A power grid operation timing behavior identification method and system based on continuous learning

Abstract

The present application relates to the technical field of electric power operation behavior recognition, in particular to a power grid operation time sequence behavior recognition method and system based on continuous learning. 1080P high-definition video data of power grid outage operation scenes is collected, covering different light, time, location and weather conditions, and key frames and time period selection and category labeling are performed by professional personnel; a target detection model is used to generate detection frame labeling, and a double-flow feature extraction network is used to obtain space-time features; a multi-strategy sample selection module is constructed, challenging samples are selected through error change rate and sample similarity calculation, and a sample memory library is established; based on the principle of continuous learning, samples of two task periods are iteratively trained; and finally the identification of specific time sequence behaviors such as climbing and connecting ground wires is realized. Through continuous learning and multi-strategy sample selection, the present application significantly improves the recognition accuracy and realizes accurate monitoring of complex power grid operation scenes, providing effective protection for the safe operation of the power system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric power operation behavior recognition, and particularly relates to a power grid operation time sequence behavior recognition method and system based on continuous learning. BACKGROUND

[0002] With the rapid development of modern power grid systems, power grid operation scenarios are becoming increasingly complex and diverse. The behaviors of operation personnel directly affect the safety, stability, and operational efficiency of power systems. In order to ensure the efficient operation and safety of power grids, how to identify, monitor, and respond to the behaviors of operation personnel in real time has become a crucial technical problem in the current power grid field. Traditional behavior recognition methods usually rely on qualitative analysis and manual monitoring, which have significant limitations. In particular, when faced with complex and variable power grid operation scenarios, the efficiency is low and the response is lagging. In addition, in the event of sudden failures or abnormal situations, traditional methods cannot quickly identify potential risks, increasing the safety hazards and operational pressure of the system.

[0003] In recent years, with the rapid development of computer vision, deep learning, sensor fusion, and other technologies, intelligent behavior recognition technology based on video data and multi-source sensor information has gradually become the mainstream solution for power grid operation behavior monitoring. Compared with traditional methods, machine learning models can automatically extract features and more quickly and accurately identify the behavior patterns of operation personnel, especially in the detection of abnormal behaviors. However, existing behavior recognition models mostly rely on pre-defined static data sets for training, lacking the ability to dynamically adapt to changes in behavior patterns in complex scenarios. This not only limits the generalization ability of the model, but also easily leads to a decrease in recognition accuracy, frequent false alarms, and other problems in actual power grid operation scenarios, affecting the stability and reliability of the system.

[0004] The highly variable nature of power grid operation scenarios requires behavior recognition models to have the ability to continuously learn and dynamically adapt to new behavior patterns. In particular, in the long-term deployment and operation, changes in behavior patterns are inevitable. If the model cannot adapt to new data distribution and behavior characteristics in a timely manner, the recognition effect will be significantly reduced. Therefore, developing a continuous learning model that can maintain efficient recognition in complex and variable scenarios is crucial. Such a model not only needs to have dynamic updating ability in long-term operation, but also should effectively deal with newly emerging behavior patterns to ensure the long-term safety of the system. At the same time, existing behavior recognition models often face a large amount of redundant data during training, leading to waste of computing resources and low training efficiency. In particular, in the power grid operation scenario, most of the collected data is for regular operation behavior, and the model has difficulty in effectively distinguishing which samples have a key impact on its performance. Therefore, under the continuous learning framework, how to accurately identify and select key samples that have a significant impact on model training has become a core technical problem to improve the efficiency and adaptability of the model. SUMMARY

[0005] In view of the problems in the prior art, the present application is proposed.

[0006] Therefore, the problem to be solved by the present application is how to accurately identify and select key samples that have a significant impact on model training, which is a core technical problem for improving the efficiency and adaptability of the model.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] In a first aspect, the present application provides a power grid operation time sequence behavior identification method based on continuous learning, which includes collecting samples containing specific time sequence behaviors in a power grid outage operation scenario, and performing key frame and time period selection and category labeling on the samples;

[0009] Performing feature extraction and processing on the samples to obtain detection frame labeled detection samples, and simultaneously obtaining original space-time features of the detection samples;

[0010] Constructing a multi-strategy sample selection module to select challenging samples for model learning, and establishing a sample memory bank;

[0011] Based on the principle of continuous learning, iteratively training samples in the sample memory bank to obtain an identification model;

[0012] Using the identification model for testing to obtain the identification result of the specific time sequence behavior in the power grid outage operation scenario.

[0013] As a preferred scheme of the power grid operation time sequence behavior identification method based on continuous learning, wherein: the collecting of samples containing specific time sequence behaviors in the power grid outage operation scenario includes: using 1080P high-definition resolution to record videos, with an average video duration of 10 minutes; video data acquisition covers different light, time, location and weather change conditions; professional personnel perform key frame and time period selection and category labeling, wherein the specific time sequence behavior category includes climbing and connecting ground wires.

[0014] As a preferred scheme of the power grid operation time sequence behavior identification method based on continuous learning, wherein: the feature extraction and processing of the samples includes: using a target detection model to process the collected video data, automatically detecting and positioning the operation personnel in the video to obtain detection frame labeled information; using a double-flow feature extraction network to perform deep-level feature extraction on the detection result to obtain spatial information and time information, thereby extracting original space-time features; performing data enhancement on the original space-time features to obtain enhanced space-time features.

[0015] As a preferred scheme of the power grid operation time sequence behavior recognition method based on continuous learning, the multi-strategy sample selection module comprises: inputting the original space-time features and the enhanced space-time features into the recognition model, performing iterative training to obtain embedding features and errors of each sample; calculating the error change rate using the errors, and selecting challenging samples according to a threshold; calculating the similarity between samples, and selecting difficult samples with the smallest similarity for training; and integrating the selected challenging samples to establish a sample memory bank.

[0016] As a preferred scheme of the power grid operation time sequence behavior recognition method based on continuous learning, the method for calculating the error change rate comprises: calculating based on the error change value of the sample between consecutive iteration rounds; considering the behavior duration of the sample and the decay effect of the historical error; and selecting samples with error change exceeding a threshold τ1.

[0017] As a preferred scheme of the power grid operation time sequence behavior recognition method based on continuous learning, the calculation of the similarity between samples comprises: calculating the similarity between samples using cosine similarity; calculating the similarity expectation of the difficult sample and other samples of the same category; and calculating the similarity expectation change rate between samples, and screening samples with large change rate according to a threshold τ2.

[0018] As a preferred scheme of the power grid operation time sequence behavior recognition method based on continuous learning, the iterative training based on the continuous learning principle comprises: training based on the samples in the sample memory bank learned in the first task period; training based on new samples selected by the multi-strategy sample selection module in the second task period; using a cross-entropy loss function to optimize the model; and dynamically updating the sample memory bank for model learning in subsequent task periods.

[0019] In a second aspect, the embodiments of the present application provide a power grid operation time sequence behavior recognition system based on continuous learning, which comprises a collection module that collects samples containing specific time sequence behaviors in a power grid outage operation scenario, and performs key frame and time period selection and category labeling on the samples;

[0020] An extraction module extracts and processes features of the samples to obtain detection samples labeled with detection boxes, and obtains original space-time features of the detection samples;

[0021] A construction module constructs a multi-strategy sample selection module, selects challenging samples for model learning, and establishes a sample memory bank;

[0022] An iteration module iteratively trains samples in the sample memory bank based on the principle of continuous learning to obtain a recognition model;

[0023] The identification module performs testing by using the identification model to obtain an identification result of a specific time sequence behavior in a power grid outage operation scene.

[0024] In a third aspect, an embodiment of the present application provides a computer device, comprising a memory and a processor, and the memory stores a computer program, wherein: the computer program instructions are executed by the processor to implement the steps of the power grid operation time sequence behavior identification method based on continuous learning according to the first aspect of the present application.

[0025] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, which stores a computer program, wherein: the computer program instructions are executed by the processor to implement the steps of the power grid operation time sequence behavior identification method based on continuous learning according to the first aspect of the present application.

[0026] The present application has the advantages that: the present application records videos in 1080P high-definition resolution in the data collection stage, and covers different light, time, position and weather conditions, forming a comprehensive and real data basis. This diversified data collection method not only improves the representativeness of the sample, but also enhances the adaptability of the model to complex environments, solving the problem of unstable recognition effect of traditional methods in complex environments.

[0027] By establishing a double-flow feature extraction network architecture: the spatial branch processes the RGB frame information; the time branch processes the optical flow information; the cooperative extraction of space-time features is realized, compared with the single feature extraction method, the capture ability of dynamic behavior is improved, and the recognition accuracy is improved by about 15%.

[0028] By innovatively constructing a multi-strategy sample selection module: an error change rate calculation mechanism is introduced, considering the sample behavior time length and historical error decay;

[0029] Cosine similarity is used to measure the similarity relationship between samples; the model can automatically select key samples with challenges, greatly reducing the redundancy of training data (about 40%), while maintaining the model performance, and significantly improving the training efficiency.

[0030] By designing a sample memory bank storage mechanism: dynamically updating the storage strategy; combining the learning process of two task periods; effectively solving the problem that the traditional model is easy to forget the learned experience, the recognition accuracy in the new scene is improved by 1.91% (from 82.78% to 84.69%) compared with the traditional method.

[0031] System overall synergy effect: from data acquisition to feature extraction, to model training, forming a complete technical chain; the precise identification of two key operation behaviors of climbing and connecting ground wire is realized; the climbing action recognition accuracy rate reaches 85.71%; the connecting ground wire action recognition accuracy rate reaches 80.49%; and reliable technical support is provided for power grid operation safety monitoring.

[0032] Actual application value: real-time behavior monitoring in complex environment is realized; continuous optimization and updating of the model are supported; the cost of manual monitoring is reduced; the operation safety is improved; and it is especially suitable for application in high-risk operation scenes such as power grid construction and maintenance.

[0033] Through the dual consideration of sample error change and similarity expectation change, the precise positioning of difficult samples is realized, which exceeds the expectation of traditional single evaluation index; the system performance is significantly improved through algorithm optimization without increasing hardware cost; the model has continuous learning ability and can continuously adapt to new operation scenes, which breaks through the limitation of traditional static model. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0035] Figure 1 Flowchart of the power grid operation time sequence behavior recognition method based on continuous learning;

[0036] Figure 2 Computer device diagram of the power grid operation time sequence behavior recognition method based on continuous learning;

[0037] Figure 3 Original video frame diagram in the power grid outage operation scene of the power grid operation time sequence behavior recognition method based on continuous learning;

[0038] Figure 4 Identification result example diagram in the power grid outage operation scene of the power grid operation time sequence behavior recognition method based on continuous learning. DETAILED DESCRIPTION

[0039] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0040] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the present application.

[0041] Secondly, the "one embodiment" or "an embodiment" referred to herein means a specific feature, structure, or characteristic under discussion. The appearance of the phrase "in one embodiment" in various places in the specification is not meant to refer to the same embodiment, nor is it meant to refer to only one embodiment or a particular embodiment. Moreover, the phrase "in one embodiment" is not meant to refer to a single feature, structure, or characteristic, but rather a specific feature, structure, or characteristic that can be included in at least one implementation of the present application.

[0042] Embodiment 1

[0043] Reference Figures 1-2 For the first embodiment of the present application, the embodiment provides a power grid operation timing behavior recognition method based on continuous learning, comprising,

[0044] S100: Collect samples containing specific timing behaviors in power grid outage operation scenarios, and perform key frame and time period selection and category labeling on the samples;

[0045] It is worth noting that in the preferred embodiment of the present application, collecting videos of power grid outage operation scenarios for labeling is a non-limiting preferred choice. Those skilled in the art can collect sample data under different power grid scenarios for labeling according to the recognition needs.

[0046] S101: Collecting samples containing specific timing behaviors in power grid outage operation scenarios includes: using 1080P high-definition resolution to record videos, with an average duration of 10 minutes for each video; video data collection covers different light, time, location and weather conditions; key frame and time period selection and category labeling are performed by professional personnel, wherein the specific timing behavior categories include climbing and connecting ground wires.

[0047] S200: Feature extraction and processing of the samples to obtain detection frame labeled detection samples, and obtain the original spatio-temporal features of the detection samples;

[0048] Preferably, based on the existing detection model, detection samples containing detection frame labels are obtained, and the original spatio-temporal features of the detection samples are obtained using the existing dual-flow feature extraction network, and the original spatio-temporal features are data enhanced to obtain enhanced spatio-temporal features.

[0049] S201: The feature extraction and processing of the sample includes: using a target detection model to process the collected video data, automatically detecting and locating the work personnel in the video to obtain detection frame label information; using a double-flow feature extraction network to perform deep-level feature extraction on the detection result to obtain spatial information and time information, thereby extracting original spatio-temporal features; performing data enhancement on the original spatio-temporal features to obtain enhanced spatio-temporal features.

[0050] S300: A multi-strategy sample selection module is constructed to select challenging samples for model learning, and a sample memory bank is established.

[0051] Preferably, the multi-strategy sample selection module is constructed to consider sample error changes and expected changes in sample similarity, select challenging samples, and perform model learning in a first task cycle, while establishing a sample memory bank to store these samples.

[0052] S301: The multi-strategy sample selection module is constructed to include: inputting the original spatio-temporal features and the enhanced spatio-temporal features into a recognition model for iterative training to obtain embedding features and errors of each sample; calculating an error change rate using the errors, and selecting challenging samples according to a threshold value; calculating the similarity between samples to select difficult samples with the smallest similarity for training; and integrating the selected challenging samples to establish a sample memory bank.

[0053] S302: The method for calculating the error change rate includes: calculating based on the error change value of the sample between consecutive iterative rounds; considering the behavior duration of the sample and the decay effect of the historical error; and selecting samples with error changes exceeding a threshold value τ1 by setting the threshold value.

[0054] S400: Based on the principle of continuous learning, the samples in the sample memory bank are used for iterative training to obtain a recognition model.

[0055] Preferably, based on the principle of continuous learning, the samples in the sample memory bank and the samples selected by the multi-strategy sample selection module in the second task cycle learning are used to perform iterative training on the recognition model again to obtain an optimal recognition model.

[0056] S401: The iterative training based on the principle of continuous learning includes: training based on the samples in the sample memory bank obtained in the first task cycle learning; training in combination with new samples selected by the multi-strategy sample selection module in the second task cycle; using a cross-entropy loss function for model optimization; and dynamically updating the sample memory bank for model learning in subsequent task cycles.

[0057] S500: The recognition model is used for testing to obtain an identification result of a specific time sequence behavior in a power grid outage work scenario.

[0058] Further, the embodiment also provides a power grid operation timing behavior recognition system based on continuous learning, comprising,

[0059] A collection module collects samples containing specific timing behaviors in power grid outage operation scenarios, and performs key frame and time period selection and category labeling on the samples;

[0060] An extraction module extracts and processes features of the samples to obtain detection sample labeled with a detection frame, and simultaneously obtains original spatio-temporal features of the detection sample;

[0061] A construction module constructs a multi-strategy sample selection module to select challenging samples for model learning, and simultaneously establishes a sample memory bank;

[0062] An iteration module iteratively trains based on the principle of continuous learning using samples in the sample memory bank to obtain a recognition model;

[0063] An identification module tests using the recognition model to obtain a recognition result of specific timing behaviors in the power grid outage operation scenario.

[0064] In summary, by using 1080P high-definition resolution for video recording in the data collection stage, and covering different light, time, location and weather conditions, a comprehensive and real data foundation is formed. This diversified data collection method not only improves the representativeness of the samples, but also enhances the adaptability of the model to complex environments, solving the problem of unstable recognition effect of traditional methods in complex environments.

[0065] By establishing a dual-flow feature extraction network architecture: the spatial branch processes RGB frame information; the temporal branch processes optical flow information; the cooperative extraction of spatio-temporal features is realized, which improves the capture ability of dynamic behaviors by about 15% compared with single feature extraction method, and improves the recognition accuracy.

[0066] By innovatively constructing a multi-strategy sample selection module: introducing an error change rate calculation mechanism, considering the sample behavior duration and historical error decay;

[0067] Cosine similarity is used to measure the similarity between samples; the model can automatically select key samples with challenges, greatly reducing the redundancy of training data (about 40% reduction), while maintaining the model performance, and significantly improving the training efficiency.

[0068] By designing a sample memory bank storage mechanism: dynamically updating the storage strategy; combining the learning process of two task periods; effectively solving the problem that traditional models easily forget learned experience, the recognition accuracy in new scenarios is improved by 1.91% (from 82.78% to 84.69%) compared with traditional methods.

[0069] System overall synergy effect: from data acquisition to feature extraction, to model training, forming a complete technical chain; it realizes the accurate identification of two key operation behaviors of climbing and connecting ground wire; the climbing action recognition accuracy rate reaches 85.71%; the connecting ground wire action recognition accuracy rate reaches 80.49%; and it provides reliable technical support for power grid operation safety monitoring.

[0070] Actual application value: it realizes real-time behavior monitoring in complex environment; supports continuous optimization and updating of the model; reduces the cost of artificial monitoring; improves the operation safety; and is especially suitable for application in high-risk operation scenes such as power grid construction and maintenance.

[0071] Through the dual consideration of sample error change and similarity expectation change, the accurate positioning of difficult samples is realized, which exceeds the expectation of traditional single evaluation index; without increasing the hardware cost, the system performance is significantly improved through algorithm optimization; the model has continuous learning ability and can continuously adapt to new operation scenes, which breaks through the limitation of traditional static model.

[0072] Embodiment 2

[0073] Reference Figure 2 - Figure 4 For the second embodiment of the application, the embodiment provides a power grid operation time sequence behavior identification method based on continuous learning. In order to verify the beneficial effects of the application, economic benefit calculation and simulation experiment are used for scientific demonstration.

[0074] In the power grid outage operation scene, the data is collected by combining the field video recording with the real-time monitoring system. The video recording adopts 1080P high-definition resolution, and the average duration of each video is about 10 minutes, which ensures the recording of the whole operation process. The video data collection covers various complex environments, including changes in illumination, time, location and weather. For example, the data is recorded at different time periods within a day, capturing scenes from bright sunlight to weak light environment; data is collected at different operation sites to ensure the diversity of location; and various weather conditions such as sunny, cloudy and rainy days are included to simulate the variability of actual working conditions. All collected video data is selected by professional personnel for key frames and time periods, and is labeled for categories, and specific time sequence behavior categories include climbing and connecting ground wire.

[0075] In the preferred embodiment of the application, the collected data includes 870 videos and 2107 action examples, including 760 pole climbing examples and 313 ground wire connecting examples, with an action data retention rate of about 1 / 6. The training set and the validation set are divided in a ratio of about 3:1, wherein the training set includes 642 videos, including 592 pole climbing actions and 272 ground wire connecting actions; the validation set includes 228 videos, including 168 pole climbing actions and 41 ground wire connecting actions.

[0076] Based on the existing detection model, a detection sample containing a bounding box annotation is obtained, and the original spatio-temporal features of the detection sample are obtained using the existing dual-stream feature extraction network, and the original spatio-temporal features are data-augmented to obtain enhanced spatio-temporal features.

[0077] Preferably, based on the existing target detection model (yolov5), the collected video data is processed to automatically detect and locate the workers in the video. The detection model accurately frames the area of the workers in each frame by generating bounding box annotation information. To ensure the reliability of the detection, the model will be verified based on multiple consecutive images, thereby effectively reducing the missed detection and false detection, especially in complex background and light conditions. The sample data after detection contains detailed worker position and boundary information;

[0078] Based on the detection results, the existing dual-stream feature extraction network (I3D) is further used to extract deep features from the detection results. The dual-stream network includes two parallel branches, one processing spatial information (RGB frame) and the other processing temporal information (optical flow), thereby extracting the original spatio-temporal features of the collected samples and data-augmenting the original spatio-temporal features (such as temporal mask, temporal translation, scale scaling, random noise, random rejection, etc.) to obtain enhanced spatio-temporal features.

[0079] A multi-strategy sample selection module is constructed, considering the sample error variation and the expected change in similarity between samples, to select challenging samples for model learning in the first task cycle, while establishing a sample memory bank to store these samples.

[0080] Preferably, the original spatio-temporal features and the enhanced spatio-temporal features are input into the recognition model for iterative training, obtaining the embedding features of each sample, and the error between the predicted results and the true results of each sample, which are input into the multi-strategy sample selection module.

[0081] The error is used to calculate the error variation rate, and the input samples are sorted according to the size of the calculated error variation rate, and the challenging samples are selected according to the threshold value.

[0082] The sample with the smallest similarity to the sample selected in the previous step is obtained, and the expected change rate of similarity between samples is calculated using the difficult sample and other samples in its category according to the true class label, and the samples with large change rates are selected according to the threshold value.

[0083] The challenging samples selected in the above two steps are integrated to establish a sample memory bank, and the iterative training of the recognition model in the first task cycle is performed.

[0084] Preferably, in order to meet the continuous learning, assuming learning two task cycles, in the model training process of the first task cycle, the recognition error of each sample is calculated as follows:

[0085]

[0086] Where N is the number of samples, y is the true result, is the predicted result. By loss evaluation on the training samples, the error value of each sample is obtained

[0087] In the preferred embodiment of the application, N = 4.

[0088] Preferably, based on the recorded error, the error change rate of the sample between two consecutive iteration rounds t and t+1 is defined, and the calculation method is as follows:

[0089]

[0090] Where, represents the error change value of the sample between the tth round and the t+1th round, is the basis of the error change rate, q i represents the behavior duration of the sample x i , and the longer the duration, the higher the complexity of the behavior. is a historical error decay function, and as the error of the sample in history decreases, its influence on the error change rate gradually decreases, where λ is the decay coefficient, controlling the influence of historical error. |·| represents the absolute value. α and γ respectively adjust the influence proportion of the basic error change item and the historical error item, and through these two parameters, the respective contributions of them to the error change rate can be balanced.

[0091] In order to determine the samples with significant error changes, a threshold τ1 is set, and the samples with error changes Δe i exceeding the threshold are selected as follows:

[0092] F = {x i |Δe i >τ1}i = 1,2, N.

[0093] Where, is the selected sample set, B is the number of selected samples, and these samples often represent the part that has not been well mastered in model learning, and have higher training value.

[0094] In the preferred embodiment of the application, α = 0.5, γ = 0.5, λ = 0.5, and τ1 = 0.4.

[0095] Preferably, for each sample x i, calculate the similarity S(x j , x i ) of it with all other samples x j in the current category, and select the sample with the smallest similarity as the difficult sample. The similarity calculation can use the cosine similarity formula:

[0096]

[0097] where v i and v j represent the embedding features of sample x i and sample x j , respectively, and ||·|| represents the 2-norm. Through calculation, the difficult sample x d with the smallest similarity with the selected sample x i is obtained, that is:

[0098]

[0099] According to the real category label, all samples in the category are extracted, and the similarity expectation of the difficult sample x d with other samples x k in the same category is calculated, as follows:

[0100]

[0101] where c represents the c-th category, and |c| is the number of samples in the c-th category. The similarity expectation change rate of the sample between two consecutive iteration rounds t and t+1 is defined, and the calculation method is as follows:

[0102]

[0103] The threshold τ2 of the similarity expectation change rate is set, and the sample set H is the number of selected samples.

[0104] In the preferred embodiment of the application, τ2=0.6.

[0105] Preferably, the sample set F of the selecting step (calculating the error change rate using the error, sorting the input samples according to the size of the calculated error change rate, and selecting the challenging sample according to the threshold) and the sample set G of the selecting step (obtaining the difficult sample with the smallest similarity with the selected sample in the previous step, calculating the similarity expectation change rate of the sample with other samples in the same category using the difficult sample according to the real category label, and screening the sample with a large change rate according to the threshold) are integrated to construct the sample memory bank The cross-entropy loss is used for iterative training of the recognition model, and the form is as follows:

[0106]

[0107] wherein y k is the true value label of the kth sample, is the prediction result of the kth sample by the recognition model.

[0108] In the preferred embodiment of the present application, the AdamW optimizer is used as the model optimizer, the initial value of the learning rate is set to 0.0001, the learning rate decay uses cosine annealing decay, and a total of 40 cycles are trained.

[0109] It is worth noting that in the preferred embodiment of the present application, the training method of the network and the parameter configuration of the learning rate optimization strategy are a non-restrictive optimal choice. Those skilled in the art can select the training method of the model and the parameter configuration according to various indicators such as recognition accuracy and efficiency.

[0110] Based on the principle of continuous learning, the samples in the sample memory library and the samples selected by the multi-strategy sample selection module in the second task period learning are used to iteratively train the recognition model again to obtain the best recognition model.

[0111] Preferably, based on the sample memory library obtained in the first task period learning, the samples stored therein and the samples selected by the multi-strategy sample selection module in the second task period learning are used to iteratively train the recognition model again to obtain the best recognition model.

[0112] The samples selected by the multi-strategy sample selection module in the second task period learning are stored or dynamically updated to the sample memory library to help the model learning in the next task period.

[0113] In the preferred embodiment of the present application, it is assumed that two task periods are learned. In the first task period, 40% of the samples are randomly selected from the training set as the input for the recognition model learning, and the remaining samples are left for the model learning in the second task period.

[0114] The best recognition model is used for testing to obtain the recognition result of the specific time sequence behavior in the power grid outage operation scenario.

[0115] Preferably, the trained best recognition model is deployed on the test system, and the newly collected sample data in the power grid outage operation scenario is used as the input to perform testing to obtain the recognition result.

[0116] For the behavior recognition model obtained by the preferred embodiment of the present application, 150 new sample data collected in the power grid outage operation scenario are used for 10 times of testing, and the average accuracy of sample recognition is recorded. The results are shown in Table 1.

[0117] In the preferred embodiment of the present application, the trained model is deployed on a test system to obtain the recognition result of the specific time sequence behavior of the model in the power grid outage operation scenario using test samples. The software configuration of the test system is python3.9.8, pytorch2.0.1, and cuda12.5, and the hardware configuration is L20 graphics card (48G video memory).

[0118] Part of the test effect diagram is as shown in Figure 4 The test result diagram shows that the behavior recognition model of the embodiment can accurately identify the specific time sequence behavior in the test sample data after continuous learning of the samples in the power grid outage operation scenario.

[0119] As shown in Table 1, the behavior recognition model of the embodiment is tested on the newly collected data in the power grid outage operation scenario, and the accuracy of climbing and connecting the ground wire in the method is 85.71% and 80.49%, respectively, and the average accuracy is 84.69%. Compared with the baseline method, the average accuracy is 82.78%, which is improved by 1.91%. The baseline method is the result without sample selection and storage. The results of the method show that the method improves the accuracy of the specific time sequence behavior recognition model by continuously learning new samples.

[0120] Table 1 Comparison of model results on test data

[0121] Behavior category Baseline method (accuracy) This method (accuracy) Climbing 84.52％ 85.71％ Connecting the ground wire 75.61％ 80.49％

[0122] The present application has the beneficial effect that compared with the prior art, it aims to develop an effective sample selection strategy based on a continuous learning framework. The strategy dynamically selects key samples with high learning value for the model by constructing a multi-strategy sample selection module, focusing on sample error changes and sample similarity expectation changes. On this basis, combined with continuous learning technology, the model is iteratively optimized, and a self-adaptive sample memory library is established to reduce the forgetting effect and improve the accuracy of specific time sequence behavior recognition in the power grid outage operation scenario. The method continuously updates and optimizes the model to ensure that it always maintains high efficiency and reliable recognition performance when facing the diversity and complexity of the power grid outage operation environment, thereby providing strong technical support for the safe and stable operation of the power system.

[0123] Embodiment 3

[0124] The embodiment also provides a computer device suitable for the case of the power grid operation time sequence behavior recognition method based on continuous learning, which includes a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the power grid forced oscillation detection and positioning method as proposed in the above embodiment.

[0125] The embodiment also provides a storage medium, which stores a computer program, and the computer program is executed by a processor to implement the power distribution network forced oscillation detection and positioning method.

[0126] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or can be a key, a trackball or a touchpad arranged on the shell of the computer device, or can be an external keyboard, a touchpad or a mouse, etc.

[0127] If the functions are implemented in the form of software function units and sold or used as independent products, the functions can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the technical solutions that essentially contribute to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0128] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a list of executable instructions for implementing logical functions, and can be specifically embodied in any computer readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch and execute instructions from the instruction execution system, device or apparatus) or in conjunction with these instructions execution system, device or apparatus. For the purpose of this specification, the "computer readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by an instruction execution system, device or apparatus or in conjunction with these instructions execution system, device or apparatus.

[0129] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via an optical scanner, then compiled, interpreted, or otherwise processed, as necessary, and stored in a computer memory.

[0130] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the embodiments described above, various steps or methods can be implemented, for example in software or firmware, stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any of the following technologies, known in the art, or combinations thereof, can be used: discrete logic circuitry having logic gates for implementing logic functions upon an application of data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and so forth.

[0131] It should be noted that the above examples are merely intended to illustrate the technical solutions of the present application and not to limit the same, and although the present application has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application, and all such modifications or replacements should be encompassed within the scope of the claims of the present application.

Claims

1. A method for power grid operation timing behavior recognition based on continuous learning, characterized in that: The method comprises the steps of: The sample collection comprises the following steps: 1080P high-definition video recording is adopted, and the average duration of each video is 10 minutes; video data acquisition covers different light, time, location and weather conditions; key frames and time period selection and category labeling are performed by professional personnel, wherein the specific time sequence behavior category includes climbing and connecting ground wire; The feature extraction and processing of the sample comprises the following steps: a target detection model is used to process the collected video data, automatically detect and locate the operating personnel in the video, and obtain detection frame labeling information; a double-flow feature extraction network is used to extract deep features of the detection result, so as to extract original space-time features; and the original space-time features are subjected to data enhancement to obtain enhanced space-time features; The multi-strategy sample selection module is constructed, and challenging samples are selected for model learning, and a sample memory bank is established; The multi-strategy sample selection module comprises the following steps: the original space-time features and the enhanced space-time features are input into the recognition model for iterative training to obtain the embedding features and the error of each sample; the error change rate is calculated by using the error, and the challenging samples are selected according to the threshold value; the similarity between the samples is calculated, and the difficult sample with the smallest similarity is selected for training; the selected challenging samples are integrated to establish the sample memory bank; Based on the principle of continuous learning, the samples in the sample memory bank are used for iterative training to obtain the recognition model; The iterative training based on the principle of continuous learning comprises the following steps: the samples in the sample memory bank learned based on the first task period are trained; the new samples selected by the multi-strategy sample selection module in the second task period are combined for training; and the recognition error of each sample is calculated as follows: The cross-entropy loss function is used for model optimization; and the sample memory bank is dynamically updated for model learning in the subsequent task period; The recognition model is used for testing to obtain the recognition result of the specific time sequence behavior in the power grid outage operation scene. , wherein, is the number of samples, is the true result, is the predicted result; by loss evaluation on the training samples, the error value of each sample is obtained and the error is recorded; Based on the recorded error, the rate of change of error of the sample between two consecutive iteration rounds is defined, which is calculated as follows: and ​ , in, Indicates the sample at the 1st Wheel and the first The change in error between wheels forms the basis of the error change rate. Indicates sample Duration of the behavior; It is a historical error decay function; as the sample's error decreases throughout history, its influence on the rate of change of error gradually decreases. It is the attenuation coefficient, which controls the influence of historical errors; Represents absolute value; and By adjusting the influence ratios of the basic error change term and the historical error term respectively, their respective contributions to the error change rate can be balanced. The method for calculating the error change rate comprises the following steps: the error change value of the sample between consecutive iteration rounds is calculated; the behavior duration and the decay of the historical error of the sample are considered; and the samples with error change exceeding the threshold value are selected by setting the threshold value τ1. The calculation of the similarity between the samples comprises the following steps: the cosine similarity is used to calculate the similarity between the samples; the similarity expectation of the difficult sample and other samples of the same category is calculated; and the similarity expectation change rate between the samples is calculated, and the samples with large change rate are screened according to the threshold value τ2.

2. The method of claim 1, wherein: The method further comprises the following steps:

3. The method of claim 2, wherein: The collection module collects the samples containing specific time sequence behaviors in the power grid outage operation scene, and performs key frame and time period selection and category labeling on the samples; 4. A power grid operation time sequence behavior identification system based on continuous learning, based on the power grid operation time sequence behavior identification method based on continuous learning in any one of claims 1-3, characterized in that: ​ ​ The extraction module extracts and processes the sample to obtain a detection sample with a bounding box label and original space-time features of the detection sample; The construction module constructs a multi-strategy sample selection module, selects challenging samples for model learning, and establishes an in-sample memory library; The iteration module iteratively trains based on the continuous learning principle using samples in the in-sample memory library to obtain an identification model; The identification module uses the identification model for testing to obtain an identification result of a specific time sequence behavior in a power grid outage operation scenario.

5. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that: The processor executes the computer program to implement the steps of the power grid operation time sequence behavior identification method based on continuous learning in any one of claims 1-3.

6. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to implement the steps of the power grid operation time sequence behavior identification method based on continuous learning in any one of claims 1-3.

Images