Annotations in condition monitoring systems contain information regarding asset history and fault characteristics in the form of unstructured text that could, if unlocked, be used for intelligent fault diagnosis. However, processing these annotations with pre-trained natural language models such as BERT is problematic due to out-of-vocabulary (OOV) technical terms, resulting in inaccurate language embeddings. Here we investigate the effect of OOV technical terms on BERT and SentenceBERT embeddings by substituting technical terms with natural language descriptions. The embeddings were computed for each annotation in a pre-processed corpus, with and without substitution. The K-Means clustering score was calculated on sentence embeddings, and a Long Short-Term Memory (LSTM) network was trained on word embeddings with the objective to recreate the output from a keyword-based annotation classifier. The K-Means score for SentenceBERT annotation embeddings improved by 40% at seven clusters by technical language substitution, and the labelling capacityof the BERT-LSTM model was improved from 88.3 to 94.2%. These results indicate that the substitution of OOV technical terms can improve the representation accuracy of the embeddings of the pre-trained BERT and SentenceBERT models, and that pre-trained language models can be used to process technical language.

In the process industry, condition monitoring systems with automated fault diagnosis methods assist human experts and thereby improve maintenance efficiency, process sustainabil-ity, and workplace safety. Improving the automated fault diagnosis methods using data and machine learning-based models is a central aspect of intelligent fault diagnosis (IFD). A major challenge in IFD is to develop realistic datasets with accurate labels needed to train and validate models, and to transfer models trained with labeled lab data to heterogeneous process industry environments. However, fault descriptions and work-orders written by domain experts are increasingly digi-tised in modern condition monitoring systems, for example in the context of rotating equipment monitoring. Thus, domain-specific knowledge about fault characteristics and severities exists as technical language annotations in industrial datasets. Furthermore, recent advances in natural language processing enable weakly supervised model optimisation using natural language annotations, most notably in the form of natural language supervision (NLS). This creates a timely opportu-nity to develop technical language supervision (TLS) solu-tions for IFD systems grounded in industrial data, for example as a complement to pre-training with lab data to address problems like overfitting and inaccurate out-of-sample gen-eralisation. We surveyed the literature and identify a con-siderable improvement in the maturity of NLS over the last two years, facilitating applications beyond natural language; a rapid development of weak supervision methods; and transfer learning as a current trend in IFD which can benefit from these developments. Finally we describe a general framework for TLS and implement a TLS case study based on Sentence-BERT and contrastive learning based zero-shot inference on annotated industry data.

We propose a novel approach to facilitate supervised fault diagnosis on unlabelled but annotated industry datasets using human-centric technical language processing and weak supervision. Fault diagnosis through Condition Monitoring (CM) is vital for high safety and resource efficiency in the green transition and digital transformation of the process industry. Learning-based Intelligent Fault Diagnosis (IFD) methods are required to automate maintenance decisions and improve decision support for analysts. A major challenge is the lack of labelled industry datasets, limiting supervised IFD research to lab datasets. However, features learned from lab environments generalise poorly to field environments due to different signal distributions, artificial induction or acceleration of lab faults, and lab set-up properties such as average frequency profiles affecting learned features. In this study, we investigate how the unstructured free text fault annotations and maintenance work orders that are present in many industrial CM systems can be used for IFD through technical language processing, based on recent advances in natural language supervision. We introduce two distinct pipelines, one based on contrastive pre-training on large datasets, and one based on a small-data human-centric approach with unsupervised clustering methods. Finally, we showcase one example of the small-data fault classification implementation on a CM industry dataset with a SentenceBERT language model, kMeans clustering, and conventional signal processing methods. Fault class imbalance and time-shift uncertainty is overcome with weak supervision through aggregates of features, and human-centric clustering is used to integrate technical knowledge with the annotation-based fault classes. We show that our model can separate cable and sensor fault recordings from bearing-related fault recordings with an F1-score of 93. To our knowledge, this is the first system to classify faults in field industry CM data based only on associated unstructured fault annotations.

Labelled industry datasets are one of the most valuable assets in prognostics and health management (PHM) research. However, creating labelled industry datasets is both difficult and expensive, making publicly available industry datasets rare at best, in particular labelled datasets.Recent studies have showcased that industry annotations can be used to train artificial intelligence models directly on industry data ( https://doi.org/10.36001/ijphm.2022.v13i2.3137 , https://doi.org/10.36001/phmconf.2023.v15i1.3507 ), but while many industry datasets also contain text descriptions or logbooks in the form of annotations and maintenance work orders, few, if any, are publicly available.Therefore, we release a dataset consisting with annotated signal data from two large (80mx10mx10m) paper machines, from a Kraftliner production company in northern Sweden. The data consists of 21 090 pairs of signals and annotations from one year of production. The annotations are written in Swedish, by on-site Swedish experts, and the signals consist primarily of accelerometer vibration measurements from the two machines.The dataset is structured as a Pandas dataframe and serialized as a pickle (.pkl) file and a JSON (.json) file. The first column (‘id’) is the ID of the samples; the second column (‘Spectra’) are the fast Fourier transform and envelope-transformed vibration signals; the third column (‘Notes’) are the associated annotations, mapped so that each annotation is associated with all signals from ten days before the annotation date, up to the annotation date; and finally the fourth column (‘Embeddings’) are pre-computed embeddings using Swedish SentenceBERT. Each row corresponds to a vibration measurement sample, though there is no distinction in this data between which sensor or machine part each measurement is from.

Industridataset med labels är bland de mest värdefulla tillgångarna att tillgå inom prognostik- och tillståndsövervaknings-forskning. Att tillverka labellade dataset är både svårt och dyrt, vilket medför att allmänt tillgängliga industridataset är sällsynta, särskilt de med labels. Studier har dock visat att industriannoteringar kan användas för att träna AI-modeller direkt på industridata ( https://doi.org/10.36001/ijphm.2022.v13i2.3137 , https://doi.org/10.36001/phmconf.2023.v15i1.3507 ), men trots att många industridataset innehåller de nödvändiga texterna så är få, om ens några, sådana dataset allmänt tillgängliga.Därför ger vi ut ett dataset innehållandes annoterade signaldata från två stora (80x10x10m) pappersmaskiner från ett pappersbruk i norra Sverige. Datan består av 21 090 par av signaler och annoteringar från ett års produktion. Annoteringarna är skrivna på svenska av experter på plats, och signalerna består huvudsakligen av accelerometervibrationsmätningar från de två maskinerna.Datasetet består av ett års annoterade vibrationsensormätningar från två pappersmaskiner, strukturerade som en Pandas dataframe och serialiserade som en pickle-fil (.pkl) samt en JSON-fil (.json). Den första kolumnen (’id’) är ID per sample; den andra kolumnen (’Spectra’) är fast-Fourier-transformerade och envelope-transformerade vibrationssignaler; den tredje kolumnen (’Notes’) är de tillhörande annoteringarna, kartlagda så att varje annotering är kopplad till alla signaler från tio dagar före annoteringsdatumet upp till annoteringsdatumet; och slutligen den fjärde kolumnen (’Embeddings’) är förberäknade text-representationer från Swedish SentenceBERT. Varje rad motsvarar ett vibrationsmätningsprov, även om det inte finns någon åtskillnad i denna data mellan vilken sensor och maskindel varje mätning kommer från.