Core Concepts
This article proposes a federated prognostic model that allows multiple users to jointly construct a failure time prediction model using their multi-stream, high-dimensional, and incomplete sensor data while keeping each user's data local and confidential.
Abstract
The key highlights and insights of the content are:
Most prognostic methods require a decent amount of historical data for model training, but in reality, the data owned by a single organization may be small or insufficient. To address this challenge, the article proposes a federated prognostic model that enables multiple users to collaboratively train a failure time prediction model using their multi-stream, high-dimensional, and incomplete sensor data while keeping each user's data local and confidential.
The proposed federated prognostic model comprises two steps: data fusion and prognostic model construction. Data fusion focuses on fusing multi-stream high-dimensional sensor signals using multivariate functional principal component analysis (MFPCA) to provide low-dimensional features. Prognostic model construction then maps the time-to-failure (TTF) to the fused features using a (log)-location-scale regression model.
To enable federated learning, the article proposes a new federated algorithm for feature extraction. It first develops a federated dominant subspace identification algorithm to detect the dominant subspace of the sensor signals. Then, it proposes a federated algorithm to compute the MFPC-scores using the detected dominant subspace and the incomplete sensor signals from multiple users.
Numerical studies indicate that the performance of the proposed federated prognostic model is the same as that of classic non-federated prognostic models and is better than models constructed by each user individually.
Stats
The article uses the following key metrics and figures to support the author's logics:
"The TTFs are computed as the first time that si(t) reaches or crosses a predefined threshold D: si(t) = −ci/ ln (˜yi) = D. This yields ln (˜yi) = −ci/D, where ˜yi is the TTF of system i."
"To mimic data acquisition errors, we add noise to the true TTFs. As a result, the observed TTFs are computed from ln (˜yi) = −ci/D + ϵi, where ϵi ∼N(0, 0.0252)."
"With the underlying degradation trajectory, the noisy discrete observations (i.e., the observed degradation signal from a condition monitoring sensor) of system i, i = 1, . . . , 100, are generated as follows: x(τi) = −ci/ ln (τi) + ε(τi), where ε(τi) ∼N(0, 0.2) is the random observation noise."