A framework for the prediction of earthquake using federated learning

Introduction

Earthquakes are catastrophic events that result in loss of life, infrastructure and have a severe impact on the economic system of a country. Numerous earthquakes occur every year around the world, killing thousands of people. Multiple computational methods have been applied for earthquake prediction, but accurate prediction of earthquakes (time and place) is still a challenging task (Jiao & Alavi, 2020). A lot of research has been conducted for earthquake prediction using different artificial intelligent (AI) approaches in the last decade (Asim et al., 2017). Tehseen, Farooq & Abid (2020a) reported that Expert systems (ES) have been widely practiced on seismic data for making earthquake predictions from 2010 to 2013. Neural network (NN) based earthquake prediction systems remained center of attention from 2013 to 2017 and were followed by an era of Machine Learning (ML) and Deep Learning (DL) based methods. Al Banna et al. (2020) has also approved that ML and DL have performed relatively better than the other computational approaches in making probabilistic forecasts about the expected magnitude of upcoming earthquakes in recent years.

Research about earthquake prediction is now significantly focused towards state-of-the-art cloud centric methodologies for efficient data communication (Lim et al., 2020). Cloud centric methodologies upload and process seismic data in a cloud server incurring high costs and unacceptable latency (Yao et al., 2019). Earthquake prediction requires multidimensional data that is naturally decentralized as data sources are located outside the cloud. Thus, the transportation of this huge seismic data to the central server also offers a big challenge (Salditch et al., 2020). Multiple streams of seismic data have been collected through sensors and aggregated at earthquake data collection centre to produce inference models. Running computational models to simulate multiple seismic data streams and transportation of these data streams as a single data model becomes computationally so hard that a research-level simulation can take weeks, even running on a supercomputer. It also evoked major issues about data privacy, data availability, data security, data protection, and data reliability. Seismic data is one of the most valuable resources, and agencies usually avoid data sharing due to data protection legislations about the sensitive nature of data (Elwood et al., 2020; Enachescu, 2007). International agencies provide open access satellite datasets with reasonable accuracy but its high cost and the issue of indigenous knowledge (IK) poses a great challenge in terms of data capturing, storage and distribution. Many societies in third-world countries lack the technicality processes of managing the IK. Due to these problems onsite regional seismic data acquisition is more popular (Zhao et al., 2020; Elwood et al., 2020).

The application of Mobile edge computing (MEC) leveraged on to bring model training closer to where data is produced. FL is an ML technique that can facilitate the earthquake prediction process by training local data models onsite and ensure data privacy at local stations. FL also applies multiple data security protocols on local data models before transmitting them to the network (McMahan & Ramage, 2017). Basically, FL is a client-server architecture that transmits local data model generated at the local station to the server machine for training global model. This feature protects the network from getting overloaded resolving latency along with other security and privacy issues (Zhao et al., 2020). Application of FL technique for earthquake prediction is also cost effective by transmitting updates only to the central server machine as compared to transmitting huge data to the server.

The main objective of this research is to ensure data availability, facilitate data collection and enforce data privacy by the application of FL technique. This research has proposed a unique framework for earthquake prediction using FL. We have collected three different datasets from Western Himalayan zone and processed them on seven different data stations to train local data model. Different ML classifiers have been trained on global data model at the central server to generate earthquake predictions. We have tested the proposed framework on instrumentally recorded dataset of Western Himalayan zone containing thirty million records from the last thirty-five years (1985–2020). The novelty of our work is that we have applied federated learning technique for earthquake prediction in Western Himalayan region by training ML based local data models onsite and transporting those models to the central server where global data model is trained by their aggregation. A stacking classifier has been trained on a global data model to generate earthquake predictions. To the best of our knowledge, this is the first research work that presents FL framework for earthquake prediction in Western Himalayan zone.

The article has been organized in the following way: “Related Work” gives a brief survey of the computational approaches used for earthquake prediction in literature. “Proposed Framework” presents the proposed FL framework for earthquake prediction. Experimental verification of the proposed framework has been performed in “Experimental Verification”. Results have been discussed in “Results and Discussion”. “Conclusion and Future Directions” concludes the article and provides future research directions.

Related work

Researchers have been exploring the field of earthquake prediction, unfolding multiple aspects of regional or global earthquake data, performing experiments, and applying multiple computational methods to evaluate their results since long time. We have discussed some of the earthquake prediction approaches used in the literature as given below:

Proposed framework

In this section, the framework for earthquake prediction using FL has been proposed. In this framework many rules, functions that are defined according to the expert judgments and mathematical formulas are applied on local stations for extracting significant parameters from local datasets. Different ML techniques have been used to train local data models by processing local datasets onsite. Local data models have been transmitted through network to the FL server where global data model is trained by aggregating local models using FedQuake algorithm. Then, meta classifier has been trained on global data model to generate earthquake predictions with more accuracy. The proposed FL framework has been presented in Fig. 1. It comprises of two basic layers namely data collection layer, data communication and alert generation layer. The architecture and layer-wise functions of proposed framework is described below:

Experimental verification

In this research, three different streams of data have been collected and processed using FL approach. For experimental verification, we have recorded climate dataset, seismic dataset and reservoir dataset from seven different stations (A, B, C, D, E, F and G) located in 100 km radial area from point 34.708° N, 72.5478° E. To perform detailed experiment, regional instrumental data from 1985 to 2020 has been collected and divided into 80% training dataset and 20% testing dataset. Experiment has been performed in two different dimensions. In first dimension, we have generated basic learning model (BLM) using relational data base system. Model has been trained on the dataset containing records from 1985 to 2015. Multivariate rules and formulas have been developed and applied to determine relationships among different data parameters for predicting the magnitude of the earthquakes encountered during 1985 till 2015 in the region. BLM has been tested on the dataset containing records from 2016 to 2020.

In second dimension, the proposed FL framework has been implemented using TensorFlow platform. Working of the client is expressed in following five steps. In the first step, dataset have been collected onsite to generate federated learning model (FLM). In FLM, seven different stations are recording data with respect to climate, seismicity, and reservoir. In the second step, some pre-processing of data is performed at edge that include normalization of data and importing multiple libraries of TensorFlow. In third step, the architecture of proposed framework has been defined that includes the type of ML classifier, batch size (b), learning rate (η) and number of trainings (e). Seven ML based local data models have been developed and their weights are updated to the central server (CS). These local data models are collected using FedQuake algorithm at the CS which is responsible to update the state of the global variables presented in Table 2 in the TFF script. As we are using one CS and seven clients for three datasets, the script will be launched eight times once at every terminal and on the server. In fourth step, data clusters have been defined using tf.train.ClusterSpec. For initialization of data clusters, a dictionary with CS and client keys is received as initialization parameters. During an iteration, script running on CS will be continuously at ‘listen and wait’ state for ‘read or write’ operation from clients. After a specific time, the script running on CS will be blocked to complete initialization of parameters and the action of client nodes will start. Working of federated clients is expressed in Table 3.

In the final step of client’s working, tf.device creates a device type object and replica.device.setter takes care of the synchronization between the clients. Every client will receive data model from the server. For generation of global data model, every client will add its updates to the received model and model will be passed back to the central server. Server asks all the clients to pull the updated model for further processing. This process continues until a good performance matrix is achieved. At the same time, regular interval check points are also loaded and maintained to evaluate the model as training model ceases and does not exist as soon as session is closed. After successfully generating global data model, meta classifier is trained on it to make earthquake predictions.

Multiple predictive modeling methods of statistics, data mining and ML are applied to train the classifiers at clients for generation of local data models. Different ML classifiers including decision tree (DT), neural network (NN), multilayer perceptron (MLP), k-nearest neighbor (KNN) and random forest (RF) have been trained on clients as presented in Fig. 2. Fuzzy Expert system (FES) as a traditional benchmark [12] and Autoregressive integrated moving average (ARIMA) method of time series has also been applied for the generation of local data model at clients. Performance of these classifiers and methods have been compared in terms of prediction accuracy, precision, recall, f1 score and rate of data loss using Eqs. (4)– (6).

(4) $$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(5) $$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(6) $$F1\left(SCORE\right)=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$$

Various parameters (P) are tuned for comparative analysis of all methods along with their values (V) are described in the Table 4.

Results and discussion

The process of FL presented in this research shows that every participating node initializes the training independently as there is no other participant in the network. Learning parameters contains the declaration of local epochs used by each participant for training. The local updates are computed and sent to the central server after specific number of epochs. The central server receives updates from every participating node, calculates their averages and aggregate them to generate the global model. The participants conduct next round of training process based on this updated model received from central server. This mechanism keeps on repeating until the given communication rounds are complete or the desired convergence level is achieved.

Results of the experiment obtained from the application of seven different ML methods have been compared in this section. We have processed three different datasets with 3,000,000 total records divided into 5 folds with batch size of 500 records collected from seven different points in Western Himalayas. Performance comparison of multiple methods has been presented in the Table 5.

To observe the behavior of machine learning models, training process has been started with a small-sized dataset containing 10,000 records. Then in next iteration 90,000 more records were added and model was trained with 100,000 records. In third iteration, full dataset containing 600,000 records has been used for training the model. To maintain consistency among iterations 600,000 more records are added in every iteration to train the model. Some ML classifiers have been trained on local data at clients are presented below:

Multilayer perceptron (MLP) belongs to a class of feedforward artificial neural network (ANN) composed of multiple layers of perceptron with threshold activation. An MLP consists of an input layer, a hidden layer and an output layer. MLP exercises backpropagation supervised learning technique for training. In the proposed framework, MLP classifier has been trained on local data at client to process regional seismic data that is not linearly separable.

k-nearest neighbor (k-NN) is one of the ML classifiers used for classification and regression. In the proposed framework, k-NN works by calculating the distance between neighbors for classification. The neighbors are selected from a pool of objects with known class or the object property value. Training data is normalized to improve accuracy of results as regional seismic data has been recorded in different scales. It assigns weight to the contributions of the neighbors and in result the nearer neighbors contribute more to the average than the more distant ones. In the proposed FL based framework for earthquake prediction, k-NN has been applied on local dataset at client due to its sensitivity to the internal structure of multi scaled regional dataset.

Decision Tree (DT) is a ML classifier used to build a tree by setting different conditions on its branches. It consists of a starting point called a root node, a point where splitting takes place called internal nodes and terminal nodes called leaves. DT classifier has been trained for earthquake prediction in the proposed framework due to its ability to handle complex data according to the expert’s description about a situation incorporated as a set of rules.

Random Forest (RF) is a ML technique that constructs multiple independent DTs under variable conditions. In this technique data arriving at root node is forwarded to all sub-trees for to predict class labels. At the end, voting is performed and the class in majority is assigned to that data. RF classifier has been trained in the proposed framework due to its ability to process big data and generate reasonable predictions with very little configuration requirements.

Performance of classifiers has been evaluated in terms of training time, accuracy, precision, and data loss by training meta classifier. Figure 3 shows that after 100,000 records KNN and RF has very low increase in accuracy indicating very moderate improvement. DT, NN and ARIMA presented a bit lower accuracy with 100,000 records of training data but as more records were added a continuous increase in accuracy has been observed. MLP and FES has shown very low accuracy with initial 100,000 records and minor improvement in accuracy has been observed when 600,000 records were added.

In case of basic learning model (BLM), centralized dataset has been used to train the model. Training of the model took place on the same node without involving any communication delay or overhead cost. It has taken more time while training as computing capacity of the single node is limited. Table 5 presents training time, accuracy obtained and emerging data loss of every model.

FLM has overcome the latency issue as clients are not generating the gradients rather small sized local models are being generated by clients and transmitted to CS. Time taken and accuracy achieved by applying FLM has been presented in the Table 5. It has been observed that less time is taken while training and transmitting the model as huge sized data is not travelling along the network. More accuracy has been obtained by the application of FL mechanism. Data loss calculated is also minimum as FL ensures data privacy and data security by the development of local data models onsite. It can be observed that the earthquake prediction model generated using FL approach is most efficient in terms of time and with minimum data loss than BLM. Accuracy of BLM with centralized data is a bit more than FLM but there is a tradeoff when data privacy, data security and data availability is more important as in the case of earthquake prediction. FL mechanism facilitated the earthquake prediction process by secured and privacy protected data collection. However, there are certain limitations associated with FL based earthquake predicting system.

During learning process, nodes frequently communicate with each other and with the FL server. Technically, FL mechanism requires sufficient local computing power, memory, and high bandwidth connections for exchanging parameters of local machine learning model. Moreover, it also needs an efficient algorithm that may reduce the number of communication rounds and send model updates in iterations as a part of training process. The provision of FL to allow clients to join and leave at any time makes it susceptible to cyber-attacks as an opponent can join the system using different aliases. FL mechanism must allow only a small fraction of the devices to be active at once and it must handle the dropped devices. The hardware variability of participating devices may also affect storage, computational, and communication capabilities of the system. FL protects on device data by sharing model updates only on the network. But transmission of model updates during the training process can still reveal sensitive information to any third party or to the server. Moreover, heterogeneity of local datasets in terms of size, population, variance with time, interoperability among nodes and node failure are the major limitation for earthquake predicting FL based systems.

We have compared the predictions made by different machine learning models with actual earthquakes that have encountered in the region presented in Table 6. We have set the threshold value of magnitude to 5.0 to show that FLM is capable to predict earthquakes that are above threshold value. It can be clearly observed from Fig. 4 that DT, NN and ARIMA have been most effective in earthquake prediction research. Meta classifier has been trained on global data model that compared the prediction accuracy of multiple ML models used to generate global data model and predicted earthquakes encountered in the selected region more accurately.

Conclusion and Future Directions

In this article, FL has been applied for earthquake prediction that enabled distributed clients to collaboratively learn a shared prediction model while keeping all the training data at its origination place. It decouples the ability to do ML from the need of storing big data in the cloud. FL leveraged model training to the edge devices where data is actually residing. The proposed FL based framework, data not only remained decentralized on client networks but some preprocessing has also been performed onsite for the extraction of significant parameters. Multiple ML techniques have been applied on local dataset collected from different stations to generate local data model comprising of only those parameters that are significant in terms of earthquake prediction. These local data models have been sent to the FL server for training of the global model. Meta classifier has been trained on global data model to improve accuracy in earthquake prediction. In the proposed architecture, data remains secure at its original place as it does not travel along the network. This feature will be beneficial for future researchers in developing secured and privacy protected systems for earthquake prediction to process data onsite without involvement of cloud. Although, most of the network security issues do not arise, but we strongly argue that data security and privacy are not binary and can present a range of threats offering new challenges for future researchers. Efficiency in terms of speed and accuracy has also been gained in the proposed architecture as big data are not moving along the network, but a small sized data model extracted from this big data has been sent on the network for further processing. However, the proposed system imposes efficiency requirements on the algorithms used to extract data model from client machines and needs to be improved in future. We have tested the proposed framework using different datasets from Western Himalayans. However, in future work the generalization ability can upgrade proposed framework to process dataset from any other region. Moreover, accuracy in earthquake prediction can be enhanced in future work by integrating federated learning with time series methods or establishing a block chain between research centers for big data and data sites. In the proposed framework, creation of large models and sending full model updates can create a bottleneck due to asymmetric internet connection speeds. New research about modern methods that might be helpful in reducing the uplink communication costs need to be investigated in future. Methods for the scalability of large sized models generated from onsite seismic data also need to be explored in future. Moreover, methods to determine how much communication is necessary in FL framework, also need to be discovered in future.

Supplemental Information