Digitise quality control to cut rework and scrap

Submitted through the COP31 Sustainable Transformation Awards · SKD Türkiye (WBCSD Global Network Partner)

Borçelik Çelik Sanayii Ticaret A.Ş. is a flat steel producer in Türkiye with more than 1,000 employees.

Quality control on the production lines relied on periodic checks by operators. On steel surfaces moving at high speed, nonconformities such as scratches, holes, oil stains and acid drops are difficult to detect visually. Where they are missed or identified late, the result is deviation from the customer order, second-quality product, rework and scrap, each of which consumes energy and raw material a second time.

The associated losses are measurable. In 2024, acid drop nonconformity on the Continuous Pickling Line (CPL) alone caused 42 coils covering 182 tonnes to deviate from order specification, and 7 coils totalling 5.5 tonnes were downgraded to second quality. Compliance with international quality management systems, product standards and customer-specific requirements made addressing this a commercial as well as an environmental priority.

Location of the initiative: Türkiye – flat steel production lines (Continuous Pickling, Electro Cleaning and Coil Slitting)

The company started the project in November 2024 and commissioned the system in December 2024 to identify process-related quality losses at an early stage. Developed under Borçelik's technical leadership and operated on-premise, the machine vision system combines line-side imaging, image pre-processing and line-specific artificial intelligence (AI) models to detect nonconformities on fast-moving steel surfaces in real time. Each valid detection is linked to production metadata and retained as traceable image evidence, supporting intervention on the line rather than downstream identification.

Quality control moved from operator-dependent periodic checks to a data-based decision mechanism operating 24 hours a day. Operators respond to system alerts instead of monitoring the surface continuously, and every alert is validated in the field by the quality teams.

BorVision has been commissioned on the Continuous Pickling Line (CPL), Coil Slitting Line 2 (CSL2) and Electro Cleaning Line (ECL). Implementation is ongoing on SPM and CGL3, while MCSL1 and BCTL2 are included in the planned rollout. In the medium term, the standardised architecture also provides a basis for deployment at other group companies.

Figure 1: BorVision Quality inspection Architecture

BorVision on-premise quality inspection architecture. Encoder-triggered line-scan imaging and controlled illumination are combined with production metadata, image pre-processing and line-specific AI models. Valid detections are recorded with coil, metre, surface region and defect information, then presented through the BorVision web interface for field review and validation.

Figure 2: CPL Acid-Drop-Related Quality Outcomes

Acid-drop-related quality outcomes on the CPL before and after BorVision commissioning. Between the 2024 baseline and 2025, tonnage deviating from customer order specifications decreased from 182.0 to 9.6 tonnes, while second-quality tonnage decreased from 5.5 to 0.9 tonnes. Reductions of 94.7% and 83.6%, respectively, are calculated on a tonnage basis.

Sustainability impact

Climate

This initiative targets Scope 1 and Scope 2 emissions from the energy consumed when material is passed through a line a second time, and Scope 3, Category 1: Purchased goods and services, through the raw material that a second pass would otherwise consume. The climate effect arises from avoided rework: each tonne that does not require a second pass avoids the energy, raw material and associated emissions of that pass, so earlier detection translates into avoided consumption.

Based on the company's calculation, the reduction potential is assessed under the GHG Protocol, covering Scope 1, Scope 2 and Scope 3, by relating avoided rework tonnage to the specific energy consumption of each line and to corporate emission factors. Results are tracked in tonnage terms, covering deviation from order, second-quality and rework tonnage, rather than as an absolute tCO2e figure, and no absolute emissions saving is reported.

Compared with the 2024 baseline year, deviation from order on the CPL line fell from 182 tonnes to 9.6 tonnes and second-quality output from 5.5 tonnes to 0.9 tonnes in 2025, reductions of 94.7% and 83.6% respectively.

Nature

Resource efficiency is the direct nature-related effect. Material that would otherwise have been reworked or scrapped remains in the product, reducing the raw material drawn in per tonne of saleable steel. The saving is tracked through avoided rework tonnage, and the company does not quantify material or water use separately.

Social

The system changed the working conditions of line operators. Rather than monitoring the steel surface continuously, they act on system alerts, which reduces the attention load of the task, improves ergonomics and supports a safer working environment from an occupational health and safety perspective.

Operators also shaped the system, defining the nonconformity types, the accept and reject criteria and the response procedures triggered by an alarm. Across production and quality teams, the change strengthened data-based decision-making.

Business impact

Benefits

The measured operational gains on the CPL line are the clearest benefit. Deviation from order fell from 42 coils and 182 tonnes in 2024 to 1 coil and 9.6 tonnes in 2025, and second-quality output from 7 coils and 5.5 tonnes to 1 coil and 0.9 tonnes, reductions of 94.7% and 83.6% in tonnage.

System performance supports these results. Overall nonconformity detection accuracy reached the approved project performance indicator of 94%. Separately, the system generated 36 acid-drop detections on the CPL in 2025, of which 32 were confirmed by field teams. This corresponds to a field-confirmation rate of 88.9% (32/36); it is reported separately from the overall detection-accuracy indicator.

Retaining the system's core software, models and on-premise infrastructure within the company avoided the recurring licence and maintenance costs associated with closed commercial systems, reduced external dependency and strengthened Borçelik's internal software and AI capability.

Customers receive more consistent product quality, and compliance with quality management systems, product standards and customer-specific requirements is easier to evidence because every alarm and validation is recorded.

Costs

The company does not disclose investment figures. The main cost drivers are hardware, comprising line scan cameras, LED lighting and on-premise computing for each line, together with the internal effort required for software and AI development, image labelling, model training and threshold optimisation, and the continuing work of field validation and false alarm analysis.

Costs are contained in three ways: the system was built in-house rather than licensed, removing recurring licence and maintenance fees; components were standardised so that each additional line requires configuration and model tuning rather than redevelopment; and development was supported by the TÜBİTAK 1711 Artificial Intelligence Ecosystem Programme and by academic and technology partnerships. The resulting low hardware and operating cost is what makes rollout to further lines financially sustainable.

Impact beyond sustainability and business

Co-benefits

The collaboration generated applicable industrial AI knowledge for the academic and technology partners as well as for the company, and gave customers more consistent product quality. Internally it created a growing labelled image dataset and a standard installation architecture that shorten commissioning on each additional line.

Potential side-effects

The main trade-offs are technical. Image quality varies with line conditions, the system can raise false alarms, and teams require an adaptation period before they rely on the alerts. These are managed by optimising cameras, lenses and lighting for each line, expanding the dataset, updating models and feeding operator and quality-team feedback back into the models through periodic performance reviews.

Typical business profile

The approach is most relevant to continuous production processes where quality is assessed visually and material moves at speed, beginning with iron and steel but extending to other continuous manufacturing lines with surface quality requirements. The functions involved are production, quality, automation and information technology, supported by an in-house smart manufacturing team. It is suited to plants operating several lines, where a standardised architecture can be reused, and to companies with sufficient in-house software capability to maintain models rather than procure them.

Approach

1. Quantify the losses: Identify where process-related quality losses concentrate and quantify them for a base year in physical terms, covering coils and tonnes deviating from order, second-quality tonnage and rework tonnage.

2. Define the detection criteria: Agree the nonconformity types and the accept and reject criteria with line operators, together with the operational response once an alarm is raised.

3. Install the detection layer: Fit line scan cameras and LED lighting on a first line and build the image processing, anomaly detection and AI classification layers on-premise, with a web interface, alarm handling and data recording.

4. Train and tune the models: Collect images, label the data, train the models and optimise detection thresholds against the balance between missed defects and false alarms.

5. Validate in the field: Confirm every alarm on the line, record confirmations and false alarms, and feed the results into model updates through periodic performance meetings.

6. Track a fixed indicator set: Monitor detection rate, false alarm rate, deviation-from-order tonnage, second-quality tonnage and rework tonnage, and compare each line with the base year.

7. Standardise before scaling: Standardise the camera, lighting, data collection, modelling, alarm, reporting and web monitoring components so that line-specific differences are handled in configuration and model tuning rather than redevelopment, then roll out line by line.

8. Secure research partnerships: Combine public research funding with partnerships with a research institute, a university and a technology provider to raise the technical maturity of the models.

Stakeholders involved

• Project leads: The Smart Manufacturing Technologies team ran the work technically, setting the detection criteria with the line functions and holding the periodic performance reviews at which model updates are agreed. Senior management owns the project and reviews investment decisions, performance indicators and rollout plans regularly under the group's climate, people and innovation sustainability strategy.

• Company functions: Production teams are responsible for operational use and process improvement, quality teams for alarm validation and quality decisions, automation teams for field integration and system continuity, and information technology teams for the data infrastructure and system reliability. Line operators defined the nonconformity types, the accept and reject criteria and the post-alarm response procedures.

• Main providers: A technology provider contributed to system development alongside Borçelik's internal teams. The collaboration operated within Borçelik's in-house development model; ownership and operational control of the core code, models and on-premise infrastructure remained with the company rather than being licensed in as a closed system.

• Other: The project was carried out under the TÜBİTAK 1711 Artificial Intelligence Ecosystem Programme, with the Artificial Intelligence Institute of TÜBİTAK BİLGEM as public research partner and Bursa Technical University as academic partner. The two partners raised the technical maturity of the detection and classification models while the company supplied the line data and process knowledge, combining public research infrastructure with private sector process knowledge.

Key parameters to consider

The building blocks, comprising line scan cameras, LED lighting, image processing and AI classification, are established technologies; the demanding element is tuning them to each line. The project ran from November 2024 to commissioning in December 2024, with the first full year of comparable results in 2025 against a 2024 base year.

Key prerequisites include:

• line-side camera and lighting installation;

• on-premise computing capacity;

• a labelled image dataset; and

• field teams able to validate alarms.

Expected operating life is open-ended, since the dataset continues to grow and models are updated continuously. Development drew on the TÜBİTAK 1711 Artificial Intelligence Ecosystem Programme.

Implementation and operations tips

The recurring difficulties are variation in image quality between lines, false alarms and the adaptation period before teams rely on the system. Camera, lens and lighting optimisation, a wider dataset, model updates and structured field feedback address all three.

Involving line operators in defining the nonconformity types and the accept and reject criteria is what made the alerts operationally credible. Wider adoption is supported by low hardware and operating cost, a standard installation approach that avoids redevelopment for each line, in-house technical capability and partnerships with public research and academia.

External links

• Group sustainability reports