An air purification detection method and system based on deep learning

Abstract

The present invention proposes an air purification detection method and system based on deep learning, the method comprising: deploying infrared optical gas detection sensors around molecular sieve adsorbers, collecting multi-dimensional data, pre-processing and randomly dividing the data into training sets and test sets; constructing an air purification detection model, pre-training the air purification detection model using the training set; detecting the purified air using the optimized air purification detection model; constructing an adaptive adjustment model to intelligently adjust the operation of the molecular sieve adsorber; constructing an air purification prediction model, using the air purification prediction model to predict the future working state and adsorption efficiency of the molecular sieve adsorber and the failures that will occur in the future air purification process, and feeding back the corresponding prediction results to the operator or the control system. The present invention realizes the intelligent monitoring and optimization of the air purification process, and significantly improves the performance, stability and prediction ability of the air purification system.

Description

Technical Field

[0001] The present invention belongs to the technical field of air purification detection, and in particular, relates to an air purification detection method and system based on deep learning. Background Art

[0002] Air detection sensors may be affected by environmental factors such as temperature, humidity, background gases, etc., resulting in inaccurate readings. During long-term operation, sensors may age or become contaminated, affecting their detection performance.

[0003] Traditional air monitoring systems may not have real-time monitoring capabilities, or the transmission and processing delays of monitoring data are long, which affects the rapid response to sudden pollution events. High-quality air monitoring equipment and maintenance costs are high, which may become an economic burden for large-scale deployment and long-term operation.

[0004] Existing systems may lack sufficient flexibility and modular design, making them difficult to integrate with other environmental monitoring or control systems, limiting the overall effectiveness and scope of application of the system. Summary of the invention

[0005] The purpose of the present invention is to propose an air purification detection method and system based on deep learning, which integrates an advanced deep neural network model and can efficiently analyze the changes in the concentrations of various pollutants in the air before and after treatment by air purification equipment, covering data collection, model training and optimization, real-time detection, adaptive adjustment and prediction functions, thereby forming a closed-loop intelligent system.

[0006] In order to achieve the above object, in a first aspect of the present invention, a deep learning-based air purification detection method is provided, the method comprising:

[0007] S1. deploying an infrared optical gas detection sensor around the molecular sieve adsorber, using the infrared optical gas detection sensor to collect multi-dimensional data, pre-processing the collected data, and randomly dividing the pre-processed data into a training set and a test set; wherein the multi-dimensional data includes temperature, pressure and humidity;

[0008] S2. Build an air purification detection model based on a deep belief network, and use the training set to pre-train the air purification detection model;

[0009] S3. Use the test set to optimize the pre-trained air purification detection model;

[0010] S4, collecting real-time multi-dimensional data and using the optimized air purification detection model to detect the purified air;

[0011] S5. Construct an adaptive adjustment model based on reinforcement learning according to the detection results, use the adaptive adjustment model to intelligently adjust the operation of the molecular sieve adsorber, and feed back the detection report and alarm to the operator or control system when the upper limit is adjusted;

[0012] S6. Construct an air purification prediction model based on machine learning based on the historical multidimensional data and the detection results of the air purification detection model, and use the air purification prediction model to predict the future working state and adsorption efficiency of the molecular sieve adsorber and the failures that may occur in the future air purification process, and feed back the corresponding prediction results to the operator or the control system;

[0013] The air purification detection model includes several layers of restricted Boltzmann machines and back propagation networks, wherein the restricted Boltzmann machine includes a visible layer and a hidden layer, m visible layer units v constitute the visible layer, and n hidden layer units h constitute the hidden layer; the energy model of the restricted Boltzmann is expressed as follows:

[0014]

[0015] Among them, E(v,h) represents the energy between the visible layer unit v and the hidden layer unit h, w ij Represents the visual layer unit v i and the hidden layer unit h i The connection weights in the middle, I represents the number of nodes in the visible layer, J represents the number of nodes in the hidden layer, and a i and b j denote the unit thresholds of the visible layer unit and the hidden layer unit, respectively, γ is the regularization strength parameter, ||W|| 2 represents the square of the Frobenius norm of the weight matrix W, that is, the sum of the squares of all weights; where W = I × J;

[0016] The joint probability function constructed according to the restricted Boltzmann energy model is expressed as follows:

[0017]

[0018] Where Z represents the partition function, and Z = ∑ v ∑ h exp(-E(v,h));

[0019] The conditional probability function is:

[0020]

[0021] Here, sigmoid represents the Sigmoid function.

[0022] Further, in S1, the preprocessing is represented as follows:

[0023] Assume that N infrared optical gas detection sensors are deployed around the molecular sieve adsorber, and each sensor measures the gas concentration C(t) at time point t. Let the reading of the jth sensor around the i-th adsorber at time point t be c ij (t), then for the adsorber, it is expressed as two matrices:

[0024] C1(t)=[c 11 (t),c 12 (t)…c 1N (t)],C2(t)=[c 21 (t),c 22 (t)…c 2N (t)]

[0025] The output of the infrared optical gas detection sensor needs to be converted into the actual gas concentration value. Assuming that the output signal of the sensor is S, the relationship between it and the gas concentration C is C = f(S), where f is the calibration function;

[0026] Data cleaning removes invalid or abnormal data points, and defines C'(t) to represent the cleaned data, which is expressed as follows:

[0027]

[0028] Among them, k represents the abnormal threshold, μ ij represents the mean, σ ij represents the standard deviation, and NaN represents invalid or abnormal data points;

[0029] The cleaned data is standardized as follows:

[0030]

[0031] Wherein, C”(t) represents the data set after standardization.

[0032] Furthermore, the visual layer unit v i and the hidden layer unit h i The connection weight w in the middle ij The update rule is as follows:

[0033] Δw ij =∈[ <v i h j > data - <v i h j >model]-2γ∈w ij

[0034] Among them, w ijRepresents the contribution of the regularization term to the weight update, which prompts the connection weight to be adjusted in the direction of minimum. <v i h j >data represents the visible layer node v calculated on the training set i and hidden layer nodes h j The expected value of the joint probability distribution of <v i h j > model Indicates that under the current parameter settings of the model, for the same node pair v i and h j The expected value of the joint probability distribution of , ∈ represents the learning rate.

[0035] Furthermore, the air purification detection model uses an unsupervised greedy layer-by-layer method to obtain weights, and trains each layer of restricted Boltzmann machines from bottom to top; wherein, in the pre-training stage, the visual layer vector value is first mapped to the hidden layer, and then the original input signal is reconstructed using the hidden unit to re-establish the visual unit, and the newly re-established visual unit is once again mapped to the hidden layer as input to obtain a new hidden layer unit, and the process is repeated; wherein, in the fine-tuning process of the air purification detection model, the last layer of the air purification detection model is set as a back propagation network, and the back propagation network is used to fine-tune the entire air purification detection model, and the feature vector input by the back propagation network is the feature vector output by the restricted Boltzmann machine, and the model parameters of the air purification detection model are adjusted.

[0036] Furthermore, in the context of air purification detection, the detection result can be regarded as the response of the air purification detection model to the current input (i.e., multidimensional data in the air purification process, such as moisture, carbon dioxide concentration, air pressure, and hydrocarbon concentration). The air purification detection model evaluates the input data through its conditional probability function, infers the hidden features about the air purification state, and then gives the detection results, such as purification efficiency, whether there is a fault, etc.

[0037] Furthermore, in said S3, the steps specifically include:

[0038] S401, define the test set as T = {(x1,y1),(x2,y2),...,(x i ,y i )}, where x i Represents input data, y i Indicates the corresponding true label;

[0039] S402, using the absolute error as the loss function Loss for training the air purification detection model, expressed as follows:

[0040]

[0041] Where N is the number of samples in the test set. represents the predicted value;

[0042] S403, using the back propagation algorithm and the gradient descent method to minimize the loss function, as shown below:

[0043]

[0044] Among them, θ t+1 represents the updated model parameters, θ represents the model parameters, α represents the learning rate, Represents the gradient of the loss function with respect to the parameters.

[0045] Furthermore, the steps of constructing the adaptive adjustment model are as follows:

[0046] The feature extraction of the preprocessed real-time multidimensional data is shown as follows:

[0047]

[0048] Among them, F (x，y) represents the element of the output feature F in the i-th row and j-th column, W (x，y) represents the element of the convolution kernel W in the pth row and qth column, and the size of the convolution kernel is x×y, V x+p-1，y+q-1 represents the element of the input feature V in the x+p-1th row and y+q-1th column, b l Represents the bias term of convolutional layer l;

[0049] The extracted features are weighted and fused using the attention mechanism to generate a purified air recognition model; among them, the attention weight It is expressed as follows:

[0050]

[0051] Wherein, e represents the base of natural logarithm, which means exponential operation on Sim(x, y); Sim(x, y) represents the similarity between feature x and feature y, k represents the kth feature, and Sim(x, k) represents the similarity between feature x and feature k;

[0052] Use supervised learning methods to introduce rule constraints and design loss function L auto as follows:

[0053] L auto =L cl +L mi +λL re

[0054] Among them, L cl represents the classification loss, λ represents the hyperparameter, L re represents the rule constraint loss, Lmi represents the fusion loss.

[0055] Furthermore, the adaptive adjustment model also includes:

[0056] An adaptive threshold algorithm is introduced to adjust the temperature, pressure and humidity of the molecular sieve adsorber according to the historical data and real-time detection data of the molecular sieve adsorber. The adaptive threshold is:

[0057] T em =μ tem +k tem ×σ tem

[0058] Pr re =μ pre +k pre ×σ pre

[0059] H um =μ hum +k hum ×σ tem

[0060] Among them, T em represents the temperature threshold, μ tem and σ tem They represent the mean and standard deviation of the historical temperature data of the molecular sieve adsorber, k tem The temperature hyperparameter is used to adjust the sensitivity of the temperature threshold. pre represents the pressure threshold, μ pre and σ tem The mean and standard deviation of the historical pressure data of the molecular sieve adsorber, k pre The hyperparameter representing the pressure is used to adjust the sensitivity of the pressure threshold; H um Represents the humidity threshold, μ hum and σ hum The mean and standard deviation of the historical humidity data of the molecular sieve adsorber, k hum A hyperparameter representing humidity, which is used to adjust the sensitivity of the humidity threshold.

[0061] Furthermore, when the upper limit is reached, the detection report and alarm are fed back to the operator or the control system, and the alarm level is set to:

[0062]

[0063] Furthermore, the S6 specifically includes:

[0064] Assume that the historical data set is H, H = {(X1, Y1), (X2, Y2), ..., (X i ,Yi )}, where X i =[P i ,T i ,H i ], represents the feature vector of the i-th data point, including the pressure P i , Temperature T i and humidity H i ; Y i Indicates the corresponding working status;

[0065] Select a machine learning model M and use the historical data set H as input for training. The goal of the training is to minimize the loss function L M , which is expressed as follows:

[0066]

[0067] Among them, θ * represents the optimal parameters of the updated machine learning model M, θ M represents the parameters of the machine learning model M, θ old represents the old parameters of the machine learning model M;

[0068] According to real-time data X new The input is expressed as follows:

[0069]

[0070] in, Represents the predicted value of the machine learning model M;

[0071] The predicted value of the machine learning model M Feedback to the operator or control system for early intervention is as follows:

[0072]

[0073] Among them, Feedback to Operator or Control System refers to the operator or control system.

[0074] In a second aspect of the present invention, an air purification detection system based on deep learning is provided, the system comprising:

[0075] A data collection module is used to deploy infrared optical gas detection sensors around the molecular sieve adsorber, collect multi-dimensional data using the infrared optical gas detection sensors, pre-process the collected data, and randomly divide the pre-processed data into a training set and a test set; wherein the multi-dimensional data includes temperature, pressure and humidity;

[0076] The air purification detection module is used to build an air purification detection model based on a deep belief network, pre-train the air purification detection model using a training set; optimize the pre-trained air purification detection model using a test set; collect real-time multi-dimensional data and use the optimized air purification detection model to detect purified air;

[0077] The detection alarm module is used to build an adaptive adjustment model based on reinforcement learning according to the detection results, use the adaptive adjustment model to intelligently adjust the operation of the molecular sieve adsorber, and feed back the detection report and alarm to the operator or control system when the upper limit is adjusted;

[0078] The detection prediction module is used to build an air purification prediction model based on machine learning based on the historical multi-dimensional data and the detection results of the air purification detection model, and use the air purification prediction model to predict the future working state and adsorption efficiency of the molecular sieve adsorber and the failures that may occur in the future air purification process, and at the same time feed back the corresponding prediction results to the operator or the control system;

[0079] The air purification detection model includes several layers of restricted Boltzmann machines and back propagation networks, wherein the restricted Boltzmann machine includes a visible layer and a hidden layer, m visible layer units v constitute the visible layer, and n hidden layer units h constitute the hidden layer; the energy model of the restricted Boltzmann is expressed as follows:

[0080]

[0081] Among them, E(v,h) represents the energy between the visible layer unit v and the hidden layer unit h, w ij Represents the visual layer unit v i and the hidden layer unit h i The connection weights in the middle, I represents the number of nodes in the visible layer, J represents the number of nodes in the hidden layer, and a i and b j denote the unit thresholds of the visible layer unit and the hidden layer unit, respectively, γ is the regularization strength parameter, ||W|| 2 represents the square of the Frobenius norm of the weight matrix W, that is, the sum of the squares of all weights; where W = I × J;

[0082] The joint probability function constructed according to the restricted Boltzmann energy model is expressed as follows:

[0083]

[0084] Where Z represents the partition function, and Z = ∑ v ∑ h exp(-E(v,h));

[0085] The conditional probability function is:

[0086]

[0087] Here, sigmoid represents the Sigmoid function.

[0088] The beneficial technical effects of the present invention are at least as follows:

[0089] The present invention uses infrared optical gas detection sensors to accurately measure key parameters such as moisture, carbon dioxide concentration, air pressure and hydrocarbon concentration, providing rich and multi-dimensional data for subsequent model training. The preprocessing step ensures the quality of the data, including noise removal, missing value filling, data standardization, etc., creating favorable conditions for model training.

[0090] The present invention utilizes an air purification detection model based on a deep belief network to learn the intrinsic pattern of data on a training set, laying the foundation for subsequent detection and prediction. The powerful characterization ability of the deep learning model helps to capture complex nonlinear relationships and improve the accuracy and robustness of detection. By optimizing the pre-trained model through the test set, the generalization ability of the model can be evaluated and the parameters can be adjusted to ensure that the model not only performs well on the training data, but also gives accurate predictions on unknown data. This step improves the reliability of the model in practical applications. The multi-dimensional data collected in real time is input into the optimized model, which can instantly evaluate the air purification effect and provide real-time feedback, which helps to promptly discover and correct potential problems and ensure the quality and efficiency of air purification.

[0091] The adaptive adjustment model based on reinforcement learning can intelligently optimize the operating parameters of the molecular sieve adsorber, dynamically adjust to cope with environmental changes, and improve the adaptability and efficiency of the system. When the system detects an abnormality or reaches the upper limit of operation, the timely alarm feedback mechanism ensures the safe operation of the system and avoids equipment damage or production interruption.

[0092] The present invention constructs a machine learning prediction model and combines historical data and test results to predict the future working state, adsorption efficiency and potential failures of the molecular sieve adsorber, providing a scientific basis for preventive maintenance, resource planning and emergency response, reducing operating costs and improving the overall reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0094] Figure 1This is a flow chart of an air purification detection method based on deep learning in the present invention.

[0095] Figure 2 This is a framework diagram of an air purification detection system based on deep learning in the present invention. DETAILED DESCRIPTION

[0096] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0097] like Figure 1 As shown, an air purification detection method based on deep learning is provided in an embodiment of the present invention, and the method includes:

[0098] S1. deploying an infrared optical gas detection sensor around the molecular sieve adsorber, using the infrared optical gas detection sensor to collect multi-dimensional data, pre-processing the collected data, and randomly dividing the pre-processed data into a training set and a test set; wherein the multi-dimensional data includes temperature, pressure and humidity;

[0099] S2. Build an air purification detection model based on a deep belief network, and use the training set to pre-train the air purification detection model;

[0100] S3. Use the test set to optimize the pre-trained air purification detection model;

[0101] S4, collecting real-time multi-dimensional data and using the optimized air purification detection model to detect the purified air;

[0102] S5. Construct an adaptive adjustment model based on reinforcement learning according to the detection results, use the adaptive adjustment model to intelligently adjust the operation of the molecular sieve adsorber, and feed back the detection report and alarm to the operator or control system when the upper limit is adjusted;

[0103] S6. Construct an air purification prediction model based on machine learning based on the historical multidimensional data and the detection results of the air purification detection model, and use the air purification prediction model to predict the future working state and adsorption efficiency of the molecular sieve adsorber and the failures that may occur in the future air purification process, and feed back the corresponding prediction results to the operator or the control system;

[0104] The air purification detection model includes several layers of restricted Boltzmann machines and back propagation networks, wherein the restricted Boltzmann machine includes a visible layer and a hidden layer, m visible layer units v constitute the visible layer, and n hidden layer units h constitute the hidden layer; the energy model of the restricted Boltzmann is expressed as follows:

[0105]

[0106] Among them, E(v,h) represents the energy between the visible layer unit v and the hidden layer unit h, w ij Represents the visual layer unit v i and the hidden layer unit h i The connection weights in the middle, I represents the number of nodes in the visible layer, J represents the number of nodes in the hidden layer, and a i and b j denote the unit thresholds of the visible layer unit and the hidden layer unit, respectively, γ is the regularization strength parameter, ||W|| 2 represents the square of the Frobenius norm of the weight matrix W, that is, the sum of the squares of all weights; where W = I × J;

[0107] The joint probability function constructed according to the restricted Boltzmann energy model is expressed as follows:

[0108]

[0109] Where Z represents the partition function, and Z = ∑ v ∑ h exp(-E(v,h));

[0110] The conditional probability function is:

[0111]

[0112] Here, sigmoid represents the Sigmoid function.

[0113] Further, in S1, the preprocessing is represented as follows:

[0114] Assume that N infrared optical gas detection sensors are deployed around the molecular sieve adsorber, and each sensor measures the gas concentration C(t) at time point t. Let the reading of the jth sensor around the i-th adsorber at time point t be c ij (t), then for the adsorber, it is expressed as two matrices:

[0115] C1(t)=[c 11 (t),c 12 (t)…c 1N (t)],C2(t)=[c 21 (t),c22 (t)…c 2N (t)]

[0116] The output of the infrared optical gas detection sensor needs to be converted into the actual gas concentration value. Assuming that the output signal of the sensor is S, the relationship between it and the gas concentration C is C = f(S), where f is the calibration function;

[0117] Data cleaning removes invalid or abnormal data points, and defines C'(t) to represent the cleaned data, which is expressed as follows:

[0118]

[0119] Among them, k represents the abnormal threshold, μ ij represents the mean, σ ij represents the standard deviation, and NaN represents invalid or abnormal data points;

[0120] The cleaned data is standardized as follows:

[0121]

[0122] Wherein, C”(t) represents the data set after standardization.

[0123] Furthermore, the visual layer unit v i and the hidden layer unit h i The connection weight w in the middle ij The update rule is as follows:

[0124] Δw ij =∈[ <v i h j > data - <v i h j > model ]-2γ∈w ij

[0125] Among them, w ij Represents the contribution of the regularization term to the weight update, which prompts the connection weight to be adjusted in the direction of minimum. <v i h j > data Represents the visible layer node v calculated on the training set i and hidden layer nodes h j The expected value of the joint probability distribution of <v i h j > model Indicates that under the current parameter settings of the model, for the same node pair v i and h j The expected value of the joint probability distribution of , ∈ represents the learning rate.

[0126] Furthermore, the air purification detection model uses an unsupervised greedy layer-by-layer method to obtain weights, and trains each layer of restricted Boltzmann machines from bottom to top; wherein, in the pre-training stage, the visual layer vector value is first mapped to the hidden layer, and then the original input signal is reconstructed using the hidden unit to re-establish the visual unit, and the newly re-established visual unit is once again mapped to the hidden layer as input to obtain a new hidden layer unit, and the process is repeated; wherein, in the fine-tuning process of the air purification detection model, the last layer of the air purification detection model is set as a back propagation network, and the back propagation network is used to fine-tune the entire air purification detection model, and the feature vector input by the back propagation network is the feature vector output by the restricted Boltzmann machine, and the model parameters of the air purification detection model are adjusted.

[0127] Furthermore, in said S3, the steps specifically include:

[0128] S401, define the test set as T = {(x1,y1),(x2,y2),...,(x i ,y i )}, where x i Represents input data, y i Indicates the corresponding true label;

[0129] S402, using the absolute error as the loss function Loss for training the air purification detection model, expressed as follows:

[0130]

[0131] Where N is the number of samples in the test set. represents the predicted value;

[0132] S403, using the back propagation algorithm and the gradient descent method to minimize the loss function, as shown below:

[0133]

[0134] Among them, θ t+1 represents the updated model parameters, θ represents the model parameters, α represents the learning rate, Represents the gradient of the loss function with respect to the parameters.

[0135] Furthermore, the steps of constructing the adaptive adjustment model are as follows:

[0136] The feature extraction of the preprocessed real-time multidimensional data is shown as follows:

[0137]

[0138] Among them, F (x，y)represents the element of the output feature F in the i-th row and j-th column, W (x，y) represents the element of the convolution kernel W in the pth row and qth column, and the size of the convolution kernel is x×y, V x+p-1 , y+q-1 represents the element of the input feature V in the x+p-1th row and y+q-1th column, b l Represents the bias term of convolutional layer l;

[0139] The extracted features are weighted and fused using the attention mechanism to generate a purified air recognition model; among them, the attention weight It is expressed as follows:

[0140]

[0141] Wherein, e represents the base of natural logarithm, which means exponential operation on Sim(x, y); Sim(x, y) represents the similarity between feature x and feature y, k represents the kth feature, and Sim(x, k) represents the similarity between feature x and feature k;

[0142] Use supervised learning methods to introduce rule constraints and design loss function L auto as follows:

[0143] L auto =L cl +L mi +λL re

[0144] Among them, L cl represents the classification loss, λ represents the hyperparameter, L re represents the rule constraint loss, L mi represents the fusion loss.

[0145] Furthermore, the adaptive adjustment model also includes:

[0146] An adaptive threshold algorithm is introduced to adjust the temperature, pressure and humidity of the molecular sieve adsorber according to the historical data and real-time detection data of the molecular sieve adsorber. The adaptive threshold is:

[0147] T em =μ tem +k tem ×σ tem

[0148] P re =μ pre +k pre ×σ pre

[0149] H um =μ hum +khum ×σ hum

[0150] Among them, T em represents the temperature threshold, μ tem and σ tem They represent the mean and standard deviation of the historical temperature data of the molecular sieve adsorber, k tem The temperature hyperparameter is used to adjust the sensitivity of the temperature threshold. pre represents the pressure threshold, μ pre and σ pre The mean and standard deviation of the historical pressure data of the molecular sieve adsorber, k pre The hyperparameter representing the pressure is used to adjust the sensitivity of the pressure threshold; H um represents the humidity threshold, μ hum and σ hum The mean and standard deviation of the historical humidity data of the molecular sieve adsorber, k hum A hyperparameter representing humidity, which is used to adjust the sensitivity of the humidity threshold.

[0151] Furthermore, when the upper limit is reached, the detection report and alarm are fed back to the operator or the control system, and the alarm level is set to:

[0152]

[0153] Furthermore, the S6 specifically includes:

[0154] Assume that the historical data set is H, H = {(X1, Y1), (X2, Y2), ..., (X i ,Y i )}, where X i =[P i ,T i ,H i ], represents the feature vector of the i-th data point, including the pressure P i , Temperature T i and humidity H i ; Y i Indicates the corresponding working status;

[0155] Select a machine learning model M and use the historical data set H as input for training. The goal of the training is to minimize the loss function L M , which is expressed as follows:

[0156]

[0157] Among them, θ * represents the optimal parameters of the updated machine learning model M, θ M represents the parameters of the machine learning model M, θold represents the old parameters of the machine learning model M;

[0158] According to real-time data X new The input is expressed as follows:

[0159]

[0160] in, Represents the predicted value of the machine learning model M;

[0161] The predicted value of the machine learning model M Feedback to the operator or control system for early intervention is as follows:

[0162]

[0163] Among them, Feedback to Operator or Control System refers to the operator or control system.

[0164] like Figure 2 As shown, an embodiment of the present invention further provides an air purification detection system based on deep learning, the system comprising:

[0165] The data collection module 101 is used to deploy infrared optical gas detection sensors around the molecular sieve adsorber, collect multi-dimensional data using the infrared optical gas detection sensors, pre-process the collected data, and randomly divide the pre-processed data into a training set and a test set; wherein the multi-dimensional data includes temperature, pressure and humidity;

[0166] The air purification detection module 102 is used to construct an air purification detection model based on a deep belief network, pre-train the air purification detection model using a training set; optimize the pre-trained air purification detection model using a test set; collect real-time multi-dimensional data and detect the purified air using the optimized air purification detection model;

[0167] The detection alarm module 103 is used to construct an adaptive adjustment model based on reinforcement learning according to the detection results, use the adaptive adjustment model to intelligently adjust the operation of the molecular sieve adsorber, and feed back the detection report and alarm to the operator or the control system when the upper limit is adjusted;

[0168] The detection prediction module 104 is used to build an air purification prediction model based on machine learning based on the historical multi-dimensional data and the detection results of the air purification detection model, and use the air purification prediction model to predict the future working state and adsorption efficiency of the molecular sieve adsorber and the failures that may occur in the future air purification process, and feed back the corresponding prediction results to the operator or the control system;

[0169] The air purification detection model includes several layers of restricted Boltzmann machines and back propagation networks, wherein the restricted Boltzmann machine includes a visible layer and a hidden layer, m visible layer units v constitute the visible layer, and n hidden layer units h constitute the hidden layer; the energy model of the restricted Boltzmann is expressed as follows:

[0170]

[0171] Among them, E(v,h) represents the energy between the visible layer unit v and the hidden layer unit h, w ij Represents the visual layer unit v i and the hidden layer unit h i The connection weights in the middle, I represents the number of nodes in the visible layer, J represents the number of nodes in the hidden layer, and a i and b j denote the unit thresholds of the visible layer unit and the hidden layer unit, respectively, γ is the regularization strength parameter, ||W|| 2 represents the square of the Frobenius norm of the weight matrix W, that is, the sum of the squares of all weights; where W = I × J;

[0172] The joint probability function constructed according to the restricted Boltzmann energy model is expressed as follows:

[0173]

[0174] Where Z represents the partition function, and Z = ∑ v Σ h exp(-E(v,h));

[0175] The conditional probability function is:

[0176]

[0177] Here, sigmoid represents the Sigmoid function.

[0178] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0179] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0180] In the description of the present application, it should be noted that the terms "upper", "lower", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or the orientations or positional relationships in which the inventive product is usually placed when in use. These are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0181] In the description of this application, it should also be noted that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0182] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An air purification detection method based on deep learning, characterized in that: The method comprises: S1. deploying an infrared optical gas detection sensor around the molecular sieve adsorber, using the infrared optical gas detection sensor to collect multi-dimensional data, pre-processing the collected data, and randomly dividing the pre-processed data into a training set and a test set; wherein the multi-dimensional data includes temperature, pressure and humidity; S2. Build an air purification detection model based on a deep belief network, and use the training set to pre-train the air purification detection model; S3. Use the test set to optimize the pre-trained air purification detection model; S4, collecting real-time multi-dimensional data and using the optimized air purification detection model to detect the purified air; S5. Construct an adaptive adjustment model based on reinforcement learning according to the detection results, use the adaptive adjustment model to intelligently adjust the operation of the molecular sieve adsorber, and feed back the detection report and alarm to the operator or control system when the upper limit is adjusted; S6. Construct an air purification prediction model based on machine learning based on the historical multidimensional data and the detection results of the air purification detection model, and use the air purification prediction model to predict the future working state and adsorption efficiency of the molecular sieve adsorber and the failures that may occur in the future air purification process, and feed back the corresponding prediction results to the operator or the control system; The air purification detection model includes several layers of restricted Boltzmann machines and back propagation networks, wherein the restricted Boltzmann machine includes a visible layer and a hidden layer, m visible layer units v constitute the visible layer, and n hidden layer units h constitute the hidden layer; the energy model of the restricted Boltzmann is expressed as follows: Among them, E(v,h) represents the energy between the visible layer unit v and the hidden layer unit h, w ij Represents the visual layer unit v i and the hidden layer unit h i The connection weights in the middle, I represents the number of nodes in the visible layer, J represents the number of nodes in the hidden layer, and a i and b j denote the unit thresholds of the visible layer unit and the hidden layer unit, respectively, γ is the regularization strength parameter, ||W|| 2 represents the square of the Frobenius norm of the weight matrix W, that is, the sum of the squares of all weights; where W = I × J; The joint probability function constructed according to the restricted Boltzmann energy model is expressed as follows: Where Z represents the partition function, and Z = ∑ v ∑ h exp(-E(v,h)); The conditional probability function is: Here, sigmoid represents the Sigmoid function.

2. The air purification detection method based on deep learning according to claim 1, characterized in that: In S1, the preprocessing is represented as follows: Assume that N infrared optical gas detection sensors are deployed around the molecular sieve adsorber, and each sensor measures the gas concentration C(t) at time point t. Let the reading of the jth sensor around the i-th adsorber at time point t be c ij (t), then for the adsorber, it is expressed as two matrices: C1(t)=[c 11 (t),c 12 (t)…c 1N (t)],C2(t)=[c 21 (t),c 22 (t)…c 2N (t)] The output of the infrared optical gas detection sensor needs to be converted into the actual gas concentration value. Assuming that the output signal of the sensor is S, the relationship between it and the gas concentration C is C = f(S), where f is the calibration function; Data cleaning removes invalid or abnormal data points, and defines C'(t) to represent the cleaned data, which is expressed as follows: Among them, k represents the abnormal threshold, μ ij represents the mean, σ ij represents the standard deviation, and NaN represents invalid or abnormal data points; The cleaned data is standardized as follows: Wherein, C”(t) represents the data set after standardization.

3. The air purification detection method based on deep learning according to claim 1, characterized in that: The visual layer unit v i and the hidden layer unit h i The connection weight w in the middle ij The update rule is as follows: Δw ij =∈[<v i h j > data -<v i h j > model ]-2γ∈w ij Among them, w ij Represents the contribution of the regularization term to the weight update, which prompts the connection weight to be adjusted in the direction of minimum. <v i h j > data Represents the visible layer node v calculated on the training set i and hidden layer nodes h j The expected value of the joint probability distribution of <v i h j > model Indicates that under the current parameter settings of the model, for the same node pair v i and h j The expected value of the joint probability distribution of , ∈ represents the learning rate.

4. The air purification detection method based on deep learning according to claim 1, characterized in that: The air purification detection model uses an unsupervised greedy layer-by-layer method to obtain weights, and trains each layer of restricted Boltzmann machines from bottom to top; wherein, in the pre-training stage, the visual layer vector value is first mapped to the hidden layer, and then the original input signal is reconstructed using the hidden unit to re-establish the visual unit, and the newly re-established visual unit is once again mapped to the hidden layer as input to obtain a new hidden layer unit, and the process is repeated; wherein, in the fine-tuning process of the air purification detection model, the last layer of the air purification detection model is set as a back propagation network, and the back propagation network is used to fine-tune the entire air purification detection model, and the feature vector input by the back propagation network is the feature vector output by the restricted Boltzmann machine, and the model parameters of the air purification detection model are adjusted.

5. The air purification detection method based on deep learning according to claim 1, characterized in that: In said S3, the steps specifically include: S401. Define the test set as T = {(x1,y1),(x2,y2),...,(x i ,y i )}, where x i Represents input data, y i Indicates the corresponding true label; S402, using the absolute error as the loss function Loss for training the air purification detection model, expressed as follows: Where N is the number of samples in the test set. represents the predicted value; S403, using the back propagation algorithm and the gradient descent method to minimize the loss function, as shown below: Among them, θ t+1 represents the updated model parameters, θ represents the model parameters, α represents the learning rate, Represents the gradient of the loss function with respect to the parameters.

6. The air purification detection method based on deep learning according to claim 1, characterized in that: The steps for constructing the adaptive adjustment model are as follows: The feature extraction of the preprocessed real-time multidimensional data is shown as follows: Among them, F (x，y) represents the element of the output feature F in the i-th row and j-th column, W (x，y) represents the element of the convolution kernel W in the pth row and qth column, and the size of the convolution kernel is x×y, V x+p-1，y+q-1 represents the element of the input feature V in the x+p-1th row and y+q-1th column, b l Represents the bias term of convolutional layer l; The extracted features are weighted and fused using the attention mechanism to generate a purified air recognition model; among them, the attention weight It is expressed as follows: Wherein, e represents the base of natural logarithm, which means exponential operation on Sim(x, y); Sim(x, y) represents the similarity between feature x and feature y, k represents the kth feature, and Sim(x, k) represents the similarity between feature x and feature k; Use supervised learning methods to introduce rule constraints and design loss function L auto as follows: THE auto =L cl +L mi +λL re Among them, L cl represents the classification loss, λ represents the hyperparameter, L re represents the rule constraint loss, L mi represents the fusion loss.

7. The air purification detection method based on deep learning according to claim 6, characterized in that: The adaptive adjustment model also includes: An adaptive threshold algorithm is introduced to adjust the temperature, pressure and humidity of the molecular sieve adsorber according to the historical data and real-time detection data of the molecular sieve adsorber. The adaptive threshold is: T em =μ tem +k tem ×s tem P re =μ pre +k pre ×s pre H um =μ hum +k hum ×s hum Among them, T em represents the temperature threshold, μ tem and σ tem They represent the mean and standard deviation of the historical temperature data of the molecular sieve adsorber, k tem represents the temperature hyperparameter, which is used to adjust the sensitivity of the temperature threshold; p re represents the pressure threshold, μ pre and σ pre The mean and standard deviation of the historical pressure data of the molecular sieve adsorber, k pre The hyperparameter representing the pressure is used to adjust the sensitivity of the pressure threshold; H um Represents the humidity threshold, μ hum and σ gum The mean and standard deviation of the historical humidity data of the molecular sieve adsorber, k hum A hyperparameter representing humidity, which is used to adjust the sensitivity of the humidity threshold.

8. The air purification detection method based on deep learning according to claim 7, characterized in that: When the upper limit is reached, the detection report and alarm are fed back to the operator or the control system. The alarm level is set as:

9. The air purification detection method based on deep learning according to claim 1, characterized in that: The S6 specifically includes: Assume that the historical data set is H, H = {(X1, Y1), (X2, Y2), ..., (X i ,Y i )}, where X i =[P i ,T i ,H i ], represents the feature vector of the i-th data point, including the pressure P i , Temperature T i and humidity H i ; Y i Indicates the corresponding working status; Select a machine learning model M and use the historical data set H as input for training. The goal of the training is to minimize the loss function L M , which is expressed as follows: Among them, θ * represents the optimal parameters of the updated machine learning model M, θ M represents the parameters of the machine learning model M, θ old represents the old parameters of the machine learning model M; According to real-time data X new The input is expressed as follows: in, Represents the predicted value of the machine learning model M; The predicted value of the machine learning model M Feedback to the operator or control system for early intervention is as follows: Among them, Feedback to Operator or Control System refers to the operator or control system.

10. An air purification detection system based on deep learning, characterized in that: The system comprises: A data collection module is used to deploy infrared optical gas detection sensors around the molecular sieve adsorber, collect multi-dimensional data using the infrared optical gas detection sensors, pre-process the collected data, and randomly divide the pre-processed data into a training set and a test set; wherein the multi-dimensional data includes temperature, pressure and humidity; The air purification detection module is used to build an air purification detection model based on a deep belief network, pre-train the air purification detection model using a training set; optimize the pre-trained air purification detection model using a test set; collect real-time multi-dimensional data and use the optimized air purification detection model to detect purified air; The detection alarm module is used to build an adaptive adjustment model based on reinforcement learning according to the detection results, use the adaptive adjustment model to intelligently adjust the operation of the molecular sieve adsorber, and feed back the detection report and alarm to the operator or control system when the upper limit is adjusted; The detection prediction module is used to build an air purification prediction model based on machine learning based on the historical multi-dimensional data and the detection results of the air purification detection model, and use the air purification prediction model to predict the future working state and adsorption efficiency of the molecular sieve adsorber and the failures that may occur in the future air purification process, and at the same time feed back the corresponding prediction results to the operator or the control system; The air purification detection model includes several layers of restricted Boltzmann machines and back propagation networks, wherein the restricted Boltzmann machine includes a visible layer and a hidden layer, m visible layer units v constitute the visible layer, and n hidden layer units h constitute the hidden layer; the energy model of the restricted Boltzmann is expressed as follows: Among them, E(v,h) represents the energy between the visible layer unit v and the hidden layer unit h, w ij Represents the visual layer unit v i and the hidden layer unit h i The connection weights in the middle, I represents the number of nodes in the visible layer, J represents the number of nodes in the hidden layer, and a i and b j denote the unit thresholds of the visible layer unit and the hidden layer unit, respectively, γ is the regularization strength parameter, ||W|| 2 represents the square of the Frobenius norm of the weight matrix W, that is, the sum of the squares of all weights; where W = I × J; The joint probability function constructed according to the restricted Boltzmann energy model is expressed as follows: Where Z represents the partition function, and Z = ∑ v ∑ h exp(-E(v,h)); The conditional probability function is: Here, sigmoid represents the Sigmoid function.

Images