Japanese Factories Keep Buying the Wrong AI Inspector. The Tool Choice Is a Data Choice.

PFN Visual Inspection, Hitachi’s Google Cloud pilot and NEC Invariant Analysis attack one inspection problem with three different bets on scarce data

Start with the task, not the vendor. A Japanese factory line produces parts that must not ship with a scratch, a stain, a bad crimp or a silent sensor fault, and the people who used to catch these things by eye are retiring faster than they are replaced. Three Japanese deployments answered that task in the past few years with three genuinely different tools: Preferred Networks Visual Inspection, Hitachi’s Omika Works pilot built on Google Cloud Visual Inspection AI, and NEC’s Invariant Analysis. Each has published, checkable output numbers. The differences between them are not differences in model quality. They are differences in what each tool assumes you can spare, and that assumption decides which one works on your line.

The three tools and what each one runs on

Preferred Networks Visual Inspection (PVI) is an image classifier built on the company’s own deep-learning stack. Its headline capability is training from very little: the official product page states a model can be trained from roughly 100 good samples and 20 defective ones, where earlier approaches needed tens or hundreds of thousands of labeled images, and annotation means sorting images into good and bad groups rather than marking defect locations, with the AI locating the flaw itself and showing it as a heatmap (Preferred Networks, link 🔗). The published demo tasks are small and concrete: carpet stain detection trained on 80 good and 40 defective images, and metal part scratch detection on 50 and 50, both documented on the English product page as of its November 2026 content. Industries named by the vendor cover automotive, electronic devices, steel, food, semiconductors and construction, and a 2022 press release documented a live demonstration of pen-nib good-or-bad judgment with Platinum Pen at the Smart Factory Japan show that October (Preferred Networks, 2022-09-27, link 🔗).

Hitachi’s Omika Works, the Hitachi factory network that builds information control systems, took a different route for one specific defect: crimped wire terminals on control panels, where a bad crimp means a contact failure in equipment designed to run for a decade. The proof of concept paired Hitachi’s industrial edge computer CE50-10 with Google Cloud’s Visual Inspection AI and reported a 100 percent defect detection rate, with a machine-learning model built in a day from about 100 sample images (Google Cloud case study, link 🔗). The engineer who ran the quality side is blunt about the old process: “visual inspection is time-consuming. Plus, no matter how careful we are in our inspection, we cannot completely prevent mistakes,” says Masayuki Takagaki of the MONOZUKURI Management Department.

Two further details in the case study matter for anyone scoping a similar build. First, the motivating constraint was not only accuracy but access: the existing inspection workflow was usable only by specific trained users, which breaks when headcount turns over, so the AI build was partly a knowledge-transfer project. Second, the versatility problem is stated openly by Hitachi’s own staff: building individual machine-learning models for every customer deployment “complicates their deployment across factories,” as Naohiko Irie of the Control System Platform Development Division put it, which is why the team turned to a cloud service for generalization and kept an edge computer for the line. Omika Works carries the WEF Global Lighthouse designation, received in 2020 as the first Japanese site so recognized, so the pilot’s home is a showcase plant with mature digital tooling, not a greenfield line (Hitachi press materials, July 2026).

NEC Invariant Analysis skips the camera entirely. The technology reads historical sensor streams from a plant, automatically extracts the stable relationships between sensor pairs, and flags the moment real-time data breaks one of those relationships. Its distinctive claim is that it needs no labeled failure history at all, that it catches “silent failures” that threshold monitoring misses, and that it is white-box: every invariant relationship is visible and simplified enough to compute fast (NEC, link 🔗). Packaged as the software product NEC Advanced Analytics Invariant Analysis, it targets plants, production lines, bridges, rail systems and water infrastructure, and NEC lists Lockheed Martin among published cases (NEC product page, link 🔗). The packaging itself tells you who is expected to run it: the product ships with an optional starter-pack support service covering operation lectures, installation help and question handling, explicitly pitched at sites with no experienced data analysts, and its version history, now at 2.7, reads like a maintenance log for industrial users, adding watch-notification filters that suppress alerts from momentary anomaly spikes and backup and restore functions alongside the analytics.

The company behind the first tool has been consolidating its industrial position through 2026. Preferred Networks, founded in 2014 and backed over the years by Toyota, Fanuc and NTT, shifted its deep-learning framework work from its own Chainer to PyTorch in March 2026, and on 2 June 2026 Mitsubishi Heavy Industries announced a business alliance with the company to co-develop mission-critical autonomous AI-driven machinery, combining Mitsubishi’s hardware, control and simulation technology with Preferred Networks’ AI foundation models, AI chips and compute infrastructure (Sina Finance, 2026-06-02, link 🔗). A vendor of inspection software that also owns chips, a supercomputer line and an industrial alliance is a different kind of counterparty from a point-solution vendor, and it matters for the maintenance-cycle question this piece returns to later.

Table 1: The three deployments against the same inspection problem

Requirement PFN Visual Inspection Hitachi x Google Cloud PoC NEC Invariant Analysis
What it inspects Product appearance (images) Product appearance (images) System behavior (sensor streams)
Data it demands 100 good / 20 bad images, sorted only About 100 sample images, one defect type Historical sensor time series
Published output 250+ customers; demo tasks with 80/40 and 50/50 image sets 100% defect detection rate in PoC; model built in a day Silent-failure detection without labeled failures; white-box model
Where it computes On-premise or edge, inference library provided Hitachi edge computer CE50-10 plus Google Cloud Runs against existing monitoring systems

Source: compiled from Preferred Networks, Google Cloud and NEC published pages cited in Sources

What the published numbers actually say

Read the outputs the way you would read a supplier’s test certificate, not a benchmark. Hitachi’s 100 percent is a real published result, and its scope is exactly one defect family on one connector type inside a proof of concept; the case study does not claim plant-wide coverage. The one-day model build is likewise scoped: about 100 sample images, one task, Google’s tooling. If your defect taxonomy has twelve entries, you are planning twelve of these builds, and the PoC result says nothing about how they degrade when lighting or suppliers change.

PFN’s number to watch is the data floor, not the customer count. Training an appearance model from 100 good and 20 bad samples, without pixel-level annotation, is the design bet that separates it from conventional machine vision, and the demo image sets (80/40 for carpet stains, 50/50 for metal scratches) are published on the product page so a buyer can replicate the scale of the task. The customer count is a messier read, which is worth knowing before you quote it: the Japanese product page says 250 or more customers, the English page says more than 300, and the company’s 2022 press release said more than 150 at that date. The trend is real; the exact figure depends on which page and when.

NEC’s published claim is the most different in kind. It does not promise a detection rate at all, because its task is not classification but relationship monitoring: the product detects when the system stops behaving like itself, which is how it claims to find silent failures that sit inside normal thresholds. The Fukuoka water purification plant work and the Lockheed Martin case are the public anchors. For an appearance defect like a scratch, it is the wrong instrument; for a drifting pump curve or a control loop that has quietly stopped responding, it is the only one of the three that applies.

The design tradeoffs behind each bet

Each tool made a bet about scarce resources, and the bets are incompatible.

  • PFN bet that labels, not images, are the scarce resource. Sorting a few hundred images into two piles is something a quality team does in an afternoon; marking defect boxes for thousands of images is a project. The heatmap output exists to make the model’s judgment legible to the floor, and the inference library is packaged so the system runs where the line runs.
  • Hitachi and Google bet that the model-build itself was the bottleneck, so they compressed it to a day on managed cloud tooling and pushed inference to an edge box. The tradeoff is a cloud dependency in an environment where many Japanese manufacturers refuse to move images off-premises, and a per-defect build cycle rather than one general model.
  • NEC bet that the sensors were already installed and talking, so the cheapest path to anomaly detection is to mine the relationships already in the historian. The tradeoff is that appearance defects leave no sensor trace, and the white-box output requires engineers willing to read invariant formulas, which is a staffing assumption dressed as a technical one.

The choices also encode different answers to the staffing question the whole category exists for. Hitachi’s own engineer described the problem precisely: “inspections aren’t always carried out the same way, and it takes an experienced eye to accurately discern these things,” as Takeshi Saito of the Industrial IoT and Robotics Engineering Department put it in the case study. PFN’s answer is to compress what the eye learned into a small sorted image set. NEC’s answer is to stop imitating the eye and monitor the machine’s self-consistency instead.

Explainability splits the same way, and it is not a rounding error in regulated settings. NEC ships white-box by construction: the invariant relationships between sensor pairs are explicit formulas, simplified for fast computation, so an engineer can read why an alert fired and challenge it. PFN ships post-hoc explainability: the model is a deep network, but the defect location is rendered as a heatmap so an inspector can verify what the model saw. Hitachi and Google’s route inherits the opacity of a managed cloud model, offset by the accessibility of the deployment: the edge computer runs the trained model where the line runs, and the build tooling lives in the cloud where a central team can manage it. If your quality organization must document why a unit was rejected, for an automotive customer or an auditor, the three answers differ enough to decide the purchase on their own.

Which tool for which line

The selection logic is short enough to run in a meeting.

  • If your defect is visible on the surface, your images exist already, and nobody has time to annotate boxes: start with the PFN approach. Its published floor of 100 good and 20 bad samples means a pilot is measured in days of sorting, and the heatmap gives your inspectors something to argue with.
  • If your defect is visible, one-dimensional, and the cost of a miss is catastrophic, the Hitachi and Google Cloud route is the one with a published 100 percent result, provided you accept a cloud-built model on an edge box and you scope your claim to that defect family when you present it internally.
  • If your problem is not a part but a system, meaning the product passes visual inspection and fails in the field, NEC Invariant Analysis is the only one of the three built for that, and it will ask for sensor histories rather than photographs.

One more check belongs in the meeting. Ask each vendor what happens when the line changes: new lighting, a supplier swap, a seasonal humidity shift. PFN’s retraining path is the shortest on paper because the data demand is smallest; Hitachi’s case study documents the initial build, not the maintenance cycle; NEC adapts continuously by design but will re-learn a relationship you actually wanted flagged. Whichever answer you cannot live with eliminates the tool, regardless of the benchmark numbers.

Run the decision as a worked example, because that is how it will actually happen. Take a mid-size electronics supplier in the Toyota supply chain, building connector assemblies with two visible defect types and a history of field failures no camera ever caught. The image tools are candidates for the first problem: sort a few hundred images into good and bad piles, follow the PFN path, and hold the vendor to its published floor of 100 good and 20 bad samples as the acceptance criterion for the pilot. If the plant already runs a historian with dense sensor coverage on the test benches, add the NEC route for the field-failure class of problem, white-box formulas and all, so the quality group can read the alerts. The Hitachi and Google Cloud route earns consideration when a single catastrophic defect family, like a crimp, justifies a dedicated model and a per-family build cycle, and when management accepts a cloud-built artifact on an edge box. The supplier that tries to buy one tool for all three problems buys the wrong one twice.

Context check against the production numbers

The tooling debate sits on a production base that METI’s monthly indices describe. In September 2026, Japan’s production index stood at 104.1 on the 2020 base, up 2.2 percent month on month and 1.9 percent year on year; shipments rose 0.6 percent on the month while inventories rose 2.5 percent, a combination that puts extra pressure on quality exits because unsold stock amplifies the cost of a slipped defect. Auto production climbed 5.5 percent on the month and semiconductor production equipment output jumped 13.8 percent in the same reading (METI, September 2026). Manufacturing value added runs at about 867.1 billion dollars, roughly 20.6 percent of GDP, on World Bank data for 2026. Every point of defect leakage on those lines is priced in scrapped units, rework hours and warranty exposure, which is why the inspector shortage keeps converting into software budgets. The three tools above are where that budget has demonstrably been spent, with published evidence, inside Japanese plants.

Data basis

Data as of November 2026. This comparison reads published product pages and case studies, not a hands-on benchmark: Preferred Networks Visual Inspection’s product pages (Japanese and English) and 2022 exhibition press release; the Google Cloud customer case study on Hitachi Omika Works, including quotations from Hitachi staff as published; NEC’s technology and product pages for Invariant Analysis; and METI production index releases for September 2026. Definitions: “detection rate” means the share of defective units flagged in the cited test scope (Hitachi: one defect family, PoC); “training data floor” means the smallest sample counts stated by the vendor for its demo tasks (PFN); “silent failure” is NEC’s term for anomalies that remain inside sensor alarm thresholds. Customer counts differ between vendor pages and are reported per page with dates. No independent re-testing was performed; every performance figure in this piece is the vendor’s own published number, attributed as such.

Limitations

  • All performance figures are vendor-published; none were independently re-run.
  • The Hitachi result covers one defect type in a proof of concept, not production-wide coverage.
  • NEC’s water-plant and defense cases are described without per-site accuracy figures.

Sources

This article was prepared by the expert team at Global Infinity Labs, drawing on insights and data provided through its MCP data services.

About Global Infinity Labs

Global Infinity Labs is an AI innovation company specializing in MCP data services.

Email: support@globalinfinitylabs.com