What I examine before a manufacturer trusts its data

Our customers expect us to show, for any serialised part, the material certificate and heat lot behind it, every operation and operator stamp, each special-process run and the inspection that accepted it, and to keep those records for as long as the aircraft programme and the contract require, which is often decades. Quality management standards for aviation suppliers, customer flow-down clauses and special-process accreditation audits all test this, and an engine customer's audit is less forgiving than any of them. I treat traceability as a performance rather than a property of a system. We measure the hours, people and corrections needed to answer a genealogy question, run a timed drill every quarter with a question set by our quality department, and report the result to the audit and risk committee. Our serial genealogy record now covers 2.6 million parts made since 2020 in six plants. I am explicit about the parts it does not yet cover.

Most of what a manufacturer calls operational technology data is collected for the machine, not for the analyst. A five-axis machining centre reports spindle load and alarms in its controller's own vocabulary; a furnace logs thermocouple readings at intervals set when it was installed; a coordinate-measuring machine writes results to a file named by whoever ran the program. Before any of that data supports a decision, someone has to establish what each signal means, how often it is missing and whether a timestamp is local or universal. I run data quality as plant work. Each plant has data stewards, half funded by my office, who own the rules for their equipment and fix problems at source. We publish completeness and validity for the fields that matter most, such as heat lot, serial number and furnace load, and we treat a falling score as a production issue, not a reporting one.

Predictive maintenance is valuable in machining when three conditions hold: the failure mode gives an early physical signal, the sensors capture it reliably and the plant has a recorded, coded history of enough failures to learn from. Spindle bearings on our five-axis centres meet the first two conditions on some machines; very few of our failure records met the third when we began. I insist on counting coded failures before approving a model, and I separate condition monitoring, which flags a measurement outside its limits, from prediction, which estimates time to failure. The first is cheaper, explainable to a maintenance planner and often enough. We use prediction on fourteen machines at Longueuil with good histories and condition monitoring on all sixty-four. I report precision, missed failures and the weekly alert load on planners, because a model that floods a planning office will be ignored however accurate it is.

In a multi-plant manufacturer, head office wants comparable figures and plants want their own detail and context. I have found that centralising the raw data rarely settles that argument. At Tessandier the plants own their operational data and its local detail. My office owns a dictionary of 46 shared definitions, among them scrap, rework, first-pass yield, overall equipment effectiveness and unplanned downtime, each with a named owner, a formula, a list of exceptions and a monthly reconciliation to the general ledger where money is involved. A plant may keep additional measures for its own use, but it reports the shared ones exactly as defined. Disputes go to a definitions committee that includes plant managers and finance, and its decisions are published. The arrangement made our first cross-plant board dashboard three quarters late. It also made the figures on it ones that plant managers no longer contest.

Much of our plant data identifies people: operator log-ins on machines, inspector stamps, badge records and camera images on inspection lines. Quebec's private-sector privacy law, modernised in recent years, expects a company to assess privacy risks before new uses and before personal information is communicated outside the province, and to report serious confidentiality incidents. Our chief executive has delegated to me in writing the role of person in charge of protecting personal information, and I report to the audit and risk committee on it each year. Separately, many of our drawings and process sheets are controlled under US export rules and Canada's Controlled Goods Program, which limits who may see them and where they may be processed. Our AI register lists 23 uses. Any use that touches product acceptance, controlled technical data or individual employees needs a named owner, a written limit and approval from quality, legal or both before it runs.