Apparatus for updating learning data for continuous learning and method therefor - Patent Application 20070122997

Abstract

To provide a device for updating training data for continual learning and a method therefor.SOLUTION: A method for updating training data includes the steps of: training a current training model using a buffer training data set previously stored in a buffer and a current training data set; calculating a degree of change from models trained in at least two previous training rounds of the current model in continual learning to the current training model; determining whether the degree of change is greater than or equal to a reference value; determining an update to the buffer training data set when the degree of change is greater than or equal to the reference value; and updating the buffer training data set by extracting at least some pieces of training data in the current training data set.SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] The present invention relates to defect diagnosis techniques, and more particularly to an apparatus and method for diagnosing defects using a continuous learning-based learning model.

[0002] The present invention relates to a training data selection technique, and more particularly to an apparatus for selecting training data for continuous training and a method therefor.

[0003] The present invention relates to a training data update technique, and more particularly to an apparatus for updating training data for continuous training and a method therefor. [Background technology]

[0004] Radiographic testing (RT) is an inspection method that detects defects inside a test object by passing radiation such as X-rays or gamma rays through the test object, depending on the conditions of use and the application, and forming an X-ray film on the test object. It is currently the most widely used non-destructive testing method for detecting internal defects.

[0005] The main task of defect diagnosis based on AI-based industrial radiography (RT) is object detection, and the accuracy of object detection must be maintained over time. When AI models are operated, transfer learning is typically used as a way to re-learn the features of new data sets that are generated over time, but this can lead to forgetting, which reduces accuracy for previous data. If all previous training data sets are retained to prevent this, storage space issues arise. Summary of the Invention [Problem to be solved by the invention]

[0006] SUMMARY OF THE INVENTION It is an object of the present invention to provide an apparatus and method for diagnosing defects using a continuous learning based learning model.

[0007] SUMMARY OF THE INVENTION It is an object of the present invention to provide an apparatus for selecting training data for continuous training and a method therefor.

[0008] SUMMARY OF THE INVENTION It is an object of the present invention to provide an apparatus for updating training data for continuous training and a method therefor. [Means for solving the problem]

[0009] To achieve the above object, a method for diagnosing defects according to a preferred embodiment of the present invention includes a step in which a learning unit trains a current learning model using a buffer learning data set and a current learning data set pre-stored in a buffer, and a step in which a detection unit detects defects in a radiological image using the current learning model when the radiological image is input.

[0010] The step of training the current training model includes a step in which the training unit loads a buffer training dataset, which is training data selected from past training datasets in continuous training based on the degree of influence it has on the predictive performance of the current training model, and a step in which the current training model is trained using the buffer training dataset and the current training dataset.

[0011] The method further includes, after the step of training the current learning model, a step in which a switching unit calculates the degree of change from multiple past learning models in learning rounds of continuous learning that are two or more learning rounds before the current learning model to the current learning model; a step in which the switching unit determines whether the degree of change is greater than or equal to a reference value; a step in which the switching unit determines an update to the buffer learning dataset if the degree of change is greater than or equal to the reference value; a step in which an update unit calculates the degree of influence of each of multiple learning data in the current learning dataset on the predictive performance of the future learning model when training a future learning model to be trained in the learning round next to the current learning model in continuous learning using the current learning dataset; and a step in which the update unit selects a portion of the multiple learning data in accordance with the order of the degree of influence and updates the buffer learning dataset.

[0012] The step of calculating the influence degree includes a step in which the update unit calculates a plasticity score indicating the probability that the predicted value of the current learning model and the predicted value of the future learning model will differ for the learning data of the current learning data set; a step in which the update unit calculates a stability score indicating the probability that the predicted value of the current learning model and the predicted value of the first past learning model learned in the first past learning round, which is the learning round before the current learning model, will differ for the learning data of the current learning data set; and a step in which the update unit calculates a weighted average of the plasticity score and the stability score as the influence degree.

[0013] The step of calculating the stability score includes: Formula

number

number

number

number

number

[0014] The step of calculating the plasticity score includes: Formula

number

number

number

number

number

[0015] The step of calculating the plasticity score includes: Formula

number

number

[0016] The step of calculating the influence degree includes: Formula

number

[0017] The step of calculating the degree of change is characterized in that the degree of change is derived based on the similarity between a first gradient vector indicating the change between the weight vector of the current learning model and the weight vector of a first past learning model learned in the learning round prior to the current learning model, and a second gradient vector indicating the change between the weight vector of the first past learning model and the weight vector of a second past learning model learned in the learning round prior to the first past learning model.

[0018] The similarity is Formula

number

[0019] To achieve the above object, an apparatus for diagnosing defects according to a preferred embodiment of the present invention includes a learning unit that trains a current learning model using a buffer learning data set and a current learning data set pre-stored in a buffer, and a detection unit that detects defects in a radiological image using the current learning model when a radiological image is input.

[0020] The learning unit is characterized in that, in continuous learning, it loads a buffer learning dataset, which is learning data selected from past learning datasets according to the degree of influence it has on the predictive performance of the current learning model, and trains the current learning model using the buffer learning dataset and the current learning dataset.

[0021] The device further includes a switching unit that calculates the degree of change from multiple past learning models in learning rounds two or more before the current learning model in consecutive learning rounds, determines whether the degree of change is equal to or greater than a reference value, and decides to update the buffer learning dataset if the degree of change is equal to or greater than the reference value; and an update unit that, when a future learning model to be learned in the learning round next to the current learning model in consecutive learning is trained using the current learning dataset, calculates the degree of influence that each of multiple learning data in the current learning dataset has on the predictive performance of the future learning model, and selects some of the multiple learning data in order of the degree of influence to update the buffer learning dataset.

[0022] The update unit calculates a plasticity score indicating the probability that the predicted value of the current learning model will differ from the predicted value of the future learning model for the learning data of the current learning data set, calculates a stability score indicating the probability that the predicted value of the current learning model will differ from the predicted value of the first past learning model learned in the first past learning round, which is the learning round before the current learning model, for the learning data of the current learning data set, and calculates a weighted average of the plasticity score and the stability score as the influence.

[0023] The update unit Formula

number

number

number

number

number

[0024] The update unit Formula

number

number

number

number

number

[0025] The update unit Formula

number

number

[0026] The update unit Formula

number

[0027] The switching unit is characterized in that it derives the degree of change according to the similarity between a first gradient vector indicating the change between the weight vector of the current learning model and the weight vector of a first past learning model learned in the learning round prior to the current learning model, and a second gradient vector indicating the change between the weight vector of the first past learning model and the weight vector of a second past learning model learned in the learning round prior to the first past learning model.

[0028] The switching unit Formula

number

[0029] To achieve the above object, a method for selecting training data according to a preferred embodiment of the present invention includes a step in which a training unit trains a current training model using a buffer training dataset and a current training dataset pre-stored in a buffer; a step in which an update unit, when training a future training model to be trained in the next training round of the current training model using the current training dataset, calculates the degree of influence of each of multiple training data in the current training dataset on the predictive performance of the future training model; and a step in which the update unit selects a portion of the multiple training data in accordance with the order of the degree of influence and updates the buffer training dataset.

[0030] The step of calculating the influence degree includes a step in which the update unit calculates a plasticity score indicating the probability that the predicted value of the current learning model and the predicted value of the future learning model will differ for the learning data of the current learning data set; a step in which the update unit calculates a stability score indicating the probability that the predicted value of the current learning model and the predicted value of the first past learning model learned in the first past learning round, which is the learning round before the current learning model, will differ for the learning data of the current learning data set; and a step in which the update unit calculates a weighted average of the plasticity score and the stability score as the influence degree.

[0031] The step of calculating the stability score includes the update unit: Formula

number

number

number

number

number

[0032] The step of calculating the plasticity score includes the steps of: Formula

number

number

number

number

number

[0033] The method further includes, before the step of calculating the influence, a step in which the learning unit constructs derived training data, which is training data derived from the current training dataset, and a step in which the learning unit trains a gradient prediction model using the derived training data.

[0034] The step of calculating the plasticity score is characterized in that the switching unit calculates the plasticity score using the gradient prediction model.

[0035] The derived learning data includes a weight vector of a target layer of a current learning model that is modified each time the learning unit divides the current learning data set into epoch units and iteratively learns the current learning model, and a target gradient vector that indicates a difference between a weight vector from a previous epoch and a weight vector from a current epoch that is modified each time the learning unit divides the current learning data set into epoch units and iteratively learns the current learning model.

[0036] The step of calculating the plasticity score includes the steps of: Formula

number

number

[0037] The step of calculating the degree of influence is performed by the update unit. Formula

number

[0038] To achieve the above object, an apparatus for selecting training data according to a preferred embodiment of the present invention includes a training unit that trains a current training model using a buffer training dataset and a current training dataset pre-stored in a buffer, and an update unit that, when training a future training model to be trained in the next training round of the current training model using the current training dataset in continuous training, calculates the degree of influence of each of multiple training data in the current training dataset on the predictive performance of the future training model, and selects a portion of the multiple training data in order of the degree of influence to update the buffer training dataset.

[0039] The update unit calculates a plasticity score indicating the probability that the predicted value of the current learning model will differ from the predicted value of the future learning model for the learning data of the current learning data set, calculates a stability score indicating the probability that the predicted value of the current learning model will differ from the predicted value of the first past learning model learned in the first past learning round, which is the learning round before the current learning model, for the learning data of the current learning data set, and calculates a weighted average of the plasticity score and the stability score as the influence.

[0040] The update unit is

number

number

number

number

number

[0041] The update unit Formula

number

number

number

number

number

[0042] The learning unit generates derived learning data, which is learning data derived from the current learning data set, and trains a gradient prediction model using the generated derived learning data, and the switching unit calculates the plasticity score using the gradient prediction model.

[0043] The derived learning data includes a weight vector of a target layer of a current learning model that is modified each time the learning unit divides the current learning data set into epoch units and iteratively learns the current learning model, and a target gradient vector that indicates a difference between a weight vector from a previous epoch and a weight vector from a current epoch that is modified each time the learning unit divides the current learning data set into epoch units and iteratively learns the current learning model.

[0044] The update unit Formula

number

number

[0045] The update unit Formula

number

[0046] To achieve the above object, a method for updating learning data according to a preferred embodiment of the present invention includes a step in which a learning unit learns a current learning model using a buffer learning dataset and a current learning dataset pre-stored in a buffer; a step in which a switching unit calculates a degree of change from a past learning model in a learning round two or more learning rounds prior to the current learning model in successive learning rounds; a step in which the switching unit determines whether the degree of change is greater than or equal to a reference value; a step in which the switching unit determines an update to the buffer learning dataset if the degree of change is greater than or equal to the reference value; and a step in which an update unit extracts at least a portion of learning data from the current learning dataset and updates the buffer learning dataset.

[0047] The step of calculating the degree of change is characterized in that the switching unit derives the degree of change based on the similarity between a first gradient vector indicating the change between the weight vector of the current learning model and the weight vector of a first past learning model learned in the learning round prior to the current learning model, and a second gradient vector indicating the change between the weight vector of the first past learning model and the weight vector of a second past learning model learned in the learning round prior to the first past learning model.

[0048] The similarity is Formula

number

[0049] The step of determining whether the degree of change is equal to or greater than a reference value is characterized in that the switching unit determines that the degree of change is equal to or greater than a reference value if the similarity between the first gradient vector and the second gradient vector is less than a reference value.

[0050] The method further includes, after the step of determining whether the degree of change is equal to or greater than a reference value, a step of the switching unit erasing the current training data set if the degree of change is less than the reference value.

[0051] The buffer training data set pre-stored in the buffer is training data selected from past training data sets in successive training according to the degree of influence on the prediction performance of the current training model.

[0052] To achieve the above object, an apparatus for updating learning data according to a preferred embodiment of the present invention includes: a learning unit that trains a current learning model using a buffer learning dataset and a current learning dataset pre-stored in a buffer; a switching unit that calculates a degree of change from a past learning model in a learning cycle two or more times before the current learning model in a learning cycle of consecutive learning, determines whether the degree of change is greater than or equal to a reference value, and if the degree of change is greater than or equal to the reference value, decides to update the buffer learning dataset; and an update unit that extracts at least a portion of learning data from the current learning dataset and updates the buffer learning dataset.

[0053] The switching unit is characterized in that it derives the degree of change according to the similarity between a first gradient vector indicating the change between the weight vector of the current learning model and the weight vector of a first past learning model learned in the learning round prior to the current learning model, and a second gradient vector indicating the change between the weight vector of the first past learning model and the weight vector of a second past learning model learned in the learning round prior to the first past learning model.

[0054] The switching unit Formula

number

[0055] The switching unit determines that the degree of change is equal to or greater than a reference value if the similarity between the first gradient vector and the second gradient vector is less than a reference value.

[0056] The switching unit may delete the current training data set if the degree of change is less than a reference value.

[0057] The buffer training data set pre-stored in the buffer is training data selected from past training data sets in successive training according to the degree of influence on the prediction performance of the current training model. [Effects of the Invention]

[0058] According to the present invention, by updating the buffer learning data set according to the predictive performance of the future learning model, and then continuously learning the learning model using the new learning data set together with the buffer learning data set, it is possible to overcome catastrophic forgetting while minimizing storage space, and continuously provide a learning model that provides high accuracy at minimal cost. [Brief explanation of the drawings]

[0059] [Figure 1] FIG. 1 is a diagram illustrating the configuration of a system for diagnosing defects using a continuous learning-based learning model according to an embodiment of the present invention. [Figure 2] FIG. 1 is a diagram illustrating the configuration of an apparatus for diagnosing defects using a continuous learning-based learning model according to an embodiment of the present invention. [Figure 3] 10A and 10B are diagrams illustrating a learning data set and a learning model according to a learning round in continuous learning according to an embodiment of the present invention. [Figure 4] 1 is a flowchart illustrating a method for selecting training data for continuous training according to an embodiment of the present invention. [Figure 5] 1 is a flowchart illustrating a method for training a gradient prediction model according to an embodiment of the present invention. [Figure 6] 1 is a flowchart illustrating a method for calculating the degree of influence of learning data on the prediction performance of a future learning model according to an embodiment of the present invention. [Figure 7] 1 is a flowchart illustrating a method for diagnosing defects using a continuous learning-based learning model according to an embodiment of the present invention. [Figure 8] 10A-10C are example screen shots illustrating a method for diagnosing defects using a continuous learning-based learning model according to an embodiment of the present invention. [Figure 9] FIG. 1 illustrates a computing device according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0060] Although the present invention can be implemented in various forms by adding various modifications, specific embodiments will be illustrated and described in detail in the detailed description, but it should be understood that this is not intended to limit the present invention to the specific embodiments, but rather to include all modifications, equivalents, or alternatives that fall within the spirit and technical scope of the present invention.

[0061] The terms used in the present invention are merely used to describe specific embodiments and are not intended to limit the present invention. A singular expression includes a plural expression unless the context clearly indicates otherwise. It should be understood that, in the present invention, terms such as "comprise" or "have" are intended to specify the presence of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification, and do not preclude the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

[0062] In particular, the terms and phrases used in the following specification and claims should not be interpreted as being limited to their ordinary or dictionary meanings, but should be interpreted as meanings and concepts that correspond to the technical idea of ​​the present invention, in accordance with the principle that the inventor can appropriately define the concepts of terms to best describe his / her invention. In particular, in the embodiments of the present invention, estimation or inference means deriving a result of calculation based on what a learning model (LM: Machine Learning Model / Deep Learning Model) has learned.

[0063] First, a system for diagnosing defects using a continuous learning-based learning model according to an embodiment of the present invention will be described. Fig. 1 is a diagram illustrating the configuration of a system for diagnosing defects using a continuous learning-based learning model according to an embodiment of the present invention. Fig. 2 is a diagram illustrating the configuration of an apparatus for diagnosing defects using a continuous learning-based learning model according to an embodiment of the present invention. Fig. 3 is a diagram illustrating a learning data set and a learning model according to a learning round in continuous learning according to an embodiment of the present invention.

[0064] Referring to FIG. 1, a system for diagnosing defects using a continuous learning-based learning model according to an embodiment of the present invention includes a radiography apparatus RTA, a scanner SC, an inspection apparatus 10, and a storage apparatus 40.

[0065] The radiographic imaging apparatus RTA performs radiography on an object to which an image quality indicator (IQI) is attached, and outputs a radiographic inspection film in which the object to which the IQI is attached is photographed. The object may be a pipe, a tube, or the like.

[0066] The scanner SC scans the radiographic inspection film to generate a digitized radiation image. Once the digitized radiation image is generated, it is input to the inspection device 10. At this time, the radiation image may be input directly from the scanner SC or may be stored in another storage medium and then input from the storage medium.

[0067] The inspection device 10 trains a current training model using a buffer training data set and a current training data set previously stored in a buffer BF through continuous training, and detects defects in a radiological image using the current training model, which is the most recent training model among the training models trained through continuous training. The inspection device 10 can then output a report including a radiological image in which defects are detected using a bounding box BB.

[0068] The storage device 40 may be a cloud server or a database server. The buffer BF of the storage device 40 is for storing data, and the buffer BF may store a buffer training data set according to an embodiment of the present invention. Here, the buffer training data set is training data selected from past training data sets used in past training rounds, with the current set as the reference point in continuous training, based on the degree of influence the past training data set has on the prediction performance of the current training model. Meanwhile, although the buffer BF has been described as a storage medium included in the storage device 40, according to another embodiment of the present invention, the buffer BF may be a storage medium included in the inspection device 10.

[0069] Referring to FIG. 2, the inspection device 10 includes a collection unit 100, a learning unit 200, a switching unit 300, an update unit 400, and a detection unit 500.

[0070] The collection unit 100 continuously collects learning data for continuous learning and constructs a current learning data set. After constructing the current learning data set, the collection unit 100 provides the constructed current learning data set to the learning unit 200.

[0071] The learning unit 200 generates a learning model through learning (deep learning or machine learning). The learning model according to an embodiment of the present invention is for detecting an object learned from an image. The learning model according to an embodiment of the present invention detects an object, i.e., a defect, learned from a radiation image. In this case, the learning model can detect an area occupied by a defect in the radiation image using a bounding box BB. Examples of such learning models include CNN, YOLO, RCNN, and Faster RCNN. The learning model includes multiple interconnected layers (or modules), and the multiple layers (or modules) perform multiple operations. The multiple layers (or modules) are connected by weights (W). That is, the output of an operation result of any one layer (or module) is applied with a weight and input to the operation of the next layer. The learning model derives an output by performing multiple operations on input data, to which weights are applied between multiple layers (or modules). In other words, the learning model performs multiple operations connected by weights between multiple layers (or modules). Such multiple operations linked by weights between multiple layers (or modules) of a learning model are referred to as "weight operations."

[0072] In particular, the learning unit 200 according to the embodiment of the present invention learns a learning model through continuous learning.

[0073] 3, in an embodiment of the present invention, continuous learning refers to a method of continuously collecting new learning data, dividing the continuously collected learning data into learning datasets including a plurality of learning data sets, and continuously learning a learning model using the learning dataset for each learning session. In an embodiment of the present invention, the learning sessions may be divided by at least one of time, the number of collected learning data sets, and a specific event. Thus, in the present invention, learning is continuously performed for each learning session using a learning dataset including a plurality of learning data sets divided by the learning session, so that the learning model is continuously updated and a learning model different from the learning model derived in the previous learning session is continuously generated.

[0074] In an embodiment of the present invention, when the reference point is the current learning round t, the learning model learned in the current learning round t is referred to as the current learning model. Also, the learning dataset for learning the current learning model is referred to as the current learning dataset. The current learning dataset includes multiple pieces of learning data collected from the first previous learning round t-1, which is the learning round previous to the current learning round t, to the current learning round t.

[0075] Furthermore, the learning model learned in the first past learning round t-1, which is the learning round before the current learning model, will be referred to as the first past learning model. Furthermore, the learning dataset for learning the first past learning model will be referred to as the first past learning dataset. The first past learning dataset includes multiple learning data collected from the second past learning round t-2, which is the learning round before the first past learning round t-1, to the first past learning round t-1.

[0076] Then, with the current learning time t as the reference, the learning models of all previous learning times t-1, t-2, ... will be referred to as past learning models. Also, with the current learning time t as the reference, the learning data sets of all previous learning times t-1, t-2, ... will be referred to as past learning data sets.

[0077] On the other hand, the learning model to be learned in the future learning session t+1, which is the next learning session of the current learning model, will be referred to as the future learning model. The learning dataset for learning the future learning model will be referred to as the future learning dataset. The future learning dataset includes multiple learning data collected from the current learning session t to the future learning session t+1.

[0078] Referring again to FIG. 2, the switching unit 300 determines whether and when to update the buffer learning data set according to an embodiment of the present invention.

[0079] The switching unit 300 derives a degree of change from multiple past learning models in learning cycles two or more prior to the current learning model among the learning cycles of the continuous learning. To this end, the switching unit 300 derives a first gradient vector indicating the change between the weight vector of the current learning model and the weight vector of a first past learning model trained in the learning cycle prior to the current learning model. Next, the switching unit 300 derives a second gradient vector indicating the change between the weight vector of the first past learning model and the weight vector of a second past learning model trained in the learning cycle prior to the first past learning model. Thereafter, the switching unit 300 derives a degree of change according to the similarity between the first gradient vector and the second gradient vector. Next, the switching unit 300 determines whether the degree of change is equal to or greater than a reference value. That is, if the similarity between the first gradient vector and the second gradient vector is less than a reference value, the switching unit 300 determines that the degree of change is equal to or greater than the reference value, and if the similarity is equal to or greater than the reference value, the switching unit 300 determines that the degree of change is equal to or greater than the reference value. If the degree of change is equal to or greater than the reference value (the similarity is less than the reference value), the switching unit 300 determines to update the buffer training data set. On the other hand, if the degree of change is less than the reference value (the similarity is equal to or greater than the reference value), the switching unit 300 deletes the current training data set.

[0080] The update unit 400 calculates the degree of influence of each of the multiple learning data sets in the current learning data set on the predictive performance of the future learning model. That is, when a future learning model, which is the learning model for the next learning round of the current learning model in continuous learning, is trained using the current learning data set, the update unit 400 calculates the degree of influence of each of the multiple learning data sets in the current learning data set on the predictive performance of the future learning model. The update unit 400 then selects some of the multiple learning data sets in the current learning data set according to the calculated order of influence and updates the buffer learning data set. At this time, the update unit 400 updates the buffer learning data set by deleting buffer learning data previously stored in the buffer BF and storing learning data selected from the multiple learning data according to the order of influence in the buffer. The update unit 400 also erases the remaining learning data from the multiple learning data sets not selected according to the order of influence.

[0081] Next, a method for selecting training data for continuous training according to an embodiment of the present invention will be described. Figure 4 illustrates a method for selecting training data for continuous training according to an embodiment of the present invention.

[0082] Referring to FIG. 4, the collection unit 100 continuously collects training data to form a current training data set in step S110.

[0083] In step S120, the learning unit 200 loads a buffer training data set pre-stored in the buffer BF, and then generates a current training model by training using the buffer training data set pre-stored in the buffer BF and the current training data set. Here, the buffer training data set pre-stored in the buffer BF is training data selected from past training data sets used in past training rounds in successive training according to the degree of influence on the prediction performance of the current training model.

[0084] In step S130, the switching unit 300 derives the degree of change from a plurality of past learning models in learning rounds two or more prior to the current learning model in the continuous learning rounds. Step S130 will be described in more detail as follows.

[0085] First, the switching unit 300 derives a first gradient vector indicating a change between the weight vector of the current learning model and the weight vector of a first past learning model learned in a learning round prior to the current learning model.

[0086] Next, the switching unit 300 derives a second gradient vector indicating a change between the weight vector of the first past training model and the weight vector of the second past training model trained in a training round prior to the first past training model, and then derives a change rate according to the similarity between the first gradient vector and the second gradient vector.

[0087] At this time, the switching unit 300 can calculate the similarity using the following equation 1. [Number 1]

number

[0088] Here, CS indicates the similarity. t is the first gradient vector between the weight vector of the current training model and the weight vector of the first past training model. And g t-1 is the second gradient vector between the weight vector of the first past training model and the weight vector of the second past training model. T t-1 is the transpose of the second gradient vector.

[0089] Next, in step S140, the switching unit 300 determines whether the degree of change is equal to or greater than a reference value. At this time, if the similarity between the first gradient vector and the second gradient vector is less than the reference value according to Equation 1, the switching unit 300 determines that the degree of change is equal to or greater than the reference value. On the other hand, if the similarity between the first gradient vector and the second gradient vector is equal to or greater than the reference value according to Equation 1, the switching unit 300 determines that the degree of change is less than the reference value.

[0090] If the result of the determination in step S140 is that the degree of change is equal to or greater than the reference value (the degree of similarity is less than the reference value), the switching unit 300 proceeds to step S150 to delete the current learning data set, and repeats steps S110 to S140 described above.

[0091] On the other hand, if the result of the determination in step S140 is that the degree of change is less than the reference value (the similarity is equal to or greater than the reference value), it is decided to update the buffer learning data set, and the process proceeds to step S160.

[0092] In step S160, the training unit 200 trains a gradient prediction model (GPM). The gradient prediction model (GPM) collects derived training data from training data used to train a current training model, and predicts gradient vectors of a future training model using the collected derived training data. Training of the gradient prediction model (GPM) will be described in more detail below.

[0093] Next, in step S170, the update unit 400 calculates the degree of influence of each of the multiple learning data sets in the current learning data set on the predictive performance of the future learning model. Specifically, when a future learning model, which is the learning model for the next learning round after the current learning model in continuous learning, is trained using the current learning data set, the update unit 400 calculates the degree of influence of each of the multiple learning data sets in the current learning data set on the predictive performance of the future learning model. The procedure for calculating such influences will be described in more detail below.

[0094] Next, in step S180, the update unit 400 selects a portion of the multiple learning data sets in the current learning data set in order of influence, and updates the buffer learning data set. At this time, the update unit 400 updates the buffer learning data set by deleting the buffer learning data previously stored in the buffer BF and storing the learning data selected from the multiple learning data in order of influence in the buffer BF. Also, in step S180, the remaining learning data not selected from the multiple learning data in order of influence is deleted.

[0095] By updating the buffer learning data set in the manner described above and continuously learning the learning model using the new learning data set together with the buffer learning data set, it is possible to overcome catastrophic forgetting while minimizing storage space and continuously provide a learning model that provides high accuracy at minimal cost.

[0096] Next, a method for training a gradient prediction model (GPM) according to an embodiment of the present invention will be described. Fig. 5 is a flowchart for explaining a method for training a gradient prediction model according to an embodiment of the present invention. Specifically, Fig. 5 is a detailed description of step S160 in Fig. 4.

[0097] 5, in step S210, the learning unit 200 selects a target layer of the current learning model. For example, it is assumed that the target layer is a convolutional layer. In step S210, the learning unit 200 divides a plurality of learning data of the current learning data set into epoch units and learns them while training the current learning model in step S120. Each time the learning unit 200 divides the learning data into epoch units and learns them, it stores a modified weight vector of the target layer of the current learning model.

[0098] As a result, in step S220, the learning unit 200 generates a plurality of derived learning data sets that are learning data derived from the current learning data set.

[0099] The derived learning data includes a weight vector of the target layer of the current learning model that is modified each time the current learning data set is divided into epoch units and the current learning model is trained iteratively, and a target gradient vector that indicates the difference between the weight vector from the previous epoch and the weight vector from the current epoch each time the current learning data set is divided into epoch units and the current learning model is trained iteratively.

[0100] Each time a plurality of learning data is divided into epoch units and learning is performed repeatedly, the weight vector of the target layer to be corrected changes as shown in Equation 2 below. [Number 2]

number

[0101] Here, n is the epoch index, LR denotes the learning rate, and G(n) denotes the gradient vector. In the present invention, the product of the learning rate LR and the gradient vector G(n) is approximated as the gradient vector. As a result, the weight vector may be expressed as [W(1), W(2), W(3), ..., W(E-1)] (E is an arbitrary positive constant). The target gradient vector corresponding to [W(1), W(2), W(3), ..., W(E-1)] may be expressed as [W(2)-W(1), W(3)-W(2), W(4)-W(3), ..., W(E)-W(E-1)].

[0102] Next, the learning unit 200 inputs the weight vectors of the derived learning data to a gradient prediction model (GPM) having untrained weights in step S230. Then, the gradient prediction model (GPM) performs a weight calculation in step S240, applying the untrained weights to the learning radiation image, and derives a weight difference vector that predicts the difference between the weight vector of the target layer in the current epoch and the weight vector in the previous epoch.

[0103] Next, in step S250, the learning unit 200 calculates a loss indicating the difference between the target gradient vector and the weighted difference vector using a loss function. Then, in step S260, the learning unit 200 performs optimization to correct the weights of the gradient prediction model (GPM) so that the loss derived by the loss function is minimized.

[0104] Thereafter, in step S270, the learning unit 200 determines whether a learning completion condition is satisfied. According to one embodiment, the learning completion condition may be when the loss calculated earlier (S250) converges and becomes equal to or less than a preset target value. If the learning completion condition is not satisfied as a result of the determination in step S270, the above-described steps S230 to S270 are repeated using multiple different training data. On the other hand, if the learning completion condition is satisfied as a result of the determination in step S270, the learning unit 200 completes learning of the gradient prediction model (GPM) in step S280.

[0105] Next, a method for calculating the degree of influence of each of multiple learning data sets in a current learning data set on the predictive performance of a future learning model according to an embodiment of the present invention will be described. Figure 6 is a flowchart for explaining a method for calculating the degree of influence of learning data sets on the predictive performance of a future learning model according to an embodiment of the present invention. That is, Figure 6 is a detailed description of step S170 in Figure 4.

[0106] In step S310, the update unit 400 calculates a plasticity score indicating the probability that the predicted value of the current learning model differs from the predicted value of the future learning model for the learning data of the current learning data set.

[0107] At this time, the update unit 400 can calculate the plasticity score according to the following Equation 3. [Number 3]

number

[0108] where:

number

number

number

number

[0109] On the other hand, since the future learning model has not yet been generated, the predicted value of the future learning model is estimated using the predicted gradient of the future learning model using a gradient prediction model (GP). As a result, the predicted value of the future learning model in Equation 3 is estimated by the following Equation 4. [Number 4]

number

[0110] where:

number

number

[0111] Next, in step S320, the update unit 400 calculates a stability score indicating the probability that the predicted value of the current learning model differs from the predicted value of the first past learning model learned in the first past learning round, which is the learning round before the current learning model, for the learning data of the current learning data set.

[0112] At this time, the update unit 400 can calculate the stability score using the following Equation 5. [Number 5]

number

[0113] where:

number

[0114] n is the index of the training data in the current training data set, and t is the index of the training round in the continuous training, where t indicates the current training round and t-1 indicates the first previous training round.

number

number

number

[0115] Next, in step S330, the update unit 400 calculates the weighted average of the plasticity score and the stability score as the influence.

[0116] At this time, the update unit 400 can calculate the influence degree by the following Equation 6. [Number 6]

number

[0117] Here, S indicates the influence, i is the index of the training data in the current training dataset, λ is a preset weight, Plasticity means the plasticity score, and Stability means the stability score.

[0118] Next, a method for diagnosing defects using a continuous learning-based learning model according to an embodiment of the present invention will be described. Figure 7 is a flowchart for explaining a method for diagnosing defects using a continuous learning-based learning model according to an embodiment of the present invention. Figure 8 is an example of a screen for explaining a method for diagnosing defects using a continuous learning-based learning model according to an embodiment of the present invention.

[0119] In FIG. 7, the training model is the current training model trained using the previously stored buffer training data set and the current training data set as described above (S120).

[0120] 7, in step S410, the radiographic apparatus RTA performs radiography on an object having an image quality indicator (IQI) attached thereto, and outputs a radiographic inspection film capturing the image of the object having the IQI attached thereto. Here, the object may be a pipe, a tube, or the like.

[0121] Then, in step S420, the scanner SC scans the radiographic inspection film to generate a digitized radiographic image.

[0122] Once the digitized radiation image is generated in this manner, the generated radiation image is input to the inspection device 10. At this time, the radiation image may be input directly from the scanner SC, or may be stored in another storage medium and then input from the storage medium.

[0123] In step S430, the detection unit 500 of the inspection apparatus 10 receives a radiation image. Then, in step S440, the detection unit 500 detects defects in the radiation image using the latest model, i.e., the current learning model, among the learning models trained by continuous learning, as described above. At this time, the current learning model can detect the area occupied by the defect in the radiation image using a bounding box BB, as shown in FIG. 8.

[0124] Next, in step S450, the detection unit 500 outputs a report including the radiation image in which defects are detected using the bounding box BB.

[0125] 9 is a diagram illustrating a computing device according to an embodiment of the present invention. The computing device TN100 of FIG. 9 may be a device described herein (e.g., inspection device 10, storage device 40, etc.).

[0126] 9, computing device TN100 may include at least one processor TN110, a transceiver TN120, and a memory TN130. Computing device TN100 may further include a storage device TN140, an input interface device TN150, an output interface device TN160, etc. The components included in computing device TN100 may be connected by a bus TN170 to communicate with each other.

[0127] Processor TN110 can execute program commands stored in at least one of memory TN130 and storage device TN140. Processor TN110 can refer to a central processing unit (CPU), a graphics processing unit (GPU), or a dedicated processor on which methods according to embodiments of the present invention are performed. Processor TN110 can be configured to implement procedures, functions, methods, etc. described in connection with embodiments of the present invention. Processor TN110 can control each component of computing device TN100.

[0128] Memory TN130 and storage device TN140 can each store various information related to the operation of processor TN110. Memory TN130 and storage device TN140 can each be configured with at least one of a volatile storage medium and a non-volatile storage medium. For example, memory TN130 may be configured with at least one of a read-only memory (ROM) and a random access memory (RAM).

[0129] The transceiver TN120 can transmit or receive wired or wireless signals and can communicate with a network.

[0130] In particular, the collection unit 100, learning unit 200, switching unit 300, update unit 400, and detection unit 500 of the inspection device 10 according to the embodiment of the present invention may be realized in the form of a program readable by a computer means, stored in memory TN130, and executed by processor TN110. Alternatively, the collection unit 100, learning unit 200, switching unit 300, update unit 400, and detection unit 500 may be lower modules of processor TN110.

[0131] Furthermore, the buffer BF according to the embodiment of the present invention may be the memory TN130 or storage device TN140 of the inspection device 10 or storage device 40.

[0132] Meanwhile, the various methods according to the above-described embodiments of the present invention can be realized in the form of a program readable by various computer means and recorded on a computer-readable recording medium. Here, the recording medium may include program instructions, data files, data structures, and the like, alone or in combination. The program instructions recorded on the recording medium may be specially designed and constructed for the present invention, or may be known and available to those skilled in the art of computer software. For example, recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include not only machine language, such as that produced by a compiler, but also high-level language codes executable by a computer using an interpreter, etc. Such hardware devices can be configured to operate as one or more software modules to perform the operations of the present invention, or vice versa.

[0133] Although one embodiment of the present invention has been described above, a person having ordinary skill in the art may modify and change the present invention in various ways by adding, changing, deleting or adding components within the scope of the concept of the present invention as set forth in the claims, and this also falls within the scope of the present invention.

Claims

1. a learning unit learning a current learning model using a buffer learning data set and a current learning data set stored in advance in a buffer; a switching unit calculating a degree of change from a past learning model of a learning cycle two or more times before a current learning model among the learning cycles of continuous learning to the current learning model; a step of the switching unit determining whether the degree of change is equal to or greater than a reference value; the switching unit determining an update to the buffer learning data set if the degree of change is equal to or greater than a reference value; an update unit extracting at least a portion of training data from the current training data set and updating the buffer training data set; Including, A method for updating training data.

2. The step of calculating the degree of change includes: the switching unit derives the degree of change according to the similarity between a first gradient vector indicating a change between the weight vector of the current learning model and the weight vector of a first past learning model learned in a learning round prior to the current learning model, and a second gradient vector indicating a change between the weight vector of the first past learning model and the weight vector of a second past learning model learned in a learning round prior to the first past learning model; The method for updating training data according to claim 1 .

3. The similarity is Formula [Number 73] is calculated by The CS is a similarity, The above g ｔ is the first gradient vector between the weight vector of the current training model and the weight vector of the first past training model, The above g ｔ－１ is a second gradient vector between the weight vector of the first past training model and the weight vector of the second past training model, The above g Ｔ ｔ－１ is the transpose vector of the second gradient vector, The method for updating training data according to claim 2.

4. The step of determining whether the degree of change is equal to or greater than a reference value includes: the switching unit determines that the degree of change is equal to or greater than a reference value if the similarity between the first gradient vector and the second gradient vector is less than a reference value; The method for updating training data according to claim 2.

5. After the step of determining whether the degree of change is equal to or greater than a reference value, If the degree of change is less than a reference value, the switching unit deleting the current training data set; further comprising: A method for updating training data according to any one of claims 1 to 4.

6. The buffer training data set pre-stored in the buffer is In continuous learning, learning data is selected from past learning data sets according to the degree of influence on the prediction performance of the current learning model. A method for updating training data according to any one of claims 1 to 4.

7. a learning unit that learns a current learning model using a buffer learning data set pre-stored in a buffer and a current learning data set; calculating a degree of change from a past learning model of a learning cycle two or more times before the current learning model among the learning cycles of the continuous learning to the current learning model; After determining whether the degree of change is equal to or greater than a reference value, a switching unit that determines an update to the buffer learning data set if the degree of change is equal to or greater than a reference value; an update unit that extracts at least a portion of training data from the current training data set and updates the buffer training data set; Including, A device for updating learning data.

8. The switching unit The degree of change is derived according to the similarity between a first gradient vector indicating the change between the weight vector of the current learning model and the weight vector of a first past learning model learned in a learning round prior to the current learning model, and a second gradient vector indicating the change between the weight vector of the first past learning model and the weight vector of a second past learning model learned in a learning round prior to the first past learning model. The device for updating training data according to claim 7.

9. The switching unit Formula [Number 74] The similarity is calculated by The CS is a similarity, The above g ｔ is the first gradient vector between the weight vector of the current training model and the weight vector of the first past training model, The above g ｔ－１ is a second gradient vector between the weight vector of the first past training model and the weight vector of the second past training model, The above g Ｔ ｔ－１ is the transpose vector of the second gradient vector, The apparatus for updating training data according to claim 8.

10. The switching unit If the similarity between the first gradient vector and the second gradient vector is less than a reference value, it is determined that the degree of change is equal to or greater than a reference value. The apparatus for updating training data according to claim 8.

11. The switching unit If the degree of change is less than a reference value, the switching unit deletes the current learning data set. An apparatus for updating training data according to any one of claims 7 to 10.

12. The buffer training data set pre-stored in the buffer is In continuous learning, learning data is selected from past learning data sets according to the degree of influence on the prediction performance of the current learning model. An apparatus for updating training data according to any one of claims 7 to 10.

Images