Strip steel surface defect classification continuous learning method of human-in-the-loop

Abstract

The invention relates to a human-in-loop strip steel surface defect classification continuous learning method, which comprises the following steps: enabling a model to autonomously discover data of a new category on a production line by adopting a new category discovery strategy based on feature comparison, and labeling the data of the new category by using a new category identification strategy of a human-in-loop; when continuous learning is carried out on the new category of streaming data, firstly, a streaming data processing method based on sampling and dynamic buffering is used for processing the data, then a continuous learning algorithm PDGAS based on prototype growth and ethnic group agency is used for learning the new category of data, and the capability of the model is expanded; according to the method, the problems that a model cannot autonomously discover and mark a new class in industrial application, so that the model is difficult to continuously learn the model, and a traditional industrial model cannot continuously grow and cannot cope with an actual industrial streaming data scene are solved.

Description

Technical Field

[0001] The invention relates to the technical field of steel surface defect recognition and classification, and in particular to a human-in-loop continuous learning method for strip steel surface defect classification. Background Art

[0002] Surface defects are common quality problems in strip production, which often mean that there is a problem in a certain link on the production line. Identifying the type of defect can, to a certain extent, reflect which link of the production line the problem occurred in, so that the staff can fix the problem in time and reduce the company's losses. However, traditional defect classification methods often require a lot of manual adjustment and optimization, or by constructing deep learning models, but these deep learning models are usually static after training, and the model weights are fixed and can no longer be adjusted. In addition, the types of defects that the model can classify are fixed at the beginning. In addition to retraining the model, it is impossible to learn new types of defects that appear on the production line. In actual engineering scenarios such as strip surface defect classification, the strip surface images continuously recorded on the production line constitute a data stream reflecting the production line situation. Changes in factors such as materials and processing technology may lead to changes in defect characteristics, but static models cannot identify and classify newly generated defect types, and it is difficult to cope with the diversity of strip surface defects and the ever-changing industrial environment. For this reason, the academic community has proposed a continuous learning method.

[0003] At present, the research on continuous learning methods can be mainly divided into replay, regularization and parameter isolation; replay continuous learning uses two modules: learning module and memory module. The learning module is mainly responsible for completing the learning task, and the memory module is responsible for replaying the required old data to the learning module. By storing the old data and replaying it when the model learns new data, the model's memory of old knowledge is awakened, catastrophic forgetting is alleviated, and the learning of new knowledge and the updating of model capabilities are better completed. It is currently the most effective method to combat catastrophic forgetting. Depending on whether the sample subset is directly stored, it can be divided into direct replay and generated replay. The direct replay method uses the previous samples stored in the memory buffer to fine-tune the network together with the current sample; the generated replay method generates the old task data by itself for replay continuous learning, but the replay method requires additional storage of the data or distribution of the old task. With the continuous increase of learning tasks, the storage cost or computing cost caused by this type of method is also increasing;

[0004] The regularization method does not need to store the data of old tasks, avoiding privacy issues and storage costs. Instead, it strengthens the constraints on the loss function of the new task when the model weight is updated, so as to protect and consolidate the learned knowledge from being forgotten, and maintain the model's memory of old knowledge without storing old data. According to the object to which the constraints are applied, it can be divided into parameter regularization and distribution regularization. The main idea of ​​the parameter regularization method is to quantify the importance of model weights, screen out model weights that are highly relevant to old tasks, and train new tasks while reducing the impact on these weights, so as to achieve the effect of suppressing catastrophic forgetting. The distribution regularization method is based on knowledge distillation loss, so that the predictions of the new and old models on the new task are similar. However, this method cannot effectively constrain the optimization process of the model on the new task. Excessive constraints on model parameters lead to poor learning effects of the model on the new task, which will lead to poor performance of the model on the new task.

[0005] Continuous learning methods based on parameter isolation solve the problem of catastrophic forgetting by assigning specific model parameters to each task, and are divided into dynamic architecture and static architecture methods. Dynamic architecture methods achieve continuous learning by freezing or growing the network. Static architecture methods only activate relevant parameters for each task without modifying the architecture. However, this type of method explicitly associates neurons with different tasks, so task boundaries must be considered during both model training and inference, which limits its application scenarios. Existing continuous learning methods all have certain limitations, and most of them are targeted at static offline scenarios and lack practical value.

[0006] In addition, most of the current continuous learning methods do not take into account the actual application scenarios. The surface streaming data of strip steel on the industrial production line has the characteristics of continuous flow of data, and the data that has flowed through will no longer appear; new types of data will appear, and the number of new and old category data is unbalanced; the data stream presents a non-independent and identically distributed form and other characteristics, which makes it difficult for the model to learn it continuously. Summary of the invention

[0007] In view of this, in order to solve the problem that traditional continuous learning methods are poor in practicality and difficult to deal with streaming, class imbalance and non-independent and identically distributed data in real industrial production scenarios, and introduce strip surface defect classification scenarios to solve the problem that traditional defect classification models cannot identify newly emerging defect types, the present invention provides a human-in-the-loop strip surface defect classification continuous learning method, comprising the following steps:

[0008] S1. Obtaining the flow data of surface defects of strip steel on industrial production lines;

[0009] S2. Processing the strip surface defect streaming data using a streaming data processing method based on sampling and dynamic buffering, including a reservoir sampling method, a memory buffer dynamic allocation method, and a random retrieval strategy;

[0010] S201, using a new class discovery strategy based on feature comparison to determine the newly acquired streaming data x j ,j≥1 whether it belongs to a category that has appeared;

[0011] By computing the input streaming data x j The cosine similarity Ψ(x j ),

[0012] If Ψ(x j )>γ, then the strip surface defect classification model determines the input loss data x j If it belongs to the category that has appeared, execute step S202, where γ represents a preset threshold;

[0013] If Ψ(x j )<γ, then the strip surface defect classification model determines the input loss data x j If it belongs to a category that has not appeared, execute step S203 and report to the data identification platform;

[0014] S202, creating a new category buffer and adjusting the size of each category buffer by using a memory buffer dynamic allocation method;

[0015] S203, design a new type of identification strategy for people in the loop, using the reservoir sampling method to determine whether to use the newly acquired streaming data x j Add to the corresponding category memory buffer;

[0016] S204, using a random retrieval strategy to retrieve an approximately independent and identically distributed playback training batch from a balanced memory buffer, and inputting the batch into a strip surface defect classification model;

[0017] S3, using the continuous learning algorithm PDGAS to optimize the strip surface defect classification model;

[0018] S301, dynamically calculate, update and store category prototypes using prototype storage and growth strategies based on feature information of streaming data batches;

[0019] S302, establishing an attraction set and a repulsion set for the streaming data for continuous learning through a group agent strategy;

[0020] S303, promoting the classifier network to optimize the feature space through the group proxy loss;

[0021] S304. Update the strip surface defect classification model.

[0022] Preferably, the cosine similarity ψ(x j ) is expressed as:

[0023]

[0024] Among them, x j Represents the input streaming data, p i represents the category prototype, f θ (x j ) represents the projection vector of the input streaming data in the feature space.

[0025] Preferably, the reservoir sampling method constructs a reservoir of size k. If the strip surface defect flow data x j If the position j≤k in the data stream, the j strip surface defect streaming data directly enters the water reservoir; if the strip surface defect streaming data x j At position j>k in the data stream, a random number i∈[1,j] is generated. When i≤k, the strip surface defect stream data x j Directly enter the reservoir and replace the original i-th data. When i≤k is not satisfied, discard the data x j .

[0026] Preferably, the memory buffer dynamic allocation method includes: if the newly acquired data x j If the data x is a category that has already appeared, the reservoir sampling method is used to determine whether to store it in the memory buffer of the corresponding category; if the newly acquired data x j If it belongs to a category that has not appeared before, the memory buffer size allocated to each category is recalculated, the memory buffer size of all categories that have appeared is adjusted, samples are deleted according to the current number of cached samples, and a new category memory buffer is created. The reservoir sampling method is used to sample the newly acquired data x j into the new category memory buffer;

[0027] The memory buffer size expression allocated to each category is:

[0028]

[0029] Among them, m represents the memory buffer size allocated to each category, K represents the total strip surface defect streaming data currently acquired, and t represents the number of categories that have appeared currently.

[0030] Preferably, the new category identification strategy of the human in the loop is based on the professional knowledge and experience of human experts. The correct category labels are marked for data that do not belong to the existing categories through a visual interface, and the marked data are transmitted to the strip surface defect classification model through a data transmission interface for continuous learning.

[0031] Preferably, the prototype storage and growth strategy stores the newly acquired streaming data batches to the closest category prototype p in the dynamically changing streaming process. c , the expression is:

[0032]

[0033] Among them, μ∈]0,1] represents the parameter that controls the update amplitude, Represents the feature mean of samples of category c in the current training streaming data batch in the feature space, expressed as:

[0034]

[0035] Among them, B c ={(x i ,y i =c)∈B} represents all sample streaming data contained in category c in the current training streaming data batch B, |B c | represents the number of sample streaming data contained in category c in the current training streaming data batch B, represents the sample streaming data contained in category c, Represents sample streaming data The corresponding feature vector.

[0036] Preferably, in the training streaming data batch B, select Other sample streaming data of the same category The sample streaming data Projected mean on feature space as a proxy for population The group agent and the category prototype p c Composition attraction set The expression is:

[0037]

[0038] The sample streaming data Projection on feature space and the ones belonging to category c are the exclusion set of other categories in the training streaming data batch B The expression is:

[0039]

[0040] in, Indicates sample streaming data of other categories.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. The method provided by the present invention processes the streaming data of strip surface defects on the industrial production line through reservoir sampling, dynamic allocation of memory buffer and random retrieval strategy, thereby enhancing the quality of the model learning category feature space and improving the continuous learning effect of the strip surface defect classification model;

[0043] 2. The present invention processes data through a streaming data processing method based on sampling and dynamic buffering, and uses a continuous learning algorithm PDGAS based on prototype growth and group agents to complete autonomous discovery and labeling of new categories of data and learn, expand the capabilities of the model, and realize semi-autonomous lifelong growth functions;

[0044] 3. The method provided by the present invention solves the problem that the traditional strip surface defect classification model cannot identify new types of defects, thereby improving industrial production efficiency and product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flow chart of a continuous learning method for classifying surface defects of strip steel in a human-in-loop loop according to the present invention;

[0046] Figure 2 It is a data processing flow chart of the present invention;

[0047] Figure 3 This is a flow chart of the continuous learning algorithm PDGAS of the present invention;

[0048] Figure 4 It is a schematic diagram of the water reservoir sampling method of the present invention;

[0049] Figure 5 A schematic diagram of a dynamic allocation strategy for a memory buffer of the present invention;

[0050] Figure 6 A schematic diagram of the random search strategy of the present invention;

[0051] Figure 7 Schematic diagram of the inventor's strategy for identifying new types of circuits. DETAILED DESCRIPTION

[0052] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0053] Example 1

[0054] Most of the current continuous learning methods do not take into account the actual application scenarios. The surface streaming data of strip steel on the industrial production line has the characteristics of continuous flow of data, and the data that has flowed through will no longer appear; new types of data will appear, and the number of new and old category data is unbalanced; the data stream presents the characteristics of non-independent and identically distributed forms, which makes it difficult for the model to continuously learn it. In order to solve the problem that the traditional defect classification model cannot identify the newly emerging defect types, the present invention provides a continuous learning method for strip steel surface defect classification with a human in the loop, comprising the following steps:

[0055] S1. Obtaining the flow data of surface defects of strip steel on industrial production lines;

[0056] S2. Processing the strip surface defect streaming data using a streaming data processing method based on sampling and dynamic buffering, including a reservoir sampling method, a memory buffer dynamic allocation method, and a random retrieval strategy;

[0057] S201, using a new class discovery strategy based on feature comparison to determine the newly acquired streaming data x j ,j≥1, whether it belongs to a category that has appeared before. The new category discovery strategy enables the model to autonomously discover new categories on the production line and report the image data of these new categories;

[0058] By computing the input streaming data x j The cosine similarity Ψ(x j ), the expression is:

[0059]

[0060] Among them, x j Represents the input streaming data, p i represents the category prototype, f θ (x j ) represents the projection vector of the input streaming data in the feature space;

[0061] If Ψ(x j )>γ, then the strip surface defect classification model determines the input loss data x j If it belongs to the category that has appeared, execute step S202, where γ represents a preset threshold;

[0062] If Ψ(x j )<γ, then the strip surface defect classification model determines the input loss data x j If it belongs to a category that has not appeared, execute step S203 and report to the data identification platform;

[0063] like Figure 4As shown in the figure, the reservoir sampling method constructs a reservoir of size k. If the strip surface defect flow data x j If the position j≤k in the data stream, the j strip surface defect streaming data directly enters the water reservoir; if the strip surface defect streaming data x j At position j>k in the data stream, a random number i∈[1,j] is generated. When i≤k, the strip surface defect stream data x j Directly enter the reservoir and replace the original i-th data. When i≤k is not satisfied, discard the data x j ,The reservoir sampling method ensures that the probability of all data of each class being ,selected is equal during the data flow, thus making the samples in the ,memory buffer more representative;

[0064] S202, creating a new category buffer and adjusting the size of each category buffer by using a memory buffer dynamic allocation method;

[0065] like Figure 5 As shown, the memory buffer dynamic allocation method includes: if the newly acquired data x j If the data x is a category that has already appeared, the reservoir sampling method is used to determine whether to store it in the memory buffer of the corresponding category; if the newly acquired data x j If it belongs to a category that has not appeared before, the memory buffer size allocated to each category is recalculated, the memory buffer size of all categories that have appeared is adjusted, samples are deleted according to the current number of cached samples, and a new category memory buffer is created. The reservoir sampling method is used to sample the newly acquired data x j To the new category memory buffer, the memory buffer dynamic allocation strategy dynamically allocates the memory buffer of a specific size evenly to each observed category, so as to cache an equal amount of data for each category and build a balanced memory buffer;

[0066] The memory buffer size expression allocated to each category is:

[0067]

[0068] Among them, m represents the memory buffer size allocated to each category, K represents the total strip surface defect streaming data currently acquired, and t represents the number of categories that have appeared currently;

[0069] S203, design a new type of identification strategy for people in the loop, using the reservoir sampling method to determine whether to use the newly acquired streaming data x j Add to the corresponding category memory buffer;

[0070] like Figure 7As shown, the new category identification strategy of human in the loop is based on the professional knowledge and experience of human experts. The correct category labels are marked for data that do not belong to the existing categories through a visual interface, and the marked data are transmitted to the strip surface defect classification model through a data transmission interface for continuous learning;

[0071] S204, using a random retrieval strategy to retrieve an approximately independent and identically distributed playback training batch from a balanced memory buffer, and inputting it into a strip surface defect classification model, such as Figure 6 As shown in Figure 2, the random retrieval strategy enables the model to learn from more diverse samples, thereby improving the learning effect of the model;

[0072] S3, using the continuous learning algorithm PDGAS to optimize the strip surface defect classification model;

[0073] S301, dynamically calculate, update and store category prototypes using prototype storage and growth strategies based on feature information of streaming data batches;

[0074] The prototype storage and growth strategy stores the newly acquired streaming data batches to the closest category prototype p in the dynamically changing streaming process. c , the expression is:

[0075]

[0076] Among them, μ∈[0,1] represents the parameter that controls the update amplitude, Represents the feature mean of samples of category c in the current training streaming data batch in the feature space, expressed as:

[0077]

[0078] Among them, B c ={(x i ,y i =c)∈B} represents all sample streaming data contained in category c in the current training streaming data batch B, |B c | represents the number of sample streaming data contained in category c in the current training streaming data batch B, represents the sample streaming data contained in category c, Represents sample streaming data The corresponding eigenvector;

[0079] Prototype storage and growth strategies solve problems such as the distribution of streaming data of the same category changing over time that cannot be solved by data processing methods;

[0080] S302, establishing an attraction set and a repulsion set for the streaming data for continuous learning through a group agent strategy;

[0081] Select the same streaming data as the sample streaming data in the training streaming data batch B. Other sample streaming data of the same category The sample streaming data Projected mean on feature space as a proxy for population Group Agent and the category prototype p c Composition attraction set The expression is:

[0082]

[0083] The sample streaming data Projection on feature space and the ones belonging to category c are the exclusion set of other categories in the training streaming data batch B The expression is:

[0084]

[0085] in, Indicates sample streaming data of other categories;

[0086] The group proxy strategy solves the problem that the model is difficult to form a feature space with clear category distribution due to the large intra-class differences and small inter-class differences of the strip surface defect streaming data, thus reducing the classification effect of the model. Based on the attraction set and repulsion set in the group proxy, the inter-class differences are effectively expanded in the feature space, while the intra-class differences are reduced, thereby improving the classification performance of the model.

[0087] S303, promoting the classifier network to optimize the feature space through the group proxy loss;

[0088] S304, updating the strip surface defect classification model;

[0089] In this embodiment, a continuous learning algorithm PDGAS based on prototype growth and population agents is used to learn new categories of data and expand the capabilities of the model. This solves the problems of the model being unable to autonomously discover and label new categories in industrial applications, the traditional industrial model being unable to grow continuously, and the existing continuous learning algorithms being unable to cope with actual industrial streaming data scenarios. Ultimately, a semi-autonomous lifelong growing strip surface defect classification system solution is obtained.

[0090] Example 2 Online continuous learning of strip surface defects

[0091] Based on the strip steel surface defect image data generated on the actual industrial production line, a strip steel surface defect dataset was constructed, and on this basis, an experiment close to the real strip steel industrial production scenario was designed.

[0092] The dataset is based on the surface defect image data of strip steel generated on a real industrial production line. There are 10 categories and 7,700 data samples in total, each of which is a 780×128 grayscale image. The ratio of training data to test data is 4:1. There is a large difference in the number of samples between different defect categories in the dataset. The category with the largest number is more than twice the number of the category with the smallest number. The number of samples between categories is in a very unbalanced state. This is a characteristic of data in real scenarios and also poses a great challenge to continuous learning algorithms.

[0093] Referring to the conventional design of experiments in the field of continuous learning, the strip surface defect dataset is randomly divided into five learning tasks by category. In order to simulate the actual industrial scenario, the dataset is processed into streaming data, and the algorithm is used to perform online continuous learning. Representative methods in the field of continuous learning are selected for comparative experiments. The experimental results are shown in Table 1:

[0094] Table 1 Experimental results of each algorithm on the strip surface defect dataset

[0095]

[0096] It can be concluded from Table 1 that the algorithm proposed in the present invention achieves the best final average accuracy performance on the real strip surface defect data set, while taking less time.

[0097] Example 3: Discovery of new categories of strip surface defect classification scenarios

[0098] In order to verify the effectiveness and operation effect of the new class discovery strategy based on feature comparison proposed in the present invention in the strip surface defect classification scenario, an experimental scenario for new class discovery was constructed; the continuous learning algorithm PDGAS proposed in the present invention was used to learn half of the categories of data on the strip surface defect dataset, and then tested on a test set containing all categories, and the number of samples correctly classified by the model (including correctly judged as known categories and correctly judged as new categories) divided by the total number of samples in the test set was used as the accuracy index to evaluate the new class discovery strategy; considering that the threshold γ in the new class discovery strategy is a parameter that directly affects the recognition accuracy, it is necessary to test the performance results of the strategy under different thresholds γ, and the experimental results are shown in Table 2:

[0099] Table 2 Experimental results of new class discovery strategy on strip surface defect classification scenario

[0100]

[0101]

[0102] From the analysis of Table 2, we can conclude that when the threshold value is reasonable, the new class discovery strategy can achieve good performance; the threshold γ is an important factor affecting the performance of the strategy: when the threshold γ is too high, even data in the learned category may be mistakenly judged as a new category; and when the threshold is set too low, data belonging to the new category may be mistakenly classified into the known category, resulting in poor performance of the strategy. Therefore, choosing an appropriate threshold γ is the key to fully exerting the capabilities of the strategy.

[0103] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A continuous learning method for classifying surface defects of steel strips in a human-in-loop system, characterized in that: The following steps are involved: S1. Obtaining the flow data of surface defects of strip steel on industrial production lines; S2. Processing the strip surface defect streaming data using a streaming data processing method based on sampling and dynamic buffering, including a reservoir sampling method, a memory buffer dynamic allocation method, and a random retrieval strategy; S201, using a new class discovery strategy based on feature comparison to determine the newly acquired streaming data x j ,j≥1 whether it belongs to a category that has appeared; By computing the input streaming data x j The cosine similarity Ψ(x j ), If Ψ(x j )>γ, then the strip surface defect classification model determines the input loss data x j If it belongs to the category that has appeared, execute step S202, where γ represents a preset threshold; If Ψ(x j )<γ, then the strip surface defect classification model determines the input loss data x j If it belongs to a category that has not appeared, execute step S203 and report to the data identification platform; S202, creating a new category buffer and adjusting the size of each category buffer by using a memory buffer dynamic allocation method; S203, design a new type of identification strategy for people in the loop, using the reservoir sampling method to determine whether to use the newly acquired streaming data x j Add to the corresponding category memory buffer; S204, using a random retrieval strategy to retrieve an approximately independent and identically distributed playback training batch from a balanced memory buffer, and inputting the batch into a strip surface defect classification model; S3, using the continuous learning algorithm PDGAS to optimize the strip surface defect classification model; S301, dynamically calculate, update and store category prototypes using prototype storage and growth strategies based on feature information of streaming data batches; S302, establishing an attraction set and a repulsion set for the streaming data for continuous learning through a group agent strategy; S303, promoting the classifier network to optimize the feature space through the group proxy loss; S304. Update the strip surface defect classification model.

2. The method for continuous learning of surface defect classification of steel strips in a human-in-loop loop according to claim 1, characterized in that: The cosine similarity Ψ(x j ) is expressed as: Among them, x j Represents the input streaming data, p i represents the category prototype, f θ (x j ) represents the projection vector of the input streaming data in the feature space.

3. The method for continuous learning of surface defect classification of steel strips in a human-in-loop loop according to claim 1, characterized in that: The reservoir sampling method constructs a reservoir of size k. If the strip surface defect flow data x j If the position j≤k in the data stream, the j strip surface defect streaming data directly enters the water reservoir; if the strip surface defect streaming data x j At position j>k in the data stream, a random number i∈[1,j] is generated. When i≤k, the strip surface defect stream data x j Directly enter the reservoir and replace the original i-th data. When i≤k is not satisfied, discard the data x j .

4. The method for continuous learning of surface defect classification of steel strips in a human-in-loop loop according to claim 1, characterized in that: The memory buffer dynamic allocation method includes: if the newly acquired data x j If the data x is a category that has already appeared, the reservoir sampling method is used to determine whether to store it in the memory buffer of the corresponding category; if the newly acquired data x j If it belongs to a category that has not appeared before, the memory buffer size allocated to each category is recalculated, the memory buffer size of all categories that have appeared is adjusted, samples are deleted according to the current number of cached samples, and a new category memory buffer is created. The reservoir sampling method is used to sample the newly acquired data x j into the new category memory buffer; The memory buffer size expression allocated to each category is: Among them, m represents the memory buffer size allocated to each category, K represents the total strip surface defect streaming data currently acquired, and t represents the number of categories that have appeared currently.

5. The method for continuous learning of surface defect classification of steel strips in a human-in-loop loop according to claim 1, characterized in that: The human-in-the-loop new category identification strategy is based on the professional knowledge and experience of human experts. The correct category labels are marked for data that do not belong to the existing categories through a visual interface, and the marked data are transmitted to the strip surface defect classification model through a data transmission interface for continuous learning.

6. The method for continuous learning of surface defect classification of steel strips in a human-in-loop loop according to claim 1, characterized in that: The prototype storage and growth strategy stores the newly acquired streaming data batches to the closest category prototype p in the dynamically changing streaming process. c , the expression is: Among them, μ∈[0,1] represents the parameter that controls the update amplitude, Represents the feature mean of samples of category c in the current training streaming data batch in the feature space, expressed as: Among them, B c ={(x i ,y i =c)∈B} represents all sample streaming data contained in category c in the current training streaming data batch B, |B c | represents the number of sample streaming data contained in category c in the current training streaming data batch B, represents the sample streaming data contained in category c, Represents sample streaming data The corresponding feature vector.

7. The method for continuous learning of surface defect classification of steel strips in a human-in-loop loop according to claim 1, characterized in that: Select the same streaming data as the sample streaming data in the training streaming data batch B. Other sample streaming data of the same category The sample streaming data Projected mean on feature space as a proxy for population The group agent and the category prototype p c Composition attraction set The expression is: The sample streaming data Projection on feature space and the ones belonging to category c are the exclusion set of other categories in the training streaming data batch B The expression is: in, Indicates sample streaming data of other categories.

Images