From 1G to 5G and beyond cellular network technologies evolve in terms of technical advancement, efficiency, and traffic. Increase in accessibility and data utilization are the notable factors that are causing high network traffic. Higher the connectivity and mobility of cellular devices, higher will be the network traffic. As the transmission rate increases, the possibility of fault occurrence also increases. Continuous monitoring of network parameters and finding the fault on time are the key factors in determining consistency of network. Cellular network which is highly dynamic than usual networks needs intelligent way of fault handling. The monitoring and fault identification human overhead are unpredictable and very high and can cause errors as well. The research in modelling intelligent network fault identification system can simplify human efforts and improve efficiency. The research is on real time data of 4G cellular network including various network parameters like uplink threshold and identify the behaviour of data usual or unusual to predict the fault occurrence. The numerical dataset on various network parameters on usual and unusual behaviours helps to simplify the feature extraction part and to reduce the complexity of the model. The model is finalized based on various machine learning-deep learning model analysis.
Research Article
Intelligent 4G Cellular Network Fault Prediction Model
https://doi.org/10.21203/rs.3.rs-1504866/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
Fault prediction
Machine learning
Deep learning
Prediction model
4G Cellular network
Bidirectional LSTM
LSTM-Conv1D
Cellular network technologies evolved from 1G to 5G and beyond, provides high network connectivity and scalability. Latest cellular network technologies provide more mobility among the network equipment. This causes more traffic and demands more reliable network management. For reliable network management the fault occurrence should be identified and resolved as early as possible. Human efforts for fault diagnosis will be more in case of high network traffic and prone to errors. Intelligent fault management techniques reduce human effort. It is highly efficient, fast and has future scope in terms of evolving intelligent technologies.
Proactive intelligent fault management ensures performance enhancement among competing service providers. The major requirement of intelligent 4G cellular network fault prediction model is in telecom industry. The competing business environment demands quality of service. Providing consistent fault management mechanism is the key component. This ensures growth in market. Mobile applications such as online shopping, banking, online classes, video conferencing etc are all part of common people. These are all demanding stable network and consistent performance. The trust factor will be mainly based on effective fault identification and even can be applied for fault prevention as next level.
However, the number of studies on self-diagnosis and healing is limited as compared to other areas of research. The correspondence between common faults and related symptoms is not well defined due to dynamic nature of data. Historical data of faults in real world mobile networks is often inaccessible to the scientific community. The available 4G cellular network real time dataset has fixed and variable parameters on normal and fault detected cases in network.
The study is focusing on analysis of intelligent technique for fault prediction of 4G cellular network data based on fixed and variable parameters on normal and fault detected cases. Here the objective is to model advanced, less complex, fault management system for predicting faults for 4G cellular network, using various machine learning models and identify the best. Developing a best fit model to the numerical data helps to formulate simplified feature extraction and prediction in comparison to the log-based feature extraction. The analysis is on machine learning models such as Naive Bayesian, SVM, CNN, LSTM, Conv1D-Bidirectional LSTM, Conv1D-Vanilla LSTM and Conv1D-Stacked Bidirectional LSTM.
Z. Tan and P. Pan et al. [1] explains fault prediction model, which is based on neural network. The logs of wireless network transmission are used as dataset. Network log is pre-processed first to use as input data. It uses time window to get the required data samples which is implemented in two levels. Sample features are extracted with the help of CNN. The convolution part of model extracts the hidden features from the provided logs. LSTM model is used for predicting the results. The combined model, which is hybrid, is developed for prediction of the network fault. The model has high accuracy. But the approach based on CNN feature extraction from logs adds more complexity to the model.
The NLP-CNN based fault prediction model, explained by W. Ji, S. Duan, R. Chen, S. Wang and Q. Ling et al. [2] uses wireless network log analysis. The CNN part of the model has embedded layer to get features from logs. The vectors obtained from this will be given as the input to next layer. As it is doing convolution for feature extraction of logs on top of embedded layer, the model has more complexity in processing real time data.
S. Rezaei, H. Radmanesh, P. Alavizadeh, H. Nikoofar and F. Lahouti et al. [7] explains unsupervised machine learning model. For fault diagnosis the paper applies expectation maximization. It uses DBSCAN along with it, which is a method of grouping data based on density, providing some random noise. The method is applied in mobile network communication to implement the self-decision-making systems. Here the different algorithms for grouping data are being compared with each other and the results are being analysed. Each data group is related to a specific condition causing the fault. Also, there will be related guidelines to solve the faults. The machine learning technique of grouping both traffic and signalling are considered as the most relevant sources of fault information. The method is affected by human error during the modelling stage. The efforts will be high and is difficult to manage highly complex system.
The study on existing papers related to fault prediction models, gives the insight of developing a most advanced, least complex, fault prediction model, using the numerical data of 4G cellular network parameters taken on normal and abnormal environment, which can be used to have more simplified feature extraction and processing.
3.1 Naive Bayesian Model
The model is developed based on Bayes’ theorem. It is a simple supervised algorithm which categories the given set of data into different classes as per the trained data. The model is based Gaussian distribution as the data set contains continuous data. For fault prediction, the model is considered at the initial stage of analysis.
3.2 SVM Model
Support Vector Machine model is based on supervised learning technique. It divides the data sets into specific categories using boundary. The method uses some data points for determining the categories of input data and these are the supporting vectors.
It is less efficient in case of very large data sets and in presence of erroneous data. The model sometimes prone to overfitting or underfitting. Overfitting happens when the boundary of separation shows more nonlinearity towards data points, trying to separate most of data points. On the other hand, the underfitting comes when the model is very poor in making decision to categorise the data points. In such cases hyper parameter tuning is required to maintain the balance between these two scenarios. The main two parameters which are very much effective in developing better SVM models are c and gamma.
The parameter c is for balancing the wrongly classified data point using a penalty value, whose severity determines how much is the amount of misclassification. This is the distance from the boundary point. The parameter gamma can be explained as distance of data point from other points. If it is less, it indicates that more similar data points are grouped together. If it is high, it is showing that the points are far. Very high gamma value may classify too far data points which leads to overfitting. The c value usually lies between 0.1 and 100 and the gamma value lies between 0.0001 to 10.
3.3 ANN Model
The model is based on neural networks which is similar to human brain. It has different layers, basically input layer, output layer and hidden layer. The inputs will be given through the input layer and the next layer input is the sum of product of weights and inputs from previous layer. There is additional bias value and activation function to trigger the functionalities. The hidden layer lies between input and output layers, which can be one to any number. More number of hidden layers increases the complexity of the model.
3.4 CNN Model
The model is based on extracting features using convolution operation. It has input convolution layer which accepts the input data and then the result will be reduced in size using pooling technique to avoid overfitting. This is further given to fully connected layer to make it flattened. The result will be single dimensional vector. Figure 1 shows the fault prediction model using CNN.
3.5 LSTM Model
Long Short-Term Memory is advanced RNN technic to overcome the short memory. LSTM has three gates, forget gate, input gate and output gate. The forget gate decides the information to discard from the cell, the input gate decides values from the input to update the memory state and the output gate decides what to output based on input and the memory of the cell [12]. This is mainly used for models which consider previous data states for result generation. The LSTM which has memory to store previous state can help on the dependability of output on previous results. There are 3 types of models:
-
Prediction of next input sequence value
-
Prediction of input sequence category
-
Generation of another output sequence
This paper focuses on the second type of LSTM application which predicts the category, (here the natural/unusual network parameters) of inputs as in Fig. 2. This is commonly used in applications such as:
-
Detect and predict the anomalies: For input containing various observation of parameters or sensors, predict if the sequence is having anomaly or is normal [12].
-
Categories DNA Sequence: For DNA sequence input having A, C, G, and T values, predict whether the sequence is for coding or non-coding region.
-
Analyse Sentiment data: Given a sequence of text such as a review or a tweet, predict whether the sentiment of the text is positive or negative
3.6 Vanilla LSTM-Conv1D Model
The model is based on deep learning techniques for feature extraction and prediction. It shows good performance as compared to the traditional ways which are found to have less accuracy and performance during the analysis of previous studies.
The approach uses 1D convolution for feature extraction. The 1D convolution has kernel only move in one dimension. Apart from the normal CNN for image data, can extract features of text, numeric and audio inputs. One of such examples is to extract the sensor data and time series data. It uses input data of two dimension. For example, change of wireless network parameter values with respect to time. Here the network parameters can be given as input to the Conv1D layer. The vanilla LSTM is LSTM model with single hidden layer. Figure 4 shows the fault prediction model using vanilla LSTM and time distributed Conv1D.
3.7 Bidirectional LSTM-Conv1D Model
For the bidirectional LSTM, as the name indicates, the input sequence is analysed in forward and backward directions. Then it combines both results. The model uses 1D convolution for feature extraction and later combines with bidirectional LSTM for predicting the fault for given network parameters. Figure 5 shows the fault prediction model using bidirectional LSTM and time distributed Conv1D.
3.8 Stacked Bidirectional LSTM-Conv1D Model
The term stacked indicates that multiple layers are combined. The model uses 1D convolution layer and multiple bidirectional LSTM layers. Time distributed layer is used to wrap convolution layer, to apply to all samples. Figure 6 shows the fault prediction model using stacked bidirectional LSTM and time distributed Conv1D.
The Naïve Bayesian model and the SVM model are implemented using Keras built in libraries. The Naïve Bayesian is implemented using GaussianNB library. In this, StandardScaler library is used for the input normalization. For SVM the MinMaxScaler library is used for normalization. The parameter tuning is done using grid search.
The other neural network and LSTM based implementations are explained below, indicating the basic design of layers and functionalities.
The CNN model has input convolution layer accepting 11 feature inputs, followed by max pooling layer. The output of this layer is flattened and given to hidden layer. This is followed by the final output dense layer.
The basic LSTM model has all the input features fed into the input layer which is followed by dropout regularization. Then this is further followed by another stack of LSTM layer. Here also the regularization is used.
For vanilla LSTM with Conv1D model, the Conv1D part is for feature extraction. The 11 features are given as input to the input layer which is time distributed Conv1D. This performs convolution operation in one direction to extract the relevant features. Max pooling reduces dimension. The flatten layer converts the outcome of this into one dimensional vector and feed to LSTM input layer.
For bidirectional LSTM with time distributed Conv1D also, the input layer is time distributed Conv1D, which extracts the relevant features. The bidirectional LSTM layer processes the input for prediction in two directions and yields better results than vanilla models during the implementation.
The stacked bidirectional LSTM with Conv1D model shows better performance than all the other models, Naïve Bayesian, SVM, ANN, CNN, LSTM, vanilla LSTM, and bidirectional LSTM. The model has three layered bidirectional LSTM stack. The extracted features from the convolution operation is fed to the first layer of the LSTM stack. Out of all the models, this model gives best results in terms of accuracy and consistency for the 4G cellular network data.
5.1 Naïve Bayesian Model
Out of all the models analysed, the model resulted in very less accuracy of 54.34%. The confusion matrix is shown in the Table 1. The true positive rate is just 36.85% and the true negative is 17.49%. This indicates the models is showing poor performance in finding the fault occurrence of the 4G cellular network data.
5.2 SVM Model
The analysed SVM models involve basic model and model with hyper parameter tuning. The hyper parameters involve c and gamma.
The accuracy of the model is better than Naïve Bayesian (72.65%). Table 2 shows the Precision, Recall and F1 measure for SVM model. It is showing f1-score of 0.84 for usual network but 0 for faulty network. This shows the requirement of tuning hyper parameter to improve the performance of the model.
Table 3 shows the Precision, Recall and F1 measure for SVM model with hyper parameter tuning. The accuracy of the model is improved to 81.295%. The model shows improved f1-score for usual and unusual categories. But the unusual category still has the low recall and f1-score which are 0.47 and 0.58 respectively.
5.3 ANN Model
Table 4 shows the accuracy and loss curves of the ANN models. The ANN based models are showing less performance and consistency. Most models are showing biased. The accuracy and loss curves are shown in Table 5. The models are underfitting to the given 4G cellular network data. The models are tried with various parameters such as epochs, output activation function, neurons, regularization, hidden layer etc. The maximum possible accuracy is shown as 78%.
5.4 CNN Model
Table 6 shows the different CNN models using Conv1D. Here the best model is the second model with dropout regularization of 0.2. It gives an accuracy of 81.67%. The CNN based models are producing good accuracy without regularization and is not showing any bias or variance. However, the model is not increasing accuracy after certain level even if tried with several hyper parameter variation.
Table 7 shows the accuracy and loss curves of the models. The CNN based models are showing comparatively better performance and consistency than previous models. Only one model is slightly biased which is the third one.
5.5 LSTM Model
Table 8 shows the different LSTM based models and Table 9 shows the accuracy and loss curves. Here the models are mostly showing high variance and accuracy.
The highest accuracy is from the first model. Without any regularization it gives 86.378% which is best accuracy among all the models. But the Table 9 is indicating the high variance of the models and hence is showing overfitting to great extent.
The best model is the third model with L2 regularization of 0.1. It gives an accuracy of 78.842% without variance.
5.6 Vanilla LSTM Conv1D Model
The model is implemented with various parameters such as number of neurons, epochs, test train splits, loss functions, output activation function, optimizer and the regularization. The dropout is the regularization mechanism which will remove unwanted weight and reduce overfitting. This L2 regularization also reduces overfitting by reducing the unimportant weight values.
The Table 10 indicates the notable models with various parameter values. The table indicates that the accuracy is maximum with second model which is using adam optimizer, mse loss function and relu activation function.
The dropout is giving less result compared to other models. The best model in this method is first one which is not using any regularization and giving 81.415% accuracy without any bias or variance.
The accuracy and loss graphs in Table 11 shows there is no bias or variance for model other than the second model as the test and validation loss curves are almost syncing with each other. This indicates that there is slight variance and overfitting for second model.
5.7 Bidirectional LSTM Conv1D Model
The bidirectional LSTM-Conv1D models with various parameters and the accuracy and loss plot table are showing in Table 12 and Table 13.
Almost all the models are giving good stable results comparing to previous models. The best accuracy result is without regularization. The model4 with dropout and L2 regularization is showing the best accuracy without any bias or variance which is 82.24%. The loss curves and accuracy curves are showing a good sync in this model. Only second model is showing slight variance.
5.8 Stacked Bidirectional LSTM Conv1D Model
The stacked bidirectional LSTM-Conv1D models with parameter changes are shown in Table 14.
The high accuracy is with model4 with dropout regularization (0.1). Table 15 explains the accuracy and loss plots. With and without regularization the models have good accuracy than the bidirectional LSTM model and other models. It is best fit for the prediction model.
The overall analysis indicates that the neural network-based models are more suitable for the fault prediction model. Among all these, CNN based feature extraction is best way of getting cellular network’s relevant features.
The LSTM model is most advanced deep learning method for finding dependency among data and predicting result. Combining both models improve overall performance by utilizing the advantages of both the methods.
The prediction models have been implemented using various machine learning models like Bayesian, SVM, ANN, CNN, LSTM. Manual feature extraction gives poor result in prediction of the cellular network fault. The models show better performance while focusing on deep learning models such as CNN, which extracts relevant features and LSTM which can predict the future based on past data as well. The time distributed Conv1D feature extraction part combined with LSTM prediction part gives the best results. This utilizes the advantages of both the techniques. As the model is trained for numerical data of cellular network parameters on usual and unusual environment, the feature extraction using Conv1D simplifies the model than log-based feature extraction.
The 4G cellular network fault prediction model provides better results using stacked bidirectional LSTM and time distributed Conv1D architecture. As the cellular network data is highly dynamic and large in amount the deep learning model will provide and efficient fault detection mechanism compare to traditional mechanisms.
Cellular network fault management using deep learning techniques is a research area of wide scope as it is less under investigation compared to other areas. The limited and incomplete real time confidential data is a major hurdle. There can be used Generative Adversarial Neural Networks (GAN) in further steps to generate data based on available one. The existing model can be evaluated in 5G cellular network data as well.
Conflict of Interest
The authors don’t have any conflict of interest.
Funding
There is no funding used for the execution of this work.
Author Contribution
Geethu N. (First Author) carried out her Post Graduate research project under the guidance of Rajesh M. (Second Author). Geethu developed the proposed system based on the guidelines provided by Rajesh and conducted experiments and obtained results. Based on these results and observations made, Geethu drafted the initial version of the paper. Rajesh reviewed the paper and suggested modifications for better accuracy and performance. After incorporating the suggested changes, final paper is prepared which is reviewed by Rajesh.
- Tan, Z., Pan, P.: "Network Fault Prediction Based on CNN-LSTM Hybrid Neural Network," 2019 International Conference on Communications, Information System and Computer Engineering (CISCE), Haikou, China, 2019, pp. 486–490, doi: 10.1109/CISCE.2019.00113
- Ji, W., Duan, S., Chen, R., Wang, S., Ling, Q., "A CNN-based network failure prediction method with logs," 2018 Chinese Control And Decision Conference (CCDC), 2018, pp. 4087–4090, doi: 10.1109/CCDC.2018.8407833
- Zheng, L., et al., "A Fault Prediction of Equipment Based On CNN-LSTM Network," 2019 IEEE International Conference on Energy Internet (ICEI), Nanjing, China, 2019, pp. 537–541, doi: 10.1109/ICEI.2019.00101
- Abdullah Al, S.M., Mamun, Juha Valimaki,Anomaly Detection and Classification in Cellular Networks Using Automatic Labeling Technique for Applying Supervised Learning,Procedia Computer Science,Volume 140,2018,Pages 186–195,ISSN 1877 – 0509,https://doi.org/10.1016/j.procs.2018.10.328
- Mulvey, D., Foh, C.H., Imran, M.A., Tafazolli, R., "Cell Fault Management Using Machine Learning Techniques," in IEEE Access, vol. 7, pp. 124514–124539, 2019, doi: 10.1109/ACCESS.2019.2938410
- Bothe, S., “ Ai Based Fault Diagnosis In Emerging Cellular Networks,”: 2019, shareok.org/handle/11244/321132
- Rezaei, S., Radmanesh, H., Alavizadeh, P., Nikoofar, H., Lahouti, F., “Automatic fault detection and diagnosis in cellular networks using operations support systems data,” NOMS 2016–2016 IEEE/IFIP Network Operations and Management Symposium, Istanbul, 2016, pp. 468–473, doi: 10.1109/NOMS.2016.7502845
- Shukla, S.K., Koley, E., Ghosh, S.D.C.: offset estimation-based fault detection in transmission line during power swing using ensemble of decision tree. IET Sci. Meas. Technol. 13, 212–222 (2019) “”.
- https://towardsdatascience.com/understanding-1d-and-3d-convolution-neural-network-keras-9d8f76e29610%20WRITTEN%20BY%20Z%C2%B2%20Little%20Data%20Science%20Diary%20Follow%2010%2010
- Jang, B., Kim, M., Harerimana, G., Kang, S.-U., Kim, J.W.: Bi-LSTM Model to Increase Accuracy in Text Classification: Combining Word2vec CNN and Attention Mechanism. Appl. Sci. 10, 5841 (2020) “”.
- https://machinelearningmastery.com/how-to-reduce-overfitting-in-deep-learning-with-weight-regularization/
- https://vel.life/%E9%98%85%E3%80%8Along-short-term-memory-networks-with-python%E3%80%8B/long-short-term-memory-networks-with-python.pdf
- https://machinelearningmastery.com/how-to-develop-lstm-models-for-time-series-forecasting/
- https://www.kaggle.com/c/anomaly-detection-in-cellular-networks
- https://towardsdatascience.com/svm-hyperparameters-explained-with-visualizations-143e48cb701b
- https://towardsdatascience.com/remaining-life-estimation-with-keras-2334514f9c61
- https://machinelearningmastery.com/how-to-reduce-overfitting-in-deep-learning-with-weight-regularization/
Table 1. Confusion matrix for Naïve Bayesian model
Table 2. Precision, Recall and F1 measure for SVM model
Table 3. Precision, Recall and F1 measure for SVM model with hyper parameter tuning
|
Mode1 |
Model2 |
Model3 |
|
|---|---|---|---|
|
Accuracy |
78.036 |
75.55 |
72.2 |
|
Validation Loss |
0.486 |
0.529 |
0.581 |
|
Regularization |
- |
Dropout (0.2) |
L2(0.1) and Dropout (0.2) |
|
Epoch |
100 |
100 |
100 |
|
Optimizer |
Adam |
Adam |
Adam |
|
Test split |
0.1 |
0.1 |
0.1 |
|
Loss function |
MSE |
MSE |
MSE |
|
Output activation function |
Sigmoid |
Sigmoid |
Sigmoid |
|
Mode1 |
Model2 |
Model3 |
Mode4 |
|
|---|---|---|---|---|
|
Accuracy |
81.479 |
81.67 |
78.47 |
74.754 |
|
Validation Loss |
0.136 |
0.13 |
0.157 |
0.178 |
|
Regularization |
- |
Dropout (0.2) |
L2(0.1) |
Dropout and L2 (0.1) |
|
Epoch |
200 |
200 |
300 |
100 |
|
Optimizer |
Adam |
Adam |
Adam |
Adam |
|
Test split |
0.1 |
0.1 |
0.1 |
0.1 |
|
Loss function |
MSE |
MSE |
MSE |
MSE |
|
Output activation function |
RELU |
RELU |
RELU |
RELU |
|
Mode1 |
Model2 |
Model3 |
Mode4 |
|
|---|---|---|---|---|
|
Accuracy |
86.378 |
83.06 |
78.842 |
77.952 |
|
Validation Loss |
0.309 |
0.38 |
0.471 |
0.487 |
|
Regularization |
- |
Dropout (0.2) |
L2(0.1) |
Dropout and L2(0.1) |
|
Epoch |
100 |
100 |
100 |
100 |
|
Optimizer |
Adam |
Adam |
Adam |
Adam |
|
Test split |
0.1 |
0.1 |
0.1 |
0.1 |
|
Loss function |
MSE |
MSE |
MSE |
MSE |
|
Output activation function |
Sigmoid |
Sigmoid |
Sigmoid |
Sigmoid |
|
Mode1 |
Model2 |
Model3 |
Mode4 |
|
|---|---|---|---|---|
|
Accuracy |
81.415 |
84.08 |
79.217 |
79.42 |
|
Validation Loss |
0.133 |
0.116 |
0.148 |
0.148 |
|
Regularization |
- |
Dropout (0.2) |
Dropout and L2(0.1) |
|
|
Epoch |
100 |
200 |
100 |
200 |
|
Optimizer |
Adam |
Adam |
Adam |
Adam |
|
Test split |
0.1 |
0.1 |
0.1 |
0.1 |
|
Loss function |
MSE |
MSE |
MSE |
MSE |
|
Output activation function |
Sigmoid |
Sigmoid |
Sigmoid |
Sigmoid |
|
Mode1 |
Model2 |
Model3 |
Mode4 |
|
|---|---|---|---|---|
|
Accuracy |
81.585 |
83.613 |
81.425 |
82.224 |
|
Validation Loss |
0.133 |
0.119 |
0.135 |
0.129 |
|
Regularization |
- |
- |
Dropout (0.2) |
Dropout (0.1) |
|
Epoch |
100 |
200 |
200 |
200 |
|
Optimizer |
Adam |
Adam |
Adam |
Adam |
|
Test split |
0.1 |
0.1 |
0.1 |
0.1 |
|
Loss function |
MSE |
MSE |
MSE |
MSE |
|
Output activation function |
Sigmoid |
Sigmoid |
Sigmoid |
Sigmoid |
|
Mode1 |
Model2 |
Model3 |
Mode4 |
|
|---|---|---|---|---|
|
Accuracy |
81.799 |
82.058 |
82.389 |
82.431 |
|
Validation Loss |
0.132 |
0.148 |
0.128 |
0.126 |
|
Regularization |
- |
- |
Dropout (0.2) |
Dropout (0.1) |
|
Epoch |
100 |
150 |
200 |
200 |
|
Optimizer |
Adam |
Adam |
Adam |
Adam |
|
Test split |
0.1 |
0.1 |
0.1 |
0.1 |
|
Loss function |
MSE |
MSE |
MSE |
MSE |
|
Output activation function |
Sigmoid |
Sigmoid |
Sigmoid |
Sigmoid |
Tables 5, 7, 9, 11, 13, 15 are available in the Supplementary Files section
No competing interests reported.