An air conditioner control method and device, electronic equipment and storage medium

Abstract

The application discloses an air conditioner control method and device, electronic equipment and storage medium, which collects multi-dimensional sensing signals according to a sampling period after network disconnection, dynamically allocates computing power mode and operation precision in combination with chip hardware resource state, calls a local pre-training model to generate a first control signal, breaks through the dependence of network disconnection on cloud algorithms, realizes offline autonomous adaptive control and robust operation of the equipment on a low-power IoT chip, and at the same time, when the communication link is restored, a cloud intention vector, a local state vector and a cloud historical instruction are acquired, a dynamic semantic mapping is constructed, and a gradual smooth target control instruction and a second control signal are generated, so that the step jump and mode conflict during cloud edge state switching are completely eliminated, the frequent start and stop of air conditioner compressors and other hardware are avoided, the service life of the equipment hardware is prolonged, and the smooth adaptive control technical effect of no conflict, no interruption and high safety under the whole working condition of network disconnection autonomy and network recovery is achieved.

Description

Technical Field

[0001] This invention relates to the field of air conditioning control technology, and in particular to an air conditioning control method, an air conditioning control device, an electronic device, and a readable storage medium. Background Technology

[0002] Intelligent control of air conditioners through cloud-edge interaction refers to using the big data analysis and model decision-making capabilities of the cloud, combined with the real-time environmental perception and instant response capabilities of the local edge chip of the air conditioner, to achieve dynamic and precise regulation of the air conditioner's operating status through two-way data collaboration between the cloud and the edge.

[0003] When a communication link is interrupted, related technologies generally use static preset rules for degradation control. However, when the fixed threshold triggering control action of static preset rules changes dynamically in multi-dimensional environmental parameters, the output control command lacks accuracy and flexibility. This can lead to indoor temperature and humidity not being able to accurately adapt to dynamic environmental changes, causing drastic fluctuations in indoor temperature and sudden changes in body temperature, as well as sudden wind noise and unnecessary additional energy consumption due to frequent switching of operating conditions.

[0004] Furthermore, when the communication link is restored after an interruption, the relevant technologies typically overwrite the current local control state with the control intent or status parameters issued by the cloud server, forcing the air conditioner to immediately execute the cloud instructions. This direct overriding switching causes a step change in control parameters of the air conditioner at the moment of cloud-edge state switching, which can easily lead to sudden frequency changes in the inverter compressor and mechanical shocks to the hardware, affecting the equipment's lifespan and the continuous stability of environmental control. Summary of the Invention

[0005] The present invention provides an air conditioning control method, apparatus, electronic device, and readable storage medium to overcome or at least partially solve the above-mentioned problems.

[0006] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide an air conditioning control method, the method being applied to the air conditioner, the air conditioner being used to perform data interaction with a cloud server, the method comprising: When it is determined that the air conditioner is in a communication link interruption state, multi-dimensional sensing signals are collected according to a preset sampling period, and standardized multi-dimensional feature vectors and data validity flag bits are generated through the multi-dimensional sensing signals. Obtain the hardware resource parameters of the air conditioner, and determine the current computing power mode and computing precision configuration of the air conditioner based on the hardware resource parameters; The pre-trained model integrated into the air conditioner is invoked to generate initial artificial intelligence decision instructions for the air conditioner through the current computing power mode, the computing precision configuration, the standardized multidimensional feature vector, and the data validity flag. The initial artificial intelligence decision instruction is converted into a first control signal for the air conditioner, and the air conditioner is controlled based on the first control signal when the air conditioner is in the communication link interruption state; In response to the restoration of the communication link of the air conditioner, the cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner was in the communication link interruption state are obtained, and a progressive smooth target control instruction is generated to eliminate the step jump when the air conditioner is in the cloud edge state switching. Based on the standardized multidimensional feature vector, the progressive smooth target control command, and the initial artificial intelligence decision command, a second control signal is generated; The air conditioner is controlled based on the second control signal when the communication link of the air conditioner is restored.

[0007] Optionally, the hardware resource parameters include central processing unit utilization, remaining static random access memory capacity, and power supply voltage data. The step of determining the current computing power mode and computing precision configuration of the air conditioner based on the hardware resource parameters includes: A three-dimensional resource vector is generated by performing vector construction on the CPU utilization, the remaining static random access memory capacity, and the power supply voltage data. The three-dimensional resource vector is weighted and calculated using a preset weighted comprehensive evaluation function to generate a resource health index; The resource health index is subjected to hysteresis state determination by combining the preset hysteresis determination duration data and the dual threshold hysteresis band mechanism, and a computing power state assessment result is generated. Based on the computing power status assessment results, a matching process is performed to determine the current computing power mode; Based on the computing power status assessment results, precision mapping is performed to generate the computing precision configuration.

[0008] Optionally, the pre-trained model comprises a three-layer neural network, and the step of generating initial artificial intelligence decision instructions for the air conditioner using the current computing power mode, the computing precision configuration, the standardized multidimensional feature vector, and the data validity flag includes: The standardized multidimensional feature vector and the data validity flag are input into the pre-trained model to generate initial feature input data; Based on the current computing power mode and the computing precision configuration, the hidden layer weights of the three-layer neural network are loaded and precision configuration is selected to generate target hidden layer weights; Based on the target hidden layer weights, forward inference calculations are performed on the initial feature input data to generate initial artificial intelligence decision instructions.

[0009] Optionally, after the step of controlling the air conditioner based on the first control signal when the air conditioner is in the communication link interruption state, the online self-evolutionary training process of the pre-trained model includes: Obtain the standardized multidimensional feature vector, the personalized user fingerprint vector expressing user preference parameters, and the environmental data within a set time window; Based on the preset equivalent coefficients, the temperature deviation, humidity deviation, wind speed perception deviation, and personnel activity status deviation in the environmental data within the set time window are uniformly normalized to generate normalized temperature dimension data. By combining the normalized temperature dimension data, the standardized multidimensional feature vector, and the personalized user fingerprint vector, residual calculation is performed to generate a multidimensional comfort residual vector and a comprehensive residual scalar. When it is determined that the current computing power mode meets the preset conditions, the output layer weights determined by the pre-trained model are used to perform backpropagation calculation on the comprehensive residual scalar to generate the hidden layer local gradient; Obtain offline duration data, and perform calculations based on the offline duration data and the comprehensive residual scalar to generate an adaptive learning rate; Obtain elastic weights to reinforce the regularization term; By combining the adaptive learning rate and the elastic weight consolidation regularization term, regularization and scaling calculations are performed on the local gradients of the hidden layer to generate gradient update values. Obtain the existing hidden layer weights of the pre-trained model, and perform update calculations on the existing hidden layer weights based on the gradient update amount to generate updated hidden layer weights; The updated hidden layer weights are loaded into the hidden layer of the pre-trained model to generate an updated pre-trained model; the updated pre-trained model is used to update the first control signal.

[0010] Optionally, the step of obtaining the cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner is in a communication link interruption state, and generating a gradual and smooth target control instruction for eliminating the step jump when the air conditioner is in a cloud-edge state switch includes: Obtain the cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner is in a communication link interruption state; Normalization and projection mapping are performed on the cloud intent vector and the local state vector to generate a normalized projected intent vector and a normalized projected state vector. The semantic distance is calculated by combining the normalized projection intention vector and the normalized projection state vector to generate a hybrid semantic distance; When the mixed semantic distance is determined to be greater than a preset threshold, the real-time operating parameters of the air conditioner are obtained, and parameter construction is performed based on the real-time operating parameters to generate a dynamic mapping matrix and a dynamic bias vector. Incremental calculation is performed by combining the cloud intent vector and the cloud historical instructions to generate a cloud intent increment; Using the dynamic mapping matrix and the dynamic bias vector, matrix mapping and bias transformation are performed on the cloud intent increment to generate a local control increment; Acquire offline duration data and perform weight calculation on the offline duration data to determine the weighted fusion coefficient based on offline duration; By combining the weighted fusion coefficients and the preset adaptive first-order inertial filtering algorithm, a smoothing calculation is performed on the local control increment to generate a progressively smooth target control command.

[0011] Optionally, the air conditioner is equipped with a four-level safety state machine, and the step of generating a second control signal based on the standardized multidimensional feature vector, the progressive smooth target control command, and the initial artificial intelligence decision command includes: Obtain the initial artificial intelligence decision instruction, the progressive smooth target control instruction, and the standardized multidimensional feature vector; Obtain the current security state determined by the four-level security state machine; Combining the standardized multidimensional feature vector and the current security state, a three-layer progressive hard constraint verification, including hard truncation, feasible region projection mapping, and safe alternative solution, is performed on the initial artificial intelligence decision instruction or the progressive smooth target control instruction to generate a preliminary verification instruction. The minimum start-stop time constraint data and frequency change rate limit data of the inverter compressor of the air conditioner are obtained, as well as the air supply transition protection sequence for the cooling and heating mode switching process of the air conditioner. Based on the minimum start-stop time constraint data, the frequency change rate limit data, and the air supply transition protection sequence, the hardware physical constraint linkage verification is performed on the preliminary verification instruction to generate a final compliance control instruction that conforms to the safety boundary and hardware physical constraints, and the final compliance control instruction is converted into the second control signal.

[0012] Optionally, the step of controlling the air conditioner based on the second control signal when the communication link of the air conditioner is restored includes: Obtain the second control signal; Based on the second control signal, the variable frequency compressor and fan terminal load of the air conditioner are driven to perform control actions; During the execution of the control actions by the variable frequency compressor and the fan terminal load, actual operating condition parameters are collected in real time, and actual operating status data of the equipment is generated based on the actual operating condition parameters; the actual operating status data of the equipment is used to update the local state vector.

[0013] Secondly, embodiments of this application provide an air conditioning control device, which is applied to the air conditioner, and the air conditioner is used to perform data interaction with a cloud server. The device includes: The signal acquisition module is used to acquire multi-dimensional sensor signals according to a preset sampling period when it is determined that the air conditioner is in a communication link interruption state, and to generate a standardized multi-dimensional feature vector and a data validity flag bit through the multi-dimensional sensor signals. The hardware resource parameter acquisition module is used to acquire the hardware resource parameters of the air conditioner and determine the current computing power mode and computing precision configuration of the air conditioner based on the hardware resource parameters. An initial artificial intelligence decision command generation module is used to call the pre-trained model integrated into the air conditioner and generate an initial artificial intelligence decision command for the air conditioner through the current computing power mode, the computing precision configuration, the standardized multidimensional feature vector and the data validity flag. The first control signal conversion module is used to convert the initial artificial intelligence decision instruction into a first control signal for the air conditioner, and to control the air conditioner based on the first control signal when the air conditioner is in the communication link interruption state; A progressive smooth target control instruction generation module is used to respond to the restoration of the communication link of the air conditioner, obtain the cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner is in the communication link interruption state, and generate progressive smooth target control instructions to eliminate the step jump when the air conditioner is in the cloud edge state switching. The second control signal generation module is used to generate a second control signal based on the standardized multidimensional feature vector, the progressive smooth target control command, and the initial artificial intelligence decision command. An air conditioning control module is used to control the air conditioner when the communication link of the air conditioner is restored based on the second control signal.

[0014] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0015] Fourthly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0016] Fifthly, embodiments of this application provide a chip, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the method as described in the first aspect.

[0017] The embodiments of the present invention have the following advantages: This invention addresses several technical challenges faced by smart home systems after network outages. These include reliance on fixed logic leading to system paralysis, heterogeneous cloud-edge state synchronization causing instruction jumps and frequent device startups / shutdowns damaging hardware, and the limited computing power of edge chips hindering the handling of complex algorithms. By collecting multi-dimensional sensor signals at sampling intervals after network outages and dynamically allocating computing power modes and computational precision based on chip hardware resource status, a first control signal is generated using a locally pre-trained model. This overcomes the dependence on cloud algorithms during network outages, enabling offline autonomous adaptive control and robust operation of devices on low-power IoT chips. Simultaneously, upon communication link restoration, dynamic semantic mapping is constructed using cloud intent vectors, local state vectors, and historical cloud instructions. This generates progressively smooth target control instructions and a second control signal, completely eliminating step jumps and mode conflicts during cloud-edge state switching. This avoids frequent startups / shutdowns of hardware such as air conditioner compressors, extends device hardware lifespan, and achieves conflict-free, uninterrupted, and highly secure smooth adaptive control across all operating conditions from network outage autonomy to network recovery. Attached Figure Description

[0018] Figure 1 This is a flowchart of the steps of an air conditioning control method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an air conditioning control system provided in an embodiment of the present invention; Figure 3 This is a structural block diagram of an air conditioning control device provided in an embodiment of the present invention; Figure 4 This is a hardware structure block diagram of an electronic device provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and are not used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, the character " / " in the specification and claims, indicating at least one of the connected objects, generally indicates that the preceding and following objects are in an "or" relationship.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Reference Figure 1 The diagram illustrates a flowchart of an air conditioning control method provided in an embodiment of the present invention, which may specifically include the following steps: Step 101: When it is determined that the air conditioner is in a communication link interruption state, multi-dimensional sensing signals are collected according to a preset sampling period, and standardized multi-dimensional feature vectors and data validity flag bits are generated through the multi-dimensional sensing signals. Step 102: Obtain the hardware resource parameters of the air conditioner, and determine the current computing power mode and computing precision configuration of the air conditioner based on the hardware resource parameters; Step 103: Invoke the pre-trained model integrated into the air conditioner, and generate initial artificial intelligence decision instructions for the air conditioner through the current computing power mode, the computing precision configuration, the standardized multidimensional feature vector and the data validity flag. Step 104: Convert the initial artificial intelligence decision instruction into a first control signal for the air conditioner, and control the air conditioner based on the first control signal when the air conditioner is in the communication link interruption state; Step 105: In response to the restoration of the communication link of the air conditioner, obtain the cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner is in the communication link interruption state, and generate a progressive smooth target control instruction to eliminate the step jump when the air conditioner is in the cloud edge state switching. Step 106: Generate a second control signal based on the standardized multidimensional feature vector, the progressive smooth target control command, and the initial artificial intelligence decision command; Step 107: Control the air conditioner based on the second control signal when the communication link of the air conditioner is restored.

[0023] In this embodiment of the invention, when the air conditioner is determined to be in a communication link interruption state, multi-dimensional sensing signals are collected according to a preset sampling period. Standardized multi-dimensional feature vectors and data validity flags are generated through the multi-dimensional sensing signals. This allows for real-time sensing of changes in the indoor and outdoor environment and the air conditioner's own operating conditions when the air conditioner is disconnected from the cloud server (in a communication link interruption state). The original sensing data is preprocessed such as noise reduction and normalization, and validity labeling is performed. This provides accurate, highly reliable, and uniformly formatted data input for local intelligent decision-making in an offline state.

[0024] Communication link interruption status refers to an offline operating state in which the network connection between the air conditioner and the cloud server fails or is interrupted, resulting in the inability to upload or download data normally.

[0025] The preset sampling period refers to the fixed time interval (e.g., once every 60 seconds) at which the system pre-sets the collection of various sensor data.

[0026] Multidimensional sensing signals refer to the raw physical signals that reflect the indoor and outdoor environment and the operating conditions of the equipment itself, which are collected by multiple sensors in a coordinated manner. These signals include indoor temperature, indoor humidity, outdoor temperature, human presence status, fan speed, compressor operating frequency, etc.

[0027] Standardized multidimensional feature vectors refer to the data representation in the form of fixed-point vectors formed by mapping the acquired multidimensional sensor signals to a uniform numerical range (such as the [0,1] interval) after preprocessing such as noise reduction and anomaly removal and linear normalization.

[0028] The data validity flag is a binary bitmap identifier used to indicate whether the data collected by each sensor is normal and valid, and whether there is data loss or fault.

[0029] In this embodiment of the invention, the hardware resource parameters of the air conditioner can be obtained, and the current computing power mode and computing precision configuration of the air conditioner can be determined based on the hardware resource parameters. This allows for real-time monitoring of the hardware resource load and power status of the edge-side main control chip (such as an MCU), dynamic scheduling of computing resources and numerical precision, resolution of the computing power bottleneck and energy consumption contradiction when low-power IoT chips run AI algorithms, and ensuring the robustness of the device under resource-constrained or extreme operating conditions.

[0030] Hardware resource parameters refer to three-dimensional monitoring data reflecting the current physical resource occupancy and health status of the air conditioner's main control chip and system, including CPU utilization, remaining SRAM memory capacity, and power supply voltage.

[0031] The current computing power mode refers to the system computing load operation level classified based on dynamic evaluation of hardware resource parameters, including high computing power mode, medium computing power mode and low computing power mode.

[0032] The computational precision configuration refers to the numerical precision of model calculations and the training / inference execution status that match the current computing power mode. It includes configurations such as INT32 full-precision computation, INT8 quantization compression computation, and pausing training to retain only inference.

[0033] In this embodiment of the invention, a pre-trained model integrated into the air conditioner can be invoked to generate initial artificial intelligence decision instructions for the air conditioner through the current computing power mode, the computing precision configuration, the standardized multidimensional feature vector, and the data validity flag. This allows the lightweight AI model integrated locally in the air conditioner to perform autonomous reasoning based on current computing power resources and valid feature data, without relying on the cloud, and to dynamically generate intelligent control strategies that adapt to environmental changes and user comfort needs.

[0034] The pre-trained model integrated into the air conditioner refers to a lightweight neural network (including an input layer, an adaptive hidden layer, and an output layer) deployed in the main control chip at the edge of the air conditioner, which has offline inference and adaptive adjustment capabilities.

[0035] The initial AI decision command refers to the preliminary adjustment command generated by the pre-trained model through forward inference calculation based on the input standardized feature vector and computing power precision configuration. This command includes the target temperature, fan speed, compressor operating frequency, and operating mode.

[0036] In this embodiment of the invention, the initial artificial intelligence decision instruction can be converted into a first control signal for the air conditioner, and the air conditioner can be controlled based on the first control signal when the air conditioner is in the communication link interruption state. This converts the initial decision instruction generated by AI into a physical control electrical signal that can directly drive the hardware actuator, thereby achieving efficient, safe, and continuous autonomous dynamic control of the air conditioner during offline and network outage periods.

[0037] The first control signal refers to the physical electrical signal or drive pulse generated after the initial artificial intelligence decision instruction has been verified by the underlying security rules and mapped by the physical drive. It is used to directly drive the operation of hardware actuators such as air conditioner compressors and fans.

[0038] In this embodiment of the invention, in response to the restoration of the communication link of the air conditioner, a cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner was in a communication link interruption state can be obtained. A progressive and smooth target control instruction can be generated to eliminate the step jump when the air conditioner is in a cloud-edge state switch. This solves the problem of state heterogeneity and instruction conflict caused by the long-term independent operation of the local system when the network is restored, prevents instruction mutations during the instantaneous switching of cloud-edge state, avoids the risk of frequent device start-up and shutdown or hardware damage, and achieves a conflict-free and seamless smooth switch.

[0039] A cloud server refers to a centralized computing server located on a remote network that is responsible for processing global smart home strategies and data archiving.

[0040] Communication link restoration refers to the re-establishment of the network connection between the air conditioner and the cloud server and the resumption of normal data interaction.

[0041] The cloud intent vector refers to a normalized vector issued by the cloud server that represents the cloud's expectations for air conditioning control, including target temperature, operating mode code, fan speed setting, and energy efficiency rating.

[0042] The local state vector refers to the vector that is maintained locally and in real time by the air conditioner and reflects the current actual physical operating status of the equipment. It includes real-time room temperature, indoor and outdoor temperature difference, compressor operating frequency, fan speed, real-time humidity and continuous operating time.

[0043] Historical cloud commands refer to historical control commands generated by the cloud server according to a predetermined strategy or remote user operation during a communication link interruption, and cached in a cloud or local circular buffer.

[0044] A step transition refers to a phenomenon where control parameters (such as target temperature and compressor frequency) change drastically or directly contradict each other at the moment of state switching due to heterogeneity and inconsistency between the cloud-based control commands and the local current actual operating state.

[0045] Progressive smooth target control instructions refer to smooth transition control instructions that are generated step by step within a set transition time according to an adaptive rate of change, based on a two-way dynamic semantic mapping between cloud intent and local state, combined with a filtering algorithm.

[0046] In this embodiment of the invention, a second control signal can be generated based on the standardized multidimensional feature vector, the progressive smoothing target control command, and the initial artificial intelligence decision command. This signal integrates the smoothing target command, local perception features, and AI decision command, and then applies physical boundary constraints (such as single-step change rate limits and absolute safety ranges) through an offline security policy pool to generate a final control signal that balances comfort, smoothness, and hardware security compliance.

[0047] The second control signal refers to the final control electrical signal generated after integrating progressive smooth target control instructions, local real-time features, and AI decision instructions, and after offline policy pool security boundary verification and compliance clamping, used to drive the air conditioning actuator during the network recovery transition period.

[0048] In this embodiment of the invention, the air conditioner can be controlled based on the second control signal when the communication link of the air conditioner is restored, so that during the smooth transition phase after the communication link is restored, the air conditioner can be driven to operate using a safe and smooth second control signal to complete the seamless transition from local autonomous mode to cloud-edge collaborative mode.

[0049] This invention addresses several technical challenges faced by smart home systems after network outages. These include reliance on fixed logic leading to system paralysis, heterogeneous cloud-edge state synchronization causing instruction jumps and frequent device startups / shutdowns damaging hardware, and the limited computing power of edge chips hindering the handling of complex algorithms. By collecting multi-dimensional sensor signals at sampling intervals after network outages and dynamically allocating computing power modes and computational precision based on chip hardware resource status, a first control signal is generated using a locally pre-trained model. This overcomes the dependence on cloud algorithms during network outages, enabling offline autonomous adaptive control and robust operation of devices on low-power IoT chips. Simultaneously, upon communication link restoration, dynamic semantic mapping is constructed using cloud intent vectors, local state vectors, and historical cloud instructions. This generates progressively smooth target control instructions and a second control signal, completely eliminating step jumps and mode conflicts during cloud-edge state switching. This avoids frequent startups / shutdowns of hardware such as air conditioner compressors, extends device hardware lifespan, and achieves conflict-free, uninterrupted, and highly secure smooth adaptive control across all operating conditions from network outage autonomy to network recovery.

[0050] For example, the scenario of a home air conditioner experiencing an unexpected network outage due to a home router crashing during operation on a summer night, followed by subsequent network recovery, will be used to illustrate the point.

[0051] At 2:00 AM, the home router suddenly crashed, and the air conditioner determined that it had entered a communication link interruption state. The air conditioner collected multi-dimensional sensor signals such as indoor temperature (26℃), humidity (55%), human presence signal, and compressor frequency at a preset sampling period of 60 seconds. After preprocessing, standardized multi-dimensional feature vectors were generated, and the validity flag of the sensor data that was sampled normally was set to a fully valid flag.

[0052] The main control chip monitors the current hardware resource parameters (CPU utilization 25%, remaining memory 30KB, power supply voltage 3.3V), determines that the resources are sufficient, sets its current computing power mode to high computing power mode, and sets the computing precision configuration to full precision INT32.

[0053] The air conditioner calls upon the local pre-trained model, combining the current high computing power mode, INT32 precision configuration, and standardized feature vectors to calculate the drop in outdoor temperature and the reduction in indoor heat load in the early morning, generating initial artificial intelligence decision instructions to reduce the compressor frequency to 35Hz and adjust the fan speed to a low setting.

[0054] The system converts the instruction into a first control signal after security verification, driving the compressor and fan to operate, so that the room can maintain a comfortable and constant temperature even without the internet, avoiding overcooling or loss of control caused by traditional fixed logic.

[0055] The router successfully restarted at 5:00 AM, and the communication link was restored. The cloud sends historical commands (such as the user's previous setting to activate powerful cooling at 22°C at 5:00 AM), corresponding to the cloud intent vector; at this time, the local air conditioner is operating at a low frequency of 26°C, corresponding to the local state vector. Directly executing the cloud command would cause the compressor frequency to jump instantly from 35Hz to 90Hz (i.e., a step jump), causing hardware impact. The system uses a first-order inertial filtering algorithm to generate a gradual, smooth target control command that progressively increases the compressor frequency from 35Hz to the target value over 120 seconds at a rate not exceeding 0.33Hz per second.

[0056] Safety verification and second control signal generation: Combining real-time feature vectors, smooth target instructions and local AI decisions, and after offline safety rule verification (confirming that the temperature and pressure are not exceeded), a second control signal is generated.

[0057] The air conditioner operates based on the second control signal and smoothly transitions to the target state set in the cloud within 120 seconds, which not only responds to the cloud control intent but also ensures the safety of the air conditioner hardware and the stable and continuous indoor comfort.

[0058] Based on the above embodiments, modified embodiments of the above embodiments are proposed. It should be noted that, in order to keep the description brief, only the differences from the above embodiments are described in the modified embodiments.

[0059] In an optional embodiment of the present invention, the hardware resource parameters include central processing unit utilization, remaining static random access memory capacity, and power supply voltage data. The step of determining the current computing power mode and computing precision configuration of the air conditioner based on the hardware resource parameters includes: A three-dimensional resource vector is generated by performing vector construction on the CPU utilization, the remaining static random access memory capacity, and the power supply voltage data. The three-dimensional resource vector is weighted and calculated using a preset weighted comprehensive evaluation function to generate a resource health index; The resource health index is subjected to hysteresis state determination by combining the preset hysteresis determination duration data and the dual threshold hysteresis band mechanism, and a computing power state assessment result is generated. Based on the computing power status assessment results, a matching process is performed to determine the current computing power mode; Based on the computing power status assessment results, precision mapping is performed to generate the computing precision configuration.

[0060] In this embodiment of the invention, by comprehensively collecting and evaluating multi-dimensional hardware parameters such as the central processing unit utilization rate, remaining static random access memory (SRAM) capacity, and power supply voltage of the air conditioner's main control chip, and combining the hysteresis determination duration and dual-threshold hysteresis band mechanism, the resource health of the main control chip is accurately quantified and evaluated. While effectively filtering out instantaneous resource fluctuation interference, the current computing power mode and computing precision configuration are adaptively and stably determined, thereby ensuring the robustness and stability of the air conditioner's operation when edge computing resources are limited or operating conditions fluctuate.

[0061] Hardware resource parameters refer to three-dimensional monitoring data that reflects the real-time physical resource status of the air conditioner's main control chip and its operating environment, including central processing unit utilization, remaining static random access memory capacity, and power supply voltage data.

[0062] Central processing unit utilization refers to the percentage of computing power of the central processing unit (CPU) of the air conditioner's main control chip that has been occupied by the system or computing tasks during the current computing cycle.

[0063] The remaining static random access memory capacity refers to the amount of SRAM available in the air conditioner main control chip that has not yet been allocated and can be used for model inference or system programs.

[0064] Power supply voltage data refers to the physical value of the input power voltage supplied to the main control chip or core control circuit of the air conditioner, which is collected in real time.

[0065] A three-dimensional resource vector is a numerical vector containing three characteristic elements, which is formed by combining the CPU utilization, remaining static random access memory capacity, and power supply voltage data in a preset dimension order.

[0066] The preset weighted comprehensive evaluation function refers to a pre-defined mathematical algorithm expression that assigns different weight coefficients to each indicator in the three-dimensional resource vector and performs weighted calculations.

[0067] The Resource Health Index is a scalar score used to quantitatively assess the sufficiency and health of chip hardware resources, obtained by weighting the three-dimensional resource vector using a weighted comprehensive evaluation function.

[0068] The preset hysteresis judgment duration data refers to the pre-set time buffer threshold for anti-shake, that is, the minimum duration for which the resource health index must remain within the new state range.

[0069] The dual-threshold hysteresis mechanism refers to a control mechanism that forms a hysteresis interval by setting an upper threshold and a lower threshold, and only triggers a state switch when the indicator crosses the threshold and meets a certain duration condition to prevent frequent state oscillations.

[0070] For example, the scenario in which the air conditioner's main control chip dynamically switches between "high computing power mode" and "medium computing power mode" based on the real-time computing power status will be used as an example for illustration.

[0071] Assume the system uses the calculated resource health index (scalar score, ranging from 0 to 100) as the basis for status assessment. The system presets a high-computing-power degradation threshold (lower limit) Tlower=65 and an upgrade threshold (upper limit) Tupper=75 in the dual-threshold hysteresis mechanism (forming a hysteresis interval of [65, 75]), and sets the preset hysteresis judgment duration Thold=5 seconds. The specific operation process is as follows: Initial state and normal operation The air conditioner's main control chip is currently in "high computing power mode" (loading full-precision INT32 model). At this time, the resource health index S monitored in real time is at a high level of 85, and the system is running stably.

[0072] Scenario 1: Anti-shake determination (filtering instantaneous load pulses) At time t1, due to concurrent processing of sensor data or local system task scheduling, CPU utilization spikes instantly, causing the resource health index S to plummet from 85 to 60 (falling below the lower threshold Tlower=65). At this point, the system does not immediately switch computing power modes, but instead starts a hysteresis timer for continuous monitoring. If the short-term concurrent task ends at t1+3 seconds, the resource health index S quickly rises back to 72 (again above the lower limit threshold of 65). Since the index fell below the lower limit for only 3 seconds, which did not reach the preset hysteresis judgment time Thold=5 seconds, the system determined that the drop was an instantaneous power ripple or temporary calculation pulse, suppressed the mode switching, and the chip continued to remain in "high computing power mode", avoiding unnecessary mode oscillation.

[0073] Scenario 2: Triggering a downgrade switch (insufficient deterministic resources) At time t2, due to the increase in chip ambient temperature leading to a slight drop in supply voltage and an increase in computational tasks, the resource health index S dropped from 85 to 58 (falling below the lower limit threshold Tlower=65), and remained below 65 for the next 5 seconds (i.e., from t2 to t2+5 seconds). The system determined that the hysteresis judgment duration requirement was met, triggered a computing power mode downgrade, and smoothly switched to "medium computing power mode" (while simultaneously switching the model to INT8 quantization precision), effectively reducing chip power consumption and heat generation.

[0074] Scenario 3: Hysteresis band anti-oscillation upgrade (preventing rapid switching of critical point) During the "medium computing power mode" operation, as tasks are released, the resource health index rises back to 70. Although 70 is higher than the downgrade threshold of 65, it is still within the lag band of [65, 75] and has not exceeded the upper limit threshold Tupper=75. Therefore, the system will not trigger an upgrade and will continue to maintain the "medium computing power mode". Only when the resource health index continues to rise significantly and exceeds Tupper=75 (for example, reaching 80), and maintains this state for Thold=5 seconds, will the system upgrade and switch back to "high computing power mode".

[0075] By leveraging the combined effect of dual-threshold hysteresis bands and hysteresis determination duration, frequent fluctuations near the critical point are isolated in the spatial dimension using the upper and lower threshold differences (a buffer space of 75-65=10 minutes), and transient noise pulses are filtered out in the temporal dimension using Thold. This avoids rapid switching between computing power mode and computing accuracy under critical conditions, ensuring the continuous and stable operation of the edge computing system.

[0076] The computing power status assessment result refers to the system computing power level assessment conclusion obtained after combining the hysteresis judgment duration data and the dual-threshold hysteresis band mechanism to smooth the execution status of the resource health index.

[0077] The current computing power mode refers to the computing load allocation level (such as high computing power mode, medium computing power mode or low computing power mode) of the air conditioner main control chip, which is determined by mode matching based on the computing power status assessment results.

[0078] The computational precision configuration refers to the model inference numerical precision and operation strategy (such as INT32 full precision, INT8 quantization precision, or pausing training to retain only inference) determined after precision mapping based on the computing power status assessment results and applicable to the current computing power mode.

[0079] For example, during operation, the air conditioner's main control chip periodically collects its own CPU utilization, remaining SRAM capacity, and power supply voltage data, and combines these three data points to construct a three-dimensional resource vector. Subsequently, the main control chip substitutes this three-dimensional resource vector into a preset weighted comprehensive evaluation function for weighted calculation, generating a resource health index that reflects the current overall resource sufficiency and stability of the chip. Next, the main control chip introduces a dual-threshold hysteresis band mechanism to determine the state of the index, and combines it with preset hysteresis judgment duration data for time buffering and anti-jitter. Only when the resource health index crosses the high / low threshold and the duration reaches the hysteresis judgment duration is the computing power status evaluation result updated. Finally, based on the updated computing power status evaluation result, the main control chip performs pattern matching and precision mapping respectively to dynamically determine the air conditioner's current computing power mode (such as high, medium, and low computing power mode) and computing precision configuration (such as full precision INT32 or quantized INT8), thereby achieving efficient adaptive scheduling of hardware computing resources.

[0080] This invention, by integrating three-dimensional indicators of CPU utilization, SRAM memory, and power supply voltage, avoids the one-sidedness of evaluating a single resource indicator and can comprehensively and accurately grasp the true health load of edge chips. At the same time, by introducing a dual-threshold hysteresis band mechanism and hysteresis judgment duration for anti-jitter decision, it effectively filters the interference of instantaneous drastic resource fluctuations or power ripple on the evaluation results, completely solving the problem of frequent oscillations and ping-pong switching between computing power mode and computing accuracy near the critical point, significantly improving the stability and robustness of the edge computing system, and ensuring the continuous and stable operation of low-power IoT chips under various complex hardware conditions.

[0081] In an optional embodiment of the present invention, the pre-trained model comprises a three-layer neural network, and the step of generating initial artificial intelligence decision instructions for the air conditioner using the current computing power mode, the computing precision configuration, the standardized multidimensional feature vector, and the data validity flag includes: The standardized multidimensional feature vector and the data validity flag are input into the pre-trained model to generate initial feature input data; Based on the current computing power mode and the computing precision configuration, the hidden layer weights of the three-layer neural network are loaded and precision configuration is selected to generate target hidden layer weights; Based on the target hidden layer weights, forward inference calculations are performed on the initial feature input data to generate initial artificial intelligence decision instructions.

[0082] In this embodiment of the invention, a lightweight pre-trained model with a three-layer neural network structure is used. Combined with the current computing power mode and computational precision configuration of the air conditioner main control chip, the target hidden layer weights that match the current computing power resources are dynamically loaded. Forward inference calculation is performed on the standardized multi-dimensional features that incorporate data validity flags. Thus, in an offline state that is independent of cloud reliance, intelligent control command generation that adapts to hardware computing power constraints and takes into account decision accuracy and real-time performance is achieved.

[0083] The pre-trained model refers to a three-layer neural network control model that is integrated and deployed in the main control chip at the edge of the air conditioner, has been trained in advance, and has offline inference capabilities.

[0084] A three-layer neural network refers to a lightweight neural network architecture consisting of an input layer, a hidden layer, and an output layer, used to perform feature mapping and decision reasoning at the edge.

[0085] Standardized multidimensional feature vectors refer to data vectors that map the acquired multidimensional sensor signals to a uniform numerical range after preprocessing such as noise reduction and normalization.

[0086] The data validity flag is a binary bitmap identifier used to indicate whether the data collected by each sensor is normal and valid, and whether there are frame drops or hardware failures.

[0087] Initial feature input data refers to the comprehensive data vector input to the input layer of the pre-trained model after the standardized multidimensional feature vector and data validity flags are fused or concatenated.

[0088] The current computing power mode refers to the computing load allocation level (such as high computing power mode, medium computing power mode or low computing power mode) determined based on the real-time hardware resource status assessment of the main control chip.

[0089] The computational precision configuration refers to the type of numerical precision for model inference that matches the current computing power mode (such as INT32 full precision or INT8 quantization precision).

[0090] The target hidden layer weights refer to the neural network weight parameters after dynamically loading and executing the precision configuration selection for the three-layer hidden layer based on the current computing power mode and computing precision configuration.

[0091] Forward inference computation refers to the computational process in which initial feature input data is passed through the input layer, matrix multiplication and addition operations and activation function mapping are performed layer by layer in combination with the target hidden layer weights, and finally the decision result is generated in the output layer.

[0092] The initial AI decision instruction refers to the preliminary decision instruction output by the pre-trained model through forward inference calculation, which includes control parameters such as target temperature, fan speed setting, and compressor operating frequency.

[0093] For example, the air conditioner main control chip first fuses the standardized multidimensional feature vectors obtained offline with the data validity flags reflecting the health status of the sensors to generate initial feature input data for the input layer of the pre-trained model. Then, based on the computing power mode and computational precision configuration determined by the current evaluation, the system dynamically loads the hidden layer weights of the three-layer neural network adapted to the current computing power from the local storage (for example, loading high-precision INT32 weights in high computing power mode and quantized INT8 weights in low computing power mode) to generate target hidden layer weights. Finally, the main control chip uses the initial feature input data as input and performs forward inference calculations such as matrix multiplication and addition and activation mapping layer by layer in the three-layer neural network using the target hidden layer weights. Finally, it outputs the initial artificial intelligence decision instructions containing parameters such as target temperature, wind speed and compressor frequency in real time at the output layer.

[0094] This invention, through the construction of a lightweight pre-trained model with a three-layer neural network structure and the integration of sensor data validity flags into the initial feature input, ensures the robustness of inference calculations when sensor data exhibits occasional anomalies or failures. Simultaneously, by dynamically loading and selecting hidden layer weights based on the chip's current computing power mode and computational precision, it achieves an elastic match between the AI ​​model's inference computing power requirements and the actual computing resources of the edge hardware. This prevents chip overheating or system crashes caused by high loads while maximizing the utilization of available computing power to ensure decision accuracy, thus guaranteeing the efficiency, real-time performance, and high reliability of intelligent decision-making in offline and network-disconnected conditions.

[0095] In an optional embodiment of the present invention, after the step of controlling the air conditioner based on the first control signal when the air conditioner is in the communication link interruption state, the online self-evolutionary training process of the pre-trained model includes: Obtain the standardized multidimensional feature vector, the personalized user fingerprint vector expressing user preference parameters, and the environmental data within a set time window; Based on the preset equivalent coefficients, the temperature deviation, humidity deviation, wind speed perception deviation, and personnel activity status deviation in the environmental data within the set time window are uniformly normalized to generate normalized temperature dimension data. By combining the normalized temperature dimension data, the standardized multidimensional feature vector, and the personalized user fingerprint vector, residual calculation is performed to generate a multidimensional comfort residual vector and a comprehensive residual scalar. When it is determined that the current computing power mode meets the preset conditions, the output layer weights determined by the pre-trained model are used to perform backpropagation calculation on the comprehensive residual scalar to generate the hidden layer local gradient; Obtain offline duration data, and perform calculations based on the offline duration data and the comprehensive residual scalar to generate an adaptive learning rate; Obtain elastic weights to reinforce the regularization term; By combining the adaptive learning rate and the elastic weight consolidation regularization term, regularization and scaling calculations are performed on the local gradients of the hidden layer to generate gradient update values. Obtain the existing hidden layer weights of the pre-trained model, and perform update calculations on the existing hidden layer weights based on the gradient update amount to generate updated hidden layer weights; The updated hidden layer weights are loaded into the hidden layer of the pre-trained model to generate an updated pre-trained model; the updated pre-trained model is used to update the first control signal.

[0096] In this embodiment of the invention, during the period when the air conditioner is offline and disconnected from the network, the comprehensive comfort residual is calculated in real time using data such as multidimensional environmental deviation, personalized user preferences, and offline duration. When the chip computing power allows, the hidden layer weights of the local pre-trained model are fine-tuned online by combining adaptive learning rate and elastic weight consolidation regularization term. This enables online self-evolution and iteration of the edge AI model without cloud support, effectively preventing catastrophic forgetting while continuously improving the personalized adaptive capability of offline control commands to user comfort preferences.

[0097] The online self-evolutionary training process of a pre-trained model refers to the process of fine-tuning and incrementally learning a pre-trained model deployed on the edge using locally collected data and computing power in an offline state where the air conditioner is disconnected from the cloud, so that it can continuously optimize and evolve itself.

[0098] Personalized user fingerprint vectors refer to parameter vectors that reflect a specific user's unique preferences for environmental comfort such as temperature, humidity, and wind speed, abstracted from the user's historical habits, settings preferences, and interaction behaviors.

[0099] Environmental data within a set time window refers to data reflecting changes in indoor and outdoor environmental conditions that are continuously collected by sensors within a pre-set time period.

[0100] The equivalent coefficient refers to the preset conversion coefficient used to convert the deviations of different physical dimensions such as temperature, humidity, wind speed perception, and personnel activity status into equivalent temperature deviation values ​​according to the comfort influence weight.

[0101] Normalized temperature dimension data refers to the comprehensive environmental comfort deviation value generated by weighting and normalizing multidimensional environmental deviations using preset equivalent coefficients, with temperature as the unified dimension.

[0102] The multidimensional comfort residual vector refers to the feature vector that reflects the gap between multidimensional environmental indicators and the user's ideal comfort, calculated by combining normalized temperature dimension data, standardized multidimensional feature vectors, and personalized user fingerprint vectors.

[0103] The comprehensive residual scalar refers to a single numerical value that reflects the absolute degree of deviation in overall environmental comfort after performing scalar fusion calculation on the values ​​of each dimension in the multidimensional comfort residual vector.

[0104] Output layer weights refer to the neural network weight parameters in the output layer of a pre-trained model used to map hidden layer features to the final decision output.

[0105] The local gradient of the hidden layer refers to the gradient value calculated during backpropagation based on the comprehensive residual scalar and the output layer weights, which is used to characterize the rate of change of the parameter error of each node in the hidden layer.

[0106] Offline duration data refers to the continuous length of time that the air conditioner has remained in an offline operating state since the communication link was interrupted.

[0107] The adaptive learning rate refers to the adjustment factor used to control the step size and learning rate of model weight updates, which is dynamically calculated based on offline duration data and comprehensive residual scalars.

[0108] The elastic weight consolidation regularization term refers to the constraint term constructed based on the elastic weight consolidation algorithm. It prevents catastrophic forgetting of the model during incremental fine-tuning by penalizing modifications to the weights of important historical knowledge.

[0109] The gradient update amount refers to the gradient change vector calculated by combining the local gradient of the hidden layer, the adaptive learning rate, and the elastic weight consolidation regularization term, which is used to correct the existing hidden layer weights.

[0110] The existing hidden layer weights refer to the neural network weight parameters currently being used in the hidden layers of the air conditioner pre-trained model.

[0111] The updated hidden layer weights refer to the new hidden layer weight parameters generated after adjusting the existing hidden layer weights by adding or subtracting them using the gradient update amount.

[0112] The updated pre-trained model refers to a local neural network model that has been reloaded with updated hidden layer weights and has updated inference control capabilities.

[0113] For example, when the air conditioner is in an offline control state, the main control chip continuously collects standardized multidimensional feature vectors, personalized user fingerprint vectors, and environmental data within a set time window. It uses equivalent coefficients to normalize temperature, humidity, wind speed, and human activity deviations into normalized temperature dimension data, and further calculates multidimensional comfort residual vectors and comprehensive residual scalars. When the chip evaluates that the current computing power mode meets the preset online training conditions, it performs backpropagation calculation on the comprehensive residual scalar based on the output layer weights to derive the local gradient of the hidden layer. At the same time, it calculates the adaptive learning rate by combining the cumulative offline time and the comprehensive residual scalar, and calls up elastic weights to consolidate the regularization term. Subsequently, the main control chip uses the adaptive learning rate and regularization term to scale and regularize the local gradient of the hidden layer to generate gradient update, thereby correcting the existing hidden layer weights and reloading them into the model's hidden layer to generate an updated pre-trained model. Finally, the updated model is used to generate and correct the first control signal, realizing closed-loop self-evolutionary control in the offline state.

[0114] In this embodiment of the invention, by unifying and normalizing the multidimensional environmental deviations using equivalent coefficients, the complexity of calculating multi-objective comfort residuals on the edge side is significantly reduced. Simultaneously, by adaptively adjusting the learning rate based on offline duration and residuals, and introducing an elastic weight consolidation regularization mechanism, the model can be ensured to rapidly and adaptively evolve according to user preferences even under conditions of limited computing power at the edge. This effectively avoids "catastrophic forgetting" and model performance degradation caused by blind updates, greatly improving the personalized comfort adjustment effect and offline intelligent self-evolution capability of air conditioning under long-term network outage conditions.

[0115] In an optional embodiment of the present invention, the step of obtaining the cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner is in a communication link interruption state, and generating a progressive smooth target control instruction for eliminating the step jump when the air conditioner is in a cloud-edge state switch includes: Obtain the cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner is in a communication link interruption state; Normalization and projection mapping are performed on the cloud intent vector and the local state vector to generate a normalized projected intent vector and a normalized projected state vector. The semantic distance is calculated by combining the normalized projection intention vector and the normalized projection state vector to generate a hybrid semantic distance; When the mixed semantic distance is determined to be greater than a preset threshold, the real-time operating parameters of the air conditioner are obtained, and parameter construction is performed based on the real-time operating parameters to generate a dynamic mapping matrix and a dynamic bias vector. Incremental calculation is performed by combining the cloud intent vector and the cloud historical instructions to generate a cloud intent increment; Using the dynamic mapping matrix and the dynamic bias vector, matrix mapping and bias transformation are performed on the cloud intent increment to generate a local control increment; Acquire offline duration data and perform weight calculation on the offline duration data to determine the weighted fusion coefficient based on offline duration; By combining the weighted fusion coefficients and the preset adaptive first-order inertial filtering algorithm, a smoothing calculation is performed on the local control increment to generate a progressively smooth target control command.

[0116] In this embodiment of the invention, when the air conditioner restores network connection from an offline state, the mixed semantic distance between the cloud intent vector and the local state vector is calculated. When the heterogeneous deviation is large, a mapping matrix and bias vector are dynamically constructed based on real-time operating parameters. The cloud intent increment is accurately mapped to the control increment adapted to the local operating conditions. Then, a weighted fusion coefficient based on the offline duration and an adaptive first-order inertial filtering algorithm are used for smooth calculation. This eliminates the step jump and mode conflict caused by the heterogeneity of cloud and edge commands at the moment of network restoration, ensuring the smoothness, security and hardware reliability of the cloud and edge state switching process.

[0117] Cloud intent vectors refer to standardized parameter vectors generated by cloud servers that represent the cloud's desired control settings for air conditioning (such as target temperature, fan speed, and operating mode).

[0118] The local state vector refers to the parameter vector that is maintained locally and in real time by the air conditioner and reflects the current actual physical operating status of the equipment (such as real-time room temperature, fan speed, compressor operating frequency, etc.).

[0119] Historical commands in the cloud refer to historical control command data generated by the cloud server according to a predetermined strategy or by remote operation by the user during the communication link interruption.

[0120] Normalized projection intent vector refers to the dimensionless vector expression derived after performing linear normalization and spatial projection mapping on the cloud intent vector.

[0121] The normalized projected state vector refers to the dimensionless vector expression derived after performing linear normalization and projection mapping on the local state vector in the same transformation space.

[0122] Hybrid semantic distance refers to a scalar value used to quantify the semantic difference or state deviation between the normalized projected intention vector and the normalized projected state vector in the feature space.

[0123] Real-time operating parameters refer to the physical parameters (such as indoor and outdoor temperature difference, compressor real-time frequency, power supply voltage, etc.) that the air conditioner collects at the moment the network is restored, reflecting the current hardware load and environmental characteristics.

[0124] A dynamic mapping matrix refers to a transformation matrix that is dynamically constructed based on real-time operating parameters and is used to linearly transform cloud-based control characteristics to local hardware control dimensions.

[0125] The dynamic bias vector refers to the compensation vector that is dynamically constructed based on real-time operating parameters and is used to correct the inherent system bias during the cloud-edge dimension transformation process.

[0126] Cloud intent increment refers to the change in cloud control intent derived by calculating the difference between the cloud intent vector and the historical cloud command execution during the network outage.

[0127] Local control increment refers to the control adjustment amount adapted to local hardware generated after matrix mapping and bias correction of cloud intent increment using dynamic mapping matrix and dynamic bias vector.

[0128] Offline duration data refers to the cumulative time elapsed from the interruption of the communication link to the re-establishment of the connection by the air conditioner.

[0129] The weighted fusion coefficient refers to the attenuation factor calculated based on offline duration data according to a preset function, used to adjust the weight ratio of local historical state and new instructions in the cloud during the smoothing filtering process.

[0130] The adaptive first-order inertial filtering algorithm refers to a digital filtering algorithm that dynamically adjusts the time constant based on the weighted fusion coefficients and performs first-order inertial smoothing on the control input increment.

[0131] The progressive smooth target control command refers to the control command generated after adaptive first-order inertial filtering smoothing calculation, which gradually approaches the target in the cloud according to the adaptive rate of change within a set transition time.

[0132] For example, when the air conditioner communication link is restored, the system first obtains the cloud intent vector issued by the cloud server, the cloud historical instructions during the offline period, and the local state vector maintained in real time. The cloud intent vector and the local state vector are then normalized and spatially projected to generate a normalized projected intent vector and a normalized projected state vector. Subsequently, the mixed semantic distance between the two is calculated. If the distance is greater than a preset threshold (indicating a large deviation in the cloud-edge state), the main control chip collects real-time operating parameters such as indoor and outdoor temperature difference and compressor frequency. The parameters are used to construct a dynamic mapping matrix and a dynamic bias vector. At the same time, the difference between the cloud intent vector and the cloud historical instructions is calculated to obtain the cloud intent increment. The dynamic mapping matrix and the dynamic bias vector are used to perform matrix transformation and bias transformation on the cloud intent increment to generate a local control increment adapted to the local hardware. Finally, the main control chip obtains the offline duration data and calculates the weighted fusion coefficient based on the offline duration. This coefficient is then introduced into an adaptive first-order inertial filtering algorithm to perform smooth filtering calculation on the local control increment, deriving a progressive smooth target control instruction that gradually transitions within a specified time, thereby smoothly driving the air conditioner to evolve towards the cloud target state.

[0133] This invention, through projection mapping and hybrid semantic distance calculation, can accurately quantify and assess the degree of discrepancy between the cloud control intent and the actual local operating state after network outage recovery. When the deviation is large, by combining real-time operating parameters to construct a dynamic mapping matrix and bias vector, the cloud intent increment is converted into a local control increment that conforms to the local physical hardware constraints, thus eliminating control conflicts caused by cloud-edge semantic heterogeneity at the source. At the same time, it innovatively combines offline duration calculation with weighted fusion coefficients and adaptive first-order inertial filtering algorithm for smooth and gradual control, completely avoiding step jumps in parameters such as target temperature or compressor frequency during state switching, preventing frequent compressor start-stop or overload damage, effectively extending hardware lifespan, and achieving seamless, conflict-free, and highly smooth switching between cloud and edge states.

[0134] In an optional embodiment of the present invention, the air conditioner is equipped with a four-level safety state machine, and the step of generating a second control signal based on the standardized multidimensional feature vector, the progressive smooth target control command, and the initial artificial intelligence decision command includes: Obtain the initial artificial intelligence decision instruction, the progressive smooth target control instruction, and the standardized multidimensional feature vector; Obtain the current security state determined by the four-level security state machine; Combining the standardized multidimensional feature vector and the current security state, a three-layer progressive hard constraint verification, including hard truncation, feasible region projection mapping, and safe alternative solution, is performed on the initial artificial intelligence decision instruction or the progressive smooth target control instruction to generate a preliminary verification instruction. The minimum start-stop time constraint data and frequency change rate limit data of the inverter compressor of the air conditioner are obtained, as well as the air supply transition protection sequence for the cooling and heating mode switching process of the air conditioner. Based on the minimum start-stop time constraint data, the frequency change rate limit data, and the air supply transition protection sequence, the hardware physical constraint linkage verification is performed on the preliminary verification instruction to generate a final compliance control instruction that conforms to the safety boundary and hardware physical constraints, and the final compliance control instruction is converted into the second control signal.

[0135] This invention constructs a dual security protection architecture that integrates a four-level security state machine and hardware physical mechanism constraints. By sequentially executing a three-layer progressive hard constraint verification of the initial artificial intelligence decision command or the progressive smooth target control command—"hard truncation-feasible region projection mapping-safe alternative solution"—and linking it with physical constraints such as the start-stop time of the variable frequency compressor, the rate of frequency change, and the air supply transition protection sequence, it ensures that the generated second control signal strictly conforms to the system logic security and hardware physical boundaries under any operating condition, completely eliminating the risk of out-of-bounds commands and protecting core hardware such as the compressor.

[0136] A Level 4 safety state machine refers to a logic control model that is pre-built into the air conditioning main control system, divides the equipment operation safety into four levels (such as normal operation level, early warning frequency limiting level, fault degradation level and emergency shutdown level), and executes state transitions according to conditions.

[0137] The current safety status refers to the specific safety operating level of the air conditioner as determined by the Level 4 safety status machine based on real-time monitoring data.

[0138] Hard truncation refers to a direct truncation control strategy that forcibly clamps the command value to the upper or lower limit of the safety threshold when the command parameter exceeds the absolute safety threshold boundary (such as extreme temperature or extreme pressure).

[0139] Feasible region projection mapping refers to a correction method that maps control command vectors that exceed multidimensional safety constraints to the feasible solution space (feasible region) enclosed by multidimensional safety boundaries through geometric or mathematical projection algorithms.

[0140] Safety alternative solution refers to a mechanism that solves and outputs alternative safety control commands based on a preset safety policy library when control commands seriously violate safety rules and cannot be corrected by truncation or projection.

[0141] Three-level progressive hard constraint verification refers to the process of performing mandatory verification and correction of input instructions by sequentially performing three levels: hard truncation, feasible region projection mapping, and safe alternative solution.

[0142] Preliminary verification instructions refer to the phased control instructions that conform to the logical safety boundaries, derived after the initial artificial intelligence decision instructions or progressive smooth target control instructions have undergone three layers of progressive hard constraint verification.

[0143] The minimum start-stop time constraint data of variable frequency compressors refers to the minimum running time allowed to be shut down after the compressor is turned on and the minimum shutdown time allowed to be turned on again after the compressor is shut down, which are preset to prevent mechanical wear or high pressure damage caused by frequent start-stop of the compressor.

[0144] Frequency change rate limit data refers to the maximum allowable change range of the compressor operating frequency per unit time, which is preset to prevent sudden changes in the speed of the variable frequency compressor from causing current surges or poor lubrication.

[0145] The air supply transition protection sequence refers to the time-series protection rules that are pre-set to prevent cold air, accumulation of residual heat / cold energy, or impact of four-way valve reversal during the switching process between cooling and heating modes of an air conditioner. These rules include actions such as fan delay and air sweep transition.

[0146] Hardware physical constraint linkage verification refers to the process of co-verifying the hardware physical mechanism and protection rules of the preliminary verification command by combining the minimum start-stop time of the variable frequency compressor, the frequency change rate limit, and the air supply transition protection sequence.

[0147] The final compliance control instruction refers to a safety compliance instruction that simultaneously satisfies the linkage verification of three-layer progressive hard constraints and hardware physical constraints, and can be directly converted into a physical drive electrical signal.

[0148] For example, the air conditioner's main control chip first acquires the initial artificial intelligence decision instruction (or progressive smooth target control instruction) and standardized multi-dimensional feature vector, and reads the current operating safety status of the air conditioner from the four-level safety state machine. Subsequently, the main control chip combines the standardized multi-dimensional feature vector with the current safety status to perform a three-layer progressive hard constraint verification on the input instruction, consisting of hard truncation, feasible region projection mapping, and safe substitution solution, generating a logically safe preliminary verification instruction. Next, the main control chip retrieves the minimum start-stop time constraint data of the variable frequency compressor, the frequency change rate limit data, and the air supply transition protection sequence for switching between cooling and heating modes, and performs hardware physical constraint linkage verification on the preliminary verification instruction to ensure that the instruction will not cause frequent compressor start-stop, drastic frequency jumps, or mode switching shocks. Finally, the main control chip converts the final compliant control instruction generated through linkage verification into a second control signal to drive the smooth and safe operation of each hardware actuator of the air conditioner.

[0149] This invention, through the introduction of a four-level safety state machine and a three-layer progressive hard constraint verification of "hard truncation - feasible region projection mapping - safe alternative solution," completely resolves the safety risks of AI models or smooth transition algorithms potentially outputting out-of-bounds instructions, ensuring the absolute safety of the control logic under any abnormal operating conditions. Simultaneously, it further superimposes hardware physical constraint linkage verification such as compressor minimum start-stop time, frequency change rate limits, and air supply transition protection, effectively avoiding frequent compressor start-stop, current overload impacts, and commutation mechanical damage. This constructs a robust hardware protection barrier from both logical and physical dimensions, greatly improving the safety and reliability of system operation and extending the service life of critical hardware equipment.

[0150] In an optional embodiment of the present invention, the step of controlling the air conditioner based on the second control signal when the communication link of the air conditioner is restored includes: Obtain the second control signal; Based on the second control signal, the variable frequency compressor and fan terminal load of the air conditioner are driven to perform control actions; During the execution of the control actions by the variable frequency compressor and the fan terminal load, actual operating condition parameters are collected in real time, and actual operating status data of the equipment is generated based on the actual operating condition parameters; the actual operating status data of the equipment is used to update the local state vector.

[0151] In this embodiment of the invention, after the air conditioning communication link is restored, the variable frequency compressor and the fan terminal load are safely driven to perform control actions through the second control signal. During the hardware response process, the actual operating condition parameters are collected in real time to dynamically update the local state vector, thereby establishing a closed-loop feedback mechanism for hardware execution and state perception, ensuring real-time synchronization and accurate tracking of the physical operating status of the equipment and the local control dimension.

[0152] The second control signal refers to the final compliant control electrical signal generated after integrating progressive smooth target control instructions, local features, and AI decision instructions, and after verification by a safety state machine and hardware physical constraints, which is used to directly drive the hardware actuator.

[0153] Communication link restoration refers to the re-establishment of the network connection between the air conditioner and the cloud server and the restoration of normal data communication.

[0154] A variable frequency compressor is a core power actuator in air conditioning systems that can dynamically adjust the motor operating frequency and refrigerant circulation flow rate according to control signals.

[0155] The fan terminal load refers to the fan actuator located in the indoor and outdoor units of an air conditioner, which is driven by a motor to rotate the fan blades to achieve heat exchange and air delivery.

[0156] Control actions refer to the specific physical response behaviors of the variable frequency compressor and fan terminal load, such as frequency adjustment, start-stop control, and wind speed switching, performed under the drive of the second control signal.

[0157] Actual operating condition parameters refer to dynamic parameters (such as compressor actual frequency, operating current, exhaust temperature, and fan speed) that are collected in real time by sensors during the execution of control actions by the hardware, reflecting the physical operating state and environmental characteristics.

[0158] Equipment actual operating status data refers to the data expression that reflects the current true physical state of the equipment after cleaning and structuring the actual operating condition parameters.

[0159] The local state vector refers to the parameter vector that is maintained and updated in real time by the local air conditioning control system, and is used to characterize the current actual operating status of the equipment and the subsequent cloud-edge state collaboration benchmark.

[0160] For example, when the air conditioning communication link is restored, the main control chip obtains the second control signal generated after security verification and converts it into a drive pulse to drive the variable frequency compressor to run at a set frequency and the fan terminal load to rotate at a set wind speed. While the variable frequency compressor and fan are running, the main control chip collects actual operating condition parameters such as the actual frequency of the compressor, operating current, exhaust temperature and actual speed of the fan in real time through onboard sensors and sampling circuits, and organizes them to generate actual operating status data of the equipment. Subsequently, the main control chip uses the actual operating status data of the equipment to refresh and overwrite the local state vector stored locally in real time, thereby completing the closed-loop control process of hardware action execution and state update.

[0161] In this embodiment of the invention, by simultaneously collecting actual operating parameters and updating the local state vector in real time while the hardware is driven by the second control signal, a real-time closed-loop feedback mechanism of "drive execution - state perception - vector update" is constructed. This completely solves the problem of the disconnect between the actual operating state of the hardware and the logical state of the system under open-loop control. At the same time, the accurately updated local state vector provides reliable reference data for subsequent cloud-edge state collaboration and progressive smooth control, ensuring the accuracy and continuity of the control response, further eliminating the risk of hardware malfunction, and improving the robustness and stability of the system.

[0162] To enable those skilled in the art to better understand the embodiments of the present invention, an example is used below to illustrate the embodiments of the present invention.

[0163] refer to Figure 2 , Figure 2 This is a schematic diagram of the structure of an air conditioning control system provided in an embodiment of the present invention; Specifically, it consists of seven functional modules: main control unit (MCU), local sensing unit, self-evolving decision module, bidirectional state translation module, offline strategy pool, computing power scheduling unit, and device execution unit.

[0164] 1. System hardware architecture and connectivity The smart home control system described in this embodiment of the invention mainly includes: a main control unit (MCU), a local sensing unit, a self-evolving decision module, a two-way state translation module, an offline strategy pool, a computing power scheduling unit, and a device execution unit.

[0165] Main control unit: The overall carrier and scheduling center of the system, which uniformly manages all hardware buses, memory flash partitions, timing scheduling of various software modules, relays all data and control signals, and ensures the basic operation of the system with a hardware watchdog.

[0166] Local sensing unit: The full-scene data acquisition input terminal, which periodically collects raw signals such as temperature and humidity, human body, doors and windows, and equipment current, performs noise reduction preprocessing and standardization, and synchronously reports sensor faults.

[0167] Self-evolving decision-making module: The core of local intelligent control, relying on a lightweight neural network to output comfort adjustment commands, and using a 15-minute short-window local gradient algorithm to iterate the model offline, adapting to dynamic changes in the environment, and can receive cloud-based preference optimization parameters.

[0168] Two-way state translation module: The core of cloud-edge interaction and smooth synchronization, realizing two-way semantic conversion between cloud intent vectors and local state vectors, real-time detection of network connectivity, caching cloud commands when the network is disconnected, and generating progressive commands through filtering algorithms when the network is restored to eliminate state switching jumps.

[0169] Offline policy pool storage: a safety fallback module that stores fixed and immutable if-then safety rules with higher priority than AI decisions. It intercepts and corrects dangerous commands that exceed limits, and independently takes over the basic control of the equipment when the AI ​​is shut down, handling abnormal operating conditions such as sensors, overload, and low voltage.

[0170] Computing power scheduling unit: Hardware resource management module, which monitors CPU, memory and power supply voltage every 2 seconds, divides the computing power into high, medium and low three-level computing power modes, controls the calculation accuracy of the self-evolution module and starts and stops the training process, so as to achieve a balance between power consumption and computing power.

[0171] Equipment execution unit: The instruction landing driver layer receives and verifies the control signals to drive loads such as air conditioners and fans, collects and transmits equipment operating conditions in real time to form a closed loop, and cuts off high-power hardware output when overloaded.

[0172] The present invention preferably uses a low-power IoT main control chip as the core hardware carrier.

[0173] The main control unit is the core of the system's global scheduling and resource management. It is responsible for the timing scheduling of various functional modules, hardware resource allocation, global data interaction, and abnormal status monitoring. It coordinates the system's network operation, offline autonomy, and network recovery synchronization across all operating conditions. The local sensing unit connects to various types of external sensing devices via I2C and GPIO hardware buses. With a fixed sampling period of 60 seconds, it collects multi-dimensional environmental and equipment parameters such as indoor temperature and humidity, human presence status, and equipment operating conditions. After performing noise reduction and anomaly removal preprocessing on the raw sensor data, it generates standardized multi-dimensional feature vectors and uploads them to the decision module, providing accurate data support for model iteration optimization and closed-loop control decisions.

[0174] The self-evolving decision-making module is the core innovative functional module of this invention. It deploys a lightweight three-layer neural network architecture on the edge device side, specifically including an input layer, an adaptive hidden layer, and an output layer. This invention uses a 15-minute environmental stability time window (N=15). Based on the residual parameters of real-time sensor data and user comfort thresholds, it relies on the local gradient descent algorithm to perform real-time fine-tuning and iteration of the neural network hidden layer weights. It does not require calling the global historical dataset for backpropagation calculations, and the overall CPU utilization is less than 10%. This offline autonomous iteration mechanism overcomes the limitations of traditional cloud-based static models, enabling it to autonomously adapt to dynamic changes in the home environment, such as seasonal changes and scene layout changes, even without network connectivity. This effectively improves the accuracy of intelligent control and the comfort performance of living spaces in long-term offline scenarios.

[0175] The bidirectional state translation module constructs a dual-space dynamic semantic mapping architecture for cloud-based intent vectors and local state vectors. The cloud generates a control intent vector containing target temperature, operating mode, and fan speed, while the local edge node maintains a device state vector in real time containing compressor operating frequency, fan speed, and indoor temperature difference. When the network link is restored, this invention abandons the traditional alignment method of fixed parameter coverage and fixed matrix mapping. Based on real-time operating conditions, it dynamically generates semantic mapping relationships, transforming cloud-based control intents into incremental control commands that local devices can recognize. Combined with a first-order inertial filtering algorithm with a filtering coefficient of 0.18, it constructs a 120-second progressive smooth control curve, completely eliminating command step transitions during cloud-edge state switching, avoiding frequent device start-stop failures, and effectively extending the hardware lifespan of smart home devices.

[0176] The computing power scheduling unit constructs a three-level dynamic computing power scheduling mechanism (high, medium, and low) based on three-dimensional hardware resource parameters: MCU CPU utilization, remaining memory capacity, and power supply voltage. The high computing power mode employs INT32 full-precision computation to ensure the accuracy of neural network iterative inference. The medium computing power mode performs INT8 quantization compression on model parameters and extends the model iteration cycle, reducing device power consumption within a controllable accuracy loss range. The low computing power mode suspends the AI ​​model training and iteration process, only fixing the basic control logic and security strategies. This effectively solves the computing power bottleneck and energy consumption contradiction of low-power IoT chips carrying intelligent algorithms, significantly lowering the hardware deployment threshold and improving system operation under extreme resource-constrained conditions.

[0177] The offline policy pool constructs a layered, redundant security protection architecture. The bottom layer is embedded with an if-then logic-based conditional action rule engine, independently responding to emergency conditions such as overheating, device overload, sensor data failure, and low-voltage power supply anomalies. The upper-layer self-evolving decision module is responsible for comfort and energy efficiency optimization control under normal operating conditions, and all AI decision output commands are constrained by the bottom-layer security rule boundaries, achieving bidirectional coupling of intelligent optimization and security safeguards. Compliant control commands, after security verification, are sent to the device execution unit, which completes the terminal device drive control and transmits device operating condition parameters back in real time, forming a complete closed-loop control logic. Ultimately, this achieves uninterrupted, highly secure, and adaptive intelligent operation of smart home devices across all operating conditions—networked, offline, and network restored.

[0178] Optionally, the self-evolutionary decision-making module can employ a lightweight reinforcement learning (RL) framework. The state space SSS is defined as the environmental parameters, the action space AAA as the device control commands, and the reward function RRR includes not only temperature error penalties but also energy consumption penalties. λP, where λ is a preset weight coefficient and P is the real-time operating power of the device. During network outages, the agent tries different actions by interacting with the environment and updates the Q-value table or policy network based on the comprehensive reward. Compared to gradient descent in purely supervised learning, RL is better at discovering the long-term energy-efficient policy combinations. When the network is restored, the locally accumulated Q-value table fragments are uploaded to the cloud for fusion training.

[0179] Optionally, a Bluetooth / UWB near-field positioning module or a fingerprint sensor can be added to the local sensing unit. During offline learning, the self-evolving decision-making module not only records environmental data but also associates it with a specific user ID. When different users are detected returning home, the module automatically loads the user's historical preference model (e.g., Zhang San prefers 24°C, Li Si prefers 26°C). Upon network recovery, the bidirectional translation module not only synchronizes the status but also the user's personalized preference parameters, achieving personalized cloud-edge synchronization for each user.

[0180] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, 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 the embodiments of the present invention.

[0181] Reference Figure 3 The diagram shows a structural block diagram of an air conditioning control device provided in an embodiment of the present invention, which may specifically include the following modules: The signal acquisition module 301 is used to acquire multi-dimensional sensing signals according to a preset sampling period when it is determined that the air conditioner is in a communication link interruption state, and to generate a standardized multi-dimensional feature vector and a data validity flag bit through the multi-dimensional sensing signals. The hardware resource parameter acquisition module 302 is used to acquire the hardware resource parameters of the air conditioner and determine the current computing power mode and computing precision configuration of the air conditioner based on the hardware resource parameters. The initial artificial intelligence decision instruction generation module 303 is used to call the pre-trained model integrated into the air conditioner and generate an initial artificial intelligence decision instruction for the air conditioner through the current computing power mode, the computing precision configuration, the standardized multidimensional feature vector and the data validity flag. The first control signal conversion module 304 is used to convert the initial artificial intelligence decision instruction into a first control signal for the air conditioner, and control the air conditioner based on the first control signal when the air conditioner is in the communication link interruption state; The progressive smooth target control instruction generation module 305 is used to respond to the restoration of the communication link of the air conditioner, obtain the cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner is in the communication link interruption state, and generate progressive smooth target control instructions to eliminate the step jump when the air conditioner is in the cloud edge state switching. The second control signal generation module 306 is used to generate a second control signal based on the standardized multidimensional feature vector, the progressive smooth target control command, and the initial artificial intelligence decision command; Air conditioning control module 307 is used to control the air conditioner when the communication link of the air conditioner is restored based on the second control signal.

[0182] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0183] In addition, embodiments of the present invention also provide an electronic device, such as... Figure 4 As shown, it includes a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404. Memory 403 is used to store computer programs; When the processor 401 executes the program stored in the memory 403, it implements any of the air conditioning control methods described in the above embodiments: The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0184] The communication interface is used for communication between the aforementioned terminal and other devices.

[0185] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0186] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0187] like Figure 5As shown, in another embodiment of the present invention, a computer-readable storage medium 501 is also provided, which stores instructions that, when executed on a computer, cause the computer to perform the air conditioning control method described in the above embodiment.

[0188] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described air conditioning control method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0189] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0190] It should be noted that, in this document, the terms include, encompass, or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "including a…" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0191] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0192] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. An air conditioning control method, characterized in that, The method is applied to the air conditioner, which is used to perform data interaction with a cloud server, and the method includes: When it is determined that the air conditioner is in a communication link interruption state, multi-dimensional sensing signals are collected according to a preset sampling period, and standardized multi-dimensional feature vectors and data validity flag bits are generated through the multi-dimensional sensing signals. Obtain the hardware resource parameters of the air conditioner, and determine the current computing power mode and computing precision configuration of the air conditioner based on the hardware resource parameters; The pre-trained model integrated into the air conditioner is invoked to generate initial artificial intelligence decision instructions for the air conditioner through the current computing power mode, the computing precision configuration, the standardized multidimensional feature vector, and the data validity flag. The initial artificial intelligence decision instruction is converted into a first control signal for the air conditioner, and the air conditioner is controlled based on the first control signal when the air conditioner is in the communication link interruption state; In response to the restoration of the communication link of the air conditioner, the cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner was in the communication link interruption state are obtained, and a progressive smooth target control instruction is generated to eliminate the step jump when the air conditioner is in the cloud edge state switching. Based on the standardized multidimensional feature vector, the progressive smooth target control command, and the initial artificial intelligence decision command, a second control signal is generated; The air conditioner is controlled based on the second control signal when the communication link of the air conditioner is restored.

2. The method according to claim 1, characterized in that, The hardware resource parameters include central processing unit utilization, remaining static random access memory capacity, and power supply voltage data. The step of determining the current computing power mode and computing precision configuration of the air conditioner based on the hardware resource parameters includes: A three-dimensional resource vector is generated by performing vector construction on the CPU utilization, the remaining static random access memory capacity, and the power supply voltage data. The three-dimensional resource vector is weighted and calculated using a preset weighted comprehensive evaluation function to generate a resource health index; The resource health index is subjected to hysteresis state determination by combining the preset hysteresis determination duration data and the dual threshold hysteresis band mechanism, and a computing power state assessment result is generated. Based on the computing power status assessment results, a matching process is performed to determine the current computing power mode; Based on the computing power status assessment results, precision mapping is performed to generate the computing precision configuration.

3. The method according to claim 1, characterized in that, The pre-trained model comprises a three-layer neural network. The step of generating initial AI decision instructions for the air conditioner using the current computing power mode, the computational precision configuration, the standardized multidimensional feature vector, and the data validity flag includes: The standardized multidimensional feature vector and the data validity flag are input into the pre-trained model to generate initial feature input data; Based on the current computing power mode and the computing precision configuration, the hidden layer weights of the three-layer neural network are loaded and precision configuration is selected to generate target hidden layer weights; Based on the target hidden layer weights, forward inference calculations are performed on the initial feature input data to generate initial artificial intelligence decision instructions.

4. The method according to claim 1, characterized in that, Following the step of controlling the air conditioner based on the first control signal when the air conditioner is in a communication link interruption state, the online self-evolutionary training process of the pre-trained model includes: Obtain the standardized multidimensional feature vector, the personalized user fingerprint vector expressing user preference parameters, and the environmental data within a set time window; Based on the preset equivalent coefficients, the temperature deviation, humidity deviation, wind speed perception deviation, and personnel activity status deviation in the environmental data within the set time window are uniformly normalized to generate normalized temperature dimension data. By combining the normalized temperature dimension data, the standardized multidimensional feature vector, and the personalized user fingerprint vector, residual calculation is performed to generate a multidimensional comfort residual vector and a comprehensive residual scalar. When it is determined that the current computing power mode meets the preset conditions, the output layer weights determined by the pre-trained model are used to perform backpropagation calculation on the comprehensive residual scalar to generate the hidden layer local gradient; Obtain offline duration data, and perform calculations based on the offline duration data and the comprehensive residual scalar to generate an adaptive learning rate; Obtain elastic weights to reinforce the regularization term; By combining the adaptive learning rate and the elastic weight consolidation regularization term, regularization and scaling calculations are performed on the local gradients of the hidden layer to generate gradient update values. Obtain the existing hidden layer weights of the pre-trained model, and perform update calculations on the existing hidden layer weights based on the gradient update amount to generate updated hidden layer weights; The updated hidden layer weights are loaded into the hidden layer of the pre-trained model to generate an updated pre-trained model; the updated pre-trained model is used to update the first control signal.

5. The method according to claim 1, characterized in that, The step of obtaining the cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner is in a communication link interruption state, and generating a progressive smooth target control instruction to eliminate the step jump when the air conditioner is in a cloud-edge state switch includes: Obtain the cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner is in a communication link interruption state; Normalization and projection mapping are performed on the cloud intent vector and the local state vector to generate a normalized projected intent vector and a normalized projected state vector. The semantic distance is calculated by combining the normalized projection intention vector and the normalized projection state vector to generate a hybrid semantic distance; When the mixed semantic distance is determined to be greater than a preset threshold, the real-time operating parameters of the air conditioner are obtained, and parameter construction is performed based on the real-time operating parameters to generate a dynamic mapping matrix and a dynamic bias vector. Incremental calculation is performed by combining the cloud intent vector and the cloud historical instructions to generate a cloud intent increment; Using the dynamic mapping matrix and the dynamic bias vector, matrix mapping and bias transformation are performed on the cloud intent increment to generate a local control increment; Acquire offline duration data and perform weight calculation on the offline duration data to determine the weighted fusion coefficient based on offline duration; By combining the weighted fusion coefficients and the preset adaptive first-order inertial filtering algorithm, a smoothing calculation is performed on the local control increment to generate a progressively smooth target control command.

6. The method according to claim 1, characterized in that, The air conditioner is equipped with a four-level safety state machine. The step of generating the second control signal based on the standardized multidimensional feature vector, the progressive smooth target control command, and the initial artificial intelligence decision command includes: Obtain the initial artificial intelligence decision instruction, the progressive smooth target control instruction, and the standardized multidimensional feature vector; Obtain the current security state determined by the four-level security state machine; Combining the standardized multidimensional feature vector and the current security state, a three-layer progressive hard constraint verification, including hard truncation, feasible region projection mapping, and safe alternative solution, is performed on the initial artificial intelligence decision instruction or the progressive smooth target control instruction to generate a preliminary verification instruction. The minimum start-stop time constraint data and frequency change rate limit data of the inverter compressor of the air conditioner are obtained, as well as the air supply transition protection sequence for the cooling and heating mode switching process of the air conditioner. Based on the minimum start-stop time constraint data, the frequency change rate limit data, and the air supply transition protection sequence, the hardware physical constraint linkage verification is performed on the preliminary verification instruction to generate a final compliance control instruction that conforms to the safety boundary and hardware physical constraints, and the final compliance control instruction is converted into the second control signal.

7. The method according to claim 6, characterized in that, The step of controlling the air conditioner based on the second control signal when the air conditioner's communication link is restored includes: Obtain the second control signal; Based on the second control signal, the variable frequency compressor and fan terminal load of the air conditioner are driven to perform control actions; During the execution of the control actions by the variable frequency compressor and the fan terminal load, actual operating condition parameters are collected in real time, and actual operating status data of the equipment is generated based on the actual operating condition parameters; the actual operating status data of the equipment is used to update the local state vector.

8. An air conditioning control device, characterized in that, The device is applied to the air conditioner, which is used to perform data interaction with a cloud server, and the device includes: The signal acquisition module is used to acquire multi-dimensional sensor signals according to a preset sampling period when it is determined that the air conditioner is in a communication link interruption state, and to generate a standardized multi-dimensional feature vector and a data validity flag bit through the multi-dimensional sensor signals. The hardware resource parameter acquisition module is used to acquire the hardware resource parameters of the air conditioner and determine the current computing power mode and computing precision configuration of the air conditioner based on the hardware resource parameters. An initial artificial intelligence decision command generation module is used to call the pre-trained model integrated into the air conditioner and generate an initial artificial intelligence decision command for the air conditioner through the current computing power mode, the computing precision configuration, the standardized multidimensional feature vector and the data validity flag. The first control signal conversion module is used to convert the initial artificial intelligence decision instruction into a first control signal for the air conditioner, and to control the air conditioner based on the first control signal when the air conditioner is in the communication link interruption state; A progressive smooth target control instruction generation module is used to respond to the restoration of the communication link of the air conditioner, obtain the cloud intent vector generated by the cloud server, the local state vector of the air conditioner, and the cloud historical instructions generated by the cloud server when the air conditioner is in the communication link interruption state, and generate progressive smooth target control instructions to eliminate the step jump when the air conditioner is in the cloud edge state switching. The second control signal generation module is used to generate a second control signal based on the standardized multidimensional feature vector, the progressive smooth target control command, and the initial artificial intelligence decision command. An air conditioning control module is used to control the air conditioner when the communication link of the air conditioner is restored based on the second control signal.

9. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method as described in any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the method as described in any one of claims 1-7.

Images