Objectives:

The aim of this work package is to provide comprehensive solutions for predicting supply capabilities of AGV vehicles and managing a fleet of AGV based on various parameters from sensors and accompanying management systems. Energy management is one of the most important problems when transporting components on AGV vehicles to production nests, including stations and machines, performing tasks for subsequent stages of production. AGV vehicles transport components to production nests. However, their batteries have to be charged periodically and in appropriate time in order to avoid situations in which AGV vehicle stops, it cannot get to the production nest and come back to the base. Therefore, one of the objectives of the WP3 is to monitor the energy consumption in realtime and predict moments when AGVs will not reach the target production nest and should be re-directed to the battery charging station. On the other hand, monitoring energy consumption of particular AGVs and other variables of the whole system may deliver useful information on possible malfunctions of AGV vehicle components, which leads to a decreased operational time of AGVs and may cause unnecessary delays in the production. Therefore, the second objective of the WP3 is to detect anomalies in the job sequence of AGV vehicles. Finally, costs of delivery tasks (measured e.g., as consumed energy or delivery time) will be estimated to manage the entire production through proper arrangement of delivery processes performed by AGV vehicles:

In order to achieve overall objective of the WP3 four specific objectives have been defined:

• Predicting electricity consumption of AGV vehicles for particular job sequence
• Detecting anomalies in job sequence performed by AGV vehicles
• Estimating the costs of delivery tasks based on given parameters of the data model
• Support to predictive maintenance by on on-line anomalies detection

Research tasks:

T3.1. Data capturing, storing, and preparing a representative data sets for training machine learning models

This task covers the identification of all data sources that can add some value to the data analysis process. The data sources provide historical data and data streams. The research works focus on:

• transformation of historical data will in native format, while data streams will be captured in time windows,
• storing the captured data in a dedicated data lake for avoiding costly structuralizing operations,
• cleaning the data to ensure its quality before training Machine Learning models,
• selecting of a subset of relevant features to avoid the curse of dimensionality.

T3.2 Development of particular Machine Learning models for predicting tasks

Predictive tasks performed within WP3 will rely on appropriate ML models. For different objectives defined and various data sets, we will create a bag of ML models that will use techniques of supervised and unsupervised learning.  The research works focus on:

• supervised techniques used on the historical data labeled with a class (e.g., information on the occurrence of failure of a component of the AGV),
• various algorithms, such as decision trees, neural networks, SVM, Bayesian networks, and others, and checking their suitability for the task,
• predicting energy consumption using regression approaches utilizing algorithms, such as linear regression, Bayesian linear regression, decision forest regression, neural network regression,
• tuning an optimal set of hyperparameters of the ML models, which might be different for each specific ML model and a data set,
• anomaly detection including testing the SVM and PCA-based methods, and the K-means algorithm,
• creation of Web services disclosing the ML models on the Cloud for testing, validating, and finally, using.

T3.3. Experimental evaluation of built predictive models and quality assessment

Predictive models evaluation to assess quality of prediction results. Evaluation techniques will depend on the ML model used. The research works focus on:

• calculation of several coefficients for regression models, such as the mean absolute error (MAE), the root mean square error (RMSE), and the coefficient of determination (R2),
• calculation of several quality measures for the supervised learning-based classification including sensitivity, specificity, accuracy, precision, F1 score, and the Matthews correlation coefficient,
• evaluation of the models in the cross-validation process,
• exploring with ROC curves and the AUC value the tradeoff between specificity and sensitivity.

T3.4 Predictive maintenance

Anomaly detection in CoBotAGV behavior will allow for avoidance of production stops by predictive maintenance. The research works focus on:

• transformation into behavioral patterns time series data collected in (T3.1) with minimized feature vectors related to basic tasks performed by CoBotAGV,
• automated pattern creation based on energy consumption, battery level, operation time, safety stops, used actuators, etc. The patterns will be continually enriched by observation,
• detection of anomalies by comparison between ongoing data versus patterns,
• comparison between historical and current parameters will allow detecting gradual degradation of parameters.

Research progress: