the deep reinforcement learning network through an experience replay mechanism and a target network update strategy, with the goal of maximizing the long-term cumulative reward composed of the immediate reward values, and optimize its network parameters until the model converges.

[0102] Based on any embodiment of the device in this application, prior to the function operation module, the device further includes: a behavior determination module, configured to determine user behavior data within a preset time period after the execution of the historical control based on the historical control parameter vector in the training samples and the new multimodal state vector after its execution; a feedback identification module, configured to identify implicit user feedback information based on the behavior data, wherein the implicit user feedback information includes at least one of the following: the user's manual adjustment operation on the parameters of the controlled online device, whether the user cancels the historical control, and the duration of the user's continuous stay in the environmental space; and an evaluation quantification module, configured to quantify the identified implicit user feedback information into scores of environmental evaluation indicators according to predefined mapping rules, wherein positive feedback corresponds to a positive score and negative feedback corresponds to a negative score.

[0103] Based on any embodiment of the device in this application, prior to the function operation module, the device further includes: an energy consumption monitoring module, configured to acquire energy consumption monitoring data from smart meters or controlled online devices in the private network before and after the execution of historical control parameter vectors in the training samples; a change detection module, configured to calculate the change in energy consumption after the execution of the historical control parameter vectors relative to before the execution based on the energy consumption monitoring data, as a basic energy consumption index; and an energy consumption quantification module, configured to generate a final operating energy consumption index based on the ratio of the change in energy consumption to a baseline energy consumption value according to predefined energy efficiency evaluation rules, wherein the baseline energy consumption value is the rated power or historical average energy consumption of the controlled online device.

[0104] Based on any embodiment of the device in this application, prior to the data acquisition module 3100, this device further includes: a message broadcasting module, configured to periodically send broadcast discovery messages to the private network and listen for device announcement messages sent by online devices within the private network to identify newly connected online devices; a capability querying module, configured to send standardized capability query instructions to the identified newly connected online devices and receive and parse the response returned by the device to obtain its device capability description information, the device capability description information including a unique device identifier, device type, supported control instruction set, and parameter range of each instruction; a capability mapping module, configured to create a corresponding driver interface and state mapping table for the device in a preset unified device capability model based on the device capability description information, thereby achieving standardized encapsulation of its device control instructions and state feedback; and a device registration module, configured to register the newly connected online devices that have completed standardized encapsulation to the list of available devices, thus completing device registration.

[0105] To address the aforementioned technical problems, embodiments of this application also provide a computer device for implementing the intelligent central control device of this application. For example... Figure 3The diagram shows the internal structure of a computer device. The computer device includes a processor, a computer-readable storage medium, a memory, a network interface, and various communication components connected via a system bus. The computer-readable storage medium stores an operating system, a database, and computer-readable instructions. The database may store control information sequences. When the computer-readable instructions are executed by the processor, they enable the processor to implement an environment-adaptive device control method. The processor of the computer device provides computing and control capabilities to support the operation of the entire computer device. The memory of the computer device may store computer-readable instructions. When these computer-readable instructions are executed by the processor, they enable the processor to execute the environment-adaptive device control method of this application. The network interface of the computer device is used for communication with a terminal. Those skilled in the art will understand that… Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0106] In this embodiment, the processor is used to execute... Figure 2 The system defines the specific functions of each module and its sub-modules. The memory stores the program code and various data required to execute these modules or sub-modules. The network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program code and data required to execute all modules / sub-modules in the environment adaptive device control device of this application. The server can call the server's program code and data to execute the functions of all sub-modules.

[0107] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the environment adaptive device control method of any embodiment of this application.

[0108] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0109] Those skilled in the art will understand that the steps, measures, and solutions in the various operations, methods, and processes discussed in this application can be alternated, modified, combined, or deleted. Furthermore, other steps, measures, and solutions in the various operations, methods, and processes discussed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted. Furthermore, steps, measures, and solutions in the prior art that are similar to those in the open-source operations, methods, and processes of this application can also be alternated, modified, rearranged, decomposed, combined, or deleted.

[0110] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An environment-adaptive device control method characterized by, The method comprises the following steps: obtaining environment measurement data provided by a plurality of online devices in a private network, which are in the same environment space and provide different modal data sources; constructing a single-source state vector corresponding to each data source, and fusing each single-source state vector into a multi-modal state vector representing the overall environment state, which contains semantic information for inferring the user activity state; inputting the multi-modal state vector into a preset parameter generation model for inference to obtain a control parameter vector corresponding to at least part of the online devices, wherein the parameter generation model is obtained by multi-objective reinforcement learning training of a reward function constructed by an environment evaluation index and an operation energy consumption index in advance; constructing a device control instruction of the part of online devices according to the control parameter vector to control the corresponding online devices, so that the adjusted overall environment state can optimize the environment evaluation index and the operation energy consumption index.

2. The ambient adaptive device control method of claim 1, wherein, The method comprises the following steps: obtaining environment measurement data provided by a plurality of online devices in a private network, which are in the same environment space and provide different modal data sources; obtaining physical environment measurement data from physical sensor devices deployed in the environment space as first environment measurement data, which includes any multiple of temperature, humidity, light intensity, and noise decibel value; obtaining image sequences from visual acquisition devices deployed in the environment space, and performing real-time analysis on the image sequences to obtain visual perception measurement data as second environment measurement data, which includes any multiple of user identity, number of people in the room, user position, and user posture; obtaining audio streams from audio acquisition devices deployed in the environment space, and performing real-time analysis on the audio streams to obtain audio perception measurement data as third environment measurement data, which includes any multiple of speech text, environmental noise characteristics, and specific event audio; 3. The ambient adaptive device control method of claim 2, wherein, obtaining current running state data of the online devices to be controlled in the private network as fourth environment measurement data, which includes any multiple of on-off state, working mode, and running parameter. The method comprises the following steps: constructing a single-source state vector corresponding to each data source, and fusing each single-source state vector into a multi-modal state vector representing the overall environment state, which contains semantic information for inferring the user activity state; constructing a low-level feature vector representing the physical environment by standardization based on the first environment measurement data; 4. The ambient adaptive device control method of claim 1, wherein, extracting a high-level semantic feature vector representing user activity by a deep learning model based on the second and third environment measurement data; using a fusion model based on an attention mechanism to weight and fuse the low-level feature vector, high-level semantic feature vector, and device state vector constructed from the fourth environment measurement data to generate the multi-modal state vector. The method further comprises the following steps: obtaining a training data set, which includes a plurality of training samples constructed based on historical data generated during daily operation of the private network, wherein each training sample includes a historical multi-modal state vector, a historical control parameter vector of the online devices executed under the historical state, and a new multi-modal state vector after executing the control; According to a predefined reward function, a corresponding immediate reward value of each training sample is calculated, the reward function taking an environment evaluation index and a running energy consumption index as input parameters, the environment evaluation index being determined according to user implicit feedback information obtained after execution of a corresponding historical control parameter vector, and the running energy consumption index being determined according to a change amount of energy consumption data before and after execution; A deep reinforcement learning network is constructed as the parameter generation model, wherein a state space of the parameter generation model is defined according to the historical multi-modal state vector, and an action space is defined according to the historical control parameter vector; The deep reinforcement learning network is iteratively trained by an experience replay mechanism and a target network update strategy, to maximize a long-term cumulative reward composed of the immediate reward values, optimize network parameters thereof, and converge the model.

5. The ambient adaptive device control method of claim 4, wherein, Before the predefined reward function is used to calculate a corresponding immediate reward value of each training sample, the reward function taking an environment evaluation index and a running energy consumption index as input parameters, the following steps are included: For a historical control parameter vector in the training sample and a new multi-modal state vector obtained after execution of the historical control parameter vector, behavior data of a user within a preset time period after the execution of the historical control is determined; Based on the behavior data, user implicit feedback information is identified, the user implicit feedback information including at least one of the following: a manual adjustment operation of the user on a controlled online device parameter, whether the user cancels the historical control, and a duration of continuous stay of the user in the environment space; According to a predefined mapping rule, the identified user implicit feedback information is quantified into a score value of the environment evaluation index, wherein positive feedback corresponds to a positive score value, and negative feedback corresponds to a negative score value.

6. The environmentally adaptive device control method of claim 4, wherein, Before the predefined reward function is used to calculate a corresponding immediate reward value of each training sample, the reward function taking an environment evaluation index and a running energy consumption index as input parameters, the following steps are included: For a historical control parameter vector in the training sample, energy consumption monitoring data is obtained from a smart meter in the private network or the controlled online device itself before and after execution of the historical control parameter vector; Based on the energy consumption monitoring data, a change amount of energy consumption after execution of the historical control parameter vector relative to before execution is calculated as a basic energy consumption index; According to a predefined energy efficiency evaluation rule, a final running energy consumption index is generated based on a ratio of the change amount of energy consumption to a benchmark energy consumption value, wherein the benchmark energy consumption value is a rated power or a historical average energy consumption of the controlled online device.

7. The environmentally adaptive device control method of any one of claims 1 to 6, wherein, Before environment measurement data provided by a plurality of online devices in a private network that are in the same environment space and provide different modal data sources is obtained, the following steps are included: Periodically sending a broadcast discovery packet to the private network and listening to device announcement packets sent by online devices in the private network to identify newly connected online devices; For the identified new access online device, a standardized capability query instruction is sent to it, and a response returned by the device is parsed to obtain device capability description information including a device unique identifier, a device type, a supported control instruction set, and a parameter range of each instruction; According to the device capability description information, a corresponding driving interface and state mapping table are created for the device in a preset unified device capability model, so as to realize standardized packaging of device control instructions and state feedback thereof; The new access online device after completing the standardized packaging is registered in an available device list, and device registration is completed.

8. An environment-adaptive device control apparatus characterized by comprising: Comprise: A data acquisition module is configured to acquire environment measurement data provided by a plurality of online devices in a private network that are in the same environment space and provide different modal data sources; A vector construction module is configured to construct a single-source state vector for each data source, and fuse each single-source state vector into a multi-modal state vector representing the overall environment state, the multi-modal state vector containing semantic information for inferring a user activity state; An inference generation module is configured to input the multi-modal state vector into a preset parameter generation model for inference to obtain a control parameter vector corresponding to at least part of the online devices, the parameter generation model being obtained by multi-objective reinforcement learning training of a reward function constructed by an environment evaluation index and an operation energy consumption index in advance; A device control module is configured to construct a device control instruction of the part of the online devices according to the control parameter vector to control the corresponding online devices, so that the adjusted overall environment state can optimize the environment evaluation index and the operation energy consumption index.

9. A smart central control device comprising a controller, the controller comprising a processor and a memory, characterized in that, The processor calls a computer program in the memory to execute the steps of the environment adaptive device control method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program implemented according to the method of any one of claims 1 to 7 in the form of computer readable instructions, and when the computer program is called and run by a computer, the steps included in the corresponding method are executed.

Images