SRPM–CNN: a combined model based on slide relative position matrix and CNN for time series classification

Research on the time series classification is gaining an increased attention in the machine learning and data mining areas due to the existence of the time series data almost everywhere, especially in our daily work and life. Recent studies have shown that the convolutional neural networks (CNN) can extract good features from the images and texts, but it often encounters the problem of low accuracy, when it is directly employed to solve the problem of time series classification. In this pursuit, the present study envisaged a novel combined model based on the slide relative position matrix and CNN for time series. The proposed model first adopted the slide relative position for converting the time series data into 2D images during preprocessing, and then employed CNN to classify these images. This made the best of the temporal sequence characteristic of time series data, thereby utilizing the advantages of CNN in image recognition. Finally, 14 UCR time series datasets were chosen to evaluate the performance of the proposed model, whose results indicate that the accuracy of the proposed model was higher than others.


Introduction
The objective things develop continuously with the passage of time, and time is an important feature that cannot be ignored in the process of changing things. Information technology is growing rapidly in recent years, and more and more data are being recorded in their respective fields, including astronomy, electronic commerce, government affairs, finance, medicine, transportation, etc. [1]. Most of these data represent the real value of something that changes and grows over time, namely time series data. Therefore, the time series data mining is becoming increasingly prevalent and gaining an importance in the data mining field. Time series classification is a key issue in time series data mining, and is different from the traditional classification due to the values of the individual feature in sequence. Therefore, the order of data often affects the results of the time series classification [2].
The aim of time series classification is to get a tag for a time series by training a model [3]. Time series classification can assist in obtaining the implicit law hidden behind the data for supporting management decision. For instance, the human behavior can be identified by time series classification model, which can be then adopted to predict target behavior [4,5]. During decades, a large number of models have been developed for the time series classification [6] and performed good results on the UCR data sets, including the Collective of Transformation Ensembles (COTE) [7,8], Bag of SFA Symbol (BOSS) [9], 1NN-DTW [10], etc. Although the 1-Nearest Neighbor (1-NN) algorithm can provide classification outcome and achieves a good performance, it ignores the internal features of the time series. Moreover, the accuracy of these traditional models for time series classification is not good enough, when the dimensions of the time series data are large and the values of data change rapidly. Owing to its greatest strength of extraction of the features from data without any preprocessor, deep learning models, especially the convolutional neural networks (CNN), have become a practical and effective method used widely in many fields, such as image classification, target detection and natural language processing (NLP) [11,12]. Besides, considering the local and global characteristics of the data, the weights of CNN can be shared to reduce the quantity of the parameters. As a result, the extracted features are closer to the actual situation and can achieve remarkable results in many tasks, which also provides a new way to improve the precision of time series classification. However, in the existing literature on the time series classification with CNN, the original time series data are input directly into the network, which ignores the differences and related information among the different points in the same time series. Meanwhile, only deepening the depth of network helps to increase the consumption of the computing resources, and it does not improve the accuracy of the CNN.
Therefore, a combined model based on the slide relative position matrix and CNN (SRPM-CNN) for time series classification is proposed in the present study. Standard deviation was utilized to extract the trend of the time series data, which not only reduced the dimension of the time series data, but also amplified the fluctuation characteristics. Specifically, the temporal characteristics of the time series data were considered and the original data of time series was converted into images using the slide relative position matrix. This removed some of the redundant features and reduced the dimension of time series data, which minimized the computing time. The converted images were input into the fully convolutional neural network for classification, and the results showed that the accuracy was higher than the relative position matrix with CNN (RPMCNN) [13] in some datasets.
The contributions of this study can be described in three aspects: first, the trend extraction of the time series further enlarged the fluctuation of time series data and obtained the data features that were conducive to the classification results. Second, original time series data were converted into images with the sliding relative position matrix for removing the redundant information and optimizing the relative position matrix. Another important contribution was to construct the CNN, which greatly decreased the number of the network parameters and achieved better results compared with the CNN with the full connection layers.
The research work is organized as follows: In Sect. 2, we have discussed the related work and background. The overall framework of the proposed method has been presented in Sect. 3. Our experimental outcomes are discussed in Sect. 4. In Sect. 5, we have presented the conclusions and future work.

Related work
Due to the values of the time series data that always vary with time, the key issue of the time series classification is to find the changing law of data. Normally, the extraction of the subsequences from original time series data is done, to obtain their changing rule and then classify these subsequences using the shapelets. Li [14] adopted the standard deviation array, to enlarge the fluctuation of the time series and isolate the optimal shapelets through the key points in the standard deviation array, and then put these optimal shapelets into the SVM (Support Vector Machine) classifier for better classification results. However, the Pruning shapelets with key points ignored the relationship among the different key points, and then extracted and isolated the optimal subsequences that lost some continuity in the information of the original data. Thus, it is essential to pay more attention to the sequence characteristics of the time series data. The traditional classification methods usually put the original data into the classifier after complex preprocessing, but recently, the convolution calculation can identify and extract the key features automatically from the original data and achieve remarkable results for the time series classification.
Since 1995, CNN has attracted the attention of the academic community [15]. Thanks to the advantage of the extracting features, CNN has made a remarkable progress in the speech and image classification, target detection, human behavior recognition and other fields in the last several years. Efforts have been undertaken to apply CNN in the time series classification. Obviously, the structure of CNN is a special structure that is completely different from the artificial neural network (ANN), which comes from the interaction between the neurons in the human brain and can transmit information to each other. The study on time series classification with CNN can be divided into two categories according to the differences on the input data, which are 1D time series data after preprocessing and 2D images transformed from the original time series.
For 1D time series data, the original time series data are input into the Deep Neural Networks (DNN) models for obtaining the results without any preprocessing process. Three DNN models, including the Multilayer Perceptron (MLP) [17], Fully Convolutional Networks (FCN) [16], and Residual Networks (ResNet) [18] are able to classify 1D time series data. Specifically, FCN is remarkable owing to its simple protocol and provides a novel thought for the future research. However, it has a limitation for extracting features only from the time domain and ignoring other factors. To observe the different features of time series data from different perspectives, the multi-scale convolutional neural networks (MCNN) [19] have achieved varied conversion methods of original time series data and classified the one-dimension time series data. The multi-scale features are extracted by one-dimensional convolution, and then combined into the fully convolution network for classification. The long time series data not only lead to low classification accuracy, but also increase some invalid information and the memory consumption.
For 2D images, the original time series data are converted into the 2D images by different transformation methods, and then CNN is employed to classify these images. For instance, Hatami proposed a transformation method based on the Recurrent Plot (RP) [23] to represent the time series data as two-dimensional images. In his study, time series was defined as a number of different periodic behaviors, and the main idea of the RP method was that the trajectory of points appeared already would return to the previous state. But the structure of the CNN in his study consisted two convolution layers and two fully connected layers, which was too simple to affect the accuracy. Wang and Oates paid more attention to the volatility of the time series, and extracted the static and dynamic information. The Gramian Angular Field (GAF) transformed the original time series data into polar coordinates and calculated the transfer probability along the time axis with the Markov Transition Field (MTF). Therefore, they used the GAF-MTF method [20,21] to convert the time series data into images and made the machine identify it on the vision. In this method, time series data was rescaled through the Min-Max normalization, so that all the values fell into the range of (− 1, 1) and then represented the rescaled time series in the polar coordinates before the image conversion phase using the Gramian Angular Field (GAF). The time series was discretized into quantile bins and then its Markov Transition Matrix was calculated for building the Markov Transition Field (MTF) images. The tiled CNN was trained on the compound GAF-MTF images and the results were obtained [22]. The precision of this classification method was not high enough to compete with the other advanced methods. It is for this reason that Chen [13] constructed the relative position matrix for transforming the time series data into grayscale images and input these images into the CNN model based on VGGNet [24]. His method contained five convolution layers and three full connection layers for higher accuracy, but it produced much redundant information and increased the consumption of the computing resources in the process of the transformation matrix. The CNN architecture with a certain complexity improved the accuracy of classification. In recent years, some classical CNN models, such as VGGNet [24], AlexNet [25], ResNet [18] and GoogLeNet [26] have demonstrated a good performance in the image recognition. The earliest proto structure of CNN was the Lenet-5 proposed by Lecun [27]. These classical networks had five convolutional parts and three full connection layers, and each convolutional part had multiple convolutional layers. Especially, the VGGNet demonstrated better generalization ability for the different tasks.
Therefore, the present work envisaged the development of a combined model based on the slide relative position matrix and CNN for classifying the time series that exhibited a same size and number of convolution kernels as VGG-Net and removed the full connection layers for reducing the computational complexity. In the proposed method, standard deviation was adopted to extract the trend features from the time series data, which made the temporal fluctuation characteristics more obvious. Subsequently, the slide relative position matrix was constructed to transform the preprocessed time series data into images. Finally, the CNN without full connection layer was trained on these images and results were analyzed.

Methodology
The overall architecture of the proposed SRPM-CNN model is shown in Fig. 1, which contains three sequential phases, trend extraction phase, image representation phase and convolution network phase.
In the trend extraction phase, a standard deviation window was set to calculate the standard deviation of time stamp, which reduced the dimension of the time series data and extracted the fluctuation trend of data.
In the image representation phase, a slide relative position matrix was employed to convert the preprocessed time series data into 2D image representation.
In the convolution phase, based on the VGGNet without fully connected layer CNN was constructed, which contains

Trend extraction
The standard deviation window is illustrated by Fig. 2. The standard deviation of these points in standard deviation window represents the standard deviation of the center point in window, and the size of window is denoted as w. Assuming the w = 5, the value of the first point in standard deviation array will be the standard deviation of five points centered on the third point in original array.
A time series is denoted as T = t 1 , t 2 , ⋯ ⋯ , t n , the value of the ith time stamp is denoted as t i and n is the number of points in this time series. First, an appropriate standard deviation window w should be determined for computing the standard deviation of points in the same window. The first standard deviation array can be defined as Eq. (1). where − t i is the average of points centered on t i point in the standard deviation window, which is given in Eq. (2) (1) Here, the standard deviation of points in f needs to be computed again, and then the second standard deviation array s is acquired according to Eq. (3): where f i is the average of points centered on f i point in standard deviation window, which is given in Eq. (4): The effect of trend extraction with standard deviation window is given in Fig. 3, which indicates the shapes of the raw time series curve, its first standard deviation curve f, and the second standard deviation curve s. Due to the trend extraction of standard deviation, the volatility of original time series data was amplified and the fluctuation trend of time series was highlighted, which makes the curve smoother while it maintains the temporal characteristics. The dimensions of original time series data can reduce from n to m after the task of standard deviation is finished. The relationship between m and n is defined as Eq. (5).
where m is the length of the second standard deviation array s.

Image representation
It is essential to consider the temporal characteristics, when transforming the second standard deviation array s into image. The correlation between the adjacent data points is stronger than others. Therefore, a (m/2)*(m/2) matrix A is constructed by slide relative position, which represents the position relationship between two points of array s. A is defined as Eq. (6).
Evidently, each point of array s connects to the m/2 points before or after itself and this matrix exploits subtraction to denote relative position among them. Each row of A represents the relationship between a particular point and m/2 time stamps after it, because they have stronger correlation. Each column of A contains the information of two points with different intervals. Moreover, compared with RPM, matrix A reduces the redundant data and improves the generalization ability, while increasing the data volume, and the computational complexity is reduced to O (m/2*m/2). Subsequently, the Min-Max normalization method is employed to obtain the gray value matrix U defined as Eq. (7).
To verify the validity of the proposed trend extraction and image representation, some public datasets from the UCR Time Series Classification Archive were chosen as the samples. Figure 4 shows the images generated from the original time series data of four different datasets with slide relative position matrix, the curves of Fig. 4 are original time series data and the graphs are their corresponding images. Furthermore, these images contain some helpful information of original time series data. For example, the dim part displays the small relative distance between two points and the light part presents the huge relative distance. The frequency of change in color from-bright-to-dim or from-dimto-bright means the volatility characteristics of the original time series. Specifically, the color changing from dim to bright represents a section of the curve with large fluctuations, and no color change means a smooth phase without drastic fluctuations. Moreover, the overall shape of the image is oblique due to the slide relative position matrix, which is helpful to find the patterns and features embedded in transformed images. Fourteen datasets were chosen as samples in this study, and the images of different clusters in these datasets are presented in Fig. 5. Two graphs in each sub-graph are graphs converted from time series data in two different clusters of each dataset, which clearly shows the obvious difference between two clusters. For example, the original data of Gun-Point are a smooth time series and the difference between the two clusters is obvious, but the original data of OSULeaf are a serrated time series and have many frequent fluctuations, and two clusters are very similar due to containing more light and shade change in images.
Subsequently, in the convolution phase, a CNN model was adopted to handle the time series classification based on the 2D images representation, which was viewed as an image classification problem.

Convolution
The CNN had an outstanding performance in the field of image recognition, and the architecture of classical CNN contained five convolution blocks and three full connection layers. To reduce the complexity of CNN models and consumption of resources, a CNN model containing five convolution blocks and a softmax layer was designed for classifying the images converted from original time series data. The width and height of the input image is related to the length of the time series, which is defined as Eq. (8).
where, n is the length of the original sequence, w is the size of the window, D is the size of the input image.
The overall structure of the proposed model in Fig. 6 removed the full connection layer, which reduced the number of parameters. The size of input image changed the flexibly with different length of time series, where the mark "conv" in the parentheses expressed the convolution block and the (8) D = (n − 2w)∕2, numbers after "@" displayed the number of convolution kernel in the present layer, respectively.
To compare the influence of the number of parameters on the networks with or without full connection layer, we performed experiments on the Lightning7 dataset and the results are given in Table 1. Compared with CNN containing two full connection layers with 4096 nodes, the structure of the proposed model reduced the parameters by nearly 99%.
The structure of the proposed model in Fig. 6 was refined into the form of Fig. 7. As shown in Fig. 7, the letter D related to the size of the input images, the CNN model contained five convolution blocks and a softmax layer. Each convolution block had four parts, including the convolution layer, batch normalization (BN) layer, rectification layer and max pooling layer. The convolution layer extracted the features from the input images and obtained the feature maps. Meanwhile, to prevent the edge information of images being lost in the convolution calculation, the input images before convolution layer were padded. The BN layer was used to dispose the gradient vanishing and exploding problem during the back propagation, which avoided the overfitting. The rectification layer adopted a RELU activation function (see Eq. 9) to add the nonlinear factors in this model. The max pool layer reduced the dimension of feature maps produced by the previous layer.
where x is the value of the BN layer.
(9) RELU = max(0, x), Fig. 6 Overall structure of the proposed model After five convolution blocks, the output of the fifth blocks was flattened to a one-dimensional vector, which was then input into the softmax layer that offered the posterior probability of the classification outcomes, in accordance with the true labels of training samples. In this study, the performance of proposed method was evaluated by accuracy (Acc) shown in Eq. (10).
where, N is the total number of samples in the test data, and n is the number of samples correctly classified.

Dataset selection and experiment settings
In the present study, 14 datasets (see Table 2) were downloaded from the UCR Time Series Classification Archive were used to evaluate the proposed SRPM-CNN model. The reasons to choose these datasets are similar to that in reference [28]. The size of training data ranged from 28 to 1000. The length of the training data ranged from 96 to 637. These datasets were standardized to normal distribution in advance.
In the proposed SRPM-CNN model, the size of each convolution kernel was set to 3 × 3 and its stride was fixed to 1 pixel, the size of pooling window was set to 3 × 3, and the stride fixed to 2 pixels was exploited to each max pool layer. Furthermore, a 2 × 2 zero padding method was added before each new convolution layer to prevent the loss of the edge information. The BN layer was after the convolution layer for regulating the processing data, which improved the generalization ability of model and prevented its overfitting. Adam, a stochastic optimization method, was employed to accelerate the training speed. The learning rate was set to 0.000005, which showed a better performance due to the small value of pixel after standardization.
The proposed SRPM-CNN model was implemented on the Tensorflow platform and run on personal computer with Intel Core i5-2000U CPU with 2.2 GHz and 12 GB global memory. The hyperparameters that were input into the model were set to {w, 4, 0.000005}, which implied the window size of standard deviation, the quantity of batch size and the learning rate during optimizing the weights, respectively. The model was fine-tuned by adjusting the parameters to achieve the better experimental results.

Comparison of models with different standard deviation window
The size of the standard deviation window w had an influence on the final classification results, when it was used to extract the trend of time series. The trend curve of time series data could not be extracted effectively if the small size of window was used for the long time series. On the contrary, some helpful characteristics of the time series data were lost, when the big size of window was used for short time series data. Figure 8 shows the different images with different values of w on the three datasets.
The w was set to a slightly large value for some longer time series data and to a bit small value for some shorter time series data. The length of the time series that extracted trend was within 86-630 (the size of input image D was within 43-315). The influence of standard deviation windows on the loss of different datasets is shown in Fig. 9. Obviously, different w in the model resulted in different decline speeds of the loss for some datasets, including the GunPoint, Beef,  GunPoint  50  150  150  2  Lightning7  70  73  319  7  Beef  30  30  470  5  OSULeaf  200  242  427  6  OliveOil  30  30  570  4  Coffee  28  28  286  2  CBF  30  900  128  3  Yoga  300  3000  426  2  ECG200  100  100  96  2  Lightning2  60  61  637  2  Fish  175  175  463  7  Trace  100  100  275  4  TwoPatterns  100  4000  128  4  Wafer  1000  6164  152  2 CBF, Trace, ECG200 and TwoPatterns. However, different w in the model exhibited few effects on loss for some datasets, e.g., the Lightning7, OSULeaf, Coffee, Fish, Lightning2, Yoga, Wafer and OliveOil. As for ECG200 and TwoPatterns, there was a lower loss, when using the original data to generate images. The influence of different standard deviation windows on the time consumption of different datasets is shown in Fig. 10. Obviously, different w in the model resulted in different consumption of time. For the same dataset, an increase in w resulted in a less time to train the model. For different datasets, the consumption of time was positively correlated with the size and the length of w. Figure 11 shows the change in the accuracy of different datasets with different values of w in the range of (0, 40). It is obvious that most datasets presented the optimal value of w, which achieved a best performance in this model. But for some datasets, the original data was directly used to construct the SRPM (means w = 0) for obtaining better results, since the length of these data is shorter than others.

Comparison of different methods
There are some classical time series classification models, including 1NN-DTW, BOSS, COTE, etc. Meanwhile, the FCN is baseline model of 1D-CNN methods, which can achieve extraordinary performance through simple proto structure and reduce complexity without any processing in feature selection. The 2D-CNN methods, including GAT-MTF and RPMCNN can transform the 1D time series data into the 2D images. UCR archive has released the accuracies of some models, including TSF, STC and ResNet. The results of the proposed SRPM-CNN model were compared with that of other 13 models in Table 3. Table 3 shows that the proposed SRPM-CNN model offered better results on most of the datasets. However, its accuracy was lower than that of others in some datasets owing to the different characteristics of the datasets and models. In view of this, some technical findings can be summarized as follows.
• The proposed SRPM-CNN performed better results on datasets with long length of time series data and balanced allocation of training sets and test sets. For example, Lightning7, Beef, OliveOil and Fish had balanced volume of train and test data. Moreover, the length of their original data was longer than that of others. The SRPM could find useful features from a longer time series and highlighted the temporal characteristic. The proposed SRPM-CNN model, not only effectively extracted the features, but also paid more attention to the ordering characteristic. • The accuracy of almost all models on seriously unbalanced dataset was relatively low. For instance, the performances of all models on Yoga were poor, because the numbers of its train data and test data presented a serious imbalance. • The proposed model performed a slightly low accuracy for datasets with short length of original time series data.
For example, in CBF and ECG200, some useful information could be lost due to short length of time series, when the proposed model extracted trends with different w and constructed the SRPM.

Conclusions
There are a lot of time series data in our daily life, but the accuracy is often low, when the CNN is applied on them for time series classification. Therefore, a combined SRPM-CNN model for time series classification was proposed in the present study. The proposed SRPM-CNN model consisted of four parts, including the data trend extraction, image representation, convolutional neural networks and performance analysis. Initially, the standard deviation window was adopted to amplify the fluctuation of the temporal characteristics of the time series data. The slide relative position matrix was constructed to code the extracted trend sequences. Subsequently, the corresponding images were transformed from the time series data and input into the full convolutional neural network for image classification. Finally, fourteen datasets were used to verify the performance of the proposed model, and the results showed that the proposed model offered better results than others.
Our future studies will be focused on the data of short time series and the time series classification with multiple features of 2D-CNN.