Computer-implemented method for finding cause of fault in industrial process
By calculating first-order and second-order correlation values, and combining deep learning algorithms and converter architecture, the problem of difficulty in identifying the root cause of failures in complex industrial processes is solved, achieving efficient and accurate failure cause analysis, which is applicable to multiple application fields.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-10
- Publication Date
- 2026-04-10
AI Technical Summary
Existing technologies struggle to effectively identify the root causes of failures in complex industrial processes, especially when the types of failures are diverse and the systems are complex. Traditional methods such as fault tree analysis and feature attribution algorithms are unable to trace chain causal relationships.
A computer-implemented method is adopted to identify the cause of failure by calculating first-order and second-order correlation values. Deep learning algorithms such as DeepSHAP, Deeplift, and integral gradient are used, combined with a converter architecture and attention mask, to analyze the correlation of data points in industrial processes, trace causal relationships, and identify the root cause.
It can simplify the process of finding the cause of failure by understanding a partial causal sequence without requiring a complete cause-effect graph, thereby improving the accuracy and efficiency of fault identification. It is applicable to fields such as industrial process control, root cause analysis, camouflaged IT, security, and telecommunications.
Smart Images
Figure CN121834729A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a computer-implemented method for finding the cause of failures in industrial processes. Background Technology
[0002] Finding the root cause of failures in complex industrial processes or facilities is a significant challenge involving multiple aspects. First, these systems are often highly technically complex. Industrial processes for manufacturing and processing workpieces consist of numerous interconnected components, machines, and control systems, all of which must work seamlessly together to ensure the final product is fault-free. Each individual part of this chain can potentially be a source of failure, making it difficult to identify the true cause of the problem.
[0003] Another factor is the wide variety of possible failure types. These failures can be mechanical, electronic, or software-related. Mechanical failures may be caused by wear and tear or malfunction of components, while electronic failures are usually attributed to problems in wiring, sensing devices, or control electronics. Software-related failures may arise from bugs or incompatibilities within the process control software. These different failure sources must be analyzed systematically and according to certain methodologies in order to determine the true root cause of the problem.
[0004] Reducing waste and minimizing failures are key objectives in industrial production to lower costs and increase efficiency. To achieve these goals, companies must utilize precise diagnostic tools and methods. This typically includes advanced measurement and monitoring technologies, such as sensors that continuously provide status data for machines and processes. However, this data not only needs to be detected but also correctly interpreted, requiring specialized analysis and simulation software.
[0005] Furthermore, human factors play a significant role. The expertise and experience of engineers and technicians are crucial for the accurate diagnosis and elimination of faults. They must be able to identify complex data patterns and understand which parameters are affected when anomalies occur. This requires not only deep technical knowledge but also strong problem-solving skills and often interdisciplinary thinking.
[0006] Finally, external factors such as material quality or environmental conditions can also influence the formation of failures. Changes in material composition or temperature fluctuations in the production environment can lead to unforeseen problems that are difficult to diagnose.
[0007] Overall, troubleshooting in complex industrial processes is a demanding and multi-dimensional process that requires a combination of advanced technology, deep expertise, and systematic methods to ensure production efficiency and minimize waste.
[0008] There are various methods for finding faults in industrial processes, and these methods vary depending on the specific application and available technology. One of the most basic methods is Fault Tree Analysis (FTA), which systematically investigates the causes of failures by breaking them down into hierarchical layers. This method is typically supported by historical data and empirical values and requires a deep understanding of the process and system.
[0009] More advanced methods involve real-time data monitoring and analysis. This involves using numerous sensors that continuously monitor parameters such as temperature, pressure, vibration, and other relevant process variables. The collected data is then analyzed using algorithms to identify anomalies. These algorithms are typically based on statistical models or machine learning and can identify patterns and deviations that indicate potential failures.
[0010] A significant advancement in this field is the use of feature attribution algorithms. These algorithms, such as Shapley values or DeepSHAP, are designed to explain the decision-making processes of complex models, such as deep neural networks. DeepSHAP combines Shapley values from cooperative game theory with deep learning models to quantify the contribution of each individual feature to the model's predictions. In industrial scenarios, this means a better understanding of which variables and process parameters contribute most to a particular fault or anomaly. This enables more targeted troubleshooting and mitigation.
[0011] DeepSHAP is particularly well-suited for analyzing large volumes of data generated by multiple sensors in modern industrial facilities. By applying DeepSHAP, the complex relationships and interactions between different process parameters can be better understood, thereby identifying the most likely factors to cause failure. This is especially valuable in highly complex processes where causal relationships are not always obvious.
[0012] Another approach is to use predictive maintenance, which employs machine learning to predict the condition of machines and facilities. By analyzing historical data, the system can learn when and under what conditions a failure might occur and proactively plan maintenance actions before an actual failure happens. This not only reduces downtime but also minimizes scrap rates and optimizes resource utilization.
[0013] Using simple "feature attribution" methods, especially DeepSHAP, Deeplift, integral gradient, etc., it is rare to find the root cause of failures in industrial processes; instead, it can only find first-order causal relationships. Summary of the Invention
[0014] Therefore, the purpose of this invention is to propose a method that can not only trace the first-order causal relationships between variables in an industrial process, but also identify the root causes of problems that occur in the process.
[0015] This objective is achieved through the subject matter of the independent claims.
[0016] According to a first aspect of the invention, this objective is achieved by a computer-implemented method for finding the cause of a failure in an industrial process. The method includes several steps. First, a dataset about the industrial process is provided, wherein the data points include at least one failure data point about a faulty process step, and wherein these data points are measurements and / or manufacturing parameters, wherein a measurement sequence is assigned to a component within the industrial process.
[0017] Each data point can be assigned to a sensor's measured value or a setting parameter of a part or component of a system that performs the industrial process.
[0018] An industrial process can be any commercially applicable process that uses at least one machine. Within the scope of this invention, a system refers to at least one machine, preferably a combination of multiple machines. A machine can perform all possible industrial processes. Industrial processes can include, but are not limited to: conveying, heating, cooling, cutting, drilling, milling, grinding, stamping, casting and injection molding, assembly, welding, painting, packaging, positioning, mixing, etc. Therefore, an industrial process may require one or more machines and / or may be distributed across multiple manufacturing facilities.
[0019] A fault data point is a data point in which at least one variable deviates from the tolerance range, such that the process can be classified as faulty. For example, if the force of the drill bit is monitored during drilling of a workpiece, this force should be within a certain tolerance range during drilling. If the force exceeds the limit of this range, the process can be classified as faulty, and the finished workpiece may also be faulty or even unusable. The first data point whose value exceeds the limit of the tolerance range may be used as a fault data point.
[0020] Which parameters to monitor in an industrial process and which data point to consider a failure point largely depends on the industrial process in which the cause of the failure is to be found.
[0021] In the next step, a first-order correlation value is determined for each data point from the first data point group, wherein the data points in the first data point group precede the faulty data points in causality.
[0022] The first group of data points may include, for example, all data points in the dataset that constitute the cause of the faulty data points. Other data points need not be considered because they are unlikely to be the cause of the faulty data points. In this context, causality can be considered as "earlier in time." However, not all earlier data points automatically constitute the cause of subsequent data points.
[0023] Within the scope of this invention, multiple correlation values are calculated. A first-order correlation value indicates a measure of the correlation between a first data point and a second data point. The second data point is a data point that depends causally on the first data point. For a first-order correlation value, this is the faulty data point. Known algorithms such as deepSHAP, deeplift, integral gradient, and Shapley value can be used to determine the first-order correlation value.
[0024] However, first-order correlation values do not depict chain causal relationships. For example, a first data point A might be highly correlated with a second data point B; therefore, second data point B is also highly correlated with the faulty data point. If only the first-order correlation value of the faulty data point is considered, data point B would be highly correlated with the faulty data point, but the correlation between data point A and the faulty data point would be low. Thus, chain causal relationships are not taken into account, making it significantly more difficult to find the cause of the fault. To address this issue, second-order correlation values are determined.
[0025] For example, consider four data points A, B, C, and D. D should be the fault data point. Data points A, B, and C form the first data point group. For this first data point group, the first-order correlation, i.e., the correlation of D, needs to be determined. Here, methods such as Shapley value, DeppShap, DeepLift, and integral gradient can be used. The sum of the first-order correlation values equals the predicted or expected value of the fault measurement calculated based on A, B, and C. For this purpose, normalization of the correlation values can be used.
[0026] Here, the predicted value, or the prediction of D, is a parameter calculated based on A, B, and C (i.e., based on causal precursors) and is intended to approximate the true measured value of D as closely as possible. Such a predicted value can be calculated using a mathematical model (if known) or its corresponding prediction from a machine learning method trained on historical data from the facility. The predicted value of data point D is referred to below as Pred. D For all data points i that need to partially propagate the correlation back to their causal predecessors, a predictive model Pred is required. i .
[0027] After calculating the first-order correlation value of D (i.e., R), 1,D,iAfter determining which data points (i = A, B, C) are correlated with the faulty data point D, we can initially identify which data points are related to it. Each of the measured data points A, B, and C may be partially influenced again by its own causal predecessor. Therefore, we calculate the second-order correlation value of D by reassigning the first-order correlation values of each data point A, B, and C to its own causal predecessor.
[0028] Therefore, the first-order correlation value of C is calculated, that is, the correlation between A and B and the predicted value Pred is determined. C The correlation between C and D. Now, the first-order correlation between C and D, i.e., R. 1,D,C , can be partially reallocated to the causal precursors of C. For this purpose, the first-order correlation value of C (i.e., R0) can be used. 1,C,i ), divided by the measured value of C (i.e., X) C Multiply by the first-order correlation value of C and D (i.e., R) 1,D,C ), and the first-order correlation value between the causal precursors of C (where i=A, B) and D (i.e., R) 1,D,i Add them together: , Where i is a causal predecessor of C, that is, A and B in this simple example.
[0029] The second-order correlation value between C and D is calculated by subtracting the correlation of the predecessors that have been reallocated to themselves from the first-order correlation between C and D, i.e., R. 2,D,C : In one implementation, this method can also be applied to the causal precursors of C to obtain higher-order correlation values for D. For example, the second-order correlation value of B with D can be partially redistributed to its own causal precursor (in this case, only A). Here, the method is similar to that for second-order correlation values: For all i and Therefore, it is quite common for a higher k-order correlation value for a data point d to be determined by redistributing the k-1 order correlation value between l and d to the causal predecessor of the data point l and correcting the correlation value of the data point l itself according to this redistribution.
[0030] When reallocating correlation values, if the absolute value of the measurement is close to 0 in the step of dividing by the measurement of the data point, the reallocated value should be set to 0 to avoid numerical problems.
[0031] If the data point does not have a causal predecessor to which the correlation value can be reassigned (i.e., data point A in this case), the method can be terminated.
[0032] Finally, based on the highest-order correlation value between each data point and D, it can be determined which data points are the cause of the faulty data points.
[0033] Compared to outlier cause analysis based on causal structure, a complete causal graph is not required. Instead, understanding a portion of the process—the causal order of the variables—is sufficient. This order is usually readily available by ranking the variables according to their occurrence over time. This eliminates complex calculations and simplifies the search for the causes of failures.
[0034] Thus, the present invention achieves its objective.
[0035] This invention can be applied in various fields. One of these fields is industrial process control, particularly in the production processes of pharmaceuticals, chemicals, and so on. Here, the relevant variables can include control parameters of the production process and measurements from the production process detected during each production step, such as temperature, pressure, acidity, etc.
[0036] Other application areas include root cause analysis, fake IT, security, and telecommunications.
[0037] This invention can also be applied in non-commercial fields, such as in diagnostics for identifying the causes of diseases, especially in determining the causes of diseases, in the natural sciences for experimental analysis, or in business management issues.
[0038] In one embodiment of the invention, the average value of each data point, respectively derived from a known fault-free component or batch, is subtracted from each data point. This also applies to the data used to train the prediction model Pred. i The proposed method uses historical training data to perform this correction on the data points. Therefore, it examines the correlation between the proposed method and the deviation from the desired behavior.
[0039] In one implementation, a prediction or classification algorithm is used to determine first-order reference values and / or second-order reference values.
[0040] The prediction algorithm is trained to predict values based on historical data. The output of such an algorithm typically includes values between 0 and 1, which are then scaled to a target range.
[0041] The classification algorithms work differently. This classification algorithm generates probability values for each class in a set of classes. These values total 1 as long as the remaining classes are available.
[0042] Depending on the problem or the industrial process to be analyzed, it is advantageous to use predictive or classification algorithms.
[0043] In one implementation, the prediction algorithm Pred i This includes neural network or converter architectures, augmented decision tree architectures, or other prediction methods trained on historical data using machine learning.
[0044] A neural network is an artificial computational structure inspired by the workings of the human brain. It consists of a hierarchical structure of neurons located in layers such as input layers, hidden layers, and output layers. Each neuron receives input, which is weighted according to connection strength, and then executes an activation function to generate an output. During training, the network's weights are iteratively adjusted by subjecting it to large amounts of data and minimizing the difference between predictions and actual results. In this way, the network learns to recognize complex patterns and relationships in the data and can subsequently react to new inputs and make predictions.
[0045] Transformer architectures, or simply transformers, are specific neural network architectures based on attention mechanisms and designed to process sequences, such as text, or time-ordered data points within the scope of this invention. These transformer architectures have proven to be extremely powerful in machine data processing.
[0046] The main idea behind the converter is that it is entirely based on an attention mechanism to model the relationships between data points in a sequence. Traditional sequence models, such as LSTM or RNN, struggle to efficiently detect long-term dependencies. The converter, however, can directly access the entire input sequence and model the relationships between data points in parallel, significantly improving model performance.
[0047] The converter architecture consists of stacked identical layers. Each layer is divided into two main components: a multi-head self-attention mechanism and a position-wise feedforward network.
[0048] The multi-head self-attention mechanism is at the heart of the transformer. Here, the input sequence is transformed into queries, keys, and values. For each data point, a so-called attention weight is calculated, which indicates how relevant other data points in the sequence are to that data point. Then, each data point can be represented by a weighted combination of the values of these other data points. Here, an embedding vector is generated.
[0049] Attention weights are computed multiple times in the so-called "head" to generate different representations of the attention context. The results are then concatenated and recombined through a linear transformation.
[0050] The position-aware feedforward network comprises two linear transformations, which are applied independently to each data point. Since each vector is processed separately, it is the non-sequential part of the architecture.
[0051] Since the converter architecture lacks built-in information about the causal order of data points, positional information is added through positional encoding. These positional codes are added to the embedding vector to account for the influence of position.
[0052] The converter architecture can be divided into an encoder and a decoder. The encoder processes the input sequence, and the decoder generates the output sequence.
[0053] Overall, the transformer architecture is advantageously capable of processing time-ordered data points in a highly parallel and efficient manner, which improves performance when searching for relationships between data points in the dataset. By using a self-attention mechanism, the model can selectively access relevant information and identify complex dependencies within the input sequence.
[0054] In one implementation, the prediction or classification algorithm uses an attention mask.
[0055] Attention masks can force the algorithm (especially the transformer) to use only causal predecessor values for value prediction.
[0056] In one implementation, the data points form a time series of one or more process parameters and / or process measurements.
[0057] The advantage of time series datasets over other datasets lies in their inherent temporal order. Any preprocessing steps, such as sorting data points, are eliminated. Furthermore, data points in a time series dataset share the same dimensionality. This consistency simplifies data point processing because it eliminates the need for parameter weighting.
[0058] In one implementation, each data point comprises a vector consisting of one or more process parameters and / or process measurements.
[0059] In this implementation, data points can have different structures. Here, data points are not limited to parameters or measurements. A single data point can include values of multiple process parameters and / or process measurements. Which process parameters and / or process measurements are specifically included depends on the corresponding process being analyzed. For example, a portion of these data points could describe the welding process, depicting the temperature at the weld point and locations spaced apart from it. After welding, the surface hardness of the workpiece is checked, causing subsequent data points to include a "hardness" parameter. Although the parameters and measurements of the data points differ, they are analyzed together to determine whether the root cause of the failure lies in the welding or in the hardness, and thus in the material structure.
[0060] In one implementation, these data points are based on defined dependencies.
[0061] If the dependencies between the various process steps and the data points therein are known, this knowledge can be used to further simplify the method. This is especially true when these data points are assigned to process steps with no causal relationship, thereby simplifying and accelerating the execution of the method.
[0062] For example, a workpiece should be assembled from two components, A and B. The process analyzed includes the manufacturing of the components and their assembly. Components A and B are manufactured independently of each other. In this case, when analyzing the process, it can be considered that A and B are manufactured independently and thus without causal relationship. If the correlation value of the process stage of manufacturing one of components A and B is determined, the data points describing the manufacturing of the corresponding other component can be ignored. This reduces the actual computational cost of the proposed method.
[0063] In another implementation, the algorithm can be applied to calculate the correlation value of the difference between a data point and one or more qualified data points. To do this, the difference between the measurement sequence of the component and one or more measurement sequences of known qualified components is calculated, and the algorithm operates based on this difference measurement sequence. For prediction, the difference between the prediction of the data points for the component and the prediction of the equivalent data points for qualified components is calculated.
[0064] In another aspect, the present invention relates to a computer program having program code so as to perform the method described above when the computer program is executed on a computer.
[0065] In another aspect, the present invention relates to a computer-readable data carrier having program code of a computer program so as to perform the method described above when the computer program is executed on a computer.
[0066] In another aspect, the present invention relates to a system for finding the cause of a failure in an industrial process, wherein the system is designed to perform the method described above.
[0067] In summary, it should be emphasized that the present invention provides a method for finding the cause of a failure in an industrial process, a corresponding computer program, a computer-readable data carrier having the computer program, and a system for performing the method.
[0068] The described design and extension schemes can be combined with each other arbitrarily.
[0069] Other possible design, extension, and implementation schemes of the present invention also include combinations of features not explicitly mentioned in the preceding or hereinafter described in relation to the embodiments of the present invention. Attached Figure Description
[0070] The accompanying drawings are intended to provide a further understanding of embodiments of the invention. These drawings illustrate the embodiments and, in conjunction with the description, serve to explain the principles and design of the invention.
[0071] Other embodiments and several of the advantages mentioned will be apparent from these accompanying drawings. The elements presented in these drawings are not necessarily shown to scale with each other.
[0072] in: Figure 1 This illustration shows how data points are divided into groups. Figure 2 This illustration schematically shows the correlation of data points within an exemplary dataset; and Figure 3 The flowchart of the method according to the embodiment is illustrated schematically.
[0073] In these accompanying drawings, unless otherwise stated, the same reference numerals denote the same or functionally identical elements, parts or components. Detailed Implementation
[0074] Figure 1 An exemplary and highly simplified dataset 10 is illustrated schematically. Datasets from real industrial processes can include hundreds or thousands of data points. To explain the principles behind the invention, a dataset with six data points A through F is shown. Here, F is a fault data point indicating a fault state. It is necessary to determine which process steps caused the fault. To do this, the correlation between the preceding data points A through E and the fault data point F is analyzed.
[0075] Data points A through E form the first data point group 12 because these data points precede the faulty data point F in time and, in principle, may constitute the cause of the faulty data point. For each data point in the first data point group 12, a first-order correlation value R1 is determined. This first-order correlation value indicates the correlation between the data point and the faulty data point F. This is in Figure 2 The diagram illustrates this. However, the first-order correlation value R1 only indicates the direct correlation between the data point and the faulty data point. To determine the indirect correlation, a further correlation value, namely the second-order correlation value R2, is determined.
[0076] For each data point in the first data point group 12, a second-order correlation value R2 is determined. That is, for example, for data point E, a second data point group 14 is formed, wherein for each data point in this data point group 14, a second-order correlation value R2 is determined. The second data point group 14 for data point E may include one or more of the preceding data points A to D.
[0077] For the remaining data points in the first data point group 12, these steps are repeated to form another second data point group 16 for data point D. This second data point group 16 includes data points A through C. This process is repeated until only data point A remains. Since there are no other data points preceding this data point, it is not necessary to generate a second-order correlation value for it.
[0078] Figure 3 The flowchart of the method according to the embodiment is illustrated schematically.
[0079] In step S10, a dataset of the industrial process to be analyzed is provided. The dataset includes at least one fault data point located at the end of the dataset in time.
[0080] In step S12, a first-order correlation value is determined for each data point that is not a faulty data point. Then, in step S14, a second-order correlation value is determined for each data point. These second-order correlation values reflect the indirect correlation between the preceding data points and the faulty data points.
[0081] Second-order correlations can be determined by decomposing the first-order correlation values of relevant data points into their own contribution and backpropagation components. The own contribution describes the direct correlation between the measured value or parameter and the faulty data point, and is the portion of the measured value or parameter's influence that cannot be explained by its causal precursors. The backpropagation component is the portion of the measured value that can be explained by its causal precursors.
[0082] Thus, in step S14, for each preceding data point, the second-order correlation value of the corresponding data point whose first-order correlation value has been partially reallocated is determined. This process is repeated for all data points preceding the faulty data point. The first data point is an exception in time because it has no indirect dependency.
[0083] Now, in step S16, using first-order and second-order correlation values, zero-order correlation values can be indicated for all data points. These zero-order correlation values, in combination, indicate the indirect and direct correlations between these data points and the faulty data points. Finally, in step S18, the cause of the fault is determined based on these zero-order correlation values. The highest zero-order correlation value indicates the degree of contribution of the corresponding data point to the fault in the faulty data point.
Claims
1. A computer-implemented method for finding the cause of failures in industrial processes. in, The method includes the following steps: - Provide a dataset (10, S10) for the industrial process, wherein the dataset (10) includes at least one measurement sequence with multiple data points, wherein the data points include at least one fault data point for a faulty process step, and wherein the data points are measurements and / or manufacturing parameters, wherein the measurement sequence is assigned to a component within the industrial process; - For each data point from the first data point group (12), a first-order correlation value (R1, S12) is determined, wherein the data points of the first data point group (12) have a causal relationship with the fault data point; - For at least one first data point in the first data point group (12): determine the second-order correlation value (R2, S14) by partially reallocating the first-order correlation value (R1) of each first data point to data points that are causally related to the corresponding first data point; and - Use the second-order correlation value (R2) to determine the cause of failure in the industrial process (S18).
2. The method according to claim 1, wherein, The method further includes: - Based on the k-1 order correlation value and the first order correlation value, the k-order correlation value of the fault data point is determined, and the k-order correlation value is also used to determine the cause of the fault.
3. The method according to any one of the preceding claims, wherein, Use prediction or classification algorithms to determine the first-order correlation value (R1) and / or the second-order correlation value (R2).
4. The method according to claim 3, wherein, The prediction or classification algorithm includes neural networks, converter architectures, or augmented decision tree architectures.
5. The method according to any one of claims 3 or 4, wherein, The prediction or classification algorithm uses attention masks.
6. The method according to any one of the preceding claims, wherein, The data points form a time series of one or more process parameters and / or process measurements.
7. The method according to any one of the preceding claims, wherein, Each data point consists of a vector composed of one or more process parameters and / or process measurements.
8. The method according to any one of the preceding claims, wherein, The data points are based on defined dependencies.
9. A computer program having program code for performing the method according to any one of claims 1 to 8 when the computer program is executed on a computer.
10. A computer-readable data carrier having program code of a computer program so as to perform the method according to any one of claims 1 to 8 when the computer program is executed on a computer.
11. A system for finding the cause of failures in an industrial process, wherein, The system is designed to perform the method according to any one of claims 1 to 8.