Air purification system, method and storage medium

Abstract

This application relates to an air purification system, method, and storage medium, applied in the field of Internet of Things (IoT) technology. The system includes: a cloud platform, a user terminal, a target vehicle, and a data acquisition module, a control module, and an air purification device within the target vehicle. The user terminal periodically collects user location information and sends the user location information and its timestamp to the cloud. The data acquisition module periodically collects vehicle location information and sends the vehicle location information and its timestamp to the cloud. The cloud platform calculates the position offset between the user terminal and the target vehicle at the current moment and at multiple historical moments. A position offset vector is generated based on the multiple position offsets, and the position offset vector is input into a first classifier to obtain a control result. When the control result indicates that the air purification device is enabled, an enable signal is sent to the control module to enable the air purification device. This application can automatically activate the air purification device.

Description

Technical Field

[0001] This application relates to the field of Internet of Things (IoT) technology, and in particular to an air purification system, method, and storage medium. Background Technology

[0002] In recent years, with the rise of the domestic RV market and the development of IoT technology, IoT RV control systems using wireless IoT, wireless terminals, and edge terminals have emerged, providing great convenience for RV users. RV users can remotely control various electrical facilities within the RV, improving the ease of use of the RV's interior facilities.

[0003] Air purification is a crucial function for improving the comfort of RV use. Due to the small interior space and limited energy supply in RVs, RV air purification differs from that of typical homes or cars. RV air purification has higher overall requirements, demanding both high-powered purification equipment to meet the needs of the large space and energy conservation to cope with situations where refueling is impossible during extended periods of travel. RV users typically turn on the purification equipment upon entering the RV or via remote control, lacking intelligent control over the equipment. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides an air purification system, method, and storage medium.

[0005] According to a first aspect of this application, an air purification system is provided, comprising: a cloud, a user terminal, a target vehicle, and a data acquisition module, a control module, and an air purification device located in the target vehicle;

[0006] The user terminal is used to periodically collect user location information and send the user location information and the timestamp of the user location information to the cloud.

[0007] The data acquisition module is used to periodically collect vehicle location information and send the vehicle location information and the timestamp of the vehicle location information to the cloud.

[0008] The cloud platform is used to calculate the positional offset between the user terminal and the target vehicle at the current moment and at multiple historical moments, based on the user location information, the timestamp of the user location information, the vehicle location information, and the timestamp of the vehicle location information; and

[0009] A position offset vector is generated based on the position offset corresponding to the current time and multiple historical times. The position offset vector is then input into a pre-trained first classifier to obtain the control result.

[0010] If the control result is "on", an "on" signal is sent to the control module so that the control module controls the air purification device to be turned on.

[0011] Optionally, the first classifier includes: a first hidden layer, a second hidden layer, and an output layer;

[0012] Specifically, the cloud is used to process the position offset vector using a first preset formula in the first hidden layer to obtain a first processed vector;

[0013] The first processing vector is processed by the second preset formula in the second hidden layer to obtain the second processing vector;

[0014] The second processing vector is processed by the third preset formula in the output layer to obtain the control result.

[0015] Optionally, the position offset vector is represented as:

[0016] ; Indicates the current time Position offset;

[0017] The cloud specifically serves to implement a first preset formula:

[0018] ;

[0019] The position offset vector is processed to obtain the first processed vector. ;

[0020] The cloud specifically serves to utilize the second preset formula:

[0021] ;

[0022] For the first processing vector The process is performed to obtain the second processing vector. ;

[0023] The cloud specifically serves to utilize a third preset formula:

[0024] ;

[0025] For the second processing vector Processing is performed to obtain control results;

[0026] in, and This represents the dimension of the position offset vector, with values ​​ranging from 0 to k. The value of can be 0~k, when , , Exceeding the boundary value When taking boundary values ;

[0027] This represents the linear weights of the first hidden layer. This represents the linear bias of the first hidden layer. This represents the linear weights of the second hidden layer. This indicates the linear bias of the second hidden layer. Represents the linear weights of the output layer. A represents the linear bias of the output layer, which is obtained through training; A() represents the activation function.

[0028] Optionally, the cloud is specifically used to train and generate the first classifier through the following steps:

[0029] The sample user location information and timestamp of the sample user location information are periodically sent by the sample user terminal, the sample vehicle location information and timestamp of the sample vehicle location information are periodically sent by the sample vehicle, and the sample control results of the target vehicle user on the control module in the sample vehicle are obtained.

[0030] Based on the sample user location information, the timestamp of the sample user location information, the sample vehicle location information, and the timestamp of the sample vehicle location information, calculate the sample position offset between the sample user terminal and the sample vehicle at the current time and at multiple historical times.

[0031] A sample position offset vector is generated based on multiple sample position offsets. The sample position offset vector is then input into a classifier initialized with network parameters to obtain the prediction control result.

[0032] The absolute value of the difference between the sample control result and the predictive control result is used as the loss value;

[0033] The network parameters are updated based on the loss value to generate the first classifier.

[0034] Optionally, the excitation function is:

[0035] , where m represents the preset adjustment coefficient.

[0036] Optionally, the target vehicle further includes: an image acquisition module and a data processing module;

[0037] The image acquisition module is used to acquire infrared thermal imaging images of the sleeping area in the target vehicle when the air purification device is turned on, and send an image sequence containing multiple infrared thermal imaging images to the data processing module.

[0038] The data processing module is used to perform multi-scale transformation on each infrared thermal imaging image in the image sequence using a Gaussian kernel function to obtain multi-scale image transformation features; and to perform gradient processing on the multi-scale image transformation features in two directions of the infrared thermal imaging image to obtain the two-dimensional gradient value corresponding to each pixel in the infrared thermal imaging image.

[0039] The difference between the two-dimensional gradient values ​​corresponding to the same pixels in adjacent infrared thermal imaging images in the plurality of infrared thermal imaging images is taken as the two-dimensional gradient difference, and a motion feature vector is generated based on the two-dimensional gradient difference; wherein, the motion feature vector includes the magnitude of the two-dimensional gradient difference and the included angle of the two-dimensional gradient difference;

[0040] The sum of the motion feature vectors corresponding to all adjacent infrared thermal imaging images is taken as the motion feature vector of the image sequence;

[0041] The motion feature vector of the image sequence is input into a pre-trained second classifier to determine the sleep result;

[0042] When the sleep result is sleep, turn off the air purifier.

[0043] Optionally, the data processing module is specifically used to perform image scaling according to the following formula:

[0044] ;

[0045] For images In scale The image transformation features are obtained by performing a scaling transformation. ;

[0046] Based on the following gradient processing formula:

[0047] ;

[0048] Gradient processing is performed on the image transformation features at multiple scales in two directions of the infrared thermal image to obtain the pixels in the infrared thermal image. exist l gradient value in direction and gradient value in the v direction ;

[0049] in, A pre-set constant; This represents the position coordinates of a pixel in an infrared thermal imaging image. Indicates the convolution kernel and The position of the corresponding element This represents the convolution operation. The variance is The Gaussian kernel function; j is 0, 1, ..., J-1, where J represents the number of scales in the multi-scale transformation.

[0050] Optionally, the data processing module is specifically used to generate a formula based on the following motion feature vectors:

[0051] ;

[0052] right and Processing is performed to obtain pixels. The magnitude of the gradient difference and included angle ;

[0053] in, Represents pixels in adjacent infrared thermal imaging images exist gradient difference in direction, Represents pixels in adjacent infrared thermal imaging images exist v The gradient difference in the direction;

[0054] Based on the modulus of the gradient difference corresponding to each pixel and included angle This generates motion feature vectors.

[0055] Optionally, the second classifier is a Naive Bayes classifier; the data processing module is specifically used to apply the Naive Bayes rule: The classification decision function of the Naive Bayes classifier:

[0056] , To determine the sleep outcome;

[0057] in, This represents the motion feature vector of the image sequence, where c indicates whether the image is sleeping (c=1 indicates sleeping, c=0 indicates not sleeping), and arg represents the motion feature vector when the image is sleeping. The value of C when the maximum value is reached; , , This was obtained through statistical analysis of sample data. This represents the probability that c in the sample data is classified as 0 or 1. This represents the joint probability of the sample motion feature vectors obtained based on the sample data. This represents the conditional probability of the sample data when c is 0 or c is 1. Indicates in Given the motion feature vector of the image sequence, the probability that c is 0 or c is 1.

[0058] According to a second aspect of this application, an air purification method is provided, comprising:

[0059] Receive user location information and timestamps of the user location information periodically collected and sent by the user terminal, and vehicle location information and timestamps of the vehicle location information periodically collected and sent by the data acquisition module in the target vehicle.

[0060] Based on the user location information, the timestamp of the user location information, the vehicle location information, and the timestamp of the vehicle location information, calculate the position offset between the user terminal and the target vehicle at the current time and at multiple historical times;

[0061] A position offset vector is generated based on the position offset corresponding to the current time and multiple historical times. The position offset vector is then input into a pre-trained first classifier to obtain the control result.

[0062] If the control result is "on", an activation signal is sent to the control module in the target vehicle so that the control module controls the activation of the air purification device in the target vehicle.

[0063] Optionally, the first classifier includes: a first hidden layer, a second hidden layer, and an output layer;

[0064] The step of inputting the position offset vector into a pre-trained first classifier to obtain the control result includes:

[0065] The position offset vector is processed by a first preset formula in the first hidden layer to obtain a first processed vector;

[0066] The first processing vector is processed by the second preset formula in the second hidden layer to obtain the second processing vector;

[0067] The second processing vector is processed by the third preset formula in the output layer to obtain the control result.

[0068] Optionally, the position offset vector is represented as:

[0069] ; Indicates the current time Position offset;

[0070] The step of processing the position offset vector using a first preset formula in the first hidden layer to obtain a first processed vector includes:

[0071] Through the first preset formula:

[0072] ;

[0073] The position offset vector is processed to obtain the first processed vector. ;

[0074] The step of processing the first processing vector using a second preset formula in the second hidden layer to obtain a second processing vector includes:

[0075] Through the second preset formula:

[0076] ;

[0077] For the first processing vector The process is performed to obtain the second processing vector. ;

[0078] The process of processing the second processing vector using a third preset formula in the output layer to obtain the control result includes:

[0079] Through the third preset formula:

[0080] ;

[0081] For the second processing vector Processing is performed to obtain control results;

[0082] in, and This represents the dimension of the position offset vector, with values ​​ranging from 0 to k. The value of can be 0~k, when , , Exceeding the boundary value When taking boundary values ;

[0083] This represents the linear weights of the first hidden layer. This represents the linear bias of the first hidden layer. This represents the linear weights of the second hidden layer. This indicates the linear bias of the second hidden layer. Represents the linear weights of the output layer. A represents the linear bias of the output layer, which is obtained through training; A() represents the activation function.

[0084] Optionally, the method further includes:

[0085] The sample user location information and timestamp of the sample user location information are periodically sent by the sample user terminal, the sample vehicle location information and timestamp of the sample vehicle location information are periodically sent by the sample vehicle, and the sample control results of the target vehicle user on the control module in the sample vehicle are obtained.

[0086] Based on the sample user location information, the timestamp of the sample user location information, the sample vehicle location information, and the timestamp of the sample vehicle location information, calculate the sample position offset between the sample user terminal and the sample vehicle at the current time and at multiple historical times.

[0087] A sample position offset vector is generated based on multiple sample position offsets. The sample position offset vector is then input into a classifier initialized with network parameters to obtain the prediction control result.

[0088] The absolute value of the difference between the sample control result and the predictive control result is used as the loss value;

[0089] The network parameters are updated based on the loss value to generate the first classifier.

[0090] Optionally, the excitation function is:

[0091] , where m represents the preset adjustment coefficient.

[0092] Optionally, the method further includes:

[0093] When the air purification device is turned on, infrared thermal imaging images of the sleeping area in the target vehicle are collected.

[0094] For each infrared thermal imaging image in the image sequence, a multi-scale transformation is performed on the infrared thermal imaging image using a Gaussian kernel function to obtain multi-scale image transformation features; gradient processing is performed on the multi-scale image transformation features in two directions of the infrared thermal imaging image to obtain the two-dimensional gradient value corresponding to each pixel in the infrared thermal imaging image.

[0095] The difference between the two-dimensional gradient values ​​corresponding to the same pixels in adjacent infrared thermal imaging images in the plurality of infrared thermal imaging images is taken as the two-dimensional gradient difference, and a motion feature vector is generated based on the two-dimensional gradient difference; wherein, the motion feature vector includes the magnitude of the two-dimensional gradient difference and the included angle of the two-dimensional gradient difference;

[0096] The sum of the motion feature vectors corresponding to all adjacent infrared thermal imaging images is taken as the motion feature vector of the image sequence;

[0097] The motion feature vector of the image sequence is input into a pre-trained second classifier to determine the sleep result;

[0098] When the sleep result is sleep, turn off the air purifier.

[0099] Optionally, the step of performing multi-scale transformation on the infrared thermal imaging image using a Gaussian kernel function to obtain multi-scale image transformation features includes:

[0100] Based on the following image scaling formula:

[0101] ;

[0102] For images In scale The image transformation features are obtained by performing a scaling transformation. ;

[0103] The step of performing gradient processing on the image transformation features at multiple scales in two directions of the infrared thermal imaging image to obtain the two-dimensional gradient value corresponding to each pixel in the infrared thermal imaging image includes:

[0104] Based on the following gradient processing formula:

[0105] ;

[0106] Gradient processing is performed on the image transformation features at multiple scales in two directions of the infrared thermal image to obtain the pixels in the infrared thermal image. exist gradient value in direction and gradient value in the v direction ;

[0107] in, A pre-set constant; This represents the position coordinates of a pixel in an infrared thermal imaging image. Indicates the convolution kernel and The position of the corresponding element This represents the convolution operation. The variance is The Gaussian kernel function; j is 0, 1, ..., J-1, where J represents the number of scales in the multi-scale transformation.

[0108] Optionally, generating a motion feature vector based on the two-dimensional gradient difference includes:

[0109] The following formula is used to generate motion feature vectors:

[0110] ;

[0111] right and Processing is performed to obtain pixels. The magnitude of the gradient difference and included angle ;

[0112] in, Represents pixels in adjacent infrared thermal imaging images exist gradient difference in direction, Represents pixels in adjacent infrared thermal imaging images exist v The gradient difference in the direction;

[0113] Based on the modulus of the gradient difference corresponding to each pixel and included angle This generates motion feature vectors.

[0114] Optionally, the second classifier is a Naive Bayes classifier; the step of inputting the motion feature vector of the image sequence into the pre-trained second classifier to determine the sleep result includes:

[0115] According to Naive Bayes' theorem: The classification decision function of the Naive Bayes classifier:

[0116] , To determine the sleep outcome;

[0117] in, This represents the motion feature vector of the image sequence, where c indicates whether the image is sleeping (c=1 indicates sleeping, c=0 indicates not sleeping), and arg represents the motion feature vector when the image is sleeping. The value of C when the maximum value is reached; , , This was obtained through statistical analysis of sample data. This represents the probability that c in the sample data is classified as 0 or 1. This represents the joint probability of the sample motion feature vectors obtained based on the sample data. This represents the conditional probability of the sample data when c is 0 or c is 1. Indicates in Given the motion feature vector of the image sequence, the probability that c is 0 or c is 1.

[0118] According to a third aspect of this application, an electronic device is provided, comprising: a processor configured to execute a computer program stored in a memory, wherein the computer program, when executed by the processor, implements the method described in the first aspect.

[0119] According to a fourth aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0120] According to a fifth aspect of this application, a computer program product is provided that, when the computer program product is run on a computer, causes the computer to perform the method described in the first aspect.

[0121] The technical solution provided in this application has the following advantages compared with the prior art:

[0122] The system periodically collects user location information via a user terminal and vehicle location information via a data acquisition module in the target vehicle. By receiving user location information, timestamps of user location information, and vehicle location information, it calculates the current positional offset between the user terminal and the target vehicle, as well as the positional offsets between the user terminal and the target vehicle at multiple historical moments. Based on this positional offset, it learns from user behavior to determine whether the target vehicle user will return, i.e., whether to pre-activate the air purification device in the target vehicle. If it is determined that the air purification device should be activated in advance, an activation signal is sent to the control module in the target vehicle, which can then activate the air purification device based on this signal. Therefore, the air purification device can be automatically activated before the target vehicle user returns, without relying on manual operation by the user, intelligently purifying the air quality in advance. When the target vehicle user returns, the air in the target vehicle has already been purified, improving the user experience. Attached Figure Description

[0123] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0124] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0125] Figure 1 This is a schematic diagram of an air purification system architecture in an embodiment of this application;

[0126] Figure 2 This is a flowchart of an air purification method in an embodiment of this application. Detailed Implementation

[0127] To better understand the above-mentioned objectives, features, and advantages of this application, the solution of this application will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0128] Many specific details are set forth in the following description in order to provide a full understanding of this application, but this application may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of this application, and not all embodiments.

[0129] See Figure 1 , Figure 1 This is a schematic diagram of an air purification system architecture in an embodiment of this application. The air purification system 100 includes: a cloud 101, a user terminal 102, a target vehicle 103 (e.g., a motorhome), and a data acquisition module 1031, a control module 1032, and an air purification device 1033 located in the target vehicle 103.

[0130] User terminal 102 is used to periodically collect user location information and send the user location information and the timestamp of the user location information to cloud 101.

[0131] User terminal 102 can periodically collect its own location information, which is user location information, and send the user location information and the timestamp of the time of collection to cloud 101 through the communication network.

[0132] The data acquisition module 1031 is used to periodically collect vehicle location information and send the vehicle location information and the timestamp of the vehicle location information to the cloud 101.

[0133] Similar to the user terminal, the data acquisition module 1031 can periodically collect its own location information. The location information of the data acquisition module 1031 is also the location information of the target vehicle 103, i.e., the vehicle location information. The vehicle location information and the timestamp of the time when the vehicle location information is collected are sent to the cloud 101 through the communication network.

[0134] It should be noted that the user terminal collects user location information and the data acquisition module collects vehicle location information using the same period and timestamp. This allows Cloud 101 to calculate the positional offset between the user terminal and the target vehicle at the same time.

[0135] Cloud 101 is used to calculate the positional offset between the user terminal and the target vehicle at the current moment and at multiple historical moments, based on user location information, timestamps of user location information, vehicle location information, and timestamps of vehicle location information. A positional offset vector is generated based on the positional offsets corresponding to the current moment and multiple historical moments. This positional offset vector is then input into a pre-trained first classifier to obtain the control result. If the control result indicates that the device is enabled, an enable signal is sent to the control module to enable the air purifier 1033.

[0136] After receiving the user location information sequence and its corresponding timestamp sequence, as well as the vehicle location information sequence and its corresponding timestamp sequence, Cloud 101 can calculate the positional offset (i.e., distance) between the user terminal and the target vehicle at the same time. Based on the positional offset between the user terminal and the target vehicle, user behavior can be learned to determine whether to turn on the air purifier. It is understandable that the closer the positional offset is to the current time, the higher the accuracy of the learned user behavior. Therefore, the positional offset corresponding to the current time and the positional offsets of multiple historical times before the current time can be used to learn user behavior. Arranging the positional offset corresponding to the current time and the positional offsets of multiple historical times before the current time in chronological order constitutes a positional offset vector. A first classifier can be pre-trained and used to process the positional offset vector to obtain the control result.

[0137] Optionally, Cloud 101 is specifically used to train and generate the first classifier through the following steps:

[0138] The system acquires the sample user location information and timestamps periodically sent by the sample user terminal, the sample vehicle location information and timestamps periodically sent by the sample vehicle, and the sample control results of the target vehicle user on the control module in the sample vehicle.

[0139] The input data includes sample user location information and timestamps, sample vehicle location information and timestamps, and the tag data includes the target vehicle user's control results on the control module in the sample vehicle. The control results can be the control results of the sample vehicle user on the air purification device in the sample vehicle. For example, when the air purification device is turned on, the control result can be represented as 1, and when the air purification device is turned off, the control result can be represented as 0.

[0140] Based on the sample user location information, timestamps of the sample user location information, sample vehicle location information, and timestamps of the sample vehicle location information, the sample position offsets between the sample user terminal and the sample vehicle at the current time and at multiple historical time points are calculated. A sample position offset vector is generated based on these offsets and input into the classifier initialized with network parameters to obtain the predictive control result. The absolute value of the difference between the sample control result and the predictive control result is used as the loss value; the network parameters are updated based on the loss value to generate the first classifier.

[0141] Air purifier 1033 may include a nano-water ion purification device. Nano-water ions, encapsulated in water, do not readily combine with oxygen and nitrogen in the air, thus allowing them to persist in the air for extended periods. Their antibacterial effect is also closely related to their composition: the water-encapsulated nano-water ions can easily adhere to the surface of various bacteria. They then allow the large number of hydroxyl radicals they contain to fully contact the bacteria, extracting hydrogen ions and combining with them to form water. This process inactivates various bacteria, achieving the effect of sterilization, deodorization, and purification. Therefore, nano-water ions have excellent air purification effects. It should be noted that air purifier 1033 can also be other types of air purifiers; this application does not limit its application to these types.

[0142] In some embodiments, the first classifier may include a first hidden layer, a second hidden layer, and an output layer. Specifically, Cloud101 processes the position offset vector using a first preset formula in the first hidden layer to obtain a first processed vector; processes the first processed vector using a second preset formula in the second hidden layer to obtain a second processed vector; and processes the second processed vector using a third preset formula in the output layer to obtain a control result.

[0143] As can be seen, the network structure consists of three layers, and all three layers use preset formulas for processing. The network structure is relatively simple, which can reduce the amount of computation and quickly complete the training of the first classifier.

[0144] Alternatively, the position offset vector can be represented as:

[0145] ;in, Indicates the current time Position offset. Cloud 101 is specifically used through the first preset formula:

[0146] ;

[0147] The position offset vector is processed to obtain the first processed vector. .

[0148] This represents the linear weights of the first hidden layer, meaning the maximum value of the four nodes in the input layer up to the value of the node in the first layer. The linear mapping parameters, This represents the linear bias of the first hidden layer. It should be noted that the first preset formula selects the maximum value of four nodes, but it can also select the maximum value of three nodes, the maximum value of five nodes, etc., and is not limited here.

[0149] Cloud 101 is specifically used through the second preset formula:

[0150] ;

[0151] For the first processing vector The process is performed to obtain the second processing vector. .

[0152] This represents the linear weights of the second hidden layer, which are the values ​​of the nodes in the first layer. To the second-level node The linear mapping parameters, This indicates the linear bias of the second hidden layer.

[0153] Cloud 101 is specifically used through the third preset formula:

[0154] ;

[0155] For the second processing vector Processing is performed to obtain control results;

[0156] in, and This represents the dimension of the position offset vector, with values ​​ranging from 0 to k. The value of can be 0~k, when , , Exceeding the boundary value At that time, boundary values ​​can be taken. ;

[0157] Represents the linear weights of the output layer. A represents the linear bias of the output layer, which is obtained through training; A() represents the activation function.

[0158] The first classifier in this application embodiment, compared with the classic fully connected neural network classifier, adds a maximization layer (i.e., the first hidden layer), which can reduce the robustness of the classifier to local noise and improve the classification performance of the classifier.

[0159] Optionally, the excitation function is:

[0160] Where m represents the preset adjustment coefficient, and the value of m can be set to 3 through experimentation.

[0161] Compared to the classic sigmoid function used in fully connected neural networks, this activation function calculation method can improve the performance of sample classification, especially increasing the correct classification rate of outliers.

[0162] The air purification system of this application embodiment can calculate the positional offset between the user terminal and the target vehicle at multiple moments. Based on this positional offset, it learns user behavior to determine whether the user of the target vehicle will return, i.e., whether to pre-activate the air purification device in the target vehicle. If it is determined that the air purification device should be activated in advance, an activation signal is sent to the control module in the target vehicle, which can then activate the air purification device based on the activation signal. Therefore, the air purification device can be automatically activated before the user returns to the target vehicle, without relying on manual operation by the user, intelligently purifying the air quality in advance. When the user returns to the target vehicle, the air inside has already been purified, improving the user experience.

[0163] The air purification system of this application embodiment can also monitor the user's sleep behavior and turn off the air purification device when the user falls asleep, so as to reduce noise and save energy.

[0164] Optionally, the target vehicle may also include an image acquisition module and a data processing module.

[0165] The image acquisition module is used to acquire infrared thermal imaging images of the sleeping area in the target vehicle when the air purification device is turned on, and send an image sequence containing multiple infrared thermal imaging images to the data processing module.

[0166] In recent years, image processing-based behavior monitoring has been widely researched and applied, achieving good results. Image processing-based behavior monitoring applications typically require a good lighting environment, i.e., suitable and stable lighting. However, target vehicles are located in enclosed spaces with relatively insufficient indoor lighting, and due to the small space, the monitored area is often obstructed and interfered with. Considering user privacy, i.e., users do not want their sleeping areas to be photographed, the image acquisition module can be an infrared thermal imaging camera, which can acquire infrared thermal images of the sleeping area.

[0167] For example, the image acquisition module can acquire infrared thermal images at regular intervals, resulting in an image sequence containing multiple infrared thermal images. This image sequence is then sent to the data processing module for analysis to determine whether the user is currently asleep. It can be seen that the image acquisition module does not upload the image sequence to the cloud via the communication network, thus protecting the user's privacy.

[0168] In related technologies, user behavior analysis is typically based on methods that utilize local features, such as optical flow analysis and gradient histogram methods. In recent years, machine learning methods based on convolutional networks have gradually emerged. Traditional methods mainly rely on statistical analysis of local features in images / videos, making them susceptible to noise interference. Furthermore, convolutional networks require training large network models, making real-time operation difficult on data processing modules with limited computing power.

[0169] To address the aforementioned issues, this application proposes a lightweight learning model that improves the robustness of local feature statistical methods to noise based on optimized local features. It also boasts high computational efficiency and enables real-time operation on the data processing module.

[0170] First, a feature representation method for image sequences based on multi-scale local features is proposed, as follows:

[0171] The data processing module is used to perform multi-scale transformation on each infrared thermal imaging image in the image sequence using a Gaussian kernel function to obtain multi-scale image transformation features.

[0172] Optionally, the data processing module is specifically used to perform image scaling according to the following formula:

[0173] ;

[0174] For images In scale The image transformation features are obtained by performing a scaling transformation. ;

[0175] in, A pre-set constant; This represents the position coordinates of a pixel in an infrared thermal imaging image. Indicates the convolution kernel and The position of the corresponding element This represents the convolution operation. The variance is The Gaussian kernel function; j is 0, 1, ..., J-1, where J represents the number of scales in the multi-scale transformation, and J=3 can be chosen.

[0176] A set of outputs formed by transforming the same image at continuous scales can be called a pyramid representation of the original image. For example, , , It is an image The pyramid representation has 3 layers.

[0177] Subsequently, gradient processing is performed on the image transformation features at multiple scales in two directions of the infrared thermal image to obtain the two-dimensional gradient values ​​corresponding to each pixel in the infrared thermal image.

[0178] Alternatively, the gradient can be processed according to the following formula:

[0179] ;

[0180] Gradient processing is performed on the image transformation features at multiple scales in two directions of the infrared thermal image to obtain the pixels in the infrared thermal image. exist l gradient value in direction and gradient value in the v direction ;

[0181] The two-dimensional gradient values ​​corresponding to each pixel in the infrared thermal imaging image obtained according to the above gradient processing formula can be called the gradient map of the image pyramid representation. Each pixel in the gradient map is a two-dimensional vector, representing the change of that pixel in two directions.

[0182] By acquiring the local features of a pixel in an infrared thermal imaging image at multiple scales, the interference of local noise on local features can be reduced, and the robustness of the features can be improved. By constructing the gradient map of the original image pyramid, the local variation features of the original image can be described.

[0183] Furthermore, the difference between the two-dimensional gradient values ​​corresponding to the same pixels in adjacent infrared thermal imaging images in multiple infrared thermal imaging images is taken as the two-dimensional gradient difference, and a motion feature vector is generated based on the two-dimensional gradient difference; wherein, the motion feature vector includes the magnitude of the two-dimensional gradient difference and the angle between the two-dimensional gradient difference.

[0184] Optionally, the data processing module is specifically used to generate formulas based on the following motion feature vectors:

[0185] ;

[0186] right and Processing is performed to obtain pixels. The magnitude of the gradient difference and included angle ;

[0187] in, Represents pixels in adjacent infrared thermal imaging images exist gradient difference in direction, Represents pixels in adjacent infrared thermal imaging images exist v The gradient difference in the direction;

[0188] Based on the modulus of the gradient difference corresponding to each pixel and included angle This generates motion feature vectors.

[0189] Based on each pixel The value is quantized to form a histogram vector. The dimension can be 128; based on each pixel The value is quantized to form a histogram vector. The dimension can be 16. The two sets of histogram vectors are merged to form the motion feature vector. .

[0190] For an image sequence, the sum of the motion feature vectors corresponding to all adjacent infrared thermal imaging images is used as the motion feature vector of the image sequence. This motion feature vector can be input into a pre-trained second classifier to determine the sleep outcome. When the sleep outcome is sleep, the air purifier is turned off.

[0191] In some embodiments, the second classifier can be pre-trained in the cloud, and the data processing module can download the second classifier from the cloud to its local machine, then use the second classifier to process the motion feature vectors of the image sequence. It can be seen that the communication between the data processing module and the cloud is unidirectional. Alternatively, if the data processing module has sufficient computing power, the second classifier can also be trained and generated locally within the data processing module.

[0192] Optionally, the second classifier is a Naive Bayes classifier; the data processing module is specifically used to apply the Naive Bayes rule: The classification decision function of the Naive Bayes classifier:

[0193] , To determine the sleep outcome;

[0194] in, This represents the motion feature vector of the image sequence, where c indicates whether it is sleeping (c=1 indicates sleeping, c=0 indicates not sleeping), and arg represents the motion feature vector when... The value of C when it reaches its maximum value; assuming If the result is true, the sleep result is 1; otherwise, the sleep result is 0.

[0195] , , It can be obtained through statistical analysis of sample data. This represents the probability that c in the sample data is classified as 0 or 1. This represents the joint probability of the sample motion feature vectors obtained based on the sample data. This represents the conditional probability of the sample data when c is 0 or c is 1. Indicates in Given the motion feature vector of the image sequence, the probability that c is 0 or c is 1.

[0196] This application does not limit the type of the second classifier. For example, the second classifier can also be a binary classification model such as a neural network model.

[0197] The air purification system of this application embodiment can be started and stopped without relying on manual operation by the target vehicle user, thus realizing intelligent and automated control of the air purification device. Table 1 shows the relevant test data of the air purification system of this application embodiment.

[0198] Table 1

[0199]

[0200] As can be seen, the air purification system in this application embodiment has a high rate of automated recognition, effectively realizing the optimized control and automatic adjustment of the vehicle-to-everything (V2X) air purification system.

[0201] This application also provides an air purification method, see [link to relevant documentation]. Figure 2 This may include the following steps:

[0202] Step S210: Receive user location information and timestamps periodically collected and sent by the user terminal, as well as vehicle location information and timestamps periodically collected and sent by the data acquisition module in the target vehicle.

[0203] Step S220: Calculate the position offset between the user terminal and the target vehicle at the current time and at multiple historical times based on the user location information, the timestamp of the user location information, the vehicle location information, and the timestamp of the vehicle location information.

[0204] Step S230: Generate a position offset vector based on the position offsets corresponding to the current time and multiple historical time points, and input the position offset vector into the pre-trained first classifier to obtain the control result.

[0205] In step S240, if the control result is "on", an "on" signal is sent to the control module in the target vehicle so that the control module controls the air purification device in the target vehicle to be turned on.

[0206] Optionally, the first classifier includes: a first hidden layer, a second hidden layer, and an output layer;

[0207] The position offset vector is input into a pre-trained first classifier to obtain the control results, including:

[0208] The position offset vector is processed by the first preset formula in the first hidden layer to obtain the first processed vector;

[0209] The first processing vector is processed by the second preset formula in the second hidden layer to obtain the second processing vector;

[0210] The second processing vector is processed by the third preset formula in the output layer to obtain the control result.

[0211] Alternatively, the position offset vector can be represented as: ;in, Indicates the current time The position offset. The position offset vector is processed using a first preset formula in the first hidden layer to obtain a first processed vector, including:

[0212] Through the first preset formula:

[0213] ;

[0214] The position offset vector is processed to obtain the first processed vector. .

[0215] The first processing vector is processed by the second preset formula in the second hidden layer to obtain the second processing vector, which includes:

[0216] Through the second preset formula:

[0217] ;

[0218] For the first processing vector The process is performed to obtain the second processing vector. .

[0219] The second processing vector is processed by the third preset formula in the output layer to obtain the control result, including:

[0220] Through the third preset formula:

[0221] ;

[0222] For the second processing vector Processing is performed to obtain control results;

[0223] in, and This represents the dimension of the position offset vector, with values ​​ranging from 0 to k. The value of can be 0~k, when , , Exceeding the boundary value At that time, boundary values ​​can be taken. ;

[0224] This represents the linear weights of the first hidden layer. This represents the linear bias of the first hidden layer. This represents the linear weights of the second hidden layer. This indicates the linear bias of the second hidden layer. Represents the linear weights of the output layer. A represents the linear bias of the output layer, which is obtained through training; A() represents the activation function.

[0225] Optionally, the above air purification methods also include:

[0226] The sample user location information and timestamp of the sample user location information periodically sent by the sample user terminal, the sample vehicle location information and timestamp of the sample vehicle location information periodically sent by the sample vehicle, and the sample control results of the target vehicle user on the control module in the sample vehicle.

[0227] Based on the sample user location information, the timestamp of the sample user location information, the sample vehicle location information, and the timestamp of the sample vehicle location information, calculate the sample position offset between the sample user terminal and the sample vehicle at the current time and at multiple historical times.

[0228] A sample position offset vector is generated based on multiple sample position offsets. The sample position offset vector is then input into a classifier initialized with network parameters to obtain the prediction control result.

[0229] The absolute value of the difference between the sample control result and the predicted control result is used as the loss value;

[0230] The network parameters are updated based on the loss value to generate the first classifier.

[0231] Optionally, the excitation function is:

[0232] , where m represents the preset adjustment coefficient.

[0233] Optionally, the above air purification methods also include:

[0234] When the air purifier is turned on, infrared thermal imaging images of the sleeping area in the target vehicle are collected.

[0235] For each infrared thermal imaging image in the image sequence, a multi-scale transformation is performed on the infrared thermal imaging image using a Gaussian kernel function to obtain multi-scale image transformation features; gradient processing is then performed on the multi-scale image transformation features in two directions of the infrared thermal imaging image to obtain the two-dimensional gradient value corresponding to each pixel in the infrared thermal imaging image.

[0236] The difference between the two-dimensional gradient values ​​corresponding to the same pixels in adjacent infrared thermal imaging images in multiple infrared thermal imaging images is used as the two-dimensional gradient difference, and a motion feature vector is generated based on the two-dimensional gradient difference; wherein, the motion feature vector includes the magnitude of the two-dimensional gradient difference and the angle between the two-dimensional gradient differences;

[0237] The sum of the motion feature vectors corresponding to all adjacent infrared thermal imaging images is taken as the motion feature vector of the image sequence;

[0238] The motion feature vectors of the image sequence are input into a pre-trained second classifier to determine the sleep result;

[0239] When the sleep result indicates sleep, turn off the air purifier.

[0240] Optionally, the infrared thermal imaging image is subjected to multi-scale transformation using a Gaussian kernel function to obtain multi-scale image transformation features, including:

[0241] Based on the following image scaling formula:

[0242] ;

[0243] For images In scale The image transformation features are obtained by performing a scaling transformation. ;

[0244] Gradient processing is performed on the image transformation features at multiple scales in two directions of the infrared thermal image to obtain the two-dimensional gradient values ​​corresponding to each pixel in the infrared thermal image, including:

[0245] Based on the following gradient processing formula:

[0246] ;

[0247] Gradient processing is performed on the image transformation features at multiple scales in two directions of the infrared thermal image to obtain the pixels in the infrared thermal image. exist gradient value in direction and gradient value in the v direction ;

[0248] in, A pre-set constant; This represents the position coordinates of a pixel in an infrared thermal imaging image. Indicates the convolution kernel and The position of the corresponding element This represents the convolution operation. The variance is The Gaussian kernel function; j is 0, 1, ..., J-1, where J represents the number of scales in the multi-scale transformation.

[0249] Optionally, generating a motion feature vector based on the two-dimensional gradient difference includes:

[0250] The following formula is used to generate motion feature vectors:

[0251] ;

[0252] right and Processing is performed to obtain pixels. The magnitude of the gradient difference and included angle ;

[0253] in, Represents pixels in adjacent infrared thermal imaging images exist gradient difference in direction, Represents pixels in adjacent infrared thermal imaging images exist v The gradient difference in the direction;

[0254] Based on the modulus of the gradient difference corresponding to each pixel and included angle This generates motion feature vectors.

[0255] Optionally, the second classifier is a Naive Bayes classifier; the step of inputting the motion feature vector of the image sequence into the pre-trained second classifier to determine the sleep result includes:

[0256] According to Naive Bayes' theorem: The classification decision function of the Naive Bayes classifier:

[0257] , To determine the sleep outcome;

[0258] in, This represents the motion feature vector of the image sequence, where c indicates whether the image is sleeping (c=1 indicates sleeping, c=0 indicates not sleeping), and arg represents the motion feature vector when the image is sleeping. The value of C when the maximum value is reached; , , This was obtained through statistical analysis of sample data. This represents the probability that c in the sample data is classified as 0 or 1. This represents the joint probability of the sample motion feature vectors obtained based on the sample data. This represents the conditional probability of the sample data when c is 0 or c is 1. Indicates in Given the motion feature vector of the image sequence, the probability that c is 0 or c is 1.

[0259] The specific details of each step in the above method have been described in detail in the corresponding system, so they will not be repeated here.

[0260] In this embodiment of the application, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the above-described air purification method.

[0261] It should be noted that the computer-readable storage medium shown in this application can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable storage medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, radio frequency, etc., or any suitable combination thereof.

[0262] In this embodiment of the application, a computer program product is also provided, which, when run on a computer, causes the computer to execute the above-described air purification method.

[0263] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0264] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An air purification system, characterized in that, include: Cloud, user terminal, target vehicle, and data acquisition module, control module, and air purification device located in the target vehicle; The user terminal is used to periodically collect user location information and send the user location information and the timestamp of the user location information to the cloud. The data acquisition module is used to periodically collect vehicle location information and send the vehicle location information and the timestamp of the vehicle location information to the cloud. The cloud is used to calculate the position offset between the user terminal and the target vehicle at the current time and at multiple historical times based on the user location information, the timestamp of the user location information, the vehicle location information, and the timestamp of the vehicle location information. as well as A position offset vector is generated based on the position offset corresponding to the current time and multiple historical times. The position offset vector is then input into a pre-trained first classifier to obtain the control result. If the control result is "on", an "on" signal is sent to the control module so that the control module controls the air purifier to be turned on. The first classifier includes: a first hidden layer, a second hidden layer, and an output layer; Specifically, the cloud is used to process the position offset vector using a first preset formula in the first hidden layer to obtain a first processed vector; The first processing vector is processed by the second preset formula in the second hidden layer to obtain the second processing vector; The second processing vector is processed by the third preset formula in the output layer to obtain the control result; The position offset vector is represented as: [O(t),O(t+T),...,O(t+(k-1)T),O(t+kT)]; O(t+kT) represents the position offset at the current time t+kT; The cloud specifically serves to implement a first preset formula: R(x1)=A(β1+∑ i (α1(x1,i).max(O(t+iT),O(t+(i+1)T),O(t+(i+2)T),O(t+(i+3)T)))) The position offset vector is processed to obtain the first processed vector R(x1); The cloud specifically serves to utilize the second preset formula: The first processing vector R(x1) is processed to obtain the second processing vector Q(x2); The cloud specifically serves to utilize a third preset formula: The second processing vector Q(x2) is processed to obtain the control result; Where x1 and x2 represent the dimensions of the position offset vector, with values ​​from 0 to k, and i has values ​​from 0 to k. When O(t+(i+3)T), O(t+(i+2)T), and O(t+(i+1)T) exceed the boundary value O(t+kT), the boundary value O(t+kT) is taken. α1 represents the linear weights of the first hidden layer, β1 represents the linear bias of the first hidden layer, α2 represents the linear weights of the second hidden layer, β2 represents the linear bias of the second hidden layer, α3 represents the linear weights of the output layer, and β3 represents the linear bias of the output layer, all of which are obtained through training; A() represents the activation function.

2. The system according to claim 1, characterized in that, Specifically, the cloud is used to train and generate the first classifier through the following steps: The sample user location information and timestamp of the sample user location information are periodically sent by the sample user terminal, the sample vehicle location information and timestamp of the sample vehicle location information are periodically sent by the sample vehicle, and the sample control results of the target vehicle user on the control module in the sample vehicle are obtained. Based on the sample user location information, the timestamp of the sample user location information, the sample vehicle location information, and the timestamp of the sample vehicle location information, calculate the sample position offset between the sample user terminal and the sample vehicle at the current time and at multiple historical times. A sample position offset vector is generated based on multiple sample position offsets. The sample position offset vector is then input into a classifier initialized with network parameters to obtain the prediction control result. The absolute value of the difference between the sample control result and the predictive control result is used as the loss value; The network parameters are updated based on the loss value to generate the first classifier.

3. The system according to claim 1, characterized in that, The excitation function is: Where m represents the preset adjustment coefficient.

4. The system according to claim 1, characterized in that, The target vehicle also includes: an image acquisition module and a data processing module; The image acquisition module is used to acquire infrared thermal imaging images of the sleeping area in the target vehicle when the air purification device is turned on, and send an image sequence containing multiple infrared thermal imaging images to the data processing module. The data processing module is used to perform multi-scale transformation on each infrared thermal imaging image in the image sequence using a Gaussian kernel function to obtain multi-scale image transformation features; and to perform gradient processing on the multi-scale image transformation features in two directions of the infrared thermal imaging image to obtain the two-dimensional gradient value corresponding to each pixel in the infrared thermal imaging image. The difference between the two-dimensional gradient values ​​corresponding to the same pixels in adjacent infrared thermal imaging images in the plurality of infrared thermal imaging images is taken as the two-dimensional gradient difference, and a motion feature vector is generated based on the two-dimensional gradient difference; wherein, the motion feature vector includes the magnitude of the two-dimensional gradient difference and the included angle of the two-dimensional gradient difference; The sum of the motion feature vectors corresponding to all adjacent infrared thermal imaging images is taken as the motion feature vector of the image sequence; The motion feature vector of the image sequence is input into a pre-trained second classifier to determine the sleep result; When the sleep result is sleep, turn off the air purifier.

5. The system according to claim 4, characterized in that, The data processing module is specifically used to perform image scaling according to the following formula: For image I(l,v) at scale 2 j Scale transformation is performed under σ to obtain the image transformation features L(l,v,2). j σ); Based on the following gradient processing formula: Gradient processing is performed on the image transformation features at multiple scales in two directions of the infrared thermal image to obtain the gradient value of pixel (l,v) in the l direction of the infrared thermal image. and gradient value in the v direction Where σ is a pre-set constant; (l,v) represents the position coordinates of a pixel in the infrared thermal image, and (l′,v′) represents the position of the element in the convolution kernel corresponding to (l,v). This represents the convolution operation, where g(l′,v′,σ) represents the variance σ. 2 The Gaussian kernel function; j is 0, 1, ..., J-1, where J represents the number of scales in the multi-scale transformation.

6. The system according to claim 4, characterized in that, The data processing module is specifically used to generate a formula based on the following motion feature vectors: For γ l (l,v) and γ v The gradient difference between pixels (l,v) is processed to obtain the magnitude ρ(l,v) of the gradient difference and the included angle. Where, γ l (l,v) represents the gradient difference of pixels (l,v) in adjacent infrared thermal images along the l direction, γ v (l,v) represents the gradient difference of pixels (l,v) in the v direction in adjacent infrared thermal imaging images; Based on the magnitude ρ(l,v) of the gradient difference corresponding to each pixel and the included angle Generate motion feature vectors.

7. The system according to claim 4, characterized in that, The second classifier is a Naive Bayes classifier; the data processing module is specifically used to apply the Naive Bayes rule: The classification decision function of the Naive Bayes classifier: c = 0, 1, to determine the sleep outcome; Among them, H mo This represents the motion feature vector of the image sequence, where c indicates whether the image is in a sleep state (c=1 indicates sleep, c=0 indicates no sleep), and arg represents the motion feature vector when P(c|H) is in a sleep state. mo The value of C when it reaches its maximum value; P(c), P(H) mo ), P(H mo |c) can be obtained statistically from the sample data, where P(c) represents the probability that c is classified as 0 or 1 in the sample data, P(H mo P(H) represents the joint probability of the sample motion feature vectors obtained based on the sample data. mo |c) represents the conditional probability of the sample data when c is 0 or c is 1, P(c|H) mo ) indicates in H mo Given the motion feature vector of the image sequence, the probability that c is 0 or c is 1.

8. An air purification method, characterized in that, The method includes: Receive user location information and timestamps of the user location information periodically collected and sent by the user terminal, and vehicle location information and timestamps of the vehicle location information periodically collected and sent by the data acquisition module in the target vehicle. Based on the user location information, the timestamp of the user location information, the vehicle location information, and the timestamp of the vehicle location information, calculate the position offset between the user terminal and the target vehicle at the current time and at multiple historical times; A position offset vector is generated based on the position offset corresponding to the current time and multiple historical times. The position offset vector is then input into a pre-trained first classifier to obtain the control result. If the control result is "on", an activation signal is sent to the control module in the target vehicle so that the control module controls the activation of the air purification device in the target vehicle. The first classifier includes: a first hidden layer, a second hidden layer, and an output layer; The step of inputting the position offset vector into a pre-trained first classifier to obtain the control result includes: The position offset vector is processed by a first preset formula in the first hidden layer to obtain a first processed vector; The first processing vector is processed by the second preset formula in the second hidden layer to obtain the second processing vector; The second processing vector is processed by the third preset formula in the output layer to obtain the control result; The position offset vector is represented as: [O(t),O(t+T),…,O(t+(k-1)T),O(t+kT)]; O(t+kT) represents the position offset at the current time t+kT; The step of processing the position offset vector using a first preset formula in the first hidden layer to obtain a first processed vector includes: Through the first preset formula: R(x1)=A(β1+∑ i (α1(x1,i).max(O(t+iT),O(t+(i+1)T),O(t+(i+2)T),O(t+(i+3)T)))) The position offset vector is processed to obtain the first processed vector R(x1); The step of processing the first processing vector using a second preset formula in the second hidden layer to obtain a second processing vector includes: Through the second preset formula: The first processing vector R(x1) is processed to obtain the second processing vector Q(x2); The process of processing the second processing vector through a third preset formula in the output layer to obtain the control result includes: Through the third preset formula: The second processing vector Q(x2) is processed to obtain the control result; Where x1 and x2 represent the dimensions of the position offset vector, with values ​​from 0 to k, and i has values ​​from 0 to k. When O(t+(i+3)T), O(t+(i+2)T), and O(t+(i+1)T) exceed the boundary value O(t+kT), the boundary value O(t+kT) is taken. α1 represents the linear weights of the first hidden layer, β1 represents the linear bias of the first hidden layer, α2 represents the linear weights of the second hidden layer, β2 represents the linear bias of the second hidden layer, α3 represents the linear weights of the output layer, and β3 represents the linear bias of the output layer, all of which are obtained through training; A() represents the activation function.

Images