Catalyst Size Vision Inspection System

Overview

A cylindrical catalyst extrudate less than 2 mm in diameter and only a few millimeters long looks insignificant, but the size distribution of millions of them directly affects reactor flow uniformity, bed pressure drop, catalytic activity and service life. Downstream customers usually set size specifications (for example “length 3–15 mm ≥95%”), and an off-spec batch can mean the whole batch is rejected.

Most lines still rely on offline sampling: finished product → sampling → manual sieving and measuring in the lab → report.

• Slow feedback: results take 2–8 hours, so a large part of the batch is already produced when a problem is found.
• Low coverage: only 0.1%–1% is sampled, hiding the distribution within a batch and missing local abnormal sections.
• Inconsistent: manual sieving and calipers introduce bias from operators, caliper zero and judgment criteria.
• Hard to trace: paper or Excel records are hard to match with process parameters, making root cause analysis difficult.

The Catalyst Size Vision Inspection System is deployed as an independent station at the end of the line. Machine vision AI models measure the diameter and length of every particle in real time and grade and sort them into A/B/C ranges, turning after-the-fact sampling into particle-by-particle inspection with real-time rejection.

Metrics

Metric	Description
Diameter	Short side of the minimum bounding rectangle, sub-pixel accuracy
Length	Skeleton arc length, supports curved pieces
Particle size distribution	Live count and share by length range
A/B/C ratios	Ratios of the three ranges and batch pass rate
Output count	Cumulative count / estimated mass, output per shift
SPC	Cp / Cpk, out-of-control points on control charts

Inspect and sort

As each particle is judged, multi-channel air jets divert it by range in real time, with ≤100 ms from imaging to air jet:

• A · Pass: in spec, continues to the product bin
• B · Rework: too long but usable, blown to the rework hopper and returned to cutting
• C · Reject: short or broken, blown to the waste bin

A/B/C limits and action mapping (air jet / alarm / record only) can be edited online per product and batch and take effect immediately without stopping the line. All operations are logged for audit.

Software

• Live inspection: shows particle contours and A/B/C results frame by frame, with batch counts per range and the latest measurements; switches between live camera and offline folder testing. Insufficient lighting is detected and flagged to keep bad data out.
• Report dashboard: particle count, pass rate, NG rate, mean length ± σ and mean diameter at a glance; length trend, range pie chart and length histogram show process stability; linked statistics and distribution tables export to Excel and are traceable by batch / shift / time.

Compared with manual sampling

Item	Manual sampling	This system
Coverage	0.1%–1% sampling	Every particle
Response	2–8 hours	≤100 ms
Consistency	Calipers, varies by operator	Uniform algorithm, sub-pixel measurement
Handling	Traced after the fact	Real-time alarm + rejection
Data	Excel / paper records	Structured database + reports
Labor	1–2 dedicated sieving staff	Automatic, monitoring only

Architecture

• Image acquisition: industrial camera + bi-telecentric lens + collimated LED backlight + vibratory feeding for an even single layer.
• Edge computing: image pre-processing, model inference, geometric measurement and range judgment.
• Application: live view, metric visualization, SPC charts, alerts, reports and traceability, with a LIMS interface reserved.

Image source, algorithms and judgment rules are pluggable interfaces, and the system can be customized for other materials, metrics and sorting rules.