Photovoltaic module shielding object identification system and method
Through the multi-resolution feature self-selecting algorithm of convolutional neural network, the problem of low accuracy of occlusion recognition of photovoltaic components is solved, efficient occlusion recognition is achieved, and recognition accuracy and detection rate are improved.
Patent Information
- Application Number
- CN202411880456.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-19
- Publication Date
- 2025-04-25
AI Technical Summary
In the prior art, the accuracy of the occlusion identification of photovoltaic modules is not high and the detection rate is low, which cannot meet the operation and maintenance needs of photovoltaic power stations.
A multi-resolution feature self-selecting algorithm based on convolutional neural network is used to generate feature maps of different scales and resolutions through the image feature extraction module, the multi-resolution feature extraction module and the feature self-selecting module. The Sigmoid function is used to calculate the occlusion score to improve the recognition accuracy.
It significantly improves the recognition accuracy and detection rate of occlusions of photovoltaic modules, enhances the sensitivity to occlusion background information, and improves the comprehensiveness of occlusion recognition.
Smart Images
Figure CN119888439A_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the technical fields of artificial intelligence, foreign object recognition, and photovoltaic string maintenance, and specifically relates to a photovoltaic module obstruction recognition system and method. Background Art
[0002] The energy of photovoltaic power generation comes directly from the sunlight, and the sunlight on the earth's surface is greatly affected by the climate. Long-term rainy, snowy, cloudy and foggy days will seriously interfere with the power generation state of the system. Another prominent point is that particulate matter (such as dust) or other fallen objects (such as bird droppings and leaves) in the environment sink on the surface of the solar cell module, and in some distributed photovoltaic systems, there may be trees and building shadows that block part of the light. These abnormal shielding will reduce the conversion efficiency of the battery module, resulting in reduced power generation and even damage to the battery panel. According to research, shielding such as bird droppings and shadows has become one of the most important problems affecting photovoltaic performance. Shielding will reduce the power generated by the photovoltaic array and cause safety hazards. When a part of a photovoltaic panel is blocked, its blocked unit will become reverse biased and exist in the circuit as a load that consumes energy. When the dust accumulation density is high, its short-circuit current, open-circuit voltage and power generation efficiency are greatly reduced. If the shielding is allowed to exist at this time, problems such as hot spots will occur at the least, reducing the life of the photovoltaic module, and even causing fires, endangering personal and property safety. Therefore, the identification of abnormal obstructions to photovoltaic modules is one of the urgent problems to be solved in the operation and maintenance of photovoltaic power stations.
[0003] The existing technology uses drones to inspect photovoltaic modules and take photos, and then analyzes the photos to identify obstructions on photovoltaic strings. However, in the process of collecting images of photovoltaic modules based on drone inspections, different obstructions may appear in different forms and sizes in the image due to factors such as shooting distance and angle. In addition, due to the influence of the natural environment, the distribution of obstructions on photovoltaic modules varies in sparseness. Obstructions may be clustered or scattered. Traditional recognition algorithms have low accuracy and detection rates, and cannot meet the requirements of photovoltaic station operation and maintenance. Summary of the invention
[0004] The purpose of this application is to overcome the defects of the prior art in that the recognition accuracy of obstructions to photovoltaic strings is low and the detection rate is low.
[0005] In order to achieve the above objectives, the present application proposes a photovoltaic module obstruction recognition system, the system comprising:
[0006] The image feature extraction module includes an m×m convolution unit and n residual units, where n is an even number, and is used to generate k, k=n / 2 feature maps of different scales;
[0007] A multi-resolution feature extraction module, comprising k 1×1 convolution units, k 3×3 convolution units and k-1 bilinear interpolation units, for processing the k feature maps of different scales generated by the image feature extraction module into k feature maps of different resolutions after feature fusion;
[0008] The feature self-selection module includes k global average pooling units and k fully connected layer units, which are used to process the feature maps of different resolutions after the k features generated by the multi-resolution feature extraction module are fused into k activated feature maps, and the feature map with the highest score is selected and calculated using the Sigmoid function as the photovoltaic module occlusion score; the sum of the scores is the sum of the probabilities of determining each occlusion category, and the category with the highest score is determined as the final occlusion category.
[0009] As an improvement of the above system, the image feature extraction module processes the image including:
[0010] First, the input image is convolved using an m×m convolution unit, and then processed by two residual units in sequence; after each two residual units, a feature map is generated as output; after each two residual units, the resolution of the generated feature map will be halved and the number of channels will be doubled.
[0011] As an improvement of the above system, the process of processing the image by the multi-resolution feature extraction module includes:
[0012] The lowest resolution feature map C obtained by the image feature extraction module k The feature map M is obtained by processing with a 1×1 convolution unit. k , for the feature map M k The feature map P is obtained by processing with 3×3 convolution units. k ;
[0013] For feature map M k It is upsampled to twice its original scale through a bilinear interpolation unit and then combined with the feature map C processed by a 1×1 convolution unit. k-1 Add each element one by one to get the feature map M k-1 , for M k-1 Use 3×3 convolutional units to get the feature map P k-1 ;
[0014] For feature map M k-1 Carry out and feature map M k After similar processing, the feature map P is obtained k-2 ;
[0015] The feature map P is obtained by the same process. 1 ;
[0016] The obtained feature map P 1 -P k That is the output of the multi-resolution feature extraction module.
[0017] As an improvement of the above system, the process of processing the image by the feature self-selection module includes:
[0018] For each feature map output by the multi-resolution feature extraction, a global average pooling unit and a fully connected layer unit are sequentially processed to obtain an activated feature map;
[0019] The activated feature map with the highest score is selected and calculated using the Sigmoid function as the PV panel obstruction score.
[0020] As an improvement to the above system, m is 7.
[0021] As an improvement to the above system, n is 8.
[0022] The present application also provides a photovoltaic module obstruction recognition method, which is implemented based on the above system and includes:
[0023] The photovoltaic string image is input into the trained system to obtain the photovoltaic module obstruction score.
[0024] Compared with the prior art, the advantages of this application are:
[0025] The photovoltaic module obstruction recognition system and method proposed in this application are based on a convolutional neural network multi-resolution feature self-selection algorithm, which introduces a feature pyramid into the obstruction recognition process to strengthen the semantic expression of obstructions at various scales and densities, and through the automatic selection of multi-resolution features, comprehensively improve the sensitivity to the background information of photovoltaic module obstructions, which can greatly improve the recognition accuracy and detection rate of photovoltaic string obstructions. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 The figure shows the photovoltaic module obstruction recognition architecture diagram;
[0027] Figure 2 Shown is the structural diagram of the image feature extraction module;
[0028] Figure 3 Shown is a structural diagram of a multi-resolution feature extraction module;
[0029] Figure 4 Shown is the structure diagram of the feature self-selection module. DETAILED DESCRIPTION
[0030] The technical solution of the present application is described in detail below with reference to the accompanying drawings.
[0031] In the problem of photovoltaic module obstruction recognition, the density and scale of obstructions will vary greatly due to natural factors, image acquisition and other factors. In addition, an object only needs to be recognized when it can obstruct the photovoltaic module. Therefore, recognition is not only related to the object category, but also to the object background information. The single-resolution features commonly used in traditional recognition algorithms are not sensitive enough to small-scale and low-density objects, and easily dilute the background information of the object, so they are not suitable for this scenario. The photovoltaic module obstruction recognition system and method proposed in this application are based on a convolutional neural network multi-resolution feature self-selection algorithm, and introduce feature pyramids into the obstruction recognition process to enhance the semantic expression of obstructions at various scales and densities; and through the automatic selection of multi-resolution features, the sensitivity to the background information of photovoltaic module obstructions is comprehensively improved.
[0032] like Figure 1 As shown, the photovoltaic module obstruction recognition system includes image feature extraction, multi-resolution feature extraction and feature self-selection modules, wherein:
[0033] 1. Image feature extraction module
[0034] like Figure 2 As shown in the figure, the image feature extraction module includes a 7×7 convolution unit and 8 residual units. In the image feature extraction module, the global features of the entire photovoltaic module image are extracted at multiple scales based on the residual network. The input image has a height of H and a width of W. After passing through the image feature extraction module, C images with different resolutions and different numbers of channels can be obtained. 1 , C 2 , C 3 and C 4 There are 4 feature maps of different scales in total. After every 2 residual units, the resolution of the generated feature map will be halved and the number of channels will be doubled. In this embodiment, 7×7 convolution units and 8 residual units are used. In other embodiments, convolution units of other sizes and other numbers of residual units may also be used. Preferably, the 7×7 convolution unit retains more image characteristics and is more suitable for image feature extraction of photovoltaic modules.
[0035] 2. Multi-resolution feature extraction module
[0036] like Figure 3 As shown in Figure 1, the multi-resolution feature extraction module includes 4 1×1 convolution units, 4 3×3 convolution units and 3 bilinear interpolation units, and finally generates feature maps of different resolutions after feature fusion: P 1 , P 2 , P 3 , P 4 .
[0037] In the multi-resolution feature extraction module, from C 4 Starting from the beginning, the semantic information of deep features is gradually integrated into the shallow layer. Since the number of channels between features of different resolutions is different, it is necessary to first unify each feature map to the same number of channels using a 1×1 convolution unit. After obtaining feature maps M with the same number of channels, 4 The fusion method is as follows: the feature map C 4 Perform 1×1 convolution to obtain feature map M 4 , for the feature map M 4 Use 3×3 convolution to get the feature map P 4 . For the feature map M 4 Use bilinear interpolation to upsample it to twice its original scale, and then combine it with the feature map C processed by 1×1 convolution. 3 Add each element one by one to get the feature map M 3 , for M 3 Use 3×3 convolution to get the feature map P 3 . For the feature map M 3 Carry out and feature map M 4 Similar processing obtains the feature map P 2 . The feature map P is obtained by the same process. 1 In the above operation, the 3×3 convolution processing is performed on the feature map M to eliminate the aliasing effect caused by upsampling and improve the receptive field.
[0038] In other embodiments, the number of 1×1 convolution units, 3×3 convolution units, and bilinear interpolation units may also be other numbers greater than 1, as long as the number of 1×1 convolution units and 3×3 convolution units is the same as half of the number of residual units of the image feature extraction module, and the number of bilinear interpolation units is 1 less than the number of 3×3 convolution units.
[0039] 3. Feature selection module
[0040] In general recognition tasks, the common method of feature fusion is direct fusion. However, if the scales of the features are different, fusion cannot be completed by direct addition. A simple feature fusion solution is to use the characteristic of the consistent number of channels between features, first use global average pooling to convert different features into vectors of the same dimension as the number of channels, and then directly add the feature vectors corresponding to each resolution element by element to obtain the fused vector. The above operation can be expressed as:
[0041]
[0042] In the feature self-selection module of this application, there is no choice of fusion before recognition, but prediction is performed at each scale at the same time, such as Figure 4As shown. After global average pooling, each feature map is directly input into the fully connected layer to obtain the activation result, and then the feature map that can obtain the maximum activation is selected as the final feature map. The above operation can be expressed as:
[0043] P = argm P a i xFC(GAP(Pi))
[0044] The feature self-selection module includes 4 global average pooling units (GAP), 4 fully connected layer units (FC), and finally generates 4 calculated feature maps a1, a2, a3, and a4. Finally, one of the feature maps with the largest value (a1, a2, a3, and a4) is selected as the result and output.
[0045] The above operation can also be regarded as the result of setting the weight of the selected feature map to 1, the weight of the remaining resolution features to 0, and then fusing them according to the weight.
[0046] In other embodiments, the number of global average pooling units and fully connected layer units may also be other numbers greater than 1, which may be the same as half of the number of residual units of the image feature extraction module.
[0047] The result is:
[0048] After completing feature selection, directly use the fully connected layer that calculates the maximum activation to get the activation value, and then use the Sigmoid function to get the final score s. The above process can be expressed as:
[0049] s=Sigmoid(FC(P))
[0050] The sum of the scores is the sum of the probabilities of determining each occlusion category, and the category with the highest score is determined as the final occlusion category.
[0051] During the training process, the final score s of the output is trained using cross entropy. Assume that the label of the training sample is y, y = 1 represents a positive sample, and y = 0 represents a negative sample, then the loss L in the training phase is:
[0052] L=y×lns+(1-y)ln(1-s)
[0053] The smaller the loss value, the better the recognition effect.
[0054] The process of training and testing the photovoltaic module obstruction recognition system includes:
[0055] Input: Photovoltaic module image I test The labeled photovoltaic module image training set I is composed of
[0056] Output: Photovoltaic module image I testThe final occluder recognition score s.
[0057] Training process:
[0058] 1) Initialize network parameters θF.
[0059] 2) Loop and perform the following operations:
[0060] 3) Randomly select a photovoltaic module image from the image training set I.
[0061] 4) Extract multi-scale global features C through the image feature extraction module 1 , C 2 , C 3 and C 4 .
[0062] 5) Extract multi-resolution features P through the multi-resolution feature extraction module 1 , P 2 , P 3 and P 4 .
[0063] 6) For multi-resolution feature P 1 , P 2 , P 3 and P 4 Perform global average pooling and input it into the fully connected layer to obtain the corresponding activation a 1 、a 2 、a 3 and a 4 .
[0064] 7) Use Sigmoid(max{a 1 , a 2 , a 3 , a 4}) Calculate the recognition score.
[0065] 8) Optimize the global network parameters based on the loss function L and the input image label.
[0066] 9) The loss function converges.
[0067] Testing process:
[0068] 1) Will I test Input into the trained system and obtain the maximum activation value a in the same way as training the network for a single iteration.
[0069] 2) Obtain the final PV panel obstruction score s through the Sigmoid function.
[0070] The present application may also provide a computer device, comprising: at least one processor, a memory, at least one network interface and a user interface. The various components in the device are coupled together through a bus system. It is understood that the bus system is used to achieve connection and communication between these components. In addition to the data bus, the bus system also includes a power bus, a control bus and a status signal bus.
[0071] The user interface may include a display, a keyboard or a pointing device, such as a mouse, a trackball, a touch pad or a touch screen.
[0072] It is understood that the memory in the embodiments disclosed in the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). The memories described herein are intended to include, but are not limited to, these and any other suitable types of memories.
[0073] In some embodiments, the memory stores the following elements, executable modules or data structures, or a subset thereof, or an extended set thereof: an operating system and applications.
[0074] The operating system includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., which are used to implement various basic services and process hardware-based tasks. The application includes various application programs, such as a media player (Media Player), a browser (Browser), etc., which are used to implement various application services. The program for implementing the method of the embodiment of the present disclosure can be included in the application.
[0075] In the above embodiment, the processor may also call a program or instruction stored in the memory, specifically, a program or instruction stored in an application program, and is used to:
[0076] Execute the steps of the above method.
[0077] The above method can be applied to a processor or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in the processor or an instruction in the form of software. The above processor may be a general processor, a digital signal processor (Digital Signal Processor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The above-disclosed methods, steps and logic block diagrams can be implemented or executed. The general processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the above-disclosed method can be directly embodied as a hardware decoding processor to execute, or the hardware and software modules in the decoding processor are combined to execute. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.
[0078] It is understood that the embodiments described in the present application can be implemented by hardware, software, firmware, middleware, microcode or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in the present application or a combination thereof.
[0079] For software implementation, the technology of the present application can be implemented by executing the functional modules (such as procedures, functions, etc.) of the present application. The software code can be stored in a memory and executed by a processor. The memory can be implemented in the processor or outside the processor.
[0080] The present application may also provide a non-volatile storage medium for storing a computer program. When the computer program is executed by a processor, each step in the above method embodiment can be implemented.
[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present application and are not intended to limit it. Although the present application is described in detail with reference to the embodiments, a person skilled in the art should understand that any modification or equivalent replacement of the technical solution of the present application does not depart from the spirit and scope of the technical solution of the present application and should be included in the scope of the claims of the present application.
Claims
1. A photovoltaic module obstruction recognition system, characterized in that: The system comprises: The image feature extraction module includes an m×m convolution unit and n residual units, where n is an even number, and is used to generate k, k=n / 2 feature maps of different scales; a multi-resolution feature extraction module, comprising k 1×1 convolution units, k 3×3 convolution units and k-1 bilinear interpolation units, for processing the k feature maps of different scales generated by the image feature extraction module into k feature maps of different resolutions after feature fusion; and The feature self-selection module includes k global average pooling units and k fully connected layer units, which are used to process the feature maps of different resolutions after the k features generated by the multi-resolution feature extraction module are fused into k activated feature maps, and the feature map with the highest score is selected and calculated using the Sigmoid function as the photovoltaic module occlusion score; the sum of the scores is the sum of the probabilities of determining each occlusion category, and the category with the highest score is determined as the final occlusion category.
2. The photovoltaic module obstruction recognition system according to claim 1, characterized in that: The process of processing the image by the image feature extraction module includes: First, the input image is convolved using an m×m convolution unit, and then processed by two residual units in sequence; after each two residual units, a feature map is generated as output; after each two residual units, the resolution of the generated feature map will be halved and the number of channels will be doubled.
3. The photovoltaic module obstruction recognition system according to claim 1, characterized in that: The process of processing images by the multi-resolution feature extraction module includes: The lowest resolution feature map C obtained by the image feature extraction module k The feature map M is obtained by processing with a 1×1 convolution unit. k , for the feature map M k The feature map P is obtained by processing with 3×3 convolution units. k ; For feature map M k It is upsampled to twice its original scale through a bilinear interpolation unit and then combined with the feature map C processed by a 1×1 convolution unit. k-1 Add each element one by one to get the feature map M k-1 , for M k-1 Use 3×3 convolutional units to get the feature map P k-1 ; For feature map M k-1 Carry out and feature map M k After similar processing, the feature map P is obtained k-2 ; The feature map P1 is obtained by the same process; The obtained feature map P1-P k That is the output of the multi-resolution feature extraction module.
4. The photovoltaic module obstruction recognition system according to claim 1, characterized in that: The process of processing the image by the feature self-selection module includes: For each feature map output by the multi-resolution feature extraction, a global average pooling unit and a fully connected layer unit are sequentially processed to obtain an activated feature map; The activated feature map with the highest score is selected and calculated using the Sigmoid function as the PV panel obstruction score.
5. The photovoltaic module obstruction recognition system according to claim 1, characterized in that: The m is 7.
6. The photovoltaic module obstruction identification system according to claim 1, characterized in that: The n is 8.
7. A photovoltaic module obstruction identification method, implemented based on any system of claims 1-6, comprising: The photovoltaic string image is input into the trained system to obtain the photovoltaic module obstruction score.