A smart speaker control method and system based on smart home

Abstract

This invention discloses a smart speaker control method and system based on smart homes. The method includes the following steps: S1: Real-time collection of environmental and user status data through a distributed sensor array to construct a dynamic scene map; S2: Based on the dynamic scene map, generating at least one hypothetical action chain representing the user's potential intention; S3: Real-time capture and parsing of explicit voice commands or implicit behavioral inversion intentions from multiple users. When multiple intentions are detected simultaneously, conflict arbitration is performed based on a dynamic priority weighting algorithm; S4: Based on the arbitration result, control commands are issued, and the system switches to local autonomous mode when the wide area network is interrupted; S5: Implicit and / or explicit user feedback on executed actions is collected and used as a reward signal to perform incremental reinforcement learning updates on the local prediction model. This invention effectively solves the problems of command conflict and resource contention, ensuring personalized experience and fairness.

Description

Technical Field

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

[0002] With the development of IoT and AI technologies, smart speakers have become the core interaction interface for smart home systems. Traditional smart speaker control methods mainly rely on voice commands issued by the user, such as the user saying "turn on the living room lights," which the speaker receives, interprets, and executes.

[0003] However, traditional solutions have the following technical problems: First, their interaction mode is a passive control based on command and response, which cannot predict user intentions and proactively provide services; second, in a multi-user home environment, they cannot effectively handle command conflicts or provide personalized responses based on user identity; and finally, they are highly dependent on the network and cloud computing power, resulting in high response latency when the network fluctuates, and they cannot utilize the distributed computing capabilities among local devices for collaborative perception. Summary of the Invention

[0004] The purpose of this invention is to provide a smart speaker control method and system based on smart home technology, so as to solve the problems of passive interaction, lack of personalized prediction and excessive reliance on cloud computing in the prior art.

[0005] This invention is achieved using the following technical solution: a smart speaker control method based on smart home, comprising the following steps: S1: Real-time collection of environmental and user status data through a distributed sensor array to construct a dynamic scene map; S2: Based on the dynamic scene map, execute a lightweight temporal prediction model at the local edge of the smart speaker to generate at least one hypothetical action chain representing the user's potential intention; S3: Capture and parse explicit voice commands or implicit behavioral inversion intentions from multiple users in real time. When multiple intentions are detected to coexist, perform conflict arbitration based on a dynamic priority weighting algorithm. S4: Based on the arbitration result, control commands are sent to one or more smart home execution devices through the distributed device capability abstraction layer, and the local autonomous mode is switched when the wide area network is interrupted. S5: Collect implicit and / or explicit user feedback on performed actions, use it as a reward signal, and perform incremental reinforcement learning updates on the local prediction model.

[0006] Furthermore, step S1 includes the following sub-steps: S111: Establish a local data acquisition bus; S112: Set the asynchronous acquisition cycle; S113: Perform real-time credibility marking on the acquired raw data stream. If the trigger frequency of a certain sensor exceeds a preset multiple standard deviation of its historical average and is not related to sensors at adjacent locations, then mark the data stream as high noise and reduce its weight in subsequent fusion calculations.

[0007] Furthermore, step S2 includes the following sub-steps: S211: Load a temporal convolutional network based on an attention mechanism as a lightweight prediction model; S212: Take the sequence of environment-user joint state tensors covering the past N minutes as the model input and output the probability distribution of high-level behavioral intentions; S213: At the current moment, perform a forward inference calculation on the tensor sequence in the input window to obtain at least one most likely user intent and its probability value at the current moment.

[0008] Furthermore, step S2 also includes the following sub-steps: S221: Query the scene-action mapping decision tree forest stored locally to obtain multiple action chains associated with the prediction intent and the success probability of each chain based on historical statistical data; S222: Based on the current environmental parameters, generate specific control parameters for the highest probability intention and generate a preliminary list of hypothetical actions; S223: Assign confidence weights to each hypothetical action based on the user's historical habits. If an action occurs frequently in the past preset period, increase its weight; otherwise, decrease its weight.

[0009] Furthermore, step S3 includes the following sub-steps: S311: When an explicit voice command is detected, perform voiceprint recognition to determine the user ID that issued the command, and perform natural language understanding to identify the command intent and slot; S312: When there is no explicit voice command, the implicit intent is inverted through the user's behavior trajectory or device operation sequence; S321: Obtain the dynamic priority weight for each user ID, which is dynamically calculated from the base priority coefficient, the current activity urgency coefficient, and the recent interaction satisfaction coefficient; S322: Determine the conflict type between the explicit instruction and the predicted intent. If it is a direct conflict, the explicit instruction has higher priority than any predicted intent and the hypothetical action of the conflict is canceled immediately. If it is a resource preemption conflict, proceed to the compromise action generation step.

[0010] Furthermore, the compromise action generation step includes: The time reuse scheme is to inform the user of the allocation plan through voice prompts, first execute short actions of low-priority intents, and then automatically switch to actions of high-priority intents. The spatial routing scheme, based on the voiceprint localization results, routes actions with different intentions to smart speakers or execution devices in different spaces for execution.

[0011] Furthermore, step S4 includes the following sub-steps: S421: Periodically sends heartbeat packets to the cloud. If no response is received after a preset number of consecutive attempts, it will automatically switch to local autonomous mode. S422: Start the local rule engine based on finite state machine, use the real-time local sensor data collected in step S1 to execute the reduced prediction logic, and execute the high-frequency rules extracted and pre-cached from the historical cloud model. S423: When the cloud connection is restored, all sensor events, decision logs and execution results stored during the local autonomous mode are timestamped and uploaded to the cloud in batches for incremental training of the model.

[0012] Furthermore, step S5 includes the following sub-steps: S511: Collect explicit feedback and / or implicit feedback, wherein explicit feedback includes affirmative or negative words in the user's voice evaluation, as well as the user's manual cancellation of the automatically executed action; implicit feedback includes changes in the user's physiological parameters within a preset time window after the action is executed, device operation reversal behavior, and acoustic feature analysis of speech without semantics. S512: Calculate the temporal difference error, compare the predicted state value before the action with the reward obtained after the action plus the value of the next state; and use the error to update only the last two layers of the local temporal convolutional network model through the backpropagation algorithm, keeping the weights of the previous layers frozen; encrypt and upload the experience replay tuple containing the current scene state, the action executed, the reward obtained and the next scene state to the cloud for periodic training of the global benchmark model.

[0013] Furthermore, it also includes the following steps: S6: Monitors safety and user physiological abnormalities, and unconditionally interrupts the automatic control process in case of abnormalities.

[0014] A smart speaker control system based on smart home technology, used to implement the aforementioned smart speaker control method, includes: The data acquisition module is used to collect environmental and user status data in real time through a distributed sensor array and to construct a dynamic scene map; The action generation module is used to execute a lightweight temporal prediction model at the local edge of the smart speaker to generate at least one hypothetical action chain representing the user's potential intention. The conflict arbitration module is used to capture and parse explicit voice commands or implicit behavioral inversion intentions of multiple users in real time. When multiple intentions are detected to coexist, conflict arbitration is performed based on a dynamic priority weighting algorithm. The command issuance module is used to issue control commands to one or more smart home execution devices through the distributed device capability abstraction layer, and switch to local autonomous mode when the wide area network is interrupted. The optimization module is used to collect implicit or explicit user feedback on the actions performed, and use it as a reward signal to perform incremental reinforcement learning updates on the local prediction model.

[0015] The beneficial effects of this invention are as follows: This invention, through a distributed device capability abstraction layer and a local autonomous mode, enables the execution of core scenario control based on pre-cached rules even when the wide area network is interrupted, significantly improving the robustness and offline availability of the system.

[0016] This invention uses implicit user feedback (such as physiological parameters, device reversal operations, and voice tone) as a reward signal for reinforcement learning. It can achieve rapid incremental learning by only updating the weights of the last layer of the model, making the control strategy more and more in line with the long-term habits of a specific family. At the same time, the cloud and edge evolve together.

[0017] This invention incorporates a physiological abnormality and safety environment monitoring mechanism, which unconditionally takes over control and triggers an emergency response in emergency situations, thereby improving the invention in multiple dimensions such as intelligence, personalization, reliability, and safety. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0022] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0023] See Figure 1 A method for controlling a smart speaker based on smart home technology includes the following steps: S1: Real-time collection of environmental and user status data through a distributed sensor array to construct a dynamic scene map; S2: Based on the dynamic scene map, execute a lightweight temporal prediction model at the local edge of the smart speaker to generate at least one hypothetical action chain representing the user's potential intention; S3: Capture and parse explicit voice commands or implicit behavioral inversion intentions from multiple users in real time. When multiple intentions are detected to coexist, perform conflict arbitration based on a dynamic priority weighting algorithm. S4: Based on the arbitration result, control commands are sent to one or more smart home execution devices through the distributed device capability abstraction layer, and the local autonomous mode is switched when the wide area network is interrupted. S5: Collect implicit and / or explicit user feedback on performed actions, use it as a reward signal, and perform incremental reinforcement learning updates on the local prediction model.

[0024] In this embodiment, step S1 includes the following sub-steps: S11: Start the multi-source heterogeneous sensor data acquisition process; S12: Integrate multimodal data to generate spatiotemporally correlated scene feature vectors.

[0025] The process of initiating the multi-source heterogeneous sensor data acquisition includes: S111: Establish a local data acquisition bus. Specifically, the smart speaker establishes a low-latency communication connection with at least one environmental sensor node (such as a temperature and humidity sensor, a light sensor, or a human infrared motion sensor), at least one user status sensor (such as a pressure sensor for a smart mattress, or a heart rate and acceleration sensor for a smart bracelet), and at least one device status sensor (such as a power monitoring module for a smart light or a travel position sensor for a smart curtain) through its built-in Zigbee, Wi-Fi, Thread, or Bluetooth Mesh protocol stack.

[0026] S112: Set asynchronous acquisition cycles. To avoid data storms, set differentiated acquisition frequencies for different types of sensors. For slowly changing environmental parameters (such as room temperature and humidity), use the first sampling frequency (e.g., once every 30 seconds). For user behavior-related state parameters (such as human movement and door magnetic switches), use the second sampling frequency (e.g., once every 200 milliseconds). For high-frequency interactive devices (such as the power on / off status of a television), use event-triggered acquisition (i.e., reporting only when the state changes).

[0027] S113: Perform preliminary labeling of sensor data reliability and conduct real-time quality assessment of the collected raw data stream. For example, if a human infrared sensor is triggered more than three times the standard deviation of its historical average in the past 24 hours and is not correlated with sensors in adjacent locations, the data stream is labeled as high noise and its weight is reduced in subsequent fusion calculations. Another example is comparing heart rate data and acceleration data reported by a smart bracelet; if the heart rate suddenly increases but the acceleration shows no change, it is determined that the bracelet may have been removed or the data is abnormal, and its reliability coefficient is temporarily reduced.

[0028] The fusion of multimodal data to generate spatiotemporally correlated scene feature vectors includes: S121: Construct a user location heatmap. This involves combining the trigger timing of multiple fixed-location human sensors (such as those at the living room sofa, bedroom doorway, and kitchen countertop), the sound source localization algorithm results from the smart speaker's own microphone array (based on Time Difference of Arrival (TDOA),) and the Bluetooth RSSI signal strength of wearable devices to generate a probabilistic user location heatmap on a two-dimensional home map. This heatmap not only indicates which room the user is in but also quantifies the probability density of the user staying in a particular area.

[0029] S122: Extract temporal behavior segments and slice the continuous sensor event stream. For example, the sequence of events such as the kitchen human body sensor being triggered, the smart faucet valve opening 5 seconds later, the water purifier TDS sensor value fluctuating 3 seconds later, and the microwave oven door magnetic switch signal 10 seconds later is packaged into a temporal behavior segment and its semantic label is initially marked as "preparing to cook".

[0030] S123: Generate an environment-user joint state tensor. Align environmental parameters (temperature, humidity, light intensity, air quality index) with user behavior segments and current device operating status (light color temperature / brightness, curtain opening degree, air conditioner set temperature) in time and space to generate a multi-dimensional joint state tensor T(x, y, z, t), where x and y represent spatial grid coordinates, z represents the sensor type dimension, and t represents the time dimension.

[0031] In this embodiment, step S2 includes: S21: Load a lightweight prediction model locally on the smart speaker; S22: Generate a chain of action hypotheses to be executed based on the predicted intent.

[0032] The loading of a lightweight prediction model locally on the smart speaker includes: S211: Model architecture selection. A Temporal Convolutional Network (TCN) based on an attention mechanism is used as the base model. Compared to Recurrent Neural Networks (RNNs), this model has higher parallel computation efficiency and can capture long-range temporal dependencies. The model has been pre-trained in the cloud using massive amounts of anonymized family data, and subsequently fine-tuned at the edge based on the family's historical behavioral logs.

[0033] S212: Model Input and Output. The model input is the joint state tensor sequence generated in step S123, covering the past N minutes (e.g., N=5). The model output is a probability distribution vector, where each dimension corresponds to a preset high-level behavioral intention, such as: preparing to sleep, preparing to leave home, engaging in entertainment activities, performing cleaning work, preparing to eat, etc.

[0034] S213: Perform a forward inference. At the current moment, the neural network accelerator (NPU) of the smart speaker performs a forward inference calculation on the tensor sequence of the input window to obtain the three most likely user intentions and their probability values ​​at the current moment.

[0035] The step of generating the action hypothesis chain to be executed based on the predicted intent includes: S221: Query the scene-action mapping knowledge base. The smart speaker locally stores a lightweight scene-action mapping library. This library is not a fixed set of if-then rules, but a forest of decision trees. For example, for the intention to prepare to fall asleep, the library associates multiple possible action chains, each with a learned success probability based on historical statistical data.

[0036] S222: Generate a preliminary list of hypothetical actions. Based on current environmental parameters, generate specific control parameters for the intent with the highest probability. For example: if the predicted intent is to prepare for sleep and the current living room light intensity is > 300 lux, then generate hypothetical action H1: The main light brightness linearly dims to 20% within 30 seconds, and the color temperature gradually changes from 4000K to 2700K. If the bedroom curtains are simultaneously detected to be open and the current time is after 21:30, then generate hypothetical action H2: The bedroom motorized curtains slowly close to 90%. If the smart speaker's microphone array detects that the current ambient noise level is below 30 decibels (indicating a quiet environment), then generate hypothetical action H3: Play preset white noise or sleep-aid music, starting at volume 15, and set to automatically stop after 30 minutes.

[0037] S223: Confidence-weighted assignment of action hypotheses. For each generated hypothetical action H1, H2, and H3, a confidence weight is assigned based on its match with the user's historical habits. For example, if system logs show that users have a 90% probability of closing the curtains within 5 minutes of dimming the lights during the same time period over the past 7 nights, then the weight of action H2 is adjusted to 0.9. If there is no record of ever closing the curtains, the weight is reduced to 0.1.

[0038] In this embodiment, step S3 includes: S31: Real-time monitoring and parsing of multiple user commands or behavioral cues; S32: Execute the multi-priority weighted arbitration algorithm.

[0039] The real-time monitoring and parsing of multiple user commands or behavioral cues includes: S311: Voiceprint and content parsing of explicit commands. When the smart speaker's microphone array detects a wake word or direct command, the following sub-steps are executed: S3111: Voiceprint Recognition. Using a lightweight speaker verification model, the voiceprint feature vector in the voice command is extracted and compared with the registered family member voiceprint database to determine the user ID of the user who issued the command.

[0040] S3112: Natural Language Understanding (NLU) performs intent classification and slot filling on instruction text. For example, for the instruction "turn the volume up," the recognized intent is "adjust volume," and the slot is "volume increase."

[0041] S312: Implicit Behavior Intent Inversion. Even without voice commands, the system infers the user's intent from their actions, which may correspond to the predicted intent in the steps described above. Figure 1 This could also lead to new intentions.

[0042] S3121: Trajectory-based intent inversion. If a user walks from the study to the bedroom and carries a connected smartphone (determined by changes in Bluetooth signal strength), the system infers that the intent is to rest in the bedroom or change workplaces.

[0043] S3122: Intent inversion based on device operation sequence. If the user manually operates the following actions in succession: turning off the living room TV, picking up the smart remote control, and switching the air conditioner mode from cooling to ventilation, the system infers that the intent is to end the movie-watching activity and prepare to leave the living room.

[0044] The execution of the multi-priority weighted arbitration algorithm includes: S321: Obtain the dynamic priority weight for each user ID. This weight is not fixed but dynamically calculated based on various real-time factors. The calculation formula can be abstractly represented as: Weight = Base Priority Coefficient × (Current Activity Urgency Coefficient) × (Recent Interaction Satisfaction Coefficient). For example, during late-night hours on a weekday, the base priority coefficient for children is temporarily lowered, while the coefficient is raised if the father needs to get up early the next day. If a user's command issued within the past hour is not executed correctly twice consecutively (e.g., due to network problems), their urgency coefficient will be artificially increased to avoid frustration.

[0045] S322: Determine the type of conflict between explicit instructions and predicted intentions.

[0046] Type A (Direct Conflict): The hypothetical action H2 (close the curtains) generated by the predicted intent directly contradicts the explicit instruction not to close the curtains. In this case, the explicit instruction has higher priority than any predicted intent. The system immediately cancels the execution of H2 and records this conflict as negative feedback for subsequent model updates.

[0047] Type B (Resource Preemption Conflict): Both the predicted intent H3 (play sleep-aid music) and the explicit instruction to play nursery rhymes require audio playback resources. In this case, the current dynamic priority weights of the two users (father and child) are compared. If the father's weight is 0.85 and the child's weight is 0.6, the father's intent is executed. However, the system does not abruptly stop but instead executes sub-step S323.

[0048] S323: Generates a compromise or negotiation action. When a resource preemption conflict occurs, the system outputs a coordination solution instead of completely obeying the highest priority.

[0049] S3231: Time reuse. The system prompts via voice: "Okay, I'll play a song first, and then switch to sleep mode in 3 minutes." It then plays a children's song, automatically lowering the volume and switching to sleep-aid music after 3 minutes.

[0050] S3232: Spatial routing. If a home has multiple smart speakers (e.g., one in the living room and one in the children's room), the system automatically routes the command to play nursery rhymes to the speaker in the children's room, while the living room speaker plays soothing music. This decision is based on voiceprint localization results—the system identifies that the child's voice mainly comes from the direction of the children's room.

[0051] S324: In the absence of explicit instructions, execute the predicted intention action with the highest confidence. If no explicit voice instruction is detected within the current time window (e.g., 10 seconds), the system assumes that the user tacitly agrees with the prediction result, and directly executes the hypothetical action chain with the highest confidence weight in step S222, and provides prompts before or after execution through minimalist voice or visual signals (e.g., the LED halo on the top of the speaker breathing slowly).

[0052] In this embodiment, step S4 includes: S41: Generate control commands independent of the distribution device; S42: Perform local offline control under distributed decision-making.

[0053] The generation of control instructions independent of the distribution device includes: S411: Convert the intent into a device capability description. Instead of directly sending a message to set the brightness of Brand A bulbs to 50%, the system sends a capability request: Lighting device: Target brightness medium level, gradient duration 5 seconds, target area coordinates (living room, sofa area).

[0054] S412: Local device capability matching and translation. A distributed capability registry within the local area network (which may reside on the smart gateway or main speaker) receives the request and searches for a list of devices that can meet the requirements of lighting devices, medium level, gradient, and living room sofa area. Subsequently, for the specific light bulb device found, its corresponding device driver is invoked to translate the abstract instructions into device-specific binary control frames (such as ZCL standard frames or proprietary protocol frames).

[0055] The local offline control under distributed decision-making includes: S421: Detects cloud connection status; the smart speaker periodically sends heartbeat packets to the cloud. If no response is received three times consecutively, it automatically switches to local autonomous mode.

[0056] S422: Start local rule engine backup. In local autonomous mode, complex cloud-based machine learning models are unavailable. However, the speaker will start a local rule engine based on a finite state machine. This engine uses real-time local sensor data collected in step S1 to execute a scaled-down version of prediction logic. For example, if a human infrared sensor has not been triggered locally for more than 30 minutes and the current time is a weekday afternoon, the core actions of the "away from home" mode will be automatically executed (turn off all lights and disconnect power to non-essential outlets). This rule is a high-frequency rule extracted from historical cloud models and has been pre-cached in local storage.

[0057] S423: When the cloud recovers, synchronize local logs. Once the network is restored, the smart speaker will timestamp all sensor events, decision logs and execution results stored during the local autonomous mode and upload them in batches to the cloud for incremental training of subsequent models.

[0058] In this embodiment, step S5 includes: S51: Acquisition of implicit and explicit feedback signals; S52: Perform incremental reinforcement learning updates.

[0059] The acquisition of the implicit and explicit feedback signals includes: S511: Acquire explicit feedback. This includes direct voice feedback from the user, such as "Okay," "Thank you," "No," and "Stop." "Okay" and "Thank you" are awarded a positive reward signal of +1; "No," "Stop," or if the user immediately manually cancels the automatically executed action (e.g., manually pressing a physical switch to turn on a light that was turned off by the speaker within 5 seconds) are awarded a negative penalty signal of -1.

[0060] S512: Collects implicit feedback; more refined feedback comes from changes in user behavior. S5121: Body comfort feedback. Assuming the system predicts the user's intention to fall asleep, it automatically sets the air conditioner temperature to 24 degrees Celsius. Within 15 minutes of this action, if the smart mattress or smart bracelet detects a significant decrease in the user's turning frequency and the heart rate variability (HRV) indicator shows a relaxed state, this is considered a strong positive implicit feedback.

[0061] S5122: Continuous action feedback, assuming the system automatically turns off the TV. If the user does not turn the TV back on within the next 2 minutes, moderate positive feedback is given. Conversely, if the user turns the TV back on within 1 minute, it is considered strong negative feedback, indicating a prediction error.

[0062] S5123: Voice intonation feedback. Even without spoken content, the smart speaker can determine the user's satisfaction level by analyzing the acoustic characteristics (pitch, duration, energy) of the meaningless interjections the user subsequently utters.

[0063] The incremental reinforcement learning update includes: S521: Calculate the temporal difference error by comparing the predicted state value before the action with the reward obtained after the action plus the value of the next state, and calculate the error. This error directly reflects the quality of this decision.

[0064] S522: Update the weights of the last two layers of the local TCN model. Using the above error, adjust the weight parameters of only the last two fully connected layers of the local TCN prediction model of the smart speaker through the backpropagation algorithm.

[0065] S523: Upload the experience replay tuple to the cloud. The complete four-tuple of this interaction (current scene state, action performed, reward obtained, next scene state) is encrypted and uploaded to the cloud. The cloud server aggregates these tuples from a massive number of households for periodic retraining of the global baseline model, and distributes the optimized model increments to smart speakers in each household during off-peak hours at night.

[0066] In this embodiment, the following steps are also included: S6: Monitor safety and user physiological abnormalities, and unconditionally interrupt the automatic control process in the event of an abnormality. This step mainly includes: S61: Monitor key indicators for abnormalities and interrupt automatic execution; S62: Provides a convenient one-click exit automation mechanism.

[0067] The monitoring of key indicators that become abnormal and the automatic execution is interrupted includes: S611: Monitoring user's physiological indicators. If the smart speaker detects a sudden increase in the user's heart rate to over 120 beats per minute via a connected wearable device, accompanied by severe shaking detected by the accelerometer (potentially indicating a fall or panic), it immediately interrupts any ongoing automated scenarios (such as turning off the lights) and proactively activates the emergency protocol: restoring the lights to 100% brightness and asking "Are you okay? Do you need help?" via the speaker. If there is no positive response within 10 seconds or if keywords such as "help" are detected, it automatically calls the preset emergency contact.

[0068] S612: Environmental safety anomaly monitoring. If the smoke sensor or carbon monoxide sensor triggers an alarm, the smart speaker will immediately disable all automation logic, turn all lights to maximum brightness, play an evacuation alarm throughout the house, and proactively shut off the gas valve and turn on the ventilation equipment. Its priority is higher than any user comfort prediction.

[0069] The convenient one-click exit automation mechanism includes: S621: Identify specific exit commands. Users can say "exit automatic mode" or "I will manually control" to permanently or for a set time suspend the active prediction function of this invention.

[0070] S622: Physical action exit. Quickly tap the top of the smart speaker three times in succession (identified by the accelerometer), or press and hold the physical mute button on the speaker for 5 seconds to temporarily disable all active predictive actions, allowing the speaker to return to the traditional passive response mode and meet the user's need for absolute control.

[0071] Furthermore, this invention also provides a smart speaker control system based on smart home technology to implement the aforementioned smart speaker control method, comprising: a data acquisition module for real-time acquisition of environmental and user status data through a distributed sensor array and construction of a dynamic scene map; an action generation module for executing a lightweight temporal prediction model at the local edge of the smart speaker to generate at least one hypothetical action chain representing the user's potential intent; a conflict arbitration module for real-time capture and parsing of explicit voice commands or implicit behavioral inversion intents of multiple users, and for performing conflict arbitration based on a dynamic priority weighting algorithm when multiple intents are detected to coexist; an instruction issuance module for issuing control commands to one or more smart home execution devices through a distributed device capability abstraction layer, and switching to local autonomous mode when the wide area network is interrupted; and an optimization module for collecting implicit or explicit user feedback on executed actions, using it as a reward signal to perform incremental reinforcement learning updates on the local prediction model.

[0072] This invention collects environmental data in real time through a distributed sensor array to construct a dynamic scene map. Based on this scene map and historical behavior sequences, a lightweight edge computing model is used to predict potential user intentions and generate at least one action hypothesis to be executed. In multi-user scenarios, intention conflict arbitration is performed using voiceprints, behavior heatmaps, and device usage priority weights. Finally, control actions are executed according to the arbitration results, and the execution results and subsequent user feedback (explicit or implicit) are used as reward signals for reinforcement learning to update the local prediction model in real time.

[0073] On the one hand, based on local edge computing and time-series prediction models, the system can predict user intentions in advance and automatically execute related actions, transforming traditional passive response interaction into a seamless proactive intelligent service, greatly reducing the frequency of user operations and waiting delays. On the other hand, in multi-user coexistence scenarios, through voiceprint recognition, behavior heatmaps, and dynamic priority weighted arbitration algorithms, the system effectively solves the problems of instruction conflict and resource contention, and supports flexible coordination strategies such as time reuse or spatial distribution, ensuring personalized experience and fairness.

[0074] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to this application.

[0075] The above embodiments describe the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Modifications and variations made by those skilled in the art without departing from the spirit and scope of the invention should be within the protection scope of the appended claims.

Claims

1. A method for controlling a smart speaker based on smart home technology, characterized in that, Includes the following steps: S1: Real-time collection of environmental and user status data through a distributed sensor array to construct a dynamic scene map; S2: Based on the dynamic scene map, execute a lightweight temporal prediction model at the local edge of the smart speaker to generate at least one hypothetical action chain representing the user's potential intention; S3: Capture and parse explicit voice commands or implicit behavioral inversion intentions from multiple users in real time. When multiple intentions are detected to coexist, perform conflict arbitration based on a dynamic priority weighting algorithm. S4: Based on the arbitration result, control commands are sent to one or more smart home execution devices through the distributed device capability abstraction layer, and the local autonomous mode is switched when the wide area network is interrupted. S5: Collect implicit and / or explicit user feedback on performed actions, use it as a reward signal, and perform incremental reinforcement learning updates on the local prediction model.

2. The smart speaker control method based on smart home as described in claim 1, characterized in that, Step S1 includes the following sub-steps: S111: Establish a local data acquisition bus; S112: Set the asynchronous acquisition cycle; S113: Perform real-time credibility marking on the acquired raw data stream. If the trigger frequency of a certain sensor exceeds a preset multiple standard deviation of its historical average and is not related to sensors at adjacent locations, then mark the data stream as high noise and reduce its weight in subsequent fusion calculations.

3. The smart speaker control method based on smart home as described in claim 1, characterized in that, Step S2 includes the following sub-steps: S211: Load a temporal convolutional network based on an attention mechanism as a lightweight prediction model; S212: Take the sequence of environment-user joint state tensors covering the past N minutes as the model input and output the probability distribution of high-level behavioral intentions; S213: At the current moment, perform a forward inference calculation on the tensor sequence in the input window to obtain at least one most likely user intent and its probability value at the current moment.

4. The smart speaker control method based on smart home as described in claim 3, characterized in that, Step S2 also includes the following sub-steps: S221: Query the scene-action mapping decision tree forest stored locally to obtain multiple action chains associated with the prediction intent and the success probability of each chain based on historical statistical data; S222: Based on the current environmental parameters, generate specific control parameters for the highest probability intention and generate a preliminary list of hypothetical actions; S223: Assign confidence weights to each hypothetical action based on the user's historical habits. If an action occurs frequently in the past preset period, increase its weight; otherwise, decrease its weight.

5. The smart speaker control method based on smart home as described in claim 1, characterized in that, Step S3 includes the following sub-steps: S311: When an explicit voice command is detected, perform voiceprint recognition to determine the user ID that issued the command, and perform natural language understanding to identify the command intent and slot; S312: When there is no explicit voice command, the implicit intent is inverted through the user's behavior trajectory or device operation sequence; S321: Obtain the dynamic priority weight for each user ID, which is dynamically calculated from the base priority coefficient, the current activity urgency coefficient, and the recent interaction satisfaction coefficient; S322: Determine the conflict type between the explicit instruction and the predicted intent. If it is a direct conflict, the explicit instruction has higher priority than any predicted intent and the hypothetical action of the conflict is canceled immediately. If it is a resource preemption conflict, proceed to the compromise action generation step.

6. The smart speaker control method based on smart home as described in claim 5, characterized in that, The compromise action generation steps include: The time reuse scheme is to inform the user of the allocation plan through voice prompts, first execute short actions of low-priority intents, and then automatically switch to actions of high-priority intents. The spatial routing scheme, based on the voiceprint localization results, routes actions with different intentions to smart speakers or execution devices in different spaces for execution.

7. The smart speaker control method based on smart home as described in claim 1, characterized in that, Step S4 includes the following sub-steps: S421: Periodically sends heartbeat packets to the cloud. If no response is received after a preset number of consecutive attempts, it will automatically switch to local autonomous mode. S422: Start the local rule engine based on finite state machine, use the real-time local sensor data collected in step S1 to execute the reduced prediction logic, and execute the high-frequency rules extracted and pre-cached from the historical cloud model. S423: When the cloud connection is restored, all sensor events, decision logs and execution results stored during the local autonomous mode are timestamped and uploaded to the cloud in batches for incremental training of the model.

8. The smart speaker control method based on smart home as described in claim 1, characterized in that, Step S5 includes the following sub-steps: S511: Collect explicit feedback and / or implicit feedback, wherein explicit feedback includes affirmative or negative words in the user's voice evaluation, as well as the user's manual cancellation of the automatically executed action; implicit feedback includes changes in the user's physiological parameters within a preset time window after the action is executed, device operation reversal behavior, and acoustic feature analysis of speech without semantics. S512: Calculate the temporal difference error, compare the predicted state value before the action with the reward obtained after the action plus the value of the next state; and use the error to update only the last two layers of the local temporal convolutional network model through the backpropagation algorithm, keeping the weights of the previous layers frozen; encrypt and upload the experience replay tuple containing the current scene state, the action executed, the reward obtained and the next scene state to the cloud for periodic training of the global benchmark model.

9. A smart speaker control method based on smart home as described in any one of claims 1 to 8, characterized in that, It also includes the following steps: S6: Monitors safety and user physiological abnormalities, and unconditionally interrupts the automatic control process in case of abnormalities.

10. A smart speaker control system based on smart home technology, used to implement the smart speaker control method according to any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to collect environmental and user status data in real time through a distributed sensor array and to construct a dynamic scene map; The action generation module is used to execute a lightweight temporal prediction model at the local edge of the smart speaker to generate at least one hypothetical action chain representing the user's potential intention. The conflict arbitration module is used to capture and parse explicit voice commands or implicit behavioral inversion intentions of multiple users in real time. When multiple intentions are detected to coexist, conflict arbitration is performed based on a dynamic priority weighting algorithm. The command issuance module is used to issue control commands to one or more smart home execution devices through the distributed device capability abstraction layer, and switch to local autonomous mode when the wide area network is interrupted. The optimization module is used to collect implicit or explicit user feedback on the actions performed, and use it as a reward signal to perform incremental reinforcement learning updates on the local prediction model.

Images