Hair dryer temperature control method and device based on intelligent sensor

Abstract

The invention provides a blower temperature control method and device based on an intelligent sensor. The blower temperature control method comprises the steps that a multi-dimensional sensor data set in the operation process of a blower is obtained; performing hair state feature analysis on the multi-dimensional sensor data set to obtain a hair state feature set; executing a user behavior pattern analysis operation based on the hair state feature set to generate a user behavior pattern feature set; calling a built-in autonomous learning module of the hair dryer to perform joint learning on the hair state feature set and the user behavior mode feature set, and outputting a hair state-behavior mode association model; and generating a personalized temperature control and wind speed adjustment strategy according to the hair state-behavior mode correlation model. According to the invention, the output adjustment parameters can be directly matched with the current hair state and the user behavior mode, and the adaptability and accuracy of temperature control and wind speed adjustment are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control, and more specifically, to a method and apparatus for temperature control of a hair dryer based on intelligent sensors. Background Technology

[0002] Currently, common hair dryer temperature control methods typically involve users manually setting temperature levels or relying on preset fixed temperature control modes. This method of adjustment is difficult to adapt to the actual condition of the user's hair and usage behavior, and can easily lead to situations where the temperature is too high and damages the hair, or the temperature is too low and affects the drying efficiency. Therefore, how to improve the accuracy and adaptability of hair dryer temperature control has become an urgent technical problem to be solved. Summary of the Invention

[0003] This invention provides a method and device for temperature control of a hair dryer based on a smart sensor.

[0004] In a first aspect, embodiments of the present invention provide a hair dryer temperature control method based on intelligent sensors. The method includes: acquiring a multi-dimensional sensor data set during the operation of the hair dryer; performing hair state feature analysis on the multi-dimensional sensor data set to obtain a hair state feature set; performing user behavior pattern analysis based on the hair state feature set to generate a user behavior pattern feature set; calling the self-learning module built into the hair dryer to jointly learn the hair state feature set and the user behavior pattern feature set to output a hair state-behavior pattern association model; and generating a personalized temperature control and fan speed adjustment strategy based on the hair state-behavior pattern association model.

[0005] Secondly, embodiments of the present invention provide a hair dryer temperature control device, comprising: a memory storing a computer program; and a processor for loading the computer program to implement the hair dryer temperature control method based on a smart sensor as described above.

[0006] This invention provides a hair dryer temperature control method based on intelligent sensors. It acquires multi-dimensional sensor datasets during hair dryer operation, merges them, and performs hair state feature analysis to obtain a hair state feature set. This transforms raw sensor data into effective features representing hair physical properties and styling needs. Based on the hair state feature set, it performs user behavior pattern analysis to generate a user behavior pattern feature set, establishing an intrinsic correlation between the objective hair state and the user's subjective usage behavior. This allows the adjustment logic to better align with user habits. The hair dryer's built-in self-learning module performs joint learning processing on the hair state feature set and the user behavior pattern feature set, outputting a hair state-behavior pattern association model. This learning enables adaptability to individual hair state differences and usage behavior changes, avoiding the limitations of relying on fixed rules or pre-stored templates. Based on this association model, it generates personalized temperature control and fan speed adjustment strategies, ensuring that the output adjustment parameters directly match the current hair state and user behavior pattern, improving the adaptability and accuracy of temperature control and fan speed adjustment. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a flowchart of a hair dryer temperature control method based on a smart sensor provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the composition of a hair dryer temperature control device provided in an embodiment of the present invention. Detailed Implementation

[0009] See also Figure 1 , Figure 1 The flowchart illustrates a hair dryer temperature control method based on intelligent sensors, provided in an embodiment of the present invention. This method can be executed by a hair dryer temperature control device, which can be the control module of the hair dryer or the hair dryer itself. The method specifically includes the following steps: Step S100: Acquire a set of multi-dimensional sensor data during the operation of the hair dryer.

[0010] A multidimensional sensor dataset is a collection of data gathered by various types of sensors during the operation of a hair dryer. These sensors monitor and collect information about the hair dryer's operating status, hair condition, and the surrounding environment from different dimensions and perspectives. Specifically, a multidimensional sensor dataset can include temperature data collected by a temperature sensor, humidity data collected by a humidity sensor, airflow sensor data, position sensor data, timestamp data, and real-time temperature change rate. Temperature data reflects the temperature of the hair dryer's air outlet, the temperature of the hair surface, and the temperature of the surrounding environment; humidity data reflects the moisture content of the hair and the humidity level of the surrounding environment; airflow sensor data monitors parameters such as the intensity, direction, and speed of airflow; position sensor data records the hair dryer's position and movement trajectory in space; timestamp data provides the time dimension for other data, allowing the data to be arranged and analyzed chronologically; and the real-time temperature change rate reflects the rate at which temperature changes over time.

[0011] In practical applications, various types of sensors can be installed on hair dryers to obtain multi-dimensional sensor data sets. For example, a temperature sensor can be installed near the air outlet to monitor the outlet temperature in real time; a position sensor can be installed on the handle to record the position and movement trajectory of the hair dryer; and humidity and airflow sensors can be installed at the air inlet or other suitable locations to collect humidity and airflow data, respectively. Simultaneously, an internal clock module records timestamp data, and the real-time temperature change rate is obtained by differential calculation of the temperature data. The data collected by these sensors is transmitted to the data processing unit inside the hair dryer for further processing and analysis.

[0012] Step S200: Perform hair state feature analysis on the multidimensional sensor data set to obtain a hair state feature set.

[0013] Hair condition feature analysis involves processing and analyzing multi-dimensional sensor data sets to extract feature information that reflects the condition of the hair. The hair condition feature set is a collection of these extracted features that can comprehensively and accurately describe the current state of the hair, including hair texture, hair volume, and target styling features.

[0014] As one implementation method, step S200 can be specifically implemented as the following steps S210~S270: Step S210: Spatiotemporally align the multidimensional sensor data set, and integrate the data collected by different types of sensors into a multidimensional data matrix under a unified spatiotemporal framework through timestamp synchronization and spatial coordinate mapping.

[0015] Spatiotemporal alignment is the process of unifying and aligning data collected by different types of sensors in time and space, enabling these data to be analyzed and processed within the same spatiotemporal framework. Timestamp synchronization involves uniformly calibrating the timestamps of data collected by various sensors to ensure temporal comparability. Spatial coordinate mapping involves uniformly transforming the spatial location information of data collected by different sensors, allowing them to be represented in the same spatial coordinate system. A multidimensional data matrix is ​​a matrix containing spatiotemporally aligned data, integrating data collected by different types of sensors into a unified data structure with spatiotemporal dimensions.

[0016] For example, firstly, the timestamps of each sensor can be calibrated using hardware clock synchronization to ensure consistent time accuracy. Then, based on the installation location and relative relationship of each sensor, a unified spatial coordinate system is established, and the spatial location information of the data collected by each sensor is mapped into this coordinate system. For instance, if a temperature sensor is installed at the air outlet of a hair dryer and an airflow sensor is installed on the side of the hair dryer, the spatial location information of their collected data is transformed into a spatial coordinate system with the hair dryer as the origin. Finally, the data, after timestamp synchronization and spatial coordinate mapping, is arranged into a multi-dimensional data matrix according to certain rules, such as chronological order and sensor type.

[0017] Step S220: Perform multimodal feature fusion on the multidimensional data matrix. Use an adaptive weight allocation mechanism to weight the temperature-related data, humidity-related data and airflow sensing data to generate multimodal fusion features. The adaptive weight allocation mechanism adjusts the weight ratio according to the signal-to-noise ratio of each modality data.

[0018] Multimodal feature fusion is the process of fusing different types of data (such as temperature-related data, humidity-related data, and airflow sensing data) from a multidimensional data matrix to extract more representative and comprehensive feature information. An adaptive weight allocation mechanism is a mechanism that dynamically adjusts the weight ratio based on the signal-to-noise ratio (SNR) of each modality. The SNR is the ratio of signal strength to noise strength; a higher SNR indicates better data quality and more effective information. By adjusting the weight ratio based on the SNR, high-quality data can play a greater role in the fusion process, thereby improving the quality of the fused features. Multimodal fusion features are the feature information obtained after multimodal feature fusion, integrating information from multiple aspects such as temperature, humidity, and airflow sensing, and can more accurately reflect the state of hair under airflow. In implementation, the SNR of each modality can be calculated first. For temperature-related data, the SNR can be obtained by statistically analyzing the temperature data over a period of time and calculating the variance of the signal and noise. A similar method can be used to calculate the SNR of humidity-related data and airflow sensing data. Then, based on the calculated signal-to-noise ratio (SNR), the weight ratios of each modality are dynamically adjusted using a pre-defined functional relationship. For example, a linear or non-linear function can be used to map the SNR to the weight ratios. Finally, the data from each modality are weighted and summed according to the adjusted weights to obtain the multimodal fusion features.

[0019] Step S230: Extract temperature and humidity response features and airflow disturbance distribution features from the multimodal fusion features. Temperature and humidity response features characterize the coupling change law of hair surface temperature and humidity with airflow. Airflow disturbance distribution features characterize the spatial distribution of turbulence intensity when airflow passes through the hair area.

[0020] Temperature and humidity response characteristics represent the coupled changes in hair surface temperature and humidity under the influence of airflow, reflecting the relationship between temperature and humidity and their changes over time. Airflow disturbance distribution characteristics represent the spatial distribution of turbulence intensity as airflow passes through the hair region.

[0021] As one implementation method, step S230 can be specifically implemented as the following steps S231~S235: Step S231: Perform joint time-domain analysis on the temperature-related data and humidity-related data in the multimodal fusion features, extract continuous time segments through a sliding window, calculate the mutual information value of temperature and humidity in each time segment, and obtain the temperature and humidity mutual information sequence.

[0022] Joint time-domain analysis is the process of jointly analyzing temperature-related and humidity-related data along the time dimension, aiming to uncover the interrelationships and patterns of change between them over time. A sliding window is used to extract continuous time segments, sliding across the time series data with a fixed window size, extracting one continuous time segment at a time. Mutual information is an indicator used to measure the correlation between two random variables, representing the amount of information one variable contains about the other. The temperature and humidity mutual information sequence is a sequence composed of the mutual information values ​​of temperature and humidity calculated within each time segment, reflecting the change in the correlation between temperature and humidity over time. For example, the size of the sliding window and the sliding step size can be determined first. For instance, the sliding window size can be set to 10 time units, and the sliding step size to 1 time unit. Then, temperature-related and humidity-related data are extracted from the multimodal fusion features and extracted using the sliding window method to obtain a series of continuous time segments. For each time segment, the mutual information value of temperature and humidity is calculated. Specifically, a probability distribution-based mutual information calculation method can be used, i.e., first estimating the joint probability distribution and marginal probability distribution of temperature and humidity, and then calculating according to the general definition formula of mutual information. Finally, the mutual information values ​​of each time segment are arranged in chronological order to obtain the temperature and humidity mutual information sequence.

[0023] Step S232: Perform differential processing on the temperature and humidity mutual information sequence to obtain the temperature and humidity mutual information change rate, and use the temperature and humidity mutual information change rate as the core component of the temperature and humidity response characteristics.

[0024] Differential processing calculates the rate of change of data by measuring the difference between two adjacent data points. The rate of change of temperature and humidity mutual information is the result obtained after differential processing of the temperature and humidity mutual information sequence, reflecting the speed at which the correlation between temperature and humidity changes over time. The rate of change of temperature and humidity mutual information is used as the core component of the temperature and humidity response characteristics because it can sensitively capture the coupled changes in hair surface temperature and humidity with the influence of airflow.

[0025] In practical calculations, the temperature and humidity mutual information sequence can be processed using first-order difference processing, that is, calculating the difference between two adjacent mutual information values. For example, for the i-th and (i+1)-th mutual information values ​​in the temperature and humidity mutual information sequence, their difference is calculated as the rate of change of temperature and humidity mutual information at the i-th time point. Combining the rates of change of temperature and humidity mutual information at all time points yields the sequence of rates of change of temperature and humidity mutual information, which serves as the core component of the temperature and humidity response characteristics.

[0026] Step S233: Perform spatial spectrum estimation on the airflow sensing data in the multimodal fusion features, and calculate the airflow energy distribution in different spatial orientations using a beamforming algorithm to obtain the airflow spatial spectrum features.

[0027] Spatial spectrum estimation is the process of analyzing airflow sensor data in a spatial dimension to estimate the distribution of airflow energy in different spatial orientations. Beamforming algorithms, used to enhance directional signals and suppress interference from other directions, form directional beams by weighted summation of signals collected from multiple sensors, thereby calculating the airflow energy distribution in different spatial orientations. Airflow spatial spectrum features are characteristic information composed of the calculated airflow energy distribution in different spatial orientations, providing a direct reflection of the airflow distribution in space.

[0028] As one implementation method, step S233 calculates the airflow energy distribution in different spatial orientations using a beamforming algorithm to obtain airflow spatial spectrum characteristics. Specifically, this can be implemented as the following steps S2331~S2335: Step S2331: The airflow sensing data in the multimodal fusion features is divided into spatial arrays, and the one-dimensional data stream collected by the distributed airflow sensor is reconstructed into a two-dimensional spatial array signal. Each array unit corresponds to the airflow sampling value in different spatial orientations.

[0029] Spatial array partitioning is the process of dividing airflow sensor data according to spatial location to form a two-dimensional spatial array structure. Distributed airflow sensors consist of multiple airflow sensors installed in space, collecting airflow sensor data as a one-dimensional data stream. Reconstructing this one-dimensional data stream into a two-dimensional spatial array signal involves rearranging the data collected by different sensors at different times according to spatial location, so that each array unit corresponds to airflow sampling values ​​from different spatial orientations. Such a two-dimensional spatial array signal can more intuitively reflect the distribution of airflow in space, providing a more suitable data structure for subsequent beamforming algorithm calculations.

[0030] For example, the division method and size of the spatial array are first determined. For instance, the space around the hair dryer can be divided into a two-dimensional grid, with each grid cell corresponding to an array cell. Then, based on the installation location and acquisition time of the distributed airflow sensors, the one-dimensional airflow sensing data is allocated to the corresponding array cells to form a two-dimensional spatial array signal.

[0031] Step S2332: Construct the array manifold matrix of the beamforming algorithm. Calculate the guide vector for each spatial orientation based on the physical arrangement of the sensors. The guide vector contains the trigonometric function relationship between the azimuth angle and the sensor spacing. Arrange the guide vectors in azimuth order to form the array manifold matrix.

[0032] The array manifold matrix describes the response characteristics of a sensor array to signals from different spatial orientations. The steering vector, representing the sensor array's response to a signal from a specific spatial orientation, incorporates the trigonometric relationship between the azimuth angle and the sensor spacing. By calculating the steering vectors for different spatial orientations and arranging them in azimuth order, the array manifold matrix can be formed. The construction of the array manifold matrix provides the foundation for subsequent calculations of the beamforming weight vector.

[0033] The steering vector needs to be determined based on the physical arrangement and azimuth of the sensors. Assuming there are M sensors in the sensor array, for a given azimuth angle θ, the steering vector a(θ) can be calculated using the following formula: a(θ) = [e^(-θ) / θ] (-j2πd1sinθ / λ) ,e (-j2πd2sinθ / λ) ,...,e (-j2πdMsinθ / λ) ] T λ represents the distance between the i-th sensor and the reference sensor, and λ represents the wavelength of the signal. By arranging the guide vectors corresponding to different azimuth angles in azimuth order, we can obtain the array manifold matrix A=[a(θ1),a(θ2),...,a(θN)], where θi represents the i-th azimuth angle and N represents the number of azimuth angles.

[0034] Step S2333: Construct a beamforming weight vector based on the array manifold matrix and the spatial array signal, and calculate the weight coefficients using the minimum variance distortionless response criterion. The weight coefficients are associated with the conjugate transpose of the array manifold matrix and the inverse of the noise covariance matrix.

[0035] The beamforming weight vector is used to weight and sum the signals of a spatial array to form a directional beam. The minimum variance distortionless response criterion is used to calculate the beamforming weight coefficients. Its goal is to minimize the variance of the output signal while ensuring a distortion-free response to the desired signal, thereby suppressing noise and interference. The weight coefficients are correlated with the conjugate transpose of the array manifold matrix and the inverse of the noise covariance matrix. This correlation allows for the calculation of the optimal weight coefficients, enabling the beamforming algorithm to effectively enhance the signal in the desired direction.

[0036] Specifically, let the spatial array signal be x(t), the array manifold matrix be A, the steering vector of the desired signal be a(θ0), and the noise covariance matrix be Rnn. According to the minimum variance distortionless response criterion, the beamforming weight vector w can be calculated using the following formula: w = (A... H Rnn -1 A) -1 A H Rnn -1 a(θ0), where A H Rn represents the conjugate transpose of the array manifold matrix. -1This represents the inverse of the noise covariance matrix.

[0037] Step S2334: Perform inner product operation on the weight vector formed by the spatial array signal and the beam to obtain the airflow energy value in different spatial orientations. Perform spatial smoothing on the energy value and eliminate array mutual coupling interference by averaging through a sliding window to obtain the preliminary airflow energy distribution.

[0038] By performing an inner product operation on the weighted vector formed by the spatial array signal and the beam, the airflow energy values ​​at different spatial orientations can be obtained. Spatial smoothing is a process that smooths the obtained airflow energy values ​​in the spatial dimension to eliminate the effects of array mutual coupling interference and noise. Sliding window averaging is a commonly used spatial smoothing method. It involves sliding a fixed-size window across space and averaging the energy values ​​within the window to obtain the smoothed energy values. The preliminary airflow energy distribution is the distribution of airflow energy at different spatial orientations after spatial smoothing.

[0039] For example, the airflow energy value P(θ) = |w| is obtained by performing an inner product operation on the spatial array signal x(t) and the beamforming weight vector w at each spatial azimuth. H x(t)| 2 , where w H This represents the conjugate transpose of the beamforming weight vector. Then, these energy values ​​are processed using a sliding window averaging method. For example, a 3×3 sliding window is set, and the average of the nine energy values ​​within the window is calculated to obtain the smoothed energy value at the center of the window. This process is applied to all energy values ​​in the space to obtain the preliminary airflow energy distribution.

[0040] Step S2335: Detect spectral peaks in the preliminary airflow energy distribution, retain the energy components corresponding to the azimuth angles whose energy values ​​are higher than the average energy by a preset multiple, arrange the retained energy components in azimuth order, and generate airflow spatial spectrum features.

[0041] Peak detection is the process of identifying peak points in the initial airflow energy distribution where the energy value exceeds a certain threshold. The average energy is the average of all energy values ​​in the initial airflow energy distribution. The preset multiple is a pre-defined coefficient used to determine how many times higher the energy value than the average energy corresponds to the energy component at which the azimuth angle should be retained. Arranging the retained energy components in azimuth order generates the airflow spatial spectrum characteristics, highlighting the main distribution areas of airflow energy in space.

[0042] For example, firstly, the average energy of the initial airflow energy distribution is calculated. Then, the average energy is multiplied by a preset factor to obtain a threshold. Each energy value in the initial airflow energy distribution is compared, and if an energy value is higher than the threshold, its corresponding azimuth and energy component are retained. Finally, the retained energy components are arranged in order of azimuth to form the airflow spatial spectrum characteristics.

[0043] Step S234: Threshold segmentation is performed on the airflow spatial spectrum features, and spatial components with energy higher than the preset energy percentage are retained and combined into airflow disturbance distribution features.

[0044] Threshold segmentation is the process of dividing the energy components in the spatial spectrum of airflow based on a preset energy percentage threshold, retaining the spatial components with energy levels higher than that threshold. The preset energy percentage is a pre-defined ratio used to determine the range of energy components to be retained. The airflow disturbance distribution feature is a feature composed of the spatial components with energy levels higher than the preset energy percentage, which can highlight areas where airflow disturbances are stronger in space.

[0045] For example, a preset energy percentage is first determined. For instance, the preset energy percentage is set to 80%. Then, the sum of all energy components in the airflow spatial spectrum characteristics is calculated. Next, the energy components are sorted in descending order of energy value, and the energy values ​​are accumulated sequentially until the accumulated energy value reaches 80% of the total energy. The spatial components corresponding to these energy components are retained and combined to form the airflow disturbance distribution characteristics.

[0046] Step S235: Concatenate the core components of the temperature and humidity response features with the airflow disturbance distribution features by feature dimension to obtain complete temperature and humidity response features and airflow disturbance distribution features.

[0047] By stitching together the feature dimensions, the information of temperature and humidity response features and airflow disturbance distribution features can be integrated to obtain complete temperature and humidity response features and airflow disturbance distribution features, so that these two features can more comprehensively reflect the heat and moisture transfer and airflow of hair under the action of airflow.

[0048] Step S240: Input the temperature and humidity response features and airflow disturbance distribution features into the hair quality feature recognition submodule, perform cross-modal feature association through a multi-scale feature interaction network, and output the hair quality features.

[0049] The hair texture feature recognition submodule is specifically designed for identifying hair texture characteristics. It receives temperature and humidity response features and airflow disturbance distribution features as input and processes them through an internal multi-scale feature interaction network. A multi-scale feature interaction network is a network structure capable of extracting and interacting with features at different scales. It can correlate and fuse these two different modalities of features—temperature and humidity response features and airflow disturbance distribution features—to uncover their intrinsic relationships. The hair texture features, output after processing by the multi-scale feature interaction network, reflect the hair's texture, elasticity, resilience, and other properties.

[0050] Multi-scale feature interaction networks can employ a convolutional neural network (CNN) architecture, comprising multiple convolutional layers, pooling layers, and fully connected layers. Different kernel sizes can be used in different convolutional layers to extract features at different scales. For example, a 3×3 kernel can be used in the first convolutional layer to extract smaller-scale features, while a 5×5 kernel can be used in the second convolutional layer to extract larger-scale features. Pooling layers downsample the features, reducing their dimensionality. Fully connected layers then fuse and classify features from different scales, outputting hair texture features. For example, hair texture features can be categorized into different types such as dry hair, normal hair, and oily hair.

[0051] Step S250: Based on the spatial distribution features and airflow disturbance distribution features of the multidimensional data matrix, generate hair coverage area contour features through a region density clustering algorithm, and perform convolution operation on the hair coverage area contour features and airflow disturbance distribution features to obtain hair volume features.

[0052] Region density clustering is an algorithm that clusters data points based on their density. It can identify the outline of hair-covered areas based on the spatial distribution characteristics of a multidimensional data matrix and the distribution characteristics of airflow disturbances. The hair-covered area outline features are generated by the region density clustering algorithm and describe the spatial coverage and shape of the hair. Convolving these features with the airflow disturbance distribution features allows for the fusion of their information, yielding hair volume features that reflect the quantity and density of hair.

[0053] In practical applications, the region density clustering algorithm can be DBSCAN (Density-Based Spatial Clustering of Applications with Noise). First, the spatial distribution characteristics of the multidimensional data matrix and the airflow disturbance distribution characteristics are used as input, and the DBSCAN algorithm is used for clustering. This algorithm divides data points into core points, boundary points, and noise points by setting a neighborhood radius and a minimum number of points. Core points are data points that contain at least the minimum number of points within their neighborhood radius; boundary points are data points that contain less than the minimum number of points within their neighborhood radius but are adjacent to a core point; noise points are data points that do not belong to any cluster. By connecting the core points and boundary points, the contour features of the hair-covered region can be obtained. Then, the contour features of the hair-covered region are convolved with the airflow disturbance distribution features.

[0054] Step S260: Based on the temperature and humidity response characteristics over time, combined with hair quality and volume characteristics, determine the target styling features through the styling feature inference submodule.

[0055] The temperature and humidity response characteristics over time reflect the rate of change of hair surface temperature and humidity over time, demonstrating the dynamic changes in heat and moisture transfer in hair under airflow. The styling feature inference submodule is specifically designed to infer the target styling features based on the input feature information. The target styling features are determined by this submodule and describe the user's desired hair style, such as straight hair, curly hair, or voluminous hair.

[0056] The styling feature inference submodule can employ rule-based inference methods or machine learning models. Rule-based inference methods judge and infer based on pre-defined rules and conditions, considering the gradient of temperature and humidity response features over time, hair texture characteristics, and hair volume characteristics. For example, if the gradient of temperature and humidity response features over time is small, the hair texture is dry, and the hair volume is low, then the target styling feature is inferred to be straight hair. Machine learning models can use neural network models, such as multilayer perceptrons (MLPs). Taking the gradient of temperature and humidity response features over time, hair texture characteristics, and hair volume characteristics as input, a trained MLP model makes predictions and outputs the target styling feature.

[0057] Step S270: Combine hair texture features, hair volume features, and target styling features into a hair state feature set.

[0058] The hair state feature set is a collection composed of hair texture features, hair volume features, and target styling features, comprehensively describing the current state of the hair and the user's desired style. By combining these three features, more comprehensive and accurate information can be provided for subsequent user behavior pattern analysis and personalized temperature control and airflow adjustment. For example, hair texture features, hair volume features, and target styling features can be arranged in a certain order to form a vector or matrix, which serves as the hair state feature set.

[0059] Step S300: Perform user behavior pattern analysis based on the hair state feature set to generate a user behavior pattern feature set.

[0060] User behavior pattern analysis involves processing and analyzing a set of hair state features to extract characteristic information that reflects a user's behavior patterns when using a hair dryer. The user behavior pattern feature set is a collection of these extracted features, describing the user's holding posture, movement trajectory, usage duration, and other behavioral patterns when using a hair dryer.

[0061] As one implementation method, step S300 can be specifically implemented as the following steps S310~S370: Step S310: Extract position sensing data and timestamp data from the multi-dimensional sensor data set to construct the spatiotemporal motion trajectory of the hair dryer during operation.

[0062] Position sensing data records the hairdryer's spatial location within a multi-dimensional sensor dataset, while timestamp data provides the temporal dimension to this position data. The spatiotemporal motion trajectory is the hairdryer's movement path in time and space during operation, constructed by combining position sensing data and timestamp data. By constructing this trajectory, one can intuitively understand the hairdryer's movement and temporal sequence during use.

[0063] For example, position sensing data and timestamp data are filtered from a multi-dimensional sensor dataset. By associating the position data corresponding to each timestamp and arranging them in chronological order, the spatiotemporal trajectory of the hair dryer can be obtained.

[0064] Step S320: Segment the spatiotemporal motion trajectory, and divide it into multiple motion trajectory segments based on the trajectory direction change rate and velocity continuity to obtain a set of trajectory segments.

[0065] The rate of change of trajectory direction is the speed at which the trajectory direction changes over time in a spatiotemporal motion trajectory, reflecting the degree of change in the hair dryer's motion direction. Velocity continuity refers to whether the velocity changes between adjacent points on the trajectory are continuous, and can be used to determine whether the hair dryer's motion is smooth. By segmenting the spatiotemporal motion trajectory based on the rate of change of trajectory direction and velocity continuity, it can be divided into multiple trajectory segments with different motion characteristics. The set of trajectory segments is then the collection of these segmented trajectory segments.

[0066] For example, firstly, the rate of change of direction and velocity between adjacent points on the spatiotemporal trajectory are calculated. The rate of change of direction can be obtained by calculating the change in the angle between the vectors of two adjacent points. The velocity can be obtained by dividing the distance between two adjacent points by the time interval. Then, the spatiotemporal trajectory is segmented according to preset thresholds for the rate of change of direction and velocity continuity. If the rate of change of direction of a certain segment exceeds the threshold, or the velocity is discontinuous, it is divided into a new trajectory segment.

[0067] Step S330: Perform correlation analysis between the trajectory segment set and the hair quality features in the hair state feature set, and calculate the kinematic feature parameters of each motion trajectory segment. The kinematic feature parameters include the trajectory segment curvature, average motion speed, and directional rotation angle.

[0068] Association analysis connects and compares the set of trajectory segments with hair quality characteristics, analyzing the interrelationships and influences between them. The curvature of the trajectory segment is the degree of bending of the motion trajectory segment, reflecting the turning behavior of the hair dryer during its movement. The average motion speed is the average speed of movement along the trajectory segment, reflecting the speed of the hair dryer's movement. The directional angle is the angle of change in direction between the starting and ending points of the motion trajectory segment, reflecting the change in the direction of the hair dryer's movement.

[0069] As one implementation method, step S330 can be specifically implemented as the following steps S331~S337: Step S331: Perform equal-interval coordinate interpolation on each motion trajectory segment in the trajectory segment set to obtain a high-density sampled trajectory coordinate sequence.

[0070] Equal-interval coordinate interpolation involves inserting new coordinate points at equal intervals along the motion trajectory segment to increase the sampling density of the trajectory. High-density sampled trajectory coordinate sequences are obtained through equal-interval coordinate interpolation, containing more coordinate points and more accurately describing the shape and characteristics of the motion trajectory segment.

[0071] For example, for each motion trajectory segment in the trajectory segment set, equally spaced coordinate interpolation is performed according to a preset interval distance. For instance, if the preset interval distance is 0.1m, for a motion trajectory segment with a length of 1m, 9 new coordinate points will be inserted. Linear interpolation can be used for coordinate interpolation, that is, the coordinates of the inserted points are calculated based on the coordinates of two adjacent points and the interval distance.

[0072] Step S332: Calculate the spatial distance between adjacent sampling points in the trajectory coordinate sequence, generate a distance sequence, and accumulate the distance sequences to obtain the trajectory segment length.

[0073] Spatial distance is the actual distance between adjacent sampling points in space within a trajectory coordinate sequence. A distance sequence is a sequence composed of these spatial distances between adjacent sampling points. By accumulating the distance sequences, the total length of the motion trajectory segment can be obtained. For example, the spatial distance between adjacent sampling points in the trajectory coordinate sequence is calculated using the Euclidean distance formula. The calculated spatial distances are then arranged sequentially to form a distance sequence. Finally, all elements in the distance sequence are summed to obtain the trajectory segment length.

[0074] Step S333: Calculate the ratio between the length of the trajectory segment and the time interval corresponding to the motion trajectory segment to obtain the preliminary average speed.

[0075] The time interval is the difference between the start and end times of a motion trajectory segment. By comparing the length of the trajectory segment with the corresponding time interval, we can obtain the preliminary average speed of that motion trajectory segment, reflecting the average moving speed of the hair dryer along that segment.

[0076] For example, the start time t1 and end time t2 of the motion trajectory segment are obtained from the timestamp data, and the time interval Δt = t2 - t1 is calculated. Then, the length L of the trajectory segment is compared with the time interval Δt to obtain the preliminary average velocity v = L / Δt.

[0077] Step S334: Perform curve fitting on the trajectory coordinate sequence, calculate the second derivative of the fitted curve, obtain the curvature distribution of the trajectory segment, and take the mean of the curvature distribution as the curvature of the trajectory segment.

[0078] The second derivative is obtained by taking the derivative of the fitted curve twice, reflecting the rate of change of the curve's curvature. The curvature distribution of the trajectory segment is composed of the distribution of the second derivative of the fitted curve along the trajectory segment. Taking the mean of the curvature distribution as the curvature of the trajectory segment can comprehensively reflect the degree of curvature of the motion trajectory segment.

[0079] For example, a polynomial curve fitting method can be used to fit the trajectory coordinate sequence. For instance, a quadratic polynomial y=ax can be used. 2The x and y coordinates of the trajectory coordinate sequence are fitted using the formula +bx+c, and the coefficients a, b, and c are solved using the least squares method. Then, the fitted curve is differentiated twice to obtain the second derivative. The values ​​of the second derivative on the trajectory segment form the curvature distribution of the trajectory segment. Finally, the mean of the curvature distribution is calculated as the curvature of the trajectory segment.

[0080] Step S335: Input the hair quality features into the feature mapping submodule to obtain the hair quality correlation coefficient. The hair quality correlation coefficient decreases linearly with the increase of the fineness and softness of the hair quality features.

[0081] The feature mapping submodule is used to map hair texture features to hair texture correlation coefficients. It generates corresponding correlation coefficients based on different attributes of the hair texture features through an internal mapping mechanism. The hair texture correlation coefficient is a coefficient related to hair texture features, decreasing linearly with the increase of fineness in the hair texture features, reflecting the degree to which hair texture affects the speed of the hairdryer and other operations.

[0082] As one implementation method, step S335, inputting hair quality features into the feature mapping submodule to obtain hair quality correlation coefficients, can be implemented as follows: Step S3351~S3355: Step S3351: Receive hair quality features, perform multi-dimensional decomposition on the hair quality features, and extract surface texture features, moisture retention features and thermal conductivity features from the hair quality features. Surface texture features represent the integrity of hair scales, moisture retention features represent the expansion rate of hair after absorbing water, and thermal conductivity features represent the rate at which hair conducts temperature.

[0083] Multidimensional decomposition involves breaking down hair texture characteristics according to different attributes and dimensions to extract surface texture features, moisture retention features, and thermal conductivity features. Surface texture features describe the integrity of the hair cuticles; higher cuticle integrity results in better surface texture features. Moisture retention features reflect the hair's swelling rate after absorbing water, demonstrating its water absorption capacity and moisturizing properties. Thermal conductivity features characterize the rate at which the hair conducts temperature, reflecting temperature changes when the hair is heated.

[0084] For example, different methods are used for multi-dimensional decomposition based on the specific representation of hair texture characteristics. For instance, if a hair texture characteristic is a vector, it can be split into three sub-vectors representing surface texture characteristics, moisture retention characteristics, and thermal conductivity characteristics. For surface texture characteristics, quantification can be performed by observing the integrity of the hair cuticles under a microscope. For moisture retention characteristics, the swelling rate can be calculated by measuring the weight change of the hair before and after water absorption. For thermal conductivity characteristics, the conduction rate can be calculated by measuring the temperature change of the hair during heating.

[0085] Step S3352: Input the surface texture features into the texture-coefficient mapping unit. Generate the first basic coefficient through the nonlinear mapping relationship between the texture features and the correlation coefficient. The higher the scale integrity in the surface texture features, the larger the first basic coefficient.

[0086] The texture-coefficient mapping unit is a unit that implements a non-linear mapping relationship between texture features and correlation coefficients. The first fundamental coefficient is a coefficient generated by the texture-coefficient mapping unit that is related to the surface texture features; the higher the scale integrity in the surface texture features, the larger the first fundamental coefficient. The texture-coefficient mapping unit can use a neural network model to implement the non-linear mapping relationship. For example, a multilayer perceptron (MLP) containing input, hidden, and output layers can be used. The surface texture features are used as input to the input layer, and the first fundamental coefficient is output through a non-linear transformation in the hidden layer and a linear transformation in the output layer. During training, a large amount of sample data can be used, and the weights of the neural network can be adjusted through backpropagation to ensure that the mapping relationship between the surface texture features and the first fundamental coefficient conforms to the expected result.

[0087] Step S3353: Input the moisture retention feature and the heat conduction feature into the coupling mapping unit, calculate the product of the two as the coupling feature value, and generate the second basic coefficient by the ratio of the coupling feature value to the preset benchmark value. The larger the coupling feature value, the smaller the second basic coefficient.

[0088] The coupling mapping unit is used to handle the coupling relationship between moisture retention characteristics and thermal conductivity characteristics. The coupling characteristic value is obtained by calculating the product of the moisture retention characteristic and the thermal conductivity characteristic. A preset benchmark value is a pre-defined reference value used for comparison with the coupling characteristic value. A second fundamental coefficient is generated by the ratio of the coupling characteristic value to the preset benchmark value; the larger the coupling characteristic value, the smaller the second fundamental coefficient.

[0089] Step S3354: Extract the real-time temperature change rate from the multi-dimensional sensor data set, map the real-time temperature change rate to the adjustment coefficient. The faster the temperature change rate, the larger the adjustment coefficient. The first base coefficient, the second base coefficient and the adjustment coefficient are weighted and summed. The weights are allocated according to the hair volume features in the hair state feature set. The greater the hair density in the hair volume features, the higher the weight of the second base coefficient.

[0090] The real-time temperature change rate is the rate at which temperature changes over time, recorded in a multi-dimensional sensor dataset. The adjustment coefficient is obtained by mapping the real-time temperature change rate; the faster the temperature change rate, the larger the adjustment coefficient. A comprehensive coefficient is obtained by weighted summing of the first base coefficient, the second base coefficient, and the adjustment coefficient. The weights are assigned based on hair volume features in the hair state feature set; the higher the hair density within the hair volume feature, the higher the weight of the second base coefficient.

[0091] For example, the real-time temperature change rate is extracted from a multi-dimensional sensor dataset. This real-time temperature change rate can be mapped to an adjustment coefficient using a linear mapping function, for example, adjustment coefficient = k * real-time temperature change rate, where k is the mapping coefficient. Then, the weights of the first base coefficient, the second base coefficient, and the adjustment coefficient are determined based on hair volume characteristics in the hair state feature set.

[0092] Step S3355: Normalize the weighted summation result to ensure it falls within the preset correlation coefficient range, and output the hair quality correlation coefficient.

[0093] Range normalization processes the weighted summation result to ensure its value range falls within a preset correlation coefficient interval. This preset interval is a pre-defined range of coefficient values, such as [0, 1]. Range normalization ensures the hair quality correlation coefficient is within a reasonable range, facilitating subsequent calculations and applications. For example, let the weighted summation result be x, and the preset correlation coefficient interval be [a, b]. First, the minimum value x of the weighted summation result is calculated. min and maximum value x max Then, normalize the range using the following formula: Hair quality correlation coefficient = a + (xx) min )*(ba) / (x max -x min ).

[0094] Step S336: Multiply the initial average speed with the hair quality correlation coefficient to obtain the adjusted average speed.

[0095] By multiplying the initial average speed by the hair quality correlation coefficient, the initial average speed can be adjusted according to different hair qualities to obtain the adjusted average speed. The hair quality correlation coefficient reflects the degree to which hair quality affects the speed of the hair dryer. When the hair is fine and soft, the correlation coefficient is small, and the adjusted average speed will decrease accordingly; when the hair is coarse and stiff, the correlation coefficient is large, and the adjusted average speed will increase accordingly. For example, multiplying the initial average speed v1 by the hair quality correlation coefficient k yields the adjusted average speed v2 = v1 * k.

[0096] Step S337: Combine the trajectory segment curvature, the adjusted average velocity, and the directional rotation angle into kinematic characteristic parameters.

[0097] Kinematic characteristic parameters are a set of parameters composed of trajectory segment curvature, adjusted average velocity, and directional angle. By combining these three parameters, the motion state of the hair dryer on each trajectory segment can be analyzed more accurately. For example, the trajectory segment curvature, adjusted average velocity, and directional angle can be arranged in order to form a vector or matrix as kinematic characteristic parameters.

[0098] Step S340: Calculate the trajectory smoothness index and velocity fluctuation coefficient based on the kinematic feature parameters, and generate the usage duration feature by combining the hair volume feature in the hair state feature set.

[0099] The trajectory smoothness index measures the smoothness of a motion trajectory and can be calculated using information such as trajectory segment curvature and directional angle from kinematic feature parameters. The velocity fluctuation coefficient reflects the fluctuation of motion velocity and can be calculated based on changes in the adjusted average motion velocity from the kinematic feature parameters. The usage duration feature is generated by combining the trajectory smoothness index, velocity fluctuation coefficient, and hair volume features from the hair condition feature set, reflecting the approximate duration of the user's use of the hair dryer.

[0100] For example, the trajectory smoothness index can be obtained by calculating the standard deviation of the curvature of the trajectory segments. The smaller the standard deviation, the smoother the trajectory. The velocity fluctuation coefficient can be obtained by calculating the variance of the adjusted average motion velocity. The larger the variance, the greater the velocity fluctuation. Then, based on the hair volume feature in the hair state feature set, different weights are assigned to the trajectory smoothness index and the velocity fluctuation coefficient. The higher the hair density in the hair volume feature, the higher the weight of the velocity fluctuation coefficient. The trajectory smoothness index and the velocity fluctuation coefficient are weighted and summed according to their weights to obtain a comprehensive index. Finally, based on this comprehensive index and a preset duration mapping function, the usage duration feature is generated.

[0101] Step S350: The kinematic feature parameters and usage duration features are fused temporally, and the evolution of the grip posture over time is learned through a gated recurrent unit network to generate grip posture change features.

[0102] Temporal fusion combines kinematic feature parameters and usage duration features in chronological order to form a feature sequence with a time dimension. Gated Recurrent Unit (GRU) networks are neural network models capable of processing sequential data and learning temporal dependencies within the data. By inputting kinematic feature parameters and usage duration features into the GRU, the network can learn the evolution of the grip posture over time, generating grip posture change features that reflect the dynamic changes in the user's grip posture during hair dryer use.

[0103] The gated recurrent unit (GRU) network comprises input gates, reset gates, and candidate hidden states. At each time step, the network updates the current hidden state based on the current input (kinematic feature parameters and usage duration features) and the hidden state from the previous time step, controlled by the input and reset gates. After processing over multiple time steps, the network learns the evolution of the grip posture over time. Finally, the hidden state sequence output by the network is processed to generate grip posture change features. For example, grip posture change features can be represented as a vector, where each element represents a certain feature of the grip posture at different time steps.

[0104] Step S360: Based on the trajectory smoothness index and speed fluctuation coefficient, combined with the grip posture change characteristics, the air outlet position movement characteristics are determined through the position change inference submodule.

[0105] The position change inference submodule is used to infer the air outlet position movement characteristics based on the input trajectory smoothness index, speed fluctuation coefficient, and grip posture change characteristics. The air outlet position movement characteristics describe the position movement of the blower's air outlet in space.

[0106] As one implementation method, step S360 can be specifically implemented as the following steps S361~S366: Step S361: Perform time-domain feature decomposition on the trajectory smoothness index, and decompose it into transient smoothness component and steady-state smoothness component through wavelet transform. The transient smoothness component represents the trajectory change feature where the value of the time scale less than the total duration of a single complete operation of the hair dryer is not less than a preset first difference. The steady-state smoothness component represents the trajectory trend feature where the difference between the time scale and the total duration of a single complete operation of the hair dryer is less than a preset second difference.

[0107] Temporal feature decomposition decomposes the trajectory smoothness index along the time dimension, extracting feature components at different time scales. Wavelet transform, a mathematical tool capable of simultaneous local analysis in time and frequency, can decompose the trajectory smoothness index into components of different frequencies and time scales. The transient smoothness component reflects the abrupt changes in the trajectory at shorter time scales, corresponding to trajectory changes where the value between the time scale and the total duration of a single complete operation of the hair dryer is not less than a preset first difference, i.e., much smaller than the total duration of a single complete operation of the hair dryer. The steady-state smoothness component reflects the trend characteristics of the trajectory at longer time scales, corresponding to trajectory changes where the difference between the time scale and the total duration of a single complete operation of the hair dryer is less than a preset second difference, i.e., close to the total duration of a single complete operation of the hair dryer.

[0108] For example, wavelet transform is used to decompose the trajectory smoothness index. For instance, a suitable wavelet basis function, such as the Daubechies wavelet, is selected. The trajectory smoothness index is used as input, and wavelet transform is used to obtain wavelet coefficients at different scales. Based on a preset first and second difference, the wavelet coefficients are divided into transient smoothness components and steady-state smoothness components. If the time scale corresponding to a wavelet coefficient is less than the total duration of a single complete operation of the hair dryer and the difference is not less than 0.1 s, it is classified as a transient smoothness component; if the difference between the time scale corresponding to a wavelet coefficient and the total duration of a single complete operation of the hair dryer is less than 0.01 s, it is classified as a steady-state smoothness component.

[0109] Step S362: Perform frequency domain feature decomposition on the speed fluctuation coefficient, extract the main fluctuation frequency components through Fourier transform, and obtain low-frequency fluctuation features and high-frequency fluctuation features. The low-frequency fluctuation features represent the speed change trend features with frequencies below the preset feature frequency threshold, and the high-frequency fluctuation features represent the speed oscillation features with frequencies above the preset feature frequency threshold. The preset feature frequency threshold is determined based on the typical speed change cycle of the blower during operation.

[0110] Frequency domain eigenvalue decomposition decomposes the velocity fluctuation coefficient along the frequency dimension, extracting characteristic components at different frequencies. Fourier transform can decompose the velocity fluctuation coefficient into sine and cosine components of different frequencies. Low-frequency fluctuation characteristics reflect the velocity's trend at lower frequencies, corresponding to velocity changes below a preset characteristic frequency threshold. High-frequency fluctuation characteristics reflect the velocity's oscillation at higher frequencies, corresponding to velocity changes above the preset characteristic frequency threshold. The preset characteristic frequency threshold is determined based on the typical velocity change period of the hair dryer during operation, thus dividing the frequency components of the velocity fluctuation into low-frequency and high-frequency parts.

[0111] For example, the velocity fluctuation coefficient is decomposed using Fourier transform. The velocity fluctuation coefficient is taken as input and its spectrum is obtained through Fourier transform. Based on a preset characteristic frequency threshold, the spectrum is divided into low-frequency and high-frequency components.

[0112] Step S363: Perform feature splicing on the transient smoothness component, steady-state smoothness component, low-frequency fluctuation feature and high-frequency fluctuation feature to obtain the basic features of position change.

[0113] Step S364: Input the basic features of position change and the features of grip posture change into a bidirectional long short-term memory network, learn the long temporal dependencies between features through a gating mechanism, and output the hidden state features of position change.

[0114] Bidirectional Long Short-Term Memory (Bi-LSTM) networks can process sequential data, simultaneously considering both forward and backward information. Gating mechanisms are a crucial structure in Bi-LSTM networks, controlling the inflow, outflow, and retention of information, enabling the network to learn long-term temporal dependencies in sequential data. By inputting basic features of positional changes and features of grip posture changes into the Bi-LSTM network, the network can learn the long-term temporal dependencies between these two features, outputting hidden state features of positional changes that contain the latent information of the positional changes.

[0115] A bidirectional Long Short-Term Memory (LSTM) network consists of forward LSTM units and backward LSTM units. At each time step, the forward LSTM unit updates the current hidden state based on the current input and the hidden state of the previous time step, while the backward LSTM unit updates the current hidden state based on the current input and the hidden state of the next time step. The forward and backward hidden states are concatenated to obtain the bidirectional hidden state at the current time step. After processing for multiple time steps, the network outputs position-change hidden state features. For example, the position-change hidden state features can be represented as a vector, where each element represents a certain feature of the hidden state at different time steps.

[0116] Step S365: Apply an attention mechanism to weight the hidden state features of position changes, highlighting the feature contributions of key time steps, to obtain weighted position features.

[0117] Attention mechanisms can assign different weights to each time step in the hidden state features of positional changes based on their importance. By weighting the hidden state features of positional changes using an attention mechanism, the feature contributions of key time steps can be highlighted, allowing important information to receive more attention and resulting in weighted positional features.

[0118] For example, the attention mechanism can employ a soft attention mechanism. First, the attention score for each time step in the positional change hidden state features is calculated. This can be achieved by mapping the positional change hidden state features to attention scores using a fully connected layer. Then, the attention scores are converted into attention weights using a softmax function. Finally, the positional change hidden state features are multiplied element-wise by the attention weights to obtain the weighted positional features. For instance, if the positional change hidden state features are a vector sequence of length T, and the attention weights are also vectors of length T, multiplying them element-wise yields the weighted positional features.

[0119] Step S366: Input the weighted position features into the fully connected network for feature mapping, and output the probability distribution of the air outlet position movement features through the normalized activation function. Take the feature corresponding to the highest probability of the probability distribution as the air outlet position movement feature.

[0120] A fully connected network can perform feature mapping on weighted positional features, transforming them into a probability distribution of vent position movement features. A normalized activation function, such as the softmax function, can convert the output of the fully connected network into a probability distribution. The feature corresponding to the highest probability in this distribution is taken as the vent position movement feature, representing the most likely scenario of vent position movement. For example, the weighted positional features are input into a fully connected network. The fully connected network contains multiple fully connected layers, each performing a linear transformation and non-linear activation on the input. The output of the final fully connected layer is processed by the softmax function to obtain the probability distribution of the vent position movement features.

[0121] Step S370: Combine the usage duration feature, grip posture change feature, and air outlet position movement feature into a user behavior pattern feature set.

[0122] The user behavior pattern feature set is a collection composed of usage duration features, grip posture change features, and air outlet position movement features, comprehensively describing the user's behavior patterns when using a hair dryer. For example, the usage duration features, grip posture change features, and air outlet position movement features can be arranged in a certain order to form a vector or matrix, which serves as the user behavior pattern feature set. For instance, if the usage duration feature is represented by a scalar, the grip posture change feature by a vector of length m, and the air outlet position movement feature by a vector of length n, then the user behavior pattern feature set can be represented as a vector of length 1 + m + n.

[0123] Step S400: Call the hair dryer's built-in self-learning module to jointly learn the hair state feature set and the user behavior pattern feature set, and output the hair state-behavior pattern association model.

[0124] The self-learning module is a built-in network component in the hair dryer, used to jointly learn from the hair state feature set and the user behavior pattern feature set, outputting a hair state-behavior pattern association model. This module adopts a hybrid architecture, combining deep neural networks (DNNs) and rule-based inference mechanisms to achieve efficient learning and accurate model output.

[0125] For example, the deep neural network part is mainly responsible for feature extraction, matching, and learning, and consists of the following key layers and modules: 1. Input Layer: Receives a joint feature vector and a historical joint feature set. The joint feature vector is obtained by aligning the hair state feature set and the user behavior pattern feature set according to their feature dimensions. The historical joint feature set is read from the hair dryer's local storage unit and contains a combination of the historical hair state feature set and the historical user behavior pattern feature set. 2. Feature Extraction Layer: This layer uses an autoencoder network to perform multi-level feature decomposition on the input joint feature vector and the historical joint feature set, breaking down the high-dimensional feature vector into a basic feature layer, a related feature layer, and an abstract feature layer. The autoencoder network consists of an encoder and a decoder. The encoder extracts feature information at different levels step by step through multiple hidden layers. For example, the first hidden layer can extract the statistical distribution features of the basic feature layer, such as mean and variance; the second hidden layer can explore the nonlinear relationships of the related feature layer; and the third hidden layer focuses on extracting high-order semantic features of the abstract feature layer. The decoder is used to reconstruct the input data to ensure that the encoder can accurately learn the important information of the features. 3. Similarity Calculation Layer: In this layer, for each layer's features, the similarity between the joint feature vector and the corresponding features in the historical joint feature set is calculated. For the basic feature layer, cosine similarity is used to measure similarity, resulting in a basic layer similarity sequence. For the associated feature layer, Euclidean distance is calculated and converted into an associated layer similarity sequence. For the abstract feature layer, KL divergence is calculated and processed through a series of steps (including difference compensation, exponential transformation, weighted adjustment, and normalization) to obtain an abstract layer similarity sequence. 4. Weighted Fusion Layer: This layer weights and fuses the similarity sequences from each layer to generate a comprehensive matching score. The weights of the weighted fusion are dynamically adjusted based on the influence of each layer's features on the model output. Gradient Boosting Decision Tree (GBDT) can be used to determine the importance score of each layer's features, thus dynamically allocating weights. For example, if the GBDT model training results show that the abstract feature layer has a greater impact on the model output, then a higher weight is assigned to the abstract layer's similarity sequences. 5. Decision Layer: The overall matching score is compared with a preset matching threshold. The preset matching threshold is optimized based on the historical model training results. When the overall matching score falls below the preset matching threshold, the joint feature vector is added to the incremental training sample pool.

[0126] The rule-based reasoning mechanism is mainly used to process explicit rules and logical judgments, and works in conjunction with the deep neural network to improve the reliability and interpretability of the model.

[0127] 1. Incremental Training Rules Module: This module manages the incremental training sample pool and defines the rules for incremental training. When a new joint feature vector is added to the incremental training sample pool, it determines when to start incremental training based on preset rules. For example, the incremental training process is triggered when the number of samples in the incremental training sample pool reaches a certain threshold, or after a certain time interval. 2. Parameter Optimization Rule Module: During the incremental training process, this module updates the weights of the model according to the adaptive learning rate adjustment mechanism and the gradient clipping strategy. The adaptive learning rate adjustment mechanism can adopt adaptive optimization algorithms such as Adagrad, Adadelta, or Adam, and dynamically adjust the learning rate according to the changes in the loss function. The gradient clipping strategy prevents gradient explosion by limiting the magnitude of the gradient, ensuring the stability of training.

[0128] For the model output and feedback part, it can include: 1. Model Output Layer: After being processed and learned through the above-mentioned layers, the optimized hair state-behavior pattern association model is finally output. This model can be used for the subsequent generation of personalized temperature control and wind speed adjustment strategies; 2. Feedback Mechanism: This module is responsible for collecting the usage feedback information of the model, such as the actual temperature control effect and the satisfaction of wind speed adjustment. According to this feedback information, the parameters and rules of the model are further optimized, forming a closed-loop learning and optimization process, continuously improving the performance and adaptability of the model. Through this hybrid architecture, the autonomous learning module can give full play to the advantages of deep neural networks in feature learning and pattern recognition, and at the same time combine the reliability and interpretability of the rule-based reasoning mechanism to accurately learn and model the association relationship between hair states and user behavior patterns;

[0129] As an implementation method, step S400 can be specifically implemented as the following steps S410~S470: Step S410: Align the feature dimensions of the hair state feature set and the user behavior pattern feature set to obtain a joint feature vector.

[0130] Feature dimension alignment is to adjust the feature dimensions of the hair state feature set and the user behavior pattern feature set to be the same, so that they can be effectively combined and analyzed. The joint feature vector is a vector formed by combining the hair state feature set and the user behavior pattern feature set after feature dimension alignment, containing the comprehensive information of hair states and user behavior patterns. Exemplarily, first analyze the feature dimensions of the hair state feature set and the user behavior pattern feature set. If the dimensions of the two are inconsistent, methods such as padding or cropping can be used for adjustment. For example, if the dimension of the hair state feature set is m and the dimension of the user behavior pattern feature set is n, and m < n, zero vectors can be filled at the end of the hair state feature set to make its dimension consistent with that of the user behavior pattern feature set. Then, the adjusted hair state feature set and user behavior pattern feature set are concatenated in a certain order to obtain the joint feature vector.

[0131] Step S420: Read the historical joint feature set from the hair dryer's local storage unit. The historical joint feature set contains a combination of historical hair state feature set and historical user behavior pattern feature set.

[0132] The historical joint feature set is a data set stored in the hair dryer's local storage unit. It records the combined features of hair state features and user behavior pattern features generated during past hair dryer use. By comparing and analyzing this combined feature set with the current joint feature vector, the correlation between hair state and user behavior patterns can be better explored. The local storage unit can be a storage device such as flash memory or hard drive inside the hair dryer, used for long-term storage of this historical data.

[0133] For example, when joint learning is required, the autonomous learning module sends a read request to the local storage unit. Upon receiving the request, the local storage unit searches for and reads the historical joint feature set according to a preset storage format and indexing method. For instance, if the historical joint feature set is stored in a database, the autonomous learning module will extract relevant data from the database using SQL queries. The retrieved historical joint feature set may include joint feature data from multiple different time points, each of which is composed of a combination of the historical hair state feature set and the historical user behavior pattern feature set at that time.

[0134] Step S430: Perform hierarchical feature matching on the joint feature vector and the historical joint feature set. The high-dimensional feature vector is decomposed into a basic feature layer, an associated feature layer and an abstract feature layer through multi-level feature decomposition. The feature similarity is calculated at each level.

[0135] Hierarchical feature matching is a method for comparing and matching joint feature vectors with historical joint feature sets. It involves decomposing high-dimensional feature vectors into multiple levels to obtain feature information at different levels, and then calculating feature similarity at each level. Multi-level feature decomposition is the process of breaking down the high-dimensional feature vectors in the joint feature vector and historical joint feature set into a basic feature layer, a related feature layer, and an abstract feature layer. The basic feature layer contains the statistical distribution characteristics of the original features, such as the mean and variance, describing the basic statistical properties of the features. The related feature layer contains the non-linear relationships between features, reflecting the interactions and dependencies between different features. The abstract feature layer contains higher-order semantic features, describing the meaning and attributes of the features at a more abstract level. By calculating feature similarity at each level, the similarity between the joint feature vector and the historical joint feature set can be evaluated more comprehensively and accurately.

[0136] As one implementation method, step S430 can be specifically implemented as the following steps S431~S436: Step S431: Perform multi-level feature decomposition on the joint feature vector, and map it to the basic feature layer, the associated feature layer and the abstract feature layer through an autoencoder network. The basic feature layer contains the statistical distribution features of the original features, the associated feature layer contains the non-linear association relationship between features, and the abstract feature layer contains high-order semantic features.

[0137] Autoencoder networks are neural network models used for feature extraction and dimensionality reduction, consisting of an encoder and a decoder. The encoder maps the joint feature vector of the input to a low-dimensional feature space, while the decoder reconstructs the original joint feature vector from the vectors in this low-dimensional feature space. By training the autoencoder network, the encoder can learn the important feature information of the joint feature vector. In the multi-level feature decomposition process, the autoencoder network can map the joint feature vector to basic feature layers, associated feature layers, and abstract feature layers.

[0138] Specifically, the encoder part of an autoencoder network can adopt a multilayer perceptron (MLP) structure, containing multiple hidden layers. Different levels of feature information can be extracted in different hidden layers. For example, in the first hidden layer, statistical distribution features of the basic feature layer, such as the mean and variance of the features, can be extracted. These statistical distribution features can be obtained by statistically calculating each feature in the joint feature vector. In the second hidden layer, nonlinear relationships between related feature layers can be extracted. Nonlinear activation functions (such as ReLU) can be used to transform the features and uncover the nonlinear relationships between them. In the third hidden layer, higher-order semantic features of the abstract feature layer can be extracted. These features may require learning through more complex neural network structures and training methods. The decoder part then reconstructs these different levels of feature information to make them as close as possible to the original joint feature vector. By continuously training the autoencoder network and adjusting its weights and parameters, the encoder can accurately extract feature information at different levels.

[0139] Step S432: Perform the same multi-level feature decomposition on each historical joint feature in the historical joint feature set to obtain the historical basic feature layer set, the historical related feature layer set, and the historical abstract feature layer set.

[0140] The process of performing multi-level feature decomposition on each historical joint feature in the historical joint feature set is the same as the process of performing multi-level feature decomposition on the joint feature vector. Using a pre-trained autoencoder network, each historical joint feature in the historical joint feature set is input into the encoder, and through the encoder's multiple hidden layers, it is mapped to the basic feature layer, the associated feature layer, and the abstract feature layer.

[0141] Since the historical joint feature set can include joint features from multiple different time points, performing this multi-level feature decomposition on each historical joint feature yields multiple basic feature layers, related feature layers, and abstract feature layers. Combining these basic feature layers corresponding to different historical joint features results in the historical basic feature layer set; combining the related feature layers results in the historical related feature layer set; and combining the abstract feature layers results in the historical abstract feature layer set. Each element in these sets corresponds to the feature information of a historical joint feature at its respective level.

[0142] Step S433: Calculate the cosine similarity between the base feature layer of the joint feature vector and each historical base feature layer in the historical base feature layer set to obtain the base layer similarity sequence.

[0143] Cosine similarity is a metric used to measure the similarity between two vectors. It assesses their similarity by calculating the cosine of the angle between the two vectors. The closer the cosine value is to 1, the more similar the two vectors are; the closer the cosine value is to 0, the less similar the two vectors are.

[0144] When calculating the cosine similarity between the base feature layer of the joint feature vector and each historical base feature layer in the historical base feature layer set, the base feature layer of the joint feature vector is treated as a vector, and each historical base feature layer in the historical base feature layer set is also treated as a vector. For each historical base feature layer vector in the historical base feature layer set, the cosine similarity formula is used for calculation. By performing this calculation on each historical base feature layer vector in the historical base feature layer set and the base feature layer vector of the joint feature vector, a series of cosine similarity values ​​are obtained. Arranging these values ​​in order yields the base layer similarity sequence.

[0145] Step S434: Calculate the Euclidean distance between the associated feature layer of the joint feature vector and each historical associated feature layer in the set of historical associated feature layers, and convert the Euclidean distance into a similarity sequence of associated layers.

[0146] Euclidean distance is a metric used to measure the distance between two vectors. It is obtained by calculating the square root of the sum of the squares of the differences between corresponding elements of the two vectors. The smaller the Euclidean distance, the closer the two vectors are. To convert Euclidean distance into a similarity metric, it needs to be transformed. When calculating the Euclidean distance between the associated feature layer of the joint feature vector and each historical associated feature layer in the historical associated feature layer set, the associated feature layer of the joint feature vector is treated as a vector, and each historical associated feature layer in the historical associated feature layer set is also treated as a vector. For each historical associated feature layer vector in the historical associated feature layer set, the Euclidean distance formula is used for calculation. After obtaining the Euclidean distance, it is converted into an association layer similarity. For example, the reciprocal of the Euclidean distance is taken, and then normalized so that its value is within the range of [0,1]. Let the Euclidean distance be d, and the converted association layer similarity be s, then s = 1 / (1+d). Each historical associated feature layer vector in the historical associated feature layer set is calculated and transformed with the associated feature layer vector of the joint feature vector to obtain a series of associated layer similarity values. These values ​​are then arranged in order to obtain the associated layer similarity sequence.

[0147] Step S435: Calculate the KL divergence between the abstract feature layer of the joint feature vector and each historical abstract feature layer in the historical abstract feature layer set, and convert the KL divergence into an abstract layer similarity sequence.

[0148] KL divergence (also known as relative entropy) is an index used to measure the degree of difference between two probability distributions. It is obtained by calculating the integral of the log-likelihood ratio of the two probability distributions. The larger the KL divergence, the greater the difference between the two probability distributions.

[0149] As one implementation method, step S435 converts the KL divergence into an abstract layer similarity sequence, which can be implemented as follows: Step S4351~S4355: Step S4351: Calculate the KL divergence between the abstract feature layer of the joint feature vector and each historical abstract feature layer in the historical abstract feature layer set to obtain the KL divergence sequence. The KL divergence is calculated by the log-likelihood ratio integral of the probability distribution of the abstract feature layer, which represents the degree of difference between the two abstract feature distributions.

[0150] First, consider the abstract feature layer of the joint feature vector and each historical abstract feature layer in the historical feature layer set as probability distributions. The elements in these feature layers can be normalized to satisfy the conditions of the probability distribution (the sum of the elements is 1 and each element is non-negative). Let the probability distribution of the abstract feature layer of the joint feature vector be P, and the probability distribution of a certain historical abstract feature layer in the historical feature layer set be Q. Then, their KL divergence is calculated as: KL(P||Q)=∑P(x)*log(P(x) / Q(x)), where x represents an element in the abstract feature layer. Performing this calculation on each historical abstract feature layer probability distribution in the historical feature layer set and the probability distribution of the abstract feature layer of the joint feature vector yields a series of KL divergence values. Arranging these values ​​in order gives the KL divergence sequence. This KL divergence sequence reflects the degree of difference between the abstract feature layer of the joint feature vector and each historical abstract feature layer in the historical feature layer set.

[0151] Step S4352: Extract the target shape feature from the hair state feature set, input the target shape feature into the difference compensation unit. The higher the shape complexity in the target shape feature, the larger the compensation coefficient output by the difference compensation unit. Multiply the KL divergence sequence with the compensation coefficient element by element to obtain the compensated KL divergence sequence.

[0152] The target styling feature describes the user's desired hairstyle. The difference compensation unit is a unit that outputs a compensation coefficient based on the target styling feature. The compensation coefficient can vary in size depending on the styling complexity of the target styling feature. Higher styling complexity indicates a more difficult hairstyle to create, requiring more attention and adjustment; therefore, the difference compensation unit will output a larger compensation coefficient.

[0153] For example, target styling features are extracted from a set of hair state features. These target styling features can include information from multiple aspects, such as the complexity of the hairstyle and whether styling products are needed. The target styling features are input into a difference compensation unit, which maps them to compensation coefficients according to a preset mapping relationship. For example, a linear or non-linear mapping function can be used to map the styling complexity index in the target styling features to the compensation coefficients. Then, the obtained compensation coefficients are multiplied element-wise with each element in the KL divergence sequence to obtain the compensated KL divergence sequence. The purpose of this is to consider both the differences at the abstract feature layer and the influence of the target styling features, thus appropriately adjusting the KL divergence.

[0154] Step S4353: Perform an exponential transformation on the compensated KL divergence sequence, and map the divergence value to the initial similarity value through a negative exponential function. The smaller the compensated KL divergence value, the closer the initial similarity value is to 1.

[0155] Exponential transformation is a method to convert divergence values ​​in a compensated KL divergence sequence into similarity values. Using a negative exponential function, the compensated KL divergence values ​​are used as input to obtain initial similarity values. The negative exponential function has the form f(x) = exp(-x), where x is the compensated KL divergence value. Due to the properties of the negative exponential function, the smaller the compensated KL divergence value, the closer the negative exponential function value is to 1; the larger the compensated KL divergence value, the closer the negative exponential function value is to 0. This exponential transformation is performed on each element in the compensated KL divergence sequence to obtain a series of initial similarity values. Arranging these values ​​in order yields the initial similarity value sequence.

[0156] Step S4354: Extract grip posture change features from the user behavior pattern feature set, input the grip posture change features into the weight unit. The higher the stability of the grip posture change features, the larger the weight value output by the weight unit. Add the initial similarity value and the weight value element by element to obtain the weighted similarity value.

[0157] The grip posture change feature is one of the features in the user behavior pattern feature set, describing how the user's grip posture changes while using a hair dryer. The weight unit can output different weight values ​​based on the stability of the grip posture change feature. Higher stability in the grip posture change feature indicates a more stable user grip posture, a greater impact on the similarity of the abstract feature layer, and therefore a larger weight value output by the weight unit.

[0158] For example, grip posture change features are extracted from a set of user behavior pattern features. These features can include various aspects, such as the frequency and magnitude of grip posture changes. The grip posture change features are input into a weighting unit, which maps them to weight values ​​according to a predefined mapping relationship. For example, a linear or non-linear mapping function can be used to map the stability index of the grip posture change features to weight values. Then, each element in the initial similarity value sequence is summed element-wise with its corresponding weight value to obtain a weighted similarity value sequence. The purpose of this is to further consider the influence of grip posture change features on the abstract feature layer similarity, and to adjust the initial similarity values ​​accordingly.

[0159] Step S4355: Perform maximum-min normalization on the weighted similarity values ​​to ensure they fall within the preset correlation coefficient range, and output the hair quality correlation coefficient.

[0160] Step S436: Use the similarity sequences of the basic layer, the similarity sequences of the associated layer, and the similarity sequences of the abstract layer as the feature similarity of each layer.

[0161] The previously calculated basic layer similarity sequences, association layer similarity sequences, and abstract layer similarity sequences are combined to form the feature similarities for each layer. These feature similarities reflect the degree of similarity between the joint feature vector and the historical joint feature set at different levels, providing a basis for subsequent comprehensive matching degree calculations.

[0162] Step S440: Weighted fusion of feature similarity at each level to generate a comprehensive matching index. The weights of the weighted fusion are dynamically adjusted according to the influence of each level of features on the model output.

[0163] Weighted fusion is a method that comprehensively calculates the similarity sequences of the base layer, the association layer, and the abstract layer. By assigning different weights to the similarity sequences of each layer and then summing them in a weighted manner, a comprehensive matching index is obtained. The weights of the weighted fusion are dynamically adjusted according to the influence of each layer's features on the model output. This means that under different circumstances, the importance of features at different layers to the model output may vary, thus requiring dynamic adjustment of the weights based on the actual situation.

[0164] For example, machine learning methods can be used to determine the influence of features at each level on the model output. For instance, a Gradient Boosting Decision Tree (GBDT) model can be used, taking the joint feature vector and historical joint feature set as input, and the model's output as the target. By training the GBDT model, feature importance scores for each level are obtained. Based on these feature importance scores, the weights of the similarity sequences at each level are dynamically adjusted. By dynamically adjusting the weights, the overall matching index can more accurately reflect the overall similarity between the joint feature vector and the historical joint feature set.

[0165] Step S450: When the overall matching degree index is lower than the preset matching threshold, the joint feature vector is added to the incremental training sample pool of the autonomous learning module. The preset matching threshold is optimized based on the historical model training effect.

[0166] The preset matching threshold is a pre-defined threshold used to determine whether the similarity between the joint feature vector and the historical joint feature set is sufficiently high. When the overall matching score is lower than the preset matching threshold, it indicates that the similarity between the joint feature vector and the historical joint feature set is low, which may represent a new combination of hair state and user behavior pattern. In this case, the joint feature vector is added to the incremental training sample pool of the autonomous learning module so that the model can be incrementally trained using this new data, thereby improving the model's adaptability and accuracy.

[0167] The preset matching threshold is optimized based on the historical model training performance. During model training, the training performance, such as accuracy and recall, is continuously recorded. The preset matching threshold is dynamically adjusted based on changes in these metrics. If the model's training performance is unsatisfactory, the preset matching threshold may need to be lowered to allow more new data to enter the incremental training sample pool; conversely, if the model's training performance is good, the preset matching threshold can be appropriately increased to reduce unnecessary new data entering the incremental training sample pool.

[0168] Step S460: Optimize the parameters of the basic association model in the autonomous learning module using the incremental training sample pool, and update the model weights through an adaptive learning rate adjustment mechanism and gradient pruning strategy to obtain the optimized hair state-behavior pattern association model.

[0169] The incremental training sample pool contains combined information on new hair states and user behavior patterns. Using this data to optimize the parameters of the basic association model in the autonomous learning module allows the model to better adapt to new situations. The adaptive learning rate adjustment mechanism is a mechanism that dynamically adjusts the learning rate based on changes in the loss function during training. The learning rate is the step size used to update weights during model training; an appropriate learning rate allows the model to converge to the optimal solution more quickly. Gradient pruning strategies are used to prevent gradient explosion by limiting the magnitude of gradients, ensuring the stability of model training.

[0170] For example, data from the incremental training sample pool is input into the basic association model for training. During training, an adaptive learning rate adjustment mechanism is used to dynamically adjust the learning rate based on changes in the loss function. For instance, adaptive optimization algorithms such as Adagrad, Adadelta, or Adam can be used, which automatically adjust the learning rate based on the gradient history of each parameter. Simultaneously, a gradient pruning strategy is used; when the norm of the gradient exceeds a preset threshold, the gradient is pruned to ensure its norm does not exceed the threshold. Through continuous iterative training and updating of the model's weights, the model's performance is maintained until it reaches a satisfactory level, resulting in an optimized hair state-behavior pattern association model.

[0171] Step S470: Use the optimized hair state-behavior pattern association model as the output hair state-behavior pattern association model.

[0172] After incremental training and parameter optimization, the resulting optimized hair state-behavior pattern association model better reflects the relationship between hair state and user behavior patterns. This optimized model is used as the final output hair state-behavior pattern association model for the subsequent generation of personalized temperature control and wind speed adjustment strategies.

[0173] Step S500: Generate personalized temperature control and wind speed adjustment strategies based on the hair state-behavior pattern association model.

[0174] The hair condition-behavior pattern association model contains the correlation information between hair condition and user behavior patterns. Based on this model, personalized temperature control and wind speed adjustment strategies can be generated to meet the needs and hair conditions of different users.

[0175] As one implementation method, step S500 generates a personalized temperature control and wind speed adjustment strategy based on the hair state-behavior pattern association model. Specifically, it can be implemented as follows: Step S510~S560: Step S510: Input the current hair state feature set and the current user behavior pattern feature set into the hair state-behavior pattern association model, and calculate the initial temperature control parameter matrix and the initial wind speed parameter matrix through the forward propagation of the model. The initial temperature control parameter matrix contains temperature control parameters for different hair areas, and the initial wind speed parameter matrix contains wind speed control parameters for different blowing periods.

[0176] The current hair state feature set and the current user behavior pattern feature set represent the feature information of the hair state and user behavior pattern corresponding to the current use of the hair dryer. These are input into the hair state-behavior pattern association model, which performs forward propagation calculations based on its internal parameters and structure. Forward propagation is the process of passing input data from the model's input layer through hidden layers to the output layer. During this process, the model performs a series of transformations and calculations on the input data, ultimately outputting the initial temperature control parameter matrix and the initial wind speed parameter matrix.

[0177] The initial temperature control parameter matrix is ​​a matrix containing temperature control parameters for different hair areas. Different temperature control parameters can be set for different parts of the hair (such as the crown, sideburns, and back of the head) to meet the needs of different areas. The initial airflow parameter matrix is ​​a matrix containing airflow control parameters for different stages of blow-drying. Different airflow control parameters can be set for different stages of blow-drying (such as the beginning, middle, and end stages) to achieve more precise airflow adjustment.

[0178] Step S520: Perform spatial correlation analysis on the initial temperature control parameter matrix, calculate the spatial correlation of temperature parameters in adjacent hair regions using the Gaussian kernel function, and smooth the initial temperature control parameter matrix based on the spatial correlation to obtain spatially smoothed temperature control parameters.

[0179] Spatial correlation analysis is the process of analyzing the spatial relationships between temperature parameters in different hair regions within an initial temperature control parameter matrix. The Gaussian kernel function is a commonly used kernel function that can be used to calculate the spatial correlation degree between temperature parameters in adjacent hair regions. Spatial correlation degree reflects the similarity and mutual influence between temperature parameters in adjacent hair regions.

[0180] For example, a Gaussian kernel function is used to process the initial temperature control parameter matrix. Let the temperature parameter of a certain hair region in the initial temperature control parameter matrix be x, and the temperature parameter of its adjacent hair regions be y. The Gaussian kernel function has the form K(x,y)=exp(-||xy|| 2 / (2*σ 2 )), where σ is the bandwidth parameter of the Gaussian kernel function. The spatial correlation between adjacent hair regions is obtained by calculating the Gaussian kernel function values ​​of their temperature parameters.

[0181] Based on the calculated spatial correlation, the initial temperature control parameter matrix is ​​smoothed. A weighted average method can be used, assigning different weights to the temperature parameters of adjacent hair regions according to their spatial correlation, and then summing these weighted averages to obtain the smoothed temperature parameters. This smoothing process is applied to each element of the initial temperature control parameter matrix to obtain spatially smoothed temperature control parameters. These spatially smoothed temperature control parameters take into account the spatial correlation between temperature parameters of adjacent hair regions, avoiding drastic fluctuations in temperature parameters and resulting in more stable and reasonable temperature control.

[0182] Step S530: Perform time-series stability analysis on the initial wind speed parameter matrix, and eliminate short-term fluctuation interference by moving average filtering to obtain time-series stable wind speed parameters.

[0183] Temporal stability analysis is the process of analyzing the time series stability of wind speed control parameters in the initial wind speed parameter matrix during different wind periods. Moving average filtering is a commonly used filtering method that eliminates short-term fluctuations by averaging the wind speed parameters within a fixed window size across the time series.

[0184] By applying this moving average filtering process to the wind speed parameters at all time steps in the initial wind speed parameter matrix, time-stable wind speed parameters are obtained. These time-stable wind speed parameters eliminate short-term fluctuations, making wind speed regulation more stable and reliable.

[0185] Step S540: Adapt the spatial smoothing temperature control parameters to the hair quality features in the hair state feature set, adjust the range of temperature parameters through the feature mapping function, and generate the adapted temperature control parameters.

[0186] Hair texture is an important feature in the set of hair condition characteristics, describing the texture and properties of the hair. Different hair textures have different tolerances and requirements for temperature. Therefore, it is necessary to adapt spatial smoothing temperature control parameters to hair texture characteristics, and adjust the range of temperature parameters through feature mapping functions to meet the needs of different hair textures.

[0187] As one implementation method, step S540 can be specifically implemented as the following steps S541~S546: Step S541: Input the hair quality characteristics into the temperature control coefficient generation network, and calculate the upper temperature adjustment coefficient and the lower temperature adjustment coefficient through a multilayer perceptron. The input layer of the temperature control coefficient generation network contains the quantized representation vector of the hair quality characteristics, the hidden layer performs nonlinear transformation through a nonlinear activation function, and the output layer maps the coefficients to a preset range through a range mapping function.

[0188] The temperature control coefficient generation network is used to generate upper and lower temperature adjustment coefficients based on hair quality characteristics. It employs a multilayer perceptron (MLP) structure, comprising an input layer, hidden layers, and an output layer. The input layer receives the quantized representation vector of hair quality characteristics, which is a vector obtained after quantizing the hair quality characteristics and contains key information about them. The hidden layer performs a nonlinear transformation on the input quantized representation vector of hair quality characteristics using a nonlinear activation function (such as ReLU) to extract latent information from the hair quality characteristics. The output layer maps the calculated coefficients to a preset range using a range mapping function. For example, the preset range for the upper and lower temperature adjustment coefficients can be [0,1]. Through this network structure and calculation process, the upper and lower temperature adjustment coefficients are obtained and used for subsequent adjustment of the spatially smoothed temperature control parameters.

[0189] Step S542: Divide the spatial smoothing temperature control parameters into matrix blocks, and divide the hair region into multiple sub-region temperature control parameter matrices according to the spatial distribution of the hair region.

[0190] Based on the spatial distribution of the hair region, the spatial smoothing temperature control parameter matrix is ​​divided into multiple sub-region temperature control parameter matrices. For example, the hair can be divided into different regions such as the top of the head, sideburns, and back of the head, with each region corresponding to a sub-region temperature control parameter matrix. This matrix segmentation allows for more precise and targeted adjustments to subsequent temperature parameters. For example, the boundaries of each sub-region are determined based on the spatial extent and division rules of the hair region. Elements belonging to the same sub-region in the spatial smoothing temperature control parameter matrix are extracted to form a sub-region temperature control parameter matrix.

[0191] Step S543: For each sub-region temperature control parameter matrix, multiply the matrix elements with the temperature upper limit adjustment coefficient to obtain the sub-region temperature upper limit.

[0192] For each sub-region temperature control parameter matrix, multiply each element by the upper temperature limit adjustment coefficient. Let t be an element in the sub-region temperature control parameter matrix, and k be the upper temperature limit adjustment coefficient. upper Then the upper limit of the sub-region temperature is t*k. upperThrough this multiplication operation, the upper temperature limit of each sub-region is adjusted according to the hair quality characteristics, making it more suitable for the actual needs of the hair in that region.

[0193] Step S544: Multiply the matrix elements of the temperature control parameter matrix of each sub-region with the lower limit adjustment coefficient to obtain the lower limit of the sub-region temperature.

[0194] Similar to calculating the upper limit of temperature for a sub-region, for each sub-region temperature control parameter matrix, each element is multiplied by the lower limit adjustment coefficient. Let t be an element in the sub-region temperature control parameter matrix, and k be the lower limit adjustment coefficient. lower Then the lower limit of the sub-region temperature is t*k lower Through this multiplication operation, the lower limit of the temperature for each sub-region is adjusted according to the hair quality characteristics, making it more suitable for the actual needs of the hair in that region.

[0195] Step S545: Limit the elements in the temperature control parameter matrix of each sub-region, replace the elements that are higher than the upper limit of the sub-region temperature with the upper limit of the sub-region temperature, and replace the elements that are lower than the lower limit of the sub-region temperature with the lower limit of the sub-region temperature.

[0196] Limiting is the process of restricting the range of elements in the temperature control parameter matrix for each sub-region. If an element is higher than the upper temperature limit of the sub-region, it is replaced with the upper temperature limit; if an element is lower than the lower temperature limit, it is replaced with the lower temperature limit. This ensures that the temperature parameters of each sub-region are within a reasonable range, avoiding damage to the hair caused by excessively high or low temperatures. For example, each element in the temperature control parameter matrix of each sub-region is iterated through, and comparison and replacement operations are performed. For instance, let an element in the sub-region temperature control parameter matrix be t, and the upper temperature limit of the sub-region be t_t. upper The lower limit of the sub-region temperature is t lower If t>t upper Then replace t with t upper If t <t lower Then replace t with t lower .

[0197] Step S546: Reorganize the temperature control parameter matrix of all sub-regions after the amplitude limiting process into a complete set of adapted temperature control parameters.

[0198] The temperature control parameter matrices of all sub-regions, after amplitude limiting, are recombined according to their original spatial distribution order to obtain the complete adapted temperature control parameters. These adapted temperature control parameters take into account the influence of hair quality characteristics on temperature parameters and ensure the rationality and stability of the temperature parameters.

[0199] Step S550: Adapt the time-series stable wind speed parameters to the usage duration features in the user behavior pattern feature set, adjust the rate of change of the wind speed parameters through a time decay function, and generate adapted wind speed parameters.

[0200] Usage duration reflects the approximate duration of a user's use of the hair dryer. The time decay function is a function that adjusts the airflow parameters based on changes over time, allowing the airflow parameters to gradually change with usage time, avoiding sudden changes in airflow that could cause discomfort to the hair.

[0201] For example, the time-stability wind speed parameter is correlated with usage duration characteristics. Based on the usage duration characteristics, the parameters of the time decay function are determined. For instance, the time decay function can take the form of an exponential decay function: v(t) = v0 * exp(-α * t), where v(t) is the wind speed parameter at time t, v0 is the initial wind speed parameter, α is the decay coefficient, and t is the usage duration. The decay coefficient α is adjusted according to the usage duration characteristics to make the rate of change of the wind speed parameter more in line with the user's needs. The time-stability wind speed parameter is adjusted using the time decay function to obtain the adapted wind speed parameter. This adapted wind speed parameter takes into account the impact of user usage duration on wind speed, making wind speed adjustment more user-friendly and comfortable.

[0202] Step S560: Jointly optimize the adapted temperature control parameters and adapted wind speed parameters, and use a multi-objective optimization algorithm to balance the temperature control accuracy and wind speed adjustment smoothness to generate personalized temperature control and wind speed adjustment strategies.

[0203] Multi-objective optimization algorithms are used to handle optimization of multiple conflicting objectives. In this invention, it is necessary to balance two objectives: temperature control accuracy and wind speed regulation smoothness. Temperature control accuracy requires that the temperature control parameters accurately meet the needs of the hair, avoiding damage caused by excessively high or low temperatures; wind speed regulation smoothness requires that the wind speed parameters change smoothly, avoiding discomfort to the hair caused by sudden changes in wind speed.

[0204] For example, a multi-objective optimization algorithm such as NSGA-II (Non-dominated Sorting Genetic Algorithm II) can be used. The adapted temperature control parameters and adapted wind speed parameters are taken as inputs, and temperature control accuracy and wind speed adjustment smoothness are taken as objective functions. Through iterative calculations of the multi-objective optimization algorithm, the optimal temperature control parameters and wind speed parameters are found, so that both objectives are well satisfied.

[0205] The final optimal combination of temperature control and wind speed parameters is the generated personalized temperature control and wind speed adjustment strategy. This strategy can achieve precise temperature control and comfortable wind speed adjustment based on the user's hair condition and behavior patterns, providing the user with a better user experience.

[0206] See also Figure 2 , Figure 2 This is a schematic diagram of a hair dryer temperature control device according to an embodiment of the present invention. The hair dryer temperature control device includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the hair dryer temperature control device, capable of parsing various instructions and processing various data within the device. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the hair dryer temperature control device. The memory 103 is a storage device within the hair dryer temperature control device, used to store programs and data. It is understood that the memory 103 here can include the built-in memory of the hair dryer temperature control device, or it can include extended memory supported by the device. The memory 103 provides storage space for storing the operating system of the hair dryer temperature control device, which is not limited in this invention.

[0207] In one embodiment, the processor 101 executes the hair dryer temperature control method based on smart sensors provided in the above embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A method for temperature control of a hair dryer based on a smart sensor, characterized in that, The method includes: acquiring a multi-dimensional sensor data set during the operation of the hair dryer; performing hair state feature analysis on the multi-dimensional sensor data set to obtain a hair state feature set; performing user behavior pattern analysis based on the hair state feature set to generate a user behavior pattern feature set; calling the hair dryer's built-in self-learning module to jointly learn the hair state feature set and the user behavior pattern feature set to output a hair state-behavior pattern association model; and generating personalized temperature control and wind speed adjustment strategies based on the hair state-behavior pattern association model.

2. The method as described in claim 1, characterized in that, The step of analyzing hair state features from the multidimensional sensor data set to obtain a hair state feature set includes: spatiotemporal alignment of the multidimensional sensor data set, integrating data collected by different types of sensors into a multidimensional data matrix under a unified spatiotemporal framework through timestamp synchronization and spatial coordinate mapping; multimodal feature fusion of the multidimensional data matrix, employing an adaptive weight allocation mechanism to weight temperature-related data, humidity-related data, and airflow sensing data to generate multimodal fusion features, wherein the adaptive weight allocation mechanism adjusts the weight ratio according to the signal-to-noise ratio of each modality; and extracting temperature and humidity response features and airflow disturbance distribution features from the multimodal fusion features. The temperature and humidity response features characterize the coupling change law of hair surface temperature and humidity with airflow, and the airflow disturbance... The distribution features characterize the spatial distribution of turbulence intensity when airflow passes through the hair region; the temperature and humidity response features and the airflow disturbance distribution features are input into the hair quality feature recognition submodule, and cross-modal feature association is performed through a multi-scale feature interaction network to output the hair quality features; based on the spatial distribution features of the multi-dimensional data matrix and the airflow disturbance distribution features, the hair coverage area contour features are generated through a region density clustering algorithm, and the hair coverage area contour features are convolved with the airflow disturbance distribution features to obtain the hair volume features; according to the time-varying gradient of the temperature and humidity response features, combined with the hair quality features and the hair volume features, the target styling features are determined through the styling feature inference submodule; the hair quality features, the hair volume features, and the target styling features are combined into a hair state feature set.

3. The method as described in claim 2, characterized in that, The extraction of temperature and humidity response features and airflow disturbance distribution features from the multimodal fusion features includes: performing joint time-domain analysis on temperature-related data and humidity-related data in the multimodal fusion features, extracting continuous time segments through a sliding window, calculating the mutual information value of temperature and humidity within each time segment to obtain a temperature and humidity mutual information sequence; performing differential processing on the temperature and humidity mutual information sequence to obtain the temperature and humidity mutual information change rate, and using the temperature and humidity mutual information change rate as the core component of the temperature and humidity response features; performing spatial spectrum estimation on the airflow sensing data in the multimodal fusion features, calculating the airflow energy distribution in different spatial orientations through a beamforming algorithm to obtain airflow spatial spectrum features; performing threshold segmentation on the airflow spatial spectrum features, retaining spatial components with energy higher than a preset energy percentage, and combining them into airflow disturbance distribution features; and concatenating the core component of the temperature and humidity response features with the airflow disturbance distribution features by feature dimension to obtain complete temperature and humidity response features and airflow disturbance distribution features.

4. The method as described in claim 1, characterized in that, The step of performing user behavior pattern analysis based on the hair state feature set to generate a user behavior pattern feature set includes: extracting position sensing data and timestamp data from the multi-dimensional sensor data set to construct the spatiotemporal motion trajectory of the hair dryer during operation; segmenting the spatiotemporal motion trajectory, dividing it into multiple motion trajectory segments based on the trajectory direction change rate and velocity continuity to obtain a trajectory segment set; performing correlation analysis between the trajectory segment set and hair quality features in the hair state feature set, and calculating the kinematic feature parameters of each motion trajectory segment, including trajectory segment curvature, average motion speed, and directional rotation angle; and so on. The trajectory smoothness index and velocity fluctuation coefficient are calculated based on the kinematic feature parameters. Combined with the hair volume feature in the hair state feature set, a usage duration feature is generated. The kinematic feature parameters and the usage duration feature are fused temporally. The evolution of the grip posture over time is learned through a gated recurrent unit network to generate a grip posture change feature. Based on the trajectory smoothness index and the velocity fluctuation coefficient, combined with the grip posture change feature, the air outlet position movement feature is determined through a position change inference submodule. The usage duration feature, the grip posture change feature, and the air outlet position movement feature are combined into a user behavior pattern feature set.

5. The method as described in claim 4, characterized in that, The step of performing correlation analysis between the trajectory segment set and the hair quality features in the hair state feature set, and calculating the kinematic feature parameters of each motion trajectory segment, includes: performing equal-interval coordinate interpolation on each motion trajectory segment in the trajectory segment set to obtain a high-density sampled trajectory coordinate sequence; calculating the spatial distance between adjacent sampling points in the trajectory coordinate sequence to generate a distance sequence, and accumulating the distance sequences to obtain the trajectory segment length; performing a ratio operation between the trajectory segment length and the time interval corresponding to the motion trajectory segment to obtain a preliminary average speed; performing curve fitting on the trajectory coordinate sequence, calculating the second derivative of the fitted curve to obtain the trajectory segment curvature distribution, and taking the mean of the curvature distribution as the trajectory segment curvature; inputting the hair quality features into the feature mapping submodule to obtain a hair quality correlation coefficient, which decreases linearly with the increase of the fineness of the hair quality features; multiplying the preliminary average speed with the hair quality correlation coefficient to obtain an adjusted average motion speed; and combining the trajectory segment curvature, the adjusted average motion speed, and the directional angle into kinematic feature parameters.

6. The method as described in claim 1, characterized in that, The process involves using the hair dryer's built-in self-learning module to jointly learn the hair state feature set and the user behavior pattern feature set, outputting a hair state-behavior pattern association model. This includes: aligning the feature dimensions of the hair state feature set and the user behavior pattern feature set to obtain a joint feature vector; reading a historical joint feature set from the hair dryer's local storage unit, the historical joint feature set containing combined features of the historical hair state feature set and the historical user behavior pattern feature set; and performing hierarchical feature matching on the joint feature vector and the historical joint feature set, decomposing the high-dimensional feature vector into a basic feature layer, an associated feature layer, and an abstract feature layer through multi-level feature decomposition, calculating features at each level. Similarity; the feature similarities at each level are weighted and fused to generate a comprehensive matching index. The weights of the weighted fusion are dynamically adjusted according to the influence of each level of features on the model output. When the comprehensive matching index is lower than a preset matching threshold, the joint feature vector is added to the incremental training sample pool of the autonomous learning module. The preset matching threshold is optimized based on the historical model training effect. The parameters of the basic association model in the autonomous learning module are optimized using the incremental training sample pool. The model weights are updated through an adaptive learning rate adjustment mechanism and a gradient pruning strategy to obtain an optimized hair state-behavior pattern association model. The optimized hair state-behavior pattern association model is used as the output hair state-behavior pattern association model.

7. The method as described in claim 6, characterized in that, The hierarchical feature matching of the joint feature vector and the historical joint feature set involves decomposing the high-dimensional feature vector into a basic feature layer, an associated feature layer, and an abstract feature layer through multi-level feature decomposition, and calculating feature similarity at each level. This includes: performing multi-level feature decomposition on the joint feature vector and mapping it to the basic feature layer, associated feature layer, and abstract feature layer using an autoencoder network. The basic feature layer contains the statistical distribution features of the original features, the associated feature layer contains the non-linear relationships between features, and the abstract feature layer contains high-order semantic features; and performing the same multi-level feature decomposition on each historical joint feature in the historical joint feature set to obtain a historical basic feature layer set and a historical associated feature layer set. The system comprises a feature layer set and a historical abstract feature layer set; calculating the cosine similarity between the basic feature layer of the joint feature vector and each historical basic feature layer in the historical basic feature layer set to obtain a basic layer similarity sequence; calculating the Euclidean distance between the associated feature layer of the joint feature vector and each historical associated feature layer in the historical associated feature layer set, and converting the Euclidean distance into an associated layer similarity sequence; calculating the KL divergence between the abstract feature layer of the joint feature vector and each historical abstract feature layer in the historical abstract feature layer set, and converting the KL divergence into an abstract layer similarity sequence; and using the basic layer similarity sequence, the associated layer similarity sequence, and the abstract layer similarity sequence as the feature similarity of each level.

8. The method as described in claim 1, characterized in that, The step of generating personalized temperature control and wind speed adjustment strategies based on the hair state-behavior pattern association model includes: inputting the current hair state feature set and the current user behavior pattern feature set into the hair state-behavior pattern association model; calculating an initial temperature control parameter matrix and an initial wind speed parameter matrix through forward propagation of the model; the initial temperature control parameter matrix containing temperature control parameters for different hair regions; and the initial wind speed parameter matrix containing wind speed control parameters for different blowing periods; performing spatial correlation analysis on the initial temperature control parameter matrix; calculating the spatial correlation degree of temperature parameters in adjacent hair regions using a Gaussian kernel function; and smoothing the initial temperature control parameter matrix based on the spatial correlation degree to obtain a spatial correlation coefficient. The process involves: 1) Smoothing the spatial temperature control parameters; 2) Performing a time-series stability analysis on the initial wind speed parameter matrix to obtain time-stable wind speed parameters; 3) Adapting the spatially smoothed temperature control parameters to the hair quality features in the hair state feature set, adjusting the temperature parameter range using a feature mapping function to generate adapted temperature control parameters; 4) Adapting the time-stable wind speed parameters to the usage duration features in the user behavior pattern feature set, adjusting the wind speed parameter change rate using a time decay function to generate adapted wind speed parameters; 5) Jointly optimizing the adapted temperature control parameters and the adapted wind speed parameters, balancing temperature control accuracy and wind speed adjustment smoothness using a multi-objective optimization algorithm to generate personalized temperature control and wind speed adjustment strategies.

9. The method as described in claim 8, characterized in that, The process of adapting the spatially smoothed temperature control parameters to the hair texture features in the hair state feature set, and adjusting the range of temperature parameters through a feature mapping function to generate adapted temperature control parameters includes: inputting the hair texture features into a temperature control coefficient generation network, calculating the upper temperature limit adjustment coefficient and the lower temperature limit adjustment coefficient through a multilayer perceptron, wherein the input layer of the temperature control coefficient generation network contains a quantized representation vector of the hair texture features, the hidden layer undergoes nonlinear transformation through a nonlinear activation function, and the output layer maps the coefficients to a preset range through a range mapping function; and performing matrix block partitioning on the spatially smoothed temperature control parameters according to the spatial distribution of the hair region. The system is divided into multiple sub-region temperature control parameter matrices. For each sub-region temperature control parameter matrix, the matrix elements are multiplied by the upper temperature limit adjustment coefficient to obtain the upper temperature limit of the sub-region. The matrix elements of each sub-region temperature control parameter matrix are multiplied by the lower temperature limit adjustment coefficient to obtain the lower temperature limit of the sub-region. The elements in each sub-region temperature control parameter matrix are subjected to amplitude limiting processing, replacing elements higher than the upper temperature limit of the sub-region with the upper temperature limit of the sub-region, and replacing elements lower than the lower temperature limit of the sub-region with the lower temperature limit of the sub-region. All the amplitude-limited sub-region temperature control parameter matrices are reorganized into complete adapted temperature control parameters.

10. A temperature control device for a hair dryer, characterized in that, include: A memory storing a computer program; a processor for loading the computer program to implement the hair dryer temperature control method based on a smart sensor as described in any one of claims 1-9.

Images