AI-based power engineering safety quality intelligent diagnosis system
By analyzing and filtering the validity and invalidity of the augmented data generated by the adversarial network, a training set was constructed, which solved the accuracy problem of the transformer fault prediction model and achieved more accurate fault prediction.
Patent Information
- Application Number
- CN202510885200.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2026-08-04
- Estimated Expiration
- 2045-06-30
AI Technical Summary
The existing transformer fault prediction models have poor fault prediction accuracy, mainly because the sample datasets extended by GAN generative adversarial networks lack diversity and heterogeneity, resulting in large generalization errors and affecting the accuracy of transformer operation and maintenance.
By acquiring normal data, fault data, and generative adversarial network augmentation data of transformers, analyzing the valid and invalid values of each augmentation data, determining its retention value, eliminating redundant or heterogeneous data, constructing a training set, and training a fault prediction model.
This improved the accuracy of the transformer fault prediction model, ensured the diversity and representativeness of the training data, and enhanced the reliability and stability of fault prediction.
Smart Images
Figure CN120833085B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of transformer fault prediction, and more specifically to an AI-based intelligent diagnostic system for power engineering safety and quality. Background Technology
[0002] Electrical engineering is related to the production, transmission, and distribution of electrical energy. It mainly includes the installation, commissioning, operation, and maintenance of electrical equipment. Transformers are an important type of electrical equipment in electrical engineering. After the installation of transformers in electrical engineering is completed, it is usually necessary to diagnose the safety and quality of the transformers during operation and to carry out timely maintenance.
[0003] Fault prediction models based on deep learning, machine learning and other AI technologies have significant advantages over traditional manual inspection and diagnosis methods in the intelligent diagnosis of safety and quality in power engineering. By monitoring and analyzing the operating data of power equipment installed in power engineering, they can predict the occurrence of equipment failures and perform timely maintenance.
[0004] However, the scale of transformer fault sample data is usually extremely small, making it difficult for AI-based fault prediction models to obtain sufficient training data. Therefore, existing methods often expand the transformer fault sample data, such as using sample datasets expanded by GAN (Generative Adversarial Network) to train transformer fault prediction models. However, GAN is prone to mode collapse, which means that the expanded sample dataset still cannot effectively cover the different types of transformer fault data distribution due to a lack of diversity. Furthermore, GAN may also generate heterogeneous sample data (due to excessive noise). Therefore, transformer fault prediction models trained directly using sample datasets expanded by GAN are prone to large generalization errors, which can lead to misdiagnosis when diagnosing the operational safety and quality of transformers after installation in power engineering, thus affecting the operation and maintenance of transformers and the stability of power transmission and transformation systems.
[0005] In other words, the fault prediction accuracy of the transformer fault prediction models provided by existing technologies is poor. Summary of the Invention
[0006] To address the technical problem of poor fault prediction accuracy in existing transformer fault prediction models, the present invention aims to provide an AI-based intelligent diagnostic system for power engineering safety and quality. The specific technical solution adopted is as follows: In a first aspect, one embodiment of the present invention provides an AI-based intelligent diagnostic system for the safety and quality of power engineering projects, the system comprising: The acquisition module is used to acquire multiple normal data, multiple fault data, and multiple extended data. The normal data is the operating data of the transformer under normal operating conditions, the fault data is the operating data of the transformer under fault conditions, and the extended data is the data obtained after inputting the multiple fault data into a generator adversarial network. The analysis module is used to obtain the valid and invalid values of each piece of expanded data. The valid value is used to indicate the degree of completeness of the corresponding expanded data to the fault type. The invalid value is used to indicate the degree to which the corresponding expanded data deviates from the data distribution of multiple associated fault data. The expanded data and the multiple associated fault data belong to the same fault type. The determination module is used to determine the reserved value of each piece of augmented data based on the valid and invalid values of each piece of augmented data. The training module is used to construct a training set based on multiple normal data, multiple fault data, and multiple expanded data after removal, and to train a fault prediction model based on the training set. The retained values and the probability of the corresponding expanded data being removed are negatively correlated.
[0007] In one embodiment, the analysis module includes an effective analysis submodule, which is used for: Obtain the feature difference value and range feature value corresponding to each augmented data, wherein the feature difference value is used to represent the degree of difference between the corresponding augmented data and the associated multiple fault data, and the range feature value is used to represent the increase in the data coverage range of the associated multiple fault data in the corresponding feature space after the addition of the corresponding augmented data; The effective value of each augmented data is obtained based on the feature difference value and range feature value corresponding to each augmented data.
[0008] In one embodiment, the effective analysis submodule includes a feature difference analysis unit, which is used for: Obtain the first feature vector of each augmented data after feature extraction, and the second feature vector of each fault data after feature extraction; Analyze the degree of difference between the first feature vector of each augmented data and its associated multiple second feature vectors to obtain the degree of difference of multiple vectors corresponding to each augmented data; The mean of the multiple vector differences corresponding to each augmented data is determined as the feature difference value corresponding to each augmented data.
[0009] In one embodiment, obtaining the first feature vector after feature extraction for each augmented data point and the second feature vector after feature extraction for each fault data point includes: Obtain multiple first feature values after feature extraction for each expanded data, and multiple second feature values after feature extraction for each fault data, wherein the multiple first feature values correspond one-to-one with multiple preset data dimensions, and the multiple second feature values also correspond one-to-one with the multiple data dimensions; Multiple first feature values extracted from each augmented data are concatenated to obtain the first concatenated vector of each augmented data, and multiple second feature values extracted from each fault data are concatenated to obtain the second concatenated vector of each fault data. The first concatenated vector of each augmented data is standardized to obtain the first feature vector corresponding to each augmented data, and the second concatenated vector of each fault data is standardized to obtain the second feature vector of each fault data.
[0010] In one embodiment, the effective analysis submodule includes a range analysis unit, the range analysis unit being used for: Obtain multiple first feature values after feature extraction for each expanded data, and multiple second feature values after feature extraction for each fault data, wherein the multiple first feature values correspond one-to-one with multiple preset data dimensions, and the multiple second feature values also correspond one-to-one with the multiple data dimensions; Aggregate the first feature value of each augmented data in each data dimension and the multiple second feature values associated with it in that data dimension to obtain the first feature set corresponding to each augmented data in each data dimension; and aggregate the multiple second feature values associated with each augmented data in each data dimension to obtain the second feature set corresponding to each augmented data in each data dimension. Analyze the differences between the first feature set and the second feature set corresponding to each augmented data point in each data dimension to obtain the range feature value corresponding to each augmented data point.
[0011] In one embodiment, the step of analyzing the difference between the first feature set and the second feature set corresponding to each augmented data in each data dimension to obtain the range feature value corresponding to each augmented data includes: Based on the ratio of the range of the first feature set and the second feature set corresponding to each augmented data in each data dimension, determine the dimension range value corresponding to each augmented data in each data dimension. The mean of the range values of multiple dimensions corresponding to each expanded data is determined as the range feature value corresponding to each expanded data, and the range values of multiple dimensions correspond one-to-one with the multiple data dimensions.
[0012] In one embodiment, the effective analysis submodule includes an effective value calculation unit, which is used for: Normalize the feature difference value corresponding to each expanded data to obtain the first normalized value corresponding to each expanded data, and normalize the range feature value corresponding to each expanded data to obtain the second normalized value corresponding to each expanded data. The product of the first normalized value and the second normalized value corresponding to each augmented data is determined as the effective value of each augmented data.
[0013] In one embodiment, the analysis module includes an invalidity analysis submodule, which is used for: A difference analysis is performed on each first time series data of each augmented data and its associated multiple second time series data under the corresponding data category to obtain multiple data difference values corresponding to each augmented data under each data category. The first time series data is the time series data of the corresponding augmented data under the corresponding data category, and the second time series data is the time series data of the fault data associated with the corresponding augmented data under the corresponding data category. From the multiple data difference values corresponding to each augmented data in each data category, the smallest data difference value is determined as the difference index value of each augmented data in that data category; Calculate the mean of multiple difference index values for each augmented data point across multiple data categories to obtain the invalid values for each augmented data point.
[0014] In one embodiment, the determining module is configured to: Normalize the valid and invalid values of each augmented data point to obtain the normalized valid and normalized invalid values of each augmented data point. The ratio of the normalized valid value to the normalized invalid value of each augmented data is determined as the retained value of each augmented data.
[0015] In one embodiment, the training module includes a training set construction unit, which is used for: Based on the retained value of each augmented data, the multiple augmented data are sorted in ascending order to obtain an augmented data sequence; Starting from the first end of the augmented data sequence, augmented data is removed sequentially so that the ratio between the first sample number and the second sample number is a preset ratio. The first sample number is the total number of the multiple fault data and the remaining augmented data after removal, and the second sample number is the number of the multiple normal data. A training set is constructed using the multiple fault data and the remaining augmented data after removal as positive samples, and the multiple normal data as negative samples.
[0016] Secondly, another embodiment of the present invention provides an AI-based intelligent diagnostic method for the safety and quality of power engineering projects, the method comprising: Multiple normal data, multiple fault data, and multiple extended data are acquired. The normal data is the operating data of the transformer under normal operating conditions, the fault data is the operating data of the transformer under fault conditions, and the extended data is the data obtained by inputting the multiple fault data into a generator adversarial network. Obtain the valid and invalid values for each piece of extended data. The valid value indicates the degree of completeness of the corresponding extended data for the fault type, and the invalid value indicates the degree to which the corresponding extended data deviates from the data distribution of multiple associated fault data. The extended data and the multiple associated fault data belong to the same fault type. Based on the valid and invalid values of each augmented data, determine the retained value for each augmented data; A training set is constructed based on multiple normal data, multiple fault data, and multiple expanded data after removal. A fault prediction model is trained based on the training set. The retained values and the probability of the corresponding expanded data being removed are negatively correlated.
[0017] Thirdly, in another embodiment of the present invention, an electronic device is provided, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method described in the second aspect above.
[0018] Fourthly, in another embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method described in the second aspect above.
[0019] The present invention has the following beneficial effects: After acquiring multiple augmented data points from the output of a generative adversarial network (GAN) to supplement and improve fault data, this invention further determines the retention value of each augmented data point by acquiring the degree of data completeness of each augmented data point for its respective fault type, as well as the degree to which each augmented data point deviates from the data distribution of multiple fault data points of the same fault type. In other words, it accurately quantifies the necessity of retaining each augmented data point, so as to identify redundant or heterogeneous data points that cannot effectively supplement or improve multiple fault data points from the multiple augmented data points. Subsequently, based on the retention values, multiple augmented data points are reasonably removed, and a training set is constructed based on the multiple augmented data points after removal, as well as multiple normal data points and multiple fault data points, to train a fault prediction model that can accurately predict faults in transformers. Attached Figure Description
[0020] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is a schematic diagram of the structure of an AI-based intelligent diagnostic system for power engineering safety and quality, provided in one embodiment of the present invention. Figure 2 This is a flowchart illustrating an AI-based intelligent diagnostic method for power engineering safety and quality, provided in one embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation
[0022] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an AI-based intelligent diagnostic system for power engineering safety and quality proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0024] The following description, in conjunction with the accompanying drawings, details the specific solution of an AI-based intelligent diagnostic system for power engineering safety and quality provided by this invention.
[0025] This invention proposes an AI-based intelligent diagnostic system for power engineering safety and quality. Please refer to [link / reference]. Figure 1 The diagram illustrates a structural schematic of an AI-based intelligent diagnostic system 100 for power engineering safety and quality, provided in an embodiment of the present invention. The system 100 includes the following modules: The acquisition module 101 is used to acquire multiple normal data, multiple fault data, and multiple supplementary data.
[0026] The normal data refers to the operating data of the transformer under normal working conditions, the fault data refers to the operating data of the transformer under fault conditions, and the supplementary data refers to the data obtained after inputting the multiple fault data into an adversarial network.
[0027] The above operating data includes data corresponding to various categories (such as current, voltage, sound, vibration, etc.), for example: voltage data when the transformer is working (obtained by voltage sensor), current data (obtained by current sensor), vibration data (obtained by vibration sensor (such as accelerometer)), and sound data (obtained by audio sensor (such as microphone).
[0028] It should be understood that the voltage data and current data mentioned above are time-series data, that is, they represent the voltage and current signals of the voltage device over a period of time; similarly, the vibration data and sound data mentioned above are also time-series data, representing the vibration and sound signals of the voltage device over a period of time.
[0029] Furthermore, the aforementioned multiple fault data points collectively correspond to multiple fault types. For example, these multiple fault types may include: phase-to-phase short circuits, winding turn-to-turn short circuits, and single-phase ground faults. In application, the number of fault data points corresponding to different fault types can be set to be consistent to avoid differences in the accuracy of the fault prediction model for different fault types due to variations in the number of fault data points. This ensures that the finally trained fault prediction model provides relatively accurate fault prediction results for various fault conditions, thereby enhancing the versatility of the finally trained fault prediction model in complex scenarios.
[0030] In some implementations, given that the transformer's operating environment is easily affected by factors such as ambient temperature, humidity, and electromagnetic interference, interpolation can be used to supplement missing values in the operating data (collected via current / voltage sensors), and the Savitzky-Golay filtering algorithm can be used to smooth noisy data in the collected operating data to remove noise and outliers, thereby ensuring the accuracy and consistency of the data.
[0031] The process of acquiring the above-mentioned multiple augmented data is as follows: multiple fault data belonging to the same fault type are used as network training samples, and the generator and discriminator in the generative adversarial network are trained in alternating iterations (first train the generator / discriminator, then train the discriminator / generator, which completes one alternating iteration training), and the generator trained in alternating iterations (with random noise as input) outputs multiple augmented data belonging to the same fault type as the input multiple fault data.
[0032] Analysis module 102 is used to obtain the valid and invalid values of each augmented data.
[0033] The valid value indicates the degree of completeness of the corresponding extended data for the fault type, and the invalid value indicates the degree to which the corresponding extended data deviates from the data distribution of the associated multiple fault data. The extended data and the associated multiple fault data belong to the same fault type.
[0034] The degree of completeness of the corresponding extended data for the fault type can also be understood as: compared with the scheme of using multiple fault data of the corresponding extended data for the fault type for model training, the increase in the prediction accuracy of the fault prediction model for the fault type of the extended data obtained by using the corresponding extended data and multiple fault data of the corresponding fault type for model training.
[0035] The higher the effective value, the more accurate and comprehensive the data distribution and fault characteristics of the corresponding fault type are after the supplementary data is added to multiple fault data of the fault type. The higher the prediction accuracy of the fault prediction model for that fault type is.
[0036] Based on the above effective value settings, among multiple augmented data, augmented data that is too similar to the original fault data can be identified (this type of augmented data can be approximately understood as a copy of the corresponding fault data). This reduces the probability of such redundant augmented data participating in model training, allowing each fault type to fully represent its data distribution characteristics through multiple corresponding fault data and augmented data, so that the fault prediction model can more accurately complete the prediction of each fault type.
[0037] The degree to which the corresponding augmented data deviates from the data distribution of the associated multiple fault data can also be understood as: compared with the scheme of using multiple fault data of the corresponding augmented data to train the model, the decrease in the prediction accuracy of the fault prediction model for the fault type of the augmented data when using the corresponding augmented data and multiple fault data of the corresponding fault type to train the model.
[0038] The higher the invalid value, the more inaccurate the data distribution and fault characteristics of the corresponding fault type are after the supplementary data is added to the fault data of the fault type. The lower the accuracy of the fault prediction model for that fault type will be.
[0039] Based on the above invalid value setting, heterogeneous data relative to multiple fault data can be identified in multiple augmented data sets. This reduces the probability of such heterogeneous augmented data participating in model training, allowing the addition of augmented data to help each fault type more fully represent its data distribution characteristics, so that the fault prediction model can more accurately complete the prediction of each fault type.
[0040] Furthermore, the analysis module 102 includes an effective analysis submodule 1021, which is used for: Obtain the feature difference value and range feature value corresponding to each augmented data, wherein the feature difference value is used to represent the degree of difference between the corresponding augmented data and the associated multiple fault data, and the range feature value is used to represent the increase in the data coverage range of the associated multiple fault data in the corresponding feature space after the addition of the corresponding augmented data; The effective value of each augmented data is obtained based on the feature difference value and range feature value corresponding to each augmented data.
[0041] The above-mentioned feature difference values are set to quantify the low-dimensional differences between the corresponding extended data and its associated multiple fault data at the data level, while the range feature values are set to quantify the high-dimensional differences between the corresponding extended data and its associated multiple fault data in the feature space. This allows for a comprehensive consideration of the extent to which the corresponding extended data is complete for the fault type, making the determined effective values more accurate and reliable.
[0042] Among them, the feature difference value and the effective value are positively correlated, and the range feature value and the effective value are also positively correlated.
[0043] It should be understood that the larger the feature difference value, the larger the position of the corresponding expanded data in the feature space (the space formed by multiple fault data). The differences between the multiple positions of the multiple fault data associated with it in the feature space are also larger. Therefore, it can fill in the areas not covered by the multiple fault data associated with it in the feature space, thereby enriching the data distribution of the fault type to which it belongs in the feature space. This increases the probability that the fault prediction model can accurately learn the fault features of the fault type to which it belongs, and ultimately enables the trained fault prediction model to have a high prediction accuracy for the fault type to which it belongs.
[0044] The larger the range feature value, the stronger the increase in the data coverage of the multiple fault data associated with it after the corresponding extended data is added. Therefore, multiple fault data with corresponding extended data can better cover the potential distribution range of the fault type they belong to in the feature space.
[0045] Furthermore, the effective analysis submodule 1021 includes a feature difference analysis unit 10211, which is used for: Obtain the first feature vector of each augmented data after feature extraction, and the second feature vector of each fault data after feature extraction; Analyze the degree of difference between the first feature vector of each augmented data and its associated multiple second feature vectors to obtain the degree of difference of multiple vectors corresponding to each augmented data; The mean of the multiple vector differences corresponding to each augmented data is determined as the feature difference value corresponding to each augmented data.
[0046] For example, the first and second feature vectors can be obtained using a feature extraction tool. For instance, the feature extraction tool can be a tool built based on a convolutional autoencoder.
[0047] In one example, the Euclidean distance between the first feature vector of each augmented data and its associated second feature vector can be determined as the vector dissimilarity of each augmented data. Accordingly, by calculating the Euclidean distance between the first feature vector of each augmented data and its associated multiple second feature vectors, the multiple vector dissimilarity of each augmented data can be obtained.
[0048] In some implementations, obtaining the first feature vector after feature extraction for each augmented data point and the second feature vector after feature extraction for each fault data point includes: Obtain multiple first feature values after feature extraction for each expanded data, and multiple second feature values after feature extraction for each fault data, wherein the multiple first feature values correspond one-to-one with multiple preset data dimensions, and the multiple second feature values also correspond one-to-one with the multiple data dimensions; Multiple first feature values extracted from each augmented data are concatenated to obtain the first concatenated vector of each augmented data, and multiple second feature values extracted from each fault data are concatenated to obtain the second concatenated vector of each fault data. The first concatenated vector of each augmented data is standardized to obtain the first feature vector corresponding to each augmented data, and the second concatenated vector of each fault data is standardized to obtain the second feature vector of each fault data.
[0049] Specifically, the data dimension is a composite dimension consisting of the type of data source and the monitored indicators. The types of data sources include current, voltage, sound, vibration, etc.; the monitored indicators include statistical indicators (such as mean, variance, standard deviation, extreme values), time-domain indicators (such as waveform factor, peak factor, etc.), frequency-domain indicators (such as peak value, average frequency, centroid of the spectrum, power spectral density, etc.).
[0050] In this implementation, feature extraction should be understood as: the process of performing data analysis using processing rules that match the detected indicators of the corresponding data dimensions for the data portion of the expanded data / fault data corresponding to the data type (such as current, voltage, sound, vibration).
[0051] For example, multiple first feature values of each augmented data after feature extraction can be concatenated end to end to obtain the first feature vector (one-dimensional vector) corresponding to each augmented data. Alternatively, multiple first feature values extracted from each augmented data can be concatenated vertically and horizontally along the data dimension to obtain the first feature vector (an m×n two-dimensional vector, where m is the total number of types corresponding to the data source and n is the total number of indicators monitored) for each augmented data.
[0052] The method for obtaining the second feature vector is the same as that for obtaining the first feature vector, so it will not be repeated here to avoid repetition.
[0053] The above standardization process can be understood as normalizing the value of each feature element in the first and second feature vectors (scaling the value to the (0, 1) interval). For example, the above normalization process can use Min-Max standardization or Z-Score standardization.
[0054] It should be noted that, compared to directly extracting feature vectors through complex operations such as convolution, the method of obtaining feature values based on simple operations such as the mean and variance, and then concatenating the feature values to form feature vectors (such as the first feature vector and the second feature vector), can support richer and more flexible indicator and type settings, while reducing the computational complexity of feature vectors and improving the efficiency of feature vector acquisition.
[0055] Furthermore, the effective analysis submodule 1021 includes a range analysis unit 10212, which is used for: Obtain multiple first feature values after feature extraction for each expanded data, and multiple second feature values after feature extraction for each fault data, wherein the multiple first feature values correspond one-to-one with multiple preset data dimensions, and the multiple second feature values also correspond one-to-one with the multiple data dimensions; Aggregate the first feature value of each augmented data in each data dimension and the multiple second feature values associated with it in that data dimension to obtain the first feature set corresponding to each augmented data in each data dimension; and aggregate the multiple second feature values associated with each augmented data in each data dimension to obtain the second feature set corresponding to each augmented data in each data dimension. Analyze the differences between the first feature set and the second feature set corresponding to each augmented data point in each data dimension to obtain the range feature value corresponding to each augmented data point.
[0056] The aforementioned range feature values are used to quantify the degree of difference in the data coverage of multiple fault data associated with the corresponding expanded data before and after the addition of the expanded data.
[0057] For example, the intersection-union ratio between the first feature set and the second feature set corresponding to each augmented data in each data dimension can be determined as the range feature value corresponding to each augmented data.
[0058] In some implementations, the step of analyzing the differences between the first feature set and the second feature set corresponding to each augmented data in each data dimension to obtain the range feature value corresponding to each augmented data includes: Based on the ratio of the range of the first feature set and the second feature set corresponding to each augmented data in each data dimension, determine the dimension range value corresponding to each augmented data in each data dimension. The mean of the range values of multiple dimensions corresponding to each expanded data is determined as the range feature value corresponding to each expanded data, and the range values of multiple dimensions correspond one-to-one with the multiple data dimensions.
[0059] In this implementation, the range of the set (the difference between the maximum and minimum values) is used to represent the data coverage of the corresponding set (such as the first feature set and the second feature set). By calculating the ratio of the ranges, the dimensional range value of the corresponding expanded data under the corresponding data dimension is determined accordingly. This accurately represents the degree of difference in the data coverage of multiple fault data associated with the corresponding expanded data before and after adding the expanded data under the corresponding data dimension. Finally, the range feature value of the corresponding expanded data is determined by the mean calculation method to minimize the influence of extreme values and make the determined dimensional range value more accurate and reliable.
[0060] Furthermore, the effective analysis submodule 1021 includes an effective value calculation unit 10213, which is used for: Normalize the feature difference value corresponding to each expanded data to obtain the first normalized value corresponding to each expanded data, and normalize the range feature value corresponding to each expanded data to obtain the second normalized value corresponding to each expanded data. The product of the first normalized value and the second normalized value corresponding to each augmented data is determined as the effective value of each augmented data.
[0061] In this approach, normalization is used to balance the numerical differences between feature difference values and range feature values. This avoids the determination of effective values being excessively influenced by feature difference values or range feature values due to their numerical differences, thus ensuring the accuracy of the final effective values. Specifically, using the product of the normalized values as the effective values simplifies the calculation of effective values as much as possible while ensuring their accuracy, thereby improving the calculation efficiency of effective values and reducing the time required to prepare the training set for the fault prediction model.
[0062] The normalization process applied to the feature difference values and range feature values can be accomplished using the Min-Max normalization method.
[0063] Furthermore, the analysis module 102 includes an invalidity analysis submodule 1022, which is used for: A difference analysis is performed on each first time series data of each augmented data and its associated multiple second time series data under the corresponding data category to obtain multiple data difference values corresponding to each augmented data under each data category. The first time series data is the time series data of the corresponding augmented data under the corresponding data category, and the second time series data is the time series data of the fault data associated with the corresponding augmented data under the corresponding data category. From the multiple data difference values corresponding to each augmented data in each data category, the smallest data difference value is determined as the difference index value of each augmented data in that data category; Calculate the mean of multiple difference index values for each augmented data point across multiple data categories to obtain the invalid values for each augmented data point.
[0064] For example, the various data categories corresponding to fault data and extended data can be: current, voltage, sound, vibration, etc.
[0065] Specifically, dynamic time warping (DTW) can be performed on the first time series data of each augmented data under each data category and the multiple second time series data associated with it under that data category to calculate multiple DTW distances between the first time series data of each augmented data under each data category and the multiple second time series data associated with it under that data category, and these distances can be used as multiple data difference values for each augmented data under each data dimension.
[0066] By selecting the smallest data difference value among multiple data difference values for each augmented data in each data category as the difference index value of the corresponding augmented data in the corresponding data category, the degree of difference of the corresponding augmented data relative to other related fault data in the corresponding data category can be accurately quantified.
[0067] By averaging multiple difference index values, the impact of extreme values can be reduced, allowing invalid values to accurately reflect the overall deviation of the corresponding expanded data from the data distribution of its associated multiple fault data under multiple data categories.
[0068] It should be understood that the larger the invalid value, the greater the degree to which the corresponding augmented data deviates from the true data distribution of its fault type. In other words, the more likely the corresponding augmented data is to be heterogeneous data that does not have the true fault characteristics of the transformer. Therefore, the corresponding augmented data should be removed to ensure the authenticity of the training samples used to train the fault prediction model.
[0069] The determination module 103 is used to determine the reserved value of each augmented data based on the valid and invalid values of each augmented data.
[0070] The above retention values are used to indicate the degree to which the corresponding supplementary data is necessary to be retained.
[0071] Specifically, the determining module 103 is used for: Normalize the valid and invalid values of each augmented data point to obtain the normalized valid and normalized invalid values of each augmented data point. The ratio of the normalized valid value to the normalized invalid value of each augmented data is determined as the retained value of each augmented data.
[0072] It should be understood that valid and invalid values will be normalized to the same interval (e.g., (0, 1)) to avoid large numerical differences between them, thereby ensuring the accuracy of the subsequently determined retained values.
[0073] The method used for normalization can be found in the foregoing description, and will not be repeated here to avoid repetition.
[0074] The training module 104 is used to construct a training set based on multiple normal data, multiple fault data, and multiple expanded data after removal, and to train a fault prediction model based on the training set, wherein the retained value and the probability of the corresponding expanded data being removed are negatively correlated.
[0075] In one example, a retention threshold (such as 2) can be preset, and augmented data below the retention threshold can be removed from multiple augmented data sets to obtain multiple augmented data sets after removal.
[0076] For example, the above fault prediction model can be a model built from a 1D-CNN neural network or a DBN neural network.
[0077] In applications, the operating data of transformers installed in power engineering can be used as input to the fault prediction model to predict transformer faults and output the fault prediction results. Then, appropriate measures can be taken based on the obtained fault prediction results.
[0078] For example, when the fault prediction result indicates that the transformer is operating normally, the input operating data can be labeled according to the actual situation (e.g., if the transformer is indeed operating normally, the input operating data is labeled as normal data; while if the transformer has a fault, the input operating data is labeled as incorrectly identified fault data) so that the fault prediction model can be iteratively updated in the future. Alternatively, if the fault prediction results indicate that the transformer is at risk of failure, the staff can be reminded to perform appropriate inspections and maintenance on the target transformer.
[0079] After acquiring multiple augmented data points from the output of a generative adversarial network (GAN) to supplement and improve fault data, this invention further determines the retention value of each augmented data point by acquiring the degree of data completeness of each augmented data point for its respective fault type, as well as the degree to which each augmented data point deviates from the data distribution of multiple fault data points of the same fault type. In other words, it accurately quantifies the necessity of retaining each augmented data point, so as to identify redundant or heterogeneous data points that cannot effectively supplement or improve multiple fault data points from the multiple augmented data points. Subsequently, based on the retention values, multiple augmented data points are reasonably removed, and a training set is constructed based on the multiple augmented data points after removal, as well as multiple normal data points and multiple fault data points, to train a fault prediction model that can accurately predict faults in transformers.
[0080] In one embodiment, the training module includes a training set construction unit, which is used for: Based on the retained value of each augmented data, the multiple augmented data are sorted in ascending order to obtain an augmented data sequence; Starting from the first end of the augmented data sequence, augmented data is removed sequentially so that the ratio between the first sample number and the second sample number is a preset ratio. The first sample number is the total number of the multiple fault data and the remaining augmented data after removal, and the second sample number is the number of the multiple normal data. A training set is constructed using the multiple fault data and the remaining augmented data after removal as positive samples, and the multiple normal data as negative samples.
[0081] For example, the preset ratio can be 1:1.5 or 1:2.
[0082] For example, if the total number of augmented data in the augmented data sequence is 1000, and when the number of augmented data becomes 500, the ratio between the number of the first sample and the number of the second sample becomes a preset ratio, then the above elimination operation is as follows: starting from the first position of the augmented data sequence, the augmented data is eliminated sequentially until the 500th position of the augmented data sequence (if the first position number is 0, then the 500th position number is 499).
[0083] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above.
[0084] This invention proposes an AI-based intelligent diagnostic method for safety and quality in power engineering. Please refer to [link / reference]. Figure 2 The diagram illustrates a flowchart of an AI-based intelligent diagnostic method for power engineering safety and quality, provided by an embodiment of the present invention. The method includes: Secondly, another embodiment of the present invention provides an AI-based intelligent diagnostic method for the safety and quality of power engineering projects, the method comprising: Step S1: Obtain multiple normal data, multiple fault data, and multiple supplementary data.
[0085] The normal data refers to the operating data of the transformer under normal working conditions, the fault data refers to the operating data of the transformer under fault conditions, and the supplementary data refers to the data obtained after inputting the multiple fault data into an adversarial network.
[0086] Step S2: Obtain the valid and invalid values for each augmented data.
[0087] The valid value indicates the degree of completeness of the corresponding extended data for the fault type, and the invalid value indicates the degree to which the corresponding extended data deviates from the data distribution of the associated multiple fault data. The extended data and the associated multiple fault data belong to the same fault type.
[0088] Step S3: Determine the retained value for each augmented data based on the valid and invalid values of each augmented data.
[0089] Step S4: Construct a training set based on multiple normal data, multiple fault data, and multiple expanded data after removal, and train a fault prediction model based on the training set.
[0090] The probability of the retained value and the corresponding expanded data being removed are negatively correlated.
[0091] The AI-based intelligent diagnostic method for power engineering safety and quality provided in the above embodiments and the AI-based intelligent diagnostic system for power engineering safety and quality belong to the same concept. The specific implementation process is detailed in the system embodiment and will not be repeated here.
[0092] This invention also provides an electronic device. Please refer to [link to relevant documentation]. Figure 3 The electronic device may include a processor 301, a memory 302, and a program 3021 stored in the memory 302 and capable of running on the processor 301.
[0093] When program 3021 is executed by processor 301, it can achieve the following: Figure 2 Any steps in the corresponding method embodiments and the achievement of the same beneficial effects will not be repeated here.
[0094] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by hardware related to program instructions, and the program can be stored in a readable medium.
[0095] This invention also provides a readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described functions. Figure 2 Any step in the corresponding method embodiment can achieve the same technical effect, and will not be repeated here to avoid repetition.
[0096] The computer-readable storage medium of this invention can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. For example, a computer-readable storage medium can be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
[0097] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.
[0098] The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.
[0099] Computer program code for performing the operations of this invention can be written in one or more programming languages or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, and C++, as well as conventional procedural programming languages such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or terminal. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0100] This invention also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned steps to achieve the AI-based intelligent diagnostic method for power engineering safety and quality provided in the above embodiments.
[0101] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0102] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
Claims
1. An AI-based power engineering safety quality intelligent diagnosis system, characterized in that, The system includes: The acquisition module is used to acquire multiple normal data, multiple fault data, and multiple extended data. The normal data is the operating data of the transformer under normal operating conditions, the fault data is the operating data of the transformer under fault conditions, and the extended data is the data obtained after inputting the multiple fault data into a generator adversarial network. The analysis module is used to obtain the valid and invalid values of each piece of expanded data. The valid value is used to indicate the degree of completeness of the corresponding expanded data to the fault type. The invalid value is used to indicate the degree to which the corresponding expanded data deviates from the data distribution of multiple associated fault data. The expanded data and the multiple associated fault data belong to the same fault type. The determination module is used to determine the reserved value of each piece of augmented data based on the valid and invalid values of each piece of augmented data. The training module is used to construct a training set based on multiple normal data, multiple fault data, and multiple expanded data after removal, and to train a fault prediction model based on the training set, wherein the retained value and the probability of the corresponding expanded data being removed are negatively correlated. The analysis module includes an effective analysis submodule, which is used to: obtain the feature difference value and range feature value corresponding to each piece of expanded data, wherein the feature difference value is used to represent the degree of difference between the corresponding expanded data and multiple associated fault data, and the range feature value is used to represent the increase in the data coverage range of the multiple associated fault data in the corresponding feature space after the addition of the corresponding expanded data; and obtain the effective value of each piece of expanded data based on the feature difference value and range feature value corresponding to each piece of expanded data. The effective analysis submodule includes a range analysis unit, which is used to: acquire multiple first feature values of each expanded data after feature extraction, and multiple second feature values of each fault data after feature extraction, wherein the multiple first feature values correspond one-to-one with multiple preset data dimensions, and the multiple second feature values also correspond one-to-one with the multiple data dimensions; aggregate the first feature value of each expanded data in each data dimension and the multiple second feature values associated with it in that data dimension to obtain a first feature set corresponding to each expanded data in each data dimension; aggregate the multiple second feature values associated with each expanded data in each data dimension to obtain a second feature set corresponding to each expanded data in each data dimension; and analyze the difference between the first feature set and the second feature set corresponding to each expanded data in each data dimension to obtain the range feature value corresponding to each expanded data. The analysis module includes an invalidity analysis submodule, which is used to: perform difference analysis on each first time-series data of each augmented data and its associated second time-series data under the corresponding data category to obtain multiple data difference values corresponding to each augmented data under each data category, wherein the first time-series data is the time-series data of the corresponding augmented data under the corresponding data category, and the second time-series data is the time-series data of the fault data associated with the corresponding augmented data under the corresponding data category; determine the smallest data difference value from the multiple data difference values corresponding to each augmented data under each data category as the difference index value of each augmented data under that data category; calculate the mean of the multiple difference index values of each augmented data under multiple data categories to obtain the invalid value of each augmented data. 2.The AI-based electric power engineering safety quality intelligent diagnosis system according to claim 1, characterized in that, The effective analysis submodule includes a feature difference analysis unit, which is used for: Obtain the first feature vector of each augmented data after feature extraction, and the second feature vector of each fault data after feature extraction; Analyze the degree of difference between the first feature vector of each augmented data and its associated multiple second feature vectors to obtain the degree of difference of multiple vectors corresponding to each augmented data; The mean of the multiple vector differences corresponding to each augmented data is determined as the feature difference value corresponding to each augmented data. 3.The AI-based electric power engineering safety quality intelligent diagnosis system according to claim 2, characterized in that, The process of obtaining the first feature vector after feature extraction for each augmented data point and the second feature vector after feature extraction for each fault data point includes: Obtain multiple first feature values after feature extraction for each expanded data, and multiple second feature values after feature extraction for each fault data, wherein the multiple first feature values correspond one-to-one with multiple preset data dimensions, and the multiple second feature values also correspond one-to-one with the multiple data dimensions; Multiple first feature values extracted from each augmented data are concatenated to obtain the first concatenated vector of each augmented data, and multiple second feature values extracted from each fault data are concatenated to obtain the second concatenated vector of each fault data. The first concatenated vector of each augmented data is standardized to obtain the first feature vector corresponding to each augmented data, and the second concatenated vector of each fault data is standardized to obtain the second feature vector of each fault data. 4.The AI-based electric power engineering safety quality intelligent diagnosis system according to claim 1, characterized in that, The step of analyzing the differences between the first feature set and the second feature set corresponding to each augmented data in each data dimension, and obtaining the range feature value corresponding to each augmented data, includes: Based on the ratio of the range of the first feature set and the second feature set corresponding to each augmented data in each data dimension, determine the dimension range value corresponding to each augmented data in each data dimension. The mean of the range values of multiple dimensions corresponding to each expanded data is determined as the range feature value corresponding to each expanded data, and the range values of multiple dimensions correspond one-to-one with the multiple data dimensions. 5.The AI-based electric power engineering safety quality intelligent diagnosis system according to claim 1, characterized in that, The effective analysis submodule includes an effective value calculation unit, which is used for: Normalize the feature difference value corresponding to each expanded data to obtain the first normalized value corresponding to each expanded data, and normalize the range feature value corresponding to each expanded data to obtain the second normalized value corresponding to each expanded data. The product of the first normalized value and the second normalized value corresponding to each augmented data is determined as the effective value of each augmented data. 6.The AI-based electric power engineering safety quality intelligent diagnosis system according to claim 1, characterized in that, The determining module is used for: Normalize the valid and invalid values of each augmented data point to obtain the normalized valid and normalized invalid values of each augmented data point. The ratio of the normalized valid value to the normalized invalid value of each augmented data is determined as the retained value of each augmented data. 7.The AI-based electric power engineering safety quality intelligent diagnosis system according to claim 1, characterized in that, The training module includes a training set construction unit, which is used for: Based on the retained value of each augmented data, the multiple augmented data are sorted in ascending order to obtain an augmented data sequence; Starting from the first end of the augmented data sequence, augmented data is removed sequentially so that the ratio between the first sample number and the second sample number is a preset ratio. The first sample number is the total number of the multiple fault data and the remaining augmented data after removal, and the second sample number is the number of the multiple normal data. A training set is constructed using the multiple fault data and the remaining augmented data after removal as positive samples, and the multiple normal data as negative samples.