CN117314892A - Incremental learning-based continuous optimization method for solar cell defect detection - Google Patents

Abstract

The invention relates to the field of computer vision technology, and specifically discloses a continuous optimization method for solar cell defect detection based on incremental learning, which constructs a defect detection model and continuously optimizes and updates the defect detection model. Specifically, it continuously inputs defect data and updates features. The extractor, auxiliary classifier, feature fusion and classifier are incrementally trained using the gradient descent method to construct an example set of unknown defect categories and adjust the example set of known defect categories. The feature extractor is trained based on the geometric center of the feature extractor. Pruning enables defect detection that greatly reduces catastrophic forgetting and continuously optimizes both old and new task detection. The advantage is that a small number of category data sets can start the defect detection process, and only a very small sample set of old data sets and an input training set need to be stored at each stage, ensuring continuous optimization and update of the defect detection model, while reducing storage and computing loads. .

Description

A continuous optimization method for solar cell defect detection based on incremental learning

Technical field

The invention relates to the field of computer vision technology, and in particular to a continuous optimization method for solar cell defect detection based on incremental learning.

Background technique

Driven by new energy and energy-saving policies, solar cells have been applied and produced on a large scale. However, when the manufacturing process fails, the batteries produced will have certain defects. Detection and analysis of defects occurring during the production process and timely adjustment of the corresponding faulty process can reduce losses. Therefore, real-time and accurate detection of battery chip defect categories can be achieved to adjust the corresponding faults. Craftsmanship is very necessary.

In the process of mass production of solar cells, if we collect their defect data, we will find that the defect data has a long-tail distribution. Some categories (head defect categories) have fewer categories but more samples for each category. The opposite is true for the tail defect categories. , which makes it difficult to obtain a balanced data set containing all defect categories at one time. Only a small number of defect category data sets can be obtained, and the remaining defect categories cannot be learned at one time. Affected by actual production factors, defect manifestations may change. As a result, known defect categories learned by the model have new feature values; adjustments to the manufacturing process may also lead to the generation of new defect categories. Therefore, a small amount of defect category data sets can be used as a basic category training set to train a basic neural network model. On the basis of retaining old knowledge, we can continuously and in stages train and learn the data of unknown defect categories that continue to arrive and the data containing new feature values. Known defect categories to continuously optimize inspection capabilities.

Currently, the intelligent defect detection of solar cells still adopts the traditional offline learning method, which must rely on a complete data set during the training phase, and can only train a model from scratch, which takes a long time to train. Offline learning methods cannot adapt to data changes in time when learning new data that continuously arrives and continuously optimizing the model. Therefore, when applied to current large-scale, high-dimensional, and constantly changing data set learning scenarios, they face update lag and high consumption of computing resources. question.

In summary, in order to solve the problems of high computing resource consumption and lag in updates when the current offline learning method is used to continuously optimize the model, a continuous optimization method for solar cell defect detection based on incremental learning is urgently needed.

Contents of the invention

The purpose of the present invention is to provide a continuous optimization method for solar cell defect detection based on incremental learning. The specific technical solutions are as follows:

A continuous optimization method for solar cell defect detection based on incremental learning, including the following steps:

Step S1, defect data continues to be input:

Input the example set and defect data. The example set is a subset of the old data set. The defect data includes unknown category data and known category data containing new feature values. The union of the example set and defect data is the training set;

Step S2, feature extractor expansion:

Build a new feature extractor and initialize the new feature extractor using the parameters of the feature extractor in the old model to obtain a new feature extractor;

Step S3, auxiliary classifier construction:

Build an auxiliary classifier to process all known category data as one category and calculate auxiliary losses on new features;

Step S4, feature fusion device construction:

Extract features of solar cell images from multiple scales through a new feature extractor, connect all the extracted features to obtain an intermediate feature map, refine the intermediate feature map based on the channel and spatial attention mechanism, input the intermediate feature map, and infer a sequence in turn One-dimensional attention map and two-dimensional attention map are used to obtain refined feature maps through the feature fusion device;

Step S5, classifier update:

Copy the parameters used for the old features from the old model classifier to build the current stage classifier, and add corresponding fully connected layer nodes according to the number of unknown defect categories at the current stage. The new parameters are randomly initialized, and the refined feature map is sent to the current stage for classification. Get the predicted probability from the processor and calculate the cross entropy loss;

Step S6, use stochastic gradient descent method for incremental training:

The old model is incrementally trained based on the training set. According to the global loss including classification loss and auxiliary loss, the stochastic gradient descent method is used to learn unknown category samples and known category samples containing new feature values. The unknown category samples are Convert to known category samples to obtain an updated new model;

Step S7, construction of unknown defect category example set and adjustment of known defect category example set:

Based on the new model, unknown defect category data and known defect category data are trained to obtain an unknown category example set and a known category example set with new features, respectively. The unknown category example set, known category example set and step S1 The example sets are merged to form a new example set, which is used for the next stage of incremental training;

Step S8, feature extractor pruning:

Calculate the geometric center of each layer of the new feature extractor in the new model, and prune the convolution kernel of this layer based on the geometric center;

Step S9: Continuously optimize and update the defect detection model:

Continuously input unknown category defect data and known category defect data, repeat steps S1 to S8, continue to update the old model, and obtain a globally unified defect detection network for the current unknown categories and known categories to achieve continuous optimization of defect detection.

Preferably, in step S1, for the continuously input defect data Perform optimization detection, where/> ,when When , the model to be updated is the basic neural network model/> ,/> From the basic category training set/> Obtained by training.

Preferably, in step S2, the feature extractor is constructed based on a multi-scale residual network.

Preferably, in step S3, the auxiliary classifier uses probability For prediction classification, the expression is as follows:

;

in, Express/> function;/> represents an auxiliary classifier whose label space is ,/> Represents a label set of unknown categories;/> represents new features;

Auxiliary loss about new features The expression is as follows:

;

in, Represents the training set;/> Represents an image;/> Represents the corresponding category label.

Preferably, in step S4, the channel and spatial attention mechanism is specifically the channel attention sub-network and the spatial attention sub-network. The specific process of obtaining the refined feature map through the feature fusion device is as follows:

Channel attention sub-network passes and/> Compress intermediate feature map/> The spatial dimension of ,/> The spatial dimension of is/> , generate two feature vectors/> and/> , whose dimensions are all/> , connect the two eigenvectors to get/> , using two fully convolutional layer pairs/> For processing, the expression is as follows:

;

;

in, Indicates the weight of each channel;/> Represents the activation function; the first fully convolutional layer The number of output channels is/> ,/> Represents the reduction ratio; the second fully convolutional layer/> The convolution kernel size is/> , the number of output channels is/> ;/> Represents channel characteristics;

The spatial attention sub-network takes channel features as input and applies average and maximum operations along the channel axis to generate two feature maps, expressed as follows:

;

;

in, and/> Represents the element average and maximum operations in the channel axis respectively;

based on and/> Get the dimension as/> Spatial attention map/> , the expression is as follows:

;

The refined feature map is obtained by multiplying the spatial attention map and channel features, and the expression is as follows:

;

in, Represents spatial refinement features, that is, refined feature maps/> .

Preferably, in step S5, predict the probability The expression is as follows:

;

in, represents the classifier;

cross entropy loss The expression is as follows:

;

in, Represents the training set;/> Represents an image;/> Represents the corresponding category label.

Preferably, in step S6, the global loss The expression is as follows:

;

in, Represents the hyperparameters that control the effect of the auxiliary classifier, for the initial model, i.e./> when, .

Preferably, in step S7, the example set update method is as follows:

Build an unknown category example set , for unknown category data set/> Randomly sample to obtain the initial unknown category example set/> , use model/> Initialize intermediate model/> , and in/> training , obtained through several iterations of stochastic gradient descent/> , the iteration process is as follows:

;

in, Indicates fine-tuning/> The learning rate,/> Represents cross entropy loss, /> Express/> gradient;

calculate on/> loss, and backpropagates the verification loss to optimize/> , get the final example set of unknown category samples/> , the propagation process is as follows:

;

in, Represents the learning rate of the propagation process of the unknown category example set,/> Represents cross entropy loss, /> Express/> gradient;

Adjust the set of examples of known categories , for known category data sets/> Obtained by random sampling/> ;Secondly, use/> Initialize intermediate model/> , and in/> On training/> , obtained through several iterations of stochastic gradient descent/> , the iteration process is as follows:

;

calculate on/> loss, and backpropagates the verification loss to optimize/> , get the final set of known category examples with new feature values/> , the propagation process is as follows:

;

in, Represents the learning rate of the propagation process of the known sample example set.

Preferably, in step S8, the geometric center is expressed as follows:

;

;

in, Express/> No./> The convolutional layer/> A convolution kernel whose dimension is/> ,/> and/> Respectively represent the first/> The number of input channels and the number of output channels of a convolutional layer;

when No./> The geometric center of the convolutional layer/> Located at page/> Among the convolution kernels of a convolutional layer, the following formula is satisfied:

;

in, Indicates the first/> A convolutional layer has a geometric center/> Convolution kernels with the same value, that is, convolution kernels that need to be pruned,/> Express/> and/> The second norm of the difference;

Calculated by the formula There are convolution kernels that need to be pruned in the convolution layer to achieve model pruning. The expression is as follows:

;

in, Indicates the first/> A convolutional layer has a geometric center/> Convolution kernels with the same or similar values, that is, convolution kernels that need to be pruned, /> Express/> and/> The second norm of the difference.

Preferably, in step S9, unknown and known category defect data are continuously input, and based on the old model, feature extractor expansion, auxiliary classifier construction, feature fusion construction, classifier update, and stochastic gradient descent method are used to increase the features in sequence. Through quantitative training, unknown defect category example set construction and known defect category example set adjustment, and feature extractor pruning steps, a globally unified defect detection network for the current unknown and known categories is obtained. , to achieve continuous optimization of defect detection;

The specific detection process is as follows:

Based on defect detection network , for all received solar cell images to be detected/> Perform defect detection; first, based on feature extractor/> , the extracted features are connected to obtain an intermediate feature map/> ,

;

Then, according to step S3, through the feature fusion device, the intermediate feature map Transform into refined feature map/> ,

;

;

will refine the feature map Send to classifier/> Predict the defect category label of the solar cell/> , and based on this, the defect category to which it belongs is obtained. The expression is as follows:

;

.

Applying the technical solution of the present invention has the following beneficial effects:

The continuous optimization method for solar cell defect detection based on incremental learning disclosed by the present invention can start the network learning process by using the initial data set, and only needs to store the old data set, a very small-scale example set and the input training set at each stage, effectively Reduces storage and computing load. During the network learning process, samples of unknown categories are continuously learned to achieve accurate detection of all defect categories. When learning unknown defect categories, the network's catastrophic forgetting of known defect categories is effectively alleviated. And for the new manifestations that may appear in known defect categories, the updates required for the detection of known defect categories are implemented.

The dynamic feature extractor expansion, auxiliary classifier construction and feature fusion device construction in the present invention are aimed at the emergence of unknown categories. In order to achieve a balance between model stability and plasticity, a new feature extractor is created when unknown categories appear. , and initialized using the parameters of the old feature extractor, retaining the knowledge of known categories while also effectively learning the characteristics of unknown categories. The auxiliary classifier construction is used to train the newly constructed feature extractor. Feature fusion effectively solves the possible confusion problem between known categories and unknown categories in incremental learning.

In addition, the present invention not only constructs an example set for the emergence of unknown categories, but also considers known categories with new feature values. When unknown categories appear and known categories change, the present invention dynamically builds unknown category example sets and adjusts existing categories. A set of examples of known categories. The present invention also proposes feature extractor pruning, which achieves substantial parameter reduction while maintaining model performance as much as possible, reduces the storage and calculation requirements of the model, and improves the practicality of the model.

In addition to the objects, features and advantages described above, the present invention has other objects, features and advantages. The present invention will be described in further detail below with reference to the drawings.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions of the prior art more clearly, the following will briefly introduce the drawings needed to describe the embodiments or the prior art. Obviously, the drawings in the following description are only For some embodiments of the present invention, those of ordinary skill in the art can also obtain other drawings based on these drawings without exerting creative efforts.

Figure 1 is a step flow chart of a method for continuous optimization of solar cell defect detection in a preferred embodiment of the present invention;

Figure 2 is a flow chart of updating the unknown category example set and the known category example set in the preferred embodiment of the present invention.

Detailed ways

The core of the invention is to provide a continuous optimization method for solar cell defect detection based on incremental learning. In the existing technology, intelligent defect detection of solar cells still uses the traditional offline learning method, which must rely on a complete data set during the training phase, and can only train a model from scratch, and the training is time-consuming. Therefore, the offline learning method cannot adapt to data changes in time when learning the continuously arriving new data and continuously optimizing the model. Therefore, when applied to the current large-scale, high-dimensional, and constantly changing data set processing scenarios, it faces update lag and high consumption of computing resources. And other issues.

The continuous optimization method for solar cell defect detection based on incremental learning disclosed by the present invention can start the network learning process by using the initial data set, and only needs to store the old data set, a very small-scale example set and the input training set at each stage, effectively Reduces storage and computing load. During the network learning process, samples of unknown categories are continuously learned to achieve accurate detection of all defect categories. When learning unknown defect categories, the network's catastrophic forgetting of known defect categories is effectively alleviated. And for the new manifestations that may appear in known defect categories, the updates required for the detection of known defect categories are implemented.

In order to enable those skilled in the art to better understand the solution of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

Example:

Please refer to Figure 1. Figure 1 is a step flow chart of a continuous optimization method for solar cell defect detection in a preferred embodiment of the present invention. Steps S1-S8 are updating the model based on features in the old model and known category data, and step S9 is continuing Update the old model to achieve a solar cell defect detection network with continuously optimized and updated models.

Referring to Figure 1, in the embodiment of the present invention, a method for continuous optimization of solar cell defect detection includes the following steps:

Step S1, defect data continues to be input:

Enter the example set and defect data of solar cell data. The example set is a subset of the old data set (the old data set is the defect data used for training in the old model). The defect data includes unknown category data and contains new features. The union of the known category data of the value, the example set and the defect data is the training set; for the continuously input defect data Carry out phased optimization testing, in which/> , when/> When , the model to be updated is the basic neural network model/> ,/> From the basic category training set/> Obtained by training.

Step S2, feature extractor expansion:

Construct a new feature extractor and use all the parameters of the existing feature extractor to initialize the new feature extractor to obtain a new feature extractor; the feature extractor is based on the multi-scale residual network (Inception-ResNet, the network model is an existing technology and will not be described in detail here). In this embodiment, in order to avoid catastrophic forgetting of old knowledge and maintain previously learned knowledge, all current feature extractors are frozen parameters. Then, in order to more effectively adapt to the feature distribution of unknown categories and thereby perform more accurate detection of unknown defect categories, a new feature extractor is constructed/> , and copy/> parameters to initialize/> .

Step S3, auxiliary classifier construction:

Construct an auxiliary classifier, process all known category data as one category, and calculate auxiliary losses related to new features; the auxiliary classifier uses probability For prediction classification, the expression is as follows:

;

in, Express/> function;/> represents an auxiliary classifier whose label space is ,/> Represents a label set of unknown categories;/> represents new features;

Auxiliary loss about new features The expression is as follows:

;

in, Represents the training set;/> Represents an image;/> Represents the corresponding category label.

Step S4, feature fusion device construction:

The features of the solar cell image are extracted from multiple scales through a new feature extractor, and all the extracted features are connected to obtain an intermediate feature map. The intermediate feature map is input based on the channel and spatial attention mechanism, and the feature fusion device infers one-dimensional attention in turn. Force map and two-dimensional attention map are used to obtain refined feature maps through feature fusion;

In this embodiment, the channel and spatial attention mechanisms are specifically the channel attention sub-network and the spatial attention sub-network. The specific process of obtaining the refined feature map through the feature fusion device is as follows:

Channel attention sub-network passes (global average pooling) and/> (Global max pooling) compresses intermediate feature maps/> The spatial dimension of ,/> The spatial dimension of is/> (C represents the number of channels of the feature map; H represents the height of the feature map; W represents the width of the feature map), generating two feature vectors/> and/> , whose dimensions are all , connect the two eigenvectors to get/> , using two fully convolutional layer pairs/> For processing, the expression is as follows:

;

;

in, Indicates the weight of each channel;/> Represents the activation function; the first fully convolutional layer The number of output channels is/> ,/> Represents the reduction ratio; the second fully convolutional layer/> The convolution kernel size is/> , the number of output channels is/> ;/> Represents channel characteristics;

The spatial attention sub-network takes channel features as input and applies average and maximum operations along the channel axis to generate two feature maps, expressed as follows:

;

;

in, and/> Represents the element average and maximum operations in the channel axis respectively;

based on and/> Get the dimension as/> Spatial attention map/> , the expression is as follows:

;

The refined feature map is obtained by multiplying the spatial attention map and channel features, and the expression is as follows:

;

in, Represents spatial refinement features, that is, refined feature maps/> .

Step S5, classifier update:

Copy the parameters used for the old features from the old model classifier to build a new model classifier, and add corresponding fully connected layer nodes according to the number of unknown defect categories in the old model. The new parameters are randomly initialized and the refined feature map is sent to The predicted probability is obtained from the classifier and the cross-entropy loss is calculated;

Specifically, the predicted probability The expression is as follows:

;

in, represents the classifier;

cross entropy loss The expression is as follows:

.

Step S6, use stochastic gradient descent method for incremental training:

The old model is incrementally trained based on the training set. According to the global loss including classification loss and auxiliary loss, the stochastic gradient descent method is used to learn unknown category samples and known category samples containing new feature values. The unknown category samples are Convert to known category samples to obtain an updated new model;

Specifically, in step S6, the global loss The expression is as follows:

;

in, Represents the hyperparameters that control the effect of the auxiliary classifier, for the initial model, i.e./> when, .

Step S7, construction of unknown defect category example set and adjustment of known defect category example set:

Based on the new model, unknown defect category data and known defect category data are trained to obtain an unknown category example set and a known category example set with new features, respectively. The unknown category example set, known category example set and step S1 The example sets are merged to form a new example set for incremental training of new models;

Specifically, as shown in Figure 2, the example set update method is as follows:

Build an unknown category example set , for unknown category data set/> Randomly sample to obtain an unknown category example set/> , use model/> Initialize intermediate model/> , and in/> On training/> , obtained through several iterations of stochastic gradient descent/> , the iteration process is as follows:

;

in, Indicates fine-tuning/> learning rate;

calculate on/> loss, and backpropagate the above verification loss to optimize/> , get the final example set of unknown category samples/> , the propagation process is as follows:

;

in, Represents the learning rate of the propagation process of the unknown category example set;

Adjust the set of examples of known categories , a known category data set for the new model/> Obtained by random sampling/> . Second, use/> Initialize and get the intermediate model/> , and in/> On training/> , obtained through several iterations of stochastic gradient descent/> , the iteration process is as follows:

;

calculate on/> loss, and backpropagates the verification loss to optimize/> , get the final set of known category examples with new feature values/> , the propagation process is as follows:

;

in, Represents the learning rate of the propagation process of the known sample example set.

Step S8, feature extractor pruning:

Calculate the geometric center of each layer of the new feature extractor in the new model, and prune the convolution kernel of this layer based on the geometric center;

Specifically, the geometric center is expressed as follows:

;

;

in, Express/> No./> The convolutional layer/> A convolution kernel whose dimension is/> ,/> and/> Respectively represent the first/> The number of input channels and the number of output channels of a convolutional layer;

when No./> The geometric center of the convolutional layer/> Located at page/> Among the convolution kernels of a convolutional layer, the following formula is satisfied:

;

in, Indicates the first/> A convolutional layer has a geometric center/> Convolution kernels with the same value, that is, convolution kernels that need to be pruned,/> Express/> and/> The second norm of the difference;

Calculated by the formula There are convolution kernels that need to be pruned in the convolution layer to achieve model pruning. The expression is as follows:

;

in, Indicates the first/> A convolutional layer has a geometric center/> Convolution kernels with the same or similar values, that is, convolution kernels that need to be pruned, /> Express/> and/> The second norm of the difference.

It should be noted that if in the same layer, the values of some convolution kernels are close to or equal to the geometric center of the layer, then these convolution kernels can be represented by other convolution kernels of the layer, and these convolution kernel pairs are pruned. The impact on network performance is very small. By pruning these convolution kernels, it is possible to maintain network performance while optimizing the storage requirements of the model and accelerating training and prediction.

Step S9: Continuously optimize and update the defect detection model:

Continuously input unknown category defect data and known category defect data, repeat steps S1 to S8, and continuously update the network to obtain a globally unified defect detection network for the current unknown categories and known categories to achieve continuous optimization of defect detection.

Specifically, unknown and known categories of defect data are continuously input, and based on the old model, feature extractor expansion, auxiliary classifier construction, feature fusion construction, classifier update, incremental training using stochastic gradient descent method, and unknown defects are sequentially performed. Class example set construction, adjustment of known defect category example sets, and feature extractor pruning steps yield a globally unified defect detection network for currently unknown and known categories. , thereby achieving continuous optimization of defect detection;

The specific detection process is as follows:

Based on defect detection network , for all received solar cell images to be detected/> Perform defect detection. First, based on the feature extractor/> , the extracted features are connected to obtain an intermediate feature map/> ,

;

Then, according to step S3, through the feature fusion device, the intermediate feature map Transform into refined feature map/> ,

;

;

will refine the feature map Send to classifier/> Predict the defect category label of the solar cell/> , and based on this, the defect category to which it belongs is obtained. The expression is as follows:

;

.

The above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims (10)

1. A continuous optimization method for solar cell defect detection based on incremental learning, which is characterized by including the following steps: Step S1, defect data continues to be input: Input the example set and defect data. The example set is a subset of the old data set. The defect data includes unknown category data and known category data containing new feature values. The union of the example set and defect data is the training set; Step S2, feature extractor expansion: Build a new feature extractor and initialize the new feature extractor using the parameters of the feature extractor in the old model to obtain a new feature extractor; Step S3, auxiliary classifier construction: Build an auxiliary classifier to process all known category data as one category and calculate auxiliary losses on new features; Step S4, feature fusion device construction: Extract features of solar cell images from multiple scales through a new feature extractor, connect all the extracted features to obtain an intermediate feature map, refine the intermediate feature map based on the channel and spatial attention mechanism, input the intermediate feature map, and infer a sequence in turn One-dimensional attention map and two-dimensional attention map are used to obtain refined feature maps through the feature fusion device; Step S5, classifier update: Copy the parameters used for the old features from the old model classifier to build the current stage classifier, and add corresponding fully connected layer nodes according to the number of unknown defect categories at the current stage. The new parameters are randomly initialized, and the refined feature map is sent to the current stage for classification. Get the predicted probability from the processor and calculate the cross entropy loss; Step S6, use stochastic gradient descent method for incremental training: The old model is incrementally trained based on the training set. According to the global loss including classification loss and auxiliary loss, the stochastic gradient descent method is used to learn unknown category samples and known category samples containing new feature values. The unknown category samples are Convert to known category samples to obtain an updated new model; Step S7, construction of unknown defect category example set and adjustment of known defect category example set: Based on the new model, unknown defect category data and known defect category data are trained to obtain an unknown category example set and a known category example set with new features, respectively. The unknown category example set, known category example set and step S1 The example sets are merged to form a new example set, which is used for the next stage of incremental training; Step S8, feature extractor pruning: Calculate the geometric center of each layer of the new feature extractor in the new model, and prune the convolution kernel of this layer based on the geometric center; Step S9: Continuously optimize and update the defect detection model: Continuously input unknown category defect data and known category defect data, repeat steps S1 to S8, continue to update the old model, and obtain a globally unified defect detection network for the current unknown categories and known categories to achieve continuous optimization of defect detection. 2. The continuous optimization method for solar cell defect detection according to claim 1, characterized in that, in step S1, for the continuously input defect data Perform optimization detection, where/> , when/> When , the model to be updated is the basic neural network model/> ,/> From the basic category training set/> Obtained by training. 3. The continuous optimization method for solar cell defect detection according to claim 2, characterized in that, in step S2, the feature extractor is constructed based on a multi-scale residual network. 4. The method for continuous optimization of solar cell defect detection according to claim 3, characterized in that, in step S3, the auxiliary classifier uses probability For prediction classification, the expression is as follows: ; in, Express/> function;/> represents an auxiliary classifier whose label space is ,/> Represents a label set of unknown categories;/> represents new features; Auxiliary loss about new features The expression is as follows: ; in, Represents the training set;/> Represents an image;/> Represents the corresponding category label. 5. The continuous optimization method for solar cell defect detection according to claim 4, characterized in that, in step S4, the channel and spatial attention mechanism is specifically a channel attention sub-network and a spatial attention sub-network, through a feature fusion device The specific process of obtaining the refined feature map is as follows: Channel attention sub-network passes and/> Compress intermediate feature map/> The spatial dimension of ,/> The spatial dimension of is/> , generate two feature vectors/> and/> , whose dimensions are all/> , connect the two eigenvectors to get/> , using two fully convolutional layer pairs/> For processing, the expression is as follows: ; ; in, Indicates the weight of each channel;/> Represents the activation function; the first fully convolutional layer The number of output channels is/> ,/> Represents the reduction ratio; the second fully convolutional layer/> The convolution kernel size is/> , the number of output channels is/> ;/> Represents channel characteristics; The spatial attention sub-network takes channel features as input and applies average and maximum operations along the channel axis to generate two feature maps, expressed as follows: ; ; in, and/> Represents the element average and maximum operations in the channel axis respectively; based on and/> Get the dimension as/> Spatial attention map/> , the expression is as follows: ; The refined feature map is obtained by multiplying the spatial attention map and channel features, and the expression is as follows: ; in, Represents spatial refinement features, that is, refined feature maps/> . 6. The method for continuous optimization of solar cell defect detection according to claim 5, characterized in that, in step S5, the prediction probability The expression is as follows: ; in, represents the classifier; cross entropy loss The expression is as follows: ; in, Represents the training set;/> Represents an image;/> Represents the corresponding category label. 7. The continuous optimization method for solar cell defect detection according to claim 6, characterized in that, in step S6, the global loss The expression is as follows: ; in, Represents the hyperparameters that control the effect of the auxiliary classifier, for the initial model, i.e./> when, . 8. The continuous optimization method for solar cell defect detection according to claim 7, characterized in that, in step S7, the example set update method is as follows: Build an unknown category example set , for unknown category data set/> Randomly sample to obtain the initial unknown category example set/> , use model/> Initialize intermediate model/> , and in/> On training/> , obtained through several iterations of stochastic gradient descent/> , the iteration process is as follows: ; in, Indicates fine-tuning/> The learning rate,/> Represents cross entropy loss, /> Express/> gradient; calculate on/> loss, and backpropagates the verification loss to optimize/> , get the final example set of unknown category samples/> , the propagation process is as follows: ; in, Represents the learning rate of the propagation process of the unknown category example set,/> Represents cross entropy loss, /> Express/> gradient; Adjust the set of examples of known categories , for known category data sets/> Obtained by random sampling/> ;Secondly, use/> Initialize intermediate model/> , and in/> Up training/> , obtained through several iterations of stochastic gradient descent/> , the iteration process is as follows: ; calculate on/> loss, and backpropagates the verification loss to optimize/> , get the final set of known category examples with new feature values/> , the propagation process is as follows: ; in, Represents the learning rate of the propagation process of the known sample example set. 9. The continuous optimization method for solar cell defect detection according to claim 8, characterized in that, in step S8, the geometric center is expressed as follows: ; ; in, Express/> No./> The convolutional layer/> A convolution kernel whose dimension is/> ,/> and/> Respectively represent the first/> The number of input channels and the number of output channels of a convolutional layer; when No./> The geometric center of the convolutional layer/> Located at page/> Among the convolution kernels of a convolutional layer, the following formula is satisfied: ; in, Indicates the first/> A convolutional layer has a geometric center/> Convolution kernels with the same value, that is, convolution kernels that need to be pruned,/> Express/> and/> The second norm of the difference; Calculated by the formula There are convolution kernels that need to be pruned in the convolution layer to achieve model pruning. The expression is as follows: ; in, Indicates the first/> A convolutional layer has a geometric center/> Convolution kernels with the same or similar values, that is, convolution kernels that need to be pruned, /> Express/> and/> The second norm of the difference. 10. The continuous optimization method for solar cell defect detection according to claim 9, characterized in that, in step S9, unknown and known category defect data are continuously input, and feature extractor expansion and auxiliary classifier are sequentially implemented based on the old model. Construction, feature fusion construction, classifier update, incremental training using stochastic gradient descent method, unknown defect category example set construction and known defect category example set adjustment, feature extractor pruning step, to obtain a set of currently unknown and already known defect categories. A globally unified defect detection network with known categories , to achieve continuous optimization of defect detection; The specific detection process is as follows: Based on defect detection network , for all received solar cell images to be detected/> Perform defect detection; first, based on feature extractor/> , the extracted features are connected to obtain an intermediate feature map/> , ; Then, according to step S3, through the feature fusion device, the intermediate feature map Transform into refined feature map/> , ; ; will refine the feature map Send to classifier/> Predict the defect category label of the solar cell/> , and based on this, the defect category to which it belongs is obtained. The expression is as follows: ; .