Method for detecting anomalies in process curves related to components using intelligent models

Abstract

Method for detecting anomalies in process curves related to a component using an intelligent model. The intelligent model (11) is located in a control unit (10) comprising a user interface (14) and a synthetic data generation module (16). In step S1, the intelligent model (11) is built and trained using a plurality of process curves (12) related to at least one component (15). In step S2, synthetic data (18) related to the process curves (12) of the at least one component (15) is generated using a hybrid model (11a) present in the synthetic data generation module (16). In step S3, the intelligent model (11) is fine-tuned using the generated synthetic data (18). In step S4, anomalies in the process curves (12) of the at least one component (15) are detected in real-time environment by the fine-tuned intelligent model (11).

Description

Technical Field

[0001] This invention relates to a method for detecting anomalies in process curves associated with at least one component using an intelligent model. Background Technology

[0002] Fastening processes are common and critical on manufacturing / assembly lines. In automotive electronics factories, they are important during control unit manufacturing and housing installation. This process involves multiple steps with user-defined targets and monitoring functions, which log information about the process. If a target is not met, the tool outputs "non-conforming." Given the limited number of engineers on shifts, the likelihood of inspecting multiple stations and identifying "every single non-conforming" is extremely low, leading to downtime and scrap.

[0003] Indian patent application IN202441006642 discloses a control unit for performing tasks related to a component. In operating mode, the control unit monitors and identifies at least one incomplete function related to the component. The control unit manually corrects the at least one incomplete function and stores data related to the at least one incomplete function. The control unit uses the stored data to build and train an intelligent module. The control unit uses the trained intelligent module to perform the component-related tasks in a real-time environment. Attached Figure Description

[0004] Figure 1 An embodiment of the present invention illustrates a control unit for detecting anomalies in a process curve using an intelligent model; and Figure 2 A flowchart of a method for detecting anomalies in a process curve using an intelligent model of a control unit, according to the present invention, is shown. Detailed Implementation

[0005] Figure 1 A control unit for detecting anomalies in process curves using an intelligent model, according to an embodiment of the present invention, is shown. The control unit 10 includes a user interface 14 and a synthetic data generation module 16. The control unit 10 constructs and trains an intelligent model 11 using multiple process curves 12 associated with at least one component 15, and generates synthetic data 18 associated with the process curves 12 of at least one component 15 of the control unit 10 using a hybrid model 11a present in the synthetic data generation module 16. The control unit 10 then fine-tunes the intelligent model 11 using the generated synthetic data 18, and detects anomalies in the process curves 12 of at least one component 15 in a real-time environment using the fine-tuned intelligent model 11.

[0006] The structure of the control unit and the working components of the control unit 10 are explained in further detail. The operation of the control unit 10 and the various components connected to it are explained in detail. The method of deploying in and operating the control unit 10 is not simply a software program itself, but involves a device with multiple hardware modules (intelligent model 12, user interface, hybrid model, synthetic data generation module) and functional types, each hardware module performing a specific task. The control unit 10 is an operational device for processing and storing information received from at least one source.

[0007] The control unit 10 is selected from a group of analysis units, including microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuits, one or more microchips or integrated circuits interconnected using a motherboard, hard-wired logic, software stored in a storage device and executed by the microprocessor, firmware, application-specific integrated circuits (ASICs) and / or field-programmable gate arrays (FPGAs), and / or any component that processes signals based on operating instructions. It should be noted that the intelligent incident analysis unit 10 can be any other type known to those skilled in the art, but is not limited to the devices described above.

[0008] Multiple success process curves 12a and failure process curves 12b are fed as inputs into the intelligent model 11 for building and training the intelligent model 11. The intelligent model 11 is any model with an internally deployed artificial intelligence method. The intelligent model 11 uses any of the network technologies, such as deep learning or machine learning, to find anomalies in the process curves 12. The construction and training of the intelligent model 11 utilizes a large number of success and failure process curves 12 under various operating conditions. For ease of understanding, the process curves 12 are associated with the tightening process of the screw 15 on the surface of the control unit 10.

[0009] Successful process curve 12a relates to the correct tightening process of screw 15 being tightened into the designated hole on control unit 10, while unsuccessful or failed process 12b relates to deviations in the tightening process curve of screw 15 being tightened into the designated hole on the surface of control unit 10. According to one understanding of the invention, process curve 12 is plotted between the rotation angle of screw 15 relative to the hole and the torque generated / applied to screw 15 during the tightening process. However, it should be understood that process curve 12, or any type of graph plotted between two physical parameters, can be considered, but is not limited to the curves described above, and the intelligent model is used to detect anomalies in the corresponding process curve / graph 12.

[0010] The control unit 10 also uses the hybrid model 11a present in the synthetic data generation unit 16 to generate domain-specific synthetic data 18, one type of which is related to the screw tightening process curve. The hybrid model 11a uses physics-based techniques and generates the synthetic data 18 based on multiple model parameters and temperatures (room temperature and material temperature). The hybrid model 11a is constructed in response to anomalous inputs to acquire the process curve 12 as output, and the process curve 12 is used as a synthetic feature to train the classifier 22 of the intelligent model 11.

[0011] When the control unit 10 fails to achieve the required accuracy, the control unit 10 uses an incremental correction model 11b, which employs a neural network to generate synthetic data 18. The control unit 10 detects anomalies in at least one process curve 12 associated with at least one component 15, which occur when at least one component 15 undergoes any of the following processes: fastening process, pressing process, riveting process, or crimping process.

[0012] After detecting an anomaly in process curve 12, control unit 10 provides corrective measures corresponding to the detected anomaly in a real-time environment based on fault tree 20a.

[0013] Figure 2 A flowchart illustrating a method for detecting anomalies in process curves using an intelligent model 11 of a control unit 10 according to the present invention is shown. The intelligent model 11 is located in the control unit 10, which includes a user interface 14 and a synthetic data generation module 16. In step S1, the intelligent model 11 is constructed and trained using multiple process curves 12 associated with at least one component 15. In step S2, synthetic data 18 associated with the process curves 12 of at least one component 15 is generated using a hybrid model 11a present in the synthetic data generation module 16. In step S3, the intelligent model 11 is fine-tuned using the generated synthetic data 18. In step S4, anomalies in the process curves 12 of at least one component 15 are detected in a real-time environment using the fine-tuned intelligent model 11.

[0014] First, an intelligent model 11 is constructed and trained using a large number of process curves 12 (successful process curves and failure / unsuccessful process curves) associated with at least one component 15, one of which is a screw 15, and these process curves 12 are generated during the tightening process of the screw 15. The intelligent model 11 thus constructed and trained is fine-tuned using synthetic data 18. The synthetic data 18 is generated by a hybrid model 11a existing in the synthetic data generation module 16, which employs a physics-based technique and generates the synthetic data 18 using some model parameters. The control unit 10 checks the required accuracy, and if the required accuracy is not achieved, additional data models are used to generate the synthetic data 18. The output of the model 11a is the process curves 12, which are used as features for training the classifier 22 of the intelligent model 11.

[0015] Now, using the trained and fine-tuned intelligent model 11, the control unit 10 receives a process curve 12 associated with at least one component 15 in a real-time environment, processes the process curve 12 to detect any anomalies. When an anomaly is detected, the control unit 10 suggests corrective measures for the corresponding detected anomaly from a pre-stored fault tree 20a (stored in the memory 20 of the control unit 10). The anomaly type is selected from a set of anomalies including screw 15 being screwed into the corresponding hole of the control unit 10 at the wrong angle, excessive torque being applied to the screw 15 during the tightening process, etc.

[0016] The method is illustrated using an example. At least one component is a screw 15, and process curve 12 relates to the tightening process of screw 15. Screw fastening is a detachable connection method used for mechanical fastening and electrical connection. Due to its simple structure, reusability, and clean process, screw fastening is one of the most common processes used in the field of mechatronics for manufacturing electronic control units (ECUs). However, a significant challenge of this operation is that the fastening of each screw connection generates a torque versus screw rotation angle curve. Now, experts in the field can use these curves to understand the behavior of the forming process and its failures and malfunctions.

[0017] Furthermore, due to the inherent characteristics of the working patterns within a manufacturing plant, processes must be monitored, and failures and malfunctions can occur at any time. When any failure occurs, maintenance personnel collect charts and consult experts to identify the problem and its solution. The provided solution is then implemented. All screws involved follow this process chain. Clearly, the analysis, prediction, and diagnosis of this process heavily rely on expert guidance. Moreover, the development process for these screws remains an under-researched area in manufacturing. In this regard, the present invention provides a solution that uses physical and machine learning models (hybrid model 11a or data model 11b) to generate synthetic data 18 to improve the classifier 22 of the intelligent model 11, thereby identifying anomalies.

[0018] The process involves multiple steps. Control unit 10 uses process curves 12 (success and failure scenarios (12a and 12b)) to build and train the intelligent model 11. Control unit 10 uses both successful (OK) and failed (Not OK) process curves (12a and 12b) as input to the intelligent model 11. These curves (12a and 12b) are generated by a machine with a low failure rate, making "failure" data very scarce (less than 1%). The parameters affecting the fastening process include geometry, material properties, and assembly conditions, making it highly complex, and due to the scarcity of "failure" curves 12b, the intelligent model 11 requires fine-tuning.

[0019] The low incidence of "failure" cases leads to data skew, which can cause classifier 22 to become biased. A key challenge here is training the intelligent model 11 without overfitting to "success" cases. Given the difficulty and time-consuming nature of generating realistic outlier data, this invention proposes using hybrid models 11a to generate synthetic data. These models 11a combine physical equations (reflecting the mechanics of the fastening process) with real data. The control unit 10 ensures that the amount of real data required to generate synthetic data 18 is kept to a minimum. This synthetic data 18 helps to balance the skew and improve the generalization ability of classifier 22 under different system settings and material compositions.

[0020] Control unit 10 addresses the data imbalance caused by the extremely low proportion of "failure" scenarios and generates useful synthetic data 18 with minimal real-world data input, especially in complex scenarios where settings, materials, and processes vary. Control unit 10 further ensures that the classifier of intelligent model 11 has good generalization ability under different conditions and settings. Control unit 10 combines data augmentation techniques, domain-driven synthetic data 18, and robust anomaly detection methods to provide synthetic data 18 that highly simulates real-world behavior under various conditions.

[0021] For physical model 11a, control unit 10 can use any constitutive model. For ease of understanding, this invention adopts the Johnson-Cook constitutive model given in the following formula.

[0022] Where A, B, C, m, and n are model parameters. This model 11a is used to predict material behavior, including fatigue and failure. Parameters obtained from the Johnson-Cook model are used to predict the flow behavior of metals, high strain rates, temperature, and deformation of various materials. As shown in the above equation, the flow stress is the plastic strain. Strain rate (using reference strain rate) The normalized value is a function of the material's melting temperature Tm and room temperature Tr. This physics-based model 11a is solved using the finite element method in tools such as Abaqus, and the solution generates a process curve corresponding to the given input.

[0023] After constructing the hybrid model 11a, the control unit 10 obtains process curves 12 for abnormal inputs. These process curves can be used as synthetic features to train the classifier 22 of the intelligent model 11. Some examples of categories of non-abnormal failure scenarios include: oscillations at 360-degree intervals, screwdriver head bending, offset between screw head and screwdriver, incorrect zero-position teaching, incorrect part position teaching, and spindle perpendicularity relative to the ECU / WPC. However, it should be understood that failure scenarios are not limited to the above and can also be any other types known in the prior art.

[0024] If the publicly available model fails to achieve the required accuracy due to the inability to model certain highly complex scenarios, the control unit 10 employs a hybrid model 11a, such as an incremental correction model 11b using neural networks or even generative adversarial networks (GANs), in addition to physics, to correct the simulation results. The control unit 10 deploys the intelligent model 11 in a real-time environment to detect anomalies (failures) based on the process curve 12. When an anomaly is detected, the intelligent model 11 directs to a fault tree 20a, which includes corrective measures compiled based on domain expertise.

[0025] Through the aforementioned process, control unit 10 can recommend parameters, helping users achieve more precise target functions at a step-by-step granularity, and enabling early detection and diagnosis. In this process, hybrid methods will enable the detection of useful patterns and provide actionable recommendations in the field. Prediction can further enrich action information along the value chain. Control unit 10 also uses physics and data to generate synthetic data 18 in situations where there are few failure exceptions and corresponding data is difficult to obtain, and translates and applies the learning outcomes to use cases such as pressing, riveting, and crimping. Control unit 10 performs automation, enabling junior staff / internships to take on more responsibility and freeing up experts to focus on more complex tasks. Control unit 10 is also integrated as a service into diagnostic / predictive platforms.

[0026] It should be understood that the embodiments described above are merely examples and do not limit the scope of the invention. Many such embodiments, as well as other modifications and variations, are contemplated in the specification. The scope of the invention is limited only by the claims.

Claims

1. A method for detecting anomalies in a process curve (12) using an intelligent model (11) present in a control unit (10), the control unit (10) including a user interface (14) and a synthetic data generation module (16). The method includes the following steps: - The intelligent model (11) is constructed and trained using multiple process curves (12) associated with at least one component (15). - Use the hybrid model (11a) present in the synthetic data generation module (16) to generate synthetic data (18) related to the process curve (12) of at least one component (15). - The intelligent model (11) is fine-tuned using the generated synthetic data (18); - Anomalies in the process curve (12) of at least one component (15) are detected in a real-time environment by a finely tuned intelligent model (11).

2. The method as described in claim 1, wherein, When an anomaly is detected in the process curve (12) of at least one component (15), corrective measures are applied from the fault tree (20a) stored in the memory (20).

3. A control unit (10) for detecting anomalies in a process curve (12) using an intelligent model (11), the control unit (10) comprising a user interface (14) and a synthetic data generation module (16), the control unit being configured to: - The intelligent model (11) is constructed and trained using multiple process curves (12) associated with at least one component (15). - Using the physical model (11a) present in the synthetic data generation module (16), generate synthetic data (18) related to the process curve (12) of at least one component (15) of the control unit (10). - The intelligent model (11) is fine-tuned using the generated synthetic data (18); - Anomalies in the process curve (12) of at least one component (15) are detected in a real-time environment by a finely tuned intelligent model (11).

4. The control unit (10) as claimed in claim 3, wherein, Multiple success process curves (12a) and failure process curves (12b) are fed into the intelligent model (11) as inputs for building and training the intelligent model (11).

5. The control unit (10) as claimed in claim 4, wherein, The hybrid model (11a) present in the synthetic data generation unit (16) generates domain-specific synthetic data (18), and one of the domain-specific synthetic data is related to the screw tightening process curve.

6. The control unit (10) as claimed in claim 5, wherein, The hybrid model (11a) generates the synthetic data (18) based on multiple model parameters and temperatures (room temperature and material temperature).

7. The control unit (10) as claimed in claim 6, wherein, The hybrid model (11a) is constructed in response to the abnormal input to obtain the process curve (12) as the output, and the process curve (12) is used as a synthetic feature to train the classifier (22) of the intelligent model (11).

8. The control unit (10) as claimed in claim 7, wherein, When the hybrid model (11a) does not meet the required accuracy, the synthetic data (18) is generated using an incremental correction model (11b) with a neural network.

9. The control unit (10) as claimed in claim 3, wherein, When the at least one component (15) undergoes any of the processes including fastening, pressing, riveting, and curling, the control unit (10) detects an anomaly in at least one process curve (12) associated with the at least one component (15).

10. The control unit (10) as claimed in claim 9, wherein, The control unit (10) is adapted to provide corrective measures corresponding to detected anomalies in a real-time environment based on a fault tree (20a).

Images