DE-102024210914-A1 - Computer-implemented method for simulating causality in a production line

Abstract

The invention relates to a computer-implemented method for simulating causality in production processes, wherein the method comprises the following steps: - Providing a structural causal model (10) for the production of products using a production line; - Initializing the structural causal model (10, S10) with the parameters with which the product is to be manufactured; - Performing a simulation for each product according to the initialized parameters (S16); and - Obtaining a simulated product sample (S26) for production from the simulation result.

Inventors

• Tim Pychynski
• Nicholas Tagliapietra
• Juergen Luettin
• Lavdim Halilaj

Assignees

• Robert Bosch Gesellschaft mit beschränkter Haftung

Dates

Publication Date

20260513

Application Date

20241113

Claims (14)

1. A computer-implemented method for simulating causality in production processes, comprising the following steps: - Providing a structural causal model (10) for the production of products using a production line; - Initializing the structural causal model (10, S10) with the parameters with which the product is to be manufactured; - Performing a simulation for each product according to the initialized parameters (S16); and - Obtaining a simulated product sample (S26) for production from the simulation result.
2. Computer-implemented method according to Claim 1 , wherein the products are divided into batches, with the structural causal model (10) being reinitialized for each batch (S12).
3. Computer-implemented method according to Claim 2 , where the production line comprises two or more parallel paths, and the simulation includes the assumption that the products are produced in parallel in the paths.
4. Computer-implemented method according to one of the preceding claims, wherein the simulation results of the products are concatenated (S20, S22).
5. Computer-implemented method according to one of the preceding claims, wherein the initialized parameters depend on the type of machine represented by the simulation, on the tolerances of the machine represented by the simulation and/or the product type of the products to be manufactured.
6. Computer-implemented method according to one of the preceding claims, wherein the simulation is an observational simulation and is used to identify dependencies of individual steps within the production line (S34).
7. Computer-implemented method according to one of the Claims 1 until 5 , where the simulation is an intervention simulation and is used to find an optimized parameterization of the production line (S48).
8. Computer-implemented method according to Claim 7 , wherein the simulation is performed in several simulation runs (S42), wherein finding the optimized parameterization (S48) involves using a cost function for each simulation run (S44), wherein the parameterization is adjusted in each simulation run (S46) to reduce the cost function.
9. Computer-implemented method according to one of the preceding claims, wherein the simulation result includes a Boolean value indicating whether the simulated product sample is within a specified error tolerance or not.
10. A computer-implemented method according to any of the preceding claims, wherein the structural causality model (10) comprises: - a set U of exogenous variables U i linked to external factors; - a set V of endogenous variables V i linked to other endogenous and/or exogenous variables; - a probability density distribution P U for each exogenous variable U i ; and - a mapping F for each exogenous variable Ui and each endogenous variable V i onto at least one further endogenous variable V j≠i ; wherein each endogenous variable V i is represented by a node, with a further final node outputting the simulation result.
11. Computer-implemented method according to Claim 10 , where each node or part of the nodes includes an additive component for simulating noise and/or errors.
12. Computer program containing program code to execute a procedure following one of the preceding steps when the computer program is run on a computer.
13. Computer-readable data carrier containing the program code of a computer program to execute a procedure according to one of the Claims 1 until 11 to be executed when the computer program is run on a computer.
14. System for simulating causality in production processes, wherein the system is configured to perform a process according to one of the Claims 1 until 11 to execute.

Description

The invention relates to a computer-implemented method for simulating causality in a production line. State of the art Optimizing product manufacturing is a crucial aspect for the manufacturing industry. Production must remain within budget and produce goods that meet all quality criteria. Simultaneously, production systems must be designed and ramped up quickly to enable a short time-to-market, which is also a critical factor for competitiveness. Therefore, manufacturing systems and processes must be continuously improved to remain competitive. Manufacturing new products requires making crucial decisions, such as defining tolerance limits for components or for production and quality control. After production begins, numerous further optimizations are regularly implemented on the production line, including adjusting and optimizing process parameters. This is often a very time-consuming and costly process, as optimization is generally not systematic and rarely utilizes all available data and information. The underlying cause-and-effect relationships between machine parameters, component parameters, and process parameters are not always known, so optimization is often carried out based on intuition and trial and error. This frequently relies on the individual expertise of specialists, which is not always available and can be quickly lost, for example, if the specialist leaves the company. The same applies to root cause analysis when problems or anomalies in production need to be explained and resolved. To solve this problem, various approaches utilize machine learning. For example, developing agents with an understanding of causality makes it possible to go beyond statistical coincidences. Furthermore, it is associated with desirable abilities such as logical reasoning and out-of-distribution generalization (Richens & Everitt, 2024). Using causality tools (Pearl, 2009), the data generation process (DGP) can be uncovered and manipulated to gain a deeper understanding of the modeled system. For example, causal analysis can be used to estimate the effects of interventions on a system, taking into account factors such as biases and missing data (Mohan & Pearl, 2019). To make progress in this area, a fair and comprehensive evaluation of causal algorithms is crucial. Benchmark tests, which analyze methods from various perspectives, are a fundamental component for advancing the field. However, making comparisons across multiple domains presents several challenges. From a practical perspective, one of the main obstacles hindering progress in understanding causality is the lack of publicly available benchmarks to support methodological evaluation. For example, when benchmarking using real-world data, the true data generation process may be partially or even completely unknown. For example, when considering an intervention within a production line, it is impossible to simultaneously observe both the potential outcomes of intervention and those of no intervention. This means that the underlying truth values of the causal estimates are unknown. Consequently, purely factual observational data are insufficient for evaluation due to the lack of counterfactual measurements. A similar challenge was highlighted by Gentzel et al. (2019), who emphasized the importance of evaluating intervention measures and subsequent tasks. However, in many cases, obtaining intervention data is either impossible or extremely costly. Curth et al. (2021) argued that algorithms that match the assumptions of the data generation process have an advantage with these specific benchmarks. However, the results may not be generalizable to other scenarios. Scalability is not only a further challenge for inference tasks, but also across the entire field of causality. The related task of causal discovery (CD), that is, the reconstruction of the causality diagram from data, suffers from similar pressures, where mathematical guarantees are often sacrificed in exchange for computational feasibility (Zheng et al., 2018b). Novel causal models are often tested on representative causal graphs with simple structural equations, which lack the complexity of the real world. In contrast to other work investigating applications of causality in medicine, genetics, and ecology, this invention focuses on the manufacturing sector, which has historically seen only experimental and isolated applications (Vukovic & Thalmann, 2022). In a theoretical and empirical evaluation of simple causal graphs by Zecevic et al. (2023), the unsolvability of marginal inference and the scaling laws of various causal models were highlighted. If the goal is to reduce the complexity of various unsolvable queries, it is possible to use solvable probabilistic models such as sum-product networks (SPNs) (Poon & Domingos, 2012). Furthermore, SPNs can be used to model causal phenomena. Rubin's "Potential Outcomes (PO)" framework (Imbens & Rubin, 2015) can be used to solve the scalability problem due to