Using context to adapt to sensor drift

TL;DR

The paper addresses sensor drift in metal-oxide gas sensors used for industrial odor sensing by introducing a context+skill neural network that learns to predict and utilize a drift context. A recurrent context pathway summarizes past sensor responses and modulates a separate odor-recognition skill, enabling automatic adaptation without recalibration. On the Vergara chemical-drift dataset, context+skill models and NN ensembles outperform drift-naive and SVM-based approaches, with context providing the most benefit as drift accumulates over time. This biologically inspired, lifelong adaptation framework demonstrates practical robustness for real-world odor sensing in changing environments.

Abstract

Lifelong development allows animals and machines to adapt to changes in the environment as well as in their own systems, such as wear and tear in sensors and actuators. An important use case of such adaptation is industrial odor-sensing. Metal-oxide-based sensors can be used to detect gaseous compounds in the air; however, the gases interact with the sensors, causing their responses to change over time in a process called sensor drift. Sensor drift is irreversible and requires frequent recalibration with additional data. This paper demonstrates that an adaptive system that represents the drift as context for the skill of odor sensing achieves the same goal automatically. After it is trained on the history of changes, a neural network predicts future contexts, allowing the context+skill sensing system to adapt to sensor drift. Evaluated on an industrial dataset of gas-sensor drift, the approach performed better than standard drift-naive and ensembling methods. In this way, the context+skill system emulates the natural ability of animal olfaction systems to adapt to a changing world, and demonstrates how it can be effective in real-world applications.

Figures (3)

• Figure 1: Recognizing odor despite sensor drift. (A.) The first sensor steady-state response is plotted for all samples of the gas acetone. The data collection periods, called batches, are labeled 1-10 and separated by vertical dashed lines. There is variability within as well as between batches. (B.) For thirty trials, two feedforward networks are trained using a single batch of training data indicated by the vertical dashed lines and colors. The mean classification accuracies evaluated on each batch are shown. The variability of response to a single odor poses a challenge to odor recognition. (C.) Performance degrades as a function of the time between the training batch and the evaluation batch. Mean accuracy along with 95% confidence interval is shown as a function of the absolute distance between the training batch and the evaluation batch. The further away the testing batch is from the training batch, the lower the generalization accuracy becomes.
• Figure 2: Neural network architectures. (A.) The batches used for training and testing illustrate the training procedure. The first $T-1$ batches are used for training, while the next unseen batch $T$ is used for evaluation. When training the context network, subsequences of the training data are selected to be processed recurrently, indicated by the labels $s$ through $p$. In all cases, training data is obtained only from the first $T-1$ batches of data. (B.) A feature vector is input to a collection of SVMs, one trained on each prior batch. Each SVM output is weighted by its corresponding coefficient, $\beta$, and the weighted sum of the output class predictions is taken to be the output, $\hat{\mathbf{y}}$, of the ensemble. (C.) A schematic of the skill model shows feedforward progression of input through two hidden layers $\mathbf{s}$ and $\mathbf{d}$ followed by the output layer $\hat{\mathbf{y}}$. (D.) A schematic of the context+skill model introduces a sequential processing of prior samples as a separate processing pathway. For each context batch from $s$ through $p-1$, one sample per odor class is chosen as a representative. The context information is then utilized by the "decision-making" layer $\mathbf{d}$ and is thus integrated into the feedforward pathway.
• Figure 3: Generalization accuracy. The generalization accuracy of each model was evaluated on batch $T$. For each model type and every batch, 30 models were trained. The line represents the average over the 30 trials, and the error bar is the 95% confidence interval. (A.) The skill and context+skill models are shown with the other models faded out. The context most contributes to performance in the later batches, which offer the longest context sequences. (B.) The SVM ensemble and NN ensemble models are shown with the results from (A) are shown faded out. Both ensemble models are variable in performance between batches.