A rapid detection system and method for the failure risk of a solar heat collection field

Abstract

The present invention discloses a method for quickly detecting the fault risk of a solar heat collection field, which includes the following steps; Step 1: By setting an infrared thermal imaging camera for real-time monitoring, an overall temperature distribution thermal map of the solar heat collection field is obtained through shooting; Step 2: Establish communication between the infrared thermal imaging camera and the computer to collect temperature images of the temperature state of the heat collection field; Step 3: Preprocess and regionally divide the overall temperature distribution thermal map of the heat collection field, and perform feature extraction to obtain a scatter plot of the fault temperature field of the collectors in each region; Step 4: Import the extracted feature information of the temperature field distribution and related characteristic parameters in each region into the feature library of the solar collector fault model, and perform two matching analyses of the feature library of the fault model in sequence, so as to improve the fault detection efficiency of the heat collection field, reduce the false alarm rate and the missed alarm rate, and ensure the safe operation of the solar heat collection system. The present invention has the characteristics of simple operation and no need to upgrade the system and add new sensors.

Description

Technical Field

[0001] The present invention relates to the technical field of large-scale solar heat collection field fault detection, and particularly relates to a method for rapidly detecting the fault risk of a solar heat collection field. Background Art

[0002] The solar heat collection field is the core component of a solar heating system and the main energy source of the system. The temperature of the working medium in the solar heat collection field is greatly affected by meteorological factors such as solar radiation intensity and outdoor air temperature, and is constantly changing. In areas with strong solar radiation and large day-night temperature differences, the temperature fluctuation of the working medium in the heat collection field intensifies, and the temperature change rate surges sharply, resulting in uneven flow distribution, large resistance change, and severe pressure fluctuation in the heat collection field. Overheating, leakage, aging, and failure are frequent faults of the collectors in the solar heat collection field. If the fault source cannot be detected and identified in the initial stage of the fault or even before the fault occurs, it will cause the heat collection capacity to decline, seriously affect the system operation performance, and reduce the reliability of the solar heating system. Therefore, accurately and rapidly detecting, identifying, and locating the faults of the solar heat collection field is an urgent task.

[0003] The existing fault detection methods for heating systems and heat collection systems generally monitor sensors such as temperature, flow rate, and pressure, and analyze the possible trends of the influence of expert knowledge and engineering experience on system faults, and use the analysis results as the basis for system fault diagnosis. This method diagnoses and analyzes the faults of the system through the correlation between the thermodynamic and hydraulic characteristics and physical signals of the system. It needs to introduce electrical signals into the interior of the collector and contact with its structural materials, and has high requirements for the number of system sensors and data acquisition conditions. Either the sensors are densely arranged in a mesh, or there are large blind areas in the system links. When a fault occurs, if the fault point is far from the monitoring point and the fault degree is not large, the signal monitoring will be seriously lagged; generally, a type of fault may cause changes in multiple characteristic parameters, so the credibility of the detection results of a single sensor is not strong; and the cost of adding sensors is high. Looking at the current engineering applications, the setting of the data acquisition points of the sensors only depends on whether there is a control requirement, and the number of points is scarce, and it is still difficult to meet the data requirements for system fault diagnosis. Summary of the Invention

[0004] In order to overcome the above-mentioned deficiencies of the prior art, the purpose of the present invention is to provide a method for rapidly detecting the fault risk of a solar heat collection field, which has the characteristics of not contacting the structure to be measured, not being restricted by the connection method of the measured heat collection field, being able to comprehensively analyze and characterize the temperature field distribution characteristics of the heat collection field, and detecting the corresponding fault risk accordingly, and has the advantages of simple operation and no need to upgrade and transform the system to add sensors. <H

[0005] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0006] A rapid detection system for the failure risk of a solar heat collection field, comprising a solar heat collection field body, an infrared thermal imaging camera and a computer processing device;

[0007] The infrared thermal imaging camera monitors the solar heat collection field body online in real time, takes an overall temperature distribution thermal map of the solar heat collection field according to a set fixed time interval, and transmits it to the computer processing device. The computer processing device preprocesses and divides the area of the temperature distribution thermal map, and extracts information to obtain the fault temperature field characteristic information of each area collector, and matches it with the characteristic information in the typical fault model feature library of the solar collector, so as to complete the detection of the fault type and degree of the solar collector.

[0008] The infrared thermal imaging camera is set at a position directly facing the solar heat collection field. Considering the problem that the spacing and inclination of each row of collectors in different heat collection fields are inconsistent, there may be occlusion of the front row on the rear row collectors in some areas. The position of the infrared thermal imaging camera is set according to the principle of the front and, if necessary, the side, to ensure that each collector is within the monitoring range of the camera.

[0009] A rapid detection method for the failure risk of a solar heat collection field, comprising the following steps;

[0010] Step 1: Temperature image acquisition;

[0011] Set the infrared thermal imaging camera to monitor in real time to obtain an overall temperature distribution thermal map of the solar heat collection field;

[0012] Step 2: Temperature image reading;

[0013] Establish communication between the infrared thermal imaging camera and the computer. The infrared thermal imaging camera and the computer are in the same local area network. Download the PC software corresponding to the camera on the computer and match the IP to achieve real-time monitoring, and collect temperature images of the temperature state of the heat collection field;

[0014] Step 3: Data acquisition, analysis and processing;

[0015] Preprocess and divide the area of the overall temperature distribution thermal map of the heat collection field, and perform feature extraction to obtain a scatter plot of the fault temperature field of each area collector;

[0016] Step 4: Fault model matching;

[0017] Import the extracted temperature field distribution characteristic information and related characteristic parameters of each area into the solar collector fault model feature library, and perform two matching analyses of the fault model feature library in sequence to improve the fault detection efficiency of the heat collection field, reduce the false alarm rate and missed alarm rate, and ensure the safe operation of the solar heat collection system.

[0018] The preprocessing part of the said Step 3 includes the following steps:

[0019] The original infrared thermal image is preliminarily preprocessed by an infrared image normalization method based on temperature difference to eliminate the influence of thermal environment factors on the temperature field of the solar collector; the images of the solar collector under different environmental factors are regarded as basic signals and processed in the spatial domain to achieve the purpose of eliminating the influence of the external thermal environment;

[0020] Then, further processing of adaptive dynamic range compression based on improved guided filtering and Sigmoid function is carried out to perform dynamic range compression and image enhancement processing on the image, and obtain a more natural result that conforms to human visual characteristics and has obvious characteristics.

[0021] The guided filter decomposes the input infrared image into a base layer B and a detail layer D. The base layer B is the processing result of the improved guided filter for the input image, which contains high-contrast information and most of the dynamic range in the input image; the detail layer D is the difference between the input image and the base layer B, which contains low-contrast details in the input image;

[0022] B = G[I h

[0023] D = I h - B

[0024] where I h is the input infrared thermal image; the symbol G[*] represents the improved guided filter, and an edge-aware weight is constructed using a variant of the first-order differential of the image;

[0025] Then, the Sigmoid adaptive compression algorithm is used to compress the dynamic range of the base layer B. On the premise of retaining the overall visual perception of the image, the gray information of the high-dynamic range image is compressed so that it can be normally displayed on traditional display devices;

[0026] B' = S[B]

[0027] where the symbol S[*] represents the Sigmoid algorithm; B' is the base layer after dynamic range compression, and the value of B' is between [0,1];

[0028] The detail layer is strengthened using a fixed gain coefficient η, and the enhanced detail layer D' is

[0029] D' = η * D

[0030] The setting of the fixed gain coefficient is recommended to be between [1, 2];

[0031] Finally, a suitable dynamic range is allocated to the base layer B' after dynamic range compression, and it is recombined with the enhanced detail layer D' to obtain the output image;

[0032] I l ​= a * B' + D'

[0033] Where I l is the output image. The parameter a is used to control the size of the dynamic range of the base layer. Using the setting value of a = 220 can avoid possible dynamic range overflows and make full use of the display dynamic range of the image, and the degree of detail enhancement is relatively reasonable, and more natural results that conform to human visual characteristics can be obtained.

[0034] The solar collector fault model feature library in the fourth step includes four typical fault judgment model feature libraries of the solar heat collection field based on the regression analysis model and three fault degree model feature libraries of typical faults in the solar heat collection field based on the deep convolutional neural network;

[0035] Four typical fault judgment model feature libraries of the solar heat collection field based on the regression analysis model, which are used to judge the four typical faults (failure, overheating, leakage, aging) of the solar heat collection field. Based on the characteristics of the solar heat collection field, a semi-physical model is selected. By analyzing the solar energy system, characteristic parameters with typical significance are selected. Based on the polynomial regression model, a model is established using the operation data under normal operating conditions of the system. The irradiance and flow rate of the solar heat collection system are selected as the input variables of the model. The expression is:

[0036] Y = f(G, Q) + ε = a0 + a1G + a2Q + a3G 2 + a4Q 2 + a5GQ

[0037] Where Y is the regression variable output model; ε is the error term; a0 a5 are the correlation coefficients of the regression model; G is the irradiance, the input parameter of the regression model, W / m 2 ; Q is the flow rate, the input parameter of the regression model, t / h;

[0038] After regression analysis, the predicted values of the model are compared with the collector temperature values under four typical fault operating conditions collected. Taking a 90% confidence level, the characteristic parameter residuals and corresponding thresholds are determined, and the residuals of the characteristic parameters are analyzed to realize the monitoring and identification of faults;

[0039] Three fault degree model feature libraries of typical faults (overheating, leakage, aging) in the solar heat collection field based on the deep convolutional neural network. The data features are extracted through one-dimensional deep convolutional neural network training, and the fault degree at the time of each fault occurrence is determined.

[0040] The four typical fault judgment model feature libraries of the solar heat collection field based on the regression analysis model:

[0041] Select a semi-physical model, select typical characteristic parameters through the analysis of the solar energy system, and establish a model based on the regression model of polynomials using the operation data under normal operating conditions of the system. Irradiance and flow rate have a great impact on the operation of the solar collector system. These two parameters are selected as the input variables of the model, and the expression is:

[0042] Y = f(G, Q) + ε = a0 + a1G + a2Q + a3G 2 + a4Q 2 + a5GQ

[0043] where ε is the error term; Y is the output model of the regression variable; a0 a5 are the correlation coefficients of the regression model; G is the irradiance, the input parameter of the regression model, W / m 2 ; Q is the flow rate, the input parameter of the regression model, t / h. The model expression of the characteristic parameters represents a function of the flow rate and irradiance;

[0044] According to the variable characteristics and relevant experience, select a binary polynomial model; use the least squares method to solve the correlation coefficients of the expression. The method adopted is to regard the regression matrix used for modeling as the row vector X. Using matrix notation, the polynomial model expression can be written as:

[0045] Y = aX + e

[0046] In the formula, X is the regression input variable vector; Y is the dependent output variable vector; a is the regression model coefficient vector; e is the error term vector; measurement errors and model errors are included in X and Y;

[0047] When the regression variables of the model are linearly independent, the residuals of the model tend to be stable; when the error terms follow a normal distribution, the least squares method is the best prediction method. At this time, the sum of the squares of the regression residuals reaches the minimum. The formula used to solve the regression coefficients is:

[0048]

[0049] In the formula, a OLS is the regression model coefficient vector; X reg is the regression matrix of the modeling data; is the transpose matrix of X reg ; Y reg is the regression vector of the dependent variable; a OLS is the coefficient vector obtained by solving the model using the least squares method. After calculating using the above formula, the specific values of each coefficient are obtained. Furthermore, the mean square error of the residuals and the regression determination coefficient of the model are calculated. These two are important indicators for judging the regression accuracy of the model.

[0050] The fault degree model feature library of three typical faults of the solar collector field based on the deep convolutional neural network:

[0051] including the following steps:

[0052] Data preparation: Audit and verify the temperature data in the database, supplement and correct the missing data and inconsistent data to ensure the correctness of the data and the consistency of the logic;

[0053] ① Processing of missing data: Due to network transmission failures, missing values may appear in the database. For the missing data, an estimation method can be used to replace the missing values with the sample mean for supplementation;

[0054] ② Processing of inconsistent data: With the help of the integrity constraint mechanism of the database management system, check the inconsistent data and then correct it with reference to the corresponding data values in the database;[[ID=!2]]

[0055] Generating a data set: Divide the prepared data into training set labels and test set labels;

[0056] Generating test set data and labels: According to different operating conditions of the solar energy system, read the temperature data values corresponding to the operating conditions from the database to generate test set data and operating condition labels under various operating conditions; among them, the operating condition labels are the same as those of the training set operating condition labels and are automatically generated by the program according to the operating conditions;

[0057] Network training:

[0058] Input the training set data into the convolutional neural network, continuously perform convolution and pooling to obtain feature vectors, and send them into the fully connected network. By calculating the output of the network and the training set labels, the network error is obtained. Using the error backpropagation algorithm, continuously correct the network weights, biases, convolution coefficients, and pooling coefficients to make the error meet the set accuracy requirements. Then the network training is completed, and a convolutional neural network model, that is, a fault degree model for typical faults of the solar collector field based on the deep convolutional neural network, is generated.

[0059] The fault degree model for typical faults of the solar collector field based on the deep convolutional neural network is specifically as follows:

[0060] ① Read in a set of training set data d, whose size is [M×1×N], where M represents the size of the training batch, and 1×N represents one-dimensional training data;

[0061] ② Perform the first convolution operation on the read training data to obtain a feature map t. Initialize the coefficients of the convolution kernel g. Let the size of g be [P×1×Q], where P represents the number of convolution kernels, and [1×Q] represents the size of the convolution kernel. The obtained convolution result is t = Σ(d*g), and the size of the feature map is [M×1×N×Q];

[0062] ③Perform max pooling operation on the feature map t obtained from the convolution operation to obtain the feature map Z. Initialize the pooling coefficient, given the pooling stride p and the pooling window size k. The size of the finally obtained feature map Z is [M×1×(N / p)×Q]. The pooling process reduces the dimension of the data;

[0063] ④Repeat the above steps ②-③, repeatedly perform convolution and pooling operations to obtain the feature vector x. At this time, the feature extraction process of the convolutional neural network is completed;

[0064] ⑤Initialize the weight matrix w and the bias b of the fully connected network. Send the extracted feature vector x into the fully connected network and perform operations with the weight matrix w and the bias b to obtain the network output y = ∑(w×x + b);

[0065] ⑥Subtract the network output y obtained by the network from the training set label l to obtain the network error e = y - l. Take the derivative of the network error and use this derivative for backpropagation to sequentially correct the weight w, bias b of the fully connected network, the pooling coefficients of each layer, and the convolution coefficients of each layer;

[0066] ⑦Repeat the above process until the network error e meets the accuracy requirements. The network training process is completed, and a convolutional neural network model is generated.

[0067] Advantages of the present invention:

[0068] The present invention uses an infrared thermal imaging camera to monitor the state of the temperature field of a solar collector, having the advantages of a large monitoring area, fast response speed, high sensitivity, etc. The infrared thermal imaging camera can measure a temperature change of 0.1°C and can be used to accurately identify the fault information of the collector, that is, the fault type, fault location, and fault degree. Develop a large-scale solar collector field fault detection system, extract the fault feature information of the collector and compare and match it with the models in the typical fault feature library in sequence to complete the type judgment, position determination, and degree determination of the fault state. Description of the Drawings

[0069] Figure 1 It is a schematic diagram of the on-site of a large-scale solar collector field.

[0070] Figure 2 It is the common connection methods of solar collectors (single-row series, single-row parallel, series-parallel U-shaped, series-parallel Z-shaped).

[0071] Figure 3 It is the general connection form of a large-scale solar collector field.

[0072] Figure 4 It is the workflow block diagram of the large-scale solar collector field fault detection method based on infrared imaging.

[0073] Figure 5This is a block diagram of the normalization processing of infrared images of solar collector fields based on temperature difference (step three).

[0074] Figure 6 This is a block diagram of the adaptive dynamic range compression process (step three) based on improved guided filtering and Sigmoid function.

[0075] Figure 7 Flowchart of the solar collector field fault degree model (step four) based on deep convolutional neural network.

[0076] Figure 8 This is a flow chart of the training process algorithm (step four) of the convolutional neural network model.

[0077] Figure 9 This is a block diagram of a rapid fault detection method for large solar collector fields based on infrared imaging.

[0078] Figure 10 The thermal infrared images are shown in Figure a (original thermal infrared image; Figure b temperature-compensated infrared thermal image). Figure 1 Figure c: Thermal imaging based on adaptive dynamic range compression processing using improved guided filtering and sigmoid function Figure 2 ; Figure d is the thermal infrared imaging image after regional division).

[0079] Figure 11 This is the temperature field distribution characteristic information diagram of the solar collector area ① (Figure a is a scatter diagram of the solar collector temperature field distribution in the collector area ①; Figure b is a temperature residual deviation diagram of the solar collector temperature field in the collector area ①).

[0080] Figure 12 This is the temperature residual diagram at the 90% confidence level for this operating condition. DETAILED DESCRIPTION

[0081] The present invention will be further described in detail below in conjunction with the embodiments.

[0082] like Figure 1 Figure 1 shows a schematic diagram of a large solar thermal field. Western my country experiences strong solar radiation and large temperature differences between day and night. Currently, solar systems are widely used, and some large solar thermal fields exist, which are prone to overheating, leakage, aging, and failure.

[0083] like Figure 2 As shown, the solar collector connection method.

[0084] There are three main connection methods for flat-plate collectors: series connection, parallel connection and hybrid connection.

[0085] The connection where the outlet of one collector is connected to the inlet of another collector is called series connection. The connection where the outlet and inlet of one collector are respectively connected to the outlet and inlet of another collector is called parallel connection. A combination of series and parallel connections can be called hybrid connection, and according to the sequence of series and parallel connections, it can be further divided into series-parallel and parallel-series.

[0086] The flow resistance of the collector parallel system is relatively small. However, if the number of parallel collectors is too large, it may cause uneven flow distribution in the system, resulting in local overheating, which can lead to the aging of the system collectors and the occurrence of overheating failures. The flow resistance of the collector series system is greater, but the flow distribution is more uniform. As the number of series-connected collectors increases, the heat collection efficiency of the collectors along the way decreases. Therefore, to reduce the failure rate of collectors and ensure the efficient operation of the heat collection system, it is necessary to reasonably design the combined connection method of the heat collection field.

[0087] In addition to the influence of the collector connection method on the system operation, the layout of the flow direction in the heat collection field also has a certain impact on the efficiency of the heat collection system and the occurrence of collector failures. The layout of the flow direction in the heat collection field can be divided into U-shaped layout and Z-shaped layout according to whether the working medium flow directions in the supply and return water pipes are the same or opposite, as Figure 2 shown. In the U-shaped layout (i.e., the different-path system), the non-uniformity of the flow distribution in the heat collection field is greater, and the failure rate of overheating and aging caused by local overheating is higher. In the Z-shaped layout (i.e., the same-path system), the resistance of the heat collection system is greater, and the initial investment increases due to the increase in the pipeline length.

[0088] During the operation of the heat collection field, the non-uniformly distributed heat collection working medium absorbs heat while flowing. Coupled with the fluctuations in solar irradiance and ambient temperature, the resistance of each branch in the heat collection field fluctuates, so the flow distribution in the heat collection field is also affected and changes. And the change in flow distribution will in turn affect the change in the temperature of the heat collection working medium. Therefore, during the operation of the heat collection field, the hydraulic state and thermal state are constantly changing, which is prone to causing various types of collector failures.

[0089] As Figure 3 shown, the connection form of a general large-scale solar heat collection field is series-parallel U-shaped layout. Compared with the single series, parallel, U-shaped, and Z-shaped layouts, although it reduces the failure rate of various types of collectors, it cannot completely eliminate the occurrence of failures.

[0090] By setting an infrared thermal imaging camera in the solar heat collection field for real-time monitoring and synchronously transmitting the infrared thermal imaging map to the computer. Among them, the data of instruments such as irradiance meters and flow meters in the solar heat collection field are synchronized with the infrared imaging map records.

[0091] As Figure 4 shown, the framework diagram of a new type of solar heat collection field fault detection method.

[0092] Use an infrared thermal imaging camera to monitor the solar energy collection field in real time and obtain the original infrared thermal image; transmit the original infrared thermal image to a computer for preprocessing including infrared image normalization based on temperature difference and adaptive dynamic range compression processing based on improved guided filtering and the Sigmoid function; further divide the roughly regionalized infrared thermal image of the solar energy collection field to obtain the local infrared thermal image of each collector, and use infrared software to select features and collect data for it; then use MATLAB data analysis software to obtain the scatter plot of the temperature field distribution characteristics of each collector; finally, compare and analyze by combining the instrument monitoring data in the solar energy collection field with the characteristic information of two typical fault models of the solar energy collection field to determine the fault information and output it, providing technical support for the qualitative analysis of the faults in the solar energy collection field.

[0093] Preprocessing of infrared thermal image:

[0094] 1) Use the infrared image normalization method based on temperature difference to perform preliminary preprocessing on the original infrared thermal image transmitted from the infrared thermal imaging camera to the computer.

[0095] Based on the transformation method, use image processing means to eliminate the influence of the thermal environment factors on the temperature field of the solar collector in the preliminary preprocessing stage of the image. This method regards the images of the collector under different environmental factors as basic signals and processes them in the spatial domain to achieve the purpose of eliminating the influence of the external thermal environment.

[0096] Such as Figure 5 shown, when the inlet temperature of the collector is constant, by inputting the original thermal environment factors (ambient temperature, solar irradiance, etc.) during image acquisition, according to software simulation, the temperature T n at a certain location of the collector in the system under this thermal environment is obtained, and the difference T n between T m and the temperature T i at the same location of the collector under the standard thermal environment factors is calculated (T i = T n - T m ), which is the temperature compensation difference that needs to be applied to the original infrared thermal imaging image. The original infrared thermal image is transformed as a whole to the same thermal environment state as in the standard infrared thermal image.

[0097] 2) Perform further processing of adaptive dynamic range compression using improved guided filtering and the Sigmoid function.

[0098] Such as Figure 6 shown:

[0099] The input infrared image is decomposed into a base layer B and a detail layer D using guided filtering. The base layer B is the result of processing the input image with improved guided filtering, which contains the high-contrast information and most of the dynamic range in the input image; the detail layer D is the difference between the input image and the base layer B, which contains the low-contrast details in the input image.

[0100] B = G[I h

[0101] D = I h - B

[0102] where I h is the input infrared thermal image; the symbol G[*] represents the improved guided filtering, and an edge-aware weight is constructed using a variant of the first-order image differential.

[0103] Then, the dynamic range of the base layer B is compressed using the Sigmoid adaptive compression algorithm. While preserving the overall visual perception of the image, the gray-scale information of the high-dynamic-range image is compressed so that it can be normally displayed on a traditional display device.

[0104] B' = S[B]

[0105] where the symbol S[*] represents the Sigmoid algorithm; B' is the base layer after dynamic range compression, and the value of B' is between [0,1].

[0106] To improve the visibility of the low-contrast details, the detail layer is enhanced using a fixed gain coefficient η, and the enhanced detail layer D' is

[0107] D' = η * D

[0108] The setting of the fixed gain coefficient is recommended to be between [1, 2], which can effectively enhance the low-contrast detail information and avoid obvious artificial traces caused by over-enhancement.

[0109] Finally, an appropriate dynamic range is allocated to the base layer B' after dynamic range compression, and it is recombined with the enhanced detail layer D' to obtain the output image.

[0110] I l = a * B' + D'

[0111] where I l is the output image, and the parameter a is used to control the size of the dynamic range of the base layer. Using the setting value of a = 220 can avoid possible dynamic range overflow and make full use of the display dynamic range of the image, and the degree of detail enhancement is relatively reasonable, and a more natural result that conforms to human visual characteristics can be obtained.

[0112] Establish a feature library for typical fault models of solar collector fields: ​

[0113] The solar collector field fault detection system processes the simulation and actual measurement results of the collector under normal operating conditions and various fault conditions, extracts the characteristic information of the collector fault temperature field and its related characteristic parameters, and establishes a characteristic library of four types (failure, overheat, leakage, aging) of typical solar collector field fault judgment models based on the regression analysis model; the solar heating system fault detection model based on the deep convolutional neural network is a general fault detection method that conforms to the data characteristics of the solar heating system, has high diagnostic accuracy, and strong feature extraction ability. In the present invention, the temperature field data of the solar collector field are monitored and extracted in real time by an infrared thermal imaging camera, stored in a database, the data features are extracted by a one-dimensional deep convolutional neural network, and the deep convolutional neural network is trained with a temperature data set for multi-class fault mode recognition, and a characteristic library of three types (overheat, leakage, aging) of solar collector field detection fault degrees based on the deep convolutional neural network is established;

[0114] 1) Characteristic library of four types of typical solar collector field fault judgment models based on the regression analysis model

[0115] Generally, in order to establish an accurate model, a physical model can be selected because the physical model is detailed and accurate. However, in practical applications, the diversity of the system makes it difficult to establish a physical model. Based on the characteristics of the solar collector field, the present invention selects a semi-physical model and selects characteristic parameters with typical significance through the analysis of the solar energy system. Based on the polynomial regression model, the model is established using the operating data under normal operating conditions of the system. Irradiance and flow rate have a great influence on the operation of the solar collector system, so the present invention selects these two parameters as the input variables of the model, and the expression is:

[0116] Y = f(G, Q)+ε = a0 + a1G + a2Q + a3G 2 + a4Q 2 + a5GQ

[0117] where ε is the error term; Y is the regression variable output model; a0 a5 are the correlation coefficients of the regression model; G is the irradiance, the input parameter of the regression model, W / m 2 ; Q is the flow rate, the input parameter of the regression model, t / h. The model expression of the characteristic parameters represents a function of the flow rate and the irradiance.

[0118] Generally, the higher the degree of the polynomial, the more accurate the simulation result. According to the variable characteristics and relevant experience, a binary polynomial model is selected; the least squares method is used to solve the correlation coefficients of the expression. The adopted method is to regard the regression matrix used for modeling as a row vector X. Using matrix notation, the polynomial model expression can be written as:

[0119] Y = aX + e

[0120] Where X is the regression input variable vector; Y is the dependent output variable vector; a is the regression model coefficient vector; e is the error term vector; measurement errors and model errors are included in X and Y.

[0121] When the regression variables of the model are linearly independent, the residual of the model tends to be stable; when the error term follows a normal distribution, the least squares method is the best prediction method. At this time, the sum of squares of the regression residuals reaches the minimum, and the formula for solving the regression coefficients is:

[0122]

[0123] Where a OLS is the regression model coefficient vector; X reg is the regression matrix of the modeling data; is the transpose matrix of X reg ; Y reg is the dependent variable regression vector; a OLS is the coefficient vector solved by the model using the least squares method. After calculation using the above formula, the specific values of each coefficient are obtained. Furthermore, the mean square error of the residuals and the regression determination coefficient of the model are calculated, and the two are important indicators for judging the regression accuracy of the model.

[0124] The input quantities selected in the present invention are flow rate and irradiance. According to the characteristics of the solar energy system, the collector temperature T i is used as a characteristic parameter. Due to errors, the model prediction value may not be equal to the actual measurement value or the actual calculated value. Generally, the difference between the two is defined as the residual. Under normal conditions, the residual is within a certain range, and this range is called the residual threshold. It is generally determined by statistical methods using historical data, and the threshold of the characteristic parameter residual is obtained under a certain confidence level. After determining the characteristic parameter residual and the corresponding threshold, analyzing the residual of the characteristic parameter can realize the monitoring and identification of faults.

[0125] 2) Fault severity model feature library for three typical faults of solar collector fields based on deep convolutional neural network

[0126] As Figure 7 shown, establishing a fault severity model feature library for three typical faults of overheating, leakage, and aging of solar collector fields based on deep convolutional neural network includes the following steps:

[0127] Data preparation: Review and verify the temperature data in the database, supplement and correct the missing data and inconsistent data to ensure the correctness of the data and the logical consistency.

[0128] ① Handling of missing data: Due to network transmission failures, missing values may appear in the database. For the missing data, an estimation method can be used to replace the missing values with the sample mean for supplementation;

[0129] ② Handling of inconsistent data: With the help of the integrity constraint mechanism of the database management system, check the inconsistent data and then correct it with reference to the corresponding data values in the database.

[0130] Generating the dataset: Divide the prepared data into training set labels and test set labels.

[0131] Generating training set data and labels: According to different operating conditions of the solar energy system, read the temperature data values corresponding to the operating conditions from the database to generate training set data and operating condition labels under various operating conditions; in specific applications, the meanings of each operating condition label are shown in Table 1:

[0132] Table 1 Meanings of labels for the solar collector field fault degree model based on deep convolutional neural network

[0133]

[0134] The program automatically generates operating condition labels according to different operating conditions;

[0135] Generating test set data and labels: According to different operating conditions of the solar energy system, read the temperature data values corresponding to the operating conditions from the database to generate test set data and operating condition labels under various operating conditions; among them, the operating condition labels are the same as those of the training set, and are automatically generated by the program according to the operating conditions.

[0136] Network training: As Figure 8 shown, the algorithm flow chart of the convolutional neural network training process. Input the training set data into the convolutional neural network, continuously perform convolution and pooling to obtain the feature vector, and send it to the fully connected network. By calculating the output of the network and the training set labels, obtain the network error, and use the error backpropagation algorithm to continuously correct the network weights, biases, convolution coefficients, and pooling coefficients to make the error meet the set accuracy requirements. Then the network training is completed, and a convolutional neural network model is generated, that is, the fault degree model of typical faults in the solar collector field based on deep convolutional neural network.

[0137] The specific steps are as follows:

[0138] ① Read in a set of training set data d, whose size is [M×1×N], where M represents the size of the training batch, and 1×N represents the one-dimensional training data;

[0139] ②Perform the first convolution operation on the input training data to obtain the feature map t. Initialize the coefficients of the convolution kernel g. Let the size of g be [P×1×Q], where P represents the number of convolution kernels, and [1×Q] represents the size of the convolution kernel. The convolution result is t = ∑(d*g), and the size of the feature map is [M×1×N×Q];

[0140] ③Perform max pooling operation on the feature map t obtained from the convolution operation to obtain the feature map Z. Initialize the pooling coefficients, given the pooling stride p and the pooling window size k. Finally, the size of the obtained feature map Z is [M×1×(N / p)×Q]. The pooling process reduces the dimension of the data;

[0141] ④Repeat the above steps ②-③, perform convolution and pooling operations repeatedly to obtain the feature vector x. At this time, the feature extraction process of the convolutional neural network is completed;

[0142] ⑤Initialize the weight matrix w and bias b of the fully connected network. Feed the extracted feature vector x into the fully connected network and perform operations with the weight matrix w and bias b to obtain the network output y = ∑(w×x + b);

[0143] ⑥Subtract the output y obtained by the network from the training set label l to obtain the network error e = y - l. Take the derivative of the network error and use this derivative for backpropagation to sequentially correct the weights w, biases b, pooling coefficients of each layer, and convolution coefficients of each layer;

[0144] ⑦Repeat the above process until the network error e meets the accuracy requirements. The network training process is completed, and a convolutional neural network model is generated.

[0145] Network testing: Input the test set data into the already trained network model and output the fault information result. The specific steps are as follows:

[0146] ①Load the already trained convolutional neural network model. At this time, the convolution kernel coefficients, pooling coefficients, network weights w, and biases b of the convolutional neural network have all been trained;

[0147] ②Input the test data set into the already trained convolutional neural network to output the operating condition label, thereby outputting the fault detection result. According to the output label, the fault type and its severity can be judged. For example, 1 represents mild overheating, 3 represents severe overheating, etc.

[0148] Solar collector field infrared fault detection system:

[0149] Such as Figure 9As shown in the figure, the solar energy collection field is monitored by an infrared thermal imaging camera, the communication between the infrared thermal imaging camera and the computer is established, and the original infrared image of the entire solar energy collection field is obtained; preprocessing is carried out to obtain the converted infrared thermal image; data collection, analysis and processing, and regional division are carried out on the converted infrared thermal image, the characteristic information of the temperature field distribution in each region is extracted, and it is compared with the characteristic information of the two typical fault model feature libraries of the solar energy collection field in turn to determine whether there is a fault. If a fault occurs, the fault information of the collector is further determined.

[0150] The fault detection system for the solar energy collection field includes four parts: temperature image acquisition, temperature image reading, data acquisition and analysis processing, and fault model matching:

[0151] Step 1: Temperature image acquisition. For the solar energy collection field, it is monitored in real time by an infrared thermal imaging camera to obtain the overall temperature distribution thermal map of the collection field;

[0152] Step 2: Establish the communication between the infrared thermal imaging camera and the computer, and read the temperature image of the temperature state of the collection field;

[0153] Step 3: Conduct preliminary preprocessing to remove the influence of the thermal environment factors to obtain an infrared thermal image Figure 1 , and conduct dynamic range compression and image enhancement processing to obtain the converted infrared thermal image Figure 2 . For the converted infrared thermal image Figure 2 Use infrared software to conduct data collection and further divide the region for analysis and processing to obtain the characteristic information of the temperature field distribution of the fault of each collector;

[0154] Step 4: Compare the characteristic information of the temperature field distribution in each region extracted with the characteristic information in the characteristic library of the typical fault judgment model of the collector, analyze and determine whether the collector has a fault and determine the fault type, further import it into the characteristic library of the fault degree model, determine the fault-related information and take corresponding measures.

[0155] Note: 1. ΔT n is the difference range between the actually collected fault temperature data set and the normal operation temperature field data set;

[0156] 2. Assume that Δt1 > Δt2 > Δt3 > Δt4, that is, when the fault is in the aging state, the temperature residual of the collector has a deviation of more than Δt1; when the fault is in the leakage state, the temperature residual of the collector has a deviation of more than Δt2; when the fault is in the overheating state, the temperature residual of the collector has a deviation of more than Δt3, and when the fault is in the failure state, the temperature residual of the collector has a deviation of more than Δt4.

[0157] Application case

[0158] In practical applications, the entire detection process is described by taking the infrared imaging diagram of a partial solar collector field as an example. The standard thermal environment parameters are set according to GB / T 4271-2007 "Test Methods for Thermal Performance of Solar Collectors": irradiance 700 W / m 2 , ambient temperature 30 °C, flow rate 0.02 m 2 / s, etc.

[0159] Step 1: Use an infrared thermal imaging camera to take an infrared imaging diagram of the solar collector field (the actual thermal environment parameters during shooting are: irradiance 750 W / m 2 , ambient temperature 34 °C, flow rate 0.02 m 2 / s, etc.) and transmit it to the computer, as shown in Figure a below.

[0160] Step 2: Preprocess the original infrared imaging diagram: ① Establish a two-dimensional heat transfer model of the solar collector, and use the fsolve function in the matlab platform to solve the temperature of the solar collector field under standard thermal environment factors and actual thermal environment factors respectively, and obtain the temperature T i to be compensated under this environment = 0.327 °C. Therefore, the infrared thermal imaging after temperature compensation is Figure 10 shown in Figure b. ② Perform further processing of improved guided filtering and adaptive dynamic range compression of the sigmoid function, and obtain the converted thermal infrared imaging through processing on the matlab platform Figure 2 , as shown in Figure c below.

[0161] Step 3: Use infrared software to collect data from the preprocessed infrared thermal imaging 2 and further analyze it by region (as shown in Figure d), and obtain the characteristic information of the fault temperature field distribution in each region, such as Figure 11 shown. Figure a is a scatter plot of the temperature field distribution of the collector in area ① of the collector, and Figure b is a deviation map of the temperature residuals of the temperature field of the solar collector in this area.

[0162] Step 4: The typical fault model feature library of the solar collector field includes two parts: four typical fault judgment models of failure, overheating, leakage, and aging of the solar collector field based on the regression analysis model, and the fault degree models of three typical faults of overheating, leakage, and aging of the solar collector field based on the deep convolutional neural network.

[0163] After the preliminary collection and processing of the fault-free operation data of the solar energy system under the standard thermal environment, a regression model is established and evaluated using the coefficient of determination and standard error mentioned above. The coefficient of determination indicates the fitting degree of the regression model and also explains the proportion of the regression model in the error. Therefore, the closer it is to 1, the better the effect. This is a comprehensive measure of the fitting degree of the regression model. A better fit is achieved when the coefficient of determination is large and the standard error is small. The calculation results of the corresponding fitting of this model are as follows: the coefficient of determination of the regression model is 95.2%, and the standard error of the mean square deviation is 2.86 °C. The results show that the regression coefficient is close to 1, indicating a good correlation and accurate simulation results.

[0164] After processing, the formula Y = f(G, Q)+ε = a0 + a1G + a2Q + a3G 2 + a4Q 2 + the regression model coefficients of the six characteristic parameters in a5GQ are shown in Table 2:

[0165] Table 2 Characteristic Parameter Table of the Regression Model

[0166]

[0167] The predicted values of the model are compared with the collector temperature values under four typical fault operation conditions collected. Taking a 90% confidence level, it is obtained that the deviation of the collector temperature residual under the aging fault operation condition exceeds 7 °C, the deviation of the collector temperature residual under the leakage fault operation condition exceeds 10 °C, the deviation of the collector temperature residual under the overheating fault operation condition exceeds 12 °C, and the deviation of the collector temperature residual under the failure fault operation condition exceeds 15 °C, that is, Δt1 = 7 °C, Δt2 = 10 °C, Δt3 = 12 °C, Δt4 = 15 °C.

[0168] As Figure 12 shown, more than 90% of the temperature residuals ΔTn in this example working condition are greater than 10 °C but less than 12 °C. Therefore, it can be concluded that there is a leakage fault in the solar collector in this area, and the points within the red frame are the leakage fault points.

[0169] For the fault degree models of three typical faults, namely overheating, leakage, and aging, of the solar collector field based on the deep convolutional neural network, we define the structure of the one-dimensional deep convolutional neural network, which specifically includes two convolutional layers, two pooling layers, and one fully connected layer. The parameters of each layer are shown in Table 3:

[0170] Table 3 Training Structure Parameter Table of the One-Dimensional Deep Convolutional Neural Network

[0171]

[0172] Extract the temperature values from the scatter plot of the temperature field distribution of the solar collector in the leakage area, and input them into the solar collector field fault severity model based on a deep convolutional neural network. The output is label 4, indicating a minor leakage fault.

[0173] In summary, a fault has occurred in this part of the solar collectors, and the fault type, fault location, and fault severity have been confirmed.

Claims

1. A rapid detection method for the failure risk of a solar heat collection field, characterized in that, It includes the following steps; Step 1: Temperature image acquisition; Set up an infrared thermal imaging camera for real-time monitoring to obtain the overall temperature distribution heat map of the solar energy collection field; Step 2: Temperature image reading; Establish communication between the infrared thermal imaging camera and the computer. The infrared thermal imaging camera and the computer are in the same local area network. Download the PC software corresponding to the camera on the computer and match the IP to achieve real-time monitoring, and collect temperature images of the temperature state of the collection field; Step 3: Data acquisition, analysis and processing; Preprocess and regionally divide the overall temperature distribution heat map of the collection field, and perform feature extraction to obtain the scatter plot of the fault temperature field of the collectors in each region; Step 4: Fault model matching; Import the extracted characteristic information of the temperature field distribution in each region and related characteristic parameters into the fault model feature library of the solar energy collector. Perform two matching analyses of the fault model feature library in sequence to improve the fault detection efficiency of the collection field, reduce the false alarm rate and missed alarm rate, and ensure the safe operation of the solar energy collection system; The fault model feature library of the solar energy collector in Step 4 includes four typical fault judgment model feature libraries of the solar energy collection field based on the regression analysis model and three fault degree model feature libraries of three typical faults of the solar energy collection field based on the deep convolutional neural network; For the four typical fault judgment model feature libraries of the solar energy collection field based on the regression analysis model, based on the characteristics of the solar energy collection field, a semi-physical model is selected. By analyzing the solar energy system, characteristic parameters with typical significance are selected. Based on the polynomial regression model, a model is established using the operation data under normal operating conditions of the system. The irradiation amount and flow rate of the solar energy collection system are selected as the input variables of the model. The expression is: Y = f(G, Q) + ε = a0 + a1G + a2Q + a3G 2 + a4Q 2 + a5GQ Where Y is the regression variable output model; ε is the error term; a0 - a5 are the regression model correlation coefficients; G is the irradiance, the input parameter of the regression model, W / m 2 ; Q is the flow rate of the input parameter of the regression model, t / h; After regression analysis and processing, compare the predicted value of the model with the collector temperature values under four typical fault operating conditions collected. Take a 90% confidence level to determine the characteristic parameter residuals and corresponding thresholds, and analyze the residuals of the characteristic parameters to achieve fault monitoring and identification; For the three fault degree model feature libraries of three typical faults of the solar energy collection field based on the deep convolutional neural network, data features are extracted through one-dimensional deep convolutional neural network training, and the fault degree at the time of each fault is determined; The three fault degree model feature libraries of three typical faults of the solar energy collection field based on the deep convolutional neural network: It includes the following steps: Data preparation: Review and verify the temperature data in the database, supplement and correct the missing data and inconsistent data to ensure the correctness of the data and the logical consistency; ① Processing of missing data: Due to network transmission failures, missing values may appear in the database. For the missing data, an estimation method is used to supplement the missing values with the sample mean; ② Processing of inconsistent data: With the help of the integrity constraint mechanism of the database management system, check the inconsistent data, and then correct it with reference to the corresponding data values in the database; Generate data sets: Divide the prepared data into training set labels and test set labels; Generate test set data and labels: According to different operating conditions of the solar energy system, read the temperature data values corresponding to the operating conditions from the database, and generate test set data and operating condition labels under various operating conditions; among them, the operating condition labels are the same as those of the training set and are automatically generated by the program according to the operating conditions. Network training: Input the training set data into the convolutional neural network, continuously perform convolution and pooling to obtain feature vectors, send them into the fully connected network, calculate the output of the network and the training set labels to obtain the network error, and use the error backpropagation algorithm to continuously correct the network weights, biases, convolution coefficients, and pooling coefficients to make the error meet the set accuracy requirements. Then the network training is completed, and a convolutional neural network model, that is, a fault degree model for typical faults of the solar collector field based on the deep convolutional neural network, is generated.

2. The rapid detection method for the failure risk of a solar heat collection field according to claim 1, wherein, The preprocessing part of the third step includes the following steps: Adopt an infrared image normalization method based on temperature difference to preliminarily preprocess the original infrared thermal image, and eliminate the influence of the thermal environment factors on the temperature field of the solar collector; regard the images of the solar collector under different environmental factors as basic signals and process them in the spatial domain to achieve the purpose of eliminating the influence of the external thermal environment. Then, perform further processing of adaptive dynamic range compression based on the improved guided filter and Sigmoid function to perform dynamic range compression and image enhancement processing on the image, and obtain a more natural result that conforms to human visual characteristics and has obvious characteristics.

3. A rapid detection method for the failure risk of a solar heat collection field according to claim 2, characterized in that The guided filter decomposes the input infrared image into a basic layer B and a detail layer D. The basic layer B is the processing result of the improved guided filter on the input image, which contains the high-contrast information and most of the dynamic range in the input image; the detail layer D is the difference between the input image and the basic layer B, which contains the low-contrast details in the input image. B = G[I h ​ D = I h - B where I h is the input infrared thermal image; the symbol G[*] represents the improved guided filter, and the edge-aware weight is constructed using a variant of the first-order differential of the image; Then, use the Sigmoid adaptive compression algorithm to compress the dynamic range of the basic layer B. On the premise of retaining the overall visual perception of the image, compress the gray information of the high-dynamic range image so that it can be normally displayed on traditional display devices. B' = S[B] where the symbol S[*] represents the Sigmoid algorithm; B' is the basic layer after dynamic range compression, and the value of B' is between [0, 1]. Strengthen the detail layer with a fixed gain coefficient η, and the enhanced detail layer D' is D' = η * D It is recommended that the setting of the fixed gain coefficient be between [1, 2]. Finally, allocate a suitable dynamic range for the basic layer B' after dynamic range compression, and recombine it with the enhanced detail layer D' to obtain the output image. I l = a * B' + D' Among which I l is the output image, and the parameter a is used to control the size of the dynamic range of the base layer.

4. A rapid detection method for the failure risk of a solar heat collection field according to claim 1, characterized in that, The feature library of the four-class typical fault judgment model for the solar collector field based on the regression analysis model: Select a semi-physical model, select characteristic parameters with typical significance through the analysis of the solar energy system, and establish a model based on the polynomial regression model using the operation data under the normal operating conditions of the system. Irradiance and flow have a great impact on the operation of the solar collector system. Select these two parameters as the input variables of the model, and the expression is: Y = f(G, Q) + ε = a0 + a1G + a2Q + a3G 2 + a4Q 2 + a5GQ where ε is the error term; Y is the regression variable output model; a0 - a5 are the regression model correlation coefficients; G is the irradiation amount, the input parameter of the regression model, W / m 2 ; Q is the flow rate, the input parameter of the regression model, t / h, and the model expression of the characteristic parameter represents a function of the flow rate and the irradiation amount; According to the variable characteristics and relevant experience, a binary polynomial model is selected; the least squares method is used to solve the correlation coefficient of the expression. The way is to regard the regression matrix used for modeling as the row vector X. Using matrix notation, the polynomial model expression is written as: Y = aX + e where X is the regression input variable vector; Y is the dependent output variable vector; a is the regression model coefficient vector; e is the error term vector; measurement errors and model errors are included in X and Y; When the regression variables of the model are linearly independent, the residual side of the model tends to be stable; when the error term follows a normal distribution, the least squares method is the best prediction method. At this time, the sum of squares of the regression residuals reaches the minimum. The formula used to solve the regression coefficients is: where a OLS is the regression model coefficient vector; X reg is the regression matrix of the modeling data; is the transpose matrix of X reg ; Y reg is the regression vector of the dependent variable; a OLS is the coefficient vector obtained by solving the model using the least squares method. After calculating using the above formula, the specific values of each coefficient are obtained, and then the mean square error of the residual and the regression determination coefficient of the model are calculated. These two are important indicators for judging the regression accuracy of the model.

5. A rapid detection method for the failure risk of a solar heat collection field according to claim 1, characterized in that The fault degree model of typical faults in the solar collector field based on the deep convolutional neural network is as follows: ① Read in a set of training set data d, whose size is [M×1×N], where M represents the size of the training batch, and 1×N represents one-dimensional training data; ② Perform the first convolution operation on the read training data to obtain the feature map t. Initialize the coefficients of the convolution kernel g. Let the size of g be [P×1×Q], where P represents the number of convolution kernels, and [1×Q] represents the size of the convolution kernel. The obtained convolution result is t = ∑(d * g), and the size of the feature map is [M×1×N×Q]; ③ Perform the max pooling operation on the feature map t obtained by the convolution operation to obtain the feature map Z. Initialize the pooling coefficients, given the pooling stride p and the pooling window size k. Finally, the size of the obtained feature map Z is [M×1×(N / p)×Q]. The pooling process reduces the dimension of the data; ④ Repeat the above steps ② - ③, repeatedly perform convolution and pooling operations to obtain the feature vector x. At this time, the feature extraction process of the convolutional neural network is completed; ⑤ Initialize the weight matrix w and bias b of the fully connected network. Send the extracted feature vector x into the fully connected network and perform operations with the weight matrix w and bias b to obtain the network output y = ∑(w × x + b); ⑥ Subtract the output y obtained by the network from the training set label l to obtain the network error e = y - l. Take the derivative of the network error and use this derivative for backpropagation to sequentially correct the weights w, biases b, pooling coefficients of each layer, and convolution coefficients of each layer; ⑦ Repeat the above process until the network error e meets the accuracy requirements, the network training process is completed, and a convolutional neural network model is generated.

Images