CN116150652A - Lightning radiation waveform classification system and method - Google Patents
Lightning radiation waveform classification system and method Download PDFInfo
- Publication number
- CN116150652A CN116150652A CN202211673315.3A CN202211673315A CN116150652A CN 116150652 A CN116150652 A CN 116150652A CN 202211673315 A CN202211673315 A CN 202211673315A CN 116150652 A CN116150652 A CN 116150652A
- Authority
- CN
- China
- Prior art keywords
- lightning
- waveform
- classification
- radiation waveform
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Measurement Of Radiation (AREA)
Abstract
The invention discloses a lightning radiation waveform classification system and a lightning radiation waveform classification method, wherein a lightning radiation waveform fragment sample is obtained through a waveform segmentation method, and a corresponding relation sample of the lightning radiation waveform fragment sample and a discharge event is used as a training data set; establishing a convolutional neural network model, performing parameter tuning and training of the convolutional neural network model based on the training data set, and obtaining a classification model after training; calculating the lightning energy spectrum of the lightning radiation waveform segment by window division to obtain a waveform set to be classified; inputting the data set to be classified into the trained classification model for classification, and obtaining a classification result. The method solves the problems of weak universality and low efficiency in lightning waveform classification, realizes accurate and efficient classification of the lightning radiation waveform, and has the advantages of strong model robustness and transparent and reliable classification process.
Description
Technical Field
The invention belongs to the technical field of lightning classification, and particularly relates to a lightning radiation waveform classification system and a lightning radiation waveform classification method.
Background
Lightning is an atmospheric discharge phenomenon with ultra-large scale, and causes great threat to human production and life, so that research and protection of lightning are more important. In the process of the initiation and development of lightning discharge, the charge transport process in different discharge stages can generate radiation electromagnetic field signals with different characteristics in space excitation, provides important indication of the type of lightning discharge for lightning positioning, and is an important component of lightning multi-physical quantity observation data. Meanwhile, the judgment of the discharge type of the lightning radiation electromagnetic field provides an important basis for inversion of lightning parameters, and plays a key role in analysis and statistics of the discharge characteristic parameters such as discharge duration, discharge current amplitude, discharge times and the like. Therefore, the accurate and efficient classification of the lightning radiation waveforms has important significance for improving the lightning detection level and perfecting the research of the lightning mechanism.
The difference of charge transport processes in the lightning discharge process enables electromagnetic radiation waveforms generated by all discharge events to have different waveform characteristics, but the distribution of key waveform parameters of the radiation waveforms is affected by weather conditions, and the parameter distribution ranges in different areas are different, so that the same waveform parameter distribution ranges of different waveforms may overlap. Meanwhile, the lightning discharge process is various in types, various electromagnetic field radiation signals are generated, and different discharge event waveforms can be similar. In summary, the overlapping of waveform parameter ranges and the similarity of discharge event waveforms create difficulties in achieving accurate waveform classification.
At present, a criterion method and a machine learning method are available for a lightning electromagnetic radiation waveform classification method. The criterion method is to count and obtain key waveform parameter distribution ranges such as rising and falling time, wave head time, half-peak width, amplitude and the like of different discharge event waveforms, and extract multi-parameter classification criteria by utilizing the difference of the waveform parameter distribution ranges of different discharge events to finish waveform classification. Based on the method, patent CN104569624a proposes a full-lightning cloud-to-ground flash identification method, and classification criteria of ground flash, narrow bipolar pulse and cloud flash signals are summarized through nine waveform parameters such as waveform rising and falling time, pulse width, forward-backward inverse peak ratio, sub-peak ratio, signal-to-noise ratio, forward-backward signal-to-noise ratio and the like. The machine learning method uses a support vector machine method (Supporting Vector Machine, SVM), a Random Forest method (RF), or the like to accomplish waveform classification. Before training the model, the two methods need to calculate the amplitude-frequency characteristics of the waveform by carrying out characteristic engineering, and select part of key data points to be combined to form a characteristic vector. The SVM method is essentially that a hyperplane is found through training so that the distance between the plane and the normal characteristic vector and the opposite characteristic vector is maximum, and therefore classification is achieved. Whereas the RF method generalizes a set of classification rules from the feature vector set by training so that the classification results under the rules contradict the data set to a minimum.
The two methods are difficult to realize accurate classification of lightning radiation electromagnetic field waveforms, and the specific reasons are as follows:
1) The multi-parameter criteria used by the criteria method are from the long-term observation and collection of lightning activity radiation waveforms in a certain area by a lightning monitoring station. The detection range of the current wide-area lightning monitoring system is usually national, and the detection range contains various climatic conditions, and due to (1) the waveform parameter ranges generated by different discharge events are overlapped (2) the waveform parameter distribution ranges of the same waveform under different climates are inconsistent, the proper and general classification criterion is difficult to obtain.
2) The machine learning method mainly has the following three problems: (1) the machine learning methods of SVM, RF, etc. require manual a priori knowledge to predetermine the feature vector composition. If the number of the features is too small, accurate classification cannot be realized, and under fitting occurs. If the number of features is too large, the model may be too focused on a feature in the classification process resulting in an overfitting. Because of the difference of climates in different areas, when the source of the data set changes, the characteristic engineering needs to be conducted again to adjust the characteristic vector composition so as to keep the model to have a good classifying effect. (2) The machine learning methods such as SVM, RF and the like are difficult to process long waveform fragments, extra effort is required to perform waveform preprocessing, and the complexity of input data is reduced through means such as downsampling, feature engineering and the like. The duration of waveforms generated by different discharge events such as cloud flash, ground flash and the like is different, the distribution range can be from microseconds to milliseconds, a unified and proper data preprocessing scheme is difficult to determine, and the processing difficulty of the method is increased. (3) The weight parameters of the SVM method in the low-dimensional linear model can be interpreted as the importance of the features, so that the explanation of the classification process is realized. However, for the classified object of the lightning radiation waveform, the feature vector dimension is high, the weight parameter corresponds to a high-dimensional space through kernel function mapping, and the method has no visual interpretability. The RF method decomposes the classification process into multiple threshold judgment and outputs the classification result in a judgment tree mode, the process of the method is clear, however, the source of the threshold used for judgment cannot be given, and visual classification basis cannot be given. The unexplained basis of classification reduces the reliability of the model and makes it difficult to make the model a classification decision reference.
In summary, the current lightning radiation waveform classification method has three problems: (1) the prior method can not directly process the original waveform, and needs to introduce manual priori knowledge to process the waveform to refine the characteristics. Waveform information suffers from certain loss in the process of summarizing waveform characteristics, and selected characteristics are manually selected, so that all kinds of lightning radiation waveforms can not be summarized comprehensively and accurately. (2) Existing methods lack versatility across different data sets. The multi-parameter criteria in the criteria method are often from one or a plurality of areas and cannot be suitable for all climatic conditions. There is also some similarity in waveforms generated by different types of discharge events. These two points make it difficult to establish a common set of composite criteria for lightning waveform classification. The classification performance in machine learning closely depends on manually pre-selected features, and an optimal feature combination under a certain data set may not achieve optimal performance under other data sets, and additional adjustment is required, so that the efficiency and the universality are lacking. (3) The existing method can perspective the calculation result of each step of the algorithm, however, the calculation result cannot explain which part of the model plays an important role in the classification result in a manner of human being usual cognition, and cannot explain how the change of the input waveform affects the output. The lack of interpretability of the model leads the model training to have a certain blindness, and is difficult to solve the problems of low universality and the like possibly existing in a targeted manner, thereby reducing the efficiency of the model training.
Disclosure of Invention
The invention aims to provide an interpretable lightning radiation waveform artificial intelligent classification method based on a convolutional neural network (Convolutional Neural Network, CNN) and a class activation method, and aims to solve the problems of weak universality, low efficiency and the like of other classification methods.
In order to achieve the purpose, the lightning radiation waveform classification system comprises a training data set building module, a convolutional neural network model training module, a data set to be classified building module and a waveform intelligent classification module; the training data set establishing module is used for obtaining a lightning radiation waveform segment sample through a waveform segmentation method and taking a corresponding relation sample of the lightning radiation waveform segment sample and a discharge event as a training data set; the convolutional neural network model training module is used for establishing a convolutional neural network model, performing parameter tuning and training of the convolutional neural network model based on the training data set, and obtaining a classification model after training; the data set to be classified establishing module is used for obtaining lightning radiation waveform fragments from the lightning radiation original waveforms transmitted back by the detector through a waveform segmentation method, calculating the lightning energy spectrum of the lightning radiation waveform fragments in a window-dividing manner, determining that the lightning radiation waveform fragments are noise waveforms when the energy distribution of all windows is smaller than a threshold value, discarding the lightning radiation waveform fragments, and taking the undelivered lightning radiation waveform fragments as a waveform set to be classified; the waveform intelligent classification module is used for inputting the data set to be classified into the trained classification model for classification, and obtaining a classification result.
A lightning radiation waveform classification method comprising the steps of:
and step 4, inputting the data set to be classified into the trained classification model for classification, and obtaining a classification result.
The beneficial effects of the invention are as follows:
(1) The convolutional neural network supports direct input of original waveforms for classification, the convolutional operation is performed through the multi-size convolutional kernel to finish the self-adaptive scale feature extraction, the Convolutional Neural Network (CNN) finishes classification through learning the physical features of the waveform core, and training parameters do not need to be repeatedly adjusted on different data sets, so that the method has higher training efficiency compared with the current other classification methods;
(2) The class activation diagram inverts the self-adaptive learned characteristics of the visual model based on the high-dimensional characteristic diagram and the weight coefficient of the full-connection layer, verifies the reliability and the universality of the classification basis, breaks the black box characteristic of the neural network model, and can guide the model to train a specific process;
(3) The Convolutional Neural Network (CNN) model is combined with the class activation method, so that the problems of weak universality and low efficiency in lightning waveform classification are solved, the accurate and efficient classification of lightning radiation waveforms is realized, and the advantages of strong model robustness and transparency and reliability in the classification process are achieved.
Drawings
FIG. 1 is a block diagram of a system of the present invention;
FIG. 2 is a waveform of the original lightning radiation with EMD and normalized;
FIG. 3 is a schematic diagram of four types of lightning radiation waveforms;
FIG. 4 is a schematic diagram of a convolutional neural network structure;
FIG. 5 shows the loss function and the change in accuracy during the training of the four-classification model;
FIG. 6a is a diagram showing the classification result of RS waveform;
FIG. 6b is a diagram showing the visual result of RS waveform classification;
FIG. 7a is a diagram illustrating the classification result of NB waveforms;
FIG. 7b is a diagram showing the result of the classification of NB waveforms according to the visualization;
FIG. 8a is a diagram showing PB waveform classification results;
FIG. 8b is a view of PB waveform classification based on visual results;
FIG. 9a is a diagram showing the classification result of IC waveforms;
Fig. 9b is a diagram showing the result of the classification of the IC waveforms according to the visualization.
The system comprises a 1-training data set establishment module, a 2-convolutional neural network model training module, a 3-data set to be classified establishment module and a 4-waveform intelligent classification module.
Detailed Description
The invention is described in further detail below with reference to the attached drawings and specific examples:
a lightning radiation waveform classification system is shown in figure 1, and comprises a training data set establishment module 1, a convolutional neural network model training module 2, a data set to be classified establishment module 3 and a waveform intelligent classification module 4;
the training data set establishing module 1 is used for obtaining a lightning radiation waveform segment sample through a waveform segmentation method, and taking a corresponding relation sample of the lightning radiation waveform segment sample and a discharge event as a training data set;
the convolutional neural network model training module 2 is used for establishing a convolutional neural network model, performing parameter tuning and training of the convolutional neural network model based on the training data set, and obtaining a classification model after training;
the to-be-classified data set establishing module 3 is used for preprocessing the original lightning radiation waveform transmitted back by the detector through a waveform segmentation method to obtain lightning radiation waveform segments, calculating the lightning energy spectrum of the lightning radiation waveform segments in a window manner, determining that the lightning radiation waveform segments are noise waveforms when the energy distribution of all windows is smaller than a threshold value, discarding the lightning radiation waveform segments, and taking the undelivered lightning radiation waveform segments as to-be-classified waveform sets;
The waveform intelligent classification module 4 is used for inputting the data set to be classified into the trained classification model to classify, obtaining a classification result, and confirming the reliability of the classification result through the calculation model contribution weight value interpretability module.
In the above technical solution, the specific implementation method of the waveform segmentation method in the training data set building module 1 is as follows:
s11, waveform filtering: filtering low-frequency signals in the original lightning radiation waveform by adopting an empirical mode decomposition method to make the waveform straight and the pulse characteristics prominent;
s12, waveform segmentation: cutting the filtered lightning radiation waveform into a plurality of mutually overlapped lightning radiation waveform fragments based on a sliding window method;
s13, normalization processing: projecting the lightning radiation waveform segment into the [ -1,1] interval by adopting a min-max normalization method to obtain a normalized lightning radiation waveform segment;
in this embodiment, the waveform segment of the lightning radiation obtained by waveform division includes 2500 data points, and the time length is 500 μs. FIG. 2 is a diagram of an original lightning radiation waveform and a filtered normalized lightning radiation waveform, wherein the low frequency component is removed, and the pulse component is fully highlighted, which is more beneficial to the subsequent waveform segmentation and recognition.
The training data set creating module 1 creates a training data set by:
and extracting a lightning radiation waveform sample from the normalized lightning radiation waveform fragment based on a criterion method, marking the discharge event type of the lightning radiation waveform sample after manual screening and inspection, and taking a corresponding relation sample of the lightning radiation waveform fragment sample S and the discharge event type marking L as a training data set.
In this embodiment, the waveforms after the preliminary screening are confirmed by the manual discrimination method and then marked, the data set after the manual screening contains 8000 samples, each discharge event category contains 2000 samples, the data set storage format is arff format, and the total number of rows and columns is 8000, wherein 1-2500 columns are data columns, 2501 is a label column, four types of discharge event typical waveforms are shown in fig. 3, the time length of the four types of waveform samples is 500us, and the four types of waveform typical features are shown, and only 0-250us part of waveforms are shown in the figure.
In the above technical solution, the criterion method in the training data set building module 1 is a comprehensive criterion formed by using waveform characteristic parameters, and the waveform characteristic parameters include waveform rising time, waveform falling time, pulse width, waveform forward-backward inverse peak ratio, sub-peak ratio and signal-to-noise ratio; the criterion threshold value is from the statistical result of the normalized lightning radiation waveform segment;
The lightning radiation waveform sample types include back-strike (RS), pre-breakdown (PB), narrow bipolar pulse (NB), cloud flash (IC).
In the above technical solution, the specific implementation method of the empirical mode decomposition method in the training data set building module 1 is as follows:
the empirical mode decomposition method obtains a series of Intrinsic Mode Functions (IMFs) for characterizing signal characteristics by decomposing nonlinear and non-stationary time sequence signals, and each IMF reflects the change rule of the data and has strong self-adaptability.
S11.1, marking local maximum value points and local minimum value points of an original lightning radiation waveform signal x (t) by using the lightning original radiation waveform signal x (t) transmitted by the detector;
s11.2, connecting the local maximum points through cubic spline function interpolation to form an upper envelope line e max (t) connecting the local minimum points to form a lower envelope e min (t) subtracting the mean of the upper envelope and the lower envelope from the original radiation waveform signal x (t) to obtain h 1 (t)：
S11.3, judging h 1 (t) if the IMF component condition is met, if not, replacing x (t) with h 1 (t) and repeating steps S11.1 and S11.2 k times until h 1k (t) the IMF component condition is satisfied, at which time h is noted 1k (t) is IMF component IMF i (t)；
S11.4 subtracting IMF component IMF from said original lightning radiation waveform signal x (t) i (t) repeating steps S11.1, S11.2 and S11.3N times until the termination condition is satisfied; the original lightning radiation waveform signal x (t) is decomposed into a series of IMF components IMF 1 ,imf 2 ,imf 3 …,imf N And residual quantity r n Linear superposition of (t):
s11.5, analyzing the IMF component IMF using Fourier transforms 1 ,imf 2 ,imf 3 …,imf N Removing IMF components with main frequency lower than a set frequency threshold value, namely 1kHZ, and accumulating and recombining the residual IMF components to obtain a filtered lightning radiation waveform.
In the above technical solution, the IMF component conditions are: in the time period of the original radiation waveform signal, the number of any two of the local maximum value point, the local minimum value point and the zero point must be equal or at most differ by one; at any moment of the original radiation waveform signal, the mean value of the upper envelope curve obtained by connecting the local maximum value and the lower envelope curve obtained by connecting the local minimum value is zero;
the termination conditions include two types: the first termination condition is the nth IMF component or the residual quantity r n (t) is less than a set threshold; the second termination condition is the residual quantity r n (t) is a constant or monotonic function, and the embodiment adopts a first termination condition, namely, calculates the standard deviation S of the decomposition results of two adjacent times d Less than 0.3.
In the above technical solution, the specific implementation manner of the sliding window method in step S12 in the training data set building module 1 is as follows:
setting the length of a sliding window to 2500 data points, aligning the time starting point of the sliding window with the time starting point of the filtered lightning radiation waveform, intercepting the lightning radiation waveform segment data in the sliding window, and obtaining a first section of lightning radiation waveform segment sample p 1 Moving the sliding window forwards for a set length, intercepting the lightning radiation waveform segment data in the sliding window, and obtaining a second section of lightning radiation waveform segment sample p 2 Moving the sliding window according to the same forward movement set length and obtaining a corresponding lightning radiation waveform segment sample until the cutting of the filtered lightning radiation waveform is completed, obtaining a lightning radiation waveform segment p 1 ,p 2 ,p 3 ...p m M is the number of lightning radiation waveform segments;
the condition of the forward movement for a set length is to ensure that two adjacent sections of lightning radiation waveform segment samples contain a proper overlapping area; the size of the forward movement set length in this embodiment is half the size of the sliding window.
Projecting the segment to [ -1,1 ] based on a min-max normalization method p' =p/(max (p) -min (p)) [ -1,1 ] ]Within the interval, a normalized waveform segment p is obtained 1 ’,p 2 ’,p 3 ’...p m ’。
In the above technical solution, the lightning radiation waveform segment p m The acquisition method of (1) comprises the following steps:
if the lightning radiation waveform segment p m Is different from the sliding window in time length, when the lightning radiation waveform segment p m Resampling the lightning radiation waveform segment p by nearest neighbor interpolation when the time length of the sliding window is more than or equal to 2/3 of the time length of the sliding window, namely 1500 points m To the length of the sliding window; when the lightning radiation waveform segment p m If the time length of the sliding window is less than 2/3 of the time length of the sliding window, namely 1500 points, discarding the lightning radiation waveform segment p m 。
And dividing each lightning radiation waveform segment according to the width of 100 data points to obtain 25 data frames, calculating the average energy value of each frame of data, wherein the average energy value calculating method is the average value of the square sum of the data point values. If the average value of all frames in the sample is smaller than the threshold value, the sample is a noise waveform, and the sample needs to be discarded. The waveforms to be classified can be saved as arff format through screening, and n rows and 2500 columns are included, wherein n is consistent with the number of the waveforms to be classified, and the 1 st column to the 2500 th column are data columns.
In the above technical solution, the specific implementation method of the convolutional neural network model training module 2 is as follows:
s21, establishing a model: as shown in fig. 4, the input of the convolutional neural network model is 2500×1 of one-dimensional time sequence data, an input layer receives the corresponding relation sample data of the lightning radiation waveform segment sample S and the discharge event type label L, the sample data is convolved by a plurality of special convolutional layers and then enters a global pooling layer to be pooled, finally enters a full-connection layer, and is activated by an activation function, namely a Softmax activation function, so as to obtain the probability that the sample data belongs to the lightning radiation waveform sample type, and the lightning radiation waveform sample type corresponding to the maximum probability is the waveform type L';
s22, model training and testing: dividing the training data set according to a proportion to obtain a training set and a testing set; setting super parameters for model training, deploying the model in a high-computation-power display card for training, and judging whether the model converges or not according to fluctuation and change of the model output accuracy; when the model is converged, the weight of parameters in the model is saved, and whether the model classification performance is balanced or not is tested through indexes such as classification speed, accuracy, precision, recall, F1 coefficient and the like.
The method for judging whether the model converges is that the model converges when the classification accuracy fluctuation in continuous several times of iteration does not exceed a set threshold, in this embodiment, the number of times is preferably 5, and the set threshold is preferably 0.5%.
In the above technical solution, the special convolution layer scans the sample data by using a convolution check to obtain a feature distribution matrix;
the global pooling layer is used for carrying out secondary treatment on the characteristic distribution matrix, so that characteristic dimension is reduced, and calculated amount is reduced;
the special convolution layer adopts maximum value pooling to reserve the most prominent features in the feature distribution matrix;
the output result of the multilayer special convolution layer further comprises a high-dimensional feature map, the high-dimensional feature map is multiplied and summed with a weight matrix in the full-connection layer to obtain a contribution weight value of each data point of the lightning radiation waveform segment sample S to the classification probability, and the visualization of the classification process is realized in a thermodynamic diagram mode;
the multi-layer special convolution layer comprises two convolution blocks, wherein the convolution blocks comprise three layers of convolution layers, and the input of each convolution block is directly connected with the second convolution layer of the corresponding convolution block through a shortcut link, so that the convergence speed is increased, and the degradation phenomenon is avoided;
The convolution layer comprises four size convolution kernels, wherein the four size convolution kernels comprise a 40 multiplied by 1 convolution kernel, a 20 multiplied by 1 convolution kernel, a 10 multiplied by 1 convolution kernel and a 1 multiplied by 1 convolution kernel, input data of the convolution layer respectively enter two links, the first link is the convolution of the input data of the convolution layer through the 1 multiplied by 1 convolution kernel, a first convolution result of the first link is obtained, the first convolution result of the first link is convolved with the 40 multiplied by 1 convolution kernel, a second convolution result of the first link is obtained, the first convolution result of the first link is convolved with the 20 multiplied by 1 convolution kernel, a third convolution result of the first link is obtained, and the first convolution result of the first link is convolved with the 10 multiplied by 1 convolution kernel, and a fourth convolution result of the first link is obtained; and the second link is a pooling result of the second link obtained by the maximum pooling layer with the size of 3 multiplied by 1 of the input data of the convolution layer, the pooling result of the second link is convolved with the convolution kernel of 1 multiplied by 1 to obtain a convolution result of the second link, and the second convolution result of the first link, the third convolution result of the first link, the fourth convolution result of the first link and the convolution result of the second link are serially connected to form output data of the convolution layer.
The high-dimensional characteristic diagram of the output of the multilayer special convolution layer is an n-layer characteristic diagram, wherein the value on a data point x in the k-layer characteristic diagram is f k (x) The method comprises the steps of carrying out a first treatment on the surface of the After the n layers of feature maps pass through the global pooling layer, each feature map is converted into a single value F k Output, since global average pooling is employed, F k Can be calculated as follows:
F k ＝∑ x f k (x)
irrespective of the activation function, one can consider the probability S that for any class C, the waveform belongs to class C C Can be obtained by the following formula:
wherein a is k c The contribution weight of the characteristic diagram representing the k layer to the model decision attribution category C corresponds toFull connection coefficients corresponding to class C in the full connection layer;
thus, a class activation graph for class C can be defined as:
M C (x) And (3) for the contribution weight value (CAM value) of the lightning radiation waveform segment sample S under the category C, the contribution weight value can directly explain the importance degree of any time sequence data point in the lightning radiation waveform segment sample S to the model judging waveform membership category C in a thermodynamic diagram mode.
Comparing the result L' output by the activation function with the discharge event type label L by the model to calculate a loss function, and updating the weight and the bias in the model through back propagation to finish iteration;
After the model is built, the training data set is trained by using the model built in S21, and the training result of the super-parametric waveform is shown in Table 1 in S22.
TABLE 1 lightning radiation waveform four-classification model classification index
FIG. 5 shows the change of the loss function value and the accuracy in the training process of the four-classification lightning radiation waveform classification model, and the model can be converged after 5 iterations, and the final classification accuracy is about 97.8%, at this time, the model training is completed, and the model is saved as a model. Pkl file.
The super-parameters used for model training comprise an Optimizer (Optimizer), a Loss Function (Loss Function), iteration times and a learning rate change strategy.
In the embodiment, the dividing ratio of the training set and the test set in the training data set is 8:2; the optimizer is set as a random gradient descent method (SGD), the loss function is selected as a cross entropy function (cross entropy), the size of training data is 64 parts each time, the training is carried out for 40 times, the forward calculation and the reverse propagation processes are carried out, the initial value of the Learning Rate (Learning Rate) is set to be 0.01, the Learning Rate change is dynamically adjusted in a Cosine annealing (Cosine Decay) mode, and the Learning Rate gradually falls from the initial value of 0.01 to 0.01 and then gradually returns to 0.01 according to the Cosine function change rule in a complete training period; and when the classification accuracy fluctuation in five continuous iterations is not more than 0.5%, the model converges, and the model parameter file with the highest accuracy on the test set is taken and stored, so that the optimal model parameter file is obtained.
The calculation method of the Accuracy Accuracy comprises the following steps:
the calculation method of the Precision rate comprises the following steps:
the Recall rate Recall calculation method comprises the following steps:
the F is 1 The coefficient calculating method comprises the following steps:
wherein:
TP (True Positive): the representative is the actual positive example, and the model is judged to be the positive example;
FP (False Positive): representing that the actual situation is positive, and judging that the model is negative;
TN (True Negative): the representative is the counterexample, and the model is judged to be the counterexample;
FN (False Negative): the representation is actually a counterexample, and the model is judged to be a positive example.
Table 1 shows the classification performance index of the trained four-classification lightning radiation waveform classification model, and the model is visibleHigher classification accuracy can be obtained in four types of RS, PB, IC, NB discharge events, and meanwhile, the Precision, recall and F are high 1 The index values of the coefficients are all close to 1, so that the model has no unbalance problem, and the classification accuracy is true and reliable.
In the technical scheme, a Pytorch1.9.1 framework is adopted for model building, parameter training, tuning and performance optimization, and model training is deployed on a Tesla A100 GPU issued by Inwinda (NVIDIA) company. When the model converges, the classification accuracy can reach 97.8%, and the single waveform classification time is 0.0397s.
In the above technical solution, the specific implementation method of the waveform intelligent classification module 4 is as follows:
s41, inputting waveforms to be classified to obtain classification results: calling a model. Pkl file stored in a classification model S2 which is trained in the convolutional neural network model training module 2, and finishing the classification of the predicted input data by using a prediction function to obtain the waveform category and the corresponding probability given by the model;
s42, calculating a model classification result class activation diagram: and calling a model. Pkl file stored in the classification model S2 which is trained in the convolutional neural network model training module 2, completing waveform visualization by using a calculation_CAM function, and displaying the importance degree of any data point in waveform data to model judgment waveform belonging to a certain discharge event class in a thermodynamic diagram mode.
Fig. 6 to 9 are examples of classification results and class activation map results of typical cases of RS, NB, PB, IC four classes of waveforms. Fig. 6-9 (a) are schematic diagrams of the waveform classification result, wherein RS in the upper text of the legend represents the true category of the sample, max-CAM represents the sample and is visualized by adopting a max-CAM method, and the lower text of the legend represents the probability of the waveform sample from different lightning event categories calculated by the CNN model. 6-9 (b) present thermodynamic diagrams of the weights contributed by each data point in the sample during classification by the model, with a color near red representing the region being more important during classification and a color near blue indicating the region contributing less weight during classification. The legend can be seen from 1) that models in different discharge event waveforms can pay attention to the main pulse part of the waveform, and the model classification is proved to be correct and reasonable according to sources. 2) In different discharge event waveforms, the distribution positions of the part with higher model attention are different, and the core physical process of the discharge event is corresponding. For example, in the RS waveform, the red high-attention portion given by the model is located at the main pulse spike; in the NB waveform, the red high attention part given by the model is distributed on bipolar peak and abrupt change process of polarity; in the PB waveform, the red high-attention part given by the model is positioned in the continuous bipolar part; the red high-interest portion given by the model in the IC waveform is located in the sparser unipolar waveform portion. The distribution corresponds to a key physical process, and the model is proved to have higher universality due to grasping the core characteristics of the waveform.
A lightning radiation waveform classification method comprising the steps of:
and step 4, inputting the data set to be classified into the trained classification model for classification, and obtaining a classification result.
What is not described in detail in this specification is prior art known to those skilled in the art. It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
It should be finally understood that the foregoing embodiments are merely illustrative of the technical solutions of the present invention and not limiting the scope of protection thereof, and although the present invention has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that various changes, modifications or equivalents may be made to the specific embodiments of the invention, and these changes, modifications or equivalents are within the scope of protection of the claims appended hereto.
Claims (11)
1. A lightning radiation waveform classification system, characterized by: the device comprises a training data set establishing module (1), a convolutional neural network model training module (2), a data set to be classified establishing module (3) and a waveform intelligent classifying module (4);
The training data set establishing module (1) is used for obtaining a lightning radiation waveform segment sample through a waveform segmentation method and taking a corresponding relation sample of the lightning radiation waveform segment sample and a discharge event as a training data set;
the convolutional neural network model training module (2) is used for establishing a convolutional neural network model, performing parameter tuning and training of the convolutional neural network model based on the training data set, and obtaining a classification model after training;
the data set to be classified establishing module (3) is used for obtaining lightning wave form fragments in the original lightning wave form transmitted back by the detector through a wave form segmentation method, calculating the lightning energy spectrum of the lightning wave form fragments in a window mode, determining that the lightning wave form fragments are noise wave forms when the energy distribution of all windows is smaller than a threshold value, discarding the lightning wave form fragments, and taking the undelivered lightning wave form fragments as a wave form set to be classified;
the waveform intelligent classification module (4) is used for inputting the data set to be classified into the trained classification model for classification, and obtaining a classification result.
2. The lightning radiation waveform classification system of claim 1, wherein: the specific implementation method of the waveform segmentation method in the training data set building module (1) comprises the following steps:
S11, filtering low-frequency signals in an original lightning radiation waveform by adopting an empirical mode decomposition method;
s12, cutting the filtered lightning radiation waveform into a plurality of mutually overlapped lightning radiation waveform fragments based on a sliding window method;
s13, projecting the lightning radiation waveform segment into the [ -1,1] interval by adopting a min-max normalization method to obtain a normalized lightning radiation waveform segment;
the training data set generating method in the training data set establishing module (1) comprises the following steps:
and extracting a lightning radiation waveform sample from the normalized lightning radiation waveform fragment based on a criterion method, marking the discharge event type of the lightning radiation waveform sample after manual screening and inspection, and taking a corresponding relation sample of the lightning radiation waveform fragment sample S and the discharge event type marking L as a training data set.
3. The lightning radiation waveform classification system of claim 2, wherein:
the criterion method in the training data set establishing module (1) is a comprehensive criterion formed by utilizing waveform characteristic parameters, wherein the waveform characteristic parameters comprise waveform rising time, waveform falling time, pulse width, waveform forward and backward anti-peak ratio, secondary peak ratio and signal to noise ratio; the criterion threshold value is from the statistical result of the normalized lightning radiation waveform segment;
The lightning radiation waveform sample types comprise a back pulse RS, a pre-breakdown PB, a narrow bipolar pulse NB and a cloud flash IC.
4. The lightning radiation waveform classification system of claim 2, wherein: the specific implementation method of the empirical mode decomposition method in the training data set building module (1) is as follows:
s11.1, marking local maximum value points and local minimum value points of an original lightning radiation waveform signal x (t) by using the lightning original radiation waveform signal x (t) transmitted by the detector;
s11.2, connecting the local maximum points through cubic spline function interpolation to form an upper envelope line e max (t) connecting the local minimum points to form a lower envelope e min (t) subtracting the mean of the upper envelope and the lower envelope from the original radiation waveform signal x (t) to obtain h 1 (t)：
S11.3, judging h 1 (t) if the IMF component condition is met, if not, replacing x (t) with h 1 (t) and repeating steps S11.1 and S11.2 k times until h 1k (t) the IMF component condition is satisfied, at which time h is noted 1k (t) is IMF component IMF i (t)；
S11.4 subtracting IMF component IMF from said original lightning radiation waveform signal x (t) i (t) repeating steps S11.1, S11.2 and S11.3N times until the termination condition is satisfied; the original lightning radiation waveform signal x (t) is decomposed into a series of IMF components IMF 1 ,imf 2 ,imf 3 …,imf N And residual quantity r n Linear superposition of (t):
s11.5, analyzing the IMF component IMF using Fourier transforms 1 ,imf 2 ,imf 3 …,imf N Removing IMF components of which the main frequency is lower than a set frequency threshold value, and accumulating and recombining the residual IMF components to obtain a filtered lightning radiation waveform.
5. The lightning radiation waveform classification system of claim 4, wherein:
the IMF component conditions are: in the time period of the original radiation waveform signal, the number of any two of the local maximum value point, the local minimum value point and the zero point must be equal or at most differ by one; at any moment of the original radiation waveform signal, the mean value of the upper envelope curve obtained by connecting the local maximum value and the lower envelope curve obtained by connecting the local minimum value is zero;
the termination conditions include two types: the first termination condition is the Nth IMF component or the residual r N (t) is less than a set threshold; the second termination condition is the residual quantity r N (t) is a constant or monotonic function.
6. The lightning radiation waveform classification system of claim 2, wherein: the specific implementation manner of the sliding window method in step S12 in the training data set building module (1) is as follows:
Aligning the time starting point of the sliding window with the time starting point of the filtered lightning radiation waveform, and intercepting the lightning radiation waveform segment data in the sliding window to obtain a first section of lightning radiation waveform segment sample p 1 Moving the sliding window forwards for a set length, intercepting the lightning radiation waveform segment data in the sliding window, and obtaining a second section of lightning radiation waveform segment sample p 2 Moving the sliding window according to the same forward movement set length and obtaining a corresponding lightning radiation waveform segment sample until the cutting of the filtered lightning radiation waveform is completed, obtaining a lightning radiation waveform segment p 1 ,p 2 ,p 3 ...p m M is the number of lightning radiation waveform segments;
the condition of the forward movement for a set length is to ensure that two adjacent sections of lightning radiation waveform segment samples contain a proper overlapping area;
projecting the segment to [ -1,1 ] based on a min-max normalization method p' =p/(max (p) -min (p)) [ -1,1 ]]Within the interval, a normalized waveform segment p is obtained 1 ’,p 2 ’,p 3 ’...p m ’。
7. A lightning radiation wave form classification system based on claim 6, whereinIn the following steps: the lightning radiation waveform segment p m The acquisition method of (1) comprises the following steps:
if the lightning radiation waveform segment p m Is different from the sliding window in time length, when the lightning radiation waveform segment p m Resampling the lightning radiation waveform segment p by nearest neighbor interpolation when the time length of the sliding window is more than or equal to 2/3 of the time length of the sliding window m To the length of the sliding window; when the lightning radiation waveform segment p m If the time length of the sliding window is less than 2/3 of the time length of the sliding window, discarding the lightning radiation waveform segment p m 。
8. The lightning radiation waveform classification system of claim 1, wherein: the specific implementation method of the convolutional neural network model training module (2) comprises the following steps:
s21, establishing a model: the input layer receives the corresponding relation sample data of the lightning radiation waveform fragment sample S and the discharge event type label L, the sample data is convolved through a plurality of special convolution layers and then enters a global pooling layer to be pooled, finally enters a full-connection layer and is activated by an activation function, so that the probability that the sample data belongs to the lightning radiation waveform sample type is obtained, and the lightning radiation waveform sample type corresponding to the maximum probability is the waveform type L';
s22, model training and testing: dividing the training data set according to a proportion to obtain a training set and a testing set; setting super-parameters used for model training, and judging whether the model converges or not according to fluctuation and change of the model output accuracy; when the model converges, the weight of parameters in the model is saved, and whether the classification performance of the model is balanced or not is tested through indexes such as classification speed, accuracy, precision, recall, F1 coefficient and the like;
The method for judging whether the model converges is that the model converges when the classification accuracy fluctuation in a plurality of continuous iterations does not exceed a set threshold.
9. The lightning radiation waveform classification system of claim 8, wherein:
the special convolution layer scans the sample data by utilizing convolution check to obtain a characteristic distribution matrix;
the global pooling layer is used for carrying out secondary treatment on the characteristic distribution matrix;
the special convolution layer adopts maximum value pooling to reserve the most prominent features in the feature distribution matrix;
the output result of the multi-layer special convolution layer further comprises a high-dimensional feature map, and the high-dimensional feature map is multiplied and summed with a weight matrix in the full-connection layer to obtain a contribution weight value of each data point of the lightning radiation waveform segment sample S to the classification probability;
the multi-layer special convolution layer comprises two convolution blocks, wherein the convolution blocks comprise three layers of convolution layers, and the input of each convolution block is directly connected with the second convolution layer of the convolution block through a shortcut link;
the convolution layer comprises four size convolution kernels, wherein the four size convolution kernels comprise a 40 multiplied by 1 convolution kernel, a 20 multiplied by 1 convolution kernel, a 10 multiplied by 1 convolution kernel and a 1 multiplied by 1 convolution kernel, input data of the convolution layer respectively enter two links, the first link is the convolution of the input data of the convolution layer through the 1 multiplied by 1 convolution kernel, a first convolution result of the first link is obtained, the first convolution result of the first link is convolved with the 40 multiplied by 1 convolution kernel, a second convolution result of the first link is obtained, the first convolution result of the first link is convolved with the 20 multiplied by 1 convolution kernel, a third convolution result of the first link is obtained, and the first convolution result of the first link is convolved with the 10 multiplied by 1 convolution kernel, and a fourth convolution result of the first link is obtained; the second link is a pooling result of the second link obtained by the largest pooling layer with the size of 3 multiplied by 1 of the input data of the convolution layer, the pooling result of the second link is convolved with the convolution kernel of 1 multiplied by 1 to obtain a convolution result of the second link, and the second convolution result of the first link, the third convolution result of the first link, the fourth convolution result of the first link and the convolution result of the second link are serially connected to form output data of the convolution layer;
The high-dimensional characteristic diagram of the output of the multilayer special convolution layer is an n-layer characteristic diagram, wherein the value on a data point x in the k-layer characteristic diagram is f k (x) The method comprises the steps of carrying out a first treatment on the surface of the After the n layers of feature maps pass through the global pooling layer, each feature map is converted into a single value F k Output, F k Can be calculated as follows:
F k ＝∑ x f k (x)
irrespective of the activation function, one can consider the probability S that for any class C, the waveform belongs to class C C Can be obtained by the following formula:
wherein a is k c Representing the contribution weight of the k-th layer feature map to the model decision attribution category C, and corresponding to the full connection coefficient of the corresponding category C in the full connection layer;
thus, a class activation graph for class C can be defined as:
M C (x) A contribution weight value of the lightning radiation waveform segment sample S under the category C; the super-parameters used for model training comprise an Optimizer (Optimizer), a Loss Function (Loss Function), iteration times and a learning rate change strategy;
the calculation method of the accuracy rate comprises the following steps:
the calculation method of the precision rate comprises the following steps:
the calculation method of the recall rate comprises the following steps:
the F is 1 The coefficient calculating method comprises the following steps:
wherein:
TP (True Positive): the representative is the actual positive example, and the model is judged to be the positive example;
FP (False Positive): representing that the actual situation is positive, and judging that the model is negative;
TN (True Negative): the representative is the counterexample, and the model is judged to be the counterexample;
FN (False Negative): the representation is actually a counterexample, and the model is judged to be a positive example.
10. The lightning radiation waveform classification system of claim 1, wherein: the specific implementation method of the waveform intelligent classification module (4) comprises the following steps:
s41, calling a classification model trained in the convolutional neural network model training module (2), and finishing the classification of the predicted input data to obtain the waveform category and the corresponding probability given by the model;
s42, calling the classification model trained in the convolutional neural network model training module (2), completing waveform visualization, and displaying the importance degree of any data point in waveform data to model judging waveform belonging to a certain discharge event class in a thermodynamic diagram mode;
the waveform intelligent classification module (4) is also used for confirming the reliability of classification results through the calculation model contribution weight value interpretability module.
11. A lightning radiation waveform classification method is characterized in that: it comprises the following steps:
step 1, obtaining a lightning radiation waveform segment sample by a waveform segmentation method, and taking a corresponding relation sample of the lightning radiation waveform segment sample and a discharge event as a training data set;
Step 2, a convolutional neural network model is established, parameter tuning and training of the convolutional neural network model are carried out based on the training data set, and a classification model after training is obtained;
step 3, obtaining a lightning wave form segment from a lightning original wave form transmitted back by a detector through a wave form segmentation method, calculating a lightning energy spectrum of the lightning wave form segment in a window manner, determining the lightning wave form segment to be a noise wave form when the energy distribution of all windows is smaller than a threshold value, discarding the lightning wave form segment, and taking the undelivered lightning wave form segment as a wave form set to be classified;
and 4, inputting the data set to be classified into the trained classification model to classify, obtaining a classification result, and confirming the reliability of the classification result through a calculation model contribution weight value interpretability module.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211673315.3A CN116150652A (en) | 2022-12-26 | 2022-12-26 | Lightning radiation waveform classification system and method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211673315.3A CN116150652A (en) | 2022-12-26 | 2022-12-26 | Lightning radiation waveform classification system and method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116150652A true CN116150652A (en) | 2023-05-23 |
Family
ID=86361103
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211673315.3A Pending CN116150652A (en) | 2022-12-26 | 2022-12-26 | Lightning radiation waveform classification system and method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116150652A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117233723A (en) * | 2023-11-14 | 2023-12-15 | 中国电子科技集团公司第二十九研究所 | Radar tracking envelope extraction method based on CNN class activation diagram |
-
2022
- 2022-12-26 CN CN202211673315.3A patent/CN116150652A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117233723A (en) * | 2023-11-14 | 2023-12-15 | 中国电子科技集团公司第二十九研究所 | Radar tracking envelope extraction method based on CNN class activation diagram |
| CN117233723B (en) * | 2023-11-14 | 2024-01-30 | 中国电子科技集团公司第二十九研究所 | Radar tracking envelope extraction method based on CNN class activation diagram |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109271374B (en) | Database health degree scoring method and system based on machine learning | |
| CN106951919A (en) | A kind of flow monitoring implementation method based on confrontation generation network | |
| CN109190665A (en) | A kind of general image classification method and device based on semi-supervised generation confrontation network | |
| CN116229380B (en) | Method for identifying bird species related to bird-related faults of transformer substation | |
| CN116559598B (en) | Smart distribution network fault positioning method and system | |
| CN107132515A (en) | A kind of point mark screening technique constrained based on multidimensional information | |
| CN105630743A (en) | Spectrum wave number selection method | |
| CN109323754A (en) | A kind of train wheel polygon fault diagnosis detection method | |
| CN110515364A (en) | A kind of cutting-tool wear state detection method based on variation mode decomposition and LS-SVM | |
| CN109188502B (en) | Beam position monitor abnormity detection method and device based on self-encoder | |
| CN107818298A (en) | General Raman spectral characteristics extracting method for machine learning material recognition | |
| CN112884134B (en) | Time domain-based convolutional neural network model for seismic facies identification and application | |
| CN116150652A (en) | Lightning radiation waveform classification system and method | |
| CN108090628A (en) | A kind of grain feelings security detection and analysis method based on PSO-LSSVM algorithms | |
| CN102841131A (en) | Intelligent steel cord conveyer belt defect identification method and intelligent steel cord conveyer belt defect identification system | |
| CN112603334B (en) | Spike detection method based on time sequence characteristics and stacked Bi-LSTM network | |
| CN110020637A (en) | A kind of analog circuit intermittent fault diagnostic method based on more granularities cascade forest | |
| CN116340746A (en) | Feature selection method based on random forest improvement | |
| Puspitasari et al. | Partial discharge waveform identification using image with convolutional neural network | |
| CN114662405A (en) | Rock burst prediction method based on few-sample measurement and ensemble learning | |
| CN109829407A (en) | Intelligent lightening recognition method based on convolutional neural networks | |
| CN117351321A (en) | Single-stage lightweight subway lining cavity recognition method and related equipment | |
| CN116340812A (en) | Transformer partial discharge fault mode identification method and system | |
| CN110503132A (en) | The Full wave shape LiDAR point cloud classification method integrated based on m ultiwavelet support vector machines WSVM | |
| CN109829408A (en) | Intelligent lightening recognition device based on convolutional neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |