A method and system for dynamically optimizing multi-stage filtration efficiency of an air purification device

Abstract

The application discloses a kind of multi-stage filtering efficiency dynamic optimization method and system of air purification equipment, it is related to air purifier technical field, the method includes: in response to equipment start signal, the real-time operating parameter of target air purification equipment is obtained;Filtering load prediction is carried out, and multiple candidate shunt ratios and multiple candidate fan speeds are generated;Filtering efficiency analysis and energy consumption influence analysis are carried out to candidate shunt ratio and candidate fan speed, obtain multiple filtering efficiency values and multiple energy consumption influence values;Filtering efficiency weight and energy consumption influence weight are generated, and candidate control combination is iteratively optimized in combination with filtering efficiency value and energy consumption influence value, obtain optimal control parameter combination, and control target air purification equipment.The technical problem that existing air purification equipment cannot dynamically adjust operating parameters according to real-time pollution load, leading to the difficulty in balancing between filtering efficiency and energy consumption is solved.

Description

Technical Field

[0001] This application relates to the field of air purifier technology, specifically to a method and system for dynamically optimizing the multi-stage filtration efficiency of an air purification device. Background Technology

[0002] With the acceleration of industrialization and the improvement of people's health awareness, air purification equipment has become an important tool for improving indoor air quality.

[0003] However, traditional air purification equipment typically employs a fixed-speed fan and constant flow ratio control strategy. Each stage of the filtration unit operates according to preset fixed parameters, unable to dynamically adjust based on actual pollution load and filter unit status. When outdoor pollutant concentrations suddenly increase or indoor activities generate large amounts of particulate matter, the fixed parameters cannot guarantee that each stage of the filtration unit is in optimal working condition. This often leads to premature saturation and clogging of the pre-filter, while the post-filter high-efficiency unit fails to fully utilize its purification efficiency. Conversely, during periods of good air quality, constant high-speed operation results in unnecessary energy waste and shortens the overall lifespan of the filter units. Summary of the Invention

[0004] This application provides a method and system for dynamically optimizing the multi-stage filtration efficiency of air purification equipment, which solves the technical problem that existing air purification equipment cannot dynamically adjust its operating parameters according to real-time pollution load, resulting in an inability to balance filtration efficiency and energy consumption.

[0005] The technical solution to the above-mentioned technical problems in this application is as follows: In a first aspect, this application provides a method for dynamically optimizing the multi-stage filtration efficiency of an air purification device, the method comprising: In response to the device start signal, the real-time operating parameters of the target air purification device are obtained, wherein the real-time operating parameters include the pressure difference before and after each stage of the filter unit, the particulate matter concentration, and the cumulative operating time. Based on the real-time operating parameters, the filter load is predicted to obtain the predicted load demand, and multiple candidate diversion ratios and multiple candidate fan speeds are generated according to the predicted load demand. Filter efficiency analysis and energy consumption impact analysis are performed on the multiple candidate diversion ratios and multiple candidate fan speeds to obtain multiple filter efficiency values ​​and multiple energy consumption impact values. Based on the current operating conditions, filter efficiency weights and energy consumption impact weights are generated. Combining the multiple filter efficiency values ​​and the multiple energy consumption impact values, multiple candidate control combinations are iteratively optimized to obtain the optimal control parameter combination. The target air purification device is then controlled based on the optimal control parameter combination.

[0006] Secondly, this application provides a multi-stage filtration efficiency dynamic optimization system for air purification equipment, comprising: The data acquisition module is used to respond to the device start signal and acquire the real-time operating parameters of the target air purification device, wherein the real-time operating parameters include the pressure difference before and after each stage of the filter unit, the particulate matter concentration, and the cumulative operating time. The load forecasting module is used to perform filtered load forecasting based on the real-time operating parameters, obtain the predicted load demand, and generate multiple candidate diversion ratios and multiple candidate fan speeds according to the predicted load demand. The parameter calculation module is used to perform filtration efficiency analysis and energy consumption impact analysis on the multiple candidate diversion ratios and multiple candidate fan speeds, and obtain multiple filtration efficiency values ​​and multiple energy consumption impact values. The parameter optimization module is used to generate filtration efficiency weights and energy consumption impact weights based on the current operating conditions, and to iteratively optimize multiple candidate control combinations by combining the multiple filtration efficiency values ​​and the multiple energy consumption impact values ​​to obtain the optimal control parameter combination, and to control the target air purification device based on the optimal control parameter combination.

[0007] This application provides one or more technical solutions, which have at least the following technical effects or advantages: This application provides a method and system for dynamic optimization of multi-stage filtration efficiency in air purification equipment. First, a complete equipment status perception system is constructed by real-time collection of multi-dimensional operating parameters such as the pressure difference before and after each filtration unit, particulate matter concentration, and cumulative operating time. Second, based on a machine learning model, the filtration load is predicted, and the pollutant accumulation trend and efficiency decay curve of each filtration unit are determined, realizing a shift from passive response to proactive prevention in control mode. Third, by generating multiple candidate diversion ratios and fan speed combinations and conducting a dual analysis of their filtration efficiency and energy consumption impact, a comprehensive control strategy evaluation framework is established. Finally, the weights of filtration efficiency and energy consumption impact are dynamically adjusted according to the current operating conditions, and an improved particle swarm optimization algorithm is used for iterative optimization, effectively reducing energy consumption while ensuring purification effect, achieving an adaptive balance between filtration efficiency and energy consumption.

[0008] Through the above technical solution, this application overcomes the technical defects of the existing air purification equipment, which has fixed and rigid operating parameters and cannot adapt to dynamic changes in pollution load, and realizes the coordinated optimization control of multi-stage filtration units under different operating conditions. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating a method for dynamically optimizing the multi-stage filtration efficiency of an air purification device, as provided in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of a multi-stage filtration efficiency dynamic optimization system for an air purification device provided in an embodiment of this application.

[0011] The components represented by each number in the attached diagram are explained below: Data acquisition module 11, load forecasting module 12, parameter calculation module 13, parameter optimization module 14. Detailed Implementation

[0012] This application provides a method and system for dynamically optimizing the multi-stage filtration efficiency of air purification equipment, which addresses the technical problem that existing air purification equipment cannot dynamically adjust its operating parameters according to real-time pollution load, resulting in an inability to balance filtration efficiency and energy consumption.

[0013] Example 1, as Figure 1 As shown in the figure, this application provides a method for dynamically optimizing the multi-stage filtration efficiency of an air purification device, including: S10: Respond to the device start signal and obtain the real-time operating parameters of the target air purification device, wherein the real-time operating parameters include the pressure difference before and after each stage of the filter unit, the particulate matter concentration and the cumulative operating time; In this embodiment, the target air purification device adopts a multi-stage filtration architecture, including a pre-filter unit, a medium-efficiency filter unit, and a high-efficiency filter unit, with each filter unit arranged in series according to the airflow direction.

[0014] The pressure difference between the front and rear ends is collected in real time by pressure sensors deployed at the air inlet and outlet ends of each stage of the filter unit, which is used to characterize the degree of clogging and the degradation of the air permeability of the filter unit.

[0015] Particulate matter concentration is obtained through particulate matter concentration sensors distributed at the air inlet, between each level of filtration unit, and at the air outlet of the equipment, forming concentration distribution data along the process, which reflects the spatial distribution characteristics of pollutants and the real-time collection efficiency of each level of filtration unit.

[0016] The cumulative running time is recorded by the internal timer of the device's main control chip and is used to assess the fatigue aging status and remaining service life of the filter unit.

[0017] In the process of acquiring real-time operating parameters, the sensor data is first validated to remove outliers caused by communication failures or environmental interference. Data that passes the validation is cached according to timestamps to build a time series dataset within a preset time window.

[0018] S20: Based on the real-time operating parameters, perform filter load prediction to obtain the predicted load demand, and generate multiple candidate diversion ratios and multiple candidate fan speeds according to the predicted load demand; In this embodiment, the filter load prediction adopts a time-series prediction model based on a long short-term memory network. This model uses the pressure difference sequence before and after each level of filter unit, the particulate matter concentration sequence, and the cumulative operating time increment within a preset time window as input features, and outputs the predicted value of filter load demand within a preset future period.

[0019] Multiple candidate flow splitting ratios are generated based on the predicted load demand. First, the feasible range of the flow splitting ratio is determined. The flow splitting ratio represents the proportion of airflow distribution to each level of the filtration unit, and its feasible range is jointly constrained by the rated processing air volume of each level of the filtration unit and the current differential pressure. Within the feasible range, discrete sampling is performed according to a preset step size to generate a set of candidate flow splitting ratios covering various strategy preferences, such as pre-filtration with a focus on low load, balanced distribution, and post-high-efficiency filtration with a focus on high load.

[0020] When generating multiple candidate fan speeds, the base speed corresponding to the predicted load demand is used as the benchmark. Within the preset fluctuation range, a set of candidate speeds is generated according to an arithmetic or geometric sequence to ensure full coverage of all operating conditions, including energy-saving low-speed operation, standard operating condition operation, and high-load high-speed operation. The candidate flow split ratio and candidate fan speed are combined in a full permutation to form a multi-dimensional candidate control combination space.

[0021] The process of predicting the filtration load based on the real-time operating parameters to obtain the predicted load requirements of each filtration unit includes: Obtain a preset feature extraction template, and extract features from the real-time running parameters based on the preset feature extraction template to obtain the current running feature vector; The current operating feature vector is input into the filter load prediction model to obtain the predicted load demand of each level of filter unit. The filter load prediction model is a machine learning model trained based on historical operating data, which is used to predict the pollutant accumulation trend and filtration efficiency decay curve of each level of filter unit in the future period based on the current operating characteristics.

[0022] In this embodiment, the preset feature extraction template first includes three dimensions: time-domain statistical features, frequency-domain energy features, and trend change features. The time-domain statistical features include the mean, variance, maximum, and minimum values ​​of the pressure difference before and after each filtering unit within a preset time window, reflecting the central tendency and dispersion of pressure difference fluctuations.

[0023] Frequency domain energy characteristics are obtained by performing a fast Fourier transform on the pressure difference sequence, extracting the energy ratio of low-frequency trend components and high-frequency fluctuation components, which are used to identify the periodic patterns and sudden disturbance characteristics of the filter unit blockage process. The Fourier transform can convert the time domain signal into a frequency domain representation. By decomposing the pressure difference sequence into a superposition of sine waves of different frequencies, it reflects the periodic patterns and frequency components hidden in the time domain data.

[0024] The trend change characteristics are obtained by using the sliding window slope calculation method to obtain the rate of pressure difference growth and the acceleration of concentration change, and to determine the evolution direction of the filtration load. The sliding window slope calculation method refers to fitting the linear trend of pressure difference or concentration change with time using the least squares method within a preset time window, and using the slope of the fitted line as a quantitative indicator of the trend change characteristics. The larger the absolute value of the slope, the more drastic the change in filtration load.

[0025] Furthermore, before inputting the extracted current running feature vector into the filtered load prediction model, the feature vector is first normalized to eliminate the numerical scale differences between features of different dimensions. For example, the Z-score standardization method is used to map each feature component to a distribution space with a mean of zero and a standard deviation of one.

[0026] Specifically, the Z-score standardization method is a commonly used data preprocessing method. For each feature component, its original value is subtracted from the mean of that feature on the training set, and then divided by the standard deviation of that feature on the training set to obtain the standardized feature value.

[0027] Furthermore, the filtration load prediction model employs a two-layer long short-term memory (LSTM) network structure. The first layer, an LSTM network with 128 hidden units, captures long-term dependencies in the input feature sequence. The second layer, an LSTM network with 64 hidden units, further extracts high-level temporal abstract features. The network output layer uses a fully connected structure, outputting predicted load demands at three time scales: 2 hours, 4 hours, and 8 hours. Each time scale includes predicted pollutant accumulation and filtration efficiency decay coefficients for each filtration unit.

[0028] Specifically, the steps for constructing the filtered load prediction model include: Collect historical operation datasets of air purifiers of the same model, including historical real-time operation parameters and corresponding historical load change labels; An initial load forecasting model architecture is constructed, which is a regression model that combines time-series features and multi-dimensional parameter inputs; Using the historical operational dataset as training samples and the historical load change labels as supervision targets, the initial load prediction model architecture is trained until convergence, thus obtaining the trained filtered load prediction model.

[0029] In this embodiment, firstly, the collection period of the historical operating dataset covers the entire life cycle of the device, including the initial high-efficiency operation stage of the newly installed filter unit, the intermediate decay stage of gradual clogging, and the final inefficient stage nearing replacement, ensuring that the training samples contain complete evolution patterns of filter performance.

[0030] Historical load change labels are obtained through offline calibration. That is, after the historical data collection period ends, the actual load demand for that period is inferred from the actual pollutant accumulation and measured filtration efficiency of each level of filtration unit, and used as the supervision target for model training.

[0031] Secondly, when constructing the initial load forecasting model architecture, an encoder-decoder structure is adopted. The encoder part consists of the aforementioned two-layer LSTM network, which is responsible for compressing the multi-dimensional time series input into a fixed-dimensional hidden state vector. The decoder part adopts an attention-enhanced LSTM network, which enables the model to adaptively focus on key time nodes in the historical sequence when predicting future loads, thereby improving the predictive response capability to sudden pollution events.

[0032] Specifically, the Adam optimization algorithm is used during model training, with an initial learning rate of 0.001. A learning rate decay strategy is introduced, where the learning rate decreases by a factor of 0.5 when the validation set loss no longer decreases after 5 consecutive training epochs. To prevent overfitting, Dropout layers are inserted between LSTM layers with a dropout rate of 0.3. L2 regularization is applied to the fully connected output layer with a regularization coefficient of 0.001. The training batch size is 64, and the total number of training epochs is 200. An early stopping mechanism is used, terminating training when the validation set loss shows no improvement after 20 consecutive training epochs, and saving the model parameters with the best validation performance.

[0033] Furthermore, based on the predicted load demand, multiple candidate diversion ratios and multiple candidate fan speeds are generated, including: Activate the control policy generator, which includes multiple policy configuration units; The predicted load demand is input into the control strategy generator, which calls multiple strategy configuration units to configure control parameters for the predicted load demand, thereby obtaining a candidate combination set of multiple candidate diversion ratios and multiple candidate fan speeds.

[0034] In this embodiment, the control strategy generator firstly incorporates three strategy configuration units: an energy-saving priority configuration unit, an efficiency priority configuration unit, and a balanced optimization configuration unit. Each configuration unit performs differentiated analysis of the predicted load demand based on different optimization objectives.

[0035] The energy-saving priority configuration unit aims to reduce total energy consumption. When generating candidate flow distribution ratios, it tends to increase the airflow distribution ratio of the pre-filter unit, utilizes the low flow resistance characteristics of the primary and secondary filters to reduce fan energy consumption, and generates candidate parameter combinations in conjunction with a lower fan speed range.

[0036] The efficiency-priority configuration unit prioritizes ensuring air quality by increasing the airflow ratio of the post-high-efficiency filter unit when the predicted load demand is high, and matching it with a higher fan speed to ensure sufficient air volume and prioritize meeting stringent purification requirements.

[0037] The balanced optimization configuration unit takes into account the trade-off between energy efficiency and performance. In the design of the flow distribution ratio, a progressive distribution strategy is adopted to make the load distribution of each level of filter unit more uniform, delay the excessively rapid decay of a single unit, and the fan speed selection takes into account both response speed and operating economy.

[0038] The candidate parameter subsets output by the three strategy configuration units are merged and deduplicated, and invalid combinations that exceed the physical constraints of the equipment are eliminated, ultimately forming a candidate control combination space that covers the multi-objective optimization direction.

[0039] The construction steps of the control strategy generator include: Collect multiple sample load demand data, construct a sample load dataset, and label each sample load data in the sample load dataset to obtain a sample control parameter label set; Build a multi-strategy configuration unit architecture; Using the sample load dataset as input and the sample control parameter annotation set as the supervision target, the architecture of the multiple policy configuration units is trained synchronously to obtain multiple policy configuration units; The multiple policy configuration units are integrated to obtain the control policy generator.

[0040] In this embodiment, the sample load dataset is first constructed to cover typical application scenarios, including daily low-load operation in offices, peak load during peak periods of personnel gathering in conference rooms, continuous high-concentration pollution load after renovation, and special conditions such as seasonal pollen outbreaks. The sample control parameter annotation set is manually annotated based on equipment operating experience and energy efficiency optimization principles. The annotation content includes the recommended diversion ratio range and fan speed range for each operating condition, as well as the expected energy efficiency level and purification efficiency level for each parameter combination.

[0041] Secondly, the strategy configuration unit architecture adopts a hybrid design combining rule-based and data-driven approaches. The bottom layer embeds a physical constraint rule library, limiting hard constraints such as the sum of the diversion ratios to always be 1 and the fan speed not exceeding the rated upper limit. The upper layer uses a lightweight neural network, with the input being a predicted load demand vector and the output being a probability distribution of candidate control parameters, generating diverse candidate combinations through sampling. The energy-saving priority configuration unit's neural network uses a 3-layer fully connected structure, with 64, 32, and 16 hidden layer neurons respectively. The output layer uses a sigmoid activation function to normalize the diversion ratios and speeds to the energy-saving tendency range. The efficiency-priority configuration unit and the balanced optimization configuration unit use similar architectures, differing only in the output layer bias settings.

[0042] During training, the three policy configuration units share the sample load dataset but employ differentiated loss function designs. Specifically, the loss function of the energy-saving priority configuration unit incorporates an energy consumption estimation term, calculating and penalizing the estimated power consumption based on the diversion ratio and rotational speed; the loss function of the efficiency priority configuration unit adds a purification efficiency estimation term, evaluating the degree of matching between load demand and processing capacity; and the balanced optimization configuration unit considers both indicators simultaneously, achieving flexible balance through adjustable weighting coefficients. Training employs batch gradient descent with a learning rate of 0.01, stopping training when the parameter generation accuracy of each configuration unit on the validation set exceeds 85%.

[0043] Furthermore, the strategy configuration units are integrated, and dynamic activation mechanisms are set for each unit to automatically select the dominant configuration unit based on the confidence level of the current predicted load demand and the urgency of the operating condition. When the prediction confidence level is higher than 0.9 and the load demand is within the normal range, the balanced optimization configuration unit dominates the generation of candidate combinations; when the prediction confidence level is lower than 0.7 or a sudden pollution event is detected, the system switches to the efficiency priority configuration unit to ensure purification reliability; when the equipment is in low-load or energy-saving mode at night, the energy-saving priority configuration unit is prioritized to reduce operating costs.

[0044] In summary, the outputs of each configuration unit are subjected to consistency verification and conflict resolution by the arbitration module to ensure the rationality and completeness of the final candidate combination set, thus obtaining the trained control strategy generator.

[0045] S30: Perform filtration efficiency analysis and energy consumption impact analysis on the multiple candidate diversion ratios and multiple candidate fan speeds to obtain multiple filtration efficiency values ​​and multiple energy consumption impact values; In this embodiment of the application, the filtration efficiency analysis is based on the filtration efficiency evaluation model, which considers the particulate matter removal efficiency, gaseous pollutant adsorption capacity and microbial inactivation capacity of each level of filtration unit. For each candidate diversion ratio and fan speed combination, the comprehensive purification efficiency index under the predicted load demand is calculated.

[0046] The energy consumption impact analysis is based on the fan speed and fan energy consumption curves, combined with the flow channel resistance characteristics corresponding to the candidate split ratios, to calculate the energy consumption impact value of each candidate combination.

[0047] Specifically, step S30 in the method is used for: A first candidate control combination is selected from the plurality of candidate diversion ratios and the plurality of candidate fan speeds, wherein the first candidate control combination includes a first candidate diversion ratio and a first candidate fan speed; The first candidate control combination and the predicted load demand are input into the filtering performance evaluation model to obtain the first filtering performance value. The first energy consumption impact value is obtained based on the first candidate fan speed and the preset fan energy consumption curve; Following the method of obtaining the first filtering efficiency value and the first energy consumption impact value of the first candidate control combination, the filtering efficiency value and energy consumption impact value of the remaining candidate control combinations are obtained, resulting in multiple filtering efficiency values ​​and multiple energy consumption impact values.

[0048] In this embodiment, firstly, the filtration efficiency evaluation model is constructed based on the series and parallel characteristics of multi-stage filtration units, and the air purification device is abstracted into a system including a pre-filter layer, a medium-efficiency filter layer, and a high-efficiency filter layer.

[0049] For a given first candidate flow split ratio, the model first calculates the actual processing air volume of each filtration unit. The flow split ratio directly determines the airflow distribution ratio entering each layer, while the first candidate fan speed determines the total air volume supply capacity. The particulate matter removal efficiency of each filtration unit is described using a graded efficiency model, which establishes functional relationships between penetration rate and filtration velocity and dust load for PM2.5, PM10, and gaseous pollutants, respectively. Excessive filtration velocity leads to an increase in particulate penetration rate, while an increase in dust load causes the filtration efficiency to exhibit a non-linear change of first increasing and then decreasing. After outputting the flow split efficiency of each filtration unit for the target pollutant, the overall purification efficiency index, i.e., the first filtration efficiency value, is calculated by combining series and parallel data.

[0050] Specifically, the primary filter layer mainly targets large particulate pollutants such as PM10, the medium-efficiency filter layer focuses on intercepting fine particulate matter such as PM2.5, and the high-efficiency filter layer is responsible for the deep purification of ultrafine particulate matter and gaseous pollutants. The overall filtration efficiency value is obtained by weighting the efficiency index of each layer according to the processing air volume.

[0051] The preset fan energy consumption curve is based on the fan similarity law, which expresses the fan energy consumption as a cubic function of the rotational speed. At the same time, a resistance correction coefficient corresponding to the split ratio is introduced. The first candidate split ratio determines the path combination of the airflow through each stage of the filter unit. The flow resistance of different paths is different. The pre-filter bias strategy has lower resistance due to the use of a low-resistance primary filter layer, while the post-filter high-efficiency bias strategy has increased resistance due to the dense structure of the high-efficiency filter material.

[0052] Specifically, the product of resistance and air volume is taken as the key influencing factor of fan energy consumption. The actual power consumption is calculated by combining the fan efficiency curve to obtain the first energy consumption impact value. After traversing all candidate control combinations, a filter efficiency value matrix and an energy consumption impact value matrix corresponding to the candidate space dimension are formed.

[0053] The steps for constructing the filtration performance evaluation model include: Based on historical control data of the same model of air purifier, a set of sample control parameters and a set of sample load requirements were collected. The filtration efficiency was obtained by testing under different sample control parameters and sample load requirements. The set of sample filtration efficiency values ​​was obtained by labeling the filtration efficiency. A filtration efficiency evaluation model architecture is constructed, wherein the input features of the filtration efficiency evaluation model architecture are control parameters and load requirements, and the output feature is the filtration efficiency value; The filter performance evaluation model architecture is trained under supervision using the sample control parameter set, sample load requirement set, and sample filter performance value set until convergence, thus obtaining the trained filter performance evaluation model.

[0054] In this embodiment, firstly, the collection of sample control parameter sets and sample load requirement sets covers the entire operating range of the equipment, including measured data under different pollution concentration backgrounds, different airflow levels, and different degrees of filter unit aging. The filtration efficiency is tested in a standard experimental chamber, using an aerosol generator to produce test particles of known concentration. The actual filtration efficiency is calculated by measuring the concentration difference between upstream and downstream sampling points. The purification capacity for gaseous pollutants is determined using the tracer gas decay method, while the microbial inactivation efficiency is evaluated using a standard strain challenge experiment.

[0055] The labeling of the sample filtration efficiency value set adopts a multi-index fusion method, which combines particulate matter removal efficiency, gaseous pollutant purification rate and microbial inactivation log reduction value into a comprehensive efficiency index ranging from 0 to 100 according to the application scenario weight. Among them, medical applications focus on the microbial index weight, industrial purification scenarios focus on the particulate matter index weight, and civil buildings adopt a balanced weight configuration.

[0056] The filtration efficiency assessment model architecture employs an ensemble learning approach, with the bottom layer consisting of three parallel sub-networks corresponding to particulate matter filtration efficiency prediction, gaseous pollutant adsorption capacity prediction, and microbial inactivation efficiency prediction, respectively. Each sub-network uses a deep neural network structure. The input layer receives a concatenated vector of control parameters and load requirements, including a three-dimensional vector of the shunting ratio, a scalar of fan speed, a scalar of dust accumulation load for each stage of the filtration unit, and a vector of predicted pollutant concentration. The hidden layers employ a four-layer fully connected structure with 128, 256, 128, and 64 neurons respectively. The ReLU activation function and batch normalization layers are used to accelerate convergence.

[0057] After the sub-network output layers generate predicted values ​​for the corresponding performance dimensions, they are input into the top-level fusion network. The fusion network adopts a learnable attention weight mechanism to dynamically adjust the contribution weights of each performance dimension according to the current load demand characteristics, and finally outputs the filtering performance value.

[0058] Specifically, the model training uses the mean squared error loss function, the optimizer is Adam, the initial learning rate is 0.001, a cosine annealing learning rate scheduling strategy is introduced, the training batch size is 128, the total training epochs are 300, and training stops when the overall performance prediction error on the validation set is less than 5%, thus obtaining the completed filtering performance evaluation model.

[0059] S40: Generate filtration efficiency weight and energy consumption impact weight based on the current operating conditions, and combine the multiple filtration efficiency values ​​and the multiple energy consumption impact values ​​to iteratively optimize multiple candidate control combinations to obtain the optimal control parameter combination, and control the target air purification device based on the optimal control parameter combination.

[0060] In this embodiment, the trade-off between filtration efficiency and energy consumption is dynamically adjusted based on the current operating conditions to achieve adaptive decision-making for multi-objective optimization. The determination of the current operating conditions comprehensively considers multiple factors such as indoor and outdoor pollutant concentration levels, personnel activity status, time period characteristics, and user preference settings.

[0061] The iterative optimization process employs an improved multi-objective particle swarm optimization algorithm, which maps candidate control combinations to particle position vectors. Each particle represents a set of feasible flow splitting ratios and fan speed configurations. The optimal control parameter combination is obtained through iterative optimization, and the target air purification equipment is controlled based on this.

[0062] Specifically, the system monitors the differential pressure change rate and particulate matter penetration rate of each filtration unit in real time. When the differential pressure change rate of any filtration unit exceeds a preset threshold or the particulate matter penetration rate exceeds a preset standard, a filter unit replacement warning is triggered. Then, based on the cumulative operating time and load prediction results of each filter unit, a filter unit remaining life prediction report is generated. The differential pressure change rate reflects the change in the inlet and outlet pressure difference of the filter unit per unit time, indicating the dynamic trend of filter media clogging rate. The particulate matter penetration rate is calculated through real-time comparison of upstream and downstream concentration monitoring points, directly characterizing the degree of degradation in the actual interception capacity of the filter unit.

[0063] Among them, the filtration efficiency weight and energy consumption impact weight are generated based on the current operating conditions, including: Obtain the current particulate matter concentration and compare it with a preset concentration threshold. If the current particulate matter concentration is higher than the first preset threshold, calculate and adjust the filtration efficiency weight. If the current particulate matter concentration is lower than the second preset threshold, the weight of the energy consumption adjustment is calculated. If the current particulate matter concentration is between the first preset threshold and the second preset threshold, the weight ratio will be dynamically adjusted according to the concentration change trend. The sum of the energy consumption impact weight and the filtration efficiency weight is 1.

[0064] In this embodiment, the preset concentration threshold is first set based on the air quality index classification standard. The first preset threshold corresponds to the moderate pollution level, and the second preset threshold corresponds to the good level.

[0065] When the real-time monitored particulate matter concentration exceeds the first preset threshold, it indicates that the current air quality is poor and the health risk to personnel is increased. At this time, the weight of filtration efficiency is increased by calculation, that is, filtration efficiency weight = (current particulate matter concentration / first preset threshold) × preset filtration efficiency weight, and energy consumption impact weight = 1 - filtration efficiency weight. The weight coefficient is tilted towards the purification efficiency direction. Usually, the filtration efficiency weight is set in the range of 0.7 to 0.9, and the energy consumption impact weight is correspondingly compressed to the range of 0.1 to 0.3 to ensure that the selection of candidate control combinations prioritizes meeting the high purification efficiency requirements.

[0066] When the particulate matter concentration is below the second preset threshold, the air quality is good and the purification demand is relatively relaxed. The weight of energy consumption is increased and the weight coefficient is adjusted towards energy-saving operation. The weight of filtration efficiency is reduced to the range of 0.2 to 0.4, and the weight of energy consumption is increased to the range of 0.6 to 0.8. Under the premise of ensuring basic purification effect, the goal is to pursue operational economy.

[0067] For the transitional range where the concentration is between the two threshold levels, a concentration change trend analysis mechanism is introduced to calculate the slope of the concentration change over the past ten minutes. If an upward trend is observed, it is predicted that the pollution will worsen, and the weights will be adjusted in advance towards filtration efficiency. If a downward trend is observed, it is predicted that the air quality will improve, and the weights will be gradually adjusted towards energy consumption optimization to avoid drastic fluctuations in weight switching and achieve smooth adaptive adjustment.

[0068] Furthermore, by combining the multiple filtering efficiency values ​​and the multiple energy consumption impact values, the multiple candidate control combinations are iteratively optimized to obtain the optimal control parameter combination, including: Multiple candidate diversion ratios and multiple candidate wind turbine speeds were used as the initial population individuals; Using the weighted comprehensive fitness of filtering efficiency and energy consumption impact as the optimization objective, an improved particle swarm optimization algorithm is used to iteratively update the population until the convergence condition is met. The improved particle swarm optimization algorithm performs the following steps in each iteration: The similarity threshold and the proportion of combined analysis are dynamically determined based on the current particulate matter concentration. According to the combined analysis ratio, multiple particles are randomly selected from the current population, and the selected particles are combined in pairs to calculate the similarity between the two particles in each pair. Select similar particle combinations with a similarity greater than or equal to the similarity threshold. For each similar particle combination, compare the fitness values ​​of the two particles and eliminate particles with low fitness. Based on the current filtering efficiency weight and energy consumption impact weight, as well as the weighted comprehensive fitness value of the eliminated particles, the search step size of the retained particles and the number of new solutions generated in subsequent iterations are dynamically adjusted. Specifically, when the filtering efficiency weight is greater than the energy consumption impact weight, the reduction in the search step size of the retained particles is positively correlated with the efficiency fitness component of the eliminated particles; when the energy consumption impact weight is greater than the filtering efficiency weight, the increase in the number of new solutions generated by the retained particles is positively correlated with the energy consumption fitness component of the eliminated particles. The retained particles are iterated according to the velocity and position update formula of the particle swarm optimization algorithm, and the adjusted search step size and the number of new solutions generated are introduced. The optimal individual obtained when the convergence condition is met is taken as the optimal combination of control parameters.

[0069] In this embodiment, the initial population is first generated using the Latin hypercube sampling method, which uniformly covers the three-dimensional feasible region of the flow split ratio and the scalar range of the wind turbine speed to ensure the diversity of the initial solution space. The population size is set to 50 to 100 particles, dynamically adjusted according to the problem complexity. The position vector of each particle is encoded as a four-dimensional real vector, with the first three dimensions corresponding to the component values ​​of the flow split ratio and the fourth dimension corresponding to the normalized value of the wind turbine speed. All components are normalized within the range of 0 to 1, and are mapped to the physical feasible region in practical applications.

[0070] Secondly, the weighted fitness function is defined as a linear weighted combination of the filtering efficiency value and the energy consumption impact value, i.e., fitness value = filtering efficiency weight × normalized filtering efficiency value + energy consumption impact weight × (1 - normalized energy consumption impact value). The normalization process scales based on the extreme values ​​within the candidate space, ensuring that the two metrics with different dimensions are on comparable orders of magnitude. The improved particle swarm optimization algorithm introduces an adaptive diversity control strategy based on the traditional velocity and position update mechanism to address the premature convergence problem common in multi-objective optimization.

[0071] Furthermore, the dynamic determination mechanism of the similarity threshold is negatively correlated with the current particulate matter concentration. When the concentration is high, the similarity threshold is lowered to the range of 0.6 to 0.7 to relax the particle elimination conditions and retain more high-fitness individuals to participate in fine search. When the concentration is low, the similarity threshold is raised to the range of 0.8 to 0.9 to strengthen the population diversity constraint and prevent excessive shrinkage of the solution space.

[0072] Secondly, the proportion of combinatorial analysis is set to 20% to 40% of the current population size. After randomly selecting particles and combining them in pairs, the similarity between particles is calculated using Euclidean distance. For combinations with similarity exceeding the threshold, fitness comparison and elimination operations are performed.

[0073] Specifically, the Euclidean distance is calculated based on the four-dimensional components of the particle position vector. The smaller the distance, the closer the two candidate control combinations are in terms of flow splitting ratio configuration and fan speed setting. When the Euclidean distance is less than or equal to a preset threshold, they are judged as similar particle combinations.

[0074] For example, to calculate the similarity between two particles, assume that the position vector of particle A is [0.35, 0.40, 0.25, 0.60] and the position vector of particle B is [0.38, 0.42, 0.20, 0.65], where the first three dimensions correspond to the flow ratio components of the primary, medium, and high efficiency layers, respectively, and the fourth dimension corresponds to the normalized value of the fan speed.

[0075] First, calculate the sum of squares of the four-dimensional component differences: the difference in the initial efficiency layer is 0.35 - 0.38 = -0.03, with a square of 0.0009; the difference in the medium efficiency layer is 0.40 - 0.42 = -0.02, with a square of 0.0004; the difference in the high efficiency layer is 0.25 - 0.20 = 0.05, with a square of 0.0025; and the difference in the fan speed is 0.60 - 0.65 = -0.05, with a square of 0.0025. The cumulative sum of squares is 0.0009 + 0.0004 + 0.0025 + 0.0025 = 0.0063. Then, calculate the square root of the Euclidean distance, which is approximately 0.0794. Since all components of the particle position vector are normalized to the interval between 0 and 1, the maximum possible distance is the diagonal length of the four-dimensional unit vector, which is 2. Therefore, the similarity can be defined as 1 minus the normalized Euclidean distance, that is, similarity = 1 - 0.0794 / 2 ≈ 0.9603.

[0076] Furthermore, the dynamic adjustment of the search step size follows the principle of balancing refinement and exploration. When the filtering efficiency weight is dominant, for individuals with high filtering efficiency among the eliminated particles, the search step size of the retained particles is reduced to 50% to 70% of the standard step size, guiding the algorithm to conduct in-depth mining in the neighborhood of the current high-quality solution. When the energy consumption influence weight is dominant, for individuals with high energy consumption fitness among the eliminated particles, the number of new solutions generated by the retained particles is increased to 1.5 to 2 times the standard number, expanding the search range to discover better energy-saving configurations.

[0077] The new solution generation employs a Gaussian perturbation mechanism. Candidate positions are sampled near the retained particle positions using an adjusted number of samples. After feasibility testing, these are incorporated into the next generation of the population. The Gaussian perturbation mechanism involves superimposing a random perturbation vector following a Gaussian distribution onto the particle's current position. The standard deviation of the perturbation amplitude is proportional to the current search step size, ensuring that the newly generated solution is both within the neighborhood of a high-quality solution and has a certain exploration range. Feasibility testing includes verification of the normalization constraints of the three-dimensional components of the split ratio (the sum of the three components must equal 1) and checks of the physical upper and lower limits of the wind turbine speed. Solutions that do not meet the constraints are projected to the nearest feasible boundary or resampled and generated.

[0078] Furthermore, the velocity-position update formula introduces a linearly decreasing inertia weight strategy, with an initial inertia weight set to 0.9 and a final value set to 0.4. During iteration, the global exploration capability is gradually reduced while the local exploitation capability is enhanced. The cognitive coefficient and social coefficient are set to 2.0 and 2.0 respectively, ensuring that particles can utilize both individual historical best experience and global best information from the population. A dual criterion is used to determine convergence: when the improvement in global best fitness is less than 0.1% within 20 consecutive generations, and the population diversity index is below a preset threshold, the algorithm is considered converged, and the decoded control parameters corresponding to the current global best particle are output as the optimal control parameter combination.

[0079] Furthermore, the target air purification equipment is controlled based on the optimal combination of control parameters. This includes resolving the optimal flow distribution ratio into flow allocation commands for each stage of the filter unit, converting the optimal fan speed into a variable frequency drive signal, and executing real-time adjustments through the equipment controller. The issuance of control commands adopts a hierarchical confirmation mechanism. First, the feasibility of the control parameter combination within the equipment's safe operation constraints is verified, including upper and lower limit protection for fan speed, over-limit protection for filter unit differential pressure, and motor overload protection. After successful verification, parameter updates are performed, and the deviation between the actual operating status and the expected target is monitored in the next control cycle. When the deviation exceeds the allowable range, a re-optimization process is triggered.

[0080] In summary, compared with existing technologies, this application introduces a dynamic similarity threshold and a combinatorial analysis ratio into the optimization algorithm, and adjusts the search step size and the number of new solutions generated according to the weight relationship, thereby improving the optimization efficiency and the quality of the solution, enabling the air purification equipment to always maintain the optimal working state in complex and ever-changing actual operating environments.

[0081] In summary, the embodiments of this application have at least the following technical effects: This application provides a method for dynamically optimizing the multi-stage filtration efficiency of an air purification device. First, a complete device status perception system is constructed by real-time collection of multi-dimensional operating parameters such as the pressure difference before and after each filtration unit, particulate matter concentration, and cumulative operating time. Second, based on a machine learning model, the filtration load is predicted, and the pollutant accumulation trend and efficiency decay curve of each filtration unit are determined, achieving a shift from a passive response to an active prevention control mode. Third, by generating multiple candidate diversion ratios and fan speed combinations and conducting a dual analysis of their filtration efficiency and energy consumption impact, a comprehensive control strategy evaluation framework is established. Finally, the filtration efficiency weight and energy consumption impact weight are dynamically adjusted according to the current operating conditions, and an improved particle swarm optimization algorithm is used for iterative optimization, effectively reducing energy consumption while ensuring purification effect, achieving an adaptive balance between filtration efficiency and energy consumption.

[0082] Through the above technical solution, this application overcomes the technical defects of the existing air purification equipment, which has fixed and rigid operating parameters and cannot adapt to dynamic changes in pollution load, and realizes the coordinated optimization control of multi-stage filtration units under different operating conditions.

[0083] Example 2, as Figure 2 As shown, based on the same inventive concept as the multi-stage filtration efficiency dynamic optimization method for an air purification device provided in Embodiment 1, this application also provides a multi-stage filtration efficiency dynamic optimization system for an air purification device, including: The data acquisition module 11 is used to respond to the device start signal and acquire the real-time operating parameters of the target air purification device, wherein the real-time operating parameters include the pressure difference before and after each stage of the filter unit, the particulate matter concentration and the cumulative operating time. The load forecasting module 12 is used to perform filtered load forecasting based on the real-time operating parameters, obtain the predicted load demand, and generate multiple candidate diversion ratios and multiple candidate fan speeds according to the predicted load demand. Parameter calculation module 13 is used to perform filtration efficiency analysis and energy consumption impact analysis on the multiple candidate diversion ratios and multiple candidate fan speeds, and obtain multiple filtration efficiency values ​​and multiple energy consumption impact values. The parameter optimization module 14 is used to generate filtration efficiency weights and energy consumption impact weights based on the current operating conditions, and to iteratively optimize multiple candidate control combinations by combining the multiple filtration efficiency values ​​and the multiple energy consumption impact values ​​to obtain the optimal control parameter combination, and to control the target air purification device based on the optimal control parameter combination.

[0084] In one embodiment, based on the real-time operating parameters, filter load prediction is performed to obtain the predicted load demand of each level of filter unit, including: Obtain a preset feature extraction template, and extract features from the real-time running parameters based on the preset feature extraction template to obtain the current running feature vector; The current operating feature vector is input into the filter load prediction model to obtain the predicted load demand of each level of filter unit. The filter load prediction model is a machine learning model trained based on historical operating data, which is used to predict the pollutant accumulation trend and filtration efficiency decay curve of each level of filter unit in the future period based on the current operating characteristics.

[0085] Furthermore, the steps for constructing the filtered load prediction model include: Collect historical operation datasets of air purifiers of the same model, including historical real-time operation parameters and corresponding historical load change labels; An initial load forecasting model architecture is constructed, which is a regression model that combines time-series features and multi-dimensional parameter inputs; Using the historical operational dataset as training samples and the historical load change labels as supervision targets, the initial load prediction model architecture is trained until convergence, thus obtaining the trained filtered load prediction model.

[0086] Furthermore, in one embodiment of the application, generating multiple candidate diversion ratios and multiple candidate turbine speeds based on the predicted load demand includes: Activate the control policy generator, which includes multiple policy configuration units; The predicted load demand is input into the control strategy generator, which calls multiple strategy configuration units to configure control parameters for the predicted load demand, thereby obtaining a candidate combination set of multiple candidate diversion ratios and multiple candidate fan speeds.

[0087] Furthermore, the construction steps of the control policy generator include: Collect multiple sample load demand data, construct a sample load dataset, and label each sample load data in the sample load dataset to obtain a sample control parameter label set; Build a multi-strategy configuration unit architecture; Using the sample load dataset as input and the sample control parameter annotation set as the supervision target, the architecture of the multiple policy configuration units is trained synchronously to obtain multiple policy configuration units; The multiple policy configuration units are integrated to obtain the control policy generator.

[0088] In one embodiment, the parameter calculation module 13 is specifically used for: A first candidate control combination is selected from the plurality of candidate diversion ratios and the plurality of candidate fan speeds, wherein the first candidate control combination includes a first candidate diversion ratio and a first candidate fan speed; The first candidate control combination and the predicted load demand are input into the filtering performance evaluation model to obtain the first filtering performance value. The first energy consumption impact value is obtained based on the first candidate fan speed and the preset fan energy consumption curve; Following the method of obtaining the first filtering efficiency value and the first energy consumption impact value of the first candidate control combination, the filtering efficiency value and energy consumption impact value of the remaining candidate control combinations are obtained, resulting in multiple filtering efficiency values ​​and multiple energy consumption impact values.

[0089] Furthermore, the steps for constructing the filtration performance evaluation model include: Based on historical control data of the same model of air purifier, a set of sample control parameters and a set of sample load requirements were collected. The filtration efficiency was obtained by testing under different sample control parameters and sample load requirements. The set of sample filtration efficiency values ​​was obtained by labeling the filtration efficiency. A filtration efficiency evaluation model architecture is constructed, wherein the input features of the filtration efficiency evaluation model architecture are control parameters and load requirements, and the output feature is the filtration efficiency value; The filter performance evaluation model architecture is trained under supervision using the sample control parameter set, sample load requirement set, and sample filter performance value set until convergence, thus obtaining the trained filter performance evaluation model.

[0090] Furthermore, in one embodiment of the application, generating filtration efficiency weights and energy consumption impact weights based on the current operating conditions includes: Obtain the current particulate matter concentration and compare it with a preset concentration threshold. If the current particulate matter concentration is higher than the first preset threshold, calculate and adjust the filtration efficiency weight. If the current particulate matter concentration is lower than the second preset threshold, the weight of the energy consumption adjustment is calculated. If the current particulate matter concentration is between the first preset threshold and the second preset threshold, the weight ratio will be dynamically adjusted according to the concentration change trend. The sum of the energy consumption impact weight and the filtration efficiency weight is 1.

[0091] Furthermore, by combining the multiple filtering efficiency values ​​and the multiple energy consumption impact values, the multiple candidate control combinations are iteratively optimized to obtain the optimal control parameter combination, including: Multiple candidate diversion ratios and multiple candidate wind turbine speeds were used as the initial population individuals; Using the weighted comprehensive fitness of filtering efficiency and energy consumption impact as the optimization objective, an improved particle swarm optimization algorithm is used to iteratively update the population until the convergence condition is met. The improved particle swarm optimization algorithm performs the following steps in each iteration: The similarity threshold and the proportion of combined analysis are dynamically determined based on the current particulate matter concentration. According to the combined analysis ratio, multiple particles are randomly selected from the current population, and the selected particles are combined in pairs to calculate the similarity between the two particles in each pair. Select similar particle combinations with a similarity greater than or equal to the similarity threshold. For each similar particle combination, compare the fitness values ​​of the two particles and eliminate particles with low fitness. Based on the current filtering efficiency weight and energy consumption impact weight, as well as the weighted comprehensive fitness value of the eliminated particles, the search step size of the retained particles and the number of new solutions generated in subsequent iterations are dynamically adjusted. Specifically, when the filtering efficiency weight is greater than the energy consumption impact weight, the reduction in the search step size of the retained particles is positively correlated with the efficiency fitness component of the eliminated particles; when the energy consumption impact weight is greater than the filtering efficiency weight, the increase in the number of new solutions generated by the retained particles is positively correlated with the energy consumption fitness component of the eliminated particles. The retained particles are iterated according to the velocity and position update formula of the particle swarm optimization algorithm, and the adjusted search step size and the number of new solutions generated are introduced. The optimal individual obtained when the convergence condition is met is taken as the optimal combination of control parameters.

Claims

1. A method for dynamically optimizing the multi-stage filtration efficiency of an air purification device, characterized in that, The method includes: In response to the device start signal, the real-time operating parameters of the target air purification device are obtained, wherein the real-time operating parameters include the pressure difference before and after each stage of the filter unit, the particulate matter concentration, and the cumulative operating time. Based on the real-time operating parameters, the filter load is predicted to obtain the predicted load demand, and multiple candidate diversion ratios and multiple candidate fan speeds are generated according to the predicted load demand. Filter efficiency analysis and energy consumption impact analysis are performed on the multiple candidate diversion ratios and multiple candidate fan speeds to obtain multiple filter efficiency values ​​and multiple energy consumption impact values. Based on the current operating conditions, filter efficiency weights and energy consumption impact weights are generated. Combining the multiple filter efficiency values ​​and the multiple energy consumption impact values, multiple candidate control combinations are iteratively optimized to obtain the optimal control parameter combination. The target air purification device is then controlled based on the optimal control parameter combination.

2. The method according to claim 1, characterized in that, Based on the real-time operating parameters, the filtration load is predicted to obtain the predicted load demand of each filtration unit, including: Obtain a preset feature extraction template, and extract features from the real-time running parameters based on the preset feature extraction template to obtain the current running feature vector; The current operating feature vector is input into the filter load prediction model to obtain the predicted load demand of each level of filter unit. The filter load prediction model is a machine learning model trained based on historical operating data, which is used to predict the pollutant accumulation trend and filtration efficiency decay curve of each level of filter unit in the future period based on the current operating characteristics.

3. The method according to claim 2, characterized in that, The steps for constructing the filter load prediction model include: Collect historical operation datasets of air purifiers of the same model, including historical real-time operation parameters and corresponding historical load change labels; An initial load forecasting model architecture is constructed, which is a regression model that combines time-series features and multi-dimensional parameter inputs; Using the historical operational dataset as training samples and the historical load change labels as supervision targets, the initial load prediction model architecture is trained until convergence, thus obtaining the trained filtered load prediction model.

4. The method according to claim 1, characterized in that, Based on the predicted load demand, multiple candidate diversion ratios and multiple candidate fan speeds are generated, including: Activate the control policy generator, which includes multiple policy configuration units; The predicted load demand is input into the control strategy generator, which calls multiple strategy configuration units to configure control parameters for the predicted load demand, thereby obtaining a candidate combination set of multiple candidate diversion ratios and multiple candidate fan speeds.

5. The method according to claim 4, characterized in that, The construction steps of the control strategy generator include: Collect multiple sample load demand data, construct a sample load dataset, and label each sample load data in the sample load dataset to obtain a sample control parameter label set; Build a multi-strategy configuration unit architecture; Using the sample load dataset as input and the sample control parameter annotation set as the supervision target, the architecture of the multiple policy configuration units is trained synchronously to obtain multiple policy configuration units; The multiple policy configuration units are integrated to obtain the control policy generator.

6. The method according to claim 1, characterized in that, Filter efficiency analysis and energy consumption impact analysis were performed on the multiple candidate diversion ratios and multiple candidate fan speeds to obtain multiple filter efficiency values ​​and multiple energy consumption impact values, including: A first candidate control combination is selected from the plurality of candidate diversion ratios and the plurality of candidate fan speeds, wherein the first candidate control combination includes a first candidate diversion ratio and a first candidate fan speed; The first candidate control combination and the predicted load demand are input into the filtering performance evaluation model to obtain the first filtering performance value. The first energy consumption impact value is obtained based on the first candidate fan speed and the preset fan energy consumption curve; Following the same method used to obtain the first filtering efficiency value and the first energy consumption impact value of the first candidate control combination, the filtering efficiency value and energy consumption impact value of the remaining candidate control combinations are obtained, resulting in multiple filtering efficiency values ​​and multiple energy consumption impact values.

7. The method according to claim 6, characterized in that, The steps for constructing the filtration performance evaluation model include: Based on historical control data of the same model of air purifier, a set of sample control parameters and a set of sample load requirements were collected. The filtration efficiency was tested under different sample control parameters and sample load requirements. The set of sample filtration efficiency values ​​was obtained based on the filtration efficiency label. A filtration efficiency evaluation model architecture is constructed, wherein the input features of the filtration efficiency evaluation model architecture are control parameters and load requirements, and the output feature is the filtration efficiency value; The filter performance evaluation model architecture is trained under supervision using the sample control parameter set, sample load requirement set, and sample filter performance value set until convergence, thus obtaining the trained filter performance evaluation model.

8. The method according to claim 1, characterized in that, Based on the current operating conditions, filter efficiency weights and energy consumption impact weights are generated, including: Obtain the current particulate matter concentration and compare it with a preset concentration threshold. If the current particulate matter concentration is higher than the first preset threshold, calculate and adjust the filtration efficiency weight. If the current particulate matter concentration is lower than the second preset threshold, the weight of the energy consumption adjustment is calculated. If the current particulate matter concentration is between the first preset threshold and the second preset threshold, the weight ratio will be dynamically adjusted according to the concentration change trend. The sum of the energy consumption impact weight and the filtration efficiency weight is 1.

9. The method according to claim 8, characterized in that, By combining the multiple filtering efficiency values ​​and the multiple energy consumption impact values, multiple candidate control combinations are iteratively optimized to obtain the optimal control parameter combination, including: Multiple candidate diversion ratios and multiple candidate wind turbine speeds were used as the initial population individuals; Using the weighted comprehensive fitness of filtering efficiency and energy consumption impact as the optimization objective, an improved particle swarm optimization algorithm is used to iteratively update the population until the convergence condition is met. The improved particle swarm optimization algorithm performs the following steps in each iteration: The similarity threshold and the proportion of combined analysis are dynamically determined based on the current particulate matter concentration. According to the combined analysis ratio, multiple particles are randomly selected from the current population, and the selected particles are combined in pairs to calculate the similarity between the two particles in each pair. Select similar particle combinations with a similarity greater than or equal to the similarity threshold. For each similar particle combination, compare the fitness values ​​of the two particles and eliminate particles with low fitness. Based on the current filtering efficiency weight and energy consumption impact weight, as well as the weighted comprehensive fitness value of the eliminated particles, the search step size of the retained particles and the number of new solutions generated in subsequent iterations are dynamically adjusted. Specifically, when the filtering efficiency weight is greater than the energy consumption impact weight, the reduction in the search step size of the retained particles is positively correlated with the efficiency fitness component of the eliminated particles; when the energy consumption impact weight is greater than the filtering efficiency weight, the increase in the number of new solutions generated by the retained particles is positively correlated with the energy consumption fitness component of the eliminated particles. The retained particles are iterated according to the velocity and position update formula of the particle swarm optimization algorithm, and the adjusted search step size and the number of new solutions generated are introduced. The optimal individual obtained when the convergence condition is met is taken as the optimal combination of control parameters.

10. A multi-stage filtration efficiency dynamic optimization system for an air purification device, characterized in that, A method for dynamically optimizing the multi-stage filtration efficiency of an air purification device according to any one of claims 1-9 includes: The data acquisition module is used to respond to the device start signal and acquire the real-time operating parameters of the target air purification device, wherein the real-time operating parameters include the pressure difference before and after each stage of the filter unit, the particulate matter concentration, and the cumulative operating time. The load forecasting module is used to perform filtered load forecasting based on the real-time operating parameters, obtain the predicted load demand, and generate multiple candidate diversion ratios and multiple candidate fan speeds according to the predicted load demand. The parameter calculation module is used to perform filtration efficiency analysis and energy consumption impact analysis on the multiple candidate diversion ratios and multiple candidate fan speeds, and obtain multiple filtration efficiency values ​​and multiple energy consumption impact values. The parameter optimization module is used to generate filtration efficiency weights and energy consumption impact weights based on the current operating conditions, and to iteratively optimize multiple candidate control combinations by combining the multiple filtration efficiency values ​​and the multiple energy consumption impact values ​​to obtain the optimal control parameter combination, and to control the target air purification device based on the optimal control parameter combination.

Images