Factory Data That Gets Collected but Never Analyzed
Collect 3,000 equipment tags every second and you accumulate 259.2 million data points a day. At just 8 bytes for the timestamp and 8 bytes for the value, that is roughly 4 GB per day before compression, or about 1.5 TB per year. Yet when the quality team asks, "What was the mold temperature when last month's defective lot was produced?", the answer often takes days.
The problem is rarely the volume of data. It is usually how the data is stored.
MES production records are event-based transactions, such as "Work order completed, 480 good units." Process data, by contrast, is a continuous stream of values along a time axis. Because their write and query patterns are fundamentally different, they need separate storage designs.
Historian vs. Time-Series DB vs. RDB: What Goes Where
| Type | Strengths | Best suited for |
|------|-----------|-----------------|
| Commercial historian | Industrial protocol connectivity, process-optimized compression, long-term retention | High-volume process tags from the control layer |
| Open-source time-series DB | High write throughput, flexible queries, low licensing cost | IoT sensors, aggregated data for analytics |
| RDB | Transactional integrity, relational joins | Work orders, lots, inspection results, tag metadata |
Four criteria should drive the choice:
Trying to run everything on a single RDB eventually leads to bloated indexes and slow queries.
Designing Tag Structures and Collection Intervals
Naming Conventions and Metadata
Following the ISA-95 hierarchy (site, line, equipment, signal) makes tag names readable and searchable, for example `P1.L2.INJ03.MOLD_TEMP`. Each tag should also carry metadata such as unit, normal range, collection interval, and source device, so that the same value is never interpreted two different ways downstream.
Choosing a Collection Method
Compression, Downsampling, and Retention
A common setup keeps raw data for a few months while retaining tiered aggregates for the long term: 1-minute aggregates (mean, min, max, standard deviation) and 1-hour aggregates. A 1-minute aggregate cuts point count to one-sixtieth of 1-second raw data; a 1-hour aggregate cuts it to one 3,600th. Storing only averages wipes out momentary spikes, so keep min and max values alongside them.
Some windows must never be compressed away. The periods just before and after alarms, and the production windows of defective lots, need a protection policy that preserves them at full resolution. Without it, root-cause analysis becomes impossible.
Linking Time-Series Data to Production Records and Lots
The key is contextualization: using MES work orders, lot start and end times, and the equipment involved to define slices of time-series data.
With this structure in place, an analysis like "compare mold temperature distributions across 10 defective lots and 10 good lots" takes just a few lines of query.
A Phased Roadmap and the POLYGLOTSOFT Approach
Rather than targeting every tag in the plant from day one, start a pilot with one or two critical machines that drive the most defects and 50 to 200 key tags. Validate naming conventions, retention policies, and lot-linking queries in the pilot, then expand line by line to avoid costly rework.
Drawing on our MES and IoT gateway expertise, POLYGLOTSOFT delivers an end-to-end design covering tag structure, collection intervals, storage architecture, and lot-linked query screens. If your equipment data is piling up without being put to use, let's start with an assessment of your current collection setup.