NAVMAT: an AI-powered pathway to knowledge sharing on material failures

The paper describes the product development methodology, the architecture, the modules and features of the operating prototype, as well as future challenges of NAVMAT, a naval materials failure management system. The fundamental compound of the knowledge management platform is the recording and classification of a failure incident. NAVMAT supports different types of users, from the first Reporter to the Analyst and Forensic Engineer through appropriate workflows. In these workflows the incident-related information can be provided as text, files, images and videos, which can be easily associated to the incident and provide structured information. NAVMAT provides a number of Artificial Intelligence (AI)-enabled helpers. During the incident recording, NAVMAT brings into play reactive real-time search which suggests related incidents and literature to facilitate the editor. It also speeds up the classification of incidents by providing AI-suggested labels, chosen from the (multi-lingual) concepts contained in the system Ontology. On the other hand, the intelligent indexing and search infrastructure of the system supports easy identification and retrieval of past incidents, reports and publications, by applying Natural Language Processing. The prototype of the described system has been embedded as a web application validated by potential Users and is being prepared for Operation in a fleet environment.


Introduction to NAVMAT
Two years ago, the concept and design of a naval materials failure management system, with the acronym NAVMAT, was introduced and announced [1].The current paper presents the steps towards the operating prototype, the development methodology, the system architecture, its modules and features, as well as future challenges.

Identifying the need
Large organisations (such as a fleet), structured on a model of frequent technical staff reallocation (crew mobility), with broadly distributed units (such as vessels at sea and port facilities) operating at extreme environments, are in need of a smart unified failure management system.Such as system should be designed so as to a) allow for direct recording -indexing of all incidents of failure, b) facilitate easy retrieval of previous incidents, c) offer linkage with failure analyses, repair reports and technical documents, d) suggest similar incidents, root causes and maintenance guides.Such a knowledge -based system is founded on the principles of materials science, failure analysis, condition monitoring and forensic engineering and can be powered by modern information technology tools such as Artificial 1.2.Engineering a solution NAVMAT approach utilizes a focused common-cause failure methodology for the naval and marine environment, to begin with.It supports decision making through appropriate Artificial Intelligence and Natural Language Processing methods.For example, reporting is facilitated through automatic classification of a report, based on its content, in alignment with the domain ontology.On the other hand, the identification of similar incidents relies on a clustering methodology, grouping different event reports and related assets based on the incident context and observations (i.e.beyond individual keywords).

Visualizing the Workflow
The authorized system users may be contributors, technical experts or just viewers.The contributor announces an incident by describing it in free text, predefined tags, any available documentation and media files (photos and videos).The experts, should a failure analysis be commissioned, further elaborate the incident by compiling and uploaded a failure analysis, a recommendation and/or a repair report.Viewers share opinion and experience.
During the incident recording, NAVMAT brings into play reactive real-time search which suggests related incidents and literature to facilitate the editor.It also speeds up the classification of incidents by providing AI-suggested labels, chosen from the (multi-lingual) concepts contained in the carefully engineered, extensible NAVMAT Ontology.On the other hand, the intelligent indexing and search infrastructure of NAVMAT supports easy identification and retrieval of past incidents, reports and publications, by applying Natural Language Processing to create a complete puzzle of previous knowledge related to a specific incident setting.

The failure incident
An incident is identified, described and associated with a component (such as a propeller), which is a part of a system (e.g.propulsion), belonging on a platform (such as a ship or a vehicle of even a land facility).The component is made out of a specific material (e.g. a copper alloy), following a suitable fabrication method (e.g.casting).The failed component (or system) operates in defined environment (mechanical loading in corrosive environment), being subjected to a loading regime (steady or dynamic).It exhibits a mode of failure initially identifiable by the observer (e.g.pitting corrosion), but further assessed and analysed only by an expert (stray-current corrosion).Should such a failure analysis take place, the root cause of failure may also be confirmed (e.g.faulty welding procedures on the boat).The full documentation of the incident and related content is classified and indexed for future reference.Each incident may be described by free text, tags (terms from the system ontology), documents and other media files (images and videos).Some incidents, if an analysis is commissioned, are further elaborated by a failure analysis report, a recommendations report or a repair report, bibliography etc.
Data reliability is affected by the knowhow of the contributor and the experience of the analyst.Thus the platform is design to be used by engineers with basic knowledge of naval, mechanical and materials engineering.

NAVMAT system development
The system is designed as an open architecture, scalable model, meaning that distinct modules perform distinct functions, and allows the addition of new modules next to the existing ones (possible new functionalities that will cover potential future needs).The implementation of the NAVMAT platform is founded on the following prerequisites: − NAVMAT is a web-based application.− The multilingual lexical and semantics resources are described in an updatable Ontology − User scenarios are well-defined before the system development − The failure incident (a "conglomerated data point") is at the core of its functionalities − Security considerations (Use of end-to-end encryption, secure login mechanisms, Data isolation: − Cyber-resilience considerations: (Rotating backups; Off-site backups; Backup servers) Based on the above principles and specifications the NAVMAT architecture is deployed around the following components:

The Ontology component
The ontology engineering User Interface (UI) allows the user to input the main information of an ontology to the system.Web protégé [4] of Stanford University is the Ontology editor with which the NAVMAT ontology was developed.NAVMAT c o n c e p t s or c l a s s e s [5] are identified in thes screenshot of Figure 1: Extract from the Ontology Editor.Level 1 presents the Ontology Classes of NAVMAT.Each incident may be characterised with tags stemming from these nine (9) classes.The description of every Class and subclasses is currently bilingual (English and Greek).It can be expanded to any other language.The first row displays the Hierarchy of the Classes, and subclasses at three Levels, using the Class "Failure" as example.The second row displays the ID (IRI) of each class or subclass, the corresponding annotations (term, synonyms) in English and Greek and the relations with other classes.
The collectively comprise the lexical and semantic resources of NAVMAT system.

The Document Analysis component
The document analysis component is meant to realize any intelligent analysis processes, enriching input data (failure incidents.It incorporates a) Incident Indexing (handling of the ID of the resource), b) summarization (composition of a multi-incident summary, containing all the related information from the input incidents), c) Classification (utilizing the ontology lexical and semantic resources), d) similarity -based retrieval (identifying resources similar to a query, e.g.similar types of incidents).

The Repository component
The document repository contains all the incidents (documents and resources) usable by the system: free text, document files (pdf, word, etc.), media files (images and videos).Each resource is associated with all aspects of the system, including e.g.source, set of related concepts, etc.The ID of each incident is utilized for every query or retrieval.

The Knowledge-based interface
The knowledge-based interface essentially refers to one or more thin clients (Web/mobile app) making use of security mechanisms (https, login, etc.) to support the user flows of the experts.Indicative The overall business flow of the NAVMAT system is depicted in the following Figure 2:

Technical specifications
NAVMAT is developed and operates on an HP ProLiant DL360 G10 server, on Ubuntu operating system.In the core of the NAVMAT platform is Drupal, a web application framework, free and opensource, distributed under the GNU General Public License, that guarantees a scalable, flexible, functional and secure environment.It has a very wide supporting community, and is established as a very good choice for complex, content heavy and high traffic web applications.

On the front-end
NAVMAT utilizes Drupal capabilities for a secure user-roles-and-permissions system.Its rich UX/UI toolset, ensures a fully functioning and seamless user experience, either on larger screens (desktop computers) or on smaller ones (tablets, mobile phones).In addition, it exploits the powerful built-in serving-receiving-content-via-APIs features (Application Programming Interface), which guarantee a flawless inter-communication between all the different parts and components of the system.

On the back-end
Established frameworks for text analysis, indexing, clustering and classification are being evaluated for the AI aspect of the platform.ElasticSearch [6] has been chosen for standard text analysis tasks, e.g.stemming and string normalization, as well as for indexing (as explained below).A number of toolkits [7][8][9] are utilised to further increase the efficiency of the system: tools for classification and clustering; tools for graph text representation and approximate matching and tools for aggregation / summarization.The indexing and retrieval is built upon a MySQL RDBMS, integrated with the NAVMAT ontology.

NAVMAT prototype and evaluation methodology
The NAVMAT application has the ambition to integrate the currently existing and fast growing expertise in failure of materials and components with decision making on Maintenance, Repair, Overhaul/Operations (MRO), utilising Informatics tools.The software streamlines the tasks of content analysis and indexing, content aggregation, incident assessment, report writing, editing and publishing into a sequential procedure.It acts like the expert's personal assistant, charged with finding the relevant pieces of information and organizing them in a way that the expert will be able to assess the causes of any materials failure, decide the optimum intervention, and generate their next report with much less effort.The NAVMAT prototype was tested with a small number of failure incidents inserted and indexed, and a corresponding number of failure analysis and assessment reports and publications.

The prototype workflows
Currently, the prototype can offer: a) recording, tagging and indexing of incidents, reports and publications, while it allows the user to suggest new terms and tags and make comments on published incidents; b) searching and retrieving information on registered incidents, reports and publications; c) summarizing search results and suggesting similar content, using a similarity index algorithm; d) aggregating failure data by Class or Subclass and publishing User statistics.
Indicative screenshots of the NAVMAT prototype have been selected to show one of several business workflows based on the fundamental content of the system, the failure incident (Figure 3).From the Overview (home) page, the User can may View or Contribute.
The Visiting user may select the I n c i d e n t s l i s t , to which he/she can apply selection filters, associated to the Ontology.The selected I n c i d e n t is then displayed to the screen with a brief and basic description of the failure event, related images, tags from the Ontology, and suggestions for failure incidents on this ship type, failure incidents on the same subsystem type, failure incident on the same component type.It also suggests overall similar recorded incidents, ranked by a similarity index algorithm.At this point the User can access the full failure R e p o r t (if available) or browse similar failure incidents.This information, similarities and suggestions, are automatically calculated for each Incident, the moment it is inserted to the system by the Contributing User.

Evaluation process
The initial evaluation of the NAVMAT platform was designed and implemented on the user scenarios identified at the project conception stage: User Authentication and authorization, incident insertion, incident and document meta-data enrichment, Search for incidents, Taxonomy and Ontology update.The evaluation process was focused on Search capabilities assessment, Focus Group Usability testing, User Satisfaction.It included the following steps: Formation of a User Group (participation of 21 engineers), Briefing and introducing the User Group to the concept (Extended User Guide preparation), Online presentation (teleconference), Collection of feedback and suggestions (qualitatively and quantitatively), assessment and incorporation of feedback to the platform.
The final evaluation of NAVMAT prototype is planned to be conducted by a broader User Group, after the developing team has addressed, processed and incorporated the suggestions and remarks of the initial evaluation.

Post -project business strategy
By the end of February 2024), the prototype is expected to be fully operational.The beneficiaries already defined (project host organization) are currently developing a sustainability plan, including the model of service provision to additional external users, after solving confidentiality issues and access rights.In addition, the architecture of the NAVMAT allows for a) direct expansion of the system to other industries (beyond naval materials) by simply enriching the system Ontology b) language extension by enhancing the vocabulary (Ontology tags) in any selected language.Finally, future modules of NAVMAT system such as an Image analysis component and a Risk analysis components are being examined by the research team.
workflows include: − CRUD (Create/Read/Update/Delete) operations on reports / incidents / documents; − requests suggestions from the Document analysis component, concerning main concepts in the text; − enrichment of inserted documents through the Document analysis component; − storage of the enriched document into the repository; − efficient searching (based on ontology) for previous related incidents and related resources (publications, videos, etc.).

Figure 2 :
Figure 2: Schematic representation of the NAVMAT business flow model and architecture [1].

Figure 3 .
Figure 3. NAVMAT screenshots: From Home Page to the Incidents and the system suggestions