Predictive maintenance requires operational data to be collected, processed and analysed in order to predict failures and optimise asset management. However, the diversity of industrial environments and the variety of data sources make it challenging to implement scalable, flexible solutions. In this project, we present MEPI (Maintenance Environment for Predictive Intelligence), a predictive maintenance dashboard designed to monitor and manage industrial assets using deep learning, survival analysis, natural language processing, and computer vision techniques. The platform enables users to configure assets using meta-models, collect real-time data from sensors and devices, and generate predictive models on demand. Maintenance tasks can be scheduled manually or automatically by integrating textual and visual information. All of this functionality is accessible via a web-based dashboard consisting of configurable KPIs and a REST API.
(IoT) sensors, engineering reports, images and Global Positioning System (GPS) signals that can be
applied to diferent domains such as farming [
The platform is currently in its final stages of development and is being validated through two case studies in the manufacturing and agriculture sectors. The aim of these pilots is to assess the system’s efectiveness in improving operational eficiency and promoting more sustainable maintenance practices by reducing unnecessary interventions and optimising the use of resources.
The development of PdM systems in Industry 4.0 is based on the convergence of several technological
advances. Deep learning techniques such as Recurrent Neural Networks (RNNs), Convolutional Neural
Networks (CNNs) and survival models such as the Cox Proportional Hazards Model or Random Survival
Forests have shown strong potential to predict equipment failures based on time series data and historical
records [
In parallel, the growing adoption of Industrial Internet of Things (IIoT) devices enables real-time
monitoring of assets through distributed sensors that collect environmental and operational data such as
temperature, humidity, pressure and usage patterns. This sensor data is essential for feeding predictive
models and triggering timely interventions [
Several commercial and research-oriented platforms have addressed PdM using AI techniques.
Cloudbased services like Amazon Lookout for Equipment analyse industrial sensor data to detect anomalies
and forecast failures at scale. IBM Maximo Predict [
While deep learning and sensor-based approaches have dominated the PdM landscape, the application
of NLP in this domain is gaining attention. Maintenance logs, technician reports and work orders contain
valuable unstructured information that can improve failure diagnosis and maintenance planning when
combined with time series data [
Figure 1 shows the overall architecture of the system. The platform is organised into four main modules. The first module is responsible for configuration and data acquisition (see 3.1), allowing companies to define metamodels and collect sensor data from assets and industrial environments, including geolocation and multimedia inputs. The second module focuses on predictive analytics (see 3.2), where time series forecasting models and demand-driven survival analysis are trained using historical and contextual data. The third module focuses on intelligent task scheduling (see 3.3), which uses NLP and employee activity tracking mechanisms to efectively manage maintenance activities. The fourth module is the dashboard (see 3.4), which provides a web-based interface for users to monitor KPIs, Factory 1
Sensor 1 Dashboard
Task scheduling
Tasks
Tasks API
Workers Automatic NLP analysis of
word reports Factory 2
Sensor 2 dCaotnaf-irgeuardaitniogn and
Data-model API data-reading
API system configuration Predictive models
API
Alerts system
Deep-learning repository predictive-model-1 predictive-model-2 predictive-model-3 configure alerts, visualise asset status and manage maintenance workflows. The following subsections describe these modules in more detail.
The platform will be ofered on a Software-as-a-Service (SaaS) model, providing basic functionality free of charge with certain usage restrictions. For example, standard users will be able to configure a limited number of assets and build predictive models with predefined settings. Advanced features such as full dashboard customisation, advanced historical data access or integration with external enterprise systems are reserved for premium users.
All services and software components are deployed using Docker containers. This strategy enables horizontal scalability and flexible orchestration of the platform according to the performance requirements and deployment constraints of each industrial customer.
The first core component of the platform is the configuration and data collection module, which is based on a metamodel-driven approach. A dedicated API allows each provider, also known as tenant, to define its own configuration by specifying diferent types of assets, associated attributes and the events to predict. This flexibility enables the system to adapt to a wide range of industrial scenarios by modeling the specific characteristics of each environment, such as operational context, location or asset category.
Once configured, the Data Reader module collects real-time information from multiple sources, including IoT sensors, mobile devices with GPS, and multimedia inputs such as images, video streams, maintenance report or work orders. All incoming data is structured according to the defined metamodels and transferred to the backend via secure REST APIs. In addition, a Historical Data API is available for querying stored information, supporting both model training and retrospective analysis. This modular and extensible architecture ensures seamless integration with existing monitoring systems and industry protocols.
In terms of data preprocessing, the numerical data are subjected to missing value imputation using the average of the two most recent sliding windows. To enable merging with numerical data, multimedia data in textual form is converted into a fixed-length array using one of the following approaches: (1) obtaining the document embedding for the entire text; (2) splitting the text into chunks, obtaining an embedding for each chunk, and concatenating them; or (3) applying a binary representation based on a configurable bag-of-words approach, where each element indicates the presence (1) or absence (0) of predefined terms in the text.
This module is responsible for generating PdM models tailored to each asset. The system monitors the flow of event data using scheduled CRON jobs and automatically triggers training processes when suficient historical and contextual information is available. Model training is fully automated and orchestrated through configurable pipelines that run periodically.
To carry out this training, the data received from the Data Reader module is transformed into a regression problem. This is necessary because the information on the events to be predicted is initially presented as a binary signal, which makes it necessary to transform it into a numerical interpretation that facilitates the training process. Therefore, this transformation is considered a hyperparameter of the training process. The system takes into account various numerical approximations, such as linear, exponential, logistic and logarithmic.
The training pipeline performs multiple train-test splits and explores diferent configurations, treating the model architecture itself (e.g. recurrent networks, convolutional networks or survival analysis models) as a tunable parameter. This allows the system to identify the most appropriate prediction strategy for each use case. The evaluated models includes Facebook Prophet, Random Forest, Support Vector Regression, Multi-layer Perceptron and Convolutional Neural Networks. Furthermore, when using time series models, such as Facebook Prophet, the training process incorporates as a hyperparameter diferent approaches to the number of predictor variables used in the prediction of events. On the one hand, the univariate approach, which uses a single input variable and, on the another hand, the multivariate approach, which integrates multiple input variables to perform the prediction.
On the other hand, when using regression models, two strategies for the construction of the training dataset are evaluated. The first one consists of prediction without incorporating delays (no lags), while the second one integrates lags, using previous values to predict the current value. In addition, the training process adjusts the size of the time window (number of lags) considered, making it, in turn, a hyperparameter of the training process.
Trained models are stored in a central versioned repository, enabling comparison, retraining or
rollback as required. The orchestration of the entire machine learning lifecycle follows best practices
inspired by platforms such as MLflow [
The table 1 shows the results obtained in terms of MSE and RMSE for the diferent regression models evaluated by the system during the training process. In particular, the results of each model are detailed for each possible event transformation, without taking into account the lags in the prediction. As can be seen, the Random Forest model with the exponential transformation of the events is the one with the best performance, achieving both a lower MSE and a lower RMSE than the other models. On the other hand, the CNN and MLP based models perform significantly worse. This is because, given the large amount of data for training, relatively simple architectures were chosen for these models to RF SVR CNN MLP
Logarithmic no lags no lags no lags no lags no lags no lags no lags no lags no lags no lags no lags no lags no lags
MSE allow reasonable training times, which limited their ability to generalize efectively during the training process.
The Task Scheduling module co-ordinates the execution of maintenance activities based on the output of the predictive models. It supports both manual and automated scheduling, taking into account asset criticality, recent maintenance reports, technician availability, geolocation and time-to-failure estimates.
A key component of this module is the NLP subsystem, which analyses textual maintenance reports,
logs and work orders generated by technicians. Based on the Stanza NLP library [
The NLP pipeline automatically classifies reports to estimate their priority level as critical, moderate
or low. This is achieved by combining MarIA sentence embeddings [
Based on the insights extracted, the system can recommend the most suitable technician for a given task by taking into account their expertise, proximity and availability. Additionally, a GPS-based tracking subsystem enables real-time supervision of interventions, and embedded computer vision tools provide visual diagnostics and augmented reality overlays to enhance field support.
To improve task scheduling, computer vision techniques are employed. Augmented reality methods are used to assist workers during maintenance activities. For instance, a web service has been developed that enables workers to scan QR codes on machines using mobile devices to access critical information. This enables them to view the machine’s maintenance manual, failure history and maintenance records. They can also view predictions generated by trained models that estimate the likelihood of the machine needing maintenance.
The dashboard module ofers a centralised interface for PdM, managing maintenance tasks and visualising predictive insights. It is highly configurable, enabling users to define KPIs relevant to their operational context, such as estimated RUL, alert frequency, asset health or intervention history.
Through the interface, end users can access real-time alerts, track scheduled and ongoing maintenance tasks, and interact with data collected from the field. The dashboard includes a Kanban-style task manager, interactive charts, geolocation maps and sensor data visualisation tools. It also supports rolebased access, providing diferent views and functionality for technicians, supervisors or administrators.
This module serves as the primary access point for end users and is closely integrated with all the platform’s other components, providing a seamless and comprehensive overview of the maintenance ecosystem.
The current version of the platform incorporates essential modules for asset configuration, data acquisition, model training, task scheduling, and dashboard interaction. Ongoing work includes a pilot deployment in water treatment plants, with a focus on the predictive cleaning of filtration systems using sensor data.
Future developments will enhance the model training (see Section 3.2) pipeline with automatic
hyper-parameter tuning and extend the NLP capabilities to process technician reports during validation
phases. First, we plan to explore the integration of emotion detection into the analysis of maintenance
reports. Technicians often express implicit signals of urgency, stress or dissatisfaction in their written
feedback, which could provide additional value for task prioritisation and risk assessment [
This work is being developed by the TECNOMOD research group at the University of Murcia, in collaboration with Neoradix Solutions S.L., within the framework of the Spanish National Call for Public-Private Collaborative R&D Projects 2021 (reference CPP2021-008465) funded by MICIU/AEI /10.13039/501100011033 and the European Union NextGenerationEU/PRTR.
During the preparation of this work, the authors used DeepL for grammatical and spelling correction. After using this tool, the authors reviewed and edited the content as needed and takes full responsibility for the publication’s content.