Bosch Line Information System KG
- Bosch Line Information System KG is a semantic integration layer that consolidates heterogeneous manufacturing data from plants, lines, and machines.
- It provides a unified, queryable knowledge graph that enhances quality monitoring, traceability, and industrial analytics through ontological structuring.
- The system employs ontology reshaping and literal-aware modeling to optimize data integration, improve query accuracy, and support NL-to-SPARQL interaction.
Bosch Line Information System KG is a large-scale, production-grade semantic integration layer for manufacturing data that consolidates heterogeneous information from plants, lines, stations, machines, processes, materials, events, transactions, and related master data into a coherent, queryable knowledge graph. In the Bosch literature, it appears both as an enterprise integration substrate for cross-plant knowledge access and as a domain-specific graph construction pattern for welding quality monitoring, where operational logs, settings metadata, and sensor-derived literals are organized for link prediction and industrial analytics (Monka et al., 30 Jul 2025, Tan et al., 2023).
1. Industrial scope and operational role
The Bosch LIS KG is designed to connect disparate transactional and master data systems into a shared semantic layer. Its stated purpose is to integrate “a wide range of data from multiple manufacturing plants, lines, and machines,” thereby supporting queryability across real production topology and activities. The integrated sources explicitly include Manufacturing Execution Systems, Enterprise Resource Planning, and Master Data systems, and the ontology is described as “an abstraction layer over the underlying manufacturing data sources,” modeled according to the relational DB schema of the source data (Monka et al., 30 Jul 2025).
At the time of the LLM-based evaluation, the LIS deployment covered more than 15 plants, over 2,700 production lines, 16,000 physical machines, and more than 400 distinct manufacturing processes. These figures establish the KG as an industrial-scale asset rather than a laboratory benchmark. The modeled content spans manufacturing structure and execution, including plants, production lines, stations, machines or equipment, processes and their configurations, materials, events or transactions, and traceability contexts. Benchmark questions also refer to quality or traceability databases and “transactional data for data analytics,” indicating that the KG links conceptual manufacturing entities with data-source metadata relevant for retrieval and downstream use (Monka et al., 30 Jul 2025).
A Bosch welding case provides a more localized but technically detailed instance of this general pattern. There, the graph is built from three line information sources: the welding protocol, welding setting metadata, and high-frequency sensor measurements streamed during the weld. This narrower graph targets quality monitoring and traceability for resistance spot welding, but it uses the same core premise: manufacturing data become more usable when modeled as a relationally structured KG rather than as isolated tables or time-series files (Tan et al., 2023).
| Aspect | Enterprise LIS KG | Welding KG slice |
|---|---|---|
| Main scope | Cross-plant manufacturing integration | Welding quality monitoring |
| Main sources | MES, ERP, Master Data systems | Protocol, settings, sensor streams |
| Reported scale | plants, lines, machines, processes | 44,801 triples, 3,342 entities, 26 relations |
The main significance of this architecture is semantic unification. The KG supplies a machine-readable schema for cross-plant analytics, maintenance planning, data engineering, and LLM-mediated access, while also supporting specialized industrial tasks such as spot-level quality estimation and traceability in welding (Monka et al., 30 Jul 2025, Tan et al., 2023).
2. Ontological structure and semantic commitments
In the welding formulation, the KG is explicitly defined as , where is the set of entities, the set of binary relations, and the set of literals, with , , and 0. The paper distinguishes object properties between entities from datatype properties attaching literals, even though the implemented literal-aware method represents discretized numeric values as entities linked by object-like edges (Tan et al., 2023).
The enterprise LIS ontology is presented through a narrower set of schema primitives that are especially relevant for query generation and ontology reduction. The structural predicates emphasized in prompting experiments are rdf:type, rdfs:label, rdfs:domain, rdfs:range, rdfs:subClassOf, and rdfs:subPropertyOf; richer variants also include rdfs:comment and [owl](https://www.emergentmind.com/topics/owl):inverseOf. These are not merely documentation artifacts. They function as the backbone for domain-specific triple patterns, taxonomic scoping, and constraint-aware SPARQL generation (Monka et al., 30 Jul 2025).
The class inventory is not published exhaustively, but the benchmark questions and dataset description imply a core schema around Plant, Line, Station, Machine or Equipment, Process, and Material. Typical structural relations include Line-in-Plant, Station-on-Line, Equipment-in-Line, Material-used-on-Line, and Process-used-by-Line. In the welding subgraph, the listed entity types are welding spot or operation, welding machine, welding program, car body part or component, car body as ID or class, process stage, material or specification, operator when available, and discretized literal buckets such as current, voltage, resistance, welding time, and power bins (Monka et al., 30 Jul 2025, Tan et al., 2023).
A notable semantic issue arises in alignment to the I40 Core Information Model via CIMM. In CIMM, cimm:Equipment is described as a logical concept used to represent hierarchical elements such as Plant, Line, and Station, whereas practitioners often use “equipment” to mean physical machines. This distinction matters operationally because SPARQL generation can fail if logical hierarchy and physical assets are conflated. The Bosch results therefore emphasize precise class and property definitions, comments, and inverse-property information as mechanisms for reducing semantic drift in query generation (Monka et al., 30 Jul 2025).
A common misconception is that a manufacturing KG is equivalent to a catalog of assets. The Bosch materials indicate a broader view: the LIS KG includes not only structural assets but also processes, configurations, materials, events, transactions, traceability links, and data-access metadata. In the welding case it further incorporates sensor-derived value ranges and quality indicators as graph elements, so the graph represents execution context as well as topology (Monka et al., 30 Jul 2025, Tan et al., 2023).
3. KG generation, ontology reshaping, and preprocessing
Bosch’s welding-oriented ontology reshaping work addresses the tension between knowledge-orientation and data-orientation in ontology engineering. The method takes a knowledge-oriented domain ontology 1, raw industrial data 2, mappings 3, a user-selected main class 4, and optional user information 5, and produces a smaller KG schema through the functional formulation
6
The stated goals are completeness, efficiency, and simplicity: the generated KG should represent the raw data completely, use as few classes, entities, and properties as necessary, and avoid redundant or dummy entities and unnecessarily deep graph structures (Zhou et al., 2022).
The reshaping procedure initializes the schema with the main class, adds classes that map to table names, identifies further classes via mapped properties and cues such as “ID” or “Name,” and then connects classes according to relations in the original ontology or according to user preferences. In Bosch welding, the industrial example centers on electric resistance welding, specifically resistance spot welding on automotive body lines. The paper’s motivating example shows how a deep ontology path such as WeldingSoftwareSystem → MeasurementModule → OperationCurveCurrent → {mean, array} can be simplified into a data-oriented schema centered on WeldingOperation with direct data properties for CurrentMeanValue and CurrentArrayValue. This reduction is intended to improve usability and queryability by removing dummy nodes (Zhou et al., 2022).
The evaluation uses an OWL 2 ontology expressible in Description Logics SHI(D), with 1249 axioms, 147 classes, 145 object properties, and 132 datatype properties. The raw data come from welding production lines in a German factory and include txt, csv, RUI, and SQL sources transformed into relational tables. The reported evaluation subset covers one large table of welding operation records plus multiple tables of sensor measurements, amounting to about 4.315 million records covering approximately 1000 welding operations and estimated to be related to 100 cars (Zhou et al., 2022).
Quantitatively, both the baseline and Onto-Reshape achieved 100% data coverage, but Onto-Reshape reduced time cost from 139.3–836.9 seconds to 21.6–128.0 seconds across the six subsets, reduced storage from 1.3–4.8 MB to 0.5–2.4 MB, and eliminated dummy entities entirely, whereas the baseline produced average dummy-entity counts between 4100.0 and 4800.0. Average root-to-leaf depth also fell from roughly 3.2–3.4 to roughly 1.2–2.1 (Zhou et al., 2022).
Within the welding KGE study, preprocessing follows a related but more task-specific pipeline. Relational tables from the welding protocol, settings, and sensor measurements are read; over 200 columns are present. Empty columns and columns with a single unique value are removed, representative features are retained based on welding domain knowledge, and sensitive production values are anonymized by using numeric ranges instead of raw values. Sensor streams at millisecond resolution are aggregated into mean values per stage and overall mean, discretized into bins, turned into entities, and linked back to spot entities through relations such as has_parameter_current_bin and has_parameter_voltage_bin (Tan et al., 2023).
This suggests that Bosch treats schema construction and feature engineering as coupled operations. In the reshaping work, the main concern is how to derive a compact data-oriented schema from a larger ontology; in the welding KGE work, the concern is how to expose process, settings, and sensor information to graph learning models through structurally meaningful entity and relation design (Zhou et al., 2022, Tan et al., 2023).
4. Knowledge access and NL-to-SPARQL over LIS
A major function of the LIS KG is to support knowledge access without requiring end users to formulate SPARQL directly. Bosch evaluates LLM-mediated NL-to-SPARQL generation over LIS and CIMM through context-aware prompting, formalizing the mapping as
7
where 8 is a natural-language question and 9 is KG context constructed from the ontology by content selection and representation (Monka et al., 30 Jul 2025).
The benchmark contains 17 real-world manufacturing questions spanning five personas: Benchmarking Engineer or Planner, Maintenance Planner, Technology Developer, and Data Engineer. Representative questions include “How many Stations are on Line X from Plant Y?”, “How many equipments are on Line Y in Plant X?”, “Give me the list of material numbers for the line X?”, “For process X: which other lines share that process within the same traceability database?”, and “For production line X: where and how can I access the transactional data for data analytics?” Gold SPARQL queries exist for both LIS and CIMM, though their exact text is not published (Monka et al., 30 Jul 2025).
The prompting framework distinguishes four ontology-content variants. Ont_A is a naive reduction using structural predicates only. Ont_B and Ont_C are context-based reductions retrieved through vector-space semantic similarity, selecting the top 25 relevant classes and properties and their neighbors; Ont_C additionally includes comments and axioms. Ont_D further augments Ont_C with ontology-based enrichment. The ontology can then be presented to the model as Graph (RDF Turtle), Table, or Table-sorted, and paired with one of three prompt templates: P_simple, P_example, or P_domain (Monka et al., 30 Jul 2025).
The central quantitative metric is hallucination accuracy based on ontology-grounded triple matches:
0
where matches and mismatches are counted by checking whether the subject, predicate, and object URIs in the generated SPARQL exist in the ontology. The paper also defines Precision, Recall, and F1 over triple patterns to characterize missing or hallucinated schema elements (Monka et al., 30 Jul 2025).
For LIS using GPT-3.5-16K, context-based reduction substantially outperforms naive reduction. Under P_simple, Graph with Ont_B reaches 0.95 accuracy, whereas Table with Ont_A reaches 0.59 and Table-sorted with Ont_A reaches 0.47. Under P_domain, Table achieves 0.97 for both Ont_C and Ont_D, and Table-sorted reaches 0.96 for Ont_C and Ont_D. The paper summarizes the overall effect as a 20–30% average improvement from context-aware content selection and an additional 5–8% gain from domain-specific examples relative to generic examples (Monka et al., 30 Jul 2025).
The practical implication is not that larger prompts are always better, but that “adequate context” is decisive. Bosch reports fewer wrong class or property URIs, fewer incorrect joins, and fewer missing constraints when prompts contain only the relevant ontology fragment, together with domain and range information and, when feasible, comments and inverse properties. This is also where the earlier semantic ambiguity around “equipment” becomes operationally significant: better schema context helps the model distinguish logical hierarchy from physical assets (Monka et al., 30 Jul 2025).
5. Welding quality monitoring as a literal-aware LIS-style subgraph
The welding case instantiates Bosch’s line-information approach in a graph used for two industrial questions: estimating welding spot diameter and assigning each welded spot to a specific car body part. Instead of treating these as standard regression and multiclass classification tasks, the study reformulates both as link prediction: 1 for diameter and 2 for traceability (Tan et al., 2023).
The experimental KG contains 44,801 triples, 3,342 entities, and 26 relations, including 18 machines, 181 program IDs, 613 carbodies, and 327 literal entities. Key relation examples include conducted_on_machine, uses_program, has_component_type, welded_for_stage, has_material, has_parameter_current_bin, has_parameter_voltage_bin, has_parameter_resistance_bin, has_parameter_welding_time_bin, has_parameter_power_bin, and belongs_to_carbody_part. Diameter, current, voltage, resistance, power or PWM, welding force or temperature when present, welding duration, and spot coordinates if present are handled through aggregation and discretization, with numeric bins turned into graph entities (Tan et al., 2023).
The evaluated KGE methods are TransE, RotatE, and AttH, together with KGE-MLP hybrid variants TransE*, DistMult*, and HolE*. The paper uses standard scoring functions, including
3
for TransE and
4
for RotatE, and trains the models with a ranking objective using corrupted tails as negatives. Embedding sizes are searched in 5, with 128 or 256 identified as the best trade-offs by validation MRR (Tan et al., 2023).
The main quantitative results are as follows.
| Task | Best reported result | Additional note |
|---|---|---|
| Diameter class prediction | TransE Hits@1 6, MRR 7 | MLP has best nrmse 8 |
| Car body part prediction | TransE Hits@1 9, MRR 0 | Hits@GroupBy3 reaches 1 |
For Q1, Hits@1 values are 0.42 ± 0.02 for TransE, 0.39 ± 0.01 for MLP, 0.31 ± 0.05 for AttH, and 0.25 ± 0.01 for RotatE. MRR is 0.65 ± 0.01 for TransE, 0.57 ± 0.04 for AttH, and 0.49 ± 0.00 for RotatE. Reported nrmse values are 0.05 ± 0.01 for MLP, 0.06 ± 0.00 for TransE, 0.06 ± 0.01 for AttH, and 0.08 ± 0.01 for RotatE. For Q2, Hits@1 values are 0.64 ± 0.01 for TransE, 0.61 ± 0.01 for MLP, 0.53 ± 0.03 for AttH, and 0.52 ± 0.01 for RotatE; Hits@GroupBy3 reaches 0.85 ± 0.01 for TransE, 0.79 ± 0.03 for AttH, and 0.81 ± 0.01 for RotatE. Test-time latency is reported as less than 0.03 seconds for all models (Tan et al., 2023).
The ablations are especially informative. The KGE-MLP hybrids underperform pure KGE on both tasks: for Q1, TransE*, DistMult*, and HolE* achieve Hits@1 values of 0.17, 0.22, and 0.21, respectively, compared with 0.42 for TransE; for Q2, the corresponding values are 0.34, 0.34, and 0.37, compared with 0.64 for TransE. Literal removal produces opposite effects across the two tasks. For diameter prediction, removing literals improves TransE from Hits@1 0.42 to 0.45 and from nrmse 0.06 to 0.04, which the paper attributes to diameter entities being isolated through rdf:type. For car body prediction, removing literals degrades performance from Hits@1 0.64 to 0.53 and from Hits@GroupBy3 0.85 to 0.78 because carbody_part is “well connected” and literals provide useful contextual structure (Tan et al., 2023).
A plausible implication is that Bosch’s graph-learning results are governed less by a generic preference for literal-aware models than by local graph topology and schema connectivity. The paper’s own interpretation is consistent with this: literal bins help when attached to well-connected prediction targets and may hinder learning when the target subgraph is structurally sparse (Tan et al., 2023).
6. Limitations, semantic issues, and future directions
Several limitations are explicit across the Bosch papers. In the welding KGE study, temporal information is present in welding timestamps and stage segmentation, but no explicit temporal KG is constructed. Sensor time series are aggregated into stage-wise and overall means and then discretized; no time-window sequences or temporal edges are modeled. The authors therefore identify temporal KGE and richer sequence-aware modeling as natural extensions (Tan et al., 2023).
Literal integration is another limitation. The welding work does not use learned continuous literal encoders or gates; instead, numeric literals are discretized and added as entities. The authors explicitly point to learned literal encoders such as LiteralE, numeric feature gates, continuous attribute models, and richer schema design for diameter connectivity as future work. They also note that assigning spots to millions of potential car bodies remains challenging, even though link prediction is preferable to direct multiclass classification when the label space is very large (Tan et al., 2023).
In the NL-to-SPARQL setting, the core issue is not only hallucination but also non-determinism, token limits, and concept ambiguity. Bosch reports that long, multi-hop queries with aggregates or nuanced semantics remain challenging, and that richer schema retrieval and structured tool use are planned. The semantic drift around the term “equipment” across LIS and CIMM is emblematic: comments and axioms help, but they do not fully eliminate ambiguity (Monka et al., 30 Jul 2025).
The ontology reshaping work identifies a different but related concern: simplifying a knowledge-oriented ontology into a smaller data-oriented schema “can not fully retain the semantics of the domain ontology.” Future work is therefore linked to uniform interpolation or forgetting in Description Logics, deeper theoretical study of reshaping, and broader investigation of ontology-mediated analytics and virtual knowledge graphs (Zhou et al., 2022).
These limitations clarify what the Bosch LIS KG is not. It is not, in the reported studies, an explicit temporal knowledge graph; it is not a guarantee that all domain semantics are preserved after schema reshaping; and it is not a self-sufficient interface that eliminates the need for ontology-aware validation in LLM-generated queries. At the same time, the reported deployment trajectory is concrete: the work is part of Bosch’s neuro-symbolic AI for Industry 4.0 efforts, the welding KG and code are open-sourced, solutions are under evaluation with Bosch Research and factory partners in Germany, and the reported inference time of less than 0.03 seconds is compatible with near-real-time monitoring dashboards (Monka et al., 30 Jul 2025, Tan et al., 2023).
Overall, Bosch Line Information System KG denotes a family of closely related semantic integration efforts: an enterprise-scale LIS ontology and KG for manufacturing knowledge access, a welding-focused literal-aware graph for industrial link prediction, and an ontology-reshaping methodology for turning knowledge-oriented ontologies into compact, data-oriented schemas. Taken together, these works position the KG as both an integration substrate and an operational analytic instrument, with demonstrated utility in query mediation, traceability, and quality monitoring, but with open technical questions around temporal representation, semantic retention, literal modeling, and robust schema-grounded LLM interaction (Monka et al., 30 Jul 2025, Tan et al., 2023, Zhou et al., 2022).