Engineering schools can do something factories cannot: break equipment on purpose. A university battery bench or bearing rig can produce hundreds of labeled fault recordings in a week. A production line sees a real fault a few times a year and will not share its recordings, because they reveal proprietary operating cycles. Federated learning trains one model across several sites by exchanging model updates instead of raw data, and it is the obvious way to join these two kinds of site. The obvious design, though, wastes the school's main asset. It treats the lab as one more client holding a fixed dataset. We propose instead that the lab act as a client whose data is chosen. The factories send back a privacy-protected summary of where the shared model is uncertain, meaning which operating conditions and which fault types. That summary becomes the list of experiments students run next, as credit-bearing project work.
The claim is concrete and can be falsified. A lab client steered by aggregate uncertainty signals from industrial clients should reach a given fault-detection accuracy on the factories' scarce real faults with far fewer physical experiments than a lab that runs a fixed or randomly chosen experiment plan. It should also beat the cheaper option of simply releasing a lab-pretrained model for factories to fine-tune. If that holds, the "dual-scenario" coupling of school and enterprise that the education literature describes qualitatively gets a measurable mechanism. That mechanism also addresses the enterprise data-hoarding problem the same literature names as the cause of superficial cooperation.
Where the two literatures stand
Work on industry-education integration describes structural couplings between schools and firms. Examples include alternating students between campus and enterprise R&D in Zhao et al.'s pharmacy MEng model, joint training bases with motor testbenches and battery-fault diagnosis modules for new-energy-vehicle majors in Xu et al., and virtual simulation labs co-run with firms in Ding et al.. These papers motivate the proposal: they identify the gap, namely that enterprises guard production data and that campus simulation drifts into what Xia calls a "theorization trap." They are almost entirely descriptive, with no controlled comparisons.
The only direct intersection is Dong et al.. It proposes federated learning for privacy-preserving personalization of student football training but implements no algorithm, model, or evaluation. It also involves no enterprise. We differ on every one of those points.
Cross-silo federated learning, meaning federation among a few organizations rather than millions of phones, is mature in industry. FedVision trains object detectors across factories and banks. MELLODDY trains a shared trunk with private task heads across ten competing pharmaceutical firms. Both assume every client's data is given and fixed, and neither includes an educational node.
For our application, the closest technical work is personalized federated fault diagnosis of rotating machinery, which clusters similar machines and combines shared layers with per-client heads. It shows that scarce-fault machines benefit most from federation, with F1 rising from 0.543 for local models to 0.746. We reuse its personalization design and the trunk-plus-private-head split of FedHealth. What we add is a client that can generate data on demand, and a channel that tells it what to generate.
Two papers constrain the design. DAC-FCF shows that generative augmentation can rescue very-low-data bearing diagnosis. The CH-BatteryGen benchmark reports that traditional methods lose about 20% of performance moving from charging to discharging data. Together they mean we must show that physical seeded faults beat synthetic ones under a matched budget, and that the lab-to-field shift is real enough for steering to matter.
Core mechanism
All sites share a model with a common encoder and a private classification head per site. Factory heads stay local, as in MELLODDY's private task heads. The lab head is public. Operating conditions are described by a coarse public schema agreed in advance. For batteries, the schema covers state-of-charge range, temperature band, and charge or discharge regime. For bearings, it covers load and speed band. Each condition bin is $b$.
After each federated round, every factory $k$ runs the current model on its local data. It computes $u_k(b,c)$, the summed predictive uncertainty that a sample in bin $b$ belongs to fault class $c$. In practice this is the summed predicted probability mass on $c$ among samples whose prediction entropy exceeds a threshold. These per-factory tables are combined by secure aggregation, so the coordinator sees only the sum. Calibrated Laplace noise is then added for differential privacy:
$$ s(b,c) = \sum_{k} u_k(b,c) + \mathrm{Lap}(\Delta/\varepsilon), $$
where $\Delta$ is the per-factory sensitivity, fixed by clipping each $u_k$, and $\varepsilon$ is the privacy budget per request round.
The lab ranks condition-and-fault cells by $s(b,c)$, discounted by how many lab samples already exist in that cell. The top cells become that week's student experiment tickets. A ticket might read: produce high-resistance faults in LFP cells under low-temperature discharge in the 20–40% state-of-charge range. The resulting labeled recordings join the lab client's data for the next rounds.
Because the lab's data is not masked, each student cohort's contribution can be measured directly by removing their batches and retraining. This gives instructors a per-student attribution that standard secure aggregation would hide. The tension between cryptographic masking and individual assessment does not have to be resolved in general: the masked parties are the factories, and the party being graded is the lab.
Research Program
The work proceeds in three stages.
The first stage is an offline simulation that makes the loop testable without hardware. On CH-BatteryGen, the lab is the pool of charging-regime segments. Running an experiment means revealing a fixed-size labeled chunk from one chosen cell of chemistry, fault type, severity, and state-of-charge band. The factories are fleet operators: discharging segments partitioned by vehicle into roughly ten clients, with faults subsampled to 0–5% to mimic the scarcity regime of the rotating-machinery study. A second instance uses CWRU bearing data as the lab pool, organized by load, fault location, and fault size. If accessible, factories are the real multi-machine pump data of the rotating-machinery study; otherwise they are the self-collected two-stage drive bench from DAC-FCF, partitioned by condition.
The second stage is a privacy analysis of the request channel. We measure how much an honest-but-curious coordinator can infer about a single factory's fault mix from the sequence of noisy tables, sweeping $\varepsilon$ and the bin granularity.
The third stage is a physical pilot at one new-energy-vehicle training base of the kind Xu et al. describe. Two or three partner fleet or maintenance operators act as factory clients, and students run ticketed battery-fault experiments over one semester. The pilot also carries a small pre-registered educational comparison. One section works from federated tickets and another from instructor-chosen tasks of equal hours. Both are scored on a held-out set of real field fault cases the students diagnose by hand. This is secondary and underpowered by design. Its value is that it would be one of the few controlled measurements in a literature that has none.
Baselines
The naive reference is local-only training: each factory trains on its own scarce data, and the lab trains on its own. The rotating-machinery study puts this at F1 0.543 on scarce-fault machines and 0.818 overall. It marks the floor any collaboration must clear.
The second baseline is the strongest non-federated competitor: pretrain on all lab data the budget allows, release the model, and let each factory fine-tune its head and upper layers locally. It uses the same total lab-experiment budget, the same backbone, and the same per-site tuning budget as our method.
The third baseline is the strongest existing federated method: the personalized hierarchical scheme of the rotating-machinery study, with the lab included as an ordinary static client. Its lab experiment plan is fixed in advance and uniform over cells, and its total experiment count equals ours. That paper reports average F1 0.855 overall and 0.746 on scarce-fault machines, against 0.831 for vanilla federated averaging. These numbers come from its pump data and are reference points, not targets on our benchmarks. We also report vanilla federated averaging, because FedHealth found a 5.3-point accuracy gap between unpersonalized and personalized models.
On CH-BatteryGen, the centralized reference is the reported BatteryMultiModalCNN result: F1 0.9280 for LFP discharging fault classification and 0.8813 for severity grading. We use that architecture as the shared backbone. We also train a centralized oracle on all pooled data, in simulation only, as the upper bound.
The key ablation isolates steering. We compare our method against the identical system whose lab picks cells uniformly at random, and against one where the lab picks cells using only its own uncertainty, with no factory signal. The second variant tests whether the factory channel carries information the lab could not obtain alone.
Fairness rules apply throughout. Every method gets the same number of lab experiments, the same number of federated rounds, and the same hyperparameter search, five configurations per method on a validation split held by each factory. Results are averaged over five seeds and factory partitions.
Our method wins if, on scarce-fault factory clients, it reaches the static-lab federated method's final F1 with at most half the lab experiments. At the full budget it must exceed both pretrain-then-fine-tune and the static-lab federated method by more than the seed standard deviation. Matching only the random-selection ablation would mean steering adds nothing, and we would report that.
Why not just…?
Why not just pretrain on lab data and let each factory fine-tune privately? This is the most serious threat. It needs no network, and the lab's data is not secret, so it can be released freely. It fails only if the lab-to-field shift is condition-specific and the lab cannot guess which conditions matter. CH-BatteryGen's roughly 20% charging-to-discharging drop for traditional methods suggests the shift is large, but it does not say where it is concentrated. The head-to-head comparison at equal experiment budget settles this directly. There is also a hybrid: pretrain-and-fine-tune plus the request channel alone, without federated weight sharing. If that hybrid matches the full system, we will report that the value lies in the request channel and that weight federation is unnecessary.
Why not just generate the missing faults synthetically instead of breaking hardware? DAC-FCF reports up to 32% gains in very-low-data bearing settings, and CH-BatteryGen itself builds physics-constrained generated data. Synthetic generation is cheap, so we run a variant where steered tickets are filled by a generator instead of physical runs, at matched sample counts. If generated faults close the gap, the federated request channel still stands, but the physical-lab and student component does not. That would weaken the education half of the claim, and we would say so.
Why not just centralize anonymized factory data at the university? Anonymizing sensor streams does not remove what makes them proprietary, namely the operating cycles themselves. MELLODDY's participants could not pool assay data under any scrubbing arrangement. FedVision also reports a bandwidth cost of 50 MB/s for centralized streaming versus under 1 MB/s federated. We include centralization only as the simulation oracle.
Why not just ask factory engineers which faults worry them? An interview-driven ticket list is the qualitative version of our signal, and it is what dual-instructor programs already do. We add it as a baseline in the pilot. Engineers' fault lists are mapped to cells and compared, as a lab plan, against the uncertainty-steered plan at equal budget. Engineers know which faults are costly. They do not know where the model is uncertain, and the comparison measures whether that difference matters.
Decisive Checks
The proposal is wrong if, in simulation, uncertainty-steered selection does not beat random selection at equal experiment budgets on scarce-fault clients. It is also wrong if pretrain-then-fine-tune matches the full federated system.
The privacy claim is wrong if, at the chosen $\varepsilon$, an attacker observing the request tables can classify an individual factory's dominant fault type well above chance. We test this with a shadow-model attack.
The pilot is informative, though not conclusive, if steered experiments transfer from simulation to a real bench. We test this by checking whether the federated model's scarce-fault F1 at partner sites improves over the semester faster than in a frozen-lab control period.
The educational comparison is reported with effect sizes and confidence intervals whatever the result, since its main purpose is to supply the controlled baseline the industry-education literature lacks.
References
- Innovation in engineering master's training model: A case study of Guangdong Pharmaceutical University in the field of biotechnology and pharmacy
- Research on the Teaching Reform Strategies of New Energy Vehicle Majors from the Perspective of Industry-Education Integration
- Research and Practice of Virtual Teaching and Research Office for Electronic Information Specialties
- The Transformation of Talent Training in Cultural Industry in Colleges and Universities under the Background of Epidemic Normalization
- Reform of Football Training Professional Courses Based on Federated Learning under the "Internet+" Model
- FedVision: An Online Visual Object Detection Platform Powered by Federated Learning
- Industry-Scale Orchestrated Federated Learning for Drug Discovery
- Personalized Federated Learning for Multi-task Fault Diagnosis of Rotating Machinery
- FedHealth: A Federated Transfer Learning Framework for Wearable Healthcare
- An Advanced Convolutional Neural Network for Bearing Fault Diagnosis under Limited Data
- Battery Fault: A Comprehensive Dataset and Benchmark for Battery Fault Diagnosis