Multi-parameter adaptive adjustment method and system for special activated carbon purification device of maternity room

Abstract

The application relates to the air purification technical field, in particular to a multi-parameter self-adaptive adjusting method and system of an active carbon purification device special for a mother-infant room, multi-modal data are collected, image and audio features are extracted, data frames are generated and periodically reported to the cloud; the cloud identifies human activity modes through a composite neural network and sets a purification constraint condition library; for a multi-layer active carbon filter core with adjustable layers, a purification efficiency model is constructed, and the optimal combination strategy of fan rotating speed and filter screen stacking layer number is obtained by intelligent optimization in combination with constraints such as noise, particulate matter concentration and ventilation volume; the filter screen is adjusted by a stepping motor and the fan is driven in linkage; the cloud performs correlation analysis based on execution results and control strategies, online identifies and calibrates model parameters, and optimizes subsequent control effects in a closed loop. The method realizes the balance of low noise, low particulate matter and appropriate ventilation, responds to sensitive needs of a mother-infant scene, and improves purification efficiency and comfort.

Description

Technical Field

[0001] This application relates to the field of air purification technology, specifically to a multi-parameter adaptive adjustment method and system for activated carbon purification devices specifically designed for mother and baby rooms. Background Technology

[0002] Air purifiers are commonly used devices to improve indoor air quality, especially for the living environments of sensitive groups such as mothers and infants, where maintaining clean and comfortable air is crucial.

[0003] Most existing smart air purifiers passively respond to single or limited environmental parameters (such as particulate matter concentration), and their control logic is usually preset and fixed. However, in the special environment of a mother and baby room, the actual needs are complex and dynamically changing. For example, in a scenario where the baby is sleeping peacefully, extremely low operating noise is required to ensure sleep quality, while maintaining a certain level of air purification efficiency to protect respiratory health; while when multiple people are present, it may be necessary to prioritize efficient particulate matter removal. Due to their fixed physical structure and control strategies, traditional purification devices often can only make a single choice between "powerful purification" (high speed, high noise) and "quiet operation" (low speed, low purification efficiency).

[0004] Therefore, existing technologies cannot dynamically adjust the core performance characteristics of the purification device itself according to real-time specific scenario requirements (such as personnel activity patterns), thereby finding a purification control method with a globally optimal collaborative operation strategy under multiple changing constraints (such as noise limits, purification speed, and ventilation volume).

[0005] In view of this, this application proposes a multi-parameter adaptive adjustment method and system for a special activated carbon purification device for mother and baby rooms. Summary of the Invention

[0006] To achieve the above objectives, this application provides a multi-parameter adaptive adjustment method and system for an activated carbon purification device specifically designed for mother and baby rooms. The specific technical solution is as follows:

[0007] The multi-parameter adaptive adjustment method for activated carbon purification devices specifically designed for mother and baby rooms includes:

[0008] The purification device collects multimodal data including sound, temperature and humidity, fan speed, particulate matter concentration and infrared images, extracts image features and audio features, generates structured data frames, and periodically uploads the structured data frames to the cloud server.

[0009] The cloud server receives structured data frames, uses a composite neural network model with an attention mechanism to identify the activity patterns of people in the room, and sets a purification constraint library based on different activity patterns.

[0010] Set up a multi-layer stacked activated carbon filter element with adjustable filter layers, and build a purification efficiency model. Extract the current purification constraints from the purification constraint library, and solve the purification efficiency model through an intelligent optimization algorithm to obtain the optimal combination control strategy of fan speed and variable filter stacking layer under the current constraints.

[0011] The cloud server sends the control commands for fan speed and filter layer number to the edge computing unit. The stepper motor drives the filter displacement mechanism to adjust the number of activated carbon filter layers, while controlling the fan speed and recording and uploading environmental monitoring data.

[0012] The cloud server performs correlation analysis on the environmental monitoring data uploaded by the purification device and the control strategy issued in the previous cycle, and uses identification algorithms to calibrate and update the purification efficiency model parameters online, thereby optimizing the purification device adjustment strategy for subsequent cycles.

[0013] Preferably, the purification device collects multimodal data including sound, temperature and humidity, fan speed, particulate matter concentration and infrared image, and preprocesses the collected multimodal data, including standardization, timestamp alignment and anomaly removal;

[0014] After preprocessing the collected data, feature extraction is performed on the infrared image and audio data. For the infrared image, a lightweight convolutional neural network is used to input the preprocessed infrared image frame into the lightweight convolutional neural network and extract the feature map before the global average pooling layer as the image feature vector.

[0015] For audio data, the audio data is first divided into short time frames. After applying the Hamming window function to each frame of audio data, a fast Fourier transform is performed to obtain the spectrum. The spectrum is then passed through a Mel filter bank, and the logarithmic energy is taken. A discrete cosine transform is then performed to obtain the Mel frequency cepstral coefficients. The Mel frequency cepstral coefficients of all audio frames within a single synchronization period are used to construct a two-dimensional audio feature matrix.

[0016] The purification device integrates all the pre-processed and feature-extracted data into a structured data frame, and uploads the structured data frame as a message to the cloud server at a preset period.

[0017] Preferably, the cloud server receives the structured data frame uploaded by the purification device, first parses the received formatted data frame, and extracts the timestamp, multi-dimensional sensor data, image feature vector, and audio feature matrix;

[0018] A composite neural network model is constructed on a cloud server to integrate and classify multimodal time-series data. The composite neural network model consists of a multimodal feature encoding layer, a time-series information fusion layer, and an activity pattern classification layer.

[0019] Preferably, in the multimodal feature encoding layer, independent deep feature extraction is performed on data of different modalities. For each time step data frame, the standardized numerical data is concatenated into a one-dimensional vector, and the one-dimensional vector is mapped to a high-dimensional unified feature space through the first fully connected layer to obtain numerical features. The extracted image feature vector is adjusted in dimension and aligned in space through the second fully connected layer to obtain image features. For the audio feature matrix, a convolution module is used to process it to capture local time-frequency patterns in the audio signal. The output of the convolution module is flattened and passed through the third fully connected layer to generate audio features.

[0020] In the temporal information fusion layer, the encoded features are fused and temporal analyzed. At each time step, the numerical feature, image feature and audio feature matrices are concatenated to form a fused feature vector. The fused feature vector of the entire time series is input into a bidirectional long short-term memory network to generate a hidden state sequence containing contextual information. At the same time, a self-attention mechanism is introduced after the Bi-LSTM layer of the bidirectional long short-term memory network to dynamically allocate weights by calculating the correlation between elements in the hidden state sequence.

[0021] Aggregate the context vectors of all time steps to obtain a context representation vector that represents the information of the entire time series;

[0022] In the activity pattern classification layer, the context representation vector is input into a classifier consisting of two fully connected layers and a Softmax activation function. The classifier outputs a probability distribution vector, and each dimension of the probability distribution vector corresponds to the probability of a predefined activity pattern of a person in the mother and baby room.

[0023] After identifying the current activity pattern of the character, the corresponding purification strategy boundary is matched and extracted from the pre-built purification constraint library.

[0024] Preferably, a variable filter structure is designed, which consists of multiple independent activated carbon filter elements of the same specifications. The activated carbon filter elements are installed in a frame with guide rails, and the number of activated carbon filter elements pushed into or removed from the main air duct is controlled by a displacement mechanism driven by a stepper motor.

[0025] Preferably, a purification efficiency model is constructed, which takes the fan speed and the number of filter stack layers as two core independent variables; the purification efficiency model includes: particulate matter removal rate model, ventilation volume model and noise model;

[0026] Based on the purification strategy boundaries corresponding to the identified activity patterns of individuals, a genetic algorithm is used to solve the purification efficiency model, specifically including:

[0027] Step a: Initialize the population by randomly generating a set of individuals, each representing a solution, encoded as a chromosome, specifically the fan speed and the number of activated carbon filter layers;

[0028] Step b: Assess fitness. For each individual in the population, calculate the predicted performance index using the purification efficiency model and substitute it into the fitness function to calculate the fitness value.

[0029] Step c: Prioritize individuals with high fitness to enter the next generation, perform crossover on the selected individuals to generate new offspring, and perform mutation on the offspring to increase population diversity and avoid getting trapped in local optima;

[0030] Step d: Repeat steps b and c until the preset maximum number of iterations is reached or the optimal fitness value of the population no longer improves after multiple generations;

[0031] When the genetic algorithm terminates, the fan speed and activated carbon filter layer number corresponding to the chromosome of the individual with the highest fitness in the population are the optimal combination control strategy of fan speed and filter stack layer number obtained under the current constraints.

[0032] Preferably, the purification device receives control commands from the cloud server, which include the target fan speed and the target number of filter stack layers. After receiving the data packet, the purification device parses and verifies its validity to confirm that the parameter values ​​are within the physical executable range of the device.

[0033] Preferably, after the instruction verification is successful, the purification device initiates a collaborative execution sequence. The collaborative execution sequence prioritizes adjusting the mechanical structure, and then adjusts the fan speed. The specific execution logic includes:

[0034] Adjust the number of activated carbon filter layers, calculate the required difference in the number of layers to be adjusted, and drive the filter displacement mechanism to adjust according to the required difference in the number of layers;

[0035] To smoothly control the fan speed, the purification device uses PWM speed regulation to adjust the fan speed. Based on the target speed, the corresponding PWM signal duty cycle is calculated and output to the fan drive board to smoothly adjust the fan speed to the target speed.

[0036] After the adjustment is completed at the purification device, it enters a stabilization period, waiting for the airflow field and environmental parameters inside the purification device to reach a new steady state. After the stabilization period ends, the purification device reactivates the sound sensor and particulate matter concentration sensor, collects and records the actual noise value and particulate matter concentration value in the current environment, packages the control command executed this time with the collected execution result data to form a feedback data frame, and uploads it to the cloud server.

[0037] Preferably, the cloud server receives and caches the feedback data frames uploaded by the purification device, uses the feedback data frames to calibrate and update the purification efficiency model, and applies the updated purification efficiency model to the next cycle of intelligent optimization to generate a control strategy that is more in line with the actual state of the device.

[0038] A multi-parameter adaptive adjustment system for a mother-and-baby room activated carbon purification device, which is used to implement the multi-parameter adaptive adjustment method of the mother-and-baby room activated carbon purification device, is characterized by including: a multi-modal data acquisition and processing module, a behavior recognition module, a constraint optimization control module, an execution and feedback module, and a model calibration module;

[0039] The multimodal data acquisition and processing module is used to collect multimodal data including sound, temperature and humidity, fan speed, particulate matter concentration and infrared image at the purification device end, extract image features and audio features, generate structured data frames, and periodically upload the structured data frames to the cloud server.

[0040] The behavior recognition module receives structured data frames from the cloud server, uses a composite neural network model including an attention mechanism to identify the activity patterns of people in the room, and sets a purification constraint condition library according to different activity patterns.

[0041] The constraint optimization control module sets up a multi-layer stacked activated carbon filter element with an adjustable number of filter layers, constructs a purification efficiency model, extracts the current purification constraints from the purification constraint condition library, solves the purification efficiency model to obtain the optimal combination control strategy of fan speed and variable filter stacking layer under the current constraints.

[0042] The execution and feedback module is described in which the cloud server sends control commands for the fan speed and the number of filter layers to the edge computing unit. The stepper motor drives the filter displacement mechanism to adjust the number of activated carbon filter layers, while controlling the fan speed and recording and uploading environmental monitoring data.

[0043] The model calibration module performs correlation analysis on the environmental monitoring data uploaded by the purification device and the control strategy issued in the previous cycle on the cloud server, and uses the identification algorithm to calibrate and update the purification efficiency model parameters online, thereby optimizing the purification device adjustment strategy for subsequent cycles.

[0044] The beneficial effects of this application are as follows: This application significantly improves data reliability and consistency through multimodal acquisition and preprocessing; image and audio feature extraction enhances state representation capabilities and solves the incompleteness of single sensing; structured data frames facilitate compression and transmission, ensuring the real-time performance and traceability of cloud processing, and providing a high signal-to-noise ratio and time-consistent input foundation for subsequent identification and control.

[0045] This application proposes a composite neural network based on an attention mechanism that can focus on key spatiotemporal cues and accurately identify activity patterns such as sleep and play. The identification results are mapped to differentiated purification constraints to build an extensible constraint library and achieve contextual adaptation. This reduces reliance on manual rules and improves the generalization ability and strategy targeting in complex scenarios.

[0046] This application introduces an adjustable-layer activated carbon filter and a purification efficiency model. Under multiple objectives such as noise, particulate matter, and ventilation volume, it adopts intelligent optimization to obtain the optimal combination of fan speed and filter layer count; taking into account purification efficiency, comfort, and energy consumption.

[0047] This application sends control commands from the cloud, and the edge device uses a stepper motor to precisely adjust the number of filter stacking layers and link the fan speed to achieve fast and repeatable execution; it also records and transmits sound and particulate matter data simultaneously to form a closed-loop verification channel.

[0048] This application performs correlation analysis between the execution results and the previous cycle strategy, and uses online identification to adaptively update model parameters to resist sensor drift, filter material aging and seasonal changes; continuously shortens the optimization convergence time and reduces ineffective adjustments; and continuously improves prediction and control accuracy by accumulating data during operation, ensuring long-term stable performance.

[0049] This application's method achieves closed-loop adaptive control of "identification—constraint—optimization—execution—calibration," simultaneously optimizing three objectives—low noise, low particulate matter, and suitable ventilation—for sensitive scenarios involving mothers and infants. Leveraging cloud-edge collaboration and online model calibration, it remains robust to environmental changes and individual differences, significantly improving comfort and safety during sleep and feeding. While ensuring purification effectiveness, it reduces energy consumption and filter material consumption, decreasing maintenance frequency and costs, and maintaining high-efficiency, explainable, and traceable operational quality over the long term. Attached Figure Description

[0050] Figure 1 Flowchart of the multi-parameter adaptive adjustment method for the activated carbon purification device for mother and baby rooms provided in this application;

[0051] Figure 2 The flowchart for human activity pattern recognition provided in this application;

[0052] Figure 3 The flowchart for solving the intelligent optimization algorithm provided in this application is shown.

[0053] Figure 4 The edge-side collaborative control and feedback flowchart provided in this application;

[0054] Figure 5 The flowchart for online calibration and optimization of the model provided in this application;

[0055] Figure 6The diagram shows the structure of the multi-parameter adaptive adjustment system of the activated carbon purification device for mother and baby rooms provided in this application. Detailed Implementation

[0056] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0057] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0058] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of this application. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0059] Example 1

[0060] Reference Figures 1 to 5 This is the first embodiment of the present application, such as Figure 1 As shown, a multi-parameter adaptive adjustment method for an activated carbon purification device specifically designed for mother and baby rooms is provided.

[0061] Step 1: The purification device collects multimodal data including sound, temperature and humidity, fan speed, particulate matter concentration and infrared images, extracts image features and audio features, generates structured data frames, and periodically uploads the structured data frames to the cloud server.

[0062] The purification device uses an embedded system with a built-in main control board to collect and process multimodal data in the mother and baby room in real time. Data collection is achieved through various types of sensor modules.

[0063] Sound data is acquired using a high signal-to-noise ratio omnidirectional MEMS microphone array, for example, with a sampling frequency of 16kHz and a quantization bit depth of 16-bit, to ensure that key acoustic events such as baby cries, adult voices, and ambient background noise can be clearly captured.

[0064] Temperature and humidity data are collected by an integrated digital temperature and humidity sensor (such as the SHT series), and the collection frequency can be set to 0.1Hz. The fan speed data is monitored in real time by a Hall effect sensor to monitor the rotor position change of the built-in DC brushless fan in the purification device, and the actual fan speed (rpm) is calculated. The data reporting frequency can be 0.5Hz.

[0065] Particulate matter concentration data is collected through particulate matter sensors based on the principle of laser scattering (such as the PMS series) to obtain the mass concentration (μg / m³) of PM2.5 and PM10 in the room air in real time, with a collection frequency of 0.1Hz; infrared images are collected through a non-contact infrared thermal imaging sensor array to detect the presence, number and approximate location of people in the room while protecting privacy.

[0066] After collecting multimodal data, the collected multimodal data is preprocessed to improve data quality and usability. The preprocessing process includes standardization, timestamp alignment, and anomaly removal.

[0067] Z-score normalization is applied to continuous numerical data (temperature, humidity, fan speed, and particulate matter concentration) to eliminate differences in dimensions and numerical ranges between data from different sensors. The formula for Z-score normalization is: ;in, This represents the standardized data value; This represents the multimodal data values ​​collected by the sensor; The arithmetic mean of the sensor data over a preset time window (e.g., the past 5 minutes) prior to the current time point; This represents the standard deviation of the data within this time window. Through Z-score standardization, all sensor data are mapped to a distribution with a mean of 0 and a standard deviation of 1, which is beneficial for the stable training and rapid convergence of the subsequent neural network model.

[0068] The collected multimodal data is timestamped and aligned using a high-precision clock embedded in the purification device's main control board as a reference. Each frame of data (including image frames and audio blocks) is assigned a uniform timestamp. Due to the different sampling rates of each sensor, a strategy combining downsampling and interpolation is employed, with the lowest common data update frequency as the baseline synchronization period. For high-frequency data (such as audio data), data within a synchronization period is considered a single data block. For low-frequency data (such as temperature and humidity), if there are no updates within a synchronization period, the valid values ​​from the previous moment are used for padding (forward padding). This ensures that at the end of each synchronization period, all modal data have corresponding values, thus forming a strictly time-aligned dataset. Timestamp alignment guarantees strong correlation between multimodal data in the time dimension, which is the foundation for effective fusion analysis.

[0069] Outliers in the multimodal data are removed using an anomaly detection algorithm based on the interquartile range (IQR). For each numerical sensor data stream, the first quartile (IQR) within a sliding time window is dynamically calculated. ) and the third quartile ( ), and calculate the interquartile range The valid range of data points is defined as follows: ,in This is a configurable coefficient, typically set to 1.5. Any data point outside this range is considered an outlier and replaced with a boundary value from the valid range. IQR's anomaly detection algorithm is robust to outliers caused by momentary sensor jitter or sudden environmental changes, preventing extreme outliers from interfering with subsequent model analysis.

[0070] After preprocessing the collected data, feature extraction is performed on the infrared image and audio data to reduce data dimensionality and extract key information.

[0071] For infrared images, a lightweight convolutional neural network (CNN), such as MobileNetV2, is used. The preprocessed infrared image frames are input into the lightweight CNN, and the feature maps before the global average pooling layer are extracted as image feature vectors. Image feature vector (For example, a 1280-dimensional vector) contains compact and high-level semantic information about the number, distribution, and posture of heat sources (i.e., human bodies) in the image. Feature extraction from infrared images not only significantly reduces data transmission volume but also effectively protects user privacy by not transmitting the original image.

[0072] For audio data, it is first segmented into short frames of 25 milliseconds with a frame shift of 10 milliseconds. After applying a Hamming window function to each frame, a Fast Fourier Transform (FFT) is performed to obtain the spectrum. The spectrum is then passed through a Mel filter bank, and the logarithmic energy is taken. Finally, a Discrete Cosine Transform (DCT) is performed, and the first 13 coefficients are taken as the Mel frequency cepstral coefficients (MFCC) of that frame. The MFCC coefficients of all audio frames within a synchronization period (e.g., 10 seconds) together constitute a two-dimensional audio feature matrix. This matrix can effectively characterize the intrinsic features of sound signals within a given time period, and is crucial for identifying specific activity patterns such as baby cries and human voices.

[0073] At the purification device, all preprocessed and feature-extracted data are integrated into a structured data frame. This frame includes: a unique device identifier, a synchronization timestamp, standardized temperature and humidity values, fan speed, particulate matter concentration, and image feature vectors. and audio feature matrix The purification device uploads this structured data frame as a message to the cloud server at a preset interval (e.g., every 10 seconds).

[0074] This step, by performing complex data preprocessing and feature extraction tasks at the edge, significantly reduces the computational burden on the cloud server and substantially decreases the amount of data transmitted over the network. It transforms the raw, heterogeneous data stream into high-quality, structured information with defined features, laying a solid data foundation for accurate human activity pattern recognition in the cloud. This ensures the responsiveness and decision-making accuracy of the entire adaptive adjustment process.

[0075] Step 2: The cloud server receives structured data frames and uses a composite neural network model incorporating an attention mechanism to identify the activity patterns of people in the room. Simultaneously, based on these different activity patterns, it sets a library of purification constraints. (See also...) Figure 2 This is a flowchart for the human activity pattern recognition process in this step.

[0076] The cloud server continuously receives structured data frames uploaded from the purification device. It first parses the received formatted data frames to extract timestamps, standardized multi-dimensional sensor data, and image feature vectors. and audio feature matrix .

[0077] To capture the dynamic temporal characteristics of the activity, continuously arriving data frames are cached as a fixed-length time series on a cloud server. For example, a sliding window containing the past 10 data frames (corresponding to 100 seconds) is constructed to form an input sequence. ,in , This represents the T-th data frame, and each data frame... A complete structured data frame representing a single moment in time; this sequence It will be used as input to the composite neural network model.

[0078] The composite neural network model used in this step is designed with a core architecture for deep fusion and classification of multimodal time-series data. The composite neural network model consists of a multimodal feature encoding layer, a time-series information fusion layer (including an attention mechanism), and an activity pattern classification layer.

[0079] In the multimodal feature encoding layer, the composite neural network model performs independent deep feature extraction on data from different modalities. For each time step... data frames The standardized numerical data such as temperature, humidity, fan speed, and particulate matter concentration are concatenated into a one-dimensional vector. This one-dimensional vector is then mapped to a higher-dimensional unified feature space through a fully connected layer to obtain the numerical features. The image feature vector extracted from the purification device. Similarly, image features are obtained by performing dimensional adjustment and spatial alignment using a fully connected layer. For audio feature matrix The composite neural network model employs a convolutional module consisting of two layers of one-dimensional convolutions and max pooling to capture local time-frequency patterns in the audio signal. The output of the convolutional module is flattened and passed through a fully connected layer to ultimately generate audio features. Through this encoding layer, the original heterogeneous multimodal data is transformed into feature vectors of the same dimension and semantic alignment at each time step.

[0080] In the temporal information fusion layer, the composite neural network model fuses and performs temporal analysis on the encoded features. At each time step... The feature vectors of the three modalities , and The features are concatenated to form a fused feature vector. Then, the fused feature vector of the entire time series is... The input is fed into a bidirectional long short-term memory (Bi-LSTM) network. This represents the fused feature vector at time step T. Bi-LSTM can learn the temporal dependencies in a sequence simultaneously from both the forward and backward directions, generating a hidden state sequence that includes contextual information. To enable the model to dynamically focus on the most informative parts of the sequence, a self-attention mechanism is introduced after the Bi-LSTM layer. This mechanism dynamically assigns weights by calculating the correlation between elements in the hidden state sequence. The specific calculation is as follows:

[0081] ;

[0082] in, For activation function, It is a time step The hidden state, It is the hidden state of time step T; It consists of all hidden states The matrix formed; These are learnable weight matrices used to map the input to the query, key, and value spaces, respectively. It represents the dimension of the key vector, used for scaling; It is the calculated attention weight matrix, representing the weights at the generation time step. When outputting, the degree of attention paid to all other positions in the sequence; It is the context vector after attention weighting.

[0083] Finally, the context vectors of all time steps are aggregated (e.g., by average pooling) to obtain a context representation vector that can represent the core information of the entire time series. Temporal modeling based on attention mechanisms enables composite neural network models to accurately capture key instantaneous events such as a baby suddenly crying or people entering or leaving a room, and assigns them higher weights, greatly improving the accuracy of recognition.

[0084] In the activity pattern classification layer, the final context representation vector is... The input is fed into a classifier consisting of a two-layer fully connected network and a Softmax activation function. The classifier outputs a probability distribution vector, where each dimension of the probability distribution vector corresponds to the probability of a predefined activity pattern of a person in the mother and baby room.

[0085] For example, the preset activity modes in this step include: no-man's-land mode, sleeping mode, care mode (such as quiet activities like feeding and changing diapers), and activity mode (such as playing, early education, and tidying up the room). The composite neural network model selects the mode with the highest probability as the recognition result of the person's activity mode in the current room.

[0086] After identifying the current activity mode, the corresponding purification strategy boundaries are matched and extracted from a pre-built purification constraint library. This library is built based on extensive research data on maternal and infant health and environmental comfort, setting a set of clear and differentiated environmental control objectives for each activity mode. Specific constraint examples are as follows: If the mode is identified as sleep mode, the strictest noise constraint is set, such as fan noise. Moderately stringent particulate matter concentration targets And the minimum fresh air volume to ensure basic physiological needs. If the activity mode is active, set a more lenient noise constraint. Strict particulate matter concentration targets (Due to dust generated during the event), and a high volume of fresh air. To cope with higher metabolic rates. If identified as unmanned mode with no noise constraints, set the most stringent particulate matter concentration target. (Powerful purification activated), fresh air volume It can be set to a lower value to save energy.

[0087] This step constructs a deep learning model that integrates multimodal information and incorporates an attention mechanism, enabling automated recognition of complex and dynamic human activity patterns within the mother-and-baby room. Compared to traditional methods relying on a single sensor or simple rules, the model constructed in this step can more deeply understand the contextual information of the scene. Furthermore, by associating the recognition results with a refined library of purification constraints, this application transforms the abstract purification task requirements into a mathematical problem with clear boundaries and optimization objectives. This provides clear, reasonable, and user-friendly input for subsequent steps to solve for the optimal control strategy, ensuring that the operation mode of the purification device truly meets the actual needs of mothers and babies.

[0088] Step 3: Set up a multi-layer stacked activated carbon filter cartridge with adjustable filter layers and construct a purification efficiency model. Extract the current purification constraints from the purification constraint library, and use an intelligent optimization algorithm to solve the purification efficiency model to obtain the optimal combination control strategy of fan speed and variable filter stacking layer under the current constraints. See also Figure 3 This is a flowchart of the intelligent optimization algorithm solution for this step.

[0089] Traditional air purifiers have a fixed number and type of filter layers, resulting in constant air resistance and single-pass filtration efficiency, making it impossible to dynamically balance purification efficiency and operating noise. To overcome this limitation, this solution designs a variable filter structure, consisting of multiple independent, identical thin-layer activated carbon filter elements mounted within a frame with guide rails. A precision displacement mechanism driven by a stepper motor (such as a rack and pinion or a lead screw slide) precisely controls the number of filter elements pushed into or removed from the main air duct. When low-noise operation is required, the number of filter layers is reduced to lower air resistance, ensuring sufficient airflow at lower fan speeds. When powerful purification is needed, the number of filter layers is increased to improve the pollutant capture rate per pass. The core purpose of this design is to transform the physical characteristics of the filter from static parameters into dynamically controllable variables, providing the purification device with a second degree of freedom besides fan speed. This makes it possible to find the globally optimal solution while meeting multiple objective constraints (noise, purification speed, airflow) in different scenarios (such as infant sleep, adult activity).

[0090] Construct a purification efficiency model, which will include the fan speed. (Unit: rpm) and number of filter stack layers (Unit: layer, integer) as two core independent variables. The purification efficiency model includes: particulate matter removal rate model, ventilation volume model, and noise model.

[0091] Constructing a particulate matter removal rate model The ability of activated carbon filters to remove particulate matter depends not only on the physical structure of the filter itself but also on the airflow velocity through it. A purification efficiency model describes the probability of particulate matter being captured when air passes through the filter.

[0092] ;

[0093] in, Indicates the single-pass removal rate; It is a removal efficiency coefficient related to the filter material and structure, which will be calibrated online in step 5; It is a wind speed influence factor, usually less than 1, which means that the higher the wind speed, the shorter the contact time between the air and the filter, and the single removal efficiency will decrease slightly; It is a natural exponential function. This model shows that increasing the number of filter layers... This will significantly improve removal efficiency, while increasing the fan speed This will reduce the efficiency of a single pass to some extent.

[0094] Constructing a ventilation volume model The ventilation volume (or the basis of clean air delivery rate, CADR) is directly affected by the fan power and air resistance.

[0095] ;

[0096] in, This is the final output ventilation volume (unit: m³ / h). It is the fan characteristic coefficient, which represents the proportional relationship between the rotational speed and the unobstructed airflow. It is a drag function, representing the number of filter stack layers. The resulting loss of air volume; It is the drag coefficient per unit number of floors; (Typically greater than 1) indicates that wind resistance increases non-linearly with increasing floor level. The ventilation volume model reveals the fan speed... It is the power source that provides airflow, while the number of filter layers... It is the main source of resistance.

[0097] Running noise model Noise is mainly generated by the operation of the fan, and is also affected by the turbulence generated when the airflow passes through the filter.

[0098] ;

[0099] in, This is the predicted operating noise (unit: dBA). It is the background noise of the device; Describe the logarithmic contribution of fan speed to noise, where It is the fan noise figure; Then the simulation is by Layered filters and high-speed airflow The turbulent noise generated by the combined effect, among which It is the turbulence noise figure. It is an index of the impact of wind speed on turbulent noise.

[0100] By constructing the above model, it is possible to control any given combination of controls. The resulting purification effect, ventilation volume, and noise level can be accurately quantified and predicted.

[0101] After obtaining the purification constraints (target particulate matter concentration) corresponding to the current character activity mode. Upper limit of target noise Minimum ventilation volume After that, the problem is transformed into a constrained multi-objective optimization problem. To solve this problem, a genetic algorithm (GA) is used as the core intelligent optimization algorithm.

[0102] First, define a comprehensive evaluation function (i.e., a fitness function). Used to evaluate a set of control strategies The fitness function is defined as follows: The fitness level is determined by the degree of superiority or inferiority of the purification effect and ventilation volume, while minimizing noise and energy consumption.

[0103] ;

[0104] in, The total amount of particulate matter removed per unit time can be approximated as: ( (This refers to the current measured particulate matter concentration). The energy consumption of the wind turbine can be approximated as... ; These are the weighting coefficients for each item, and these coefficients can be dynamically adjusted according to different activity modes (e.g., baby sleep mode). The weight will be very high.

[0105] It is a penalty term used to handle constraints:

[0106] ;

[0107] in, With a relatively large penalty coefficient, this penalty term causes a sharp drop in the fitness function value when the predicted noise or ventilation volume violates the constraints, thereby eliminating solutions that do not meet the constraints during the evolutionary process. This represents the function that takes the maximum value.

[0108] The solution process of a genetic algorithm includes:

[0109] Step a: Initialize the population by randomly generating a set of populations with a size of [number missing]. Each individual represents a solution, encoded as a chromosome. .in Within the allowable speed range of the fan Random value is selected from within. Within an adjustable range of filter layers Randomly select an integer from the given information.

[0110] Step b: Assess fitness for each individual in the population. The predicted performance index was calculated using a purification efficiency model and then substituted into the fitness function. Calculate its fitness value.

[0111] Step c: Chromosome selection, crossover, and mutation, using roulette wheel selection or tournament selection to prioritize individuals with high fitness for the next generation. Crossover is then performed on the selected individuals (e.g., crossover with...). Perform arithmetic cross on the values. (Single-point crossover of values) to generate new offspring individuals. Then, mutation operations are performed on the offspring individuals (e.g., on...). Add a Gaussian perturbation to the value, for The values ​​are randomly added or subtracted (to increase population diversity and avoid getting trapped in local optima).

[0112] Step d: Algorithm iteration and termination. Repeat steps b and c until the preset maximum number of iterations is reached or the optimal fitness value of the population no longer significantly improves over multiple generations.

[0113] When the genetic algorithm terminates, the chromosome corresponding to the individual with the highest fitness in the population is... This refers to the combined control strategy of optimal fan speed and filter stacking layer number obtained under the current constraints.

[0114] This step establishes a precise physical performance model and transforms the complex multi-objective purification requirements into a well-defined mathematical optimization problem, achieving a shift from qualitative requirements to quantitative solutions. This step also employs intelligent optimization techniques such as genetic algorithms, which can efficiently search and locate the globally optimal or near-optimal solution that satisfies all constraints within a vast array of potential control strategy combinations. This ensures that every adjustment of the purification device maximizes the core functions of air purification and circulation while meeting the stringent requirements of specific scenarios (such as ensuring a baby's peaceful sleep), thus achieving true intelligent and adaptive adjustment.

[0115] Step 4: The cloud server sends control commands for fan speed and filter layer count to the edge computing unit. The edge unit uses a stepper motor to drive the filter displacement mechanism to adjust the number of activated carbon filter layers, simultaneously controlling the fan speed and recording and uploading environmental monitoring data. See also... Figure 4 This is the flowchart for edge-side collaborative control and feedback in this step.

[0116] The purification device receives control commands from the cloud server in real time, and the commands explicitly include the target fan speed. (Unit: rpm) and target filter stacking layers (Unit: Layer) Two core parameters. Upon receiving the data packet, the purification device immediately parses and verifies its validity, confirming that the parameter values ​​are within the device's physically executable range (e.g., , ), and These are the adjustable lower and upper limits of the fan speed, respectively. and These are the minimum and maximum filter stacking layers, respectively, to prevent abnormal instructions from damaging the hardware.

[0117] After the command verification is successful, the purification device initiates a coordinated execution sequence. This sequence prioritizes adjusting the mechanical structure and then the fan speed to ensure a smooth transition to the purification state.

[0118] The specific execution logic includes: adjusting the number of filter stack layers and smoothly controlling the fan speed.

[0119] To adjust the number of filter stack layers, the embedded system at the purification device first accesses and reads the current filter layer status variable from its internal storage. Then calculate the difference in the number of layers that need to be adjusted. .like This drives the filter displacement mechanism for adjustment. The filter displacement mechanism can consist of a stepper motor, a precision lead screw, and a filter support frame. The purification device end adjusts according to... The sign (positive or negative) determines the rotation direction of the stepper motor (forward rotation increases the number of layers, reverse rotation decreases the number of layers), and calculates the total number of step pulses required. , ;in, This is the total number of pulses that need to be sent to the stepper motor driver; This is a pre-calibrated constant representing the number of stepper motor pulses required to move each filter layer. It is determined by the lead screw and the stepper motor step angle. The purification device sends pulses at a precise frequency to the stepper motor driver via its GPIO (General Purpose Input / Output) interface. A pulse drives the filter tray to the target position. After adjustment, the status variable is updated. In the process of adjusting the number of filter layers, the layer number command is converted into a precise pulse count, which ensures the physical positioning accuracy and repeatability of the filter stack thickness.

[0120] The fan speed is smoothly controlled. After the filter layer adjustment is completed or it is confirmed that no further adjustment is needed, the purification device adjusts the fan speed. To achieve precise control of the DC brushless fan, PWM (Pulse Width Modulation) speed regulation is used. The purification device adjusts the fan speed according to the target speed. The duty cycle of the corresponding PWM signal is calculated using a linear mapping function. , ;in, The target PWM duty cycle (typically 0-100%). It is the minimum duty cycle required for the fan to start; Is the wind turbine in? The minimum stable speed corresponding to the duty cycle; This is the speed-duty cycle conversion factor, representing the percentage increase in duty cycle required for every 1 rpm increase in speed. This factor is also pre-calibrated. The PWM generator built into the microcontroller at the purification device will generate a value with this duty cycle. The square wave signal is output to the fan drive board, thereby smoothly adjusting the fan speed to [the desired speed]. Using PWM speed control instead of simple voltage regulation provides a wider speed range and higher energy efficiency.

[0121] After the above two actions are completed at the purification device, a brief period of stability is entered. (For example, this stabilization period can be set to 10 seconds), waiting for the internal airflow field and environmental parameters of the device to reach a new steady state. After the stabilization period ends, the sound sensor and particulate matter concentration sensor at the purification device end are reactivated to collect and record the actual noise value in the current environment. and particulate matter concentration value Finally, the purification device packages the executed control commands and the collected execution result data into a single package containing {control commands: {R, L}, execution result: { , The feedback data frame with timestamp is sent and uploaded to the cloud server.

[0122] This step transforms cloud-based strategies into precise electromechanical control command sequences at the edge, enabling reliable regulation of the air purification device's physical state. Through a step-by-step execution and smooth adjustment design, the stability of the device's operating state switching and the comfort of the user experience are guaranteed. More importantly, by immediately collecting and uploading environmental feedback data after execution, a complete closed loop from decision-making to execution to feedback is constructed, providing a high-quality, strongly correlated real-time data source for the adaptive optimization of the purification efficiency model.

[0123] Step 5: On the cloud server, perform correlation analysis between the environmental monitoring data uploaded by the purification device (as the execution result) and the control strategy issued in the previous cycle. Then, use an identification algorithm to calibrate and update the purification efficiency model parameters online, optimizing the purification device adjustment strategy for subsequent cycles. See also... Figure 5 This is a flowchart of the online calibration and optimization process for the model in this step.

[0124] The cloud server continuously receives and caches feedback data frames uploaded by the purification device. These data frames contain the control strategy of the previous cycle (target fan speed). With the target number of filter layers The environmental monitoring results (actual measured noise values) after the implementation of this strategy and particulate matter concentration The cloud server then uses a unique timestamp or task ID to establish a precise causal relationship between the control strategy and the execution result, forming a complete {input: (R, L), output: ()} , Data pairs with strong correlations provide an effective basis for model parameter identification and calibration. By establishing such explicit causal relationships, the actual physical effects of specific control strategies can be accurately assessed, providing unbiased data input for subsequent parameter calibration.

[0125] The purification efficiency model is updated online using feedback data frames. Essentially, the purification efficiency model is a set of mathematical functions describing the relationship between inputs (fan speed, number of filter layers) and outputs (noise, particulate matter concentration, ventilation volume). As activated carbon filters become saturated and fan bearings wear, the purification efficiency model parameters drift. To address this issue, this step employs a recursive least squares method with a forgetting factor as the core identification algorithm to update the purification efficiency model parameters online.

[0126] Assume the noise sub-model in the purification efficiency model can be expressed as: ,in It is based on the control input The constructed regression vector, This is the transpose of the regression vector. This is the parameter vector of the noise model to be identified. In the... During each control cycle, the cloud server performs the following update steps.

[0127] Calculate the prediction error using the previous period ( Model parameters and the input of the current cycle (Depend on and (Construction), calculate the predicted value of noise. Then calculate the predicted value and the actual measured value. Error between , ;in, Representative at the The bias in the prediction of the cycle model; It is the measured noise value of this period uploaded from the edge end; Control commands issued based on this cycle The constructed regression vector, for example, can be... This is a polynomial form; yes The parameter vector of the noise model after periodic updates.

[0128] Update the gain matrix and parameter vector based on the prediction error. Update the parameters using a recursive formula:

[0129] ;

[0130] in, It is the gain matrix, which determines the weight of the prediction error on the parameter update. This is the new parameter vector after this update; It is the covariance matrix of the parameter estimates, reflecting the uncertainty of the parameter estimates; It is the forgetting factor ( This is used to reduce the weight of older data in parameter estimation, thereby enabling the model to better track time-varying parameters. For example, It can be set to 0.99, which means that more emphasis is placed on the data performance of the most recent 100 periods. It is an identity matrix.

[0131] The parameters of both the particulate matter removal rate model and the ventilation volume model are updated in parallel using the same recursive identification algorithm. After the model parameters are updated, the cloud server uses the new parameter vector... Replace old parameters The updated purification efficiency model is stored in the model library. This updated model will be applied to the next cycle of intelligent optimization to generate a control strategy that better suits the current state of the device.

[0132] This step establishes a closed-loop learning mechanism encompassing strategy execution, effect feedback, and model calibration. This transforms the air purifier's control model from static to autonomously learning and adapting to physical changes such as filter aging and fan performance degradation, continuously refining its understanding. The online calibration mechanism ensures that the air purifier makes decisions based on the most accurate model throughout its entire lifecycle, continuously outputting optimal control strategies, improving long-term purification efficiency and user comfort, and achieving self-adaptation and intelligence.

[0133] Example 2

[0134] Reference Figure 6 This is the second embodiment of the present application, which provides a multi-parameter adaptive adjustment system for an activated carbon purification device specifically for mother and baby rooms.

[0135] The system includes: a multimodal data acquisition and processing module, a behavior recognition module, a constraint optimization control module, an execution and feedback module, and a model calibration module.

[0136] The multimodal data acquisition and processing module is used to collect multimodal data including sound, temperature and humidity, fan speed, particulate matter concentration and infrared images at the purification device end, extract image features and audio features, generate structured data frames, and periodically upload the structured data frames to the cloud server.

[0137] The behavior recognition module receives structured data frames from the cloud server, uses a composite neural network model including an attention mechanism to identify the activity patterns of people in the room, and sets a purification constraint condition library according to different activity patterns.

[0138] The constraint optimization control module sets up a multi-layer stacked activated carbon filter element with an adjustable number of filter layers, constructs a purification efficiency model, extracts the current purification constraints from the purification constraint condition library, solves the purification efficiency model to obtain the optimal combination control strategy of fan speed and variable filter stacking layer under the current constraints.

[0139] The execution and feedback module involves the cloud server sending control commands for fan speed and filter layer count to the edge computing unit. The stepper motor drives the filter displacement mechanism to adjust the number of activated carbon filter layers, while simultaneously controlling the fan speed and recording and uploading environmental monitoring data.

[0140] The model calibration module performs correlation analysis on the environmental monitoring data uploaded by the purification device and the control strategy issued in the previous cycle on the cloud server, and uses the identification algorithm to calibrate and update the purification efficiency model parameters online, thereby optimizing the purification device adjustment strategy for subsequent cycles.

[0141] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings or direct couplings or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0142] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of this application without departing from the spirit and scope of protection of the claims. All of these variations are within the protection scope of this application.

Claims

1. A multi-parameter adaptive adjustment method for an activated carbon purification device specifically designed for mother and baby rooms, characterized in that: include: The purification device collects multimodal data including sound, temperature and humidity, fan speed, particulate matter concentration, and infrared images. The collected multimodal data undergoes preprocessing, including standardization, timestamp alignment, and anomaly removal. Image and audio features are extracted. For infrared images, a lightweight convolutional neural network is used to input the preprocessed infrared image frames and extract the feature map before the global average pooling layer as the image feature vector. For audio data, the audio data is first segmented into short frames. After applying a Hamming window function to each frame, a fast Fourier transform is performed to obtain the spectrum. The spectrum is then passed through a Mel filter bank, and the logarithmic energy is taken, followed by a discrete cosine transform to obtain the Mel frequency cepstral coefficients. The Mel frequency cepstral coefficients of all audio frames within a single synchronization period are combined to form a two-dimensional audio feature matrix. Generate structured data frames and upload them as messages to the cloud server at preset intervals; The cloud server receives structured data frames, first parses the received formatted data frames, and extracts timestamps, multi-dimensional sensor data, image feature vectors, and audio feature matrices. A composite neural network model consisting of a multimodal feature encoding layer, a temporal information fusion layer, and an activity pattern classification layer is used to identify the activity patterns of people in a room. The multimodal feature encoding layer performs independent deep feature extraction on data of different modalities. For each time step data frame, the standardized numerical data is concatenated into a one-dimensional vector, which is then mapped to a high-dimensional unified feature space through a first fully connected layer to obtain numerical features. The extracted image feature vector is then adjusted in dimension and aligned in space through a second fully connected layer to obtain image features. For the audio feature matrix, a convolutional module is used to process the data and capture local time-frequency patterns in the audio signal. The output of the convolutional module is flattened and passed through a third fully connected layer to generate audio features. At each time step, the temporal information fusion layer concatenates the numerical feature, image feature, and audio feature matrices to form a fused feature vector. The encoded multimodal features are then input into a bidirectional long short-term memory network to generate a hidden state sequence containing contextual information. A self-attention mechanism is introduced after the Bi-LSTM layer of the bidirectional long short-term memory network to dynamically allocate weights by calculating the correlation between elements in the hidden state sequence. The context vectors of all time steps are aggregated to obtain a context representation vector representing the information of the entire time series. In the activity pattern classification layer, the context representation vector is input into a classifier consisting of two fully connected layers and a Softmax activation function. The classifier outputs a probability distribution vector, and each dimension of the probability distribution vector corresponds to the probability of a predefined activity pattern of a person in the mother and baby room. At the same time, a purification constraint condition library is set according to different person activity patterns. After identifying the current person activity pattern, the corresponding purification policy boundary is matched and extracted from the pre-built purification constraint condition library. A multi-layered activated carbon filter cartridge with adjustable filtration layers is configured, and a purification efficiency model is constructed. This model uses fan speed and the number of filter layers as two core independent variables, including a particulate matter removal rate model, a ventilation volume model, and a noise model. Current purification constraints are extracted from a purification constraint library, and the scenario constraints are transformed into a fitness function with a penalty term. A genetic algorithm is used for population initialization, individual selection, crossover, and mutation iterations, specifically including: Step a: Initialize the population by randomly generating a set of individuals, each representing a solution, encoded as a chromosome, specifically the fan speed and the number of activated carbon filter layers; Step b: Assess fitness. For each individual in the population, calculate the predicted performance index using the purification efficiency model and substitute it into the fitness function to calculate the fitness value. Step c: Prioritize individuals with high fitness to enter the next generation, perform crossover on the selected individuals to generate new offspring individuals, and perform mutation operations on the offspring individuals; Step d: Repeat steps b and c until the preset maximum number of iterations is reached or the optimal fitness value of the population no longer improves after multiple generations; When the genetic algorithm terminates, the fan speed and activated carbon filter layer number corresponding to the chromosome of the individual with the highest fitness in the population are the optimal combination control strategy of fan speed and variable filter stacking layer under the current constraints obtained by solving the purification efficiency model. The cloud server sends the control commands for fan speed and filter layer number to the edge computing unit. The stepper motor drives the filter displacement mechanism to adjust the number of activated carbon filter layers, while controlling the fan speed and recording and uploading environmental monitoring data. The cloud server performs correlation analysis on the environmental monitoring data uploaded by the purification device and the control strategy issued in the previous cycle. It also uses the recursive least squares method with forgetting factor to calibrate and update the purification efficiency model parameters online, and optimizes the purification device adjustment strategy for subsequent cycles.

2. The multi-parameter adaptive adjustment method for the activated carbon purification device for mother and baby rooms according to claim 1, characterized in that, The design incorporates a variable filter structure consisting of multiple independent activated carbon filter elements of identical specifications. These activated carbon filter elements are installed within a frame equipped with guide rails. The number of activated carbon filter elements pushed into or removed from the main air duct is controlled by a displacement mechanism driven by a stepper motor.

3. The multi-parameter adaptive adjustment method for the activated carbon purification device for mother and baby rooms according to claim 2, characterized in that, The purification device receives control commands from the cloud server, which include the target fan speed and the target number of filter stack layers. After receiving the data packet, the purification device parses and verifies its validity to confirm that the parameter values ​​are within the device's physical executable range.

4. The multi-parameter adaptive adjustment method for the activated carbon purification device for mother and baby rooms according to claim 3, characterized in that, After the command verification is successful, the purification device initiates a collaborative execution sequence. The collaborative execution sequence prioritizes adjusting the mechanical structure, and then adjusts the fan speed. The specific execution logic includes: Adjust the number of activated carbon filter layers, calculate the required difference in the number of layers to be adjusted, and drive the filter displacement mechanism to adjust according to the required difference in the number of layers; To smoothly control the fan speed, the purification device uses PWM speed regulation to adjust the fan speed. Based on the target speed, the corresponding PWM signal duty cycle is calculated and output to the fan drive board to smoothly adjust the fan speed to the target speed. After the adjustment is completed at the purification device, it enters a stabilization period, waiting for the airflow field and environmental parameters inside the purification device to reach a new steady state. After the stabilization period ends, the purification device reactivates the sound sensor and particulate matter concentration sensor, collects and records the actual noise value and particulate matter concentration value in the current environment, packages the control command executed this time with the collected execution result data to form a feedback data frame, and uploads it to the cloud server.

5. The multi-parameter adaptive adjustment method for the activated carbon purification device for mother and baby rooms according to claim 4, characterized in that, The cloud server receives and caches feedback data frames uploaded by the purification device, uses the feedback data frames to calibrate and update the purification efficiency model, and applies the updated purification efficiency model to the next cycle of intelligent optimization to generate a control strategy that is more in line with the actual state of the device.

6. A multi-parameter adaptive adjustment system for a mother-and-baby room activated carbon purification device, used to implement the multi-parameter adaptive adjustment method for the mother-and-baby room activated carbon purification device according to any one of claims 1-5, characterized in that, include: Multimodal data acquisition and processing module, behavior recognition module, constraint optimization and control module, execution and feedback module, and model calibration module; The multimodal data acquisition and processing module is used to collect multimodal data including sound, temperature and humidity, fan speed, particulate matter concentration and infrared image at the purification device end, extract image features and audio features, generate structured data frames, and periodically upload the structured data frames to the cloud server. The behavior recognition module receives structured data frames from the cloud server and uses a composite neural network model consisting of a multimodal feature encoding layer, a temporal information fusion layer, and an activity pattern classification layer to identify the activity patterns of people in the room. The multimodal feature encoding layer performs independent deep feature extraction on data from different modalities. The temporal information fusion layer inputs the encoded multimodal features into a bidirectional long short-term memory network to generate a hidden state sequence containing contextual information. A self-attention mechanism is introduced after the Bi-LSTM layer of the bidirectional long short-term memory network to dynamically allocate weights by calculating the correlation between elements in the hidden state sequence. Simultaneously, a purification constraint condition library is set according to different activity patterns of people. The constraint optimization control module is equipped with a multi-layer stacked activated carbon filter element with an adjustable number of filter layers, and constructs a purification efficiency model. The purification efficiency model takes the fan speed and the number of filter stack layers as two core independent variables, including a particulate matter removal rate model, a ventilation volume model, and a noise model. The module extracts the current purification constraints from the purification constraint condition library, transforms the scenario constraints into a fitness function with a penalty term, and uses a genetic algorithm for population initialization, individual selection, crossover, and mutation iteration to solve the purification efficiency model and obtain the optimal combination control strategy of fan speed and variable filter stack layer under the current constraints. The execution and feedback module is described in which the cloud server sends control commands for the fan speed and the number of filter layers to the edge computing unit. The stepper motor drives the filter displacement mechanism to adjust the number of activated carbon filter layers, while controlling the fan speed and recording and uploading environmental monitoring data. The model calibration module performs correlation analysis on the environmental monitoring data uploaded by the purification device and the control strategy issued in the previous cycle on the cloud server, and uses the recursive least squares method with forgetting factor to calibrate and update the purification efficiency model parameters online, thereby optimizing the purification device adjustment strategy for subsequent cycles.

Images