Method for automatically performing a root cause analysis in a production plant

Abstract

The invention relates to a method (100) for automatically performing a root cause analysis in a production plant (1), comprising the following steps: - Determining (101) at least two sub-areas (2) for root cause analysis based on a grouping of sensors (3), wherein at least one sensor (3) is assigned to a respective sub-area (2) that is located in the same area of ​​the production plant (1), - Encoding (102) measurements of at least one sensor (3) of a respective sub-area (2) into a latent representation (4) using a machine learning model, - Determining (103) a posterior distribution of the latent representation (4) of each respective sub-area (2) for a case in which a defective product is produced by the production plant (1), - Determining (104) a respective contribution of each sub-area (2) to a production of the defective product, whereby a replacement of each individual sub-area (2) and a replacement of all possible combinations of the sub-areas (2) is tested on the basis of the corresponding latent representations (4), while for the remaining sub-areas (2) the determined posterior distribution of the latent representation (4) is used, - Performing (105) the root cause analysis in the production plant (1) based on the respective determined contribution of the sub-areas (2). Furthermore, the invention relates to a computer program, a device and a storage medium for this purpose.

Description

[0001] The invention relates to a method for automatically performing a root cause analysis in a production plant. Furthermore, the invention relates to a computer program, a device, and a storage medium for this purpose. State of the art

[0002] Automated root cause analysis in production facilities represents an advanced approach to efficiently and accurately identifying the root causes of errors, defects, or problems in the production process. By utilizing automated systems and algorithms, these analyses enable faster and often more precise identification of problem causes than traditional, manual methods.

[0003] At the heart of automated root cause analysis lies the continuous collection and processing of data. Sensors, machine logs, and other information-rich sources continuously provide data about the production process. Modern analytical tools based on artificial intelligence (AI) and machine learning (ML) are used to identify patterns and anomalies in this data. These systems can understand complex relationships between different variables and automatically identify potential causes of deviations or errors.

[0004] Previous approaches to root cause analysis in large structural causal models (SCMs) are generally very computationally intensive. These approaches typically rely on the precise estimation of a game-theoretic quantity known as a Shapley value. Shapley values ​​play a crucial role in estimating the contribution of each potential cause to the observed system failure within a framework that applies a graded and attributed contribution principle.

[0005] In mathematical terms and in their exact formulation, Shapley values ​​grow exponentially with the number of nodes in a structural causal model (SCM), also known as a graphical causal model. This, in turn, makes calculating Shapley values ​​for SCMs with more than 10 to 20 nodes—far fewer than the graphs used in manufacturing (several hundred measurement nodes)—impractical.

[0006] According to current state-of-the-art procedures, estimates for these quantities are calculated (e.g., by sampling, least-squares fitting, early stopping, etc.) to obtain a reliable estimate with minimal computational effort. However, empirical studies of these known methods have shown that their performance—that is, either the prediction performance or the computational load—is highly unpredictable and cannot be guaranteed for efficient use in practical scenarios. Furthermore, it should be noted that incorporating expert knowledge into such estimates is not straightforward and, in many cases, cannot be easily automated. Disclosure of the invention

[0007] The invention relates to a method with the features of claim 1, a computer program with the features of claim 8, a device with the features of claim 9, and a computer-readable storage medium with the features of claim 10. Further features and details of the invention will become apparent from the respective dependent claims, the description, and the drawings. Features and details described in connection with the method according to the invention naturally also apply in connection with the computer program, the device, and the computer-readable storage medium according to the invention, and vice versa, so that a reciprocal reference is always possible with regard to the disclosure of the invention.

[0008] The invention relates in particular to a method for automatically performing a root cause analysis in a production plant, comprising the following steps, wherein the steps can be performed repeatedly and / or sequentially. The production plant can in particular be described by a "Directed Acyclic Graph" (DAG). This model, or graph, in particular comprises nodes and directed edges, wherein the edges between the nodes have only one direction. This means in particular that each edge points from one node to another and thus represents a non-symmetric relationship (e.g., "A leads to B," but not, or not necessarily, "B leads to A"). Furthermore, the graph preferably does not include any cycles. This means in particular that it is impossible to start from a node, follow a sequence of edges, and return to the same node.

[0009] In a first step, preferably at least two sub-areas are defined for root cause analysis based on a grouping of sensors. In particular, at least one sensor is assigned to each sub-area, which is located in the same area of ​​the production plant. For example, one sub-area could correspond to a station in the production plant where drilling takes place, and another sub-area could correspond to a station where quality assessment is performed. The sensors are preferably represented by nodes in the model or graph.

[0010] In a further step, measurements from at least one sensor of a respective sub-area are preferably encoded into a latent representation using a machine learning model. Within the scope of the present invention, the measurements from the at least one sensor of a sub-area can also be referred to as a multidimensional or multivariate observation of a sub-area. The latent representation is, in particular, a compressed representation of these measurements from the at least one sensor. This step advantageously allows a large number of sensor measurements, or the corresponding nodes, to be combined for the sub-area, thereby significantly reducing the complexity of the model or graph representing the production plant.

[0011] In a further step, a posterior distribution of the latent representation of each respective sub-area is preferably determined for a case in which a defective product is produced by the production plant. In other words, for each sub-area, it is determined how the sensor measurements might turn out or appear when the defective product is produced by the production plant. The "posterior distribution" is, in particular, a probability distribution calculated after data observation. In the context of an autoencoder, the posterior distribution refers specifically to the distribution of the latent representation after input data has been processed. This input data could, for example, result from measurements taken by the production plant's sensors during operation in which the defective product is produced by the plant.

[0012] In a further step, the respective contribution of each sub-area to the production of the defective product is preferably determined. In particular, replacing each individual sub-area and replacing all possible combinations of the sub-areas is tested based on the corresponding latent representations, preferably a respective prior distribution of the latent representation. For the remaining sub-areas, the determined posterior distribution of the latent representation is preferably used. In simplified terms, this determines in which cases, i.e., through which sub-areas or combinations of sub-areas, a defective product results. Within the scope of the present invention, it is also conceivable to simulate several runs and determine a number of resulting defective products. The defective product can, for example, be located at a final node in the model, or...The graph representing the production plant can be identified. This last node can represent an end-of-line analysis of a respective produced product.

[0013] In a further step, a root cause analysis is preferably carried out in the production plant based on the specific contribution of each sub-area. In particular, at least one sub-area is identified that is responsible for or contributes to a failure or impairment of the production plant or to the production of the defective product.

[0014] In particular, the aforementioned method for automatically performing a root cause analysis can provide a control signal for controlling a technical system, especially a production plant. Specifically, the control signal can be used to control a fail-safe operation. Fail-safe means that a system, especially a production plant, can detect when it is no longer functioning using the described method and then enter a passive or safe state. In other words, this is a method for detecting the occurrence of errors in systems and providing a control signal that brings the system into a safe state in response.The control signal can be used in particular to establish a safe state, taking into account the sub-area that is responsible for or contributes to a failure or impairment of the production plant or to the production of the defective product.

[0015] In particular, providing the control signal can also include transmitting the control signal to a control unit of the system, especially a control unit of the production plant. This transmission can be wireless or wired, for example, both via the internet and locally.

[0016] Furthermore, it is conceivable that the machine learning model is a variational autoencoder and that the encoding is performed using an encoder module of the variational autoencoder. A variational autoencoder (VAE) is, in particular, a type of autoencoder, a neural network designed for unsupervised learning. The variational autoencoder (VAE) preferably comprises an encoder module and a decoder module. The encoder module can convert input data, especially sensor measurements, into a smaller, condensed representation, i.e., the latent representation. The decoder module can then reconstruct the original data from this condensed representation. The VAE differs from a conventional autoencoder, in particular, by introducing a probabilistic interpretation of the latent space.Instead of directly encoding input data into a fixed point in latent space, the VAE preferably models the encoding as a probability distribution. This can mean that for each input, a distribution of possible encodings is generated in latent space, for example, a normal distribution with a mean and a standard deviation. The advantages of a VAE include the following: Due to its probabilistic nature and regularization, the VAE is particularly less prone to overfitting to training data. Furthermore, VAEs can generate new data that resemble the training data by drawing random points from latent space and transforming them into realistic data using the decoder module. Additionally, examining the latent space can provide insights into the structure and underlying patterns of the data.

[0017] A further advantage is that, during coding, the sensor measurements can be specific to the regular operation of the production plant, so that the latent representation provides a condensed representation of the measurements of at least one sensor during regular operation. Regular operation is defined as when no defective products are manufactured by the production plant and the plant runs without impairment. This allows, advantageously, in the step of determining the respective contribution of each sub-area to the production of the defective product using the machine model, the respective sub-areas to be replaced by the latent representations that represent regular measurements for those sub-areas.

[0018] It is also advantageous to calculate Shapley values ​​for each sub-area when determining its respective contribution to the production of the defective product. The Shapley value of a sub-area is determined, in particular, by calculating an average marginal contribution of that sub-area across all possible combinations of sub-areas. This may require considering all possible subsets of sub-areas and evaluating how adding the sub-area in question changes the production of the defective product compared to these subsets. It may also require considering all possible subsets of the sub-areas and evaluating how adding the subset in question changes the marginal contribution of the sub-area under consideration to the production of the defective product.Shapley values ​​can advantageously provide a fair method to assess the contribution of each sub-area, as they take into account all possible combinations and interactions between the sub-areas.

[0019] Furthermore, within the scope of the invention, it is conceivable that carrying out the root cause analysis includes the following step: - Comparing the respective contributions to identify at least one sub-area with the highest contribution.

[0020] The at least one sub-area with the highest contribution can then be examined in more detail. This can, for example, also be done automatically, as described below, or, advantageously, this knowledge can be used to initiate corresponding investigations by specialists in the at least one sub-area with the highest contribution.

[0021] Preferably, the procedure may further include the following step: - Calculating the respective contribution of the measurements of the at least one sensor in the at least one sub-area with the highest contribution, in order to further carry out the root cause analysis in the production plant on the basis of the respective calculated contribution of the measurements of the at least one sensor.

[0022] This contribution of the individual sensor measurements can also be determined by calculating the respective Shapley values. Therefore, the root cause analysis can advantageously be carried out in two stages: first at the level of the sub-areas, and then within the at least one critical sub-area with the highest contribution. This can advantageously reduce the computational effort required to identify a critical sensor, since not every single measurement from the sensors of the entire production plant needs to be considered, but only those from the at least one critical sub-area.

[0023] Furthermore, the invention may provide that the method further comprises the following step: - Defining a parameter, wherein the parameter determines how many sub-areas with the highest contribution are examined in more detail following the determination based on the determination of the respective contribution of individual measurements of the at least one sensor.

[0024] For example, the parameter can specify that the three sub-areas with the highest contribution should be examined in more detail subsequently. This allows for the subsequent investigation effort to be tailored to individual needs. In a critical case, for instance, the parameter can be set to a low value to quickly identify the cause of the faulty production. In less critical cases, on the other hand, a more detailed investigation of several sub-areas can be carried out.

[0025] The invention also relates to a computer program, in particular a computer program product, comprising instructions which, when executed by a computer, cause the computer to execute the method according to the invention. Thus, the computer program according to the invention offers the same advantages as those described in detail with reference to a method according to the invention.

[0026] The invention also relates to a data processing device configured to execute the method according to the invention. The device can, for example, be a computer that executes the computer program according to the invention. The computer can have at least one processor for executing the computer program. Alternatively, a non-volatile data storage device can be provided in which the computer program is stored and from which the computer program can be read by the processor for execution.

[0027] The invention may also relate to a computer-readable storage medium which contains the computer program according to the invention and / or includes instructions which, when executed by a computer, cause the computer to execute the method according to the invention. The storage medium is, for example, designed as a data storage device such as a hard drive and / or non-volatile memory and / or a memory card. The storage medium can, for example, be integrated into the computer.

[0028] Furthermore, the method according to the invention can also be implemented as a computer-implemented method.

[0029] Further advantages, features, and details of the invention will become apparent from the following description, in which exemplary embodiments of the invention are described in detail with reference to the drawings. The features mentioned in the claims and in the description can each be essential to the invention individually or in any combination. The drawings show: Fig. 1 a schematic visualization of a method, a device, a storage medium and a computer program according to exemplary embodiments of the invention, Fig. 2 a schematic representation of a production plant with several sub-areas, Fig. 3 a schematic representation of a production plant with several sub-areas, each of which is encoded as a latent representation.

[0030] In Fig. Figure 1 shows a method 100, a device 10, a storage medium 15 and a computer program 20 according to exemplary embodiments of the invention.

[0031] Fig. Figure 1 shows in particular a method 100 for the automated execution of a root cause analysis in a production plant 1. In a first step 101, at least two sub-areas 2 are determined for the root cause analysis based on a grouping of sensors 3, wherein at least one sensor 3 is assigned to each sub-area 2, which is located in the same area of ​​the production plant 1. In a second step 102, measurements of the at least one sensor 3 of each sub-area 2 are encoded into a latent representation 4 using a machine learning model. In a third step 103, a posterior distribution of the latent representation 4 of each respective sub-area 2 is determined for a case in which a defective product is produced by the production plant 1.In a fourth step 104, the respective contribution of each sub-area 2 to the production of the defective product is determined. This involves testing the replacement of each individual sub-area 2 and the replacement of all possible combinations of sub-areas 2 based on the corresponding latent representations 4. For the remaining sub-areas 2, the determined posterior distribution of the latent representation 4 is used. In a fifth step 105, the root cause analysis is carried out in the production plant 1 based on the respective determined contribution of the sub-areas 2.

[0032] Fig. Figure 2 shows a schematic representation of a production plant 1 with several sub-areas 2. The sub-areas 2 can, for example, represent different stations within the production plant 1, such as a drilling station or a quality assessment station. Various sensors 3 are located within the sub-areas 2, providing measurements.

[0033] Fig. Figure 3 shows a schematic representation of a production plant 1, in which the measurements of the sensors 3 in the sub-areas 2 are each encoded as a latent representation 4.

[0034] The method according to exemplary embodiments of the invention utilizes, in particular, a structure of production plants 1, especially conveyor belts in production plants 1, to shift the computational load from estimating the Shapley values ​​to the outlier detection model. The latter is expected to behave more robustly with increasing dimensionality and thus contribute to dividing the entire root-cause analysis (RCA) problem into smaller, more manageable problems in a hierarchical approach.

[0035] In situations where the system under test follows a good structure (e.g., in production facilities 1), the present invention, according to exemplary embodiments, can lead to an increase in performance. In other words, the invention, according to exemplary embodiments, can provide a middle ground between Shapley value estimation at the node level (one extreme) and the model for detecting outliers with very high dimensions (the other extreme).

[0036] The present invention is in particular more robust and less computationally intensive than prior art approaches in certain target systems, for example in assembly lines of production plants 1.

[0037] The present invention can be used, according to exemplary embodiments, for the analysis of data resulting from the acquisition of a sensor 3. The sensor 3 can detect measurements of the environment in the form of sensor signals, which may be provided, for example, by specific data, namely sensor measurements (continuous or categorical) from stations in a production plant 1. Examples of acquired sensor data would be pressure, brightness, force, or similar parameters.

[0038] The present invention can be used, according to exemplary embodiments, to calculate a control signal for the control of a technical system, such as fail-safe operation. Anomalies can be detected, followed by identification of their cause. Furthermore, countermeasures can be derived, for example, by alerting process engineers and / or line planners about the actual cause of the faulty behavior, in order to direct maintenance efforts, prevent recurring failures, reduce the scrap rate early enough, and minimize production downtime.

[0039] For example, some production plants 1 operate with very low reject rates, e.g., <1%, to reduce manufacturing costs or waste and increase annual yield. Errors in the early stages of these production plants 1 can lead to an increased reject rate, for example, up to 5-10%. In possible scenarios, such production plants 1 can be operated with these relatively high reject rates until the source of the error is identified and a time window for downtime and maintenance is available. In other, more extreme scenarios, production may be halted for a longer period until the source of the error is identified and rectified. In both cases, automated cause-and-effect analysis can offer enormous potential for cost minimization, as such sources of error can be identified earlier and more reliably.However, existing automated cause-and-effect analysis approaches scale poorly with the size of the target systems, i.e., for example, the production plants 1. If such target systems have a good structure, which is the case for many production plants 1, an automated cause-and-effect analysis according to embodiments of the invention can benefit from this structure to reduce the computation time and costs required to identify the actual sources of error.

[0040] As part of a regular cause-and-effect analysis, the following procedure is preferably followed. If abnormal behavior of a system is observed (e.g., an assembly line that produces a large number of defective parts, i.e., parts that fail the quality inspection at the end of the line), process engineers and line planners preferably initiate investigations to find the actual source of the error, initiate maintenance work, and prevent a recurrence.

[0041] Cause-and-effect analysis can be automated if data—e.g., sensor measurements from various sub-areas throughout the entire production process—is available. Classical approaches—e.g., sensitivity analysis and feature relevance—correlate observed failures with the measurements.

[0042] Furthermore, there are cause-and-effect analyses based on causal structures. However, correlation-based approaches tend to fail, particularly when searching for relevant, intuitive explanations for outliers (e.g., correlating pressure drop with IDs and counters, or suggesting color brightness as a possible cause of pressure drop). Moreover, such approaches can be prone to error propagation, which can lead to maintenance work being carried out in the wrong places or sub-areas, potentially resulting in recurring errors and consequently higher maintenance costs and downtime.

[0043] Within the scope of the present invention, a hierarchical causal cause-and-effect analysis is described according to exemplary embodiments. Prior art automated cause-and-effect analyses that utilize the causal semantics of the system are based, for example, on quantities such as Shapley values, the estimation of which can be computationally intensive for medium to large graphs. According to the present invention, this limitation can be addressed by utilizing the structure of the existing technical system, such as a conveyor belt, and decomposing the entire cause-and-effect problem into smaller problems.

[0044] A graphical causal model of the system under test can be a graphical representation G=(U, E), which is based in particular on a set of nodes U and edges E. Each node V i For example, it represents a measuring point in the system.

[0045] In this graphical representation, a single leaf node (V) can betarget This represents a key performance indicator for the entire process, also known as End-of-Line (EoL). In the context of production lines, this performance indicator can be a crucial quality assessment test that either directs the manufactured component to the field or indicates the need for scrapping.

[0046] Scrapped parts can represent costly failures in the production process. Therefore, it is preferable to identify the nodes responsible for these failures and the contribution of each node V. i to determine the cause of the observed failure.

[0047] The contribution of the j-th node V j to the failure of V target can be estimated using the Shapley value contribution function: ϕ(j):=1n!∑σC(j|Ictx) =∑I⊆U\{j}1n(|I|n−1)C(j|I)

[0048] Here, ϕ(j) is in particular the valuation function, U the set of all nodes and C(j|J) The marginal contribution function (also called the set function). The number of terms in the above expression can grow rapidly with the system size, since (almost) all possible subsets of the node set are possible. J⊆U\{j} This must be taken into account, which can make an exact estimate of the above size practically impossible for graphs larger than a handful of measurement nodes. Therefore, numerical approximations such as sample-based approaches or model-based estimators are used.

[0049] Causal cause-and-effect analysis approaches rely in particular on two requirements for the graphical model: directed edges and acyclic structures, which are commonly referred to as directed acyclic graphs (DAGs).

[0050] In the present invention, according to exemplary embodiments, the structure of a technical system such as a production line is used to reduce the dimensionality of the cause-and-effect analysis problem (i.e., in particular, the number of nodes in the graphical model). By grouping measurement nodes originating from the same sub-area 2 or the same location, the size of the graph can be reduced while maintaining acyclicity. The aggregated graph preferably remains a directed acyclic graph (DAG), but particularly with multivariate measurement nodes.

[0051] One particular challenge lies in applying a set function to the multivariate nodes of the aggregated graph. The set function preferably quantifies the marginal contribution of a node j to the reward (or, in this context, to the end-of-life failure) within the context of a set of nodes. J. This is achieved in particular by quantifying the achievable reward or loss before and after node j of the group. J accession.

[0052] Removing the contribution of node j from a group of nodes in a causal model is possible, in particular, by reselecting node j from its nominal causal mechanism, taking its parents into account. This can, in a sense, be considered an intervention. The set function can be defined as: C(j|I):=−log Prd(NI∪{j}){g(N)≥g(n)} +log Prd(NI){g(N)≥g(n)}. where the second term in particular represents the logarithmic probability (reward) in determining J The node is fixed, while the first term with node j is also fixed.

[0053] For univariate nodes, defining a node in a graphical model is possible, for example, by resampling, since the empirical distribution provides a very good estimate of the distribution of that node.

[0054] For multivariate nodes, the empirical distribution may require a very large amount of data to effectively represent the true distribution. This problem can be solved by encoding the multivariate observation into a low-dimensional latent representation using a standard distribution family. For example, a variational auto-encoder (VAE) with an isotropic multivariate normal distribution on the latent representation can be used.

[0055] The hierarchical causal RCA algorithm according to exemplary embodiments of the present invention looks, for example, as follows:

[0056] Given a structure DAG G˜=(U˜:={Yi}i=1N,ε˜) of N sub-areas 2, a data set of a regular operation D, an error measurement y fail (a form of Y), an end-of-line (EoL) node y out , and a parameter k. The parameter k expresses, in particular, how many sub-areas 2 should subsequently be examined in more detail.

[0057] For each sub-area 2 i, the following pseudo-algorithm is performed:

[0058] In other words, the Variational Auto-Encoder is trained with normal values, i.e., values ​​in a regular operation of production plant 1, so that in the next step the sub-areas 2 can be replaced with these normal values.

[0059] The posterior error distribution, or posterior distribution, expresses in particular what the individual x-values ​​look like when an error occurs, i.e., when the defective product is produced.

[0060] In the next step, preferably, omitting a specific sub-area 2 is tested, and the number of failures or defective components that would result is determined. Subsequently, omitting all possible combinations of sub-areas 2 is tested in an analogous manner. This is illustrated by the pseudo-code below:

[0061] Let the interventional distribution Prd(J~)(yout) be: Prd(J˜)(yout)=P(yout|{zicnf}i∈J˜,{zifail}i∉J˜)

[0062] Let the set function for subset 2 be j: C(j|J˜)=logPrd(J˜)(yout)Prd(J˜∪{j})(yout)

[0063] Subsequently, error contributions of sub-areas 2 can be calculated based on Shapley values: ϕ(j):=1n!∑σC(j|Ictx) =∑I⊆U\{j}1n(|I|n−1)C(j|I)

[0064] This allows the contribution of each sub-area 2 to the failure to be calculated.

[0065] For the k sub-areas 2 with the highest contribution, Shapley values ​​can then be determined again based on the individual measurements within the sub-areas 2.

[0066] As output, within the scope of the present invention, the k sub-areas 2 with the highest contribution and within the k sub-areas 2 the measuring nodes or sensors 3 with the highest contribution can subsequently be provided.

[0067] The preceding explanation of the embodiments describes the present invention solely by way of examples. Naturally, individual features of the embodiments can be freely combined with one another, provided this is technically feasible, without departing from the scope of the present invention.

Claims

[1] Method (100) for automatically performing a root cause analysis in a production plant (1), comprising the following steps: - Determining (101) at least two sub-areas (2) for root cause analysis based on a grouping of sensors (3), wherein at least one sensor (3) is assigned to a respective sub-area (2) that is located in the same area of ​​the production plant (1), - Encoding (102) measurements of at least one sensor (3) of the respective sub-area (2) into a latent representation (4) using a machine learning model, - Determining (103) a posterior distribution of the latent representation (4) of each respective sub-area (2) for a case in which a defective product is produced by the production plant (1), - Determining (104) a respective contribution of each sub-area (2) to a production of the defective product, whereby a replacement of each individual sub-area (2) and a replacement of all possible combinations of the sub-areas (2) is tested on the basis of the corresponding latent representations (4), while for the remaining sub-areas (2) the determined posterior distribution of the latent representation (4) is used, - Performing (105) the root cause analysis in the production plant (1) based on the respective determined contribution of the sub-areas (2). [2] Method (100) according to claim 1, characterized by , that the procedure includes providing a control signal for controlling the production plant based on a result of the root cause analysis in the production plant (1), wherein the root cause analysis is based on the respective determined contribution of the sub-areas (2). [3] Method according to any one of the preceding claims, characterized by that the machine learning model is a variational auto-encoder and that the encoding is performed using an encoder module of the variational auto-encoder. [4] Method (100) according to any one of the preceding claims, characterized by , that in the context of coding (102) the measurements of the at least one sensor (3) are specific for a regular operation of the production plant (1), so that a compressed representation of the measurements of the at least one sensor (3) in the regular operation of the production plant (1) is provided by the latent representation (4). [5] Method (100) according to any one of the preceding claims, characterized by , that in determining (104) the respective contribution of each sub-area (2) to the production of the defective product Shapley values ​​for the respective sub-areas (2) are calculated. [6] Method (100) according to any one of the preceding claims, characterized by , that performing (105) the root cause analysis includes the following step: - Comparing the respective contributions to identify at least one sub-area (2) with the highest contribution. [7] Method (100) according to claim 6, characterized by , that the procedure (100) further comprises the following step: - Calculating the respective contribution of the measurements of the at least one sensor (3) in the at least one sub-area (2) with the highest contribution, in order to further carry out the root cause analysis in the production plant (1) on the basis of the respective calculated contribution of the measurements of the at least one sensor (3). [8] Method (100) according to claim 7, characterized by , that the procedure (100) further comprises the following step: - Defining a parameter, wherein the parameter determines how many sub-areas (2) with the highest contribution are examined in more detail following the determination (104) based on the determination of the respective contribution of individual measurements of the at least one sensor (3). [9] Computer program (20) comprising instructions which, when the computer program (20) is executed by a computer (10), cause it to execute the method (100) according to any of the preceding claims. [10] Device (10) for data processing which is configured to carry out the method (100) according to any one of claims 1 to 8. [11] Computer-readable storage medium (15) comprising instructions which, when executed by a computer (10), cause it to perform the steps of the method (100) according to any one of claims 1 to 8.

Images