METHOD AND SYSTEM FOR QUALITY CONTROL IN INDUSTRIAL MANUFACTURING

Description

[0001] The invention relates to a method and system for quality control in industrial production for one or more production processes for manufacturing at least one workpiece and / or product.

[0002] Monitoring the quality of ongoing production to ensure consistent product quality for customers is crucial to the success of any manufacturer. Deficiencies in quality control are therefore not only costly, they can also permanently damage a manufacturer's reputation. However, especially in mass production, it is virtually impossible to perform comprehensive quality control for every product delivered to the customer.

[0003] Quality control is therefore often carried out at two levels. At the first level, a rather superficial quality control is often carried out by the machine operator himself in order to cover the entire production run, but with insufficient accuracy. At the second level, a detailed quality control is carried out by trained personnel in order to achieve a high level of accuracy, but only for selected individual pieces in a production run. The number of individual pieces inspected and the timing of an inspection are often based on empirical values ​​for a product, the machines used and the materials used. Although this approach has been a proven practice for decades, it has shortcomings, particularly because the inspection at the second level only produces knowledge and a wealth of experience about the small selection of individual pieces inspected.However, it is assumed that defects identified in a selection of individual pieces will be found in the entire production run since the last quality control was performed. Therefore, if defects are identified in a selection of individual pieces, either all individual pieces in a production run are examined or, which is often more cost-effective, the entire production run is discarded as scrap.

[0004] Furthermore, in-depth quality control at the second level often only takes place after customer complaints, making it reactive. Furthermore, the current quality control system is based on the assumption that defects identified in a single piece manufactured in a mass production facility apply to the entire production run and not just to that single piece. Furthermore, quality control itself is not uniform in quality, as it often depends on the training level and daily condition of the personnel deployed.

[0005] Modern machine tools are equipped with a multitude of sensors that monitor the machine and processing status during production. There is also a growing trend toward storing and processing measured data to develop new optimization opportunities and a deeper understanding of the underlying process steps through in-depth data analysis. Well-trained experts, thanks to their in-depth understanding of a particular domain and the processes taking place there, and the use of new software tools, are able to identify details in the production process of a product and / or workpiece in a way that was previously impossible.With regard to quality control, these experts can use the data and detailed insights into the mechanisms of the production processes to develop tailor-made software-controlled solutions that detect defects in individual workpieces in near real time and / or optimize the respective individual process and / or the entire process. However, such customized software solutions are often very complex and therefore costly and time-consuming, and frequently require the use of optical sensors, which are often rather expensive. On the other hand, there is a strong need to improve quality control in production plants through automation technologies.

[0006] US 2018 / 293721 A1 discloses a system having an imaging subsystem including an energy source and a detector. The energy source is configured to generate energy directed toward a sample, and the detector is configured to detect energy from the sample and generate images responsive to the detected energy. The components include a first learning-based model for generating simulated contours for the patterns based on a pattern for the sample input by the computer subsystems to the first learning-based model. The second learning-based model generates actual contours for the patterns in the acquired images of the patterns formed on the sample, which are input by the computer subsystems to the second learning-based model.The computer subsystems are configured to compare the actual contours with the simulated contours and detect defects in the patterns formed on the sample based on the results of the comparison.

[0007] US 2017 / 277174 A1 discloses a processing machine system including a processing machine; a numerical control device that generates a command for driving an axis of the processing machine; an interior space information acquisition unit that acquires interior space information of the numerical control device; a preliminary determination unit that determines acceptance / rejection of a workpiece processed by the processing machine based on a comparison result between the interior space information and a threshold value; a final determination unit that determines acceptance / rejection of workpieces that are targets of accuracy inspection, including at least one workpiece determined as accepted / rejected by the preliminary determination unit based on a measurement result regarding an accuracy of the workpiece;and a threshold updating unit that updates the threshold used for determination processing by the preliminary determination unit based on determination results by the preliminary determination unit and the final determination unit;

[0008] Scalable, automated defect detection during the manufacture / production of a workpiece and / or product is the first step toward fully automated, data-driven quality control in factories and production plants. However, it is often not possible to collect sufficient data to represent the entire diversity of machining processes used to manufacture workpieces and products, as the collected data is only very specific to the respective production site, the machine type used, and the respective NC control system. One solution is to develop customized IT solutions for each possible scenario. These solutions include collecting and labeling the data for each process step, as well as a testing phase and a failure phase, until the developed IT solutions deliver satisfactory results. However, such an approach is very time-consuming and costly.Furthermore, it can be assumed that a mature manufacturing process provides a large quantity of workpieces of sufficient quality and that defective workpieces are the exception rather than the rule.

[0009] The object underlying the invention is to provide a method and a system for quality control, in particular in industrial manufacturing and production, for one or more production processes for manufacturing at least one workpiece and / or product, which is characterized by high reliability and safety and is also cost-effective in the implementation phase and during production operation.

[0010] This object is achieved according to the invention with regard to a method by the features of patent claim 1, and with regard to a system by the features of patent claim 8. The further claims relate to preferred embodiments of the invention.

[0011] According to a first aspect, the invention relates to a method for quality control according to claim 1.

[0012] In an advantageous further development, the learning model is designed as a deep neural network, since only limited systematic solution knowledge is available for the detection of defective workpieces and a deep neural network with several layers is suitable for processing a large amount of partly imprecise input information into a concrete result.

[0013] Advantageously, the meta-learning algorithm is designed as an agnostic meta-learning algorithm that trains the learning model using a gradient method.

[0014] In a further embodiment, a standardized mean value of a specified measured value for a workpiece and / or product and / or production process is defined as the basis for calculating a deviation.

[0015] Advantageously, the currently generated data has a data label. This allows the currently generated data to be used for supervised training of a learning model.

[0016] In a further embodiment, the sensors use non-optical methods to generate data.

[0017] Advantageously, an anomaly is detected when the deviation exceeds or falls below a specified limit.

[0018] According to a second aspect, the invention relates to a quality control system according to claim 8.

[0019] In an advantageous further development, the learning model is designed as a deep neural network.

[0020] Advantageously, the meta-learning algorithm is designed as an agnostic meta-learning algorithm that trains the learning model using a gradient method.

[0021] In an advantageous embodiment, a standardized mean value of a specified measured value for a workpiece and / or product and / or production process is defined as the basis for calculating a deviation.

[0022] In a further embodiment, the currently generated data has a data identifier (label).

[0023] Advantageously, the sensors use non-optical methods to generate data.

[0024] In a further embodiment, an anomaly is detected if the deviation exceeds or falls below a specified limit.

[0025] According to a third aspect, the invention relates to a computer program product which contains one and / or more executable computer codes for carrying out the method according to one or more of claims 1 to 7.

[0026] The invention is explained in more detail below with reference to embodiments shown in the drawing.

[0027] Showing: Figure 1 shows a block diagram for explaining an embodiment detail of the system according to the invention; Figure 2 shows a flowchart for explaining a method according to the invention; and Figure 3 shows a schematic representation of a computer program product according to the invention.

[0028] Additional features, aspects and advantages of the invention or its embodiments will become apparent from the detailed description taken in conjunction with the claims.

[0029] According to the present invention, a deep learning model 100 is used to detect defects in workpieces in a production series. An example of such a learning model 100 is a convolutional neural network. A deep learning model 100 refers to a class of optimization methods for artificial neural networks that have multiple "hidden layers" between the input layer and the output layer, thus exhibiting a comprehensive internal structure. Such a learning model 100 is optimized for rapid adaptation to a new target task, even when only limited input data from a normal manufacturing and / or machining process is available. The learning model can be programmed using frameworks such as TensorFlow or PyTorch.

[0030] According to the invention, an agnostic meta-learning algorithm 200 is used to train and calibrate the learning model 100, as described, for example, by Chelsea Finn et al. (Chelsea Finn, Pieter Abbeel, Sergey Levine: Model-Agnostic Meta-Learning for Fast Adaption of Deep Networks, July 18, 2017). The agnostic meta-learning algorithm 200 is capable of training the learning model 100 using a gradient method to initialize the parameterized deep learning model 100.

[0031] The following describes the data preparation, model training, and model application for a use case in more detail. The detection of anomalies and / or defects during the manufacturing and / or processing of workpieces and / or other products is based exclusively on sensor signal data generated by data acquisition devices and sensors in a production facility.

[0032] By training with the agnostic meta-learning algorithm 200, the deep learning model 100 is able to adapt to a new manufacturing process after the training phase, using only data from workpieces of sufficiently good quality. Ideally, the learning model 100 achieves the necessary adaptation to a new manufacturing process flow using the data generated during the test runs during the calibration phase.

[0033] According to the present invention, the learning model 100 is calibrated only on the basis of data on normal states of a workpiece and / or product (normal state data) and in the context of processing preferably non-visual sensor data for the purpose of anomaly detection in manufacturing processes.

[0034] The data preparation for training the learning model 100 is explained in more detail below. To train the learning model 100, a sufficient number of sensors is required in a production facility. For example, these sensors can measure the torques of the various axes on a milling machine and control deviations. Furthermore, data from a sufficient number of workpieces and machining processes is required. Furthermore, a data label should be provided to mark data indicating anomalies.

[0035] Preferably, an expert will label the data with data identifiers and a corresponding note. One example is torque measurements of a milling spindle. For example, the control deviations of the Z-axis of 100 workpieces and eight different machining processes are examined. The machining processes include various rough machining processes, such as cutting a pocket into the workpiece, and finishing processes such as smoothing the workpiece's upper surface.

[0036] As a preprocessing step of the method according to the invention, the various data signals generated during each process step are recorded on a scale according to their deviation from a specified standard value and thus scaled. For some data signals, it may also be useful to subtract the data value of the data signal of the representative data for normal processing (standard processing) from the newly generated data signals in order to record the deviation more clearly. This is particularly useful when the value of the deviation is small compared to the value of the data signal. Thus, the deviations from a standardized mean value of a specified measured variable for a workpiece and / or product are recorded.

[0037] The learning model 100 is trained using the agnostic meta-learning algorithm 200. The algorithm 200 for training the model can be represented as pseudocode as follows: 1. Prepare the data. 2. Randomly initialize the learning model parameters θ. 3. as long as not converging 4. Extracting a set of n processes P i ~ p (P) 5. Für every P i is to perform: 6. Extract k training examples for normal machining behavior from process P i . 7. Evaluate the anomaly detection performance of the learning model using these examples. 8. Update the learning model parameters θ with respect to θ' using a gradient descent method. 9. End for 10. Extract nxk training examples for anomaly detection from the extracted set of n processes. 11. Update the learning model parameters θ with respect to the error introduced by the model parameterized by θ' due to the gradient descent method. 12. End as long as

[0038] During the training phase, the learning model 100 is optimized for the application. If the learning model 100 is used to detect anomalies in a new workpiece and / or a new machining process, the invention assumes that sufficient data related to normal machining was collected during the calibration phase of the machine and the process, for example, by producing test workpieces. According to the invention, the method steps of the inner training loop are used by the meta-learning algorithm 200 to quickly adapt the learning model parameters 100 to the new situation (i.e., new workpiece and / or new manufacturing process): 1. Preprocess the calibration data. 2. Initialize the learning model 100 with the parameters θ found during the learning model training. 3. Take k training examples for the normal machining behavior according to the new process Pi. 4. Update the learning model parameters θ to θ' using the same configured gradient method used during training.

[0039] When the learning model 100 enters production mode, the calibrated learning model 100 is used to detect anomalies in the data currently generated by the sensors (live data) for a workpiece and / or product in a specific production process: 1. Preprocess the currently generated data. 2. Initialize the learning model 100 with the parameters θ ' found during model calibration. 3. Predict anomaly probabilities for the currently generated data using the learning model 100.

[0040] The learning model 100 is trained under conditions similar to those that would occur in a real production scenario. Advantageously, the learning model 100 is trained in a laboratory environment with a sufficient amount of data with a data identifier from various machining processes and / or manufacturing steps, such as the use of different machines. The agnostic meta-learning algorithm 200 enables the learning model 100 to quickly adapt to a new production environment and / or production processes.

[0041] Advantageously, in each training iteration, a process step and a limited data set from the normal manufacturing and machining processes used in the production of workpieces are used to calibrate the learning model 100. The learning model 100 is therefore capable of detecting anomalies in individual workpieces from sensor signal data generated during a new manufacturing process that was not used during the training phase of the learning model 100. When the learning model 100 is deployed in production, anomalies that fall below or exceed a specified threshold are then quickly and efficiently detected for each individual workpiece for which current data is generated.

[0042] Fig. 3 shows a flow chart of an inventive method for quality control in industrial production.

[0043] In a step S10, a learning model 100 is created for one or more production processes of at least one workpiece and / or product.

[0044] In a step S20, the learning model 100 is trained and initialized with a meta-learning algorithm 200.

[0045] In a step S30, the learning model 100 is calibrated with standardized data of at least one production process for the at least one workpiece and / or product.

[0046] In a step S40, currently generated data of the at least one production process for at least one currently manufactured workpiece / product are passed on to the learning model (100), wherein the data are generated by sensors.

[0047] In a step S50, the learning model 100 compares the currently generated data with the normalized data and identifies deviations.

[0048] In a step S60, the learning model scales the deviations between the currently generated data and the normalized data.

[0049] In a step S70, the learning model reports the presence of an anomaly in the currently produced workpiece / product.

[0050] Fig. 3 schematically illustrates a computer program product 300 containing one and / or more executable computer codes 350 for carrying out the method according to the first aspect of the invention.

[0051] The present invention can be used very cost-effectively because the calibration effort of the learning model 100 for a new manufacturing process is low and can be easily implemented. Furthermore, the present invention enables guided quality control compared to planned quality control in industrial manufacturing. The time required to react to defects during ongoing production can be significantly shortened, and the amount of product scrap can be considerably reduced. The present invention can be used advantageously for high-volume and mass production because each workpiece can be inspected using high-resolution data. The present invention can be used to address the singularities inherent in every production process, every machine, and every location.

Claims

1. Method for quality control in industrial manufacturing for one or more production process / es for producing at least one workpiece and / or product, comprising: - creating (S10) a learning model (100) for at least one production process for the workpiece and / or product; - training and initializing (S20) the learning model (100) using a meta-learning algorithm (200); - calibrating (S30) the learning model (100) using normalized data of the at least one production process for the at least one workpiece and / or product; - forwarding (S40) currently generated data of the at least one production process for at least one currently produced workpiece / product to the learning model (100), the data being generated by sensors; - the learning model (100) comparing (S50) the currently generated data with the normalized data and finding deviations; - the learning model (100) scaling (S60) the deviations between the currently generated data and the normalized data; and - the learning model (100) communicating (S70) the presence of an anomaly for the currently produced workpiece / product.

2. Method according to Claim 1, characterized in that the learning model (100) is in the form of a deep neural network.

3. Method according to Claim 1 or 2, characterized in that the meta-learning algorithm (200) is an agnostic meta-learning algorithm that trains the learning model (100) by means of a gradient method.

4. Method according to one of the preceding claims, characterized in that a normalized mean value of a stipulated measured variable for a workpiece and / or product and / or production process is defined as the basis for calculating a deviation.

5. Method according to one of the preceding claims, characterized in that the currently generated data have a data identifier.

6. Method according to one of the preceding claims, characterized in that the sensors use nonoptical methods for data generation.

7. Method according to one of the preceding claims, characterized in that an anomaly is found if the deviation is above or below a stipulated limit value.

8. System for quality control in industrial manufacturing for one or more production process / es for producing at least one workpiece and / or product, comprising: a learning model (100) for at least one production process for the workpiece and / or product, the learning model (100) being designed to be trained and initialized by a meta-learning algorithm (200), and the learning model (100) being designed to be calibrated using normalized data of the at least one production process for the at least one workpiece and / or product, and one or more sensors designed to generate current data of a production process for a currently produced workpiece and to forward said data to the learning model (100), the learning model (100) being designed to compare the currently generated data with the normalized data, to find deviations and to scale the deviations between the currently generated data and the normalized data, and the learning model (100) being designed to communicate the presence of an anomaly for the currently produced workpiece / product.

9. System according to Claim 8, characterized in that the learning model (100) is in the form of a deep neural network.

10. System according to Claim 8 or 9, characterized in that the meta-learning algorithm (200) is in the form of an agnostic meta-learning algorithm that trains the learning model (100) by means of a gradient method.

11. System according to one of preceding Claims 8-10, characterized in that a normalized mean value of a stipulated measured variable for a workpiece and / or product and / or production process is defined as the basis for calculating a deviation.

12. System according to one of preceding Claims 8-11, characterized in that the currently generated data have a data identifier.

13. System according to one of preceding Claims 8-12, characterized in that the sensors are designed to use nonoptical methods for data generation.

14. System according to one of preceding Claims 8-13, characterized in that an anomaly is found if the deviation is above or below a stipulated limit value.

15. Computer program product (300) containing one or more executable computer codes (350) for performing the method according to one or more of Claims 1 to 7.