CASE STUDY

From Reactive to Predictive:

AI-Driven Fault Detection That Cut Unplanned

Machine Downtime by Over 65%

+20%

Fault Detection Rate

improvement over reactive regime

£101k

Annual Saving

estimated fleet-wide (18 machines)

<8 mo

Payback Period

full fleet deployment

Sector: Technical & Industrial Textile Weaving   |  Location: Northwest England   |  Fleet: 18 production machines   |  Pilot duration: 16 weeks

The Challenge

A Northwest England manufacturer of technical and industrial woven textiles — supplying filtration fabrics, geotextile reinforcement cloth, and industrial conveyor belting — was losing significant production time and revenue to machine breakdowns it could neither predict nor prevent.

Across its 18 rapier loom weaving machines, the business recorded 312 unplanned stoppages in a single 12-month period. With an average of 47 minutes lost per stoppage, and some machines suffering more than 28 hours of unplanned downtime per year, the financial and operational impact was substantial.

"We knew the machines were going to break. We just never knew which one, or when. By the time an alarm went off, the damage was already done."  —  Site Maintenance Manager

Three failure modes accounted for nearly 90% of unplanned stops: gradual tension drift in the weft feeder mechanism, wear in the rapier gripper head, and lateral misalignment of the heddle frames. All three are progressive — they develop over hours or days before causing a failure. Yet without continuous monitoring, every breakdown came as a surprise.

Impact Area

Pre-Pilot Position

Unplanned stoppages (fleet/yr)

312 events

Average stoppage duration

47 minutes

Total unplanned downtime

~245 hours per year (fleet-wide)

Worst single machine (annual)

28.4 hours unplanned downtime

Fabric waste from loom defects

~3.2 tonnes per year

Estimated annual revenue impact

~£118,000

Maintenance approach

Reactive + fixed-interval time-based only

The DAG Approach

Data Ascend Group deployed a proof-of-concept predictive maintenance system on the highest-impact machine in the fleet — selected because it accounted for a disproportionate share of total downtime and ran the widest variety of fabric constructions, making it the most mechanically demanding asset on the weaving floor.

The approach was built on three principles: non-intrusive installation (no modification to OEM control systems), edge-first processing (data analysed at the machine, not in the cloud), and progressive intelligence (the system becomes more accurate as it accumulates operational data).

Multi-Modal Condition Monitoring

A bespoke sensor array was designed and retrofitted to the pilot machine, capturing seven independent data streams simultaneously:

hat We Measured

Why It Matters

Vibration (3 locations, 3 axes each)

Detects bearing degradation, gear wear, and mechanical imbalance 8–14 hours before failure

Motor current draw (main drive + weft feeder)

Reveals increased mechanical load — an early signature of tensioner spring fatigue and rapier wear

Acoustic emission (rapier head zone)

Captures high-frequency stress events invisible to vibration sensors — correlates with heddle frame drift

Warp beam tension (2 measurement points)

Quantifies tension variance that precedes warp breakage and pattern defects

Component temperatures (4 points)

Thermal rise in gearbox and drive components flags lubrication failure and bearing degradation

Loom speed (pick counter)

Baseline reference — speed deviations under load indicate drivetrain stress

Machine Learning at the Edge

All sensor data was processed in real time by an industrial-grade edge computing gateway mounted directly at the machine — removing dependence on network connectivity and ensuring sub-second response times. DAG's ML pipeline comprised three layers of intelligence:

ML Layer

Function

Baseline characterisation

Three weeks of healthy-operation data established normal behaviour profiles for all seven signal streams across all running conditions and fabric types.

Anomaly detection

A deep-learning autoencoder continuously compared live signals against baseline. Reconstruction error above a calibrated threshold triggered an alert — catching deviations the machine's own controls could not detect.

Fault classification

A gradient-boosted classifier mapped detected anomalies to specific failure modes, enabling maintenance teams to attend with the right parts and skills — not just a general alarm.

The system does not replace your engineers. It tells them exactly what is wrong, on which machine, with enough advance notice to plan the response.

Real-Time Health Dashboard

Every machine receives a continuously updated health score (0–100), visible on a shop floor tablet and a web portal accessible from any device. A three-tier alert system — advisory, warning, and critical — gives maintenance teams graduated lead time to act, from a planned check at the next shift break through to an immediate targeted intervention.

The Results

The active monitoring phase ran for 10 weeks following the initial baseline and model training period. Results on the pilot machine were measured and recorded by DAG engineers throughout.

83%

Pre-Failure Detection

of fault events caught before production impact

6.8 hrs

Mean Advance Warning

average time between alert and failure

68%

Downtime Reduction

on pilot machine vs. prior 12-month baseline

Fault Detection Performance

Metric

Before DAG Pilot

During Pilot

Fault detection rate

~62%

▲ 82%  (+20 pts)

Faults caught before production impact

0%

▲ 83%

Mean advance warning time

None

▲ 6.8 hours

False positive alert rate

N/A

6.2%  (2 events)

Mean response time to fault

47 min (reactive)

▲ Planned — avg. 28 min

Downtime Reduction on Pilot Machine

Failure Mode

Pre-Pilot (annualised)

During Pilot (annualised)

Weft feeder tension drift

6.5 events/yr

▲ 1.0 event/yr

Rapier head wear

5.0 events/yr

▲ 1.5 events/yr

Heddle frame misalignment

3.6 events/yr

▲ 2.0 events/yr

Total unplanned downtime

28.4 hrs/yr

▲ ~9.1 hrs/yr  (−68%)

The Financial Case for Fleet-Wide Deployment

The pilot demonstrated a clear, measurable return on a single machine. Extrapolated across the full 18-machine fleet — applying the same fault detection improvement and downtime reduction rates — the business case for full deployment is compelling.

Financial Item

Estimated Value

Downtime reduction saving (18 machines)

~£82,000 per year

Reduction in fabric waste and scrap costs

~£19,000 per year

Total estimated annual saving

~£101,000 per year

Full fleet deployment cost (hardware, integration, Year 1 support)

~£68,000

Simple payback period

Under 8 months

5-year net financial benefit

~£437,000

A sub-8-month payback and over £400,000 in five-year net benefit — from a technology that requires no changes to your existing machines or control systems.

These figures are based on the client's own production rate and scrap cost data, applied to the fault reduction rates measured during the pilot. They do not include secondary benefits such as reduced emergency maintenance labour, extended component service life from planned replacement, or improved OEE scores for customer reporting purposes.

Why Data Ascend Group

This engagement was not a software demonstration. It was a ground-up engineering project — sensor selection, signal processing, model design, edge deployment, and operational integration — delivered by a team with hands-on manufacturing and industrial AI experience.

DAG Capability

What It Means in Practice

Multi-modal sensor engineering

We design and specify sensor arrays for the real machine environment — not generic IoT kits. Every sensor placement is chosen for mechanical relevance.

Edge-first ML deployment

Our models run at the machine. No dependency on cloud connectivity, no latency, no raw data leaving your site.

Domain-specific model design

We build models trained on your machines' actual behaviour — not pre-trained generic models that miss the nuances of your process.

Interpretable outputs

We don't give your engineers a black-box score. Every alert identifies the fault type, the affected component, and the recommended action.

Non-intrusive retrofit

Our systems connect to existing machines without modifying OEM controls, voiding warranties, or requiring production stoppages for installation.

Progressive accuracy

The system improves continuously as it accumulates more operational and fault data from your specific fleet.

Data Ascend Group works exclusively with manufacturers in the Northwest England region, combining deep industrial domain knowledge with applied machine learning capability. We are listed on the Made Smarter Technology Providers Directory and are experienced in supporting manufacturers through funded digitalisation programmes.

Could This Work for Your Operation?

The DAG predictive maintenance system is applicable to any manufacturing environment where machines run continuously, failure modes are progressive rather than sudden, and unplanned downtime has a direct cost. You do not need a digital CMMS, an existing IoT infrastructure, or in-house data science capability.

You may be a good candidate if:

•       Your maintenance team spends more time responding to breakdowns than planning for them

•       You have recurring failure modes on specific machines that your engineers already recognise but cannot reliably predict

•       Unplanned stoppages are causing you to miss delivery commitments or generate fabric waste or scrap

•       Your machines are between 5 and 20 years old and lack built-in condition monitoring

•       You are exploring Made Smarter funding and need a technology partner with a proven, deployable solution

A DAG scoping assessment takes half a day on site and is offered at no charge to qualifying manufacturers. We will identify your highest-impact failure modes, quantify your current downtime cost, and give you an honest assessment of whether predictive maintenance is the right intervention.

Talk to Data Ascend Group

Book a free, no-obligation scoping visit for your site.

info@dataascendgroup.com  |  dataascendgroup.com  |  Made Smarter Technology Provider

Client identity anonymised by mutual agreement. Results reflect measured pilot data on a single machine over a 10-week active monitoring period. Fleet-wide financial projections are estimates based on pilot results applied proportionally across the 18-machine fleet using client-supplied production cost data.