A defrosting control method for air source heat pump based on texture features and HBA-DELM algorithm

Abstract

The present invention discloses a defrost control method for an air source heat pump based on texture features and an HBA-DELM algorithm, comprising the following steps: collecting sample images and constructing an image dataset classified into positive and negative sample images; performing image preprocessing and illumination preprocessing on the sample images in the image dataset to obtain a preprocessed image dataset; extracting local HOG features and global GLCM features from the sample images in the preprocessed image dataset, and weightedly fusing the two features to obtain multi-scale fusion features; training the multi-scale fusion feature information using a deep extreme learning machine, and optimizing the deep extreme learning machine using the honey badger algorithm to construct an HBA-DELM algorithm model; and performing defrost detection on an air source heat pump system using the HBA-DELM algorithm model and performing corresponding defrost control based on the defrost detection results. The present invention can accurately identify the frosting state of an air source heat pump and perform corresponding defrost operations, thereby reducing the occurrence of false defrost accidents and energy waste.

Description

Technical Field

[0001] The present invention belongs to the technical field of heat pump defrost control, and in particular relates to an air source heat pump defrost control method based on texture features and an HBA-DELM algorithm. Background Art

[0002] Air source heat pump systems are widely used in heating and cooling systems in the Yangtze River Basin. They have the advantages of high efficiency, low cost, easy installation, and environmental friendliness. In winter, the refrigerant in the evaporator tubes of the air source heat pump absorbs heat from the outdoor air to raise the water temperature. The surface temperature of the evaporator is much lower than the dew point temperature of the air. Water in the air is separated and condenses on the evaporator surface. When the evaporator surface temperature is below 0°C, the dew condenses into frost, causing the evaporator to frost. As the frost layer continues to thicken, the heat transfer resistance between the outdoor air and the refrigerant will gradually increase. At the same time, the frost layer will also hinder the flow of air between the heat exchange fins, causing the heat exchange performance of the outdoor heat exchanger to deteriorate. This will cause the evaporation temperature of the air source heat pump system to drop, and the compressor suction pressure to drop, which will lead to increased compressor operating energy consumption and a decrease in the system performance index (COP). In severe cases, it may even cause shutdown, which greatly reduces the quality of the air source heat pump.

[0003] Therefore, timely defrosting control of air source heat pumps contributes to the efficient operation of the heating system and the protection of the units. However, in the actual operation of air source heat pump units, "defrosting without frost" and "frosting without defrosting" often occur, which cannot achieve the purpose of timely defrosting control. Therefore, a precise on-demand defrosting control method is needed to improve the "false defrosting" behavior. The traditional defrosting control method is to collect data on frosting factors such as air temperature, humidity, pressure, wind speed, compressor power, and air cleanliness to determine the degree of frost and the starting point of the defrost operation. The degree of defrosting is determined by the changes in frosting-related quantities such as evaporator temperature, heating capacity, and compressor power, and the end point of defrosting is determined. This type of method is widely used in air source heat pump defrost control due to its simple operation. However, this type of method is greatly affected by environmental parameters, and the detection equipment also has problems such as reduced detection accuracy and detection loss after long-term use.

[0004] On the other hand, the image recognition method can also be used to identify the frosting state of the evaporator surface of the air source heat pump. When the evaporator surface is frosted, the roughness, uniformity and fineness of the surface change, that is, the texture features of the evaporator surface change. The image recognition method determines whether the air source heat pump needs to be defrosted by identifying the texture features. The advantage of image recognition is that the recognition accuracy is high, but there are also some shortcomings that affect the accuracy of image recognition. For example, the image recognition model will affect the image recognition accuracy due to the selection of setting parameters during the training process, and the quality of the collected image will affect the extraction of texture features during the image acquisition process, thereby affecting the accurate recognition of the frosting state of the evaporator surface. CN107506699A discloses a SAR image classification method based on texture features and DBN. The method extracts GLCM features and GMRF features from sample images and combines them into a new combined feature. The classification results are obtained by training using a deep belief network (DBN) to achieve the purpose of image recognition, but ignores the influence of sample image quality and training model parameters on recognition accuracy. Summary of the Invention

[0005] In view of the above problems, the present invention provides an air source heat pump defrosting control method based on texture features and HBA-DELM algorithm, which is used to solve the problem of incorrect defrosting of air source heat pumps in winter.

[0006] The present invention adopts the following technical solutions:

[0007] A defrosting control method for an air source heat pump based on texture features and an HBA-DELM algorithm comprises the following steps:

[0008] S1. Collecting sample images, and dividing the collected sample images into positive and negative data sets based on critical defrost control points to form an image data set;

[0009] S2. performing image preprocessing and illumination preprocessing on the sample images in the image dataset to obtain a preprocessed image dataset;

[0010] S3, extracting local HOG features and global GLCM features from the sample images in the preprocessed image dataset, and performing weighted fusion on the two features to obtain multi-scale fusion feature information;

[0011] S4. Cross-validation of deep extreme learning machine is used to perform feature training on the extracted multi-scale fusion feature information, and the deep extreme learning machine is optimized by the Honey Badger algorithm to construct the HBA-DELM algorithm model;

[0012] S5. Use the HBA-DELM algorithm model to perform defrost detection on the air source heat pump system;

[0013] S6. Perform defrost control based on the defrost detection result in step S5.

[0014] Preferably, in step S1, the sample images are images of the evaporator fins at different time periods, different operating states and different ambient temperatures under natural light, and the negative sample images contained in the negative data set are images of the evaporator fins collected when the air source heat pump system needs to perform a defrost operation.

[0015] Preferably, the image preprocessing in step S2 includes: grayscale conversion of the sample image, compression of the image grayscale and data enhancement.

[0016] Preferably, in step S2, the light preprocessing includes:

[0017] S2.1. Select a sample image in RGB format at a certain time point under standard illumination in the image dataset;

[0018] S2.2. Convert the sample image in RGB format to a sample image in YUV color space;

[0019] S2.3. Extracting the brightness component V from the sample image converted to the YUV color space;

[0020] S2.4. Arrange all the extracted brightness components V in descending order, and calculate the brightness average value based on the brightness components before and after the middle brightness component, which account for 25% of the total brightness components. The brightness average value V ave is calculated as follows:

[0021] V ave =V ref / V num ,

[0022] Among them, V ave Indicates the average brightness, V ref Indicates the total brightness value of the brightness components before and after the middle brightness component, which account for 25% of the total brightness component. num Represents the total number of pixels of the brightness component before and after the middle brightness component, which accounts for 25% of the total brightness component, and defines the illumination compensation coefficient:

[0023] L cf =255 / V ave ;

[0024] S2.5. Repeat steps S2.1 to S2.4 until the illumination compensation coefficient L at all time nodes in the sample image is obtained. cf ;

[0025] S2.6. Create an empty set Used to store the illumination compensation coefficient L at each time node cf Right now To form a lighting compensation template library;

[0026] S2.7. Perform illumination compensation on the sample images in the image dataset based on the illumination compensation coefficients in the illumination compensation template library, thereby obtaining a preprocessed image dataset.

[0027] Preferably, the steps between step S4 and step S5 further include:

[0028] Based on the illumination compensation coefficients in the illumination compensation template library described in step S2.6, illumination compensation is performed on the image with poor lighting at the corresponding time node during the actual image acquisition process.

[0029] Preferably, in step S3, the step of extracting local HOG features includes:

[0030] A. Calculate the gradient magnitude and gradient direction angle of each pixel of the sample image to obtain the local contour information of the sample image. The specific calculation method of the gradient magnitude and gradient direction angle is as follows:

[0031] I x (x,y)=I(x+1,y)-I(x-1,y),

[0032] I y (x,y)=I(x,y+1)-I(x,y-1),

[0033]

[0034]

[0035] Among them, I x (x,y) and I y (x,y) represents the gradient in the x and y directions respectively, M(x,y) represents the gradient amplitude, and δ(x,y) represents the gradient direction angle;

[0036] B. Divide the image into multiple cell units according to the local contour information of the sample image, and the pixel size of the cell unit is m*m;

[0037] C. Statistical gradient histogram of each cell unit;

[0038] D. Group each n*n cell unit into a sliding image block of (m*n)*(m*n) pixels, and calculate the gradient features within each sliding image block;

[0039] E. Combine the gradient features in all sliding image blocks to obtain the local HOG features of the sample image.

[0040] Preferably, in step S3, the step of extracting global GLCM features includes:

[0041] a. Calculate the joint probability density of the sample image in multiple preset angle directions respectively;

[0042] b. calculating the texture features corresponding to the sample image in the multiple preset angular directions based on the joint probability density in the multiple preset angular directions, wherein the texture features are nonlinear correlation statistics;

[0043] c. Calculate the mean and standard deviation of the nonlinear correlation statistics of the sample image in multiple preset angle directions to obtain the global GLCM features.

[0044] Preferably, in step S3, the expression of the multi-scale fusion feature information is:

[0045] T=β1T HOG +β2T GLCM ,

[0046] Among them, T represents the multi-scale fusion feature information, β1 and β2 are weight coefficients, T HOG represents the local HOG feature, T GLCM Represents the global GLCM feature.

[0047] Preferably, in step S4, the step of optimizing the deep extreme learning machine using the honey badger algorithm includes:

[0048] S4.1, initialize the population and random positions;

[0049] S4.2. Use the fitness function to calculate the position X of each honey badger i The corresponding fitness value f i ;

[0050] S4.3. Best location to save prey X prey , and then get the target fitness f prey ;

[0051] S4.4. Update the fitness value of the honey badger individual, compare the updated fitness value with the fitness value before the update, if the updated fitness value is less than the fitness value before the update, then use the updated fitness value as the fitness value of the honey badger individual, and use the updated position of the honey badger individual corresponding to the updated fitness value as the position of the honey badger individual; otherwise, the honey badger individual maintains the fitness value and position before the update.

[0052] S4.5, repeat step S4.4 until the fitness value of the honey badger individual is less than or equal to the target fitness value f preyThis indicates that the honey badger individual has reached the optimal position for the prey, otherwise the iteration continues until the preset number of iterations is reached.

[0053] Preferably, the position update stage of the honey badger individual includes an exploration stage and an exploitation stage, and the position update mathematical model of the exploration stage is as follows:

[0054] X new =X prey +F*β*I*X prey +F*r3*α*d i *|cos(2π*r4)*[1-cos(2π*r5)]|,

[0055] The position update mathematical model of the mining stage is as follows:

[0056] X new =X prey +F*r6*α*d i ,

[0057] Among them, X new represents the updated position of the honey badger individual, X prey represents the optimal location of the prey, F represents the sign of changing the search direction, β≥1 represents the ability of the honey badger to obtain food, I represents the odor intensity of the prey, d i represents the distance between the prey and the individual position of the honey badger, α represents the decreasing density factor, and r3, r4, r5 and r6 are all random numbers between [0,1].

[0058] The beneficial effects of the present invention are as follows: the image recognition method is used for defrost control to reduce detection errors, and secondly, the HBA-DELM algorithm model optimized by the HBA algorithm is used to further improve the recognition accuracy of the frosting state on the evaporator surface, and finally, a lighting compensation coefficient template is set to improve the problem of image recognition errors caused by poor lighting, thereby achieving accurate defrost control. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0060] Figure 1 A flowchart of an air source heat pump defrosting control method based on texture features and HBA-DELM algorithm provided in Example 1;

[0061] Figure 2This is a structural diagram of the deep extreme learning machine (DELM) algorithm model;

[0062] Figure 3 This is a flowchart of the deep extreme learning machine (DELM) algorithm model optimized by the honey badger algorithm (HBA) according to the present invention;

[0063] Figure 4 This is a structural diagram of a frost state identification system for an air source heat pump provided in Example 2. DETAILED DESCRIPTION

[0064] The following describes the embodiments of the present invention through specific embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0065] Example 1:

[0066] like Figure 1 As shown, the present invention provides an air source heat pump defrosting control method based on texture features and HBA-DELM algorithm, which specifically includes:

[0067] In step S1, sample images are first collected and classified. Evaporator fin images are collected under natural lighting at different time periods, operating conditions, and ambient temperatures to construct an image dataset. The images are correctly divided into positive and negative datasets. The negative sample images in the negative dataset are images of the evaporator fins collected when the air source heat pump system requires defrosting.

[0068] Step S2, image dataset preprocessing includes image preprocessing and illumination preprocessing to construct a preprocessed image dataset, wherein the image preprocessing includes grayscale conversion of the sample image, compression of the image grayscale and data enhancement.

[0069] Data enhancement includes filtering the sample images and mirror-flipping them at angles of 90, 180, and 270 degrees to increase the sample size. To reduce computational complexity, the sample images are grayscale compressed to 16 levels. Preferably, the grayscale can be quantized to 32 levels to allow for more detailed observation of changes in the local frost level on the evaporator.

[0070] In step S2, illumination preprocessing involves collecting background information of sample images in the image dataset under a standard illumination environment, establishing illumination compensation coefficients at each time point to form an illumination compensation template, and then performing illumination compensation on the sample images in the image dataset based on the illumination compensation coefficients at the corresponding time points, thereby obtaining a preprocessed image dataset. The time points should include those with representative illumination distribution, such as morning, noon, and evening. The specific steps of illumination preprocessing include:

[0071] S2.1. Select a sample image in RGB format at a certain time point under a standard lighting environment in the image dataset;

[0072] S2.2. Convert the sample image in RGB format to a sample image in YUV color space;

[0073] S2.3. Extracting the brightness component V from the sample image converted to the YUV color space;

[0074] S2.4. Arrange all the extracted brightness components V in descending order, and calculate the brightness average value based on the brightness components before and after the middle brightness component, which account for 25% of the total brightness components. The brightness average value V ave is calculated as follows:

[0075] V ave =V ref / V num ,

[0076] Among them, V ave Indicates the average brightness, V ref Indicates the total brightness value of the brightness components before and after the middle brightness component, which account for 25% of the total brightness component. num Represents the total number of pixels of the brightness component before and after the middle brightness component, which accounts for 25% of the total brightness component, and defines the illumination compensation coefficient:

[0077] L cf =255 / V ave ;

[0078] S2.5. Repeat steps S2.1 to S2.4 until the illumination compensation coefficient L at all time nodes in the sample image is obtained. cf ;

[0079] S2.6. Create an empty set Used to store the illumination compensation coefficient L at each time node cf Right now To form a lighting compensation template library;

[0080] S2.7. Perform illumination compensation on the sample images with poor lighting at the corresponding time node based on the illumination compensation coefficient in the illumination compensation template library, thereby obtaining a preprocessed image dataset.

[0081] It should be noted here that the illumination compensation coefficient of a certain time node may not be included in the illumination compensation template library, and it will be gradually added in the subsequent actual image acquisition process to update the illumination compensation template library, so that images with poor lighting in the subsequent image acquisition process can also select the illumination compensation coefficient at the time node from the illumination compensation template library for illumination compensation.

[0082] Another way to improve lighting is to equip an auxiliary light source during image acquisition to maintain consistent ambient brightness during image acquisition to improve image lighting.

[0083] In step S3, the step of extracting the local HOG feature vector includes:

[0084] A. Calculate the gradient magnitude and gradient direction angle of each pixel of the sample image to obtain the local contour information of the sample image. The specific calculation method of the gradient magnitude and gradient direction angle is as follows:

[0085] I x (x,y)=I(x+1,y)-I(x-1,y),

[0086] I y (x,y)=I(x,y+1)-I(x,y-1),

[0087]

[0088]

[0089] Among them, I x (x,y) and I y (x,y) represents the gradient in the x and y directions respectively, M(x,y) represents the gradient amplitude, and δ(x,y) represents the gradient direction angle;

[0090] B. Divide the image into multiple cell units according to the contour information of the sample image, and the pixel size of the cell unit is 8*8;

[0091] C. Statistical gradient histogram of each cell unit;

[0092] D. Each 2*2 cell unit is formed into a 16*16 pixel sliding image block, and the gradient features within each sliding image block are calculated;

[0093] E. Combine the gradient features of all sliding image blocks to obtain the local HOG feature vector of the sample image.

[0094] In step S3, the step of extracting the global GLCM feature vector includes:

[0095] a. Calculate the co-occurrence matrix of the sample image in four angular directions by rotating counterclockwise around the vertical axis. The image is compressed to 16 gray levels, the interval d is set to 1, and the sliding window size is 8*8. The joint probability density formula in each direction is as follows:

[0096] P 0° ={f(x1,y1)=i,f(x2,y2)=j; |x1-x2|=d&|y1-y2|=0},

[0097] P 45° ={f(x1,y1)=i,f(x2,y2)=j,x1-x2=y1-y2=d||x1-x2=y1-y2=-d},

[0098] P 90° ={f(x1,y1)=i,f(x2,y2)=j; |x1-x2|=0&|y1-y2|=d},

[0099] P 135° ={f(x1,y1)=i,f(x2,y2)=j;x1-x2=d,y1-y2=-d||x1-x2=-d,y1-y2=d},

[0100] Among them, P θ represents the joint probability density in the direction of θ (θ is 0°, 45°, 90° and 135° respectively), f(x1,y1) represents the pixel point with coordinate position (x1,y1) and grayscale i, f(x2,y2) represents the pixel point with coordinate position (x2,y2) and grayscale j, d is the distance between them, and θ is the direction angle.

[0101] b. Calculate the texture features of the sample image in multiple preset angular directions based on the joint probability density in multiple preset angular directions. The texture features are nonlinear correlation statistics. The nonlinear correlation statistics include contrast Con, homogeneity Homogeneity, correlation Corr, and angular second moment ASM. The specific calculation formula is as follows:

[0102] Contrast:

[0103] Homogeneity:

[0104] Dependencies:

[0105] Angular second moment:

[0106] Among them, in the correlation formula, μ i 、μ j , σ i and σ j The definitions are:

[0107]

[0108]

[0109]

[0110]

[0111] Where, P θ,d (i, j) represents the joint probability of gray level j appearing with i as the starting point when the spatial distance d and direction θ are given, μ i and μ j are the mean values of pixels with grayscale i and j, σ i and σ j are the variances of pixels with grayscale i and j, respectively.

[0112] c. Calculate the means and standard deviations of the four nonlinear correlation statistics in the four directions to reduce the differences in the nonlinear correlation statistics in different directions and form the GLCM feature vector.

[0113] Another method is to calculate the nonlinear correlation statistics corresponding to the sample image in multiple preset angular directions based on the joint probability density in multiple preset angular directions. The nonlinear correlation statistics include contrast Con, homogeneity Homogeneity, correlation Corr, and angular second moment ASM. The four nonlinear correlation statistics in the four directions are combined into a global GLCM feature (i.e., steps a and b of obtaining the GLCM feature vector in this embodiment). This method of extracting global GLCM features can obtain feature information in more dimensions.

[0114] The joint expression of the multi-scale fusion feature information in step S3 is: T = β1T HOG +β2T GLCM ,

[0115] Among them, T represents the multi-scale fusion feature information, β1 and β2 are weight coefficients, T HOG represents the local HOG feature, T GLCM Represents the global GLCM feature.

[0116] Step S4: Optimize the DELM mathematical model using the HBA algorithm to obtain an HBA-DELM algorithm model, and input the multi-scale fusion feature information T extracted above into the HBA-DELM algorithm model for training.

[0117] The Honey Badger Algorithm (HBA) optimization is to perform optimization by simulating the intelligent foraging behavior of the honey badger. It has the characteristics of strong optimization ability and fast convergence speed. The steps of the honey badger foraging behavior are as follows:

[0118] S4.1, initialize the population and random positions;

[0119] S4.2. Use the fitness function to calculate the position X of each honey badger i The corresponding fitness value f i , where the fitness function is the classification error rate of the training set and the validation set:

[0120] fitness=2-acc0-acc1,

[0121] Among them, acc0 is the accuracy of the training set, and acc1 is the accuracy of the validation set;

[0122] S4.3. Best location to save prey X prey , and then get the target fitness f prey ;

[0123] S4.4. Update the fitness value of the honey badger individual and compare the updated fitness value with the fitness value before the update. If the updated fitness value is less than the fitness value before the update, the updated fitness value will be used as the fitness value of the honey badger individual and the position of the honey badger individual. Otherwise, the honey badger individual will maintain the position and fitness value before the update.

[0124] S4.5, repeat step S4.4 until the updated fitness value is less than or equal to the target fitness value f prey This indicates that the honey badger individual has reached the optimal position of the prey, otherwise it continues to search iteratively until the condition is met or the maximum number of iterations is reached.

[0125] The step of determining the updated individual position of the honey badger corresponding to the updated fitness value in the above step S4.4 includes: first determining the position update stage of the honey badger individual, and adopting the corresponding position update mathematical model according to the position update stage of the honey badger individual, and then updating the position of the honey badger individual based on the position update mathematical model to obtain the updated individual position of the honey badger.

[0126] In the foraging behavior of honey badgers, they forage based on the smell of their prey. The intensity of their foraging is related to the concentration of their prey and the distance between them. The formula for the intensity of the prey's smell includes:

[0127]

[0128] S=(X i -X i+1 ) 2 ,

[0129] d i =X prey -X i ,

[0130] Where I is the scent intensity of the prey; S is the concentration intensity of the prey; d i Is the best prey X prey With honey badger individual X i distance; r1 is a random number in the range [0,1].

[0131] At the same time, a decreasing density factor α is used for iteration to ensure a smooth transition from exploration to mining;

[0132]

[0133] Where t represents the current number of iterations, t max is the maximum number of iterations, C≥1 and C=2 in this embodiment.

[0134] The position update process in the HBA optimization algorithm is divided into two parts, namely the "exploration phase" and the "exploitation phase". When r ≤ 0.5, the exploration phase begins. Otherwise, the honey badger population follows the honey guide honey badger to the hive and enters the exploitation phase. Using a parameter F that changes the search direction allows honey badgers to take advantage of the opportunity to strictly scan the search space. F is the flag for changing the search direction. Its mathematical expression is as follows:

[0135]

[0136] When r2≤0.5, the individual enters the exploration phase, and the position update is performed using the following mathematical model:

[0137] X new =X prey +F*β*I*X prey +F*r3*α*d i *|cos(2π*r4)*[1-cos(2π*r5)]|,

[0138] On the contrary, the honey badger follows the honey guide to reach the prey location for mining. The mathematical expression is as follows:

[0139] X new =X prey +F*r6*α*d i ,

[0140] Among them, X new represents the updated position of the honey badger individual, X prey represents the optimal location of the prey, F represents the sign of changing the search direction, β≥1 represents the ability of the honey badger to obtain food, I represents the odor intensity of the prey, d i represents the distance between the prey and the current position of the honey badger individual, α represents the decreasing density factor, and r2, r3, r4, r5 and r6 are all random numbers between [0,1].

[0141] Preferably, each honey badger has a different ability to obtain food, and its ability is related to the above-mentioned prey intensity. The expression of the ability β of the honey badger to obtain food can also be set as r7 is a random number between [0,1].

[0142] like Figure 2 As shown in the figure, the DELM network model structure consists of an input layer, several hidden layers, and an output layer. First, multiple extreme learning machine-autoencoders (ELM-AEs) are used for unsupervised pre-training. Then, the output weights of each ELM-AE are used to initialize the entire DELM network model. Meanwhile, the least squares method is used to update parameters during the unsupervised pre-training of the ELM-AEs. However, only the output layer weight parameters are updated, while the input layer weights and biases are not. As a result, the recognition effect of the final DELM network model is affected by the random input weights and random biases of each ELM-AE. Therefore, the Honey Badger Algorithm (HBA) is used to optimize the input weights and random bias parameters to improve the network accuracy of the DELM.

[0143] Honey Badger Algorithm (HBA) optimization flow chart is as follows Figure 3 As shown, Figure 3 On the left is the defrosting process of the HBA-DELM algorithm model. First, the data is input and preprocessed. Second, the HBA algorithm is used to optimize the DELM model to obtain the optimal parameters and build the optimal DELM model. The optimal model recognizes the preprocessed input data. If the recognition result meets the constraints, defrosting is performed. If the constraints are not met, the next round of data input and recognition is restarted. Figure 3 On the right is the HBA algorithm used to optimize the DELM algorithm step in the defrosting process. The HBA algorithm optimization process includes: HBA parameter initialization, followed by establishing a fitness function to describe the relationship between honey badgers and prey. Then the honey badger population and the guide honey badger individual update their positions according to the prey intensity and concentration. The next step is to calculate the updated fitness value and update the current optimal position. If the updated optimal position meets the constraints, the optimal parameters are output. Otherwise, the honey badger position update cycle continues to meet the constraints and output the optimal parameters.

[0144] In step S5, the trained and optimized HBA-DELM algorithm model performs online, real-time monitoring of the evaporator's frosting status. First, real-time image acquisition is performed, and illumination compensation is performed on poorly lit images captured during the actual image acquisition process. Illumination compensation coefficients at corresponding time points are selected from the illumination compensation template library to compensate for these poorly lit images, thereby improving image recognition accuracy during real-time monitoring. Multi-scale fusion feature information is then extracted from the compensated images and input into the trained HBA-DELM algorithm model, enabling accurate detection of the evaporator's frosting status and, based on the detection results, precise defrost control of the air-source heat pump.

[0145] Example 2:

[0146] like Figure 4 As shown, the present invention also provides an air source heat pump frosting state recognition system, based on the air source heat pump defrosting control method based on texture features and HBA-DELM algorithm described in Example 1, the system includes an image data acquisition module S001, an image data processing module S002, a feature extraction and feature fusion module S003, and a state recognition module S004 connected in sequence;

[0147] The image data acquisition module S001 is used to collect the original images of the evaporator fins at various operating times;

[0148] The image data processing module S002 is used to perform a series of image preprocessing on the collected original image of the evaporator fin, such as image grayscale conversion, image brightness compensation, image grayscale compression, and obtain a preprocessed image set of the evaporator fin after illumination preprocessing;

[0149] The feature extraction and feature fusion module S003 is used to extract local HOG features and global GLCM features from the preprocessed image, and fuse the extracted features through a feature fusion mechanism to obtain a multi-scale fusion feature matrix of the evaporator frosting state;

[0150] The state recognition module S004 inputs the obtained fusion feature matrix into the deep extreme learning machine training model, thereby outputting the evaporator frosting state recognition result.

[0151] The intelligent defrosting module S005 performs defrosting control according to the recognition result of the state recognition module, and performs defrosting if the collected image is diagnosed as a negative sample image.

[0152] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes or substitutions that are not conceived through creative work should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection defined in the claims.

Claims

1. A defrosting control method for air source heat pump based on texture features and HBA-DELM algorithm, characterized in that: The following steps are involved: S1. Collecting sample images, and dividing the collected sample images into positive and negative data sets based on critical defrost control points to form an image data set; S2. performing image preprocessing and illumination preprocessing on the sample images in the image dataset to obtain a preprocessed image dataset; S3, extracting local HOG features and global GLCM features from the sample images in the preprocessed image dataset, and performing weighted fusion on the two features to obtain multi-scale fusion feature information; S4. Cross-validation of deep extreme learning machine is used to perform feature training on the extracted multi-scale fusion feature information, and the deep extreme learning machine is optimized by the Honey Badger algorithm to construct the HBA-DELM algorithm model; S5. Use the HBA-DELM algorithm model to perform defrost detection on the air source heat pump system; S6. Performing defrost control based on the defrost detection result in step S5; In step S4, the step of optimizing the deep extreme learning machine using the honey badger algorithm includes: S4.1, initialize the population and random positions; S4.

2. Use the fitness function to calculate the position X of each honey badger i The corresponding fitness value f i ; S4.

3. Best location to save prey X prey , and then get the target fitness f prey ; S4.

4. Update the fitness value of the honey badger individual, and compare the updated fitness value with the fitness value before the update. If the updated fitness value is less than the fitness value before the update, the updated fitness value is used as the fitness value of the honey badger individual, and the updated position of the honey badger individual corresponding to the updated fitness value is used as the position of the honey badger individual. Otherwise, the honey badger individual maintains the fitness value and position before the update; S4.5, repeat step S4.4 until the fitness value of the honey badger individual is less than or equal to the target fitness f prey This indicates that the honey badger individual has reached the optimal position for the prey, otherwise the iteration continues until the preset number of iterations is reached.

2. The air source heat pump defrosting control method based on texture features and HBA-DELM algorithm according to claim 1, characterized in that: In step S1, the sample images are images of the evaporator fins at different time periods, different operating states and different ambient temperatures under natural light. The negative sample images contained in the negative data set are images of the evaporator fins collected when the air source heat pump system needs to perform a defrost operation.

3. The air source heat pump defrosting control method based on texture features and HBA-DELM algorithm according to claim 1, characterized in that: The image preprocessing in step S2 includes: grayscale conversion of the sample image, compression of the image grayscale and data enhancement.

4. The air source heat pump defrosting control method based on texture features and HBA-DELM algorithm according to claim 1, characterized in that: In step S2, the illumination preprocessing includes: S2.

1. Select a sample image in RGB format at a certain time point under standard illumination in the image dataset; S2.

2. Convert the sample image in RGB format to a sample image in YUV color space; S2.

3. Extracting the brightness component V from the sample image converted to the YUV color space; S2.

4. Arrange all the extracted brightness components V in descending order, and calculate the brightness average value based on the brightness components before and after the middle brightness component, which account for 25% of the total brightness components. The brightness average value V ave is calculated as follows: V ave =V ref / V num , Among them, V ave Indicates the average brightness, V ref Indicates the total brightness value of the brightness components before and after the middle brightness component, which account for 25% of the total brightness component. num Represents the total number of pixels of the brightness component before and after the middle brightness component, which accounts for 25% of the total brightness component, and defines the illumination compensation coefficient: L cf =255 / V ave ; S2.

5. Repeat steps S2.1 to S2.4 until the illumination compensation coefficient L at all time nodes in the sample image is obtained. cf ; S2.

6. Create an empty set Used to store the illumination compensation coefficient L at each time node cf Right now To form a lighting compensation template library; S2.

7. Perform illumination compensation on the sample images in the corresponding image dataset based on the illumination compensation coefficients in the illumination compensation template library, thereby obtaining a preprocessed image dataset.

5. The air source heat pump defrosting control method based on texture features and HBA-DELM algorithm according to claim 4, characterized in that: The steps between step S4 and step S5 also include the following steps: Perform illumination compensation on the image actually acquired at the corresponding time node based on the illumination compensation coefficient in the illumination compensation template library described in step S2.

6.

6. The air source heat pump defrosting control method based on texture features and HBA-DELM algorithm according to claim 1, characterized in that: In step S3, the step of extracting local HOG features includes: A. Calculate the gradient magnitude and gradient direction angle of each pixel of the sample image to obtain the local contour information of the sample image. The specific calculation method of the gradient magnitude and gradient direction angle is as follows: I x (x,y)=I(x+1,y)-I(x-1,y), I y (x,y)=I(x,y+1)-I(x,y-1), Among them, I x (x,y) and I y (x,y) represents the gradient in the x and y directions respectively, M(x,y) represents the gradient amplitude, and δ(x,y) represents the gradient direction angle; B. Divide the image into multiple cell units according to the local contour information of the sample image, and the pixel size of the cell unit is m*m; C. Statistical gradient histogram of each cell unit; D. Group each n*n cell unit into a sliding image block of (m*n)*(m*n) pixels, and calculate the gradient features within each sliding image block; E. Combine the gradient features in all sliding image blocks to obtain the local HOG features of the sample image.

7. The air source heat pump defrosting control method based on texture features and HBA-DELM algorithm according to claim 1, characterized in that: In step S3, the step of extracting the global GLCM features includes: a. Calculate the joint probability density of the sample image in multiple preset angle directions respectively; b. calculating the texture features corresponding to the sample image in the multiple preset angular directions based on the joint probability density in the multiple preset angular directions, wherein the texture features are nonlinear correlation statistics; c. Calculate the mean and standard deviation of the nonlinear correlation statistics of the sample image in multiple preset angle directions to obtain the global GLCM features.

8. The air source heat pump defrosting control method based on texture features and HBA-DELM algorithm according to claim 1, characterized in that: In step S3, the expression of the multi-scale fusion feature information is: T=β1T HOG +β2T GLCM , Among them, T represents the multi-scale fusion feature information, β1 and β2 are weight coefficients, T HOG represents the local HOG feature, T GLCM Represents the global GLCM feature.

9. The air source heat pump defrosting control method based on texture features and HBA-DELM algorithm according to claim 1, characterized in that: The position update phase of a honey badger includes an exploration phase and an exploitation phase. The mathematical model for position update in the exploration phase is as follows: X new =X prey +F*β*I*X prey +F*r3*a*d i *|cos(2π*r4)*[1-cos(2π*r5)]|, The position update mathematical model of the mining stage is as follows: X new =X prey +F*r6*α*d i , Among them, X new represents the updated position of the honey badger individual, X prey represents the optimal location of the prey, F represents the sign of changing the search direction, β≥1 represents the ability of the honey badger to obtain food, I represents the odor intensity of the prey, d i represents the distance between the prey and the honey badger individual, α represents the decreasing density factor, and r3, r4, r5 and r6 are all random numbers between [0,1].

Images