Welcome to an update on CTMTDS activities.
Highlights #
- Research Fellow for Theme 1 Michael Stewart PhD was accepted, congratulations to Dr Michael Stewart.
- Chief investigator for Theme 1 Dr Wei Lui is now Associate Professor Wei Lui – congratulations Wei.·
- PhD Candidates Yingying Yang, Tyler Bikaun, and Ryan Leadbetter completed their first candidacy milestones. We now have five of our PhD Candidates working towards Milestone 2.·
- Tim Pesch joined the Centre in October. Tim will complete his PhD at UWA within Theme 2 supervised by Adriano Polpo and Ed Cripps. Welcome to Tim.·
- Alex Rohl, Ryan Malone and Sandy Spiers submitted their honours theses and presentations. All three of our honours students presented their research to our industry partners in October.·
- Débora Corrêa and Melinda Hodkiewicz were invited to present on Prediction of Hydraulic Failure on Excavators 2020 at the Intelligent Maintenance Conference. ·
- Melinda Hodkiewicz, Tim French and Tom Smoker presented “Digitalization and Reasoning Over Engineering Textual Data Stored in Spreadsheet Tables” with Johan Kluwer at the 4th IFAC Workshop on Advanced Maintenance Engineering, Service and Technology.
- Ryan Loxton was invited to present at the WA branch of the Statistical Society of Australia in November; presenting “Optimisation in Practice”.
- CTMTDS was well represented at the last European Safety and Reliability Conference in Venice.
- Adriano Polpo and Melinda Hodkiewicz presented a conference paper with Paul Castle and Janet Ham titled “Interpretable Survival Models for Predictive Maintenance”.
- Melinda Hodkiewicz, Tim French Wei Liu and Caitlin Woods presented a conference paper with Yiyang Gao titled “Pipeline for machine reading of unstructured maintenance work order records”•
- Michael Stewart, Wei Liu, Rachel Cardell-Oliver and Mark Griffin presented a conference paper titled “Cleaning and visualization of unstructured text in safety records”•
- Melinda Hodkiewicz was invited to deliver a Plenary Lecture – “Maintenance in an industry 4.0 and COVID-19 world – the case for remote operations”.
CTMTDS Activities #
Maintenance Work Order Annotation Gold Standard Test set #
Utilising Michael Stewart s Redcoat Software Tool, researchers in Theme 1 are meeting weekly to annotate a set of 10,000 maintenance work orders. By working on the annotation task simultaneously, they are ensuring consistency in the annotation resulting in a high-quality dataset. This work is supporting the Knowledge Graph for Maintenance project, Ziyu Zhao work on integrating domain knowledge for neural technical language processing, Tyler Bikaun work on deep active learning for technical language, Laura Peh work on grammar induction for technical language, and Frederick Subere-Albawy work on evaluating language models on technical language.
Redcoat is a pivotal software application in Theme 1; in this past quarter, this team have consistently annotated small subsets of 400 work orders each week, with 5 annotators per work order. The plan is to use this dataset to train a model to construct knowledge graphs from work orders from BHP Nickel West.
Knowledge Graph Construction from Maintenance Work Orders #
Continuing his work from last quarter on Knowledge Graph Construction from Maintenance Work Orders; Michael improved the software to enhance significantly its ability to visualise and query work orders. The two main features that have been implemented are:
-
Entity hierarchies: Several entity types (Item and Activity) have been organised into hierarchies. The items are automatically placed into a hierarchy constructed from ISO-15926, while the activities are organised into key activities such as Repair, Replace and Inspect. This enables the aggregation of nodes in the graph, opening up many potential queries.
-
Aggregation: The graph can now be aggregated based on the entity hierarchies. For example, it is now possible to collapse every type of “pump” into a single node so that one may easily view the number of outgoing relationships from all pumps to every other type of entity. Using this, industry can access important information in a few clicks, such as “how many pumps have been replaced in the last month?”, or “what failure modes have been observed on pumps?”.
Managing streamed sensor data for mobile equipment prognostics #
Many standard Machine Learning Models fail to produce results for data sets with issues that include inconsistent and asynchronous recordings, missing data, and uncertainties in the ground truth for the dependent variable. Faced with these challenges, Débora Corrêa has been working on a novel solution; to create a data frame of the sensor data, fleet management and maintenance work order data and demonstrate the application of a log-linear mean regression model based on the residual lifetime of the asset. The model is built considering the values of the available covariates (sensors and alarms) which indicates the current state condition of the asset. Debora is planning to test this with one of our industry partners in the next few months.
IDEA Tool #
The objective of this project is to develop a tool to evaluate the usefulness of data and test its reliability and efficiency. Ayham is proposing a tool that does a preliminarily analysis of data in order to verify and test its efficiency and reliability to solve the encountered problem. The tool will identify data that is unfit for purpose, and indicate what additional information is needed. That is, the aim of the tool is not to build a model, it is instead to test if there is enough data to build a model.
The tool will consist of three levels of assessment:
- Metadata assessment
- Linear statistical and probabilistic assessment
- Nonlinear assessment
Development of Optimisation Models for Scheduling #
In collaboration with one of our industry partners, Hoa Bui has developed a model designed to solve dimensionality challenges.
This model is now being tested with industry data and showing itself to be quicker and more efficient than the existing methods in the literature for problems of this type. This not only results in the positive outcomes for the project but also brings exciting academic results and publications.
Workshops organised #
GitLab workshop #
On the 14 August, Andrew Rohl led a half-day workshop on Git and Gitlab to demonstrate version control and the use of the CTMTDS code repository. The workshop is based on Software Carpentry lessons, tailored to the CTMTDS environment and also addresses the challenges of version control with Jupyter notebooks.
Seq2KG: An End-to-End Neural Model for Domain-Agnostic Knowledge Graph Construction from Text #
Michael Stewart presented on a paper entitled “Seq2KG: An End-to-End Neural Model for Domain Agnostic Knowledge Graph (not Text Graph) Construction from Text” at the 17th International Conference of Principles of Knowledge Representation and Reasoning (KR2020).
Researcher Catch-up #
Our Researcher Presentations have continued fortnightly virtually. These meetings are arranged by our research fellows and consist of two 30-minute presentations given by researchers in the Centre.
They are attended by the Centre Chief Investigators, Research Fellows, HDR candidates and industry partners associated with the Centre
Random effects with changepoints – Dr Edward Cripps #
Ed discussed random effects models with change points.
Two different data sets were discussed
- Longitudinal cognitive science data assessed against two learning group theories.
- Incremental Theorists - those that believe that learning is incremental, thus negative feedback is taken as constructive criticism that can be part of the learning
- Entity Theorists – those that believe that ability is ingrained, thus if performance is poor, it is taken as a reflection of their ability and performance spiral downwards on negative feedback. The belief is that the groups will learn differently, so statistically, Ed wanted to examine the data of the two groups to determine if there is a difference and where that change point is as to learning behaviour.
- Rotational bearing vibrations
Ed described data taken from bearings, looking at the remaining useful life of the bearings' over time and ways in which we can model this using longitudinal data.
Using these data sets as examples, Ed demonstrated how a Random Effects model can be used to predict a change point with populations of individual items that have potentially different behaviours. A random-effects model is useful for maintenance data that can be generated by a common mechanism; however, this data may be producing different behaviours with each item.
Unlocking knowledge from technical text using deep active learning and entity typing – Tyler Bikaun #
Much valuable information is captured in the unstructured fields of maintenance work orders. Currently, maintenance and reliability engineers have to spend considerable time locating, reading and interpreting these fields to extract information; this is then combined with structured fields to understand the failures and work done on their equipment. Collating a picture, over time, of each piece of equipment (as well as systems of equipment) is done manually, if at all. In a plant with thousands of pieces of equipment, it is challenging to find patterns and difficult to spot emergent behaviour. Due to the speed and volume of information captured by information technology systems, it is becoming ever more challenging to identify and confirm out of the expected behaviour (a.k.a abnormal), which is necessary to make strategic decisions. Insight into the behaviour of assets and maintenance activities can be extracted by data-mining maintenance records. Current systems only focus on classifying the text as a failure event, ignoring the rest of the information. Advanced information extraction techniques such as entity typing have shown promising results in extracting fine-grained semantic content from text, but these techniques rely on deep neural networks, which require large labelled datasets to be successful. In domains such as industrial maintenance, these datasets do not exist and are costly to acquire. Typical approaches to tackle this data scarcity problem include distance supervision, crowdsourcing, semi-supervised learning and active learning. For technical texts in industrial settings, these are challenging to use due to unavailability of external knowledge bases, dependency on tacit knowledge, and lack of similar cross-domain datasets. Active learning which aims to overcome data scarcity by reducing the dependence on large datasets with the aid of human-supervisors shows the most promise under such constraints. This project aims to develop a framework that marries active learning with entity typing. Tyler will investigate theoretical and practical considerations in using technical language processing in real-world settings. Successfully applying advanced information extraction techniques in industrial maintenance will empower maintainers to better understand their assets and enhance maintenance strategy development.
Supporting Organisational Transformation - Supporting Theme 4 case study research proposal – Dr Keyao Li #
Effective and efficient maintenance management ensures that assets are productive and meet business needs. However, maintenance practices in the mining industry to date are slow to evolve, and asset downtime remains stubbornly high. In light of the innovations in edge computing, artificial intelligence and Industry 4.0, the Centre for Transforming Maintenance through Data Science (CTMTDS) serves to transform maintenance by conceiving a new digitally-driven maintenance management system.
Although technological innovations have clear benefits, their full value depends on successful adoption and implementation. Many studies show technological innovations fail if they are not supported by end-users and other stakeholders. Some factors that influence users’ attitudes and intentions toward new technology include its perceived usefulness, perceived ease-of-use, perceived value, perceived risk on system use, social influence, effects of trust and more (Davis et al. 1989; Venkatesh and Davis, 2000; Venkatesh and Bala, 2008; Wang and Wang, 2010; Khazaei, 2020). The investigation of critical factors for technological innovation implementation in the unique maintenance management work environment is of great value.
Two data science tools have been recently implemented in CTMTDS industry partner organisations: one is a risk-based maintenance planning tool, and the other one is a pipeline wall thickness risk assessment tool. To improve maintenance planning across the industry, it is essential to understand and disseminate the multiple factors that contribute to innovation implementation such as these.
This proposed research project aims to identify the factors that enhance the implementation of data science innovations in mining maintenance management system. The recent implementation of data science tools provides the setting for two case studies. Each case study will explore the factors that enabled or hindered the implementation of the data science project. Results of the case studies will inform other projects in the CTMTDS and promote broader use of digitally-driven maintenance management system by the organisations. The objectives of the case studies include:
- Gain insights from participants who contributed to the innovation
- Identify the factors that supported or hindered data science innovation.
- Produce case studies to share key learnings about the innovation process
Prediction and Simulation: Same old ideas, fancy new buzzwords – Prof Michael Small #
Michael Small presented on the core techniques of Theme 2 and how existing methodologies will be used in research focus for this Theme. Using the data set that was presented by Dr Ed Cripps in the previous catch up, Michael focused his presentation on deterministic applied mathematics dynamical systems and machine learning methods. Michael discussed how these systems are used successfully in prediction and the importance of these methodologies for research carried out in the Centre.
Visualising and querying maintenance work orders through a maintenance knowledge graph – Dr Michael Stewart #
The written descriptions of maintenance work orders capture vital information, such as indicators of failure modes and end of life events, which are directly tied to MTTF values. Despite their prevalence and the significance of the information contained within, work order descriptions are seldom used for routine analysis. In this presentation Michael provided an interactive demonstration of Theme 1's work on constructing the very first knowledge graph for maintenance, while also showcasing his approach to developing an annotated training dataset of work orders with which to train state-of-the-art deep learning models. The knowledge graph interface built as part of this research allows for the visualisation and querying of work orders, presenting vital information about work done, as well as failure modes for each asset, in a highly accessible way.
Detecting asset cascading failures from work orders – Dr Ayham Zaitouny #
Maintenance logs record unscheduled work done on assets and maintenance work orders include much information describing the assets' health, maintenance regime and performance. This project aims to identify co-occurrence of work, and hence co-occurrence of failure amongst an organisations' assets. Asset cascading failures are a common phenomenon in industry that can occur for one asset or among different assets. Diagnosing these cascading incidents is crucial to understand the system and improve the management of maintenance. In this study, a mathematical approach based on complex network analysis has been conducted to identify cascading failures by using work orders. The proposed approach has been implemented on pump data, and several cascading incidents have been detected. Operators have confirmed a number of these incidents as known issues, demonstrating the potential success of the approach. Other structures within the network analysis identified previously unrecognised issues. In this presentation, we show a summary of the method's results and discuss the potential extension of this research.
Optimal Maintenance Scheduling via Mathematical Programming – Yingying Yang #
In any asset-intensive industry, maintenance activities are inevitable and costly, and hence it is essential to investigate better ways of scheduling maintenance activities. Conducting maintenance may require certain parts of the plant or the whole plant to be shut down, which usually leads to large financial loss.
Therefore, optimal scheduling of maintenance activities is highly desirable as it will potentially reduce downtime and increase profitability. However, since the plant is a complex interconnected network of subsystems, it is exceptionally challenging to optimise maintenance schedules and provide synergistic opportunities for sharing utilities, manpower, and minimising travel between locations. Moreover, the timing of maintenance shutdowns may be affected by external factors, such as production planning, inventory management and even market demands, which make maintenance scheduling problems even more challenging to tackle. In this research, mathematical optimisation models will be developed to minimise the total cost and maximise synergies and profits in maintenance operations.
Since the models are difficult to solve for large-scale data sets, Yingying will develop new mathematical methods and novel algorithms to reduce dimensionality and solve the models more efficiently. Furthermore, the proposed models and algorithms will be implemented in optimisation software using real data provided by mining companies. The expected outcomes of her research will enable planners in the mining industry to build optimal maintenance schedules and shutdown plans, which will bring large financial benefits to the industry.
Managing Streamed Sensor Data for Mobile Equipment Failure Prediction – Dr Débora Correa #
The ability to wirelessly stream data from sensors on mobile equipment provides opportunities to assess asset condition proactively. Melinda and Débora presented streaming data that was drawn from a mining industry case study. This data is for one single excavator over nine months and has 57 numerical sensor variables and 41 binary indicators describing the conditions and status of different subsystems. In addition, data is available from the fleet management and maintenance work order systems.
They focused on the hydraulic subsystem, which has potential failure events reported in the period of the data. There are significant issues with the data due to the large volume, inconsistent and asynchronous recording from different sensors, 57% of rows have missing data, and uncertainties in the ground truth for the dependent variable (hydraulic subsystem failure).
Melinda and Débora will created a data frame of the sensor data, fleet management and maintenance work order data and demonstrated the application of log-linear mean regression model based on the residual lifetime of the asset.
The model is built considering the values of the available covariates (sensors and alarms) which indicates the current state condition of the asset.
Decision Support for Prognostics of Complex Systems: A Practical Approach Using Bayesian Networks – Ryan Leadbetter #
Research on prognostic health management tools for use in condition-based maintenance policies offers to increase the availability of assets while also decreasing maintenance costs and unplanned downtime events. However, these benefits have not been seen by the mining industry because these tools are not compatible with the standard industry practice of condition-based maintenance. For such benefits to be realised in the mining industry, new tools are needed that can combine heterogeneous data and expert knowledge for prognostics of complex systems. Bayesian networks are a possible approach which can fill this gap, but they have not been used to model complex mining assets for prognostic health management.
Ryan's research will develop decision support systems based on Bayesian networks which will be constructed using datasets and expert elicited knowledge from three mining companies. From these companies three complex systems have been identified: overland conveyors, haul trucks, and heavy haulage locomotives. This research will use an object-oriented Bayesian network approach to diagnose and forecast the health condition of the asset from condition monitoring and system operation data. The structure of the network as well as the probability distributions for the nodes will be elicited from domain expert knowledge. An efficient validation methodology will be devised to build user confidence in the decision support system's predictions. The decision support system will then be partially deployed in industry to assess performance and usability. The deployment will allow for a more accurate prediction of future asset health condition which will increase the quality of maintenance planners' decision making. Developing a general methodology for the construction of decision support systems for prognostic health management will provide a clear approach for companies to follow to move towards an effective company-wide condition-based maintenance policy.
Prediction of Pump Maintenance using Markov Models on Free Text Work Orders – Alex Rohl #
Alex presented on the importance for deriving a reliable model that can predict the future maintenance of assets in order to establish an optimal replacement policy
Alex discussed methods of deriving a model from unstructured short-text work orders that relate to maintenance on a specific piece of equipment within a large collection of physical assets. He did this by:
- Applying natural language processing techniques, translating each short-text into a generalised "concept", which de-scribes a group of similar work orders.
- Using the transitions from each concept to the next for each asset, in order to create a discrete-state continuous-time Markov model.
- Verifying this process numerically by analysing a dataset, provided by industry
Using Ordinal Networks and Autoencoder Based Neural Networks for High-Frequency Bearing Failure Detection – Ryan Malone #
Failure detection serves as a powerful tool for increasing the operational availability for industry assets. Unplanned failures can result in long periods of downtime and a decrease in overall production. Ryan presented a new methodology for failure detection, specifically applying it to time series bearing data. The dynamics of the system are characterised by order patterns within the time series and then converted to a network representation. These properties undergo a process of smoothing and standardising, and are applied to an autoencoder based neural network. One class-classification is used to characterise a normal or degrading operating state.
Ryan found the proposed method was successfully capable of providing failure warnings and outperformed the defined benchmark. This is a significant result as it shows the proposed method is successfully capable of capturing failure related system dynamics. This is unique as it detects system degradation via shape-based methods whereas the literature primarily uses amplitude-based methods. Further, being designed flexibly, the proposed method is easily applied to other pieces of failure or anomaly detection.
Maintenance Scheduling for Wash Tanks – Sandy Spiers #
Sandy presented his honours research, which is to support the development of the optimal maintenance schedule for a group of assets for the industry partner. The cost to the industry partner when one of their assets is out of operation is manageable.
However, when two assets are out of operation, the cost per day is far more than the sum of the operation costs and can be up to ten times more. Thus, the scheduling challenge is to ensure that the likelihood of more than one asset being out of operation for any reason is very low. These assets are highly dependent on each other. Sandy research reviewed a range of possible solution methods. At moderate to large scales, the industry scheduling problem becomes extremely hard to do even using the best of the currently available scheduling tools. He presented a model he believes provides a solution, close to the real world optimal. Sandy demonstrated the ability of his model to respond efficiently and accurately when exposed to future changes and the inclusion of new constraints as they arise.
Maintenance in an industry 4.0 and COVID-19 world – the case for remote operations – Melinda Hodkiewicz #
Maintenance is being refactored due to advances associated with Industry 4.0 and the need to work remotely brought on by the COVID crisis. The Western Australian (WA) economy has had positive growth in the last 12 months; this goes against global trends. A key contributor to this has been investments made by the iron ore producers in automation and remote operations. Melinda presentation describes how the adoption of new technology and processes by the miners has allowed them to survive and thrive in the current climate. It explores how the Industry 4.0 platforms now in place here in WA are also a springboard to the next generation of maintenance management for remote operations.
Integrating Domain Knowledge for Neural Technical Language Processing – Ziyu Zhao #
Structured data is widely available in the maintenance domain in many forms including asset hierarchies and FMEA (Failure Mode and Effects Analysis) tables. There is a significant volume of valuable knowledge in this data that can assist machine learning systems in making accurate predictions on unstructured text, such as identifying entities (such as items, activities and observations). Ziyu’ s research focuses on incorporating structured domain knowledge into neural networks via probabilistic logic programming. In her presentation Ziyu provided an overview of her case study on multi-label hierarchical entity typing, demonstrating her preliminary results. Ziyu also provided an outlook to future directions of her research that will be of significant benefit to the maintenance domain.
Webinars #
Complex network approaches for modelling systems dynamics.
Débora Correa organised a virtual Minisymposium (Complex network approaches for modelling systems dynamics as part of Dynamics Days Digital 2020), and several seminars ( https://sites.google.com/view/uwa-complex/home ). They are relevant to the complex systems approach that we are applying in the centre projects.
IN THE SPOTLIGHT #
Hoa Bui was interviewed, as the 2020 Winner of the AustMS WIMSIG Maryam Mirzakhani Award. This interview was published in the September issue of the Australian Mathematics Society Gazette, please see this link to view Hoa interview.
Melinda Hodkiewicz received a Certificate of Appreciation for the delivery of her plenary lecture at the European Safety and Reliability Conference on ‘Maintenance in an industry 4.0 and COVID-19 world –the case for remote operations’. She was also acknowledged for organising the special session on Natural Language Processing, Knowledge Graphs and Ontologies, which included presenters from the CTMTDS, China, UWA, Denmark, Norway, and Italy. It is the first time that topics such as NLP and text analysis for safety and maintenance have been introduced in such a structured way in to this reliability community. It has generated a lot of interest for research of Theme 1. Congratulations Melinda
Congratulations to Prof Andrew Rohl on receiving the Paul G. Dunn Research Development Award. This award is in recognition of a Head of Area who fosters, supports and helps build significant capability and capacity at Curtin University over an extended period.
Stay tuned for our next issue where we will cover Master Class Series.
The Centre plans to offer a series of masters classes focussing on the specialist areas utilised for research within the Centre. The next issue will include information about the proposed sessions.