Fan self-adaptive energy-saving control method and system for active noise reduction

Abstract

The invention relates to the technical field of fan intelligent control and active noise reduction, in particular to a fan self-adaptive energy-saving control method and system for active noise reduction, and the method comprises the steps: determining the rotating speed of a target fan through a thermal inertia prediction model; a step of establishing a personalized psychological acoustic model through online reinforcement learning and virtual sensing to determine a sound quality target; and seamless adjustment of noise reduction parameters is realized through the predictive control model when the wind speed changes. According to the invention, through deep fusion of prediction control and online learning, the energy-saving and noise-reducing performance of the fan is collaboratively optimized. The method not only can predictively adjust the wind speed to achieve high efficiency and energy conservation, but also can learn user preferences to provide personalized acoustic comfortable experience, ensures that the noise reduction effect is stable and seamless in the dynamic speed change process, and overcomes various defects of a traditional control mode.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control and active noise reduction of fans, and in particular to a fan adaptive energy-saving control method and system for active noise reduction. Background Art

[0002] Fan systems are key equipment in HVAC and data centers, used to regulate ambient temperature and air quality.

[0003] Existing fan energy-saving controls are mostly passive and inefficient. Active noise reduction technologies focus solely on reducing noise loudness, ignoring the personalized needs of sound quality. Furthermore, the noise reduction effect can deteriorate when the fan changes speed.

[0004] Therefore, it is urgent to solve the problem of how to achieve predictive and personalized coordinated control of fan energy saving and noise reduction.

[0005] To this end, a fan adaptive energy-saving control method and system for active noise reduction are proposed. Summary of the Invention

[0006] The purpose of the present invention is to provide a fan adaptive energy-saving control method and system for active noise reduction. By establishing a prediction model that can predict thermal load and collaborative control parameters, and combining it with an adaptive noise quality control method that can learn users' personalized acoustic preferences online, the problems of mutual interference between energy saving and noise reduction, poor user acoustic experience and slow system dynamic response in the existing technology are solved, and the unity of high energy efficiency, high comfort and high stability of fan operation is achieved.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A fan adaptive energy-saving control method and system for active noise reduction, comprising:

[0009] Establishing a thermal inertia prediction model, inputting environmental parameters representing the indoor environmental state and load parameters representing the heat source state, predicting the heat dissipation demand of the fan within a future preset time period, and calculating a target fan speed that meets the heat dissipation demand;

[0010] An online reinforcement learning method is used to establish and dynamically optimize a personalized psychoacoustic model; wherein the feedback signal of the online reinforcement learning method is derived from the interpretation of user operation behavior; virtual sensing technology is used to perceive the user's context, and based on the context and the personalized psychoacoustic model, the optimal sound quality target is determined and an active noise reduction algorithm is executed;

[0011] A predictive control model is established. When the target fan speed is different from the current speed, the target fan speed is input into the predictive control model to determine the optimal parameters corresponding to the target fan speed; and the parameters of the active noise reduction algorithm are synchronously adjusted to the optimal parameters.

[0012] Preferably, the environmental parameters include temperature, humidity, static pressure and CO2 concentration; the load parameters include the real-time calculated load of the server as a thermal load indicator.

[0013] Preferably, the thermal inertia prediction model includes:

[0014] A data acquisition and preprocessing unit, configured to receive and process real-time and historical data of the environmental parameters and load parameters;

[0015] A feature extraction and correlation analysis unit, configured to learn the nonlinear mapping relationship and time delay characteristics between the environmental parameters, load parameters, and heat dissipation requirements during an offline training phase;

[0016] A time series prediction core unit, configured to predict the load state within the preset future time period according to the output of the feature extraction and correlation analysis unit during operation;

[0017] The demand calculation unit is used to convert the predicted load state into a quantitative value of the heat dissipation demand and calculate the target fan speed.

[0018] Preferably, the personalized psychoacoustic model includes:

[0019] an acoustic feature analysis unit for extracting acoustic features representing sound quality from the residual noise generated by the active noise reduction algorithm;

[0020] A context perception unit, which uses the virtual sensing technology to determine the user's current application context;

[0021] A user preference learning unit, which uses the online reinforcement learning method to use the interpreted user operation behavior as a feedback signal to establish and update a user model that represents the user's preference for different acoustic features in different application scenarios;

[0022] The sound quality target generation unit is used to comprehensively calculate the current optimal sound quality target based on the acoustic characteristics, the application scenario and the user model.

[0023] Preferably, the active noise reduction algorithm calculates and generates an anti-phase noise sound field that cancels out the original fan noise based on the acoustic sensing of the original fan noise according to the current optimal sound quality target;

[0024] The residual noise is a synthetic sound field formed by acoustic interference between the original fan noise and the anti-phase noise sound field within a preset target noise reduction area. The sensing result of the residual noise is used as feedback to dynamically adjust the active noise reduction algorithm.

[0025] Preferably, the predictive control model includes:

[0026] A speed input unit, configured to receive the target fan speed;

[0027] a parameter mapping unit, which obtains a mapping model through offline training, wherein the mapping model represents the correspondence between different fan speeds and the optimal parameters of the active noise reduction algorithm, and the parameter mapping unit is used to determine the optimal parameters through the mapping model according to the input target fan speed;

[0028] A parameter output unit is used to output the determined optimal parameters.

[0029] Preferably, the predictive control model includes:

[0030] A speed input unit, configured to receive the target fan speed;

[0031] a parameter mapping unit, which obtains a mapping model through offline training, wherein the mapping model represents the correspondence between different fan speeds and the optimal parameters of the active noise reduction algorithm, and the parameter mapping unit is used to determine the optimal parameters through the mapping model according to the input target fan speed;

[0032] A parameter output unit is used to output the determined optimal parameters.

[0033] Preferably, the user operation behavior includes: an instruction to manually adjust the fan speed; an instruction to turn off the fan; and an instruction to switch the fan operation mode.

[0034] A fan adaptive energy-saving control system for active noise reduction, comprising:

[0035] Demand prediction and speed control module: establishes a thermal inertia prediction model, inputs environmental parameters representing the indoor environmental state and load parameters representing the heat source state, predicts the heat dissipation demand of the fan within a preset time period in the future, and calculates the target fan speed that meets the heat dissipation demand;

[0036] Adaptive Noise Quality Control Module: This module uses online reinforcement learning to establish and dynamically optimize a personalized psychoacoustic model, where the feedback signal of the online reinforcement learning method is derived from the interpretation of user operation behavior. It uses virtual sensing technology to perceive the user's context, and based on this context and the personalized psychoacoustic model, it determines the optimal sound quality target and runs the active noise reduction algorithm.

[0037] Transient collaborative control module: establish a predictive control model, when the target fan speed is different from the current speed, input the target fan speed into the predictive control model, determine the optimal parameters corresponding to the target fan speed; synchronously implement the active noise reduction algorithm.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. By establishing a thermal inertia prediction model, this invention can predict future trends in thermal load changes, enabling proactive and smooth adjustment of fan speed. This avoids the frequent starts and stops or drastic speed fluctuations caused by the passive response lag of traditional control methods. This predictive feedforward control ensures that the fan consistently operates near the minimum energy consumption point that meets demand, significantly improving the overall energy efficiency of the system.

[0040] 2. By incorporating online reinforcement learning and a personalized psychoacoustic model, this invention transcends the limitations of traditional noise reduction technologies, which focus solely on reducing volume. The system learns and adapts to a specific user's sound preferences. Combined with virtual sensing technology, it dynamically adjusts noise reduction targets based on the user's context, proactively shaping residual noise to a more acceptable level. This achieves a leap from noise reduction to optimizing sound quality, providing users with a personalized, high-quality acoustic comfort experience.

[0041] 3. By establishing a predictive control model for transient management, this invention effectively addresses the technical challenge of active noise reduction system performance degradation or even failure during wind turbine speed changes. Before the wind speed changes, the system pre-calculates the optimal noise reduction parameters for the new operating conditions and adjusts them synchronously, ensuring seamless control. This eliminates noise leakage and system instability caused by sudden changes in operating conditions, improving the system's reliability and stability during dynamic operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A method flow chart of a fan adaptive energy-saving control method for active noise reduction proposed in an embodiment of the present invention;

[0043] Figure 2 This is a structural diagram of the thermal inertia prediction model proposed in an embodiment of the present invention;

[0044] Figure 3This is a system structure diagram of a fan adaptive energy-saving control system for active noise reduction proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] Example 1

[0047] See also Figures 1 to 3 The present invention provides a fan adaptive energy-saving control method for active noise reduction, and the technical solution is as follows:

[0048] A fan adaptive energy-saving control method for active noise reduction, comprising:

[0049] Establishing a thermal inertia prediction model, inputting environmental parameters representing the indoor environmental state and load parameters representing the heat source state, predicting the heat dissipation demand of the fan within a future preset time period, and calculating a target fan speed that meets the heat dissipation demand;

[0050] An online reinforcement learning method is used to establish and dynamically optimize a personalized psychoacoustic model; wherein the feedback signal of the online reinforcement learning method is derived from the interpretation of user operation behavior; virtual sensing technology is used to perceive the user's context, and based on the context and the personalized psychoacoustic model, the optimal sound quality target is determined and an active noise reduction algorithm is executed;

[0051] A predictive control model is established. When the target fan speed is different from the current speed, the target fan speed is input into the predictive control model to determine the optimal parameters corresponding to the target fan speed; and the parameters of the active noise reduction algorithm are synchronously adjusted to the optimal parameters.

[0052] Furthermore, the environmental parameters include temperature, humidity, static pressure and CO2 concentration; and the load parameters include a real-time calculated load of the server as a thermal load indicator.

[0053] In this embodiment, the server real-time calculation load is defined as the CPU packet power consumption obtained through the server out-of-band management interface, in watts (W).

[0054] By introducing diverse parameters such as CO2 concentration and server load, the system can accurately distinguish between real heat sources such as occupants and equipment, greatly improving the accuracy of predicting future cooling needs, thereby achieving more precise and efficient energy-saving control.

[0055] Further, if Figure 2 As shown, the thermal inertia prediction model includes:

[0056] A data acquisition and preprocessing unit, configured to receive and process real-time and historical data of the environmental parameters and load parameters;

[0057] A feature extraction and correlation analysis unit, configured to learn the nonlinear mapping relationship and time delay characteristics between the environmental parameters, load parameters, and heat dissipation requirements during an offline training phase;

[0058] A time series prediction core unit, configured to predict the load state within the preset future time period according to the output of the feature extraction and correlation analysis unit during operation;

[0059] The demand calculation unit is used to convert the predicted load state into a quantitative value of the heat dissipation demand and calculate the target fan speed.

[0060] Before deploying this method, a one-time offline calibration of the target device is required to determine the key parameters in the thermal inertia prediction model. The calibration method includes:

[0061] Place the target server in a controlled environment and use software to run it at different computing load levels. Use a power meter to measure the server's actual power consumption under each load. After deducting the baseline power consumption, establish a "computing load - heat generation power" relationship model. This determines the baseline cooling requirement and load-to-watt conversion factor.

[0062] In the target chassis or environment, set the fan to run at a series of fixed speeds. At each speed point, measure the temperature of the fully loaded server and record the stable core temperature. Based on thermal resistance theory, establish a mapping between fan speed and heat dissipation capacity.

[0063] Combining the above two calibration models and the user's tolerance to noise, the heat dissipation demand thresholds and corresponding fan speeds for dividing the three intervals into low, medium, and high are determined.

[0064] This calibration method establishes a refined thermal management model by precisely measuring and calculating the relationships between load and heat generation, and between fan speed and heat dissipation capacity. This enables the system to intelligently and efficiently adjust fan speed based on actual cooling needs while meeting user noise requirements, achieving the optimal balance between energy consumption and heat dissipation.

[0065] Preferably, the time series prediction core unit in this embodiment adopts a long short-term memory network (LSTM) model;

[0066] The LSTM model architecture of the core time series forecasting unit is as follows: the input data consists of environmental and load parameters from the past 60 minutes (sampled once per minute), forming an input sequence with a shape of (60, 4) (4 features). The network consists of two stacked LSTM layers, each with 128 memory cells. A fully connected layer with 32 neurons and a ReLU activation function is connected after the LSTM layer. Finally, an output layer is used to predict the average load status for the next 15 minutes. During training, the Adam optimizer is used, with an initial learning rate of 0.001 and a batch size of 64.

[0067] During the offline training phase, the feature extraction and correlation analysis unit collects at least 1000 hours of historical data including environmental parameters, load parameters, and corresponding power consumption / temperature changes as a training set, and trains the LSTM network using a backpropagation algorithm to learn the time dependency and nonlinear relationship between various input parameters and heat dissipation requirements;

[0068] The calculation process of the demand calculation unit is divided into two steps:

[0069] (1) Convert the input load parameter representing the heat source state into a clear heat dissipation demand value expressed in watts.

[0070] Set a baseline cooling requirement. This value represents the cooling capacity required for the equipment at minimal load or with only ambient heat. Based on the percentage of the predicted server load, add additional cooling requirements above the baseline cooling requirement. For example, when the predicted server load is between 20% and 80%, the cooling requirement increases by two watts for every percentage point increase in load.

[0071] To ensure calculation stability and rationality, two upper and lower thresholds are set. When the predicted server load is below 20%, the cooling requirement is fixed at a baseline of 50 watts and does not decrease. When the predicted server load is above 80%, the cooling requirement is fixed at a maximum of 180 watts and does not increase.

[0072] (2) Convert the heat dissipation requirement value into a specific target fan speed in revolutions per minute:

[0073] When the calculated cooling demand is less than 70 watts, it indicates that the system is in a low load state and quietness is more important. Therefore, the target fan speed is set to a lower speed, such as 1,000 revolutions per minute.

[0074] When the calculated cooling demand is between 70 watts and 130 watts, the system enters normal operation. At this time, the target fan speed is set to a medium speed, such as 1,800 revolutions per minute, to achieve a balance between heat dissipation and noise.

[0075] When the calculated heat dissipation demand exceeds 130 watts, it indicates that the system is under high load and heat dissipation is the top priority. At this time, the target fan speed is set to a higher speed, such as 2,500 revolutions per minute, to ensure sufficient heat dissipation performance.

[0076] This fuzzy load forecast is converted into precise cooling requirements, and the fan speed is then set in intervals. This refined control strategy achieves an intelligent balance between efficient cooling and user requirements for low noise, improving system energy efficiency and user experience.

[0077] By breaking down the prediction model into data processing, feature learning, and core prediction components, the model can deeply learn and accurately model the nonlinear and time-delay characteristics of heat transfer. This improves the accuracy and reliability of future cooling demand forecasts and provides a solid technical foundation for smoother and more efficient feedforward energy-saving control.

[0078] Furthermore, the personalized psychoacoustic model includes:

[0079] an acoustic feature analysis unit for extracting acoustic features representing sound quality from the residual noise generated by the active noise reduction algorithm;

[0080] A context perception unit, which uses the virtual sensing technology to determine the user's current application context;

[0081] A user preference learning unit, which uses the online reinforcement learning method to use the interpreted user operation behavior as a feedback signal to establish and update a user model that represents the user's preference for different acoustic features in different application scenarios;

[0082] The sound quality target generating unit is used to comprehensively calculate the current optimal sound quality target based on the acoustic characteristics, the application scenario and the user model.

[0083] In this embodiment, the reinforcement learning method employs the Q-learning algorithm. The state is defined as a vector consisting of the current user context (e.g., "working" or "resting") and the sharpness and loudness values ​​of the residual noise. The action space is a set of preset sound quality objectives, such as {Goal A: prioritize reducing sharpness; Goal B: prioritize reducing loudness}. The reward function is defined as follows: A manual adjustment instruction from the user is received.

[0084] Query the thermal inertia prediction model: The system immediately queries the current thermal inertia prediction model to determine whether the 1800 RPM speed before the user adjustment meets the predicted heat dissipation requirements.

[0085] Intention Judgment and Reward Allocation: Case A: If the model determines that 1800 RPM meets or exceeds the current cooling requirements, but the user still chooses to lower the fan speed, the user is deemed dissatisfied with the noise and is given a reward of -5.

[0086] Case B: If the model determines that 1800 RPM does not meet the cooling requirement at the time, and the user chooses to increase the fan speed, then this action is considered to meet the cooling requirement. This action is not used as a feedback signal for reinforcement learning, and the reward is 0.

[0087] Case C: If the model determines that 1800 RPM meets the cooling requirements, but the user still chooses to increase the fan speed, this may indicate that the user is not dissatisfied with the current acoustic characteristics and is seeking a lower temperature. This behavior can also be excluded from feedback (reward is 0), or a smaller positive reward (such as +0.5) can be given to indicate that the current sound is acceptable.

[0088] Through this logic, the system can effectively filter out operations driven by thermal comfort, so that the reinforcement learning feedback signal more purely reflects the user's preference for sound quality.

[0089] The virtual sensing technology works by analyzing audio signals collected by the device's microphone. If a continuous human voice signal within the 300Hz-3000Hz range is detected, the user context is considered "conversational." If no discernible human voice is detected and the ambient noise level is below 40dB, the user context is considered "quiet."

[0090] The acoustic characteristics include loudness calculated according to the ISO 532-1 standard and sharpness calculated according to the DIN 45692 standard.

[0091] According to psychoacoustic research, loudness values ​​are divided into three levels: 'low' (less than 0.5), 'medium' (0.5 to 1.5), and 'high' (greater than 1.5);

[0092] The sharpness value is divided into two levels: 'soft' (less than 1.2) and 'sharp' (greater than or equal to 1.2);

[0093] The user model is a Q table, where the rows represent 'states' and the columns represent 'actions'. The value Q(s, a) in the table represents the expected reward of performing action a in state s.

[0094] This invention breaks through the limitations of traditional noise reduction technologies, which focus solely on reducing volume, by breaking down the acoustic model into multiple components: feature analysis, contextual awareness, and online preference learning. This model learns and adapts to the hearing preferences of specific users, dynamically adjusting noise reduction targets based on their context, and proactively shaping residual noise into a more palatable form. This transition from noise reduction to personalized tuning provides users with a highly intelligent acoustic comfort experience.

[0095] Furthermore, the active noise reduction algorithm calculates and generates an anti-phase noise sound field that cancels out the original fan noise based on the acoustic sensing of the original fan noise according to the current optimal sound quality target;

[0096] The residual noise is a synthetic sound field formed by acoustic interference between the original fan noise and the anti-phase noise sound field within a preset target noise reduction area. The sensing result of the residual noise is used as feedback to dynamically adjust the active noise reduction algorithm.

[0097] The target noise reduction area is defined as a spherical space with a radius of 15 cm centered on the user's typical head position. Residual noise in the area is sensed by arranging three orthogonal error microphones within the space.

[0098] Preferably, the active noise reduction algorithm adopts the FxLMS algorithm, and the parameters that need to be adjusted are the step size factor μ and the 32 coefficients of the noise reduction filter W(z).

[0099] Before the offline training process of the mapping model begins, the secondary channel must be identified. In a semi-anechoic chamber, a broadband white noise signal is played through a noise-canceling loudspeaker while being collected by an error microphone. Using the LMS adaptive filtering algorithm, the impulse response from the loudspeaker input to the microphone output is identified and used as the digital filter model S(z) for the secondary channel. This is used to filter the reference signal. The error microphone is fixed at the center of the target noise reduction area.

[0100] The offline training process of the mapping model is as follows: the fan is placed in a semi-anechoic chamber. The fan speed is set from 500RPM to 3000RPM, with a step of 50RPM, and traverses 50 speed points. At each speed point, a multi-objective automatic optimization algorithm is run. The algorithm searches for and determines a Pareto optimal parameter set at the current speed with the goal of simultaneously minimizing a set of psychoacoustic parameters representing sound quality. The final mapping model stores the correspondence between different speeds and a set of optional optimal parameters. At runtime, the parameter mapping unit selects which set of parameters to load based on the current sound quality target.

[0101] The parameter mapping unit is a lookup table, in which the index is the fan speed and the value is the corresponding optimal parameter set; for speeds not specified in the table, the parameters of adjacent speed points are linearly interpolated to obtain the speeds.

[0102] During operation, the optimal parameters can be loaded by quickly looking up the table and interpolating according to the real-time speed without the need for complex real-time calculations, achieving active control of fan noise in all working conditions with precise, efficient and better sound quality.

[0103] This invention explicitly targets the active noise reduction algorithm as optimizing sound quality and establishes a closed-loop feedback loop based on residual noise sensing. This design ensures the accuracy, purposefulness, and adaptability of the noise reduction process, enabling the system to continuously correct deviations and steadily shape the sound field to an optimal comfort state that meets user preferences, achieving a higher level of intelligent acoustic control.

[0104] Furthermore, the predictive control model includes:

[0105] A speed input unit, configured to receive the target fan speed;

[0106] a parameter mapping unit, which obtains a mapping model through offline training, wherein the mapping model represents the correspondence between different fan speeds and the optimal parameters of the active noise reduction algorithm, and the parameter mapping unit is used to determine the optimal parameters through the mapping model according to the input target fan speed;

[0107] A parameter output unit is used to output the determined optimal parameters.

[0108] By employing an offline-trained parameter mapping model, the present invention pre-determines the relationship between different speeds and optimal noise reduction parameters. This eliminates the need for time-consuming online learning when wind speeds change, allowing the system to instantly determine and set the optimal parameters for the new operating conditions. This eliminates performance degradation and noise leakage during speed changes, ensuring seamless, efficient, and stable control.

[0109] Furthermore, the user operation behavior includes: an instruction to manually adjust the fan speed; an instruction to turn off the fan; and an instruction to switch the fan operation mode.

[0110] By using users' routine operations on wind speed, switches, etc. as learning signals, the system's personalized learning process is made transparent and imperceptible to users, improving the product's usability and intelligent experience.

[0111] The present invention predicts the heat dissipation demand through the thermal inertia prediction model, smoothly adjusts the fan speed in advance, avoids energy waste caused by response lag, and significantly improves energy efficiency. Secondly, it uses online reinforcement learning to establish a personalized psychoacoustic model, which can learn the user's preference for sound and shape the residual noise to be more acceptable, achieving a leap from simple noise reduction to optimizing sound quality, and providing a personalized high-quality acoustic experience. Finally, through the predictive control model, the noise reduction parameters are pre-determined and synchronously adjusted before the fan changes speed, solving the problem of deterioration of the effect during the speed change process and ensuring stability and reliability in dynamic operation.

[0112] Example 2

[0113] The application scenario of this embodiment is set as the heating, ventilation and air conditioning (HVAC) system of a smart car (especially an electric car). Figure 3 As shown, its core goal is to minimize fan energy consumption to increase driving range while ensuring the thermal comfort of drivers and passengers, and to provide a quiet, comfortable and personalized cabin acoustic environment.

[0114] In this embodiment, the demand prediction and speed control module is adaptively adjusted according to the characteristics of the vehicle environment.

[0115] In addition to the temperature and humidity sensor data inside the vehicle, the environmental parameters also integrate the data from the daylight sensor installed on the dashboard to quantify the heat load caused by solar radiation.

[0116] The load parameters are replaced with parameters that better reflect the dynamics of the heat source in the vehicle; specifically, they include: the number of passengers read through the vehicle CAN bus, the air conditioning temperature set by the user, and the battery pack temperature fed back by the power battery management system.

[0117] This embodiment uses a hybrid prediction method that combines physical models with machine learning as a thermal inertia prediction model. A simplified in-vehicle thermodynamic model is used to perform preliminary thermal load predictions. A support vector regression (SVR) model is used to learn and compensate for the physical model's prediction errors based on real-time historical data.

[0118] The demand calculation unit converts the predicted total heat load, modified by the SVR model, into a target fan speed. The conversion rules are calibrated based on the vehicle's HVAC system air duct characteristics, establishing a "heat dissipation power - fan speed" mapping curve. For example, when the predicted heat load is less than 150W, the target speed is set at 1200 RPM; when the predicted heat load is between 150W and 350W, linear interpolation is used to adjust the speed between 1200 and 2400 RPM; and when the predicted heat load is above 350W, the target speed is set at the maximum of 3000 RPM.

[0119] In this embodiment, the adaptive noise quality control module makes full use of the in-vehicle infotainment system and vehicle bus data:

[0120] The situational awareness unit achieves more accurate situational judgment through multi-source information fusion.

[0121] Real-time vehicle speed is obtained via the CAN bus. When the vehicle speed exceeds 80 km / h, the increased masking effect of road and wind noise can be considered a "high-speed driving" situation, and the control of fan noise can be appropriately relaxed.

[0122] The system communicates with the vehicle's infotainment system to obtain the Bluetooth phone status. When the system detects an active hands-free call, it determines it as a "call" context.

[0123] By monitoring the volume of the media player, when the volume is higher than a preset threshold, it is determined to be in the "audio and video entertainment" situation.

[0124] Acoustic Feature Analysis Unit: In addition to loudness, this embodiment also introduces "fluctuation strength" as another key psychoacoustic evaluation indicator. Fluctuation strength is used to measure the slowly changing characteristics of noise, which is likely to cause irritation at specific frequencies.

[0125] User preference learning unit: Deep Q network (DQN) is used for reinforcement learning. The state (State) consists of the current situation (such as 'high-speed driving', 'on the phone') and the loudness and fluctuation intensity level of the residual noise. The action (Action) is to select a preset sound quality target (such as 'prioritize reducing loudness', 'prioritize reducing fluctuation intensity'). The reward (Reward) mechanism is similar to that of Example 1. When the user manually intervenes (increases or decreases) the wind speed automatically set by the system, the system will determine whether the behavior is caused by acoustic discomfort and give corresponding positive or negative rewards. DQN uses neural networks to approximate the Q function, can handle more complex and high-dimensional state spaces, and has more accurate learning strategies.

[0126] In this embodiment, the transient cooperative control module for achieving seamless adjustment of noise reduction parameters adopts different parameter mapping implementation methods:

[0127] Parameter mapping unit: A radial basis function (RBF) network is used to characterize the mapping relationship between the fan speed and the optimal parameters of the active noise reduction algorithm (step size factor μ and filter W(z) coefficient).

[0128] In the offline training stage, by running the optimization algorithm at different speeds, a series of "speed-optimal parameter" data pairs are obtained, and these data are used to train the RBF network.

[0129] During operation, when the target speed changes, the trained RBF network quickly and smoothly interpolates and calculates the optimal parameter set for the new speed. Compared to lookup tables, RBF networks offer better nonlinear fitting capabilities and generalization. Compared to MLPs, they are faster to train and have a simpler structure, making them particularly suitable for automotive applications that require fast response.

[0130] By applying the present invention to the HVAC system of an intelligent vehicle and specifically adopting a hybrid thermodynamic prediction model, multi-dimensional context perception integrating vehicle bus data, preference learning based on DQN, and parameter mapping based on an RBF network, the present invention successfully demonstrated its ability to achieve coordinated control of energy saving, comfort, and quietness in an in-vehicle environment, and has high practical value and market prospects.

[0131] Example 3

[0132] This embodiment is intended to illustrate another specific implementation of the present invention. Its basic principles and processes are the same as those of Example 1, but there are differences in the technical implementation, parameter selection and application scenarios of some modules to demonstrate the wide applicability of the technology of the present invention.

[0133] The application scenario of this embodiment is set as a high-end smart home or modern office environment. Its core goal is to provide users with an indoor environment experience that combines high energy efficiency and personalized acoustic comfort.

[0134] In this embodiment, the thermal inertia prediction model is intended to predict and manage the thermal environment within a living or office space.

[0135] Environmental parameters: including indoor temperature and humidity;

[0136] Load parameters: Replaced with parameters that better reflect the changes in heat sources in home or office scenes. Specifically including:

[0137] The number of people and their activity levels are estimated using cameras installed indoors combined with lightweight gesture recognition algorithms, or indirectly inferred using CO2 concentration sensors.

[0138] Status of major heat-generating devices: Obtain real-time power consumption data of major electrical devices such as computers, TVs, and lighting systems through smart sockets;

[0139] External environmental data: By accessing the Internet weather service API, we can obtain information such as outdoor temperature and solar radiation intensity to predict external heat input through windows and walls.

[0140] This embodiment uses a gated recurrent unit (GRU) network as the core model for time series prediction. As a variant of LSTM, the GRU offers similar prediction performance, but with a simpler internal structure and fewer parameters. This allows for higher computational efficiency while maintaining prediction accuracy, making it more suitable for deployment in resource-constrained smart home gateway devices. This GRU model consists of a 100-unit GRU layer followed by a fully connected layer, used to predict the comprehensive heat load for the next 30 minutes.

[0141] The conversion rules of the demand calculation unit have been adjusted according to the home scene. The predicted number of people, equipment power consumption and external heat gain are weighted and summed by their respective heat production coefficients (for example, about 100W for a sitting adult and about 150W for a computer running) to obtain a total predicted heat load in watts (W). The fan speed is set according to the total predicted heat load. For example: when the heat load is less than 200W, it is set to 800RPM (quiet mode); when it is between 200W and 450W, it is set to 1500RPM (balanced mode); when it is higher than 450W, it is set to 2200RPM (strong cooling mode).

[0142] The personalized psychoacoustic model uses different perception technologies, acoustic indicators and learning algorithms. The situational awareness unit uses multi-sensor fusion technology instead of relying solely on microphone audio analysis. The Bluetooth signal strength is used to sense whether the user's smartphone or wearable device is in the near field and to determine whether the user is in the room. By analyzing the activity status of smart devices (such as computers and TVs), it is inferred whether the user is in a "focused work", "audio and video entertainment" or "resting" situation. For example, when it is detected that the user's mobile phone is nearby and the computer is active, the situation is determined to be "focused work". The acoustic feature analysis unit introduces different psychoacoustic parameters. In addition to loudness, this embodiment also introduces roughness as a key indicator. Roughness is used to measure the unpleasantness caused by changes in the temporal structure of sound, which is crucial for evaluating the quality of fan noise.

[0143] The user preference learning unit adopts the Actor-Critic reinforcement learning framework.

[0144] State: Consists of the current user context (e.g., 'working', 'resting'), the loudness and roughness level of the residual noise.

[0145] Actor Network: A small neural network that takes in the current state and outputs a probability distribution over the optimal sound quality targets (e.g., 70% probability of reducing loudness, 30% probability of reducing harshness).

[0146] Critic network: Another small neural network that estimates the long-term rewards of the actions chosen by the actor in the current state and is used to guide the update of the actor network.

[0147] Rewards: The reward mechanism is similar to that of Example 1. It measures user satisfaction with the current sound quality by interpreting the user's manual intervention to bypass the system's automatically set wind speed. The actor-critic model enables smoother and more stable policy learning and is particularly well-suited for handling continuously changing acoustic environments.

[0148] In this embodiment, the predictive control model used to achieve seamless adjustment of noise reduction parameters employs a different implementation approach for its core parameter mapping unit. A small multilayer perceptron neural network is used to directly learn and characterize the complex nonlinear mapping relationship between fan speed and the optimal parameters of the active noise reduction algorithm.

[0149] During the offline calibration phase, the optimal parameter sets (step size μ and filter coefficient W(z)) of the FxLMS algorithm obtained at different speeds are collected. The MLP network is trained using the speed as input and the optimal parameter set as output.

[0150] When the fan speed needs to be adjusted, the speed input unit sends the new target speed to the trained MLP network. The network will instantly calculate the corresponding optimal noise reduction parameters, which are then output by the parameter output unit.

[0151] Compared with lookup tables, the MLP-based parameter mapping model has better generalization capabilities and can accurately generate optimal parameters for speed points that do not appear in offline calibration, making the adjustment of noise reduction parameters smoother and more accurate, thereby further improving the user experience during the gear shifting process.

[0152] In summary, Example 3, through its adaptation of specific application scenarios, sensing technologies, and algorithmic models, further demonstrates the flexibility and advancement of the present invention's technical solution. This demonstrates that the present invention is not only applicable to industrial environments like data centers, but also can achieve coordinated control of fan energy savings and a personalized, comfortable acoustic experience in civilian scenarios like smart homes, demonstrating its broad application prospects.

[0153] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A fan adaptive energy-saving control method for active noise reduction, characterized in that: The following steps are involved: Establishing a thermal inertia prediction model, inputting environmental parameters representing the indoor environmental state and load parameters representing the heat source state, predicting the heat dissipation demand of the fan within a future preset time period, and calculating a target fan speed that meets the heat dissipation demand; An online reinforcement learning method is used to establish and dynamically optimize a personalized psychoacoustic model; wherein the feedback signal of the online reinforcement learning method is derived from the interpretation of user operation behavior; virtual sensing technology is used to perceive the user's context, and based on the context and the personalized psychoacoustic model, the optimal sound quality target is determined and an active noise reduction algorithm is executed; A predictive control model is established. When the target fan speed is different from the current speed, the target fan speed is input into the predictive control model to determine the optimal parameters corresponding to the target fan speed; and the parameters of the active noise reduction algorithm are synchronously adjusted to the optimal parameters.

2. The method for adaptive energy-saving control of a fan for active noise reduction according to claim 1, characterized in that: The environmental parameters include temperature, humidity, static pressure and CO2 concentration; the load parameters include the real-time calculated load of the server as a thermal load indicator.

3. The method for adaptive energy-saving control of a fan for active noise reduction according to claim 1, characterized in that: The thermal inertia prediction model includes: A data acquisition and preprocessing unit, configured to receive and process real-time and historical data of the environmental parameters and load parameters; A feature extraction and correlation analysis unit, configured to learn the nonlinear mapping relationship and time delay characteristics between the environmental parameters, load parameters, and heat dissipation requirements during an offline training phase; A time series prediction core unit, configured to predict the load state within the preset future time period according to the output of the feature extraction and correlation analysis unit during operation; The demand calculation unit is used to convert the predicted load state into a quantitative value of the heat dissipation demand and calculate the target fan speed.

4. The method for adaptive energy-saving control of a fan for active noise reduction according to claim 1, characterized in that: The personalized psychoacoustic model include: an acoustic feature analysis unit for extracting acoustic features representing sound quality from the residual noise generated by the active noise reduction algorithm; A context perception unit, which uses the virtual sensing technology to determine the user's current application context; A user preference learning unit, which uses the online reinforcement learning method to use the interpreted user operation behavior as a feedback signal to establish and update a user model that represents the user's preference for different acoustic features in different application scenarios; The sound quality target generating unit is used to comprehensively calculate the current optimal sound quality target based on the acoustic characteristics, the application scenario and the user model.

5. The method for adaptive energy-saving control of a fan for active noise reduction according to claim 4, characterized in that: The active noise reduction algorithm calculates and generates an anti-phase noise sound field that cancels out the original fan noise based on the acoustic sensing of the original fan noise according to the current optimal sound quality target; The residual noise is a synthetic sound field formed by acoustic interference between the original fan noise and the anti-phase noise sound field within a preset target noise reduction area. The sensing result of the residual noise is used as feedback to dynamically adjust the active noise reduction algorithm.

6. The method for adaptive energy-saving control of a fan for active noise reduction according to claim 1, characterized in that: The predictive control model includes: A speed input unit, configured to receive the target fan speed; a parameter mapping unit, which obtains a mapping model through offline training, wherein the mapping model represents the correspondence between different fan speeds and the optimal parameters of the active noise reduction algorithm, and the parameter mapping unit is used to determine the optimal parameters through the mapping model according to the input target fan speed; A parameter output unit is used to output the determined optimal parameters.

7. The fan adaptive energy-saving control method for active noise reduction according to claim 4, characterized in that: The user operation behavior includes: an instruction to manually adjust the fan speed; an instruction to turn off the fan; and an instruction to switch the fan operation mode.

8. A fan adaptive energy-saving control system for active noise reduction, characterized in that: The following steps are involved: Demand prediction and speed control module: establishes a thermal inertia prediction model, inputs environmental parameters representing the indoor environmental state and load parameters representing the heat source state, predicts the heat dissipation demand of the fan within a preset time period in the future, and calculates the target fan speed that meets the heat dissipation demand; Adaptive Noise Quality Control Module: This module uses online reinforcement learning to establish and dynamically optimize a personalized psychoacoustic model, where the feedback signal of the online reinforcement learning method is derived from the interpretation of user operation behavior. It uses virtual sensing technology to perceive the user's context, and based on this context and the personalized psychoacoustic model, it determines the optimal sound quality target and runs the active noise reduction algorithm. Transient collaborative control module: establish a predictive control model, when the target fan speed is different from the current speed, input the target fan speed into the predictive control model, determine the optimal parameters corresponding to the target fan speed; synchronously implement the active noise reduction algorithm.

Images