Intelligent control method and device of air conditioner, electronic equipment and storage medium

Abstract

The application relates to the technical field of intelligent control of air conditioners, and discloses an intelligent control method and device of an air conditioner, an electronic device and a storage medium, wherein the method comprises: analyzing radar data through a first offline AI model to realize the perception of a posture, analyzing a Bluetooth dynamic feature through a second offline AI model to realize the identification of an authorized identity and an active intention, fusing the identification results of the two and calculating through a confidence score model, comprehensively combining the double evidences of physical existence and identity signals, and generating a control instruction for the air conditioner.The application has the beneficial technical effects that the false triggering rate caused by pets, mobile interference or only equipment left is reduced, the user privacy safety is guaranteed, network delay is eliminated, and a personalized strategy more in line with the user habits is provided.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for air conditioners, and more particularly to an intelligent control method, device, electronic equipment, and storage medium for air conditioners. Background Technology

[0002] Currently, intelligent control of air conditioners mainly relies on single-sensor technology or cloud-based solutions. Single-sensor-based solutions, such as human infrared (PIR) or simple radar, can only detect movement and cannot distinguish between interference from users and pets, easily leading to false triggers. In addition, solutions relying on networks and the cloud have risks of response delays and privacy leaks.

[0003] Therefore, existing technologies struggle to make accurate judgments in offline environments, resulting in inaccurate air conditioning control and a poor user experience. Summary of the Invention

[0004] Based on this, it is necessary to propose an intelligent control method, device, electronic equipment, and storage medium for air conditioners to address the existing intelligent control problems.

[0005] A method for intelligent control of an air conditioner, the method comprising:

[0006] Periodically collect target feature data of the target perception area through the preset human interaction detection module, and collect Bluetooth signal strength data through the Bluetooth module;

[0007] Based on the target feature data, human body recognition is performed through a preset first offline AI model to obtain a first recognition result that includes human body presence determination and posture classification.

[0008] Based on the Bluetooth signal strength data, dynamic feature analysis is performed using a preset second offline AI model to obtain a second recognition result representing the authorized terminal's intention to approach.

[0009] The first recognition result and the second recognition result are fused to obtain a fused recognition result;

[0010] The fusion recognition result is input into a preset confidence scoring model to calculate the comprehensive confidence score, and the comprehensive confidence score is compared with a preset threshold to obtain the comparison result.

[0011] Control commands for the air conditioner are generated based on the first identification result, the second identification result, and the comparison result.

[0012] Further, the step of performing human body recognition based on the target feature data using a preset first offline AI model to obtain a first recognition result including human presence determination and posture classification includes:

[0013] Based on the target feature data, a preset first offline AI model is used to identify whether a human target exists within the target perception area;

[0014] If a human target exists within the target perception area, then extract the human target data from the target feature data;

[0015] The human target data is subjected to posture recognition by a preset first offline AI model to obtain posture classification results, thereby obtaining the first recognition result.

[0016] Further, the step of obtaining a second recognition result representing the authorized terminal's intention to approach based on the Bluetooth signal strength data and through dynamic feature analysis using a preset second offline AI model includes:

[0017] Determine whether the Bluetooth signal strength data originates from a bound authorized terminal;

[0018] If the signal originates from a bound authorized terminal, the dynamic change characteristics of the Bluetooth signal strength within a preset time window are extracted.

[0019] Based on the aforementioned dynamic change characteristics, the proximity behavior pattern of the authorized terminal is analyzed using a preset second offline AI model;

[0020] The confidence level representing the user's intention to actively approach is output as the second identification result based on the proximity behavior pattern.

[0021] Further, the step of generating control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result includes:

[0022] Determine whether there is a human target in the first identification result, and whether the Bluetooth signal strength data in the second identification result belongs to the bound authorized terminal;

[0023] When the first identification result determines that a human target exists, and the second identification result confirms that the Bluetooth signal originates from a bound authorized terminal, the posture features in the first identification result, the dynamic feature confidence level in the second identification result, and the environmental context features are combined to form a multi-dimensional feature vector as the fused identification result.

[0024] Further, the step of generating control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result includes:

[0025] When the overall confidence level is lower than a first preset threshold, a command is generated to control the air conditioner to turn off.

[0026] When the overall confidence level reaches or exceeds the second preset threshold, a command to control the air conditioner to turn on is generated; wherein the second preset threshold is greater than the first preset threshold;

[0027] When the first identification result determines that the human target has left, but the Bluetooth signal strength data indicates that the authorized terminal is still within the preset range, an instruction is generated to control the air conditioner to enter the delayed shutdown process.

[0028] Further, the step of generating control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result includes:

[0029] If the first identification result determines that there is no human target and the second identification result determines that there is no authorized terminal signal, then a first instruction is generated to control the air conditioner to turn off or remain off.

[0030] If the first identification result determines that a human target exists, and the overall confidence level is lower than the first threshold, then a second instruction is generated to control the air conditioner to turn off or remain off.

[0031] If the first identification result determines that a human target exists, the second identification result determines that the Bluetooth signal strength data comes from an authorized terminal, and the overall confidence level reaches a second threshold, then a third instruction is generated to control the air conditioner to turn on or switch to a preset mode; wherein, the second threshold is greater than the first threshold.

[0032] If the first identification result determines that there is no human target, and the second identification result determines that there is an authorized terminal signal, then a fourth instruction is generated to control the air conditioner to enter the delayed shutdown process.

[0033] Furthermore, after the step of generating control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result, the method further includes:

[0034] Receive user operation data for the air conditioner and the corresponding Bluetooth signal strength data when performing control operations to form positive samples;

[0035] Using the positive samples, the parameters of the confidence scoring model are optimized through an online learning algorithm to obtain an optimized confidence scoring model.

[0036] An intelligent control device for an air conditioner, the device comprising:

[0037] The acquisition module is used to periodically acquire target feature data of the target perception area through a preset human interaction detection module, and to acquire Bluetooth signal strength data through a Bluetooth module.

[0038] The recognition module is used to perform human body recognition based on the target feature data and through a preset first offline AI model to obtain a first recognition result including human body presence determination and posture classification.

[0039] The analysis module is used to perform dynamic feature analysis based on the Bluetooth signal strength data through a preset second offline AI model to obtain a second recognition result representing the intention of the authorized terminal to approach.

[0040] The fusion module is used to fuse the first recognition result and the second recognition result to obtain a fused recognition result;

[0041] The calculation module is used to input the fusion recognition result into a preset confidence score model to calculate the comprehensive confidence score, and compare the comprehensive confidence score with a preset threshold to obtain a comparison result;

[0042] The generation module is used to generate control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result.

[0043] An electronic device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0044] Periodically collect target feature data of the target perception area through the preset human interaction detection module, and collect Bluetooth signal strength data through the Bluetooth module;

[0045] Based on the target feature data, human body recognition is performed through a preset first offline AI model to obtain a first recognition result that includes human body presence determination and posture classification.

[0046] Based on the Bluetooth signal strength data, dynamic feature analysis is performed using a preset second offline AI model to obtain a second recognition result representing the authorized terminal's intention to approach.

[0047] The first recognition result and the second recognition result are fused to obtain a fused recognition result;

[0048] The fusion recognition result is input into a preset confidence scoring model to calculate the comprehensive confidence score, and the comprehensive confidence score is compared with a preset threshold to obtain the comparison result.

[0049] Control commands for the air conditioner are generated based on the first identification result, the second identification result, and the comparison result. A computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0050] Periodically collect target feature data of the target perception area through the preset human interaction detection module, and collect Bluetooth signal strength data through the Bluetooth module;

[0051] Based on the target feature data, human body recognition is performed through a preset first offline AI model to obtain a first recognition result that includes human body presence determination and posture classification.

[0052] Based on the Bluetooth signal strength data, dynamic feature analysis is performed using a preset second offline AI model to obtain a second recognition result representing the authorized terminal's intention to approach.

[0053] The first recognition result and the second recognition result are fused to obtain a fused recognition result;

[0054] The fusion recognition result is input into a preset confidence scoring model to calculate the comprehensive confidence score, and the comprehensive confidence score is compared with a preset threshold to obtain the comparison result.

[0055] Control commands for the air conditioner are generated based on the first identification result, the second identification result, and the comparison result.

[0056] The beneficial effects of this invention are as follows: By analyzing radar data through a first offline AI model, attitude perception is achieved. Simultaneously, by analyzing Bluetooth dynamic characteristics through a second offline AI model, the authorized identity and active intent are identified. The identification results of the two are fused and calculated using a confidence scoring model, combining dual evidence of physical presence and identity signals. This reduces the false trigger rate caused by pets, mobile interference, or simply device remnants. The first offline AI model, the second offline AI model, and the confidence scoring model all run locally, enabling the generation of control commands directly based on the first identification result, the second identification result, and the comparison result without relying on a cloud network. This ensures user privacy and security, eliminates network latency, and provides a more personalized strategy that aligns with user habits. Attached Figure Description

[0057] 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 these drawings without creative effort.

[0058] in:

[0059] Figure 1 This is an application environment diagram of an intelligent control method for air conditioning in one embodiment;

[0060] Figure 2 This is a flowchart of an intelligent control method for an air conditioner in one embodiment;

[0061] Figure 3 This is a structural block diagram of an intelligent control device for an air conditioner in one embodiment;

[0062] Figure 4 This is a structural block diagram of an electronic device in one embodiment. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Figure 1 This is a diagram illustrating the intelligent control application environment of an air conditioner in one embodiment. (Refer to...) Figure 1 The intelligent control method for this air conditioner is applied to an intelligent control system. This intelligent control system includes a terminal 110 and an air conditioner 120. The terminal 110 and the air conditioner 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; the mobile terminal can be at least one of a mobile phone, tablet, or laptop. The air conditioner 120 can be a standalone air conditioner or a cluster of multiple air conditioners. The terminal 110 provides Bluetooth signal strength data, and the air conditioner 120 collects Bluetooth data and target feature data, and generates corresponding control commands.

[0065] like Figure 2 As shown, in one embodiment, an intelligent control method for an air conditioner is provided. This method can be applied to both a terminal and an air conditioner; this embodiment illustrates its application to a terminal. The intelligent control method for the air conditioner specifically includes the following steps:

[0066] S1: Periodically collect target feature data of the target perception area through the preset human interaction detection module, and collect Bluetooth signal strength data through the Bluetooth module;

[0067] S2: Based on the target feature data, human body recognition is performed through a preset first offline AI model to obtain a first recognition result that includes human body presence determination and posture classification;

[0068] S3: Based on the Bluetooth signal strength data, dynamic feature analysis is performed using a preset second offline AI model to obtain a second recognition result representing the authorized terminal's intention to approach;

[0069] S4: Fuse the first recognition result with the second recognition result to obtain a fused recognition result;

[0070] S5: Input the fusion recognition result into the preset confidence score model to calculate the comprehensive confidence score, and compare the comprehensive confidence score with the preset threshold to obtain the comparison result;

[0071] S6: Generate control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result.

[0072] As described in step S1 above, target feature data of the target perception area is periodically collected through a preset human interaction detection module, and Bluetooth signal strength data is collected through a Bluetooth module. Raw observation data is obtained simultaneously or in a defined sequence from both sensing paths according to a preset period (configurable, e.g., a sampling rate of 20Hz per frame or other suitable values). The human interaction detection module scans the perception area and outputs raw echo features, such as target distance, velocity (Doppler), micro-motion amplitude, echo energy distribution, or distance-velocity spectrum; a timestamp is added at the acquisition end, and the data is transmitted to the local processor in frames. The Bluetooth module periodically or passively listens to the broadcast and connection information of the bound device, collecting dynamic information such as signal strength, device ID, connection / disconnection events, and scan timestamps. This requires consideration of frame synchronization of the hardware interface, sampling rate matching, and primary data integrity verification, as well as basic anti-collision and anti-frame loss measures. Specifically, the human interaction detection module includes a radar submodule and an HID (Human Interaction Detection System) submodule. The radar submodule can be used to make a preliminary judgment on the presence of human information. When a human is present, the HID submodule is activated to collect the human's posture information. In addition, in a preferred embodiment, activating the HID submodule can also be combined with the matching result of the Bluetooth signal. That is, the conditions for activating the HID submodule to perform detection are met only when the radar submodule detects a human and the detected Bluetooth signal is a signal emitted by a bound authorized terminal.

[0073] As described in step S2 above, based on the target feature data, human body recognition is performed using a preset first offline AI model to obtain a first recognition result including human presence determination and posture classification. On the local main controller, the offline-deployed AI model performs deep semantic analysis on the radar preprocessed data. First, the radar data undergoes time-domain / frequency-domain preprocessing, such as bandpass filtering, window functions, frame synchronization, background static component suppression, and noise thresholding, to extract micro-motion features, multi-frame range-Doppler maps, or temporal micro-motion waveforms. Then, the features from multiple consecutive frames are input into the first offline AI model (such as a lightweight 1D-CNN, compressed convolutional network, or small temporal network). The model outputs two key types of information: one is the human presence determination (present / non-present / uncertain), and the other is the posture classification when the presence is determined (e.g., standing, walking, sitting, sleeping, etc.). This step is usually accompanied by confidence output and a time smoothing mechanism (such as sliding window voting, confidence filtering) to reduce instantaneous misjudgments, and the confidence level and category of the model output are recorded for subsequent fusion. During implementation, it is also necessary to set the model input length, inference latency limit, and memory usage limit to ensure that online real-time performance matches embedded resources. The first and second offline AI models are lightweight artificial intelligence models deployed on the local control chip of the air conditioner, which can perform inference independently without data interaction with the cloud.

[0074] As described in step S3 above, based on the Bluetooth signal strength data, dynamic feature analysis is performed using a preset second offline AI model to obtain a second identification result representing the authorized terminal's approach intention. The presence and behavioral intention of the authorized device (such as a user's mobile phone) are assessed based on the Bluetooth signal. First, the source of the Bluetooth data is determined locally to be a bound or authorized terminal (device ID matching). During matching, dynamic features such as RSSI sequence, RSSI rate of change, peak value, fluctuation variance, and short-term surges or decays within the time window are extracted. Behavioral pattern analysis is performed using a second offline AI model (such as a lightweight LSTM, GBDT, or a small time series classifier). The model outputs indicators representing "active approach confidence" or "near-field dwell confidence," and can provide an inference as to whether the device is in a "held / left behind / moving" state.

[0075] As described in step S4 above, the first recognition result and the second recognition result are fused to obtain a fused recognition result. The fusion step is not a simple concatenation, but rather a dynamic construction of a multi-dimensional feature vector based on the semantics and confidence of the two recognition results, followed by unified representation. The fusion module receives the first recognition result (human presence marker, posture category, corresponding confidence and time window statistics) and the second recognition result (authorized terminal matching status, active approach confidence, RSSI statistical features, and time label). Fusion can employ a combination of rule-based processing (e.g., setting base scores based on prior rules) and machine learning methods: the rule layer assigns initial weights to key combinations (e.g., "human presence and device ID matching"); the machine learning layer takes the original and statistical features of both paths as input and outputs a more stable fused representation. During the fusion process, feature alignment (timestamp consistency), missing value handling (e.g., temporary Bluetooth loss), and dynamic adjustment of feature weights (based on the scenario or historical behavior) must also be considered. The final fusion recognition result is structured data containing multi-dimensional feature vectors and synthetic labels (such as "high probability of authorized presence / suspected stranger / environmental interference") and corresponding fusion confidence scores, providing a unified input for the confidence scoring model and decision-making unit.

[0076] As described in step S5 above, the fused recognition result is input into a preset confidence scoring model for calculation to obtain a comprehensive confidence score. This comprehensive confidence score is then compared with a preset threshold to obtain a comparison result. The confidence scoring model quantitatively evaluates the fused feature vector and outputs a comprehensive confidence score between 0 and 1, representing the degree of credibility of "authorized personnel present." The scoring model may include a rule layer (base score accumulation, confidence lower bound) and a learning layer (logistic regression, random forest, or neural network). The learning layer can calculate the final weights based on historical labeled data or online fine-tuning results. During scoring, the confidence score of each input feature, time consistency, environmental noise indicators, and anomaly detection results are considered. For example, if radar shows strong interference but Bluetooth confidence is high, the scoring strategy will be adjusted. After calculation, the comprehensive confidence score is compared with a set of preset thresholds (e.g., low threshold for denial, high threshold for confirmation, and intermediate range for observation) to generate a categorical comparison result (e.g., low / medium / high or pass / fail / pending confirmation). The comparison results include timestamps and diagnostic information, enabling the state machine to take appropriate control actions or enter observation mode in the next step. Specifically, the rule layer generates a base score based on preset rules (e.g., "human presence +10 points, Bluetooth authorization +30 points"). The learning layer is a logistic regression model that takes a fused feature vector (e.g., pose confidence, RSSI rate of change, etc.) as input and outputs an adjustment coefficient. The final comprehensive confidence score is normalized to 0-1 after multiplying the base score by the adjustment coefficient. The learning layer is trained using offline collected labeled data to obtain the confidence scoring model.

[0077] As described in step S6 above, control commands for the air conditioner are generated based on the first identification result, the second identification result, and the comparison result. The decision-making unit is typically implemented using a state machine or strategy engine, formulating a refined control strategy based on a combination of three elements: for example, when a human body is present and in a sleeping posture with high overall confidence, an energy-saving comfort command to "increase the temperature and decrease the fan speed" is generated; when a human body is detected but the device is not matched and the overall confidence is low, a conservative strategy of "keeping the air conditioner off and recording the event" is adopted; when a human body briefly leaves but the Bluetooth device is still present and the confidence is in the middle range, a "delayed shutdown" is executed to avoid frequent switching. The generated control commands may include on / off, mode switching, temperature and fan speed settings, and delay parameters, along with execution priorities and safety constraints (such as maximum / minimum temperature difference limits). Furthermore, the decision-making process can trigger log recording, alarms, or user notifications, and new behavioral samples can be sent back to the local learning module for subsequent model fine-tuning.

[0078] In one embodiment, step S2, which involves performing human body recognition based on the target feature data using a preset first offline AI model to obtain a first recognition result including human presence determination and posture classification, includes:

[0079] S201: Based on the target feature data, identify whether a human target exists within the target perception area using a preset first offline AI model;

[0080] S202: If a human target exists within the target perception area, extract the human target data from the target feature data;

[0081] S203: The human target data is subjected to posture recognition by a preset first offline AI model to obtain posture classification results, thereby obtaining the first recognition result.

[0082] As described in step S201 above, based on the target feature data, a preset first offline AI model is used to identify whether a human target exists within the target perception area. First, the raw target feature data acquired by the radar is preprocessed, including time-domain / frequency-domain filtering to suppress noise, static background echo removal to highlight micro-motion signals, frame synchronization, and timestamp correction to ensure consistency across multiple frames. After preprocessing, an input representation is constructed, commonly in the form of range-Doppler spectra, micro-motion temporal waveforms, or energy distribution maps for multiple consecutive frames. This input is then fed into a locally deployed first offline AI model (e.g., a lightweight temporal convolutional network or a small CNN+pooling structure). After training, the model can capture the spectral features and spatiotemporal distribution of human micro-motions within a multi-frame context. The model outputs a judgment value and corresponding confidence level for the presence of a human body, and can perform preliminary segmentation and confidence ranking for multi-target scenes. It should also include post-processing steps, such as voting based on sliding windows or confidence smoothing to reduce instantaneous false alarms, and distinguishing between short-term interference and real human presence based on threshold and time consistency rules, so as to provide stable presence judgment results for subsequent target extraction and pose recognition.

[0083] As described in step S202 above, if a human target exists within the target perception area, the human target data is extracted from the target feature data. After confirming the existence of a human target, a subset of features related to the human body is separated from the continuous perception data for subsequent pose analysis. First, based on the target position, distance range, or Doppler peak value output in S201, the region of interest is determined, and the relevant frames are cropped or windowed in time and space. For multi-target scenarios, target tracking and data association algorithms (such as Kalman filtering or simple Hungarian matching) need to be executed to maintain the continuous trajectory of each target and avoid confusion between different target features. Subsequently, the extracted target data is standardized, including amplitude normalization, temporal length alignment (interpolation or truncation), and necessary denoising and feature enhancement (such as short-time Fourier transform, envelope extraction, micro-motion amplitude statistics, etc.). The timestamp, target ID, and confidence information of the extraction window are recorded, and a compensation strategy is triggered in abnormal situations (such as target occlusion or signal abrupt change), such as continuing the data of the previous window or entering observation mode, to ensure the continuity and availability of the data supplied to the pose recognition module.

[0084] As described in step S203 above, the human target data is subjected to posture recognition using a preset first offline AI model to obtain posture classification results, thereby obtaining the first recognition result. Using the target data extracted and preprocessed in S202 as input, the posture recognition sub-model is called to perform fine classification of the human body's current or short-term activity state. Posture recognition models typically employ structures with strong temporal awareness (such as lightweight LSTM, 1D-CNN temporal modules, or temporal convolution + attention mechanisms) to capture the patterns of micro-motion spectrum changes over time, thereby distinguishing between sitting, sleeping, standing, and walking states. The model outputs the posture category and corresponding confidence score, and can simultaneously output the time point of posture switching or a smoothed probability curve. To improve robustness, posture recognition should also incorporate confidence thresholds and temporal consistency checks: posture transitions are only confirmed when the category confidence score exceeds the threshold for several consecutive frames, avoiding misjudgments caused by instantaneous noise. Simultaneously, the posture recognition module must meet embedded operating constraints, controlling inference latency and memory usage, and employing quantization, pruning, or window downsampling techniques when necessary to balance real-time performance and accuracy. Finally, the presence determination of the human body, the posture classification, and their respective confidence and time information are merged to form a complete first recognition result, which is then used by the fusion and decision-making module.

[0085] In one embodiment, step S3, which involves performing dynamic feature analysis based on the Bluetooth signal strength data using a preset second offline AI model to obtain a second recognition result representing the authorized terminal's intention to approach, includes:

[0086] S301: Determine whether the Bluetooth signal strength data originates from a bound authorized terminal;

[0087] S302: If the signal originates from a bound authorized terminal, extract the dynamic change characteristics of the Bluetooth signal strength within a preset time window;

[0088] S303: Based on the aforementioned dynamic change characteristics, the proximity behavior pattern of the authorized terminal is analyzed using a preset second offline AI model;

[0089] S304: Output the confidence level representing the user's intention to actively approach based on the proximity behavior pattern as the second identification result.

[0090] As described in step S301 above, it is determined whether the Bluetooth signal strength data originates from a bound authorized terminal. First, a whitelist of authorized terminals is maintained locally (e.g., a mapping between device MAC or device ID and user account), and compared based on device identification information reported in Bluetooth scanning or connection events. To balance privacy and security, typically only pre-paired / bound and recorded identifiers are compared, rather than those uploaded to the cloud. The determination process includes extracting the identifier field from received broadcast or connection packets, verifying timestamps and signal integrity, and protecting against short-term repetitive or spoofing attempts (e.g., detecting abnormal broadcast frequencies or unresolved identifiers). In cases with multiple authorized terminals, multiple matching entries can be recorded simultaneously and sorted by priority (e.g., primary device priority, most recent activity time priority). For cases where a match cannot be made or the match is uncertain, it should be marked as "unauthorized or pending confirmation," and this flag should be passed to subsequent steps to determine whether to continue dynamic feature extraction or enter an observation / alarm strategy.

[0091] As described in step S302 above, if the signal originates from a bound authorized terminal, the dynamic change characteristics of the Bluetooth signal strength within a preset time window are extracted. After confirming the authorized terminal, the RSSI sequence and related broadcast / connection events of the device are collected locally within a preset time window (e.g., 1s, 3s, or 5s, configurable). To improve feature robustness, the original RSSI needs to be preprocessed: filtering out isolated outliers (short-term spikes), low-pass filtering for smoothing, window-based denoising, and normalization; if necessary, corrections are made in conjunction with scanning time slots and antenna directivity. Subsequently, a set of dynamic statistical and temporal features are extracted from the processed RSSI sequence, such as instantaneous rate of change (delta), maximum / minimum values, peak occurrence time, variance, short-term rise slope, moving average, energy spectrum characteristics, and dwell time; the regression coefficient of RSSI with time can also be calculated to characterize the approach / departure trend. If the environment is complex, multipath or signal jitter indicators and scanning interval distribution can be added as auxiliary features. All features are time-stamped and cached for use by the second offline AI model.

[0092] As described in step S303 above, based on the dynamic change characteristics, the proximity behavior pattern of the authorized terminal is analyzed using a preset second offline AI model. The temporal and statistical features obtained in S302 are input into the second offline AI model deployed locally for pattern recognition. This model can be a lightweight temporal model (such as a quantized LSTM or 1D-CNN) or a tree model combined with temporal feature engineering, specifically trained to distinguish behavioral patterns such as "actively approaching," "leaving," and "staying / lingering." During inference, the model considers the temporal consistency and confidence output of features, and usually uses thresholding and post-processing strategies (such as continuous frame consistency verification and short-term smoothing) to reduce erroneous judgments caused by noise. In addition, the model can use environmental context (such as time period and historical behavior patterns) for lightweight weighting, outputting behavior pattern labels and their confidence levels. To address the non-ideal nature of BLE signals, the model training and deployment phases include multi-scenario data to enhance adaptability, and the model inference meets the latency and storage constraints of embedded devices.

[0093] As described in step S304 above, the confidence level representing the user's intention to actively approach is output as the second identification result based on the approach behavior pattern. The model's behavior pattern judgment is converted into a quantitative indicator that can be used for fusion—the confidence level of active approach. This confidence level is comprehensively corrected based on the model's output probability, the completeness of input features (such as whether the samples are sufficient and whether there is packet loss), temporal consistency (whether the approach trend is continuously displayed in several consecutive windows), and anomaly detection results, ultimately forming a confidence level value in the range of 0-1. At the same time, the second identification result also includes auxiliary information (such as the authorized terminal ID, the most recent RSSI value, the timestamp, and the behavior pattern label) for the fusion module to refer to during comprehensive judgment. If the confidence level is low or the data is incomplete, a "pending confirmation" status can be output and an enhanced scan or delayed observation strategy can be triggered; if the confidence level is high, the result will serve as strong evidence to increase the weight of the "authorized presence" judgment in the fusion and confidence scoring stages. All processing is completed locally to ensure privacy and real-time performance.

[0094] In one embodiment, step S4, which generates control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result, includes:

[0095] S401: Determine whether there is a human target in the first identification result, and whether the Bluetooth signal strength data in the second identification result belongs to the bound authorized terminal;

[0096] S402: When the first identification result determines that there is a human target, and the second identification result confirms that the Bluetooth signal comes from the bound authorized terminal, the posture features in the first identification result, the dynamic feature confidence in the second identification result, and the environmental context features are combined to form a multi-dimensional feature vector as the fusion identification result.

[0097] As described in step S401 above, it is determined whether a human target exists in the first identification result, and whether the Bluetooth signal strength data identified in the second identification result belongs to a bound authorized terminal. Simultaneously, the first identification result from the radar sensing channel and the second identification result from the Bluetooth sensing channel are analyzed in parallel. For the first identification result, the validity of its "human presence determination" needs to be verified. This verification is not only based on a single output, but also examines its continuity within a time sliding window, the trend of confidence changes, and the quality indicators of the underlying radar data (such as signal-to-noise ratio). At the same time, it must be ensured that this determination has excluded typical known non-human interference patterns (such as periodic fan rotation or plant swaying). For the second identification result, a strict identity verification of the Bluetooth signal source is required. The specific process includes: extracting the device identifier from the received broadcast or connection request packet and accurately comparing it with the locally securely stored list of authorized devices; to prevent identification difficulties caused by randomized device MAC addresses, other fingerprint features of the device (such as broadcast intervals and service UUID lists) are also needed for auxiliary verification. Several edge cases also need to be handled: for example, when the confidence level for determining the presence of a human body falls within a set fuzzy range (e.g., 0.4 to 0.6), the system will not immediately reject the application, but will instead perform trend analysis based on the state at previous moments; when the Bluetooth signal is intermittent or weak, leading to uncertainty in matching, a signal quality assessment will be initiated, potentially triggering an active scanning request. Only when both conditions clearly meet the preset verification criteria will the system determine whether it is qualified to perform subsequent feature-level fusion; otherwise, it will switch to the low-confidence processing branch or maintain the current system state. This rigorous admission judgment constitutes the first quality and reliability checkpoint in the entire control decision-making process.

[0098] As described in step S402 above, when the first identification result determines the presence of a human target and the second identification result confirms that the Bluetooth signal originates from a bound authorized terminal, the posture features in the first identification result, the dynamic feature confidence level in the second identification result, and the environmental context features are combined to form a multi-dimensional feature vector as the fused identification result. Its execution prerequisite is that the verification in S401 yields a positive result. When it is confirmed that the current scenario is "radar detects a human body" and "Bluetooth signal originates from an authorized terminal," a high-dimensional feature construction stage is entered. The goal of this stage is to integrate heterogeneous data from different physical domains with different statistical characteristics into a unified mathematical representation rich in semantic information, namely, the "fused identification result." Specifically, fine-grained posture-related features are first extracted from the first identification result. This includes not only the final classification label (such as "sitting"), but also its multi-dimensional probability distribution, the key frequency energy of the micro-Doppler spectrum, the duration of the posture, and the instantaneous features of posture switching. Secondly, a quantitative representation of Bluetooth dynamic features is extracted from the second recognition result, primarily concerning the confidence value of "active approach intention," supplemented by temporal statistics such as the gradient of this confidence change and the stability measure of the RSSI sequence. Finally, environmental context features are simultaneously injected, which may include the current time (used to determine the sleep / rest period), indoor baseline temperature, recent air conditioning operating mode, and the environmental noise level estimated by the human interaction detection module. Before combination, all the above features are standardized to eliminate the influence of dimensions and strictly time-aligned to ensure they correspond to the same observation time or the same time window. Subsequently, these features are systematically combined into a fixed-length multidimensional feature vector. The structure of this vector considers the physical correlation between features; for example, the radar's micro-motion amplitude features and the Bluetooth RSSI rate of change features are correlated and encoded in the time dimension to enhance the model's ability to recognize the "human-carried device movement" pattern. The final multidimensional feature vector, the fused recognition result, provides a comprehensive and structured high-level input for the subsequent confidence scoring model. The environmental context includes, but is not limited to, a combination of one or more of the following information: timestamps, indoor and outdoor temperatures, humidity, light intensity, current air conditioning operating mode, historical control data, and environmental noise levels and interference source information estimated by the human interaction detection module.

[0099] In one embodiment, step S6, which generates control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result, includes:

[0100] S601: When the overall confidence level is lower than the first preset threshold, a command to control the air conditioner to turn off is generated;

[0101] S602: When the overall confidence level reaches or exceeds the second preset threshold, a command to control the air conditioner to turn on is generated; wherein the second preset threshold is greater than the first preset threshold;

[0102] S603: When the first identification result determines that the human target has left, but the Bluetooth signal strength data indicates that the authorized terminal is still within the preset range, an instruction is generated to control the air conditioner to enter the delayed shutdown process.

[0103] As described in step S601 above, when the overall confidence level is lower than the first preset threshold, a command to control the air conditioner to turn off is generated. The trigger condition is that the overall confidence level output by the fusion confidence scoring model is lower than the first preset threshold (e.g., 0.3 or lower). This usually means that there is a contradiction or insufficiency in the evidence chain between the radar and Bluetooth, such as the radar detecting a slight movement but Bluetooth not matching, or a weak Bluetooth signal matching but the radar determining it as interference. In this case, instead of simply turning off the air conditioner directly, a decision-making process involving multiple verification layers is initiated: First, the first identification result is reviewed. If the radar clearly determines it to be a "non-human target" (such as a pet or fan), a shutdown command is immediately generated; if the radar determines it to be a "human" target but the confidence level is contradictory, the details of the second identification result are queried, such as whether the Bluetooth signal is from a new device or whether the signal is intermittent. Secondly, referring to recent status history (such as whether there has been authorization confirmation in the past 5 minutes), if it is a brief anomaly, it may enter the "observation and waiting" state instead of shutting down immediately. The final shutdown command may be accompanied by gradual stopping logic (such as switching to the air supply mode first and then shutting down) to avoid sudden temperature changes. At the same time, the event will be marked as a "low confidence event" and recorded in the local log for subsequent offline analysis or model optimization.

[0104] As described in step S602 above, when the overall confidence level reaches or exceeds a second preset threshold, a command to control the air conditioner to turn on is generated; wherein, the second preset threshold is greater than the first preset threshold. The core condition is that the overall confidence level reaches or exceeds a higher second preset threshold (e.g., 0.75), which requires a high degree of coordination between radar and Bluetooth evidence. For example, the radar continuously detects human features with stable posture, while the Bluetooth dynamic features exhibit a clear active approach mode and a continuous and stable signal. Once the condition is met, the air conditioner is not simply turned on, but a refined control sequence is executed: First, a preset personalized mode template is selected based on the posture classification (e.g., sleeping, sitting, active) in the first identification result; for example, if the posture is "sleeping," the command is "turn on to 26℃, low fan speed, silent mode"; if it is "active," the command is "turn on to 24℃, automatic fan speed." Second, the system fine-tunes the parameters based on the environmental context (e.g., current indoor temperature, time) and sends a coded command with verification through the infrared control module. To prevent frequent switching, the triggering of this command is typically accompanied by a debouncing delay (e.g., the confidence level is higher than the threshold for two consecutive cycles), and after the initial activation, the first confidence threshold is briefly increased to maintain operational stability. This demonstrates the complete intelligent chain of this invention from "perception" to "personalized execution," achieving precise energy saving while ensuring user experience.

[0105] As described in step S603 above, when the first identification result determines that the human target has left, but the Bluetooth signal strength data indicates that the authorized terminal is still within the preset range, an instruction is generated to control the air conditioner to enter a delayed shutdown process. The triggering condition is: the first identification result clearly shows that the human target has left the radar's effective sensing area (e.g., multiple consecutive frames of distance values ​​exceeding the threshold or the disappearance of micro-motion), but the second identification result continuously detects that the authorized terminal signal is stably present within the preset near-field range (e.g., RSSI strength remains above -70dBm and fluctuates smoothly). At this time, the air conditioner will not be immediately shut off, but rather a configurable delayed shutdown process will be initiated. This process first dynamically calculates the delay time (e.g., 5-15 minutes) based on the user's historical behavior model (e.g., statistics on the duration of short daily absences) and environmental factors (e.g., indoor-outdoor temperature difference). During the delay, two sensors are continuously monitored: if the radar re-detects the human body or the Bluetooth signal dynamics indicate rapid approach, the delay is immediately canceled and operation resumes; if there are no signs of return before the delay ends, a smooth shutdown instruction is executed (e.g., first increasing the set temperature by 2°C, then shutting off the compressor). At the same time, the characteristics of this "exit-return" pattern were recorded to optimize the user personalization model.

[0106] In one embodiment, step S4, which generates control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result, includes:

[0107] S411: If the first identification result determines that there is no human target and the second identification result determines that there is no authorized terminal signal, then a first instruction to control the air conditioner to turn off or remain off is generated.

[0108] S412: If the first identification result determines that there is a human target, and the overall confidence level is lower than the first threshold, then a second instruction is generated to control the air conditioner to turn off or remain off.

[0109] S413: If the first identification result determines that there is a human target, the second identification result determines that the Bluetooth signal strength data comes from an authorized terminal, and the comprehensive confidence level reaches the second threshold, then a third instruction is generated to control the air conditioner to turn on or switch to a preset mode; wherein, the second threshold is greater than the first threshold;

[0110] S414: If the first identification result determines that there is no human target, and the second identification result determines that there is an authorized terminal signal, then a fourth instruction is generated to control the air conditioner to enter the delayed shutdown process.

[0111] As described in step S411 above, if the first identification result determines that there is no human target and the second identification result determines that there is no authorized terminal signal, a first command to control the air conditioner to turn off or remain off is generated. When the first identification result determines, based on radar data analysis, that there is no human target (e.g., the radar signal only contains environmental background noise, static object reflection, or periodic interference patterns), and the second identification result simultaneously confirms that there is no Bluetooth signal from any bound authorized terminal (including no broadcast signal or signal strength below the minimum detection threshold), a first command will be generated to control the air conditioner to turn off or remain off. This decision-making process includes a multi-layer verification mechanism: First, the radar data needs to be analyzed continuously over multiple frames (e.g., within 10 seconds) to ensure that the target absence determination has temporal stability and avoid misjudgment due to brief signal attenuation; second, the Bluetooth module needs to perform active scanning and eliminate occasional interference signals to confirm that all terminals in the authorized device list are unresponsive. Before executing the command, the environmental context (such as recent air conditioner operation records and indoor temperature change rate) will be taken into account for final confirmation: if the air conditioner is already off, the first command is "keep off"; if the air conditioner is running, a smooth shutdown command will be generated (such as switching to ventilation mode first and then shutting down completely).

[0112] As described in step S412 above, if the first identification result determines the presence of a human target and the overall confidence level is lower than the first threshold, a second instruction is generated to control the air conditioner to turn off or remain off. Handling ambiguous scenarios such as "the presence of an unauthorized target" or "insufficient confidence" is a key logic for preventing false triggering of the system. When the first identification result clearly determines the presence of a human target (radar detects features matching the micro-motion spectrum of a human body), but the overall confidence level obtained after fusion is lower than the first threshold (e.g., 0.4), a second instruction will be generated to maintain the air conditioner in its off state. Low confidence usually stems from two situations: first, although the radar detects human features, the second identification result indicates that the Bluetooth signal is not matched or the matched signal is abnormal (e.g., the signal jumps drastically); second, although the Bluetooth is matched, the radar features are contradictory (e.g., multiple target interference is detected). In this case, instead of simply rejecting the signal, an auxiliary judgment process is initiated: checking the posture details in the first identification result; if the posture is "high-frequency movement" and the signal features match the pet activity pattern library, the shutdown decision is strengthened; if it is determined to be "sitting still" but the Bluetooth is abnormal, a short-term enhanced scan may be triggered for reassessment. The execution of the second instruction is accompanied by a status marker (such as "stranger warning"), and the data characteristics of the scenario (radar spectrum fragments, Bluetooth signal sequences) are encrypted and stored in the local log for subsequent offline model optimization.

[0113] As described in step S413 above, if the first identification result determines the presence of a human target, the second identification result determines that the Bluetooth signal strength data originates from an authorized terminal, and the overall confidence level reaches a second threshold, then a third instruction is generated to control the air conditioner to turn on or switch to a preset mode; wherein, the second threshold is greater than the first threshold. This is the core execution link for achieving "precise authorized control," and its triggering conditions are the most stringent, requiring the simultaneous fulfillment of three conditions: radar confirms the presence of a human body, Bluetooth confirms it is a bound device, and the fusion confidence level reaches a higher second threshold (e.g., 0.8). This represents obtaining highly collaborative evidence across sensors, such as the radar continuously outputting a stable "human sitting" characteristic and the Bluetooth signal exhibiting a typical "actively approaching and dwelling" mode. After the conditions are met, it is not simply turned on, but a personalized startup sequence is executed: First, based on the posture classification (e.g., sleeping, working, active) in the first identification result, the corresponding preset mode template is called; second, combined with the device ID in the second identification result, the specific user is identified, and the user's preference settings (e.g., temperature, wind speed, swing angle) are loaded; the final generated third instruction is an infrared control code stream containing a complete set of parameters. Before sending the command, environmental adaptation adjustments are made, such as dynamically adjusting the start-up fan speed based on the difference between the current indoor temperature and the target temperature. To prevent momentary interference, the conditions are typically required to be met continuously for more than a set duration (e.g., 15 seconds). By combining multi-dimensional high-confidence verification with personalized strategies, seamless and convenient control is achieved while avoiding the accidental activation problems common in single-sensor solutions.

[0114] As described in step S414 above, if the first identification result determines that there is no human target, and the second identification result determines that there is an authorized terminal signal, then a fourth instruction is generated to control the air conditioner to enter the delayed shutdown process. The triggering condition is that the radar determines that the human target has left the effective area (e.g., the target distance continuously exceeds the 5-meter threshold or the micro-motion signal disappears), but the Bluetooth module continuously detects that the authorized terminal signal remains present (e.g., the RSSI strength is stable within the near field range). At this time, the air conditioner will not be shut down immediately, but a fourth instruction will be generated to start the intelligent delayed shutdown process. This process comprises three sub-stages: dynamic delay calculation, regression monitoring, and gradual shutdown control. First, the delay time (usually adjustable from 3-10 minutes) is calculated based on the user's historical behavior model (such as statistics on daily short-term absences) and real-time signal quality (Bluetooth signal stability). During the delay, radar and Bluetooth signals are continuously monitored. If a human return is detected or the Bluetooth signal rapidly strengthens (indicating the user's return), the delay is immediately canceled. If there are no signs of return before the delay ends, a gradual shutdown is executed (e.g., gradually adjusting the temperature to an energy-saving value before turning off the compressor). Simultaneously, the characteristics of this "absence mode" (such as absence time and signal attenuation curve) are recorded to optimize the user behavior model. This mechanism effectively solves the problem of frequent power on / off cycles in scenarios such as temporary item retrieval or short-term absences, significantly improving user comfort.

[0115] In one embodiment, after step S6 of generating control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result, the method further includes:

[0116] S701: Receive user operation data for the air conditioner and the corresponding Bluetooth signal strength data during the control operation to form a positive sample;

[0117] S702: Using the positive samples, the parameters of the confidence scoring model are optimized through an online learning algorithm to obtain an optimized confidence scoring model.

[0118] As described in step S701 above, the system receives user operation data for the air conditioner and the Bluetooth signal strength data corresponding to the control operation, forming a positive sample. A set of multimodal data (including radar features, Bluetooth features, and environmental context) corresponding to the moment the user actively operates the air conditioner is labeled as a positive sample, representing the ideal scenario where the system should respond correctly. Specifically, when the user actively operates the air conditioner through traditional methods (such as remote control, voice command, or physical buttons), a timestamp is captured synchronously, and a multimodal data sequence within a time window (e.g., 30 seconds before the operation to 5 seconds after the operation) is extracted from this timestamp. The system records the continuous output of the first identification result (including human presence determination, posture classification, and its probability distribution), the complete dynamics of the second identification result (Bluetooth signal strength sequence, device matching status, and active proximity confidence), and environmental context data (e.g., time, room temperature, and noise level) within this time period. After cleaning (removing abnormal frames), aligning (synchronizing time based on the operation time), and labeling, this data forms a structured positive sample. All samples are de-identified (removing personally identifiable information) and encrypted before storage, and associated with anonymous user IDs to support user-specific modeling. This mechanism ensures that the learning data comes directly from the actual behavioral feedback of users, providing real and reliable supervision signals for subsequent model optimization.

[0119] As described in step S702 above, the parameters of the confidence scoring model are optimized using the positive samples through an online learning algorithm to obtain the optimized confidence scoring model. Using the positive samples collected in S701, the confidence scoring model is incrementally optimized online without interrupting service, enabling continuous adaptation to user habits and environmental changes. The optimization process employs lightweight online learning algorithms (such as online random forests, incremental logistic regression, or mini-batch gradient descent), which are designed for embedded environments and feature low memory usage and single-sample update capabilities. Specifically, the feature vector of each new positive sample is input into the current confidence scoring model, and the loss between the model's prediction result (i.e., whether the user will perform an action) and the actual action label is calculated. Subsequently, based on this loss, the model parameters (such as the weights of split nodes in the decision tree and regression coefficients) are fine-tuned through the algorithm's internal mechanisms. To achieve stable learning, multiple protection mechanisms are introduced: First, a learning rate decay strategy is set to prevent overfitting caused by a single sample; second, a sliding window validation is used to periodically evaluate the performance of the optimized model with recent samples, and new parameters are only permanently saved when performance improves (such as an increase in prediction accuracy); finally, data from different users are trained in isolation to prevent personalized adaptation from interfering with each other. The optimized model parameters are securely written to non-volatile memory, and subsequent fusion recognition results will use the updated model for confidence scoring, gradually identifying and reinforcing user-specific behavioral patterns (such as specific home routes and common resting postures), continuously reducing the misjudgment rate and improving control accuracy.

[0120] Reference Figure 3 The present invention also provides an intelligent control device for an air conditioner, the device comprising:

[0121] The acquisition module 902 is used to periodically acquire target feature data of the target perception area through a preset human interaction detection module, and to acquire Bluetooth signal strength data through a Bluetooth module.

[0122] The recognition module 904 is used to perform human body recognition based on the target feature data and through a preset first offline AI model to obtain a first recognition result including human body presence determination and posture classification.

[0123] Analysis module 906 is used to perform dynamic feature analysis based on the Bluetooth signal strength data through a preset second offline AI model to obtain a second recognition result representing the intention of the authorized terminal to approach.

[0124] The fusion module 908 is used to fuse the first recognition result and the second recognition result to obtain a fused recognition result;

[0125] The calculation module 910 is used to input the fusion recognition result into a preset confidence score model to calculate the comprehensive confidence score, and compare the comprehensive confidence score with a preset threshold to obtain a comparison result;

[0126] The generation module 912 is used to generate control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result.

[0127] In one embodiment, the identification module 904 includes:

[0128] The human target recognition submodule is used to identify whether a human target exists within the target perception area based on the target feature data and a preset first offline AI model.

[0129] The human target data extraction submodule is used to extract human target data from the target feature data if a human target exists within the target perception area.

[0130] The first recognition result acquisition submodule is used to perform posture recognition on the human target data through a preset first offline AI model to obtain posture classification results, thereby obtaining the first recognition result.

[0131] In one embodiment, the analysis module 906 includes:

[0132] The judgment submodule is used to determine whether the Bluetooth signal strength data originates from a bound authorized terminal;

[0133] The dynamic change feature extraction submodule is used to extract the dynamic change features of Bluetooth signal strength within a preset time window if the signal originates from a bound authorized terminal.

[0134] The proximity behavior pattern analysis submodule is used to analyze the proximity behavior pattern of the authorized terminal based on the dynamic change characteristics and through a preset second offline AI model.

[0135] The second identification result acquisition submodule is used to output a confidence level representing the user's intention to actively approach as the second identification result based on the proximity behavior pattern.

[0136] In one embodiment, the fusion module 908 includes:

[0137] The identification submodule is used to determine whether there is a human target in the first identification result, and whether the Bluetooth signal strength data in the second identification result belongs to the bound authorized terminal;

[0138] The combination submodule is used to combine the posture features in the first identification result, the dynamic feature confidence level in the second identification result, and the environmental context features in the second identification result to form a multi-dimensional feature vector as the fusion identification result when the first identification result determines that there is a human target and the second identification result confirms that the Bluetooth signal comes from the bound authorized terminal.

[0139] In one embodiment, the generation module 912 includes:

[0140] A control air conditioner shutdown command generation submodule is used to generate a control air conditioner shutdown command when the overall confidence level is lower than a first preset threshold.

[0141] An air conditioner start command generation submodule is used to generate an air conditioner start command when the overall confidence level reaches or exceeds a second preset threshold; wherein the second preset threshold is greater than the first preset threshold.

[0142] The submodule for generating instructions to control the air conditioner to enter the delayed shutdown process is used to generate instructions to control the air conditioner to enter the delayed shutdown process when the first identification result determines that the human target has left, but the Bluetooth signal strength data indicates that the authorized terminal is still within the preset range.

[0143] In one embodiment, the fusion module 908 includes:

[0144] The first instruction generation submodule is used to generate a first instruction to control the air conditioner to turn off or keep it off if the first identification result determines that there is no human target and the second identification result determines that there is no authorized terminal signal.

[0145] The second instruction generation submodule is used to generate a second instruction to control the air conditioner to turn off or keep it off if the first identification result determines that there is a human target and the overall confidence level is lower than the first threshold.

[0146] The third instruction generation submodule is used to generate a third instruction to control the air conditioner to turn on or switch to a preset mode if the first identification result determines that there is a human target, the second identification result determines that the Bluetooth signal strength data comes from an authorized terminal, and the comprehensive confidence level reaches a second threshold; wherein, the second threshold is greater than the first threshold.

[0147] The fourth instruction generation submodule is used to generate a fourth instruction to control the air conditioner to enter the delayed shutdown process if the first identification result determines that there is no human target and the second identification result determines that there is an authorized terminal signal.

[0148] In one embodiment, the intelligent control device for the air conditioner further includes:

[0149] The positive sample formation module is used to receive user operation data on the air conditioner and the corresponding Bluetooth signal strength data when performing control operations to form positive samples;

[0150] The optimization module is used to optimize the parameters of the confidence scoring model using the positive samples and an online learning algorithm to obtain an optimized confidence scoring model.

[0151] Figure 4 An internal structural diagram of an electronic device in one embodiment is shown. This electronic device can specifically be a terminal, an air conditioner, or a computer device. Figure 4 As shown, the electronic device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement an intelligent control method for the air conditioner. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement the intelligent control method for the air conditioner. Those skilled in the art will understand that... Figure 4 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 electronic device to which the present application is applied. The specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0152] In one embodiment, an electronic device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0153] Periodically collect target feature data of the target perception area through the preset human interaction detection module, and collect Bluetooth signal strength data through the Bluetooth module;

[0154] Based on the target feature data, human body recognition is performed through a preset first offline AI model to obtain a first recognition result that includes human body presence determination and posture classification.

[0155] Based on the Bluetooth signal strength data, dynamic feature analysis is performed using a preset second offline AI model to obtain a second recognition result representing the authorized terminal's intention to approach.

[0156] The first recognition result and the second recognition result are fused to obtain a fused recognition result;

[0157] The fusion recognition result is input into a preset confidence scoring model to calculate the comprehensive confidence score, and the comprehensive confidence score is compared with a preset threshold to obtain the comparison result.

[0158] Control commands for the air conditioner are generated based on the first identification result, the second identification result, and the comparison result.

[0159] By analyzing radar data using a first offline AI model, attitude perception is achieved. Simultaneously, by analyzing Bluetooth dynamic characteristics using a second offline AI model, authorized identity and active intent are identified. The identification results of the two models are fused and calculated using a confidence scoring model, combining dual evidence of physical presence and identity signals. This reduces the false trigger rate caused by pets, mobile interference, or simply device abandonment. The first offline AI model, the second offline AI model, and the confidence scoring model all run locally, enabling control commands to be generated directly based on the first identification result, the second identification result, and the comparison result without relying on the cloud network. This ensures user privacy and security, eliminates network latency, and provides more personalized strategies that conform to user habits.

[0160] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the following steps:

[0161] Periodically collect target feature data of the target perception area through the preset human interaction detection module, and collect Bluetooth signal strength data through the Bluetooth module;

[0162] Based on the target feature data, human body recognition is performed through a preset first offline AI model to obtain a first recognition result that includes human body presence determination and posture classification.

[0163] Based on the Bluetooth signal strength data, dynamic feature analysis is performed using a preset second offline AI model to obtain a second recognition result representing the authorized terminal's intention to approach.

[0164] The first recognition result and the second recognition result are fused to obtain a fused recognition result;

[0165] The fusion recognition result is input into a preset confidence scoring model to calculate the comprehensive confidence score, and the comprehensive confidence score is compared with a preset threshold to obtain the comparison result.

[0166] Control commands for the air conditioner are generated based on the first identification result, the second identification result, and the comparison result.

[0167] By analyzing radar data using a first offline AI model, attitude perception is achieved. Simultaneously, by analyzing Bluetooth dynamic characteristics using a second offline AI model, authorized identity and active intent are identified. The identification results of the two models are fused and calculated using a confidence scoring model, combining dual evidence of physical presence and identity signals. This reduces the false trigger rate caused by pets, mobile interference, or simply device abandonment. The first offline AI model, the second offline AI model, and the confidence scoring model all run locally, enabling control commands to be generated directly based on the first identification result, the second identification result, and the comparison result without relying on the cloud network. This ensures user privacy and security, eliminates network latency, and provides more personalized strategies that conform to user habits.

[0168] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0169] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0170] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for intelligent control of an air conditioner, characterized in that, The method includes: Periodically collect target feature data of the target perception area through the preset human interaction detection module, and collect Bluetooth signal strength data through the Bluetooth module; Based on the target feature data, human body recognition is performed using a preset first offline AI model to obtain a first recognition result that includes human body presence determination and posture classification; specifically, it includes: based on the target feature data, identifying whether a human target exists within the target perception area using a preset first offline AI model; if a human target exists within the target perception area, extracting human target data from the target feature data; and performing posture recognition on the human target data using the preset first offline AI model to obtain a posture classification result, thereby obtaining the first recognition result. Based on the Bluetooth signal strength data, dynamic feature analysis is performed using a preset second offline AI model to obtain a second identification result representing the authorized terminal's intention to approach. Specifically, this includes: determining whether the Bluetooth signal strength data originates from a bound authorized terminal; if it originates from a bound authorized terminal, extracting the dynamic change features of the Bluetooth signal strength within a preset time window; based on the dynamic change features, analyzing the approach behavior pattern of the authorized terminal using a preset second offline AI model; and outputting a confidence level representing the user's active approach intention as the second identification result based on the approach behavior pattern. The first recognition result and the second recognition result are fused to obtain a fused recognition result; The fusion recognition result is input into a preset confidence scoring model to calculate the comprehensive confidence score, and the comprehensive confidence score is compared with a preset threshold to obtain the comparison result. Control commands for the air conditioner are generated based on the first identification result, the second identification result, and the comparison result. Receive user operation data for the air conditioner and the corresponding Bluetooth signal strength data when performing control operations to form positive samples; Using the positive samples, the parameters of the confidence scoring model are optimized through an online learning algorithm to obtain an optimized confidence scoring model.

2. The intelligent control method for air conditioning according to claim 1, characterized in that, The step of generating control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result includes: Determine whether there is a human target in the first identification result, and whether the Bluetooth signal strength data in the second identification result belongs to the bound authorized terminal; When the first identification result determines that a human target exists, and the second identification result confirms that the Bluetooth signal originates from a bound authorized terminal, the posture features in the first identification result, the dynamic feature confidence level in the second identification result, and the environmental context features are combined to form a multi-dimensional feature vector as the fused identification result.

3. The intelligent control method for air conditioning according to claim 1, characterized in that, The step of generating control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result includes: When the overall confidence level is lower than a first preset threshold, a command is generated to control the air conditioner to turn off. When the overall confidence level reaches or exceeds the second preset threshold, a command to control the air conditioner to turn on is generated; wherein the second preset threshold is greater than the first preset threshold; When the first identification result determines that the human target has left, but the Bluetooth signal strength data indicates that the authorized terminal is still within the preset range, an instruction is generated to control the air conditioner to enter the delayed shutdown process.

4. The intelligent control method for air conditioning according to claim 1, characterized in that, The step of generating control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result includes: If the first identification result determines that there is no human target and the second identification result determines that there is no authorized terminal signal, then a first instruction is generated to control the air conditioner to turn off or remain off. If the first identification result determines that a human target exists, and the overall confidence level is lower than the first threshold, then a second instruction is generated to control the air conditioner to turn off or remain off. If the first identification result determines that a human target exists, the second identification result determines that the Bluetooth signal strength data comes from an authorized terminal, and the overall confidence level reaches a second threshold, then a third instruction is generated to control the air conditioner to turn on or switch to a preset mode; wherein, the second threshold is greater than the first threshold. If the first identification result determines that there is no human target, and the second identification result determines that there is an authorized terminal signal, then a fourth instruction is generated to control the air conditioner to enter the delayed shutdown process.

5. An intelligent control device for an air conditioner, characterized in that, The device includes: The acquisition module is used to periodically acquire target feature data of the target perception area through a preset human interaction detection module, and to acquire Bluetooth signal strength data through a Bluetooth module. The recognition module is used to perform human body recognition based on the target feature data using a preset first offline AI model, and obtain a first recognition result including human body presence determination and posture classification; specifically, it includes: based on the target feature data, identifying whether there is a human target in the target perception area using a preset first offline AI model; if there is a human target in the target perception area, extracting human target data from the target feature data; and performing posture recognition on the human target data using the preset first offline AI model to obtain a posture classification result, thereby obtaining the first recognition result; The analysis module is used to perform dynamic feature analysis based on the Bluetooth signal strength data using a preset second offline AI model to obtain a second identification result representing the intention of the authorized terminal to approach. Specifically, it includes: determining whether the Bluetooth signal strength data originates from a bound authorized terminal; if it originates from a bound authorized terminal, extracting the dynamic change features of the Bluetooth signal strength within a preset time window; analyzing the approach behavior pattern of the authorized terminal based on the dynamic change features using a preset second offline AI model; and outputting a confidence level representing the user's active approach intention as the second identification result based on the approach behavior pattern. The fusion module is used to fuse the first recognition result and the second recognition result to obtain a fused recognition result; The calculation module is used to input the fusion recognition result into a preset confidence score model to calculate the comprehensive confidence score, and compare the comprehensive confidence score with a preset threshold to obtain a comparison result; The generation module is used to generate control commands for the air conditioner based on the first identification result, the second identification result, and the comparison result; The positive sample formation module is used to receive user operation data on the air conditioner and the corresponding Bluetooth signal strength data when performing control operations to form positive samples; The optimization module is used to optimize the parameters of the confidence scoring model using the positive samples and an online learning algorithm to obtain an optimized confidence scoring model.

6. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the steps of the intelligent control method for the air conditioner as described in any one of claims 1 to 4.

7. An electronic device, characterized in that, The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the intelligent control method for the air conditioner as described in any one of claims 1 to 4.

Images