Deep learning photovoltaic cell welding effect identification method and system

Abstract

The invention discloses a photovoltaic cell welding effect identification method and system based on deep learning, and the method comprises the steps: S1, collecting a surface image of a welded cell, and carrying out the denoising of the surface image, and obtaining a denoised image; s2, training an auto-encoder network based on the denoised image, and extracting depth features of the denoised image; s3, constructing a welding effect identification network, taking the denoised image and the corresponding depth features as input of the welding effect identification network, and obtaining the category probability of the welding effect; s4, constructing a weighted cross entropy loss function and optimizing parameters of the welding effect recognition network by using a gradient descent method based on momentum to obtain a trained welding effect recognition network; and S5, detecting the surface image of the welded battery piece to be recognized by using the trained welding effect recognition network to obtain a welding effect recognition result. According to the invention, the automation level of battery piece welding defect detection can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cell welding effect recognition, and in particular to a deep learning method and system for photovoltaic cell welding effect recognition. Background Art

[0002] With the growing demand for renewable energy, the photovoltaic industry has developed rapidly, and the quality control of photovoltaic modules has become the key to the development of the industry. In the production process of photovoltaic modules, the quality inspection of photovoltaic workpieces is particularly important to ensure their performance and reliability. The photovoltaic module system is complex, involving cells, tempered glass, backplane, aluminum alloy frame, sealing components, etc. Before the assembly is completed, the welding of the cell needs to be inspected accordingly. Since the quality of welding largely determines the quality of photovoltaic modules, it is necessary to ensure that the connection between the cell and other components is stable and reliable, reduce contact resistance, and improve the conversion efficiency and power generation performance of photovoltaic cells. Therefore, it is necessary to check whether the surface of the cell after welding is smooth and bright, without tin beads, burrs, false soldering, desoldering and over-soldering. The traditional photovoltaic workpiece inspection method mainly relies on manual visual inspection, which has the disadvantages of low efficiency, strong subjectivity, difficulty in unifying judgment standards, and easy fatigue. In addition, with the expansion of production scale and the increase in market demand, the requirements for detection accuracy and speed are also getting higher and higher.

[0003] In recent years, machine vision technology has been gradually applied to the field of automated inspection to improve the accuracy and efficiency of inspection. However, the existing automated inspection systems mostly use traditional image processing algorithms, such as edge detection and threshold segmentation. These methods are not robust enough in complex environments and are difficult to cope with the diversity and complexity of battery cell welding defects. At the same time, these algorithms have limited capabilities in feature extraction and defect recognition, and are difficult to generalize to new defect types. Summary of the invention

[0004] In view of this, the present invention provides a deep learning method for identifying photovoltaic cell welding effects, aiming to improve the level of automation of cell welding defect detection, enhance recognition accuracy and robustness through deep learning technology, and meet the needs of the modern photovoltaic industry for efficient and precise quality control.

[0005] To achieve the above purpose, the present invention provides a deep learning photovoltaic cell welding effect recognition method, comprising the following steps:

[0006] S1: Collect the surface image of the battery cell after welding and denoise it to obtain a denoised image;

[0007] S2: Train the autoencoder network based on the denoised image to extract the deep features of the denoised image;

[0008] S3: construct a welding effect recognition network, use the denoised image and the corresponding deep features as inputs of the welding effect recognition network, and obtain the category probability of the welding effect;

[0009] S4: construct a weighted cross entropy loss function and use the momentum-based gradient descent method to optimize the parameters of the welding effect recognition network to obtain a trained welding effect recognition network;

[0010] S5: Use the trained welding effect recognition network to detect the surface image of the battery cell after welding to obtain the welding effect recognition result.

[0011] As a further improvement method of the present invention:

[0012] Optionally, in step S1, collecting the surface image of the battery cell after welding and denoising it to obtain the denoised image includes:

[0013] S11: Grayscale conversion:

[0014] The collected surface image I of the battery cell after welding is grayed to obtain the grayed image I gray , specifically:

[0015] l gray (m,n)=0.299·I red (m,n)+0.587·I green (m,n)+0.114·I blue (m,n)

[0016] Where (m,n) is the pixel position; I gray (m,n) is the grayscale image I gray The pixel value at (m,n); I red (m,n),I green (m,n) and I blue (m, n) are the pixel values ​​of the red channel image, green channel image and blue channel image of the surface image I of the battery cell after welding at (m, n);

[0017] S12: Image denoising:

[0018] Use Gaussian filtering to remove the grayscale image I gray The noise in is:

[0019]

[0020] Among them, I smooth (m,n) is the denoised image I smoothThe pixel value at (m,n); G(i,j) is the Gaussian kernel function, where (i,j) is the Gaussian kernel index, i is the Gaussian kernel horizontal index, and j is the Gaussian kernel vertical index. σ is the standard deviation of the Gaussian kernel; k is the radius of the Gaussian kernel.

[0021] Optionally, in step S2, training an autoencoder network based on the denoised image to extract deep features of the denoised image includes:

[0022] S21: Build an autoencoder network:

[0023] The autoencoder consists of an encoder and a decoder, specifically:

[0024] Z=f encoder (I smooth )

[0025]

[0026] Among them, f encoder and f decoder are encoder function and decoder function respectively; Z is the denoised image I smooth Low-dimensional encoding of ; is the reconstructed image;

[0027] S22: Training the autoencoder network:

[0028] The training of the autoencoder network aims to minimize the reconstruction loss, which is:

[0029]

[0030] Among them, Loss rec is the loss value of reconstruction loss; ‖·‖ 2 Calculate the L2 norm;

[0031] Minimize the reconstruction loss Loss based on the stochastic gradient descent method rec The trained autoencoder network is obtained, and the stochastic gradient descent is specifically as follows:

[0032]

[0033] Among them, W rec is the weight of the autoencoder network; η is the learning rate of the autoencoder network; Loss is the loss value of the reconstruction loss rec About the weight W of the autoencoder network rec The gradient of

[0034] S23: Extract deep features:

[0035] The denoised image is input into the trained autoencoder network. The output of the encoder function in the autoencoder network is the deep features of the denoised image, specifically:

[0036]

[0037] in, The encoder function in the trained autoencoder network; For use The denoised image I is obtained smooth Low dimensional encoding.

[0038] Optionally, in step S3, a welding effect recognition network is constructed, and the denoised image and the corresponding depth features are used as inputs of the welding effect recognition network to obtain the category probability of the welding effect, including:

[0039] The welding effect recognition network consists of a sparse mask generation subnetwork, a feature fusion subnetwork, and a recognition subnetwork, specifically:

[0040] S31: Construct sparse mask generation subnetwork:

[0041] Sparse mask subnetwork to extract deep features of denoised images As input, the output sparse mask S is:

[0042]

[0043] Among them, f mask is the sparse mask generation function, the output sparse mask is the same as the denoised image I smooth They have the same size and each pixel value in the sparse mask is either 0 or 1;

[0044] S32: Construct feature fusion sub-network:

[0045] Use the generated sparse mask S to guide the denoised image I smooth Perform sparse convolution, specifically:

[0046] O sparse =f sparse_conv (I smooth |S)

[0047] Among them, O sparse is the result of sparse convolution; f sp arse_ conv is the sparse convolution function; f sp arse_ conv (I smooth |S) represents the denoised image I based on the sparse mask S smoothPerform convolution, that is, perform convolution operation only at the pixel position where the pixel value of the sparse mask S is 1;

[0048] Dense convolution is specifically:

[0049] O dense =f dense_conv (I smooth )

[0050] Among them, O dense is the result of dense convolution; f dense _ conv is the dense convolution function; f dense _ conv (I smooth ) represents the denoised image I smooth Perform dense convolution, that is, after denoising the image I smooth Convolution is performed on all pixel positions;

[0051] The results of sparse convolution and dense convolution are fused through weighted fusion operation, specifically:

[0052] O fusion =α·O sparse +(1-α)·O dense

[0053] Among them, O fusion is the weighted fusion result; α is the weighting parameter;

[0054] S33: Construct recognition subnetwork:

[0055] The weighted fusion result O fusion Input to the recognition subnetwork to obtain the category probability, specifically:

[0056] Y=f detection (O fusion )

[0057] Among them, Y is the probability of welding effect category, including the probability that the surface of the battery cell after welding belongs to normal welding and belongs to different categories of welding problems; f detection To identify the function, a convolutional neural network is used.

[0058] Optionally, in step S4, a weighted cross entropy loss function is constructed and a momentum-based gradient descent method is used to optimize the parameters of the welding effect recognition network to obtain a trained welding effect recognition network, including:

[0059] S41: Construct loss function:

[0060] The weighted cross entropy is used as the loss function of the welding effect recognition network. The weighted cross entropy is specifically:

[0061]

[0062] Among them, Loss det is the loss value of weighted cross entropy loss; c is the category number, c = 1, 2, ..., C, C is the number of categories contained in the welding effect; y true.c When the true category is the cth category, it is 1, otherwise it is 0. The true category is manually labeled; Y c is the probability of the welding effect category Y being the cth category; w c is the weight of category c;

[0063] S42: Building momentum-based gradient descent:

[0064] Initialize the momentum term V to 0 and update the momentum term in each iteration as follows:

[0065]

[0066] Among them, β is the momentum coefficient; Loss is the loss value of weighted cross entropy loss det About the weight W of welding effect recognition network det The gradient of

[0067] The updated momentum term V is used to update the weights of the welding effect recognition network, specifically:

[0068] W det ←W det -δ·V

[0069] Among them, δ is the learning rate of the welding effect recognition network;

[0070] Step S42 is iteratively executed until a predetermined number of iterations is reached, thereby obtaining a trained welding effect recognition network.

[0071] Optionally, in step S5, the trained welding effect recognition network is used to detect the surface image of the battery cell after welding to obtain the welding effect recognition result, including:

[0072] The surface image of the battery cell after welding to be identified is input into the trained welding effect recognition network, and the surface image of the battery cell after welding to be identified has no manual labeling of welding effect categories. The welding effect category probability obtained according to the output of the welding effect recognition network is Determine the type of welding effect, specifically:

[0073]

[0074] Among them, the argmax function obtains The dimension number with the largest median value is used to determine whether there is a welding problem on the surface of the battery cell after welding, and the specific type of the welding problem if there is a welding problem.

[0075] The present invention also discloses a bad workpiece identification system based on deep learning, comprising:

[0076] Denoising module: collects the surface image of the battery cell after welding and performs denoising;

[0077] Encoding module: train the autoencoder network based on the denoised image;

[0078] Network construction module: construct welding effect recognition network;

[0079] Network training module: construct a weighted cross entropy loss function and use the momentum-based gradient descent method to optimize the parameters of the welding effect identification network;

[0080] Network application module: Use the trained welding effect recognition network to detect the surface image of the battery cell after welding;

[0081] Beneficial effects:

[0082] The welding effect recognition method of the present invention realizes automated image preprocessing, feature extraction and defect recognition through deep learning technology, significantly improving the accuracy and efficiency of detection. The autoencoder network and optimized convolution structure are used to accurately capture the microscopic features of the battery cell surface, and the network parameters are optimized through the weighted cross entropy loss function and the momentum-based gradient descent method, so as to quickly and accurately identify the welding effect of the battery cell surface in a complex production environment.

[0083] By introducing sparse mask generation and feature fusion mechanisms, the welding effect recognition network of the present invention can enhance the robustness of the recognition of the welding effect on the surface of the battery cell while maintaining computational efficiency. The combination of Gaussian filtering and sparse convolution enables the system to stably output accurate recognition results when facing various welding problems on the surface of the battery cell, effectively reducing the situation of misjudgment and missed judgment, and improving the overall quality of the battery cell. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 The present invention is a flowchart of a deep learning method for identifying the welding effect of photovoltaic cells according to an embodiment of the present invention. DETAILED DESCRIPTION

[0085] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.

[0086] Embodiment 1: A deep learning method for identifying the welding effect of photovoltaic cells, such as Figure 1 As shown, the following steps are included:

[0087] S1: Collect the surface image of the battery cell after welding and denoise it to obtain the denoised image:

[0088] S11: Grayscale conversion:

[0089] The collected surface image I of the battery cell after welding is grayed to obtain the grayed image I gray , specifically:

[0090] I gray (m,n)=0.299·I red (m,n)+0.587·I green (m,n)+0.114·I blue (m,n)

[0091] Where (m,n) is the pixel position; I gray (m,n) is the grayscale image I gray The pixel value at (m,n); I red (m,n),I green (m,n) and I blue (m, n) are the pixel values ​​of the red channel image, green channel image and blue channel image of the surface image I of the battery cell after welding at (m, n);

[0092] S12: Image denoising:

[0093] Use Gaussian filtering to remove the grayscale image I gray The noise in is:

[0094]

[0095] Among them, I smooth (m,n) is the denoised image I smooth The pixel value at (m,n); G(i,j) is the Gaussian kernel function, where (i,j) is the Gaussian kernel index, i is the Gaussian kernel horizontal index, and j is the Gaussian kernel vertical index. σ is the standard deviation of the Gaussian kernel, which is 1.1 in this embodiment; k is the radius of the Gaussian kernel, which is 3 in this embodiment.

[0096] The image quality can be effectively enhanced through grayscale and Gaussian filtering preprocessing. Grayscale makes the brightness information of the image more concentrated, which is convenient for subsequent processing; while Gaussian filtering removes random noise in the image through smoothing operations, making the visual effect of the image clearer and providing higher quality image data for feature extraction and defect identification. The denoised image reduces noise interference, making the detailed features of the battery cell more prominent.

[0097] S2: Train the autoencoder network based on the denoised image to extract the deep features of the denoised image:

[0098] S21: Build an autoencoder network:

[0099] The autoencoder consists of an encoder and a decoder, specifically:

[0100] Z=f encoder (I smooth )

[0101]

[0102] Among them, f encoder and f decoder are encoder functions and decoder functions respectively. This embodiment uses the encoder function and decoder function in the U-Net network architecture to construct; Z is the denoised image I smooth Low-dimensional encoding of ; is the reconstructed image;

[0103] S22: Training the autoencoder network:

[0104] The training of the autoencoder network aims to minimize the reconstruction loss, which is:

[0105]

[0106] Among them, Loss rec is the loss value of reconstruction loss; ‖·‖ 2 Calculate the L2 norm;

[0107] Minimize the reconstruction loss Loss based on the stochastic gradient descent method rec The trained autoencoder network is obtained, and the stochastic gradient descent is specifically as follows:

[0108]

[0109] Among them, W rec is the weight of the autoencoder network; η is the learning rate of the autoencoder network, which is 0.001 in this embodiment; Loss is the loss value of the reconstruction loss rec About the weight W of the autoencoder networkrec The gradient of

[0110] S23: Extracting deep features:

[0111] The denoised image is input into the trained autoencoder network. The output of the encoder function in the autoencoder network is the deep features of the denoised image, specifically:

[0112]

[0113] in, The encoder function in the trained autoencoder network; For use The denoised image I is obtained smooth Low dimensional encoding.

[0114] The autoencoder network can automatically learn useful feature representations from raw data, avoiding the complexity and subjectivity of manually designing feature extraction algorithms in traditional methods. This method can discover complex patterns and structures in the data, providing a rich feature basis for subsequent welding effect identification. By learning the intrinsic representation of the data, the autoencoder can capture more generalized features, which enables the trained model to maintain a high recognition accuracy when facing new, unseen data.

[0115] S3: Build a welding effect recognition network, use the denoised image and the corresponding deep features as the input of the welding effect recognition network, and obtain the category probability of the welding effect:

[0116] The welding effect recognition network consists of a sparse mask generation subnetwork, a feature fusion subnetwork, and a recognition subnetwork, specifically:

[0117] S31: Construct sparse mask generation subnetwork:

[0118] Sparse mask subnetwork to extract deep features of denoised images As input, the output sparse mask S is:

[0119]

[0120] Among them, f mask is the sparse mask generation function, the output sparse mask is the same as the denoised image I smooth The mask has the same size and each pixel value in the sparse mask is 0 or 1. In this embodiment, the mask generation function uses the same structure as the decoder function of the autoencoder network;

[0121] S32: Construct feature fusion sub-network:

[0122] Use the generated sparse mask S to guide the denoised image I smooth Perform sparse convolution, specifically:

[0123] O sparse =f sparse_conv (I smooth |S)

[0124] Among them, O sparse is the result of sparse convolution; f sparse_conv is a sparse convolution function, which in this embodiment is composed of a convolution layer and a nonlinear activation layer, and the nonlinear activation layer uses a nonlinear rectification function; sparse_conv (I smooth |S) represents the denoised image I based on the sparse mask S smoott Perform convolution, that is, perform convolution operation only at the pixel position where the pixel value of the sparse mask S is 1;

[0125] Dense convolution is specifically:

[0126] O aense =f dense_conv (I smooth )

[0127] Among them, O densh is the result of dense convolution; f dense_conv is a dense convolution function, which in this embodiment is composed of a convolution layer and a nonlinear activation layer, and the nonlinear activation layer uses a nonlinear rectification function; dense_conv (I smooth ) represents the denoised image I smooth Perform dense convolution, that is, after denoising the image I smooth Convolution is performed on all pixel positions;

[0128] The results of sparse convolution and dense convolution are fused through weighted fusion operation, specifically:

[0129] O fusion =α·O sparse +(1-α)·O dense

[0130] Among them, O fusion is the weighted fusion result; α is the weighting parameter, which is 0.4 in this embodiment;

[0131] S33: Construct recognition subnetwork:

[0132] The weighted fusion result O fusion Input to the recognition subnetwork to obtain the category probability, specifically:

[0133] Y=f detection (O fusion)

[0134] Among them, Y is the probability of welding effect category, including the probability that the surface of the battery cell after welding belongs to normal welding and belongs to different categories of welding problems; f detection To identify the function, the VGG-16 convolutional neural network architecture is used in this embodiment.

[0135] Through the sparse mask generation subnetwork, the model can learn the important feature areas in the image and realize feature selection. Sparse mask allows the network to focus only on the key parts of the image, thereby improving the recognition accuracy and computational efficiency. Sparse convolution is only calculated at the pixel position where the mask is 1, which greatly reduces the computational complexity of the convolution operation. Compared with traditional dense convolution, sparse convolution can significantly reduce the complexity of the model and resource consumption at runtime. The feature fusion subnetwork combines the results of sparse convolution and dense convolution, making full use of the advantages of both convolutions. This fusion strategy enables the network to capture local details and global context information at the same time, enhancing the expressiveness of features.

[0136] S4: Construct a weighted cross entropy loss function and use the momentum-based gradient descent method to optimize the parameters of the welding effect recognition network to obtain the trained welding effect recognition network:

[0137] S41: Construct loss function:

[0138] The weighted cross entropy is used as the loss function of the welding effect recognition network. The weighted cross entropy is specifically:

[0139]

[0140] Among them, Loss det is the loss value of the weighted cross entropy loss; c is the category number, c=1, 2, ..., C, C is the number of categories included in the welding effect, in this embodiment C=6, c=1, 2, ..., 6 represent normal, with solder beads, with burrs, cold soldering, desoldering and over-soldering respectively; y true.c When the true category is the cth category, it is 1, otherwise it is 0. The true category is manually labeled; Y c is the probability of the welding effect category Y being the cth category; w c is the weight of category c;

[0141] S42: Building momentum-based gradient descent:

[0142] Initialize the momentum term V to 0 and update the momentum term in each iteration as follows:

[0143]

[0144] Wherein, β is the momentum coefficient, which is 0.95 in this embodiment; Loss is the loss value of weighted cross entropy loss det About the weight W of welding effect recognition network det The gradient of

[0145] The updated momentum term V is used to update the weights of the welding effect recognition network, specifically:

[0146] W det ←W det -δ·V

[0147] Wherein, δ is the learning rate of the welding effect recognition network, which is 0.001 in this embodiment;

[0148] Step S42 is iteratively executed until a predetermined number of iterations is reached, thereby obtaining a trained welding effect recognition network.

[0149] The weighted cross entropy loss function can effectively solve the problem of class imbalance by assigning weights to different categories, so that the model pays more attention to minority categories or difficult-to-identify welding problems during training. The momentum-based gradient descent method reduces the shock during training and speeds up the convergence speed by accumulating historical gradient information, so that the network parameters are closer to the optimal solution faster.

[0150] S5: Use the trained welding effect recognition network to detect the surface image of the battery cell after welding to obtain the welding effect recognition result:

[0151] The surface image of the battery cell after welding to be identified is input into the trained welding effect recognition network, and the surface image of the battery cell after welding to be identified has no manual labeling of welding effect categories. The welding effect category probability obtained according to the output of the welding effect recognition network is Determine the type of welding effect, specifically:

[0152]

[0153] Among them, the argmax function obtains The dimension number with the largest median value is used to determine whether there is a welding problem on the surface of the battery cell after welding, and the specific type of the welding problem if there is a welding problem.

[0154] Embodiment 2: The present invention also discloses a defective workpiece identification system based on deep learning, comprising the following five modules:

[0155] Denoising module: collects the surface image of the battery cell after welding and performs denoising;

[0156] Encoding module: train the autoencoder network based on the denoised image;

[0157] Network construction module: construct welding effect recognition network;

[0158] Network training module: construct a weighted cross entropy loss function and use the momentum-based gradient descent method to optimize the parameters of the welding effect identification network;

[0159] Network application module: Use the trained welding effect recognition network to detect the surface image of the battery cell after welding.

[0160] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0161] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0162] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A deep learning method for identifying the welding effect of photovoltaic cells, characterized in that: The following steps are involved: S1: Collect the surface image of the battery cell after welding and denoise it to obtain a denoised image; S2: Train the autoencoder network based on the denoised image to extract the deep features of the denoised image; S3: construct a welding effect recognition network, use the denoised image and the corresponding deep features as inputs of the welding effect recognition network, and obtain the category probability of the welding effect; S4: construct a weighted cross entropy loss function and use the momentum-based gradient descent method to optimize the parameters of the welding effect recognition network to obtain a trained welding effect recognition network; S5: Use the trained welding effect recognition network to detect the surface image of the battery cell after welding to obtain the welding effect recognition result.

2. The photovoltaic cell welding effect identification method based on deep learning according to claim 1 is characterized in that: The step S1 includes the following steps: S11: Grayscale conversion: The collected surface image I of the battery cell after welding is grayed to obtain the grayed image I gray , specifically: I gray (m,n)=0.299·I red (m,n)+0.587·I green (m,n)+0.114·I blue (m,n) Where (m,n) is the pixel position; I gray (m,n) is the grayscale image I gray The pixel value at (m,n); I red (m,n),I green (m,n) and I blue (m, n) are the pixel values ​​of the red channel image, green channel image and blue channel image of the surface image I of the battery cell after welding at (m, n); S12: Image denoising: Use Gaussian filtering to remove the grayscale image I gray The noise in is: Among them, I smooth (m,n) is the denoised image I smooth The pixel value at (m,n); G(i,j) is the Gaussian kernel function, where (i,j) is the Gaussian kernel index, i is the Gaussian kernel horizontal index, and j is the Gaussian kernel vertical index. σ is the standard deviation of the Gaussian kernel; k is the radius of the Gaussian kernel.

3. The photovoltaic cell welding effect identification method based on deep learning according to claim 2 is characterized in that: The step S2 includes the following steps: S21: Build an autoencoder network: The autoencoder consists of an encoder and a decoder, specifically: Z=f encoder (I smooth ) Among them, f encoder and f decoder are encoder function and decoder function respectively; Z is the denoised image I smooth Low-dimensional encoding of ; is the reconstructed image; S22: Training the autoencoder network: The training of the autoencoder network aims to minimize the reconstruction loss, which is: Among them, Loss rec is the loss value of the reconstruction loss; ‖·‖2 calculates the L2 norm; Minimize the reconstruction loss Loss based on the stochastic gradient descent method rec The trained autoencoder network is obtained, and the stochastic gradient descent is specifically as follows: Among them, W rec is the weight of the autoencoder network; η is the learning rate of the autoencoder network; Loss is the loss value of the reconstruction loss rec About the weight W of the autoencoder network rec The gradient of S23: Extracting deep features: The denoised image is input into the trained autoencoder network. The output of the encoder function in the autoencoder network is the deep features of the denoised image, specifically: in, The encoder function in the trained autoencoder network; For use The denoised image I is obtained smooth Low dimensional encoding.

4. The photovoltaic cell welding effect identification method based on deep learning according to claim 3 is characterized in that: The step S3 includes the following steps: The welding effect recognition network consists of a sparse mask generation subnetwork, a feature fusion subnetwork, and a recognition subnetwork, specifically: S31: Construct sparse mask generation subnetwork: Sparse mask subnetwork to extract deep features of denoised images As input, the output sparse mask S is: Among them, f mask is the sparse mask generation function, the output sparse mask is the same as the denoised image I smoith They have the same size and each pixel value in the sparse mask is either 0 or 1; S32: Construct feature fusion sub-network: Use the generated sparse mask S to guide the denoised image I smooth Perform sparse convolution, specifically: O sparse =f sparse_conv (I smooth |S) Among them, O sparse is the result of sparse convolution; f sparse_conv is the sparse convolution function; f sparse_conv (I smooth |S) represents the denoised image I based on the sparse mask S smooth Perform convolution, that is, perform convolution operation only at the pixel position where the pixel value of the sparse mask S is 1; Dense convolution is specifically: O dense =f dense_conv (I smooth ) Among them, O dense is the result of dense convolution; f dense _ conv is the dense convolution function; f dense _ conv (I smooth ) represents the denoised image I smooth Perform dense convolution, that is, after denoising the image I smooth Convolution is performed on all pixel positions; The results of sparse convolution and dense convolution are fused through weighted fusion operation, specifically: The fusion =α·s parse +(1-a)·O dense Among them, O fusion is the weighted fusion result; α is the weighting parameter; S33: Construct recognition subnetwork: The weighted fusion result O fusion Input to the recognition subnetwork to obtain the category probability, specifically: Y=f detection (O fusion ) Among them, Y is the probability of welding effect category, including the probability that the surface of the battery cell after welding belongs to normal welding and belongs to different categories of welding problems; f derection To identify the function, a convolutional neural network is used.

5. The photovoltaic cell welding effect identification method based on deep learning according to claim 4 is characterized in that: The step S4 includes the following steps: S41: Construct loss function: The weighted cross entropy is used as the loss function of the welding effect recognition network. The weighted cross entropy is specifically: Among them, Loss det is the loss value of weighted cross entropy loss; c is the category number, c = 1, 2, ..., C, C is the number of categories contained in the welding effect; y true.c When the true category is the cth category, it is 1, otherwise it is 0. The true category is manually labeled; Y c is the probability of the welding effect category Y being the cth category; w c is the weight of category c; S42: Building momentum-based gradient descent: Initialize the momentum term V to 0 and update the momentum term in each iteration as follows: Among them, β is the momentum coefficient; Loss is the loss value of weighted cross entropy loss det About the weight W of the welding effect recognition network det The gradient of The updated momentum term V is used to update the weights of the welding effect recognition network, specifically: W det ←W det -δ·V Among them, δ is the learning rate of the welding effect recognition network; Step S42 is iteratively executed until a predetermined number of iterations is reached, thereby obtaining a trained welding effect recognition network.

6. The photovoltaic cell welding effect identification method based on deep learning according to claim 5 is characterized in that: The step S5 includes the following steps: The surface image of the battery cell after welding to be identified is input into the trained welding effect recognition network, and the surface image of the battery cell after welding to be identified has no manual labeling of welding effect categories. The welding effect category probability obtained according to the output of the welding effect recognition network is Determine the type of welding effect, specifically: Among them, the argmax function obtains The dimension number with the largest median value is used to determine whether there is a welding problem on the surface of the battery cell after welding, and the specific type of the welding problem if there is a welding problem.

7. A deep learning photovoltaic cell welding effect recognition system, characterized in that: include: Denoising module: collects the surface image of the battery cell after welding and performs denoising; Encoding module: train the autoencoder network based on the denoised image; Network construction module: construct welding effect recognition network; Network training module: construct a weighted cross entropy loss function and use the momentum-based gradient descent method to optimize the parameters of the welding effect identification network; Network application module: Use the trained welding effect recognition network to detect the surface image of the battery cell after welding; To realize a deep learning method for identifying welding effects of photovoltaic cells as described in any one of claims 1-6.

Images