Laying the Foundation for Industrial Safety: Haizhi Technology Cooperates with an Operator to Build the "Ontology Base" for Work Safety Supervision

The national standard series "Artificial Intelligence - Interconnection of Agents" was officially released a few days ago, marking that the interconnection and interoperability of intelligent agents has entered the era of standardization. However, in the in-depth stage of digital transformation of industries such as low-altitude economy, urban governance, energy and manufacturing, there are still widespread pain points including data silos, lagged judgment of complex relationships, and difficulties in cross-domain risk control. 

In practice, HZDATA Technology builds a solid base for massive associated data relying on graph databases, realizes in-depth integration and full-factor association of multi-source heterogeneous data by using knowledge graphs, and constructs interpretable business rule constraints with Ontology to drive intelligent agents to achieve deterministic execution. This full-stack technical system effectively eliminates risks such as hallucinations and uncontrollability of large models in the process of industrial implementation. Its technical direction is highly consistent with the requirements of national standards, and can provide solid support for all industries to quickly build a cross-domain integrated intelligent governance framework.

This series of articles will deeply disassemble HZDATA's real implementation cases, share our practical experience in solving heterogeneous system collaboration and complex business logic association, and hope to provide a reference model for refined and intelligent collaborative governance in all walks of life. 

Industry Practices 

On September 7, the Ministry of Industry and Information Technology issued the "15th Five-Year Plan" for the Development of the Information and Communication Industry, explicitly proposing to promote the in-depth integration of 5G and the industrial Internet, accelerate the innovative application of information and communication technologies in work safety and other fields, and comprehensively expand the scale of digital and intelligent applications in the production sector. With the in-depth integration of 5G and industrial production, industrial 5G private networks are accelerating their extension from basic network connections to core businesses such as production, safety and management.

However, the data involved in industrial work safety comes from a wide range of sources and in diverse formats, including not only text information such as regulations, accident cases and inspection records, but also a large amount of unstructured data such as on-site pictures, surveillance videos and audios. Such information has long been scattered in different business systems and storage media, making it difficult to be uniformly retrieved, associated and reasoned. How to convert scattered multi-modal work safety supervision information into retrievable, associable and inferable knowledge capabilities has become an important challenge for the current intelligent construction of industrial safety.

To address this industry pain point, HZDATA Technology has reached in-depth cooperation with a certain operator research institute. The two sides will jointly build an integrated work safety supervision ontology base featuring "multi-modal retrieval + knowledge graph" targeting the industrial safety monitoring scenario of integrated equipment on industrial 5G private networks. This cooperation is an important practice for HZDATA Technology to expand its core technical capabilities such as "graph-model integration" to the vertical scenario of work safety.

1. The Dilemma of Multi-modal Integration: How to Make Work Safety Supervision Data "Interconnected" 

In the process of industrial 5G private networks accelerating their penetration into the front line of production, the safety monitoring of integrated equipment is evolving from "being visible" to "being understandable and predictable". Behind this lies the problem of whether massive multi-modal work safety supervision data can be truly put into use.

The data base of industrial work safety is inherently "fragmented".

One investigation of industrial safety hazards involves multiple pieces of information such as equipment operation parameters, operator qualifications, inspection records, historical accident cases, emergency response plans, and safety regulation clauses. These data are scattered in production management systems, safety monitoring platforms, equipment operation and maintenance databases, and document management systems, with different formats, inconsistent standards and no interconnection. When safety supervisors want to restore the full context of a hazard investigation, they often need to switch back and forth between multiple systems and perform manual comparison, which greatly reduces efficiency and accuracy.

The deeper challenge lies in the technical aspect.

Heterogeneous data such as texts, images, audios and videos are not only completely different in format and structure, but also have separate retrieval methods — keyword search for texts, label matching for images, and timeline scrolling for videos, lacking unified semantic association with each other. When the work safety supervision scenario requires "one-time retrieval with all multi-modal outputs", or it is necessary to associate an inspection record with relevant historical accident cases, on-site images and regulatory basis, the traditional retrieval method of separate databases and tables is far from sufficient.

At the same time, the implicit knowledge hidden behind multi-modal data — the association between equipment and hazards, the causality between accidents and response measures, the correspondence between inspection records and regulatory norms, has never been systematically refined and precipitated. It can only remain in the literal meaning of documents or the minds of experienced personnel, and is difficult to be retrieved, reused and reasoned.

As industrial safety monitoring shifts from "people looking for data" to "data being put into use", a set of underlying technical solutions that can not only break through the barriers of multi-modal data, but also convert implicit associations into inferable knowledge is in urgent need.

2. Breaking the Dilemma with "Graph-Model Integration": Upgrading from Retrieval to Knowledge 

In response to this demand, relying on its self-developed core technology of "graph-model integration", HZDATA Technology has injected two core capabilities into the work safety supervision knowledge base platform: breaking data barriers through a multi-modal unified retrieval engine, and activating implicit knowledge combined with dynamic knowledge graph construction technology, so as to truly enable massive work safety supervision data to be "interconnected, accurately found and inferable".

The value of this solution is specifically reflected in the following two aspects:

Unified Retrieval: Enabling Scattered Data to Be "Found at One Go".

The platform performs unified access and processing of multi-source heterogeneous work safety supervision data through a unified retrieval engine, and incorporates different types of information such as texts, pictures, audios and videos into a unified knowledge resource system. Relying on the unified capabilities of data access, processing and retrieval, work safety supervision data has realized the transformation from "distributed storage" to "unified organization and unified retrieval", which greatly improves the efficiency of hazard investigation and emergency response.

For example, when a safety supervisor inputs a description of equipment abnormality, the platform can automatically retrieve and associate the equipment's inspection images in the past three months, voice alarm records of similar faults, and corresponding disposal specifications, shortening the work that originally required manual cross-system comparison for several hours to the minute level, so that work safety related information can be quickly discovered and invoked.

Knowledge Graph: Enabling Implicit Associations to Be "Explicit and Inferable".

To fully release the in-depth value of work safety supervision data, the platform further combines HZDATA's self-developed GraphRAG technology to automatically extract entities, relationships and associated knowledge in the work safety supervision field from original data, and organizes the information originally scattered in documents, databases and business systems into a structured "work safety supervision knowledge network".

Taking a confined space operation as an example, the platform can automatically associate the operation approval form, personnel qualification certificates, gas detection records, emergency rescue plan, and historical accident cases into a complete "safety relationship network", allowing safety supervisors to grasp the full picture of the operation and potential risk points at a glance. This innovation breaks data silos, enables scattered information to truly form an inferable and traceable knowledge system, and realizes the leap of implicit associations from "experience hidden in human minds" to "explicit and inferable knowledge".

In the future, relying on the further integration of unified retrieval, knowledge graph and AI capabilities, this integrated work safety supervision ontology base will continue to expand rich applications such as intelligent question answering, risk correlation analysis, auxiliary hazard judgment, accident knowledge tracing, and safety decision assistance.

This will promote the industrial safety monitoring to further move to a new stage of "being understandable, associable and decision-aided" from "being visible and searchable", provide a more solid digital foundation for factory construction and intelligent industrial production safety, and fully escort the high-quality development of the industry.

This article is from the WeChat Official Account "HZDATA Technology", Author: HZ, and is published with authorization from 36Kr.