Development of a Machine Learning Based Fault Detection Model for Received Signal Level in Telecommunication Enterprise Infrastructure

Wireless telecommunication infrastructure can fail without notice for maintenance action. However, these failures may not always result in a complete downtime, but rather degraded performance that can be difficult to pinpoint without specialized tools. Received signal level (RSL) is an important metrics of the quality of a wireless connection and can be used to determine the strength of the signal from the transmitting to the receiving device. It is important to regularly monitor the RSL to ensure that the network is operating at peak performance and to proactively address any issues that may arise.

As businesses rely more and more on the networks to increase operational effectiveness and foster long-term business growth, telecommunications are a crucial component of today's business world. Through the power of telecommunication, businesses have experienced improved collaboration, enhanced communication and maximum productivity. Telecommunications networks have evolved into an important medium that provides the necessary platform for this electronic data exchange. Yet, in the current digital ecosystem, organizations can use mobile communication to speed up workflow and productivity while allowing employees to use their devices to access particular applications, reply to emails, work on presentations, and take part in teleconference calls. This is made possible through the telecommunication service providers infrastructure.

The telecommunications network is responsible for carrying all internet data and can be comprised of various technologies, including satellites, microwaves, and mobile networks like 5G. Efficient communication relies on telecommunication infrastructure, which enables individuals and organizations to communicate via wired and wireless connections, phone, internet, and other mediums. As demand for connectivity increases globally, customers and end-users expect modern telecommunication networks to operate with minimal downtime, making carriers that can offer nearly 100% uptime the preferred choice for many consumers.

Telecommunication network failure of enterprise customers’ equipment are avoidable, with careful planning, constant monitoring and maintenance in addition with spare parts available at all times. To circumvent the problem of constant outages which may lead to reputational damage, telecommunication service providers started searching for better ways to improve service delivery by applying the right maintenance strategies.

The majority of telecommunication service providers began to use reactive or preventive maintenance procedures [1]. Reactive or breakdown maintenance is the strategy whereby no maintenance is carried out until machine breaks down. While predictive maintenance involves performing maintenance based on predictions derived from the analysis of key degradation parameters. The downside of both strategies is that reactive maintenance allows failure to occur before actions are taken [2] while preventive maintenance acts irrespective of the current state of the equipment to be maintained [3]. This means that the customers network will be shutdown to perform the maintenance whether the service is fine or not. Usually, customers are always upset with this form of maintenance decisions as their network will be interrupted when they are at the middle of a major task. This could also lead to loss of revenue for bigger businesses. Predictive maintenance (PdM) technology on the other hand has been widely used to manage the health of equipment and predict equipment failure so that organizations can schedule maintenance in advance to avoid unplanned equipment downtime and improve customer service [3-5]. In the light of this, the predictive maintenance strategy seems more appropriate for the telecommunications provider as it relies on the data to make decisions when equipment will be maintained before a customer complains. Maintenance actions can then be initiated to replace or repair the faulty components so that the associated unit can continuously perform its intended function throughout its useful life. Consequently, equipment failures must be discovered and corrected to avoid service disruption [6] and should be performed proactively to reduce maintenance costs and increase equipment uptime to the maximum extent practicable. This fact necessitates a shift in maintenance techniques from diagnostics to a prognostics approach.

The problem statement of this research is that the telecommunications industry is always looking for ways to maintain its infrastructure and provide reliable service while keeping costs to a minimum. Many researchers have engaged in reviewing ways to improving maintenance system for about twenty years now [7]. The reactive maintenance approach is only suitable when a failure occurs causing a sudden shutdown and end users are always unhappy with the interruption of service. The preventive maintenance (PvM) technique emerged as one of the prevalent strategies used by most telecommunication businesses. However, the annual cost is considerably high since more replacement spare must be purchased and stored. The success of this strategy hinges on having all replacement spares on hand for scheduled maintenance tasks on time. Predictive maintenance strategies work without a schedule for servicing or parts replacement. The main idea is to predict a failure and then replace the spares before their lifespan expires. If there is any evidence of deterioration, the maintenance personnel will perform the replacement appropriately. Predictive maintenance reduces downtime, optimizes spare parts inventory, and extends the longevity of the equipment [8, 9]. In a traditional approach, Network Operations Center (NOC) personnel manually monitor the receive signal levels of the wireless enterprise customers and then engage a field support personnel to investigate any deviation in the signal parameters. However, with the increasing complexity of telecommunication systems and the rapid growth of the network, this manual approach becomes time-consuming. PdM applications are increasingly using intelligence, and model-based approaches due to the fact of it’s more efficient to data-driven decision-marking and real-time applications. The science of artificial intelligence has seen the emergence of machine learning method in the field of predictive maintenance. This problem can be mitigated by developing of a machine learning based fault detection model for received signal level in telecommunication enterprise infrastructure that can automate the process of monitoring and alert network operators of any deviations from optimal signal levels. Therefore, the problem to be addressed in this research work is identify system metric that indicate an impending system failure and develop a machine learning model that can accurately monitor the system metric and provide timely alerts to network operators for effective telecommunication maintenance purposes.

This section presented the conceptual framework, the materials and methods employed in achieving the research goals.

3.1 Conceptual framework

The conceptual framework is made up of three modules: a data collection module, data analytics module and a machine learning management module. The Data Analytics Module is responsible for collecting the received signal level data from the telecommunication network and verifying that the data is correctly formatted for further modeling. The machine learning module extracts characteristics from the preprocessed dataset to train and assess the model's performance as shown in Figure 1, while Table 1 shows the Material and software applications used in carryout this research.

1.png

Figure 1. Conceptual framework

3.2 Data generation approach

To simulate the received signal level data, the following methods were investigated and in this study a physics-based simulation method using the Path Loss 5.0 software was utilized. Table 2 contains the parameters for the Transmitter (Tx) and Receiver (Rx) which are utilized in the simulation of a vector element for a functional telecommunication link using the pathloss 5.0 software. The table provides specific information on the transmit and receiver antenna gain, transmit power output, frequency, polarization, coordinates, and antenna height which are used to determine the path length. These parameters were supplied by the microwave radio manufacturer. On the other hand, the parameters in Table 3 were employed to generate data that could be used for testing the machine learning model. Therefore, the values in both tables are distinct since they were generated using parameters from different base stations.

Table 1. Material and software applications

Table 2. Parameters for simulating training datasets

Table 3. Parameters for simulating testing datasets

Simulating data for received signal level (RSL) involves generating values for the signal strength that represents a real-world scenario and the following steps were followed:

a) Determine the frequency of transmission.

b) Select a propagation model to determines how the signal strength changes over the distance.

c) Select the transmit power according to the manufacturer’s specifications.

d) Choose the antenna heights and define the environments.

e) Generate data using pathloss 5.0.

f) Validate the data to ensure the simulated data accurately represents the real-world scenario.

3.2.1 Pathloss 5.0

Pathloss 5.0 is a robust software program that is useful for designing, optimizing, and planning radio networks. It offers an array of tools to evaluate and simulate radio propagation and to anticipate coverage and signal strength which allows users to input the base station (transmitter) and users (receiver) end parameters such as the base station location, antenna height, transmit power, frequency band, modulation scheme. Once the transmitter and receiver parameters have been set up, Pathloss 5.0 can be used to simulate the propagation of radio signals between the two ends and predict the expected signal strength, quality, and coverage area of the link.

The software uses advanced algorithms to model the effects of various factors such as terrain, building clutter, and atmospheric conditions on radio propagation, allowing users to optimize the placement and configuration of transmitters and receivers for maximum performance. In addition, Pathloss 5.0 provides tools for analyzing link performance, including link budget analysis and interference analysis, which allow users to fine-tune the transmitter and receiver parameters and improve link performance.

To simulate the wireless links, the pathloss software was launched on a computer and a new project environment was created. The transmitter and receiver parameters were set up by specifying locations, antenna height, antenna gain, transmitter power, and frequency. Digital Terrain data was imported and the propagation model was defined. Finally, the pathloss was calculated and a link budget was generated for the wireless link.

3.2.2 Data validation

The free space path loss (FSPL) equation is employed to estimate the attenuation of a radio signal as it moves through space, whereas the received signal level (RSL) equation is utilized to determine the strength of a radio signal that a receiver antenna receives in a wireless communication system. The RSL is crucial in evaluating the dependability and quality of the communication link and is often compared to the minimum signal level required for the communication system to work properly. These equations can be represented as shown in Eq. (1) and Eq. (2) [35] and are used to generate the sample raw data in Table 4 and Table 5.

$L_{\mathrm{FSL}}=92.45+20 \log \left(f_{\mathrm{GHz}}\right)+20 \log \left(d_{\mathrm{km}}\right)[\mathrm{dB}]$       (1)

where,

$f=$  the frequency of the signal in gigahertz frequency $(\mathrm{GHz})$

$d=$ the distance between the transmitter and receiver in $(\mathrm{km})$

Similarly,

$\begin{gathered}\text { Received signal strength (RSL) }=\mathrm{P}_{\mathrm{o}}-\mathrm{L}_{\mathrm{ctx}}+ \\ \quad \mathrm{G}_{\mathrm{atx}}-\mathrm{L}_{\mathrm{crx}}+\mathrm{G}_{\text {arx }}-\mathrm{FSL}-\mathrm{L}_{\mathrm{m}}[\mathrm{dBm}]\end{gathered}$         (2)

where,

$\mathrm{P}_{\mathrm{o}}$ the transmitted power in $(\mathrm{dBm})$

$\mathrm{L}_{\text {ctx}}=$ Transmitter antenna loss $(\mathrm{dB})$

$\mathrm{G}_{\mathrm{atx}}=$  the gain of the transmitting antenna (dBi)

$\mathrm{L}_{\mathrm{crx}}=$  Receiver antenna loss $(\mathrm{dB})$

$\mathrm{G}_{\mathrm{arx}}=$ the gain of the receiving antenna (dBi)

$\mathrm{L}_{\mathrm{m}}=$ system losses (dB)

FSL $=$  free-space pathloss (dB)

Using the simulation parameters from Table 2 and Eqs. (1)-(2) to validate the parameters.

$\mathrm{F}=10 \mathrm{GHz}$ and $\mathrm{d}=1.71 \mathrm{~km}$

Free space pathloss $(F S P L)=92.45+20 \log (10)+20 \log$ (1.71)

FSPL $=92.45+20+4.6599=117.11 \mathrm{~dB}$

Similarly, with Eq. (3)

$\begin{aligned} & \text { Received Signal Level (RSL) }=\mathrm{P}_{\mathrm{o}}-\mathrm{L}_{\text {ctx }}+\mathrm{G}_{\text {atx }}- \mathrm{L}_{\text {crx }}+\mathrm{G}_{\text {arx }}-\mathrm{FSL}-\mathrm{L}_{\mathrm{m}}\end{aligned}$        (3)

where,

$\mathrm{P}_{\mathrm{o}}$ = 21dBm,

$\mathrm{L}_{\mathrm{ctx}}$ = 2.43dB

$\mathrm{G}_{\mathrm{atx}}$ = 16dBi

$\mathrm{L}_{\mathrm{crx}}$ = 0.00dB

$\mathrm{G}_{\mathrm{arx}}$ = 33.9dBi

FSL = 117.11dB

$\mathrm{L}_{\mathrm{m}}$ = 0.02dB

$\begin{gathered}\mathrm{RSL}=21-2.43+16-0.00+33.9-117.11-0.02=-48.66 \mathrm{dBm}\end{gathered}$

In general, the mathematical model for simulating the RSL values is shown in Eq. (4).

$\begin{aligned} \mathrm{RSL}= & \mathrm{P}_{\mathrm{o}}-\mathrm{L}_{\mathrm{ctx}}+\mathrm{G}_{\mathrm{atx}}-\mathrm{L}_{\mathrm{crx}}+\mathrm{G}_{\mathrm{arx}}-(92.45+ \\ & \left.20 \log \left(f_{\mathrm{GHZ}}\right)+20 \log \left(d_{\mathrm{km}}\right)\right)-\mathrm{L}_{\mathrm{m}}\end{aligned}$        (4)

Table 4. Sample generated training dataset

Table 5. Sample generated test dataset

3.2.3 Data understanding

To conduct the research, simulated data was generated using pathloss 5.0 software, and a mathematical model obtained from the RSL transmission equations was used to verify the accuracy of the dataset. This additional step of validation using the RSL transmission equations helped ensure the reliability of the results before utilizing them in machine learning models. The RSL dataset comprises 5000 instances with 10 attributes, such as customer node IDs, distance (in kilometers), transmitting frequency (in GHz), free space path loss (in dB), transmit power (in dBm), transmit antenna gain (in dBi), transmitter loss (in dB), miscellaneous losses, remote antenna gain (in dBi), and received signal level (in dBm).

3.2.4 Data preprocessing

Data preprocessing is a fundamental technique aimed at reducing the complexity of data for easier processing. It covers the entire process of preparing data for analysis, including data wrangling and other tasks such as scaling and normalizing data, converting categorical variables to numerical variables, and dividing data into training and validation sets [36, 37].

3.3 Feature selection and extraction

It has been demonstrated that feature selection for high-dimensional data analysis is effective and advantageous by selecting a subset of important characteristics from a dataset using feature selection.

The objective of selecting features from a given set of data is to choose the most insightful and pertinent features for a specific issue while lowering the data's dimensions and computing complexity [38]. On the other hand, feature extraction typically transforms the raw data into easily recognizable features [39]. For this investigation, the six labelled input features indicated in Table 6 together with their data types have been chosen.

Table 6. Input features

3.4 Data normalization

Data standardization was performed on the dataset using Standard Scaler library in Python before incorporating the input into the ML algorithms. The main idea is to standardize the data by adjusting its scale and distribution so that each value has a mean of 0 and a standard deviation of 1 as shown in Eq. (5):

$z=\frac{X-\mu}{\sigma}$       (5)

where,

Z: the standard score

X: the raw score or observation being standardized

μ: the mean of the data

σ: the standard deviation of the data

3.4.1 Data division

Nguyen et al. [40] did a study where the Performance evaluation of several machine learning algorithms is done by partitioning the dataset into different proportions for training and testing. The RMSE, MAE, and R-squared metrics were used to assess the model's effectiveness and gauge each metric's predictive potential. It was observed that the training/testing ratios of 70/30 outperformed other ratios. The train-test split's aim is to ensure that the model is trained and tested on various data sets in order to avoid overfitting [41].

3.5 Model selection

The act of choosing the most appropriate statistical or machine learning model for a specific problem or dataset is known as model selection. It is an important step in the machine learning modeling process, as the choice of models can greatly impact the accuracy and interpretability of the results. Received signal level prediction is a supervised learning regression problem because the predicted output is continuous-valued [39]. Hence, neural networks, support vector regression (SVR), K-NN, and random forest regression have shown substantial performance in received signal level predictions [42-46].

3.5.1 Hyperparameter tuning

A very common approach for hyperparameter tuning is grid search, where a range of hyperparameters are specified for model training and evaluation with each hyperparameter combined together. Random search is another approach, where a random subset of hyperparameters is chosen and the model is trained and evaluated with those hyperparameters. There are also more advanced methods such as Bayesian optimization and gradient-based optimization. However, the process to tune the hyperparameter can be time-consuming and computationally expensive, especially for large and complex models [47-50].

3.5.2 Model development

The Python and machine learning libraries were imported into the Jupyter notebook environment and stored in the same directory as the 'RSL_Data.csv' file for preprocessing. The RSL data was split into features and a target variable, with input values comprising the features and the RSL values as the target (output) variable. The features were standardized by scaling them to have a mean of 0 and a standard deviation of 1.

The data was split into two subsets, using the train_test_split function. One subset was used for training the machine learning model, and the other subset was used for evaluating the performance of the model. The split was random and involved using 70% of the data for training and the remaining 30% for validation. A random seed value was also set, which ensured that the dataset was split consistently each time the code was executed. This level of consistency is crucial for reproducibility since it guarantees that the same random split is used during each execution.

For each classifier, an instance was created, and the models were fitted to the provided training data. After being trained via the fit method, the model was employed to make predictions on new data using the predict method. Additionally, the performance of the trained model was assessed on the validation dataset, which was used to fine-tune the model's hyperparameters.

To fine-tune the model’s performance, the GridSearchCV class was employed. This class exhaustively searches a specified hyperparameter space, which is specified as a list of dictionaries. The GridSearchCV class receives inputs used to evaluate the performance of the model for each hyperparameter combination. It then carries out a cross-validation procedure, utilizing the specified number of folds, to estimate the model's performance for each hyperparameter combination. The result of the GridSearchCV is a fitted GridSearchCV object, which contains information about the search process, the estimator object with the best hyperparameters found during the search. Using domain knowledge, the trained model was applied to a new testing dataset to generate predictions, which were then classified as indicating either a fault or no fault considering the rule-based algorithm that had been defined.

3.6 Performance evaluation metrics

Based on the literature [51-53], different approaches exist to assess the performance of the proposed models. In this study, the target variable is a continuous parameter; $f_i$ (which indicates the goodness of fit between the observed and predicted values), the mean squared error (MSE), the mean absolute error and the root mean squared error have been used to evaluate the performance of the models as shown in Eq. (6) to Eq. (9).

Lower values for RMSE, MSE, MSE and higher values for $R^2$ are indicative of a better predictive performance per machine learning approach.

Where, $N$ represents the number of samples, $f_i$ is the actual value predicted by the model, $y_i$ denotes the predicted value and $\bar{y}$ is mean value of $y_i$.

$R^2=1-\frac{\sum_{i=1}^N\left(y_i-f_i\right)^2}{\sum_{i=1}^N\left(y_i-\bar{y}\right)^2}$       (6)

MSE $=\frac{1}{N} \sum_{i=1}^N\left(y_i-f_i\right)^2$       (7)

$\mathrm{RMSE}=\sqrt{\frac{1}{N} \sum_{i=1}^N\left(y_i-f_i\right)^2}$         (8)

$\operatorname{MAE}=\frac{1}{N} \sum_{i=1}^N\left|y_i-f_i\right|$        (9)