Case Study: Failure Prediction in a Manufacturing Plant

Project Context - Real Challenge, Concrete Solution

The following case study presents a real implementation of a predictive maintenance system in a medium-sized manufacturing plant, where GCAIT served as the technology and implementation partner. The project demonstrates how a systematic approach, from diagnosis through pilot to scaling, can deliver measurable business benefits.

Client - Plant Characteristics

The client was a medium-sized manufacturing plant in the automotive industry, specializing in automotive component production. The plant employed approximately 500 people and achieved revenues of around PLN 200 million annually. Production operated in shift mode - three shifts, 24 hours a day, 7 days a week, meaning machines operated virtually non-stop.

This intensive machine utilization meant that every failure caused significant losses - not only in lost production but also in contractual penalties for late deliveries and potential loss of customer trust.

Challenge - Chronic Downtime Problems

The client was struggling with chronic problems of unplanned downtime on the mechanical processing production line. Key machines - CNC machining centers - experienced failures an average of 2-3 times per month, translating to approximately 30-40 failures annually.

Each failure caused average downtime lasting 4-8 hours, which at a downtime cost of approximately PLN 25,000 per hour meant that a single failure cost the company between PLN 100,000 and PLN 200,000. On an annual basis, the estimated cost of downtime was PLN 1.5-2 million, representing a significant problem for plant profitability.

The previous preventive strategy, involving machine inspections every 3 months, was ineffective. Failures occurred despite regular inspections, showing that inspection schedules weren't aligned with actual machine condition. Moreover, some parts replaced during inspections were in good condition, meaning wasted resources - replacing parts that could have worked much longer.

Phase 1: Diagnosis - Understanding the Problem Before Solving It

The first project phase lasted 4 weeks and aimed to thoroughly understand the problem, infrastructure, and available data. This enabled designing a solution tailored to plant specifics and ensuring the project would be based on real needs, not just general assumptions.

Infrastructure Audit - Understanding the Environment

We began with a detailed infrastructure audit to understand what machines and systems we were dealing with. The plant contained 12 CNC machining centers from different manufacturers - DMG Mori, Mazak, Fanuc - meaning a diversity of technologies and communication protocols.

Machine ages ranged from 3 to 15 years, showing that some machines were relatively new while others were already advanced in age. This diversity required an individual approach to each machine, as older machines may have different failure patterns than new ones.

Existing systems included SCADA (Siemens WinCC) for process monitoring, a proprietary MES system for production management, and CMMS (IBM Maximo) for maintenance management. This showed the plant already had some digital infrastructure that could be leveraged.

Available data included machine logs from SCADA systems, failure history in the CMMS system, and process parameters in the MES system. This was a good starting point, but analysis showed this data wasn't being used to predict failures - it was only recorded after the fact.

Failure History Analysis - Identifying Patterns

We analyzed 3 years of failure history from the CMMS system, enabling pattern identification and understanding which failures were most frequent and costly.

Analysis showed the most common failures involved spindles (35% of all failures), hydraulic systems (25%), controls (20%), and other components (20%). This showed spindles were the biggest problem, both in terms of frequency and cost.

Spindle failures were the most costly - not only due to frequency but also due to the longest repair times and expensive spare parts. Spindle repair often required bringing in external specialists and spare parts, extending downtime.

Importantly, analysis showed that spindle failures were often preceded by increases in temperature and vibration. This was a key discovery - if we could detect these increases in advance, we could predict failures and plan repairs during scheduled stops instead of reacting to failures at the most inconvenient times.

GC AIT Recommendation - Strategic Pilot Approach

Based on the diagnosis, we recommended a pilot on the 3 oldest machines with the highest failure rates, focusing on spindle monitoring. This approach allowed verification of assumptions on a small scale before deciding on full deployment, reducing project risk and allowing solution adjustment to plant specifics.

Phase 2: Pilot - Verifying the Concept in Practice

The second project phase lasted 12 weeks and aimed to verify the concept on a small scale before deciding on full deployment. The pilot allowed testing the solution in real production conditions, identifying problems, and adjusting the solution to plant specifics.

Sensor Installation - Building Monitoring Infrastructure

On each of the 3 pilot machines, we installed a comprehensive monitoring system enabling continuous machine condition tracking. Vibration sensors - 3-axis accelerometers mounted on the spindle and main bearings - enabled detection of changes in vibration characteristics, which are often the first signal of bearing or spindle problems.

PT100 temperature sensors mounted on the spindle and in hydraulic oil enabled temperature monitoring, which is a key machine condition indicator. Temperature increases can indicate cooling problems, bearing wear, or hydraulic oil issues.

Current sensors measuring spindle motor power consumption enabled detection of load changes that may indicate problems with tools, material, or the machine itself. An IoT gateway aggregated data from all sensors, buffered it in case of communication problems, and transmitted it to the analytical system.

Data Integration - Connecting Different Sources

Key to system effectiveness was integrating data from many different sources. New sensors provided real-time data at a frequency of 1 measurement per second, enabling detection of rapid changes in machine condition.

CNC controller data retrieved via OPC UA contained process parameters such as rotational speed, feed, cutting depth, and machine alarms. This enabled correlation of process parameters with machine condition and detection of patterns leading to failures.

Service history from the CMMS Maximo system contained information about previous repairs, replaced parts, and service costs. This enabled understanding machine history and identifying failure patterns. MES system data - production orders, recipes, process parameters - enabled understanding operational context and correlating process parameters with machine condition.

Building the AI Model - From Data to Predictions

The predictive model was built using advanced machine learning techniques. The Ensemble algorithm combined Random Forest, which is effective at detecting patterns in static data, with LSTM (Long Short-Term Memory), which is specifically designed for time series analysis and detecting long-term trends.

Features used by the model included vibration RMS (root mean square), which is a good indicator of overall vibration level, frequency spectrum that enables identification of specific problems (e.g., bearing problems often manifest in specific frequency bands), temperature, and 7/14/30-day trends that enable detection of long-term changes in machine condition.

Model training used 2 years of historical data from the CMMS system and 6 weeks of real-time data from new sensors. This combination of historical and real data enabled building a model that understands both long-term patterns and current machine condition.

Model validation used 20% of data for testing and cross-validation, ensuring the model wasn't overfitted and would perform well on new data.

Pilot Results - First Successes

During the 12-week pilot, the system demonstrated its effectiveness. Three failures were predicted 2-7 days in advance, enabling scheduling repairs for weekends when production was already stopped. This showed the system can significantly reduce the impact of failures on production.

One false alarm occurred, which was used to correct the model - this showed the system learns and improves over time. Most importantly, no unexpected failures occurred on pilot machines, showing the system effectively detects problems in advance.

Estimated savings from the pilot were approximately PLN 200,000 - mainly from avoided downtime. This showed that investment in the system can pay back quickly.

Phase 3: Scaling - Extending Success Across the Entire Plant

After a successful pilot that confirmed solution effectiveness, we proceeded to extend the system to all machines in the plant. The scaling phase lasted 8 weeks and used pilot experience to accelerate deployment and ensure solution consistency.

Extension to All Machines - Leveraging Experience

After the successful pilot, we installed the system on the remaining 9 machines. We used pilot experience to accelerate deployment and ensure solution consistency.

Sensor installation standardization enabled fast and consistent deployment on all machines. We developed detailed installation procedures ensuring sensors were mounted in the same places and in the same way on all machines, facilitating later analysis and data comparison.

OPC UA integration templates for different controllers enabled fast integration with machines from different manufacturers. We developed a standard data mapping template that could be easily adapted to specific controller characteristics, significantly accelerating the integration process.

Transfer learning - using the pilot model as a base for new machines - enabled fast model adaptation to new machines. Instead of training models from scratch, we used knowledge from pilot models and adapted them to new machine specifics, significantly accelerating the process and improving prediction quality.

Process Integration - Incorporating the System into Daily Work

Key to project success was incorporating the system into daily work processes so it became an integral part of operations, not just an additional tool.

CMMS integration enabled automatic creation of service orders on AI system alerts. This eliminated the need for manual order entry and ensured all alerts were properly tracked and executed. Technicians received orders directly in the CMMS system they already used daily.

Email and SMS notifications for technicians and managers ensured important alerts were immediately visible even when technicians weren't at computers. This was particularly important in shift mode when technicians may be in different parts of the plant.

Dashboards tailored to different user needs - a duty view showing current alerts and machine status, a maintenance manager view showing trends and statistics, a management view showing key metrics and ROI - ensured each user received information appropriate to their needs.

Training - Preparing the Team

A comprehensive training program ensured all users received appropriate skills for using the system.

Maintenance technicians underwent 2-day workshops on system operation and alert interpretation. They learned how to interpret different alert types, assess their importance, and what actions to take in response to different signals. This was crucial as technicians are end users of the system and their skills determine solution effectiveness.

Operators underwent 4-hour basic training that enabled them to understand how the system works and interpret basic alerts. This was important as operators are often the first to notice machine problems.

Management received a dashboard and report presentation that enabled them to understand how to use the system for strategic decision-making and monitoring maintenance effectiveness.

Results After 12 Months

Key Metrics

Metric	Before	After	Change
Unplanned Downtime	28 annually	8 annually	-71%
MTBF (days)	45	85	+89%
Average Repair Time (h)	6.2	4.1	-34%
Spare Parts Costs	PLN 480k	PLN 320k	-33%
OEE Availability	87%	94%	+7pp

Financial Analysis

• Total Investment: PLN 420,000 (sensors, integration, licenses, training)
• Year 1 Savings: PLN 980,000 (downtime + parts)
• ROI: 133% in the first year
• Payback Period: 5 months

Challenges and Lessons - What We Learned During the Project

Every AI implementation project brings valuable lessons. Below we present the main challenges we faced and how we solved them, which may be helpful for other projects.

Challenge 1: Historical Data Quality - Foundation for Analysis

The problem was that failure history in the CMMS system was incomplete - not all interventions were recorded, descriptions were inconsistent, some failures were classified as "other" without details. This made building predictive models difficult, as they require high-quality historical data.

The solution included workshops with technicians where we standardized failure classification and taught technicians how to properly record interventions. We also supplemented history based on interviews with experienced technicians who remembered details of previous failures. This showed how important data quality is at the beginning of a project.

Challenge 2: Technician Resistance - Change Management

The problem was that some technicians perceived the system as "control" or a threat to their expertise. They feared the system would replace their knowledge and experience, or would be used to evaluate their work.

The solution included involving technicians in the alert validation process - technicians were asked to assess whether alerts were correct and useful. This showed the system is a tool supporting their knowledge, not replacing it. We also rewarded technicians for observations and feedback, showing their knowledge was valuable and appreciated.

Challenge 3: False Alarms - Balance Between Sensitivity and Precision

The problem was that initial models generated too many alerts, leading to "alarm fatigue" - technicians started ignoring alerts because most were false. This reduced system effectiveness because even real alerts were ignored.

The solution included tuning alert thresholds, introducing alert levels (warning vs critical), and a feedback loop from technicians enabling continuous model improvement. This showed how important it is to find a balance between system sensitivity (detecting all problems) and precision (only real problems).

Challenge 4: Legacy Machine Integration - Backward Compatibility

The problem was that the two oldest machines didn't have OPC UA, only the older Modbus protocol. This required additional integration that wasn't planned at the project start.

The solution included installing a Modbus→OPC UA gateway that converted data from Modbus protocol to OPC UA, and mapping Modbus registers to an OPC UA semantic data model. This showed how important solution flexibility is and the ability to integrate with different technologies.

What's Next? - Future Possibilities

After successful predictive maintenance system deployment, the client plans to extend the system to additional areas, showing how success of one project can open doors to further innovation.

Assembly lines - robots and presses - are the next area where the system can deliver benefits. Robots and presses have their specific failure patterns that can be predicted using similar techniques, enabling benefits extension across the entire plant.

Process parameter optimization is a natural extension of failure prediction. Instead of just predicting failures, the system can also optimize process parameters such as rotational speed, feed, temperature to maximize efficiency and quality while minimizing machine wear.

Quality prediction - detecting process drift before defects occur - is another area where AI can deliver benefits. The system can detect changes in process parameters leading to defect creation before they actually occur, enabling real-time process correction.

Summary - Key Success Factors

This case study shows that predictive maintenance is not science fiction or the future - it's proven technology that delivers measurable benefits today. The GC AIT team's experience confirms that the key success factors of this project can be applied in other implementations.

Clear business goals - downtime reduction, measurable ROI - were the project foundation. From the start, we knew what we wanted to achieve and how to measure it, enabling informed decision-making and project success evaluation.

Pilot before scaling - proving value on a small sample - enabled verification of assumptions and solution adjustment before full deployment. This reduced project risk and increased probability of success.

People engagement - technicians as partners, not change victims - was crucial for success. Technicians were involved in the process from the start, their feedback was valuable, and their knowledge was appreciated. This showed the system is a tool supporting their work, not replacing it.

Process integration - system embedded in daily work - ensured the system was actually used, not just installed. CMMS integration, notifications, dashboards - all this made the system an integral part of daily work.

Continuous improvement - feedback loop, iterative improvements - enabled continuous system enhancement. The system wasn't static - it learned and improved over time, increasing its effectiveness and user trust.

If you're considering a similar implementation in your plant - let's talk. Our experts will help you follow the same path, from diagnosis through pilot to scaling, leveraging experience from similar projects and avoiding typical pitfalls. Every plant is different, but the fundamental principles of success remain the same.