Power transmission line defect category prediction method and storage medium
By adopting a combination of super feature extractor and linear classifier in transmission line defect detection, combined with a new classifier with auxiliary losses, the problem of data imbalance and catastrophic forgetting in the old and new categories is solved, the stability and plasticity of the model are achieved, and online updates and efficient optimization are supported.
Patent Information
- Application Number
- CN202510112990.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-24
- Publication Date
- 2025-05-13
AI Technical Summary
The existing technology has the problem of data imbalance between old and new categories in the detection of transmission line defects, which leads to the model being prone to catastrophic forgetting during incremental learning, wasted computing resources and time verbose.
Using a combination of super feature extractor and linear classifier, by extracting features on old and new categories and aggregating them into parameterized super features, combined with a new classifier with auxiliary loss, the old feature extractor is frozen to fix the intrinsic structure of previous data, reducing catastrophic forgetting.
The stability and plasticity of the transmission line defect category prediction model is improved, the problem of catastrophic forgetting in class incremental learning is effectively solved, and the online update and efficient optimization of the model are realized.
Smart Images
Figure CN119989201A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of defect detection, and in particular to a method and storage medium for predicting defect categories of power transmission lines. Background Art
[0002] With the continuous development of social economy, the investment and construction speed of power system is also accelerating. As far as the power system is concerned, transmission lines, as the main medium for transmitting electricity, are of great significance to the safe and reliable operation of the power grid. However, in the existing technology, when performing defect detection on small target equipment of transmission lines, it is impossible to obtain a complete set of samples at one time, which causes an imbalance of new and old category data. The model must be retrained, resulting in a waste of computing resources and a lengthy time.
[0003] Among the existing incremental update methods, regularization methods penalize changes in important weights of previously learned models, knowledge distillation retains network outputs for available data, and structure-based methods keep old parameters unchanged while assigning more parameters to new categories. However, all of these methods either sacrifice model plasticity for stability or are easily forgotten due to feature degradation of old categories, resulting in a large performance gap between models trained on all data and the previous state-of-the-art models. In addition, in the existing incremental update methods, new data features generated during the incremental learning process will have a feature impact on old data features, leading to the phenomenon of forgetting old tasks when learning new tasks; as the number of categories increases, network parameters will increase, resulting in a larger network parameter storage space, which not only increases the computational and storage costs, but also causes catastrophic forgetting in the detection network, reducing detection accuracy, and may also aggravate catastrophic forgetting, affecting model performance. Summary of the invention
[0004] The purpose of the present invention is to provide a method and storage medium for predicting the category of power transmission line defects. First, feature extraction is performed on a super feature extractor, and the extracted features are formed into parameterized super features and input into a linear classifier for prediction of the category of power transmission line defects. The present invention improves the stability and plasticity of the power transmission line defect category prediction model in the case of new samples, and effectively solves the problem of catastrophic forgetting in incremental learning. The present invention is implemented by the following technical solutions.
[0005] In a first aspect, the present invention provides a method for predicting a type of transmission line defect, comprising:
[0006] The acquired public dataset with transmission line defects is preprocessed for noise reduction and expansion;
[0007] The public data set with power transmission line defects after noise reduction and expansion preprocessing is input into a category-incremental target detection framework; wherein the category-incremental target detection framework includes a modular deep classification network; the modular deep classification network includes a serially connected super feature extractor and a linear classifier;
[0008] A super feature extractor is used to extract features on the new and old categories, and the extracted features are aggregated into parameterized super features; wherein the new and old categories include the power transmission line defect categories pre-identified by the category incremental target detection framework and the power transmission line defect categories newly introduced by the category incremental target detection framework in the current incremental learning stage;
[0009] The super features are input into the linear classifier to perform transmission line defect category prediction.
[0010] Optionally, the formation of the super feature extractor includes constructing a new feature extractor for each new task and integrating it with the feature extractor of the old task to form a super feature extractor.
[0011] Optionally, during the process of incremental learning and updating of the category-incremental target detection framework, the label space at time t-1 is defined as , ,
[0012] in, Represents the label space of the old task at time t-1. The label space is the set of all categories learned by the category incremental target detection framework in the incremental learning process until time t-1; represents the set of categories learned at the i-th time step; represents the union of the set of categories learned from the first time step i=1 to all time steps t-1; the entire formula It means that at time t-1, the label space of the category incremental target detection framework is composed of the union of all categories learned in previous time steps. This means that the model can recognize all the categories learned from the beginning to t-1 at time t-1.
[0013] Each category set There is a form ,in, Represents a collection of categories The input image of the kth class in, is the corresponding input label, where k is a positive integer and k ≥ 1.
[0014] Optionally, during the process of category incremental learning and updating, at time t, the category incremental target detection framework encounters a new task, and the label space changes from Updated to , ,
[0015] in, represents the union of the old task label space and the new task label space encountered by the category incremental target detection framework at time t, represents the label space of the new task introduced at time t.
[0016] Optionally, the extracting features from the new and old categories using a super feature extractor and aggregating the extracted features into parameterized super features includes:
[0017] At time t, when performing defect detection on a new task, for each new task label space , the category-incremental object detection framework creates a new feature extractor for this task ; For fast adaptation, the feature extractor The weight parameters are inherited from the old task feature extractor at time t-1 .
[0018] At time t, all old task label spaces The feature extractors of the old feature extractor set , the old feature extractor at time t and the new feature extractor at time t Super feature extractor at composition time t , where the old feature extractor at time t is , is a feature extractor with time step i;
[0019] For the input image x, the super feature extractor at time t The extracted super features are , super features Calculated by the following formula:
[0020] .
[0021] Optionally, the step of inputting the super feature into a linear classifier to perform transmission line defect category prediction includes:
[0022] To reduce catastrophic forgetting, freeze the old feature extractor at time t , the old feature extractor The intrinsic structure of the previous data is captured to freeze the parameters of the extractor and the related parameters of batch normalization before time t. Since these parameters are not updated, this can fix the previously learned category representation and prevent the learning of new features from forgetting old features.
[0023] The new features parameterized by the new feature extractor are used to enhance the new feature extractor so as to preserve the intrinsic structure of the old features in the subspace of the previously learned old task, and finally the structure is reused by the classifier, further effectively solving the problem of catastrophic forgetting in class incremental learning.
[0024] Utilizing a novel classifier with auxiliary loss Super Features For classification prediction, a new classifier with auxiliary loss An auxiliary classifier is introduced To regularize the category-incremental object detection framework, a new classifier with auxiliary loss Reusing the intrinsic structure of old task features effectively improves the ability of the category-incremental object detection framework to distinguish features and learn diverse new categories.
[0025] New Classifier with Auxiliary Loss Super Features Probability prediction is done using the following formula:
[0026] ,
[0027] A new classifier with auxiliary loss Super Features The predicted probability distribution of , x is the input image, y is the label of the input image x;
[0028] Optionally, the predicted probability is given by the following formula Perform the following calculations:
[0029] ,
[0030] in, , represents the predicted category output by the category increment target detection framework, is the predicted probability Get the value of the input image label y corresponding to the maximum value.
[0031] and Parameters inherited from old features , to retain old knowledge and randomly initialize its newly added parameters.
[0032] Optionally, in order to improve the model's ability to distinguish features and learn diverse new categories, the present invention introduces an auxiliary classifier To operate the new feature During training, the auxiliary classifier introduced The auxiliary classifier is used to constrain the modular deep classification network to focus on learning the features of the new task. The probability prediction of the features of the new task is performed using the following formula:
[0033] ,
[0034] In the formula, For auxiliary classifier The predicted probability distribution for the input image x, is the auxiliary classifier at time t The process of classifying the input image x first passes through the feature extractor Extract features and then pass them through auxiliary classifier Make category predictions.
[0035] The Softmax function will assist the classifier The output of is converted into a probability distribution, which is used to predict the category in the new task. The purpose of classifying new features separately is to enhance the recognition ability of new features while ensuring the recognition ability of old features.
[0036] Optionally, the novel classifier with auxiliary loss The scalable loss is calculated by the following formula:
[0037] ,
[0038] in, is the control auxiliary classifier The hyperparameters of the effect, is the auxiliary loss introduced, A new classifier with auxiliary loss of loss, For the new classifier with auxiliary loss Scalable loss. Since the sample set of equipment is unbalanced in terms of new and old category data and the types of defects are numerous and complex when performing defect detection, the introduction of auxiliary loss can improve the model's ability to distinguish features and learn diverse new categories. Scalable loss plays an important role in incremental learning. By balancing the learning of new and old tasks, it effectively solves the problem of catastrophic forgetting, improves the stability and adaptability of the model, and enables it to maintain good performance in a constantly changing task environment.
[0039] In a second aspect, the present invention provides a computer-readable storage medium having a computer program / instruction stored thereon, wherein when the computer program / instruction is executed by a processor, the steps of the method for predicting the defect category of a transmission line described in the first aspect are implemented.
[0040] Beneficial Effects
[0041] (1) When predicting the defect category of a transmission line, it is impossible to obtain a complete set of sample sets at one time, which leads to an imbalance in the new and old category data. The model must be retrained, which wastes computing resources and takes a long time. The online update technology of the transmission line defect category prediction model based on incremental learning is studied, which is conducive to online updating of the model, improving the efficiency of defect category prediction, and realizing continuous optimization and efficient updating of the model.
[0042] (2) Auxiliary loss is introduced to improve the model's ability to distinguish features and learn diverse new categories. Compared with the existing incremental model update method, this invention improves the stability and plasticity of the model when new samples are added, and effectively solves the problem of catastrophic forgetting in incremental learning. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is a schematic diagram of a flow chart of an online updating method of a category incremental target detection framework in one embodiment of the present invention;
[0044] Figure 2 Shown is a schematic diagram of an expandable incremental learning framework of the present invention;
[0045] Figure 3 Shown is a schematic diagram of the modular deep classification network structure of the present invention;
[0046] Figure 4 Shown is the experimental effect diagram of the present invention. DETAILED DESCRIPTION
[0047] The following is a further description in conjunction with the accompanying drawings and specific embodiments. In the description of the present invention, it should be understood that the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features.
[0048] Example 1
[0049] This embodiment introduces a method for predicting the type of transmission line defects. Figure 1 As shown, including the following:
[0050] The acquired public dataset with transmission line defects is preprocessed for noise reduction and expansion;
[0051] The public data set with power transmission line defects after noise reduction and expansion preprocessing is input into a category-incremental target detection framework; wherein the category-incremental target detection framework includes a modular deep classification network; the modular deep classification network includes a serially connected super feature extractor and a linear classifier;
[0052] A super feature extractor is used to extract features on the new and old categories, and the extracted features are aggregated into parameterized super features; wherein the new and old categories include the power transmission line defect categories pre-identified by the category incremental target detection framework and the power transmission line defect categories newly introduced by the category incremental target detection framework in the current incremental learning stage;
[0053] The super features are input into the linear classifier to perform transmission line defect category prediction.
[0054] In practical applications, the present invention constructs a new feature extractor for each new task and integrates it with the feature extractor of the old task to form a super feature extractor network to extract features from the new and old tasks, thereby overcoming the phenomenon of forgetting the old task when learning the new task. In addition, the present invention freezes the old feature extractor when predicting a new defect category, thereby fixing the internal structure of the previous data and making the old defect category data unaffected by the execution of the new task, thereby effectively reducing catastrophic forgetting. Scalable category incremental object detection framework such as Figure 2 As shown, the framework can adaptively learn and identify different categories from the continuously acquired new transmission line defect data, and realize the online update of the transmission line defect category prediction model.
[0055] Example 2
[0056] Based on Example 1, this example introduces a specific implementation process of a method for predicting the type of power transmission line defects, such as Figure 3 As shown, specifically including the following:
[0057] Figure 3 This is a schematic diagram of a modular deep classification network of the present invention, wherein the modular deep classification network includes a super feature extractor and a linear classifier connected in series; the following operations are performed in the super feature extractor and the linear classifier, respectively.
[0058] 1. Feature extraction stage
[0059] The formation of the super feature extractor includes constructing a new feature extractor for each new task and integrating it with the feature extractor of the old task to form a super feature extractor.
[0060] In the process of incremental learning and updating of the category incremental target detection framework, the label space at time t-1 is defined as , ,
[0061] in, Represents the label space of the old task at time t-1, and the label space category incremental target detection framework is the set of all categories learned in the incremental learning process until time t-1; represents the set of categories learned at the i-th time step; represents the union of the set of categories learned from the first time step i=1 to all time steps t-1; the entire formula It means that at time t-1, the label space of the category incremental target detection framework is composed of the union of all categories learned in previous time steps. This means that the model can recognize all the categories learned from the beginning to t-1 at time t-1.
[0062] Each category set There is a form ,in, Represents a collection of categories The input image of the kth class in, is the corresponding input label, where k is a positive integer and k ≥ 1; in practical applications, k is upper bounded to the number of all defect categories of the transmission line, reflecting the idea of online updating.
[0063] In the process of category incremental learning and updating, at time t, the category incremental target detection framework encounters a new task, and the label space changes from Updated to , ,
[0064] in, represents the union of the old task label space and the new task label space encountered by the category incremental target detection framework at time t, represents the label space of the new task introduced at time t.
[0065] The method of extracting features from the new and old categories using a super feature extractor and aggregating the extracted features into parameterized super features includes:
[0066] At time t, when performing defect detection on a new task, for each new task label space , the category-incremental object detection framework creates a new feature extractor for this task ; For fast adaptation, the feature extractor The weight parameters are inherited from the old task feature extractor at time t-1 .
[0067] At time t, all old task label spaces The feature extractors of the old feature extractor set , the old feature extractor at time t and the new feature extractor at time t Super feature extractor at composition time t , where the old feature extractor at time t is , is the feature extractor with time step i.
[0068] For the input image x, the super feature extractor at time t The extracted super features are , super features Calculated by the following formula:
[0069] .
[0070] The step of inputting the super feature into the linear classifier to perform transmission line defect category prediction includes:
[0071] To reduce catastrophic forgetting, freeze the old feature extractor at time t , the old feature extractor It captures the intrinsic structure of the previous data to freeze the old feature extractor before time t The parameters of and batch normalization related parameters. Since these parameters are not updated, this can make the previously learned category representation fixed and prevent learning new features from forgetting old features.
[0072] The new features parameterized by the new feature extractor are used to enhance the new feature extractor so as to preserve the intrinsic structure of the old features in the subspace of the previously learned old task, and finally the structure is reused by the classifier, further effectively solving the problem of catastrophic forgetting in class incremental learning.
[0073] 2. Classifier prediction stage
[0074] Utilizing a novel classifier with auxiliary loss Super Features For classification prediction, a new classifier with auxiliary loss An auxiliary classifier is introduced To regularize the category-incremental object detection framework, a new classifier with auxiliary loss Reusing the intrinsic structure of old task features effectively improves the ability of the category-incremental object detection framework to distinguish features and learn diverse new categories.
[0075] New Classifier with Auxiliary Loss Super Features Probability prediction is done using the following formula:
[0076] ,
[0077] A new classifier with auxiliary loss Super Features The predicted probability distribution of , x is the input image, y is the label of the input image x;
[0078] The predicted probability is given by the following formula Perform the following calculations:
[0079] ,
[0080] in, , represents the predicted category output by the category increment target detection framework, is the predicted probability Get the value of the input image label y corresponding to the maximum value.
[0081] and Parameters inherited from old features , to retain old knowledge and randomly initialize its newly added parameters.
[0082] In order to improve the model's ability to distinguish features and learn diverse new categories, this paper introduces an auxiliary classifier To operate the new feature During training, the auxiliary classifier introduced The auxiliary classifier is used to constrain the modular deep classification network to focus on learning the features of the new task. The probability prediction of the features of the new task is performed using the following formula:
[0083] ,
[0084] In the formula, For auxiliary classifier The predicted probability distribution for the input image x, is the auxiliary classifier at time t The process of classifying the input image x first passes through the feature extractor Extract features and then pass them through auxiliary classifier Make category predictions.
[0085] The Softmax function will assist the classifier The output of is converted into a probability distribution, which is used to predict the category in the new task. The purpose of classifying new features separately is to enhance the recognition ability of new features while ensuring the recognition ability of old features.
[0086] The new classifier with auxiliary loss The scalable loss is calculated by the following formula:
[0087] ,
[0088] in, is the control auxiliary classifier The hyperparameters of the effect, is the auxiliary loss introduced, A new classifier with auxiliary loss of loss, For the new classifier with auxiliary loss Scalable loss. Since the sample set has imbalanced data of old and new categories when the equipment performs defect detection, and the types of defects are numerous and complex, the introduction of auxiliary loss can improve the model's ability to distinguish features and learn diverse new categories.
[0089] Figure 4 It is one of the transmission line defect categories of the present invention. The method of the present invention can effectively integrate new and old features, adapt to the increase of categories, and maintain the recognition ability of previously learned categories.
[0090] Example 3
[0091] This embodiment introduces a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method for predicting the defect category of a power transmission line as introduced in Embodiment 1 or 2 are implemented.
[0092] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0093] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0094] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0095] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.
[0096] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.
Claims
1. A method for predicting the defect category of a transmission line, characterized in that: include: The acquired public dataset with transmission line defects is preprocessed for noise reduction and expansion; The public data set with power transmission line defects after noise reduction and expansion preprocessing is input into a category-incremental target detection framework; wherein the category-incremental target detection framework includes a modular deep classification network; the modular deep classification network includes a serially connected super feature extractor and a linear classifier; A super feature extractor is used to extract features on the new and old categories, and the extracted features are aggregated into parameterized super features; wherein the new and old categories include the power transmission line defect categories pre-identified by the category incremental target detection framework and the power transmission line defect categories newly introduced by the category incremental target detection framework in the current incremental learning stage; The super features are input into the linear classifier to perform transmission line defect category prediction.
2. The method for predicting the type of power transmission line defects according to claim 1, characterized in that: The formation of the super feature extractor includes constructing a new feature extractor for each new task and integrating it with the feature extractor of the old task to form a super feature extractor.
3. The method for predicting the defect category of a power transmission line according to claim 1, characterized in that: In the process of incremental learning and updating of the category incremental target detection framework, the label space at time t-1 is defined as , , in, Represents the label space of the old task at time t-1. The label space is the set of all categories learned by the category incremental target detection framework in the incremental learning process until time t-1; represents the set of categories learned at the i-th time step; represents the union of the set of categories learned from the first time step i=1 to all time steps t-1; Each category set There is a form ,in, Represents a collection of categories The input image of the kth class in, is the corresponding input label, where k is a positive integer and k ≥ 1.
4. The method for predicting the defect category of a power transmission line according to claim 3, characterized in that: In the process of category incremental learning and updating, at time t, the category incremental target detection framework encounters a new task, and the label space changes from Updated to , , in, represents the union of the old task label space and the new task label space encountered by the category incremental target detection framework at time t, represents the label space of the new task introduced at time t.
5. The method for predicting the type of power transmission line defects according to claim 4, characterized in that: The method of extracting features from the new and old categories using a super feature extractor and aggregating the extracted features into parameterized super features includes: At time t, when performing defect detection on a new task, for each new task label space , the category-incremental object detection framework creates a new feature extractor for this task ; At time t, all old task label spaces The feature extractors of the old feature extractor set , the old feature extractor at time t and the new feature extractor at time t Super feature extractor at composition time t , where the old feature extractor at time t is , is a feature extractor with time step i; For the input image x, the super feature extractor at time t The extracted super features are , super features Calculated by the following formula: 。 6. The method for predicting the type of power transmission line defects according to claim 5, characterized in that: The step of inputting the super feature into the linear classifier to perform transmission line defect category prediction includes: Utilizing a novel classifier with auxiliary loss Super Features For classification prediction, a new classifier with auxiliary loss Introduced an auxiliary classifier To regularize the category-incremental target detection framework, New Classifier with Auxiliary Loss Super Features Probability prediction is done using the following formula: ， A new classifier with auxiliary loss Super Features The predicted probability distribution of , x is the input image, and y is the label of the input image x.
7. The method for predicting the type of power transmission line defects according to claim 6, characterized in that: The predicted probability is given by the following formula Perform the following calculations: ， in, , represents the predicted category output by the category increment target detection framework, is the predicted probability Get the value of the input image label y corresponding to the maximum value.
8. The method for predicting the type of power transmission line defects according to claim 6, characterized in that: The auxiliary classifier introduced The auxiliary classifier is used to constrain the modular deep classification network to focus on learning the features of the new task. The probability prediction of the features of the new task is performed using the following formula: ， In the formula, For auxiliary classifier The predicted probability distribution for the input image x, is the auxiliary classifier at time t The process of classifying the input image x first passes through the feature extractor Extract features and then pass them through auxiliary classifier Make category predictions.
9. The method for predicting the type of power transmission line defects according to claim 1, characterized in that: The new classifier with auxiliary loss The scalable loss is calculated by the following formula: ， in, is the control auxiliary classifier The hyperparameters of the effect, is the auxiliary loss introduced, A new classifier with auxiliary loss of loss, A new classifier with auxiliary loss Scalable loss.
10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: include: Memory, for storing computer programs / instructions; A processor, configured to execute the computer program / instructions to implement the steps of the method for predicting the defect category of a power transmission line according to any one of claims 1 to 7.