Fan adaptive swing head method based on AI space perception

Abstract

The application discloses a fan self-adaptive swing head method based on AI space perception. A low-resolution infrared image frame is input into a space thermal field time sequence perception network, and a thermal distribution feature map is extracted through inter-frame difference time sequence convolution; a decoding layer generates a discrete space occupancy probability sequence through multi-scale pooling; a boundary extraction module extracts left and right boundaries exceeding a threshold and maps the left and right boundaries to a fan mechanical coordinate system, and outputs a target swing head boundary and a probability sequence; a trajectory modulation output head converts the probability into corresponding angular velocity according to a formula, and generates an asymmetric angular velocity control sequence; and finally, the angular velocity control sequence is transmitted to a motor driving unit to realize self-adaptive swing head. Under the premise of protecting privacy, the application realizes dynamic locking of air supply boundaries and stepless speed change intelligent air supply, improves comfort and reduces energy consumption.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent fan control, and in particular to a fan adaptive oscillation method based on AI spatial perception. Background Technology

[0002] With the rapid development of smart home technology, traditional household electric fans are undergoing a profound transformation from mechanization to intelligence. Early electric fans typically used fixed mechanical oscillation mechanisms, sweeping air globally within a preset maximum angle range at a constant angular velocity. This traditional air delivery mode could not sense the distribution of people in the environment, resulting in a large amount of ineffective airflow in unoccupied areas. This not only wasted energy but also failed to meet the personalized thermal comfort needs of different users in complex spaces. In recent years, in order to improve the intelligence level and air delivery efficiency of fans, various adaptive fan oscillation control methods based on sensors and control algorithms have emerged in the industry. However, many significant defects and technical bottlenecks still exist in practical applications.

[0003] Firstly, some existing smart fans employ control strategies based on simple zoning and binary speed regulation. For example, some solutions calculate the angle difference by recording the offset angle of a person's position, or use simple presence sensing technology to rigidly divide the space in front of the fan into "occupied space" and "unoccupied space." In terms of control logic, these methods typically use a dual-speed scanning algorithm, i.e., slow oscillation airflow in areas determined to be occupied, and fast skipping logic in unoccupied areas. This either-or binary control mode has obvious drawbacks: on the one hand, the distribution of people and the state of heat radiation in the real space are continuous and dynamic, and simple threshold zoning cannot accurately reflect the complex situation of people's activities; on the other hand, the frequent and abrupt switching of the motor between "normal speed" and "fast skipping" not only leads to extremely poor fan smoothness, producing obvious mechanical jerks and noise, but also accelerates the wear and tear of the motor and transmission mechanism, seriously affecting the user experience and equipment lifespan.

[0004] Secondly, to achieve higher accuracy in personnel positioning, some existing technologies have introduced high-precision sensing hardware, such as high-definition camera image recognition modules or millimeter-wave radar. While image recognition-based solutions can accurately pinpoint personnel locations and even associate them with identity information for differentiated airflow, the introduction of cameras in private home settings such as bedrooms and living rooms has raised strong concerns about privacy leaks. Furthermore, image processing requires significant computational resources, substantially increasing hardware costs. While millimeter-wave radar solutions possess a certain degree of penetration capability, their radar signal processing algorithms are extremely complex and prone to misjudgment when distinguishing between subtle human movements and environmental disturbances (such as pet activity or swaying curtains), making it difficult to achieve an ideal balance between cost and reliability.

[0005] In addition, some solutions focus on evaluating the comfort of the macroscopic environment or simply tracking a single heat source. For example, some AI-based decision-making methods construct human comfort evaluation models by integrating environmental data such as temperature and humidity, or use dual-objective optimization algorithms to find a balance between energy consumption and comfort. However, these methods mostly remain at the level of adjusting macroscopic parameters, lacking real-time high-frequency perception and trajectory-level control of the microscopic spatial thermal field distribution. Traditional automatic tracking fans based on thermal imaging principles often only use simple temperature difference guidance control, which can only rigidly control the fan head to point at the hottest heat source. They cannot generate reasonable dynamic airflow boundaries in a wide field of view containing multiple people, nor can they finely adjust the speed according to the actual occupancy probability of each position.

[0006] In summary, existing intelligent fan control methods generally face technical challenges such as high cost and privacy violations of sensing technologies, overly coarse spatial state assessment, and rigid and abrupt motor speed control. The industry urgently needs an adaptive oscillation control method that can achieve high-precision spatial thermal field perception using low-cost sensors while protecting user privacy, and can deeply and smoothly integrate spatial distribution probability with motor motion trajectory to achieve truly efficient, smooth, and intelligent airflow. Summary of the Invention

[0007] In view of this, the present invention provides a fan adaptive oscillation method based on AI spatial perception. The purpose is to accurately capture dynamic heat distribution and generate fine spatial occupancy probability by combining a low-resolution infrared sensor with a deeply customized spatial thermal field temporal perception network. Then, the perceived probability distribution is deeply and smoothly fused with the mechanical motion trajectory of the fan. Under the premise of protecting user privacy, it can realize intelligent air delivery with dynamic locking of air delivery boundary and continuous stepless speed change, thereby improving user thermal comfort and reducing ineffective energy consumption.

[0008] To achieve the above objectives, this invention provides a fan adaptive oscillation method based on AI spatial awareness, comprising the following steps: B1: Input consecutive low-resolution infrared sensor image frames into the initial feature extraction layer of the spatial thermal field temporal sensing network, perform inter-frame differential temporal convolution operation, and extract spatial thermal distribution feature maps. B2: Input the spatial heat distribution feature map into the decoding layer of the spatial heat field temporal sensing network, perform multi-scale spatial hierarchical pooling operation, and generate a discrete spatial occupancy probability sequence of the corresponding fan front field of view; B3: Input the discrete space occupancy probability sequence into the occupancy boundary extraction module of the spatial thermal field time sequence sensing network, extract the leftmost and rightmost spatial positions that exceed the occupancy threshold and map them to the fan mechanical coordinate system, and output the target oscillation boundary angle and the space occupancy probability sequence within the boundary angle range. The target oscillation boundary angle includes the target oscillation left boundary angle and the target oscillation right boundary angle. B4: Input the target head-shaking boundary angle and the spatial occupancy probability sequence within the boundary angle range into the trajectory modulation output head of the spatial thermal field temporal sensing network. Within the boundary angle range, convert the spatial occupancy probability of each spatial position into the corresponding angular velocity according to the occupancy inverse mapping angular velocity modulation formula to generate a continuous asymmetric angular velocity control sequence. B5: Transmits a continuous asymmetric angular velocity control sequence to the fan's motor drive unit, continuously adjusting the motor's speed and direction, driving the fan head to adaptively oscillate within the target oscillation boundary angle.

[0009] As a further improvement of the present invention: Optionally, step B1 further includes: A data acquisition channel for a low-resolution infrared sensor is established to continuously acquire infrared image frames within the field of view in front of the fan at a preset sampling frequency, resulting in a continuous sequence of infrared image frames. A spatial thermal field temporal sensing network is constructed, which consists of an initial feature extraction layer, a decoding layer, an occupancy boundary extraction module, and a trajectory modulation output head connected in series. The initial feature extraction layer consists of convolution operation units connected in series. The continuous sequence of infrared image frames is input to the initial feature extraction layer, and inter-frame difference temporal convolution operation is performed: first, the infrared image frames at adjacent time points are subtracted pixel by pixel to obtain an inter-frame difference matrix; then, a convolution kernel operation is applied to the inter-frame difference matrix to extract local feature representations reflecting the change of spatial heat distribution over time; the local feature representations are combined to form a spatial heat distribution feature map.

[0010] Optionally, step B2 further includes: The spatial thermal distribution feature map is input to the decoding layer of the spatial thermal field temporal sensing network. This decoding layer is composed of pooling convolutional units with progressively decreasing resolution, connected in series. During multi-scale spatial hierarchical pooling, pooling convolutional kernels with different receptive field sizes are used to extract the spatial distribution response at each scale along the spatial dimension of the feature map. These responses are then concatenated along the channel dimension and output via a fully connected layer, yielding the occupancy probability values ​​for each spatial location within the field of view in front of the fan, thus obtaining a discrete spatial occupancy probability sequence. This discrete spatial occupancy probability sequence includes... The component, the first Each component Indicates the first [number]th [unit] in the field of view in front of the fan. The probability that a spatial location is occupied by people. , This represents the total number of spatial divisions within the field of view.

[0011] Optionally, step B3 further includes: The discrete space occupancy probability sequence is input into the occupancy boundary extraction module, which extracts each component of the discrete space occupancy probability sequence. With preset occupancy threshold Perform a comparison one by one; when there exists a satisfying When the spatial location meets the conditions, if the infrared sensor detects a person, the fan will automatically start running, and the leftmost spatial location index that meets the conditions will be determined. With the rightmost spatial location index Specifically: ; in, Indicates the first The probability of spatial occupancy at each spatial location. This indicates the preset occupancy threshold. This represents the total number of spatial divisions within the field of view. This represents the leftmost spatial location index that satisfies the occupancy threshold condition. This represents the rightmost spatial location index that satisfies the occupancy threshold condition; when no such condition exists... When the spatial location of the condition is determined, the boundary extraction module does not output the boundary index, determines that the infrared light does not detect personnel, triggers the default wind sweeping strategy, controls the fan to oscillate at a constant angular velocity within the global maximum boundary angle range, and if no personnel are detected for ten minutes, the fan will automatically stop running, and the subsequent dynamic modulation steps B4 and B5 will not be executed. Index the leftmost spatial position With the rightmost spatial location index By mapping the linear coordinates to the fan's mechanical coordinate system, the left boundary angle of the target oscillation is obtained. Angle with the right boundary of the target head The linear coordinate transformation formula is as follows: ; in, Indicates spatial location index The corresponding angle value in the fan's mechanical coordinate system. This represents the left-hand limit angle of the fan's mechanical coordinate system. This represents the right-hand limit angle of the fan's mechanical coordinate system; when Pick Time of income The value is the left boundary angle of the target head shake. ,when Pick Time of income The value is the right boundary angle of the target head shake. Simultaneously, extract the index range from the discrete space occupancy probability sequence. The subsequence within, the subsequence relative to the left boundary angle of the target head shake Target head shaking right boundary angle Together they constitute the input for step B4.

[0012] Optionally, step B4 further includes: The target's left boundary angle Target head shaking right boundary angle and index range The spatial occupancy probability subsequence within the range is input to the trajectory modulation output head, which is configured for the index range. Index of each spatial location within The angular velocity of the fan head at the angle corresponding to the index is calculated according to the inverse mapping angular velocity modulation formula: ; in, Indicates the spatial position index of the fan head. Angular velocity at the corresponding angle This indicates the preset lower limit of angular velocity. This indicates the preset upper limit of angular velocity. Indicates the first The probability of spatial occupancy at a given spatial location; when hour, Pick ;when hour, Pick ;when When the value is between 0 and 1 exist to The distance is determined linearly according to the formula; based on the angle of the fan head from the left boundary of the target oscillation. To the right boundary angle of the target head The scan order will determine the index range. The corresponding spatial location index i Arranged sequentially, a continuous asymmetric angular velocity control sequence is obtained. The angular velocity value at each position in the asymmetric angular velocity control sequence is determined independently position by position based on the spatial occupancy probability of that position.

[0013] Optionally, step B5 further includes: A continuous asymmetric angular velocity control sequence is transmitted to the fan's motor drive unit. The motor drive unit outputs a drive current in real time based on the angular velocity command values ​​at each moment in the asymmetric angular velocity control sequence, adjusting the motor's speed and direction. When the fan head moves from the target oscillation left boundary angle... Shaking the right boundary angle of the target During directional movement, the motor drive unit moves from index [index] according to the asymmetric angular velocity control sequence. to The angular velocity command value in the direction drives the motor to rotate in the forward direction; when the fan head reaches the right boundary angle of the target oscillation... Then, the motor drive unit controls the motor to commutate, starting from the index in the asymmetric angular velocity control sequence. to The angular velocity command value in the direction drives the motor to rotate in the opposite direction; this process repeats, driving the fan head to rotate at the left boundary angle of the target oscillation. To the right boundary angle of the target head It continuously performs adaptive head shaking.

[0014] Compared with the prior art, the present invention has at least the following beneficial effects: First, this invention does not directly apply conventional image recognition algorithms in the low-level feature extraction stage. Instead, it specifically designs an initial feature extraction mechanism that combines inter-frame differencing with temporal convolution for low-resolution infrared data. This improvement not only completely avoids the risk of privacy leakage in home scenarios from the physical source by utilizing the low-resolution characteristics, but more importantly, by introducing pixel-level differencing operations between adjacent time steps before convolution operations, the algorithm can effectively filter static background thermal noise in the environment, greatly enhancing the network's sensitivity to the changing state of human body dynamic thermal radiation over time. This provides highly targeted, high-quality dynamic features for subsequent spatial perception.

[0015] Secondly, this invention innovatively introduces multi-scale spatial hierarchical pooling operations in the decoding stage, abandoning the binary, coarse-grained partitioning logic of traditional methods. By using pooling convolution kernels with different receptive field sizes to extract spatial distribution responses at each scale and performing channel stitching, this algorithm structure can adapt to the thermal radiation diffusion characteristics of people at different distances and with different body types, accurately outputting continuous and detailed occupancy probability sequences for each spatial position within the field of view, achieving high-resolution digital representation of complex spatial personnel distribution patterns.

[0016] More importantly, this invention constructs a deep linkage mechanism from spatial probability perception to mechanical trajectory modulation. Each algorithm module is interconnected, forming a highly collaborative control loop. The discrete spatial occupancy probability sequence output by the pre-processor network is not only used to accurately cut out the optimal oscillation boundary encompassing all personnel, avoiding the waste of global sweeping; this sequence also directly serves as underlying data support, directly converting the occupancy probability of each microscopic spatial location into the corresponding motion angular velocity through a specially designed occupancy inverse mapping angular velocity modulation logic. This design allows for seamless integration of perception results and mechanical execution; the higher the probability, the slower the fan sweeps, and the lower the probability, the faster it sweeps. From the precise extraction of dynamic thermal features and the refined generation of multi-scale probabilities to boundary locking and asymmetric continuous speed control based on probability sequences, the entire solution deeply binds the structured improvement of artificial intelligence algorithms with the control of the fan's physical and mechanical characteristics. This integrated design completely breaks the rigidity and jerking of traditional fan mechanical sweeping, achieving truly smooth, delicate, and highly adaptive air delivery that conforms to the human body distribution, significantly improving air delivery efficiency and the smoothness of equipment operation. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an embodiment of the AI-based spatial awareness fan adaptive oscillation method of the present invention. Figure 2 This is a schematic diagram of the spatial occupancy probability distribution and the target boundary extraction effect. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0019] Example 1: An AI-based spatial awareness-based adaptive fan oscillation method, such as... Figure 1 As shown, it includes the following steps: B1: Input consecutive low-resolution infrared sensor image frames into the initial feature extraction layer of the spatial thermal field temporal sensing network, perform inter-frame differential temporal convolution operation to extract spatial thermal distribution feature maps, including: A data acquisition channel for a low-resolution infrared sensor is established to continuously acquire infrared image frames within the field of view in front of the fan at a preset sampling frequency, resulting in a continuous sequence of infrared image frames. A spatial thermal field temporal sensing network is constructed, which consists of an initial feature extraction layer, a decoding layer, an occupancy boundary extraction module, and a trajectory modulation output head connected in series. The initial feature extraction layer consists of convolutional operation units connected in series. In this embodiment, the initial feature extraction layer consists of two convolutional operation units connected in series. The first convolutional operation unit uses a two-dimensional convolutional layer with a kernel size of 3×3, 16 output channels, a stride of 1, and edge padding of 1. The second convolutional operation unit uses a two-dimensional convolutional layer with a kernel size of 3×3, 32 output channels, a stride of 1, and edge padding of 1. Each convolutional operation unit is followed by a batch normalization layer and a ReLU activation function. The continuous infrared image frame sequence is input into the initial feature extraction layer, and an inter-frame difference temporal convolution operation is performed: first, the infrared image frames at adjacent time points are subtracted pixel by pixel to obtain the inter-frame difference matrix; the inter-frame difference matrix The calculation formula is: ; in, Indicates time The inter-frame difference matrix, Indicates time Infrared image frame matrix, Indicates time Infrared image frame matrix, The values ​​of each element in the data reflect the change in thermal radiation intensity between two adjacent frames at the corresponding spatial location. Then, a convolution kernel operation is applied to the inter-frame difference matrix to extract local feature representations that reflect the change of spatial heat distribution over time; the local feature representations are combined to form a spatial heat distribution feature map.

[0020] Optionally, to suppress the interference of sensor thermal noise on the inter-frame difference matrix under low signal-to-noise ratio operating conditions and improve the sensitivity of the difference results to changes in the thermal radiation of real people, a time-domain moving average smoothing preprocessing is applied to the original infrared image frame sequence before performing the inter-frame difference operation: taking time... and its predecessor The original infrared image frames at each time point are averaged pixel by pixel to obtain the smoothed infrared image frame matrix. The formula for calculating the time-domain moving average smoothing is as follows: ; in, Indicates time The smoothed infrared image frame matrix obtained after time-domain moving average smoothing. This represents the number of frames in the time-domain moving average smoothing window, in this embodiment. Set to 3; Indicates the frame offset number within the smoothing window. ; Indicates time The acquired raw infrared image frame matrix; after introducing the above-mentioned time-domain moving mean smoothing preprocessing, the inter-frame difference matrix. Change to and The calculation is performed by subtracting pixels one by one to effectively suppress the interference of sudden noise pulses in a single frame on the difference results, while retaining the thermal distribution change information caused by the movement of real people.

[0021] B2: Input the spatial thermal distribution feature map into the decoding layer of the spatial thermal field temporal sensing network, perform multi-scale spatial hierarchical pooling operations, and generate a discrete spatial occupancy probability sequence corresponding to the fan's front field of view, including: The spatial thermal distribution feature map is input to the decoding layer of the spatial thermal field temporal sensing network. The decoding layer is composed of pooling convolutional units with progressively decreasing resolution connected in series. In this embodiment, the decoding layer consists of three pooling convolutional units connected in series: the first pooling convolutional unit includes a max pooling layer with a kernel size of 2×2 and a stride of 2, and a convolutional layer with a kernel size of 3×3, 64 output channels, and a stride of 1; the second pooling convolutional unit includes a max pooling layer with a kernel size of 2×2 and a stride of 2, and a convolutional layer with a kernel size of 1×3, 128 output channels, and a stride of 1; the third pooling convolutional unit includes an average pooling layer with a kernel size of 6×6 and a stride of 6, and a convolutional layer with a kernel size of 1×1, 64 output channels, and a stride of 1; each pooling convolutional unit is followed by a batch normalization layer and a ReLU activation function in sequence. When performing multi-scale spatial hierarchical pooling operations, pooling convolution kernels with different receptive field sizes are used on the spatial dimension of the feature map to extract the spatial distribution response at each scale. The extraction process of the spatial distribution response at each scale is as follows: First scale spatial distribution response Spatial heat distribution feature map obtained from the first pooling convolution unit Processed, With 64 channels, it captures the overall thermal distribution pattern within a large receptive field; second-scale spatial distribution response. The second pooling convolution unit Processed, The number of channels is 128, which preserves the fine positional response in the horizontal direction while compressing the spatial dimension; the third-scale spatial distribution response The third pooling convolution unit Processed, The number of channels is 64, the spatial dimension is compressed to 1×1, and the highest level abstract representation of the global thermal field distribution is output. After stitching the spatial distribution responses at various scales along the channel dimension, the occupancy probability values ​​of each spatial position within the fan's forward field of view are output via a fully connected layer, resulting in a discrete spatial occupancy probability sequence; the stitching operation... , , After applying global average pooling to each vector and flattening them into one-dimensional vectors, they are then concatenated along the channel dimension to obtain a multi-scale fused feature vector. The fully connected layer will Mapped to Each output component is constrained to a certain value by the Sigmoid activation function. The interval is used to obtain the occupancy probability value of each spatial location; in this embodiment, the output dimension of the fully connected layer is... ; The discrete space occupancy probability sequence includes The component, the first Each component Indicates the first [number]th [unit] in the field of view in front of the fan. The probability that a spatial location is occupied by people. , This represents the total number of spatial divisions within the field of view; in this embodiment... Set it to 32.

[0022] Optionally, to improve the stability of the discrete spatial occupancy probability sequence between consecutive sampling times and avoid drastic fluctuations in probability values ​​caused by single-frame detection noise, which would affect the stability of subsequent boundary extraction, after the discrete spatial occupancy probability sequence is output by the decoding layer, a temporal exponential moving average smoothing process is introduced for each component of the sequence. The update formula for the temporal exponential moving average smoothing is as follows: ; in, Indicates time For the The occupancy probability of each spatial location is smoothed by a time-series exponential moving average. Indicates time For the Smoothed occupancy probability of a spatial location Indicates the decoding layer at time The output of the first The original occupancy probability of a spatial location. This represents the exponentially smoothed decay coefficient, in this embodiment... Set to 0.7; after introducing time-series exponential moving average smoothing, the discrete spatial occupancy probability sequence used in subsequent step B3 is changed to the smoothed sequence. This is to improve the consistency and noise resistance of boundary extraction results in the temporal dimension.

[0023] B3: Input the discrete spatial occupancy probability sequence into the occupancy boundary extraction module of the spatial thermal field temporal sensing network, extract the leftmost and rightmost spatial positions exceeding the occupancy threshold and map them to the fan mechanical coordinate system, output the target oscillation boundary angle and the spatial occupancy probability sequence within the boundary angle range, wherein the target oscillation boundary angle includes the target oscillation left boundary angle and the target oscillation right boundary angle, including: The discrete space occupancy probability sequence is input into the occupancy boundary extraction module, which extracts each component of the discrete space occupancy probability sequence. With preset occupancy threshold A comparison is performed one by one; in this embodiment, the preset occupancy threshold is... Set to 0.5; when there exists a condition that satisfies this condition. When the spatial location meets the conditions, if the infrared sensor detects a person, the fan will automatically start running, and the leftmost spatial location index that meets the conditions will be determined. With the rightmost spatial location index Specifically: ; in, Indicates the first The probability of spatial occupancy at each spatial location. This indicates the preset occupancy threshold. This represents the total number of spatial divisions within the field of view. This represents the leftmost spatial location index that satisfies the occupancy threshold condition. This represents the rightmost spatial location index that satisfies the occupancy threshold condition; when no such condition exists... When the spatial location of the condition is determined, the boundary extraction module does not output the boundary index, determines that the infrared light does not detect personnel, triggers the default wind sweeping strategy, controls the fan to oscillate at a constant angular velocity within the global maximum boundary angle range, and if no personnel are detected for ten minutes, the fan will automatically stop running, and the subsequent dynamic modulation steps B4 and B5 will not be executed. Index the leftmost spatial position With the rightmost spatial location index By mapping the linear coordinates to the fan's mechanical coordinate system, the left boundary angle of the target oscillation is obtained. Angle with the right boundary of the target head The linear coordinate transformation formula is as follows: ; in, Indicates spatial location index The corresponding angle value in the fan's mechanical coordinate system. This represents the left-hand limit angle of the fan's mechanical coordinate system. This represents the right-hand limit angle of the fan's mechanical coordinate system; when Pick Time of income The value is the left boundary angle of the target head shake. ,when Pick Time of income The value is the right boundary angle of the target head shake. In this embodiment, Set to -60° Set to +60°, with the fan's mechanical coordinate system taking the front of the fan head as 0°, leftward deflection as negative, and rightward deflection as positive; Simultaneously extract the index range from the discrete space occupancy probability sequence. The subsequence within, the subsequence relative to the left boundary angle of the target head shake Target head shaking right boundary angle Together they constitute the input for step B4.

[0024] Optionally, to prevent sudden jitter in the head-shaking range caused by frequent contraction of the boundary index due to personnel briefly leaving the local area of ​​the field of view or sensor noise disturbance, when determining the boundary index... and Subsequently, a boundary extension protection mechanism was introduced to protect against [various issues]. To the left and Expand to the right One index unit is used to obtain the extended protection left boundary index. Right boundary index after extended protection The extended formula is: ; in, This represents the left boundary index after extended protection. This represents the right boundary index after extended protection. In this embodiment, the boundary index extension amount is represented. Set to 2; and These respectively indicate truncating the expanded result to the valid range of the index. The maximum and minimum value operations within the boundary; after introducing the boundary extension protection mechanism, the boundary index used in subsequent step B4 is changed to and The corresponding target head-shaking left boundary angle Angle with the right boundary of the target head The spatial occupancy probability subsequence used in step B4 is recalculated using the linear coordinate transformation formula and correspondingly changed to the index range. subsequence within. like Figure 2 The figure shows the spatial occupancy probability distribution and boundary extraction effect when there are two people in the field of view, located at approximately -20° and +30° respectively. The figure demonstrates the accurate locking of the effective head-shaking range by the probability threshold truncation and boundary expansion protection mechanism.

[0025] B4: Input the target head-shaking boundary angle and the spatial occupancy probability sequence within the boundary angle range into the trajectory modulation output head of the spatial thermal field temporal sensing network. Within the boundary angle range, convert the spatial occupancy probability of each spatial position into the corresponding angular velocity according to the occupancy inverse mapping angular velocity modulation formula, generating a continuous asymmetric angular velocity control sequence, including: The target's left boundary angle Target head shaking right boundary angle and index range The spatial occupancy probability subsequence within the range is input to the trajectory modulation output head. If the boundary extension protection mechanism is enabled, it is replaced with the extended angle and index. The trajectory modulation output head is configured for the index range. Index of each spatial location within The angular velocity of the fan head at the angle corresponding to the index is calculated according to the inverse mapping angular velocity modulation formula: ; in, Indicates the spatial position index of the fan head. Angular velocity at the corresponding angle This indicates the preset lower limit of angular velocity. This indicates the preset upper limit of angular velocity. Indicates the first The probability of spatial occupancy at a given spatial location; when hour, Pick ;when hour, Pick In this embodiment, Set to 5° / s, Set to 30° / s; when When the value is between 0 and 1 exist to The distance is determined linearly according to the formula; based on the angle of the fan head from the left boundary of the target oscillation. To the right boundary angle of the target head The scan order will determine the index range. The corresponding spatial location index i Arranged sequentially, a continuous asymmetric angular velocity control sequence is obtained. The angular velocity value at each position in the asymmetric angular velocity control sequence is determined independently position by position based on the spatial occupancy probability of that position. The continuous asymmetric angular velocity control sequence is denoted as... The sequence length is The elements and index range in the sequence The spatial locations within the room correspond one-to-one.

[0026] B5: Transmits a continuous asymmetric angular velocity control sequence to the fan motor drive unit, continuously adjusting the motor speed and direction to drive the fan head to adaptively oscillate within the target oscillation boundary angle, including: A continuous asymmetric angular velocity control sequence is transmitted to the fan's motor drive unit. The motor drive unit outputs a drive current in real time based on the angular velocity command values ​​at each moment in the asymmetric angular velocity control sequence, thereby adjusting the motor's speed and direction. In this embodiment, the motor drive unit uses pulse width modulation to output the drive current, and the angular velocity command values... With pulse width modulation duty cycle The mapping relationship between them is as follows: ; in, Indicates the spatial position index of the fan head. The pulse width modulation duty cycle at the corresponding angle. Indicates the lower limit of the corresponding angular velocity. The minimum duty cycle, Indicates the upper limit of the corresponding angular velocity. The maximum duty cycle in this embodiment Set to 20%, The setting is 90%; the motor drive unit acquires the current angular position of the fan head in real time through an incremental rotary encoder installed on the motor shaft. Used for direction switching detection; When the fan head moves from the left boundary angle of the target oscillation Shaking the right boundary angle of the target During directional movement, the motor drive unit moves from index [index] according to the asymmetric angular velocity control sequence. to The angular velocity command value in the direction drives the motor to rotate in the forward direction; when the fan head reaches the right boundary angle of the target oscillation... Then, the motor drive unit controls the motor to commutate, starting from the index in the asymmetric angular velocity control sequence. to The angular velocity command value in the direction drives the motor to rotate in the opposite direction; the angle at the right boundary of the head-shaking motion to reach the target. The determination criteria are: The angle at the left boundary of the head when reaching the target The determination criteria are: When the corresponding conditions are met, the reversal operation is triggered immediately. This process is repeated, driving the fan head to oscillate at the left boundary angle of the target. To the right boundary angle of the target head It continuously performs adaptive head shaking.

[0027] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0028] 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 the present invention, 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) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0029] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A fan adaptive oscillation method based on AI spatial awareness, characterized in that, Includes the following steps: B1: Input consecutive low-resolution infrared sensor image frames into the initial feature extraction layer of the spatial thermal field temporal sensing network, perform inter-frame differential temporal convolution operation, and extract spatial thermal distribution feature maps. B2: Input the spatial heat distribution feature map into the decoding layer of the spatial heat field temporal sensing network, perform multi-scale spatial hierarchical pooling operation, and generate a discrete spatial occupancy probability sequence of the corresponding fan front field of view; B3: Input the discrete space occupancy probability sequence into the occupancy boundary extraction module of the spatial thermal field time sequence sensing network, extract the leftmost and rightmost spatial positions that exceed the occupancy threshold and map them to the fan mechanical coordinate system, and output the target oscillation boundary angle and the space occupancy probability sequence within the boundary angle range. The target oscillation boundary angle includes the target oscillation left boundary angle and the target oscillation right boundary angle. B4: Input the target head-shaking boundary angle and the spatial occupancy probability sequence within the boundary angle range into the trajectory modulation output head of the spatial thermal field temporal sensing network. Within the boundary angle range, convert the spatial occupancy probability of each spatial position into the corresponding angular velocity according to the occupancy inverse mapping angular velocity modulation formula to generate a continuous asymmetric angular velocity control sequence. B5: Transmits a continuous asymmetric angular velocity control sequence to the fan's motor drive unit, continuously adjusting the motor's speed and direction, driving the fan head to adaptively oscillate within the target oscillation boundary angle.

2. The AI-based spatial awareness-based adaptive fan oscillation method according to claim 1, characterized in that, Step B1 includes: A data acquisition channel for a low-resolution infrared sensor is established to continuously acquire infrared image frames within the field of view in front of the fan at a preset sampling frequency, resulting in a continuous sequence of infrared image frames. A spatial thermal field temporal sensing network is constructed, which consists of an initial feature extraction layer, a decoding layer, an occupancy boundary extraction module, and a trajectory modulation output head connected in series. The initial feature extraction layer consists of convolution operation units connected in series. The continuous sequence of infrared image frames is input to the initial feature extraction layer, and inter-frame difference temporal convolution operation is performed: first, the infrared image frames at adjacent time points are subtracted pixel by pixel to obtain an inter-frame difference matrix; then, a convolution kernel operation is applied to the inter-frame difference matrix to extract local feature representations reflecting the change of spatial heat distribution over time; the local feature representations are combined to form a spatial heat distribution feature map.

3. The AI-based spatial awareness-based adaptive fan oscillation method according to claim 1, characterized in that, Step B2 includes: The spatial thermal distribution feature map is input to the decoding layer of the spatial thermal field temporal sensing network. This decoding layer is composed of pooling convolutional units with progressively decreasing resolution, connected in series. During multi-scale spatial hierarchical pooling, pooling convolutional kernels with different receptive field sizes are used to extract the spatial distribution response at each scale along the spatial dimension of the feature map. These responses are then concatenated along the channel dimension and output via a fully connected layer, yielding the occupancy probability values ​​for each spatial location within the field of view in front of the fan, thus obtaining a discrete spatial occupancy probability sequence. This discrete spatial occupancy probability sequence includes... The component, the first Each component Indicates the first [unit] in the field of view in front of the fan The probability that a spatial location is occupied by people. , This represents the total number of spatial divisions within the field of view.

4. The AI-based spatial awareness-based adaptive fan oscillation method according to claim 1, characterized in that, Step B3 includes: The discrete space occupancy probability sequence is input into the occupancy boundary extraction module, which extracts each component of the discrete space occupancy probability sequence. With preset occupancy threshold Perform a comparison one by one; when there exists a satisfying When the spatial location meets the conditions, if the infrared sensor detects a person, the fan will automatically start running, and the leftmost spatial location index that meets the conditions will be determined. With the rightmost spatial location index Specifically: ； in, Indicates the first The probability of spatial occupancy at each spatial location. This indicates the preset occupancy threshold. This represents the total number of spatial divisions within the field of view. This represents the leftmost spatial location index that satisfies the occupancy threshold condition. This represents the rightmost spatial location index that satisfies the occupancy threshold condition.

5. The AI-based spatial awareness-based adaptive fan oscillation method according to claim 4, characterized in that, Step B3 further includes: When there is no satisfying When the spatial location of the condition is determined, the boundary extraction module does not output the boundary index, determines that the infrared light does not detect personnel, triggers the default wind sweeping strategy, controls the fan to oscillate at a constant angular velocity within the global maximum boundary angle range, and if no personnel are detected for ten minutes, the fan will automatically stop running, and the subsequent dynamic modulation steps B4 and B5 will not be executed. Index the leftmost spatial position With the rightmost spatial location index By mapping the linear coordinates to the fan's mechanical coordinate system, the left boundary angle of the target oscillation is obtained. Angle with the right boundary of the target head The linear coordinate transformation formula is as follows: ； in, Indicates spatial location index The corresponding angle value in the fan's mechanical coordinate system. This represents the left-hand limit angle of the fan's mechanical coordinate system. This represents the right-hand limit angle of the fan's mechanical coordinate system; when Pick Time of income The value is the left boundary angle of the target head shake. ,when Pick Time of income The value is the right boundary angle of the target head shake. Simultaneously, extract the index range from the discrete space occupancy probability sequence. The subsequence within, the subsequence relative to the left boundary angle of the target head shake Target head shaking right boundary angle Together they constitute the input for step B4.

6. The AI-based spatial awareness-based adaptive fan oscillation method according to claim 1, characterized in that, Step B4 includes: The target's left boundary angle Target head shaking right boundary angle and index range The spatial occupancy probability subsequence within the range is input to the trajectory modulation output head, which is configured for the index range. Index of each spatial location within The angular velocity of the fan head at the angle corresponding to the index is calculated according to the inverse mapping angular velocity modulation formula: ； in, Indicates the spatial position index of the fan head. Angular velocity at the corresponding angle This indicates the preset lower limit of angular velocity. This indicates the preset upper limit of angular velocity. Indicates the first The probability of spatial occupancy at a given spatial location; when hour, Pick ;when hour, Pick ;when When the value is between 0 and 1 exist to The distance is determined linearly according to the formula; based on the angle of the fan head from the left boundary of the target oscillation. To the right boundary angle of the target head The scan order will determine the index range. The corresponding spatial location index i Arranged sequentially, a continuous asymmetric angular velocity control sequence is obtained. The angular velocity value at each position in the asymmetric angular velocity control sequence is determined independently position by position based on the spatial occupancy probability of that position.

7. The AI-based spatial awareness-based adaptive fan oscillation method according to claim 1, characterized in that, Step B5 includes: A continuous asymmetric angular velocity control sequence is transmitted to the fan's motor drive unit. The motor drive unit outputs a drive current in real time based on the angular velocity command values ​​at each moment in the asymmetric angular velocity control sequence, adjusting the motor's speed and direction. When the fan head moves from the target oscillation left boundary angle... Shaking the right boundary angle of the target During directional movement, the motor drive unit moves from index [index] according to the asymmetric angular velocity control sequence. to The angular velocity command value in the direction drives the motor to rotate in the forward direction; when the fan head reaches the right boundary angle of the target oscillation... Then, the motor drive unit controls the motor to commutate, starting from the index in the asymmetric angular velocity control sequence. to The angular velocity command value in the direction drives the motor to rotate in the opposite direction; this process repeats, driving the fan head to rotate at the left boundary angle of the target oscillation. To the right boundary angle of the target head It continuously performs adaptive head shaking.

Images