A Classification Method, Device and Storage Medium for Non-Stationary Time Series Data

Abstract

The present application discloses a classification method, device and storage medium for non-stationary time series data, which are used to improve the accuracy of classifying non-stationary time series data. The classification method for non-stationary time series data disclosed in the present application includes: preprocessing the non-stationary time series data; inputting the preprocessed non-stationary time series data into a preset model; performing a convolution operation to obtain preliminary features of the data; performing a pooling operation to extract stable feature data; inputting the stable feature data into an LSTM network for processing to obtain full connection layer feature parameters; and inputting the full connection layer feature parameters into a full connection layer to confirm the category to which the non-stationary time series data belongs. The present application also provides a classification device and storage medium for non-stationary time series data.

Description

Technical Field

[0001] This application relates to the field of computing technology, and in particular, to a classification method, device, and storage medium for non-stationary time series data. Background Art

[0002] Time series have many application scenarios in fields such as science, nature, and economy. For non-stationary time series data, it is very important to classify it before use. However, in the prior art, the classification of non-stationary time series data is unstable and the classification accuracy is low. Summary of the Invention

[0003] In view of the above technical problems, embodiments of this application provide a classification method, device, and storage medium for non-stationary time series data to improve the stability and accuracy of classifying non-stationary time series data.

[0004] In a first aspect, a classification method for non-stationary time series data provided by an embodiment of this application includes:

[0005] Preprocess the non-stationary time series data;

[0006] Input the preprocessed non-stationary time series data into a preset model;

[0007] Perform a convolution operation to obtain preliminary features of the data;

[0008] Perform a pooling operation to extract stable feature data;

[0009] Input the stable feature data into an LSTM network for processing to obtain feature parameters of the fully connected layer;

[0010] Input the feature parameters of the fully connected layer into the fully connected layer to confirm the category to which the non-stationary time series data belongs.

[0011] Preferably, the preprocessing includes:

[0012] Clean the outliers from the non-stationary time series data;

[0013] Label the non-stationary time series data.

[0014] In the present invention, as a preferred example, the non-stationary time series data includes one or a combination of the following:

[0015] Non-stationary time series data of high-voltage direct current discharge, HVDC;

[0016] Non-stationary time series data of subway stray current interference, RV;

[0017] Time series data related to outliers, OPV;

[0018] Non-stationary time series data NV of natural discharge interference;

[0019] Non-stationary time series data HVAC of high-voltage AC discharge.

[0020] In the present invention, as a preferred example, the preset model includes:

[0021] The preset model is a DCL deep learning neural network;

[0022] The DCL deep learning neural network is based on the deep learning network framework DenseNet and is jointly composed of a convolutional neural network CNN and a long short-term memory neural network LSTM.

[0023] Preferably, the DCL deep learning neural network includes:

[0024] A first part, a second part, and a third part. The output of the first part is the input of the second part, and the output of the second part is the input of the third part;

[0025] Among them, the first part is a 5-layer CNN convolutional layer, including the first convolutional neural network Conv1D-1, the second convolutional neural network Conv1D-2, the third convolutional neural network Conv1D-3, the fourth convolutional neural network Conv1D-4, and the fifth convolutional neural network Conv1D-5;

[0026] The second part is an 8-layer CNN convolutional layer, including the sixth convolutional neural network Conv1D-6, the seventh convolutional neural network Conv1D-7, the eighth convolutional neural network Conv1D-8, the ninth convolutional neural network Conv1D-9, the tenth convolutional neural network Conv1D-10, the eleventh convolutional neural network Conv1D-11, the twelfth convolutional neural network Conv1D-12, and the thirteenth convolutional neural network Conv1D-13;

[0027] The third part is a long short-term memory neural network LSTM.

[0028] Preferably, the second part includes:

[0029] The connection method of the 8-layer CNN convolutional layer of the second part adopts a dense neural network structure;

[0030] The dense neural network structure includes:

[0031] The 8-layer convolutional layers of the second part are connected in sequence, and the output of the previous convolutional layer is used as the input of all subsequent convolutional layers.

[0032] Preferably, in the present invention, the performing of the convolution operation to obtain the preliminary features of the data includes:

[0033] Inputting the non-stationary time-series data of a predetermined length into the first part, specifically including:

[0034] The data enters the Conv1D-1, and after passing through the [1, 2, 1] convolution kernel, convolution is performed in a sliding window manner to enhance the features of the data;

[0035] After the output of the Conv1D-1 enters the Conv1D-2, and after passing through the [2, 1, 1] convolution kernel, convolution is performed in a sliding window manner to enhance the features of the data;

[0036] After the output of the Conv1D-2 enters the Conv1D-3, and after passing through the [1, 1, 2] convolution kernel, convolution is performed in a sliding window manner to enhance the features of the data;

[0037] After the output of the Conv1D-3 enters the Conv1D-4, and after passing through the [1, 0, 1] convolution kernel, convolution is performed in a sliding window manner to eliminate the features of the data, making the sharp parts no longer sharp;

[0038] After the output of the Conv1D-4 enters the Conv1D-5, and after passing through the [1, 1, 0] convolution kernel, convolution is performed in a sliding window manner to eliminate the features of the data, making the sharp parts no longer sharp;

[0039] The output of the Conv1D-5 obtains the preliminary features of the data.

[0040] Preferably, the performing of the pooling operation to extract the stable feature data includes:

[0041] Inputting the output data of the first part into the second part;

[0042] In the second part, the convolution kernels of the sixth to thirteenth convolutional neural networks are respectively: [1, 0], [0, 1], [1, 2], [2, 1], [0.2], [2, 0], [1, 2] and [2, 1].

[0043] Preferably, the convolution processing process of the sixth to thirteenth convolutional neural networks sequentially includes:

[0044] The convolution processing sequentially includes the following 10 steps, where the output of the previous step is the input of the next step;

[0045] First step: Sixth-layer convolution processing, seventh-layer convolution processing, eighth-layer convolution processing, ninth-layer convolution processing, tenth-layer convolution processing, eleventh-layer convolution processing, twelfth-layer convolution processing, thirteenth-layer convolution processing;

[0046] Second step: Sixth-layer convolution processing, eighth-layer convolution processing, tenth-layer convolution processing, twelfth-layer convolution processing;

[0047] Third step: Seventh-layer convolution processing, ninth-layer convolution processing, eleventh-layer convolution processing;

[0048] Fourth step: Sixth-layer convolution processing;

[0049] Fifth step: Ninth-layer convolution processing, twelfth-layer convolution processing;

[0050] Sixth step: Seventh-layer convolution processing, tenth-layer convolution processing, eleventh-layer convolution processing;

[0051] Seventh step: Sixth-layer convolution processing;

[0052] Eighth step: Tenth-layer convolution processing;

[0053] Ninth step: Sixth-layer convolution processing;

[0054] Tenth step: Thirteenth-layer convolution processing.

[0055] Preferably, inputting the fully connected layer feature parameters into the fully connected layer to confirm the category to which the non-stationary time series data belongs includes:

[0056] Inputting the output of the second part into a long short-term memory neural network LSTM to obtain first-category data;

[0057] Determining the category to which the non-stationary time series data belongs according to the first-category data;

[0058] The category to which it belongs includes one of the following:

[0059] Non-stationary time series data of high-voltage DC discharge HVDC;

[0060] Non-stationary time series data of subway stray current interference RV;

[0061] Time series data related to outliers OPV;

[0062] Non-stationary time series data of natural discharge interference NV;

[0063] Non-stationary time series data of high-voltage AC discharge HVAC.

[0064] Using the method provided by the present invention, training is carried out for five types of non-stationary time series data, namely HVDC, HVAC, OPV, RV, and NV. The data are all one-dimensional tensors. First, a one-dimensional CNN is used to enhance the internal structural features of the data. Then, a preset Densenet network structure is used to gradually make the implicit features of the data explicit, and ensure that the data features will not become sharp, and that some unobvious features will not be completely hidden. Finally, LSTM is used to sharpen the previously located features to complete the positioning of the feature segments. Thus, the classification of such time series data is completed.

[0065] In a second aspect, an embodiment of the present application further provides a classification device for non-stationary time series data, including:

[0066] A preprocessing module, configured to preprocess the non-stationary time series data and input the preprocessed non-stationary time series data into a preset model;

[0067] A first module, configured to obtain preliminary features of the data;

[0068] A second module, configured to extract stable feature data;

[0069] A third module, configured to input the stable feature data into an LSTM network for processing to obtain feature parameters of a fully connected layer;

[0070] A confirmation module, configured to input the feature parameters of the fully connected layer into a fully connected layer to confirm the category to which the non-stationary time series data belongs.

[0071] In a third aspect, an embodiment of the present application further provides a classification device for non-stationary time series data, including: a memory, a processor, and a user interface;

[0072] The memory is used to store a computer program;

[0073] The user interface is used to interact with the user;

[0074] The processor is used to read the computer program in the memory. When the processor executes the computer program, the classification method for non-stationary time series data provided by the present invention is implemented.

[0075] In a fourth aspect, an embodiment of the present application further provides a processor-readable storage medium. The processor-readable storage medium stores a computer program, and when the processor executes the computer program, the classification method for non-stationary time series data provided by the present invention is implemented.

[0076] The advantages of the present invention are as follows: By using the method of the present invention, first, a one-dimensional CNN is adopted to enhance the internal structural features of the data. Then, a network structure based on Densenet is used to gradually make the implicit features in the data explicit, and ensure that the data features will not become sharpened, and that some unobvious features will not be completely recessive. In the last step, LSTM is used to sharpen the previously located features to complete the positioning of the feature segments. Thus, the classification of non-stationary time series data is completed, improving the accuracy of the classification of non-stationary time series data, as well as the classification efficiency and compatibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0078] Figure 1 Schematic diagram of the classification method for non-stationary time series data provided by the embodiment of the present application;

[0079] Figure 2 Schematic diagram of the features of non-stationary time series data provided by the embodiment of the present application Figure 1 ;

[0080] Figure 3 Schematic diagram of the features of non-stationary time series data provided by the embodiment of the present application Figure 2 ;

[0081] Figure 4 Schematic diagram of the deep learning network structure and processing flow provided by the embodiment of the present application;

[0082] Figure 5 Schematic diagram of the classification device for non-stationary time series data provided by the embodiment of the present application;

[0083] Figure 6 Schematic diagram of the structure of another classification device for non-stationary time series data provided by the embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0084] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0085] The following explains some terms appearing in the text:

[0086] 1. In the embodiments of the present invention, the term "and / or" describes the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally represents an "or" relationship between the associated objects before and after.

[0087] 2. In the embodiments of the present application, the term "plurality" refers to two or more, and other quantifiers are similar.

[0088] 3. HVDC, the abbreviation of Higher Voltage DC, describes the non-stationary time series data of high-voltage DC discharge. Its characteristic is that the voltage will show regular sudden increases or decreases during a certain period of time, indicating the absence of the outer wall coating at this time.

[0089] 4. RV, the abbreviation of Random Voltage, describes the non-stationary time series data of subway stray current interference. Its characteristic is large fluctuations up and down. It is caused by the discharge of the subway track to the ground in the West-East Gas Pipeline Network around the city, and there happens to be a pipeline nearby, which is an accidental factor causing the discharge.

[0090] 5. OPV, the abbreviation of One Point Voltage, describes the time series data related to outliers. Its characteristic is large fluctuations at a single point, which is the waveform of the time series data generated during equipment failures.

[0091] 6. NV, the abbreviation of Nature Voltage, describes the non-stationary time series data of natural discharge interference. Its characteristic is violent fluctuations at a single point or multiple points. It is caused by natural discharge hitting the ground near the pipeline, and the occurrence probability is extremely low.

[0092] 7. HVAC, the abbreviation of High Voltage AC, describes the non-stationary time series data of high-voltage AC discharge. Its characteristic is periodic violent fluctuations within a certain period of time. It is caused by the discharge of the high-voltage line network near the pipeline and occurs on some fixed lines.

[0093] 8. CNN: Convolutional Neural Network, represented by Conv2D.

[0094] 9. RNN: Recurrent Neural Network.

[0095] 10. LSTM: Long Short-Term Memory Neural Network.

[0096] 11. Densenet: A deep learning network framework with both breadth and depth, which is at the same level as the GoogleNet and Inception architectures.

[0097] 12. Conv2D-1: The first layer of Convolutional Neural Network.

[0098] 13. Conv2D-2: The second convolutional neural network.

[0099] 14. Conv2D-3: The third convolutional neural network.

[0100] 15. Conv2D-4: The fourth convolutional neural network.

[0101] 16. Conv2D-5: The fifth convolutional neural network.

[0102] 17. Conv2D-6: The sixth convolutional neural network.

[0103] 18. Conv2D-7: The seventh convolutional neural network.

[0104] 19. Conv2D-8: The eighth convolutional neural network.

[0105] 20. Conv2D-9: The ninth convolutional neural network.

[0106] 21. Conv2D-10: The tenth convolutional neural network.

[0107] 22. Conv2D-11: The eleventh convolutional neural network.

[0108] 23. Conv2D-12: The twelfth convolutional neural network.

[0109] 24. Conv2D-13: The thirteenth convolutional neural network.

[0110] 25. Sliding window: A method of operating on data. Usually, an empty frame with a length shorter than the original data is selected and placed at the first position of the original data to obtain a set of data. Then it is placed at the second position and another set of data is obtained, and so on, resulting in a series of data with a length equal to the length of the empty frame of the original data.

[0111] 26. Convolution kernel: Usually a sequence segment of the type [1,1], [1,0,1], or [1,2,1]. A convolution kernel with three sequence lengths is a 1*3 convolution kernel, and a convolution kernel with two lengths is a 1*2 convolution kernel.

[0112] For example, in the West-East Gas Pipeline Project, data processing is the core. The potential data monitored by 211 test piles is collected and arranged in chronological order, thus obtaining the following five non-stationary time series data (i.e., HVDC, RV, OPV, NV, and HVAC). The characteristic of non-stationary time series data is that after differencing, the data does not converge, and there is a certain degree of volatility in its internal law. The main reason for such changes is that when the data is monitored, the test piles are connected to each other, and there is a phenomenon of wave superposition. Among the potentials, if multiple situations occur crosswise, the superposition of potentials will be formed. In addition to extreme high-voltage discharges, equipment failures are also among the reasons for outliers. Such outlier data is relatively hidden and difficult to be completely detected. Due to various miscellaneous data interferences caused by unexpected situations, it is often difficult to classify the above five non-stationary time series data. In the prior art, the classification of non-stationary time series data is unstable and the classification accuracy is low.

[0113] Traditional deep learning methods will be adopted for the classification of non-stationary time series data. However, the traditional deep learning method RNN is very unstable when training the time series data of this patent. Only considering the use of multiple layers of CNN is difficult to release the performance. Only considering LSTM to improve the accuracy will result in an unstable classification effect. Compared with the traditional method, the present invention does not adopt RNN as the core of the DCL network because the advantage of RNN lies in the prediction of time series data. RNN has two disadvantages in classification. One is that the memory is short. As the network deepens, the influence of the stored content will gradually weaken, resulting in the loss of some important information. The other is that there is no way to control which information to remember and which to discard, which makes some unimportant information increase the training burden of the network and increase interference.

[0114] In view of the above technical problems, the present invention proposes a classification method based on non-stationary time series data to improve the stability and accuracy of the classification of non-stationary time series data.

[0115] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0116] It should be noted that the display order of the embodiments of the present application only represents the sequence of the embodiments, and does not represent the superiority or inferiority of the technical solutions provided by the embodiments.

[0117] Embodiment 1

[0118] See Figure 1 , a schematic diagram of a classification method for non-stationary time series data provided by the embodiments of the present application, asFigure 1 As shown in the figure, the method includes steps S101 to S103:

[0119] S101. Preprocess the non-stationary time series data;

[0120] As a preferred example, the preprocessing in this step includes:

[0121] Clean the outliers from the non-stationary time series data;

[0122] Annotate the non-stationary time series data.

[0123] In the embodiments of the present invention, it is necessary to preprocess the non-stationary time series data to clarify the characteristics of the target sequence data and ensure that the data annotation is correct. One of the characteristics of non-stationary time series data is the fluctuating changes between data, and its internal law is hidden. However, there is a type of outlier data, such as Figure 2 As shown in the figure, this point is an outlier. In deep learning, the convolution operation often belongs to an interval operation. For such outliers, it is easy to cause errors. That is to say, for Figure 3 the HVDC data shown in the figure, the error of the outlier will result in a 30% probability of misidentifying the wrong data as HVDC data. Therefore, cleaning the outliers from the non-stationary time series data and annotating it can increase the accuracy of data type judgment.

[0124] It should be noted that in the embodiments of the present invention, as a preferred example, the non-stationary time series data includes one or a combination of the following:

[0125] Non-stationary time series data HVDC of high-voltage direct current discharge;

[0126] Non-stationary time series data RV of subway stray current interference;

[0127] Time series data OPV related to outliers;

[0128] Non-stationary time series data NV of natural discharge interference;

[0129] Non-stationary time series data HVAC of high-voltage alternating current discharge.

[0130] That is to say, the non-stationary time series data to be classified includes one of the above 5 types of data or several of them. After the non-stationary time series data classification method provided by the embodiments of the present invention, a specific data is identified and classified as one of the above 5 types of data.

[0131] S102. Input the preprocessed non-stationary time series data into a preset model;

[0132] As a preferred example, the preset model includes:

[0133] The preset model is a DCL deep learning neural network;

[0134] The DCL deep learning neural network is based on the deep learning network framework DenseNet and is jointly composed of a convolutional neural network CNN and a long short-term memory neural network LSTM.

[0135] CNN has certain performance in time series data classification tasks. However, a single network structure is difficult to adapt to various time series data. Only by combining models to learn from each other's strengths can the stability of the network be achieved. But it is necessary to design a best-performing HVDC time series data classification model based on breadth, depth, stability, and versatility. Taking three deep learning methods for processing time series data: LSTM, RNN, and CNN as examples, three combinations can be obtained, namely: RNN+LSTM; CNN+LSTM; CNN+RNN. However, this combination lacks a framework as a foundation and it is difficult to unleash its potential in deep learning. In the embodiments of the present invention, DenseNet is used as the neural network architecture because it is effective in terms of breadth and depth and has higher stability at the same time.

[0136] Based on DenseNet neural network, on the basis of RNN+LSTM, CNN+LSTM, and CNN+RNN, three neural networks are respectively constructed: DRL (RNN+LSTM+DenseNet), DRC (CNN+RNN+DenseNet), and DCL (CNN+LSTM+DenseNet). The performance of DCL exceeds that of DCR and DRL. The DCL adopted in the present invention is used. The DCL deep learning neural network is jointly composed of DenseNet+CNN+LSTM. The DenseNet network is adopted, and its core network is connected by thirteen one-dimensional CNNs. Finally, the parameters output by the CNN are input into the LSTM to obtain the deep learning framework of DCL in the embodiments of the present invention. The composition of DCL is as Figure 4 shown. The innovation of DCL is to use the DenseNet architecture to design and train a neural network for time series data. LSTM is placed at the later stage of the neural network as an output to strengthen the parameter features of the neural network. By using CNN as an intermediate hidden layer, the stability of the network can be greatly increased.

[0137] As Figure 4 shown, the DCL deep learning neural network includes:

[0138] The first part, the second part, and the third part. The output of the first part is the input of the second part, and the output of the second part is the input of the third part;

[0139] Among them, the first part is a 6-layer CNN convolutional layer, including the first convolutional neural network Conv1D-1, the second convolutional neural network Conv1D-2, the third convolutional neural network Conv1D-3, the fourth convolutional neural network Conv1D-4, and the fifth convolutional neural network Conv1D-5;

[0140] The second part is an 8-layer CNN convolutional layer, including the sixth convolutional neural network Conv1D-6, the seventh convolutional neural network Conv1D-7, the eighth convolutional neural network Conv1D-8, the ninth convolutional neural network Conv1D-9, the tenth convolutional neural network Conv1D-10, the eleventh convolutional neural network Conv1D-11, the twelfth convolutional neural network Conv1D-12, and the thirteenth convolutional neural network Conv1D-13;

[0141] The third part is a long short-term memory neural network LSTM.

[0142] S103. Perform a convolution operation to obtain the preliminary features of the data;

[0143] In this step, as described in S102, the first part is composed of convolutional layers of Conv1D. The main purpose is to initially perform a convolution operation on the data to obtain a smaller convolutional layer. First, the HVDC time series data group is used as the input of the model. The convolution kernel is a 1*2 convolution kernel. Through the convolution operations of Conv1D-1 to Conv1D-5, the preliminary extraction of HVDC data features is completed.

[0144] As a preferred example, in this step, performing a convolution operation to obtain the preliminary features of the data includes:

[0145] Input the non-stationary time series data of a predetermined length into the first part, specifically including:

[0146] The data enters Conv1D-1. After passing through a [1, 2, 1] convolution kernel, a sliding window method is used for convolution to achieve feature enhancement of the data;

[0147] After the output of Conv1D-1 enters Conv1D-2, after passing through a [2, 1, 1] convolution kernel, a sliding window method is used for convolution to achieve feature enhancement of the data;

[0148] After the output of Conv1D-2 enters Conv1D-3, after passing through a [1, 1, 2] convolution kernel, a sliding window method is used for convolution to achieve feature enhancement of the data;

[0149] The output of the Conv1D-3 enters the Conv1D-4. After passing through a [1, 0, 1] convolutional kernel, deconvolution is performed in a sliding window manner to eliminate the features of the data, making the sharp parts less sharp.

[0150] The output of the Conv1D-4 enters the Conv1D-5. After passing through a [1, 1, 0] convolutional kernel, deconvolution is performed in a sliding window manner to eliminate the features of the data, making the sharp parts less sharp.

[0151] The output of the Conv1D-5 obtains the preliminary features of the data.

[0152] Among them, the predetermined length is set as needed. Taking the input of a non-stationary time series data column with a length of 740 into the first part as an example, for standardization conversion, after passing a one-dimensional tensor of 1*N (N = 144*5) through the first 5 layers of CNN (i.e., Conv1D-1 to Conv1D-5), the preliminary convolutional features are obtained. The specific process is as follows:

[0153] First, after the data enters the Conv1D-1 and completes the convolution operation, a neural network max pooling process based on a 2*1 matrix with a stride of 1 is used to reconstruct the matrix, obtaining a new tensor. In this method, a process of cleaning with Shannon entropy is added outside the convolutional layer. This cleaning reduces the data width by 8 units and can quantify the discrete features of the data in digital form. The cleaning is performed from left to right and from top to bottom, and the sliding window moves in by 1 unit. Then it enters the Conv1D-2. After completing the convolution operation, a neural network max pooling process based on a 2*1 matrix with a stride of 1 is used to reconstruct the matrix. Then it enters the Conv1D-3. After completing the convolution operation, a neural network max pooling process based on a 2*1 matrix with a stride of 1 is used to reconstruct the matrix. Then it enters the Conv1D-3. After completing the convolution operation, a neural network max pooling process based on a 2*1 matrix with a stride of 1 is used to reconstruct the matrix. Then it enters the Conv1D-4. After completing the convolution operation, a neural network max pooling process based on a 2*1 matrix with a stride of 1 is used to reconstruct the matrix. Then it enters the Conv1D-5. After completing the convolution operation, a neural network max pooling process based on a 2*1 matrix with a stride of 1 is used to reconstruct the matrix. After the preliminary convolution in the first part, the preliminary features of the data are obtained.

[0154] S104. Perform a pooling operation to extract stable feature data;

[0155] Input the output data of the first part into the second part;

[0156] In the second part, the convolutional kernels of the sixth to thirteenth convolutional neural networks are respectively: [1,0], [0,1], [1,2], [2,1], [0.2], [2,0], [1,2] and [2,1].

[0157] That is to say, the second part is composed of 6 layers from Conv1D-6 to Conv1D-13, and the intermediate connection method adopts a dense neural network structure. In the embodiment of the present invention, in the dense neural network structure, the data output from the Conv1D-6 layer will be used by Conv1D-7 to Conv1D-13 respectively, and the data of Conv1D-7 will be used by Conv1D-8 to Conv1D-13 respectively. And so on, the data of Conv1D-8 will be used by Conv1D-9 to Conv1D-13 respectively, the data of Conv1D-9 will be used by Conv1D-10 to Conv1D-13 respectively, the data of Conv1D-10 will be used by Conv1D-11 to Conv1D-13 respectively, the data of Conv1D-11 will be used by Conv1D-12 to Conv1D-13 respectively, and the data of Conv1D-12 will be used by Conv1D-13. Such a network structure is used to simplify the model architecture and improve the model efficiency. Otherwise, more layers of CNN may be required to achieve the same function. For example, each time the JSD operation is performed on the data, the window length is 8 and the sliding window step size is l. The strategy adopted by the input Conv2D-7 to Conv2D-13 is to inherit the parameter data output by the previous layer of neural network. And so on, through the continuous iteration of Conv2D-6 - Conv2D-13, the effective feature segments of the data are finally located.

[0158] It should be noted that since the data is in the Densenet architecture, there is a situation of skip convolution. That is to say, the result after Conv1D-9 runs will be returned to Conv1D-8. Next, a specific processing process of the second part is given below:

[0159] The convolutional processing sequentially includes the following 10 steps, where the output of the previous step is the input of the next step;

[0160] The first step: the sixth layer convolutional processing, the seventh layer convolutional processing, the eighth layer convolutional processing, the ninth layer convolutional processing, the tenth layer convolutional processing, the eleventh layer convolutional processing, the twelfth layer convolutional processing, the thirteenth layer convolutional processing;

[0161] The second step: the sixth layer convolutional processing, the eighth layer convolutional processing, the tenth layer convolutional processing, the twelfth layer convolutional processing;

[0162] The third step: the seventh layer convolutional processing, the ninth layer convolutional processing, the eleventh layer convolutional processing;

[0163] Fourth step: Sixth-layer convolution processing;

[0164] Fifth step: Ninth-layer convolution processing, Twelfth-layer convolution processing;

[0165] Sixth step: Seventh-layer convolution processing, Tenth-layer convolution processing, Eleventh-layer convolution processing;

[0166] Seventh step: Sixth-layer convolution processing;

[0167] Eighth step: Tenth-layer convolution processing;

[0168] Ninth step: Sixth-layer convolution processing;

[0169] Tenth step: Thirteenth-layer convolution processing.

[0170] After the second part, a new sequence after convolution operation is output from the 13-layer neural network. This sequence obtains a simple feature from the five-convolution-layer simple neural network in the first part and a more obvious feature in the Densenet architecture. The above structure is a simple structure designed for data, which is efficient and concise at the same time.

[0171] S105. Input the stable feature data into an LSTM network for processing to obtain fully connected layer feature parameters;

[0172] In this step, the output of the second part is input into the long short-term memory neural network LSTM to obtain the first category of data;

[0173] Determine the category to which the non-stationary time series data belongs according to the first category of data;

[0174] The category to which it belongs includes one of the following:

[0175] Non-stationary time series data of high-voltage DC discharge HVDC;

[0176] Non-stationary time series data of subway stray current interference RV;

[0177] Time series data related to outliers OPV;

[0178] Non-stationary time series data of natural discharge interference NV;

[0179] Non-stationary time series data of high-voltage AC discharge HVAC.

[0180] That is to say, after passing through the second part, through the LSTM, the sharpened attributes of the features are released. For example, the 0s in the convolutional kernel flatten the features so that the features will not be sharpened. After the sharpened features pass through the convolutional operation, they will fall into the state of local optimal solution, which is not conducive to our finding obvious features. That is, in terms of the result, if the convolutional kernel does not perform the flattening operation on the data, it will affect the final classification accuracy.

[0181] S106. Input the feature parameters of the fully connected layer into the fully connected layer to confirm the category to which the non-stationary time series data belongs.

[0182] In this step, the data is input into the fully connected layer, and the category to which the time series data belongs is confirmed through the activation function. That is, the sharpened new sequence obtained after passing through the LSTM is converted into unique category data (i.e., one of the following: non-stationary time series data of HVDC for high-voltage DC discharge; non-stationary time series data of RV for subway stray current interference; non-stationary time series data of OPV related to outliers; non-stationary time series data of NV for natural discharge interference; non-stationary time series data of HVAC for high-voltage AC discharge) after passing through the fully connected layer, thus completing the classification work.

[0183] For the DCL deep learning neural network provided in the embodiments of the present invention, the following Figure 4 gives a specific example:

[0184] In the first 5 layers of the neural network of CNN + LSTM, first input the input data into the LSTM model. The parameter configuration of the LSTM is as follows: the embedding_dim parameter is 27, the hidden_dim parameter is 17, the num_layers parameter is 3, the output_size parameter is 2, the padding parameter is 1. Next, use the word vector conversion method torch.nn.Embedding, and the internal length parameter is configured as 180. Next, put the parameters processed by the previous embedding method into the LSTM model, and the output result is directly used as the input of the first layer of CNN.

[0185] The first part includes 5 layers of CNN convolutional layers, including the first convolutional neural network Conv1D-1, the second convolutional neural network Conv1D-2, the third convolutional neural network Conv1D-3, the fourth convolutional neural network Conv1D-4, and the fifth convolutional neural network Conv1D-5. The parameter configuration is as follows:

[0186] The first layer is Conv1D-1. The configured value of the Max_len parameter is 17, the hidden layer of hidden_dim is configured with 40, the kernel_size parameter is 3, and the bias parameter is True. The first layer of the neural network uses the operation of average pooling. After the CNN completes the first layer of convolution and before accurately entering the second layer of convolution, it is necessary to perform the average pooling operation on the data. Then it enters the second layer Conv1D-2.

[0187] The second layer is Conv1D-2. The configured value of the Max_len parameter is 40, the hidden layer of hidden_dim is configured with 70, the kernel_size parameter is 3, and the bias parameter is True. The maximum pooling is configured with the parameter stride as None, the padding parameter as 0, the ceil_mode parameter as False, and the count_include_pad parameter as True. The second layer of the neural network uses the operation of maximum pooling. After the operation is completed, it will enter the input of the third layer Conv1D-3.

[0188] The third layer is Conv1D-3. The configured value of the Max_len parameter is 70, the hidden layer of hidden_dim is configured with 100, the kernel_size parameter is 3, and the bias parameter is True. The maximum pooling is configured with the parameter stride as None, the padding parameter as 0, the ceil_mode parameter as False, and the count_include_pad parameter as True. The third layer of the neural network uses the operation of maximum pooling. After the operation is completed, it will enter the input of the fourth layer Conv1D-4.

[0189] The fourth layer is Conv1D-4. The configured value of the Max_len parameter is 100, the hidden layer of hidden_dim is configured with 130, the kernel_size parameter is 3, and the bias parameter is True. The maximum pooling is configured with the parameter stride as None, the padding parameter as 0, the ceil_mode parameter as False, and the count_include_pad parameter as True. The fourth layer of the neural network uses the operation of maximum pooling. After the operation is completed, it will enter the input of the fifth layer Conv1D-5.

[0190] The fifth layer is Conv1D-5. The Max_len parameter is configured with a value of 130, the hidden_dim hidden layer is configured with 160, the kernel_size parameter is 3, and the bias parameter is True. The max pooling is configured with the stride parameter being None, the padding parameter being 0, the ceil_mode parameter being False, and the count_include_pad parameter being True. The fourth layer of the neural network uses the max pooling operation. After the operation is completed, it will enter the input of the sixth layer Conv1D-6.

[0191] As a preferred example, after the first part, it can also include: performing a regularization operation on the output data, the dorpout parameter is configured with 0.5, the linear method uses hidden_dim = 160, and output_size is used as the input for the operation. output_size is the length dimension of the data after being output from the fifth layer Conv1D-5; next is the configuration of the activation function. The softmax is used as the activation function, and the dim parameter of the softmax is configured with 1. After the above operations on the data, the data is finally output as a one-dimensional eigenvalue. This value has a value range, and after the data is operated, it will be placed in one of the value ranges. This process is the process of data classification.

[0192] The second part includes 8 CNN convolutional layers, namely the sixth layer convolutional neural network Conv1D-6, the seventh layer convolutional neural network Conv1D-7, the eighth layer convolutional neural network Conv1D-8, the ninth layer convolutional neural network Conv1D-9, the tenth layer convolutional neural network Conv1D-10, the eleventh layer convolutional neural network Conv1D-11, the twelfth layer convolutional neural network Conv1D-12, and the thirteenth layer convolutional neural network Conv1D-13. The specific parameter configurations are as follows:

[0193] The sixth layer is Conv1D-6. The Max_len parameter is configured with a value of 160, the hidden_dim hidden layer is configured with 190, the kernel_size parameter is 3, and the bias parameter is True. The max pooling is configured with the stride parameter being None, the padding parameter being 0, the ceil_mode parameter being False, and the count_include_pad parameter being True. The sixth layer of the neural network uses the max pooling operation. After the operation is completed, it will enter the input of the seventh layer Conv1D-7.

[0194] The seventh layer is Conv1D-7. The configured value of the Max_len parameter is 190, the hidden layer of hidden_dim is configured with 220, the kernel_size parameter is 3, and the bias parameter is True. The max pooling is configured with the stride parameter being None, the padding parameter being 0, the ceil_mode parameter being False, and the count_include_pad parameter being True. The seventh neural network layer uses the max pooling operation. After the operation is completed, it will enter the input of the eighth layer, Conv1D-8.

[0195] The eighth layer is Conv1D-8. The configured value of the Max_len parameter is 220, the hidden layer of hidden_dim is configured with 250, the kernel_size parameter is 3, and the bias parameter is True. The max pooling is configured with the stride parameter being None, the padding parameter being 0, the ceil_mode parameter being False, and the count_include_pad parameter being True. The eighth neural network layer uses the max pooling operation. After the operation is completed, it will enter the input of the ninth layer, Conv1D-9.

[0196] The ninth layer is Conv1D-9. The configured value of the Max_len parameter is 250, the hidden layer of hidden_dim is configured with 280, the kernel_size parameter is 3, and the bias parameter is True. The max pooling is configured with the stride parameter being None, the padding parameter being 0, the ceil_mode parameter being False, and the count_include_pad parameter being True. The ninth neural network layer uses the max pooling operation. After the operation is completed, it will enter the input of the tenth layer, Conv1D-10.

[0197] The tenth layer is Conv1D-10. The configured value of the Max_len parameter is 280, the hidden layer of hidden_dim is configured with 310, the kernel_size parameter is 3, and the bias parameter is True. The max pooling is configured with the stride parameter being None, the padding parameter being 0, the ceil_mode parameter being False, and the count_include_pad parameter being True. The tenth neural network layer uses the max pooling operation. After the operation is completed, it will enter the input of the eleventh layer, Conv1D-11.

[0198] The eleventh layer is Conv1D-11. The configured value of the Max_len parameter is 310, the hidden_dim hidden layer is configured with 340, the kernel_size parameter is 3, and the bias parameter is True. The max pooling is configured with the stride parameter being None, the padding parameter being 0, the ceil_mode parameter being False, and the count_include_pad parameter being True. The eleventh neural network layer adopts the max pooling operation. After the operation is completed, it will enter the input of the twelfth layer Conv1D-12.

[0199] The twelfth layer is Conv1D-12. The configured value of the Max_len parameter is 340, the hidden_dim hidden layer is configured with 370, the kernel_size parameter is 3, and the bias parameter is True. The max pooling is configured with the stride parameter being None, the padding parameter being 0, the ceil_mode parameter being False, and the count_include_pad parameter being True. The eleventh neural network layer adopts the max pooling operation. After the operation is completed, it will enter the input of the thirteenth layer Conv1D-13.

[0200] The thirteenth layer is Conv1D-13. The configured value of the Max_len parameter is 370, the hidden_dim hidden layer is configured with 400, the kernel_size parameter is 3, and the bias parameter is True. The max pooling is configured with the stride parameter being None, the padding parameter being 0, the ceil_mode parameter being False, and the count_include_pad parameter being True. The twelfth neural network layer adopts the max pooling operation. After the operation is completed, it will enter the input of the thirteenth layer again.

[0201] Next, the parameters processed by the neural network embedding method output from the thirteenth layer Conv1D-13 in the previous step are put into the LSTM model for processing.

[0202] First, the model inputs a 13-layer neural network (i.e., a total of 13 layers of neural network composed of the first part and the second part) for training, and then enters the LSTM. After going through one cycle, a total of 14 layers are passed through (i.e., 13 layers of Conv1D plus 1 layer of LSTM, a total of 14 layers). The following uses the numbers 1 to 13 to represent the first layer of neural network Conv1D-1 to the thirteenth layer of neural network Conv1D-13 for illustration. That is, first enter from 1 and come out from 13, and then enter the LSTM. After the LSTM, 14 layers of operation are completed;

[0203] After passing through 13 layers, the data re-enters the 6th layer, then enters the 8th layer after the 6th layer, the 8th layer enters the 10th layer, the 10th layer enters the 12th layer, and the 12th layer enters the 7th layer. Finally, the 10th layer enters the 13th layer. By using Densenet, the neural network can be folded to improve utilization. A neural network with a depth of 13 layers achieves the effect of a 35-layer network. The specific sequence is as follows (M→N means that after the output of the Mth layer of the neural network, it enters the Nth layer of the neural network, and both M and N are integers from 1 to 13):

[0204] 13→6→8→10→12;

[0205] 12→7→9→11;

[0206] 11→6;

[0207] 6→12;

[0208] 12→7→11;

[0209] 11→6;

[0210] 6→10;

[0211] 10→13;

[0212] After completing the above operations, the output data is regularized. The dorpout parameter is configured to 0.5, and the linear method uses hidden_dim = 160 and output_size as inputs for calculation. The output_size is the length dimension of the data after output from the fifth layer of the CNN. Next is the configuration of the activation function. The softmax function is used as the activation function, and the dim parameter of the softmax is configured to 1. After the above data calculations, we finally output the data as a one-dimensional eigenvalue, thus completing the classification of non-stationary time series data.

[0213] Using the method of the present invention, first, one-dimensional CNN is used to improve the internal structural features of the data. Then, a network structure based on Densenet is used to gradually make the implicit features of the data explicit and ensure that the data features do not become sharpened and that some unobvious features do not completely become implicit. In the last step, LSTM is used to sharpen the previously located features to complete the positioning of the feature segments. Thus, the classification of non-stationary time series data is completed, improving the accuracy of the classification of non-stationary time series data, as well as the classification efficiency and compatibility.

[0214] Embodiment 2

[0215] Based on the same inventive concept, the embodiment of the present invention also provides a classification device for non-stationary time series data, as Figure 5 shown. The device includes:

[0216] The preprocessing module 501 is configured to preprocess the non-stationary time series data and input the preprocessed non-stationary time series data into a preset model;

[0217] The first module 502 is configured to obtain preliminary features of the data;

[0218] The second module 503 is configured to extract stable feature data;

[0219] The third module 504 is configured to input the stable feature data into an LSTM network for processing to obtain fully connected layer feature parameters;

[0220] The confirmation module 505 is configured to input the fully connected layer feature parameters into a fully connected layer to confirm the category to which the non-stationary time series data belongs.

[0221] As a preferred example, the preprocessing module 501 is further configured to perform preprocessing according to the following steps:

[0222] Clean the outliers from the non-stationary time series data;

[0223] Label the non-stationary time series data.

[0224] The non-stationary time series data includes one or a combination of the following:

[0225] Non-stationary time series data of HVDC for high-voltage DC discharge;

[0226] Non-stationary time series data of RV for subway stray current interference;

[0227] Time series data related to outliers, OPV;

[0228] Non-stationary time series data of NV for natural discharge interference;

[0229] Non-stationary time series data of HVAC for high-voltage AC discharge.

[0230] The preset model includes:

[0231] The preset model is a DCL deep learning neural network;

[0232] The DCL deep learning neural network is based on the deep learning network framework DenseNet and is jointly composed of a convolutional neural network CNN and a long short-term memory neural network LSTM. That is:

[0233] The DCL deep learning neural network includes:

[0234] The first part, the second part, and the third part, where the output of the first part is the input of the second part, and the output of the second part is the input of the third part;

[0235] Among them, the first part is a 5-layer CNN convolutional layer, including the first convolutional neural network Conv1D-1, the second convolutional neural network Conv1D-2, the third convolutional neural network Conv1D-3, the fourth convolutional neural network Conv1D-4, and the fifth convolutional neural network Conv1D-5;

[0236] The second part is an 8-layer CNN convolutional layer, including the sixth convolutional neural network Conv1D-6, the seventh convolutional neural network Conv1D-7, the eighth convolutional neural network Conv1D-8, the ninth convolutional neural network Conv1D-9, the tenth convolutional neural network Conv1D-10, the eleventh convolutional neural network Conv1D-11, the twelfth convolutional neural network Conv1D-12, and the thirteenth convolutional neural network Conv1D-13;

[0237] The third part is a long short-term memory neural network LSTM.

[0238] The second part includes:

[0239] The connection method of the 8-layer CNN convolutional layer of the second part adopts a dense neural network structure;

[0240] The dense neural network structure includes:

[0241] The 8 convolutional layers of the second part are connected in sequence, and the output of the previous convolutional layer is used as the input of all subsequent convolutional layers.

[0242] Among them, the composition of DCL is the same as that in the first embodiment, and will not be elaborated in this embodiment.

[0243] As a preferred example, the first module 502 is further configured to obtain the preliminary features of the data in the following manner:

[0244] Input the non-stationary time series data of a predetermined length into the first part, specifically including:

[0245] The data enters the Conv1D-1, and after passing through a [1, 2, 1] convolution kernel, deconvolution is performed in a sliding window manner to achieve data feature enhancement;

[0246] After the output of the Conv1D-1, it enters the Conv1D-2, and after passing through a [2, 1, 1] convolution kernel, deconvolution is performed in a sliding window manner to achieve data feature enhancement;

[0247] The output of the Conv1D-2 enters the Conv1D-3. After passing through a convolutional kernel of [1, 1, 2], convolution is performed in a sliding window manner to enhance the features of the data.

[0248] The output of the Conv1D-3 enters the Conv1D-4. After passing through a convolutional kernel of [1, 0, 1], convolution is performed in a sliding window manner to eliminate the features of the data, making the sharp parts less sharp.

[0249] The output of the Conv1D-4 enters the Conv1D-5. After passing through a convolutional kernel of [1, 1, 0], convolution is performed in a sliding window manner to eliminate the features of the data, making the sharp parts less sharp.

[0250] The output of the Conv1D-5 obtains the preliminary features of the data.

[0251] As a preferred example, the second module 503 is further configured to extract stable feature data according to the following method:

[0252] Input the output data of the first part into the second part.

[0253] In the second part, the convolutional kernels of the sixth to thirteenth convolutional neural networks are respectively: [1, 0], [0, 1], [1, 2], [2, 1], [0.2], [2, 0], [1, 2], and [2, 1].

[0254] It should be noted that the preset model in the second embodiment is the same as the DCL model in the first embodiment, and will not be elaborated here.

[0255] It should be noted that the preprocessing module 501 provided in this embodiment can achieve all the functions included in step S101 in the first embodiment, solve the same technical problems, and achieve the same technical effects, and will not be elaborated here;

[0256] It should be noted that the first module 502 provided in this embodiment can achieve all the functions included in step S102 in the first embodiment, solve the same technical problems, and achieve the same technical effects, and will not be elaborated here;

[0257] It should be noted that the second module 503 provided in this embodiment can achieve all the functions included in step S103 in the first embodiment, solve the same technical problems, and achieve the same technical effects, and will not be elaborated here;

[0258] It should be noted that the third module 504 provided in this embodiment can achieve all the functions included in step S104 in the first embodiment, solve the same technical problems, and achieve the same technical effects, and will not be elaborated here;

[0259] It should be noted that the confirmation module 505 provided in this embodiment can implement all the functions included in step S105 in Embodiment 1, solve the same technical problems, and achieve the same technical effects, which will not be elaborated here;

[0260] It should be noted that the device provided in Embodiment 2 and the method provided in Embodiment 1 belong to the same inventive concept, solve the same technical problems, and achieve the same technical effects. The device provided in Embodiment 2 can implement all the methods in Embodiment 1, and the same parts will not be elaborated.

[0261] Embodiment 3

[0262] Based on the same inventive concept, an embodiment of the present invention further provides a classification device for non-stationary time series data, as Figure 6 shown, the device includes:

[0263] including a memory 602, a processor 601, and a user interface 603;

[0264] The memory 602 is used to store computer programs;

[0265] The user interface 603 is used to interact with users;

[0266] The processor 601 is used to read the computer program in the memory 602. When the processor 601 executes the computer program, it realizes:

[0267] Preprocess the non-stationary time series data;

[0268] Input the preprocessed non-stationary time series data into a preset model;

[0269] Perform a convolution operation to obtain preliminary features of the data;

[0270] Perform a pooling operation to extract stable feature data;

[0271] Input the stable feature data into an LSTM network for processing to obtain feature parameters of the fully connected layer;

[0272] Input the feature parameters of the fully connected layer into the fully connected layer to confirm the category to which the non-stationary time series data belongs.

[0273] Among them, in Figure 6Among them, the bus architecture may include any number of interconnected buses and bridges, specifically, various circuits represented by one or more processors represented by processor 601 and a memory represented by memory 602 are linked together. The bus architecture may also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, etc., which are well known in the art, and thus will not be further described herein. The bus interface provides an interface. Processor 601 is responsible for managing the bus architecture and general processing, and memory 602 may store data used by processor 501 when performing operations.

[0274] Processor 601 may be a CPU, ASIC, FPGA, or CPLD, and processor 601 may also adopt a multi-core architecture.

[0275] When processor 601 executes the computer program stored in memory 602, it implements any of the classification methods for non-stationary time series data in Embodiment 1.

[0276] It should be noted that the preset model in Embodiment 3 is the same as the DCL model in Embodiment 1, and will not be elaborated herein.

[0277] It should be noted that the device provided in Embodiment 3 and the method provided in Embodiment 1 belong to the same inventive concept, solve the same technical problems, and achieve the same technical effects. The device provided in Embodiment 3 can implement all the methods in Embodiment 1, and the same parts will not be elaborated.

[0278] This application also proposes a processor-readable storage medium. Among them, this processor-readable storage medium stores a computer program, and when the processor executes the computer program, it implements any of the classification methods for non-stationary time series data in Embodiment 1.

[0279] It should be noted that the division of units in the embodiments of this application is illustrative, only a logical function division, and there may be other division methods in actual implementation. In addition, in each embodiment of this application, each functional unit may be integrated in a processing unit, or each unit may exist physically alone, or two or more units may be integrated in one unit. The above integrated units may be implemented in the form of hardware or in the form of software functional units.

[0280] Those skilled in the art should understand that the embodiments of this application can be provided as a method, a system, or a computer program product. Therefore, this application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories and optical memories, etc.) containing computer-usable program codes.

[0281] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing device produce a means for implementing the specified functions in one or more of the flows Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the specified functions in one or more of the blocks.

[0282] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction means for implementing the specified functions in one or more of the flows Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the specified functions in one or more of the blocks.

[0283] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalent technologies, this application is also intended to include these modifications and variations.

Claims

1. A classification method for non-stationary time series data, characterized in that, Including: Preprocessing the non-stationary time series data; Inputting the preprocessed non-stationary time series data into a preset model; Performing a convolution operation to obtain preliminary features of the data; Performing a pooling operation to extract stable feature data; Inputting the stable feature data into an LSTM network for processing to obtain feature parameters of the fully connected layer; Inputting the feature parameters of the fully connected layer into the fully connected layer to confirm the category to which the non-stationary time series data belongs; The preprocessing includes: Cleaning outliers from the non-stationary time series data; Labeling the non-stationary time series data; The non-stationary time series data includes one or a combination of the following: Non-stationary time series data of high-voltage DC discharge HVDC; Non-stationary time series data of subway stray current interference RV; Time series data related to outliers OPV; Non-stationary time series data of natural discharge interference NV; Non-stationary time series data of high-voltage AC discharge HVAC; The preset model includes: The preset model is a DCL deep learning neural network; The DCL deep learning neural network is based on the deep learning network framework DenseNet and is jointly composed of a convolutional neural network CNN and a long short-term memory neural network LSTM.

2. The method according to claim 1, characterized in that, The DCL deep learning neural network includes: A first part, a second part, and a third part. The output of the first part is the input of the second part, and the output of the second part is the input of the third part; Among them, the first part is a 5-layer CNN convolutional layer, including the first convolutional neural network Conv1D-1, the second convolutional neural network Conv1D-2, the third convolutional neural network Conv1D-3, the fourth convolutional neural network Conv1D-4, and the fifth convolutional neural network Conv1D-5; The second part is an 8-layer CNN convolutional layer, including the sixth convolutional neural network Conv1D-6, the seventh convolutional neural network Conv1D-7, the eighth convolutional neural network Conv1D-8, the ninth convolutional neural network Conv1D-9, the tenth convolutional neural network Conv1D-10, the eleventh convolutional neural network Conv1D-11, the twelfth convolutional neural network Conv1D-12, and the thirteenth convolutional neural network Conv1D-13; The third part is a long short-term memory neural network LSTM.

3. The method according to claim 2, characterized in that, The second part includes: The connection method of the 8-layer CNN convolutional layer of the second part adopts a dense neural network structure; The dense neural network structure includes: The 8-layer convolutional layers of the second part are connected in sequence, and the output of the previous convolutional layer is used as the input of all subsequent convolutional layers.

4. The method according to claim 2, characterized in that, The performing a convolution operation to obtain preliminary features of the data includes: Inputting non-stationary time series data of a predetermined length into the first part, specifically including: The data enters the Conv1D-1, and after passing through a [1, 2, 1] convolutional kernel, a sliding window is used for convolution to achieve feature enhancement of the data; The output of the Conv1D-1 enters the Conv1D-2. After passing through a [2,1,1] convolutional kernel, deconvolution is performed in a sliding window manner to enhance the features of the data. The output of the Conv1D-2 enters the Conv1D-3. After passing through a [1,1,2] convolutional kernel, deconvolution is performed in a sliding window manner to enhance the features of the data. The output of the Conv1D-3 enters the Conv1D-4. After passing through a [1,0,1] convolutional kernel, deconvolution is performed in a sliding window manner to eliminate the features of the data, making the sharp parts less sharp. The output of the Conv1D-4 enters the Conv1D-5. After passing through a [1,1,0] convolutional kernel, deconvolution is performed in a sliding window manner to eliminate the features of the data, making the sharp parts less sharp. The output of the Conv1D-5 obtains the preliminary features of the data.

5. The method according to claim 2, wherein The pooling operation is performed to extract stable feature data, including: The output data of the first part is input into the second part. In the second part, the convolutional kernels of the sixth to thirteenth convolutional neural networks are respectively: [1,0], [0,1], [1,2], [2,1], [0.2], [2,0], [1,2], and [2,1].

6. The method according to claim 2, characterized in that, The convolutional processing process of the sixth to thirteenth convolutional neural networks successively includes: The convolutional processing successively includes the following 10 steps, where the output of the previous step is the input of the next step. The first step: the sixth layer convolutional processing, the seventh layer convolutional processing, the eighth layer convolutional processing, the ninth layer convolutional processing, the tenth layer convolutional processing, the eleventh layer convolutional processing, the twelfth layer convolutional processing, the thirteenth layer convolutional processing. The second step: the sixth layer convolutional processing, the eighth layer convolutional processing, the tenth layer convolutional processing, the twelfth layer convolutional processing. The third step: the seventh layer convolutional processing, the ninth layer convolutional processing, the eleventh layer convolutional processing. The fourth step: the sixth layer convolutional processing. The fifth step: the ninth layer convolutional processing, the twelfth layer convolutional processing. The sixth step: the seventh layer convolutional processing, the tenth layer convolutional processing, the eleventh layer convolutional processing. The seventh step: the sixth layer convolutional processing. The eighth step: the tenth layer convolutional processing. The ninth step: the sixth layer convolutional processing. The tenth step: the thirteenth layer convolutional processing.

7. The method according to claim 2, characterized in that, The feature parameters of the fully connected layer are input into the fully connected layer to confirm the categories to which the non-stationary time series data belongs, including: The output of the second part is input into the long short-term memory neural network LSTM to obtain the first category data. The category to which the non-stationary time series data belongs is determined according to the first category data. The categories include one of the following: The non-stationary time series data of high-voltage DC discharge, HVDC. The non-stationary time series data of subway stray current interference, RV. The time series data related to outliers, OPV. The non-stationary time series data of natural discharge interference, NV. The non-stationary time series data of high-voltage AC discharge, HVAC.

8. A classification device for non-stationary time series data, characterized in that, including: A preprocessing module, configured to preprocess the non-stationary time series data and input the preprocessed non-stationary time series data into a preset model. The first module is configured to obtain preliminary features of the data; The second module is configured to extract stable feature data; The third module is configured to input the stable feature data into an LSTM network for processing to obtain fully connected layer feature parameters; The confirmation module is configured to input the fully connected layer feature parameters into a fully connected layer to confirm the category to which the non-stationary time series data belongs; The preset model includes: The preset model is a DCL deep learning neural network; The DCL deep learning neural network is based on the deep learning network framework DenseNet and is jointly composed of a convolutional neural network CNN and a long short-term memory neural network LSTM; The preprocessing includes: Performing outlier cleaning on the non-stationary time series data; Annotating the non-stationary time series data; The non-stationary time series data includes one or a combination of the following: Non-stationary time series data HVDC of high-voltage DC discharge; Non-stationary time series data RV of subway stray current interference; Time series data OPV related to outliers; Non-stationary time series data NV of natural discharge interference; Non-stationary time series data HVAC of high-voltage AC discharge.

9. A classification device for non-stationary time series data, characterized in that It includes a memory, a processor, and a user interface; The memory is used to store computer programs; The user interface is used to interact with the user; the processor is used to read the computer program in the memory. When the processor executes the computer program, the classification method of the non-stationary time series data as described in any one of claims 1 to 7 is implemented.

Images