Intelligent inspection of packaging quality with TwinCAT Machine Learning
Instant noodles can be found in just about every food store in China. In order to reduce the number of products with packaging errors and the resulting customer complaints, a large Chinese producer of instant noodles decided to use the Beckhoff control technology including TwinCAT Machine Learning. This made intelligent and reliable real-time inspection of the packaging quality possible.
Tianjin FengYuLingKong the Electrical and Mechanical Equipment Co., Ltd., a system integrator specializing in industrial automation, was awarded the contract to supply a state-of-the-art packaging inspection system to one of China's largest producers of fast food and beverages. According to the food manufacturer, it has the world's largest instant noodle production lines, with an average throughput of around 500 packs per minute per line and an annual total throughput of up to 4.8 billion packs.
High customer satisfaction through error-free packaging
Packaging and sealing are essential tasks in the manufacture of instant noodles. A noodle pack usually contains a pre-cooked noodle block plus several small sachets with spices, e.g. flavoring powders, sauces and dried vegetables. In the production line, these individual sachets are arranged on a conveyor belt moving at high speed, placed on the associated noodle block and then forwarded together to the cross-cutting sealing and packaging machine.
During the production process there are various factors that can potentially lead to the seasoning sachets slipping between two noodle blocks and being cut open by the cutting machine or being packed separately in two packets side by side. Such defective products would result in consumer complaints and damage to the company's reputation, for which reason delivery of such products to dealers should be reduced as far as possible. Since the machine type upgraded by Tianjin FengYu already produced with a very low error rate before, another aspect of quality control is critical: It must be ensured that only the defective and not the defect-free products are reliably sorted out.
Because the processes inside the sealing and packaging machine cannot be viewed and any packaging defects often are also not visible from the outside during the subsequent optical inspection, it is very difficult to find out the exact cause of the errors mentioned above. It is therefore virtually impossible to avoid defective products on principle. In order to avoid delivering substandard goods despite that, instant noodle manufacturers use highly automated quality inspection devices with minimum latency.
Tianjin FengYu supported the end user in this project with the fast implementation of a high-performance quality control system. First of all, the mechanical and electrical parts of the machine were examined. According to Tianjin FengYu, in the process they found out that due to its openness and flexibility, PC-based control technology from Beckhoff is ideally suited for acquiring the necessary analysis data. The system integrator installed several sensors inside the machine and was able to acquire the first datasets for prototypical analysis quickly and easily via TwinCAT Scope View. The subsequent data analysis showed a certain sensitivity of the measurands that occurred when a machine had wrongly cut a seasoning sachet. However, such disruptive events resulting in faulty products could not be reliably detected using conventional engineering methods. The reasons are several uncertainty factors that can affect the data, such as machine vibrations as well as changes in the packaging material or the conveying speed and the cutting tension. To find a solution for this demanding analysis application, Tianjin FengYu decided to use TwinCAT Machine Learning as a means of implementing data-based engineering with machine learning (ML).
Machine learning in industrial applications
The basis for an inspection of products are classifiers, using which flawless products can be distinguished from faulty products. If ML-based classifiers are to be used, a mathematic model is trained on the basis of exemplary data so that from then on correct decisions can be made without explicit programming.
Beckhoff offers various tools and open interfaces to support the entire engineering cycle from data acquisition and model training through to the deployment of the learned model directly within the control system:
- Data acquisition: The quantity and quality of the data have crucial effects on ML applications. A wide range of I/O and software products from Beckhoff enable the acquisition of almost any data. Various functions of the TwinCAT software, such as Scope View, Database Server, Data Agent and Analytics Logger, enable the data to be saved on an Industrial PC, in local or remote databases or the cloud.
- Model training: The data acquired must initially be preprocessed in order to find a correlation between the data and the desired results or to strengthen the correlation. Subsequently, a suitable ML algorithm is identified and parameterized for model training. Beckhoff recommends the use of open and established ML frameworks for this purpose, such as PyTorch, Keras or Scikit-learn. Finally, the trained model can be saved as a file in the standard exchange format ONNX (Open Neural Network Exchange). The ONNX file describes the operations and parameters of the trained model and can then be converted into a binary format (BML) that is better suited for serialization in TwinCAT.
- Model deployment: The TwinCAT runtime environment for machine learning (TF3800 and TF3810) can load the trained model files (BML format) dynamically into the controllers, where the models can be run in real-time with execution cycles of less than 1 ms. In this way, the results of the inference (execution of a trained ML model) can be directly processed, transmitted to the output devices via the ultrafast EtherCAT communication and so the machine can be controlled in real-time.
The detection of faulty products in the instant noodle production line was implemented exactly according to the three-step method described above. First of all, the sensor data were acquired via EL1xxx or EL3xxx EtherCAT digital and analog input terminals and TwinCAT Scope View. Subsequently, the ML model was trained via the open source framework Scikit-learn, and the model description file was generated from it. The necessary preprocessing of sensor data was implemented with TwinCAT Condition Monitoring in the control system. Then the corresponding BML file was deployed to a CX51xx Embedded PC, which runs the model in real-time with the help of the TwinCAT Machine Learning runtime and outputs the inference results for the identification of faulty products via an EL2xxx EtherCAT digital output terminal. According to Tianjin FengYu, the system openness is a great advantage of the Beckhoff control technology and was very beneficial here, because it could be integrated with the existing third-party main controller of the production line with no great effort.
Open platform accelerates algorithm development
The validation of ML algorithms is often a time-consuming and laborious process because of the necessary tests and the associated frequent visits to end customers. With the TwinCAT open software platform, however, algorithms can be validated efficiently without requiring direct access to the machine. The data recorded on a production machine is separated into training and validation data before model training is started. Only the training data set is then used to train the ML algorithm. The validation data set can initially be used in the training environment to test how well the learned algorithm performs on unknown data.
After successfully porting the data preprocessing and integrating the ML algorithm into TwinCAT, validation can be performed based on the production code. The code is executed on a test system – or the actually used Embedded or Industrial PC – and the validation dataset is streamed to the TwinCAT real-time environment for use as a virtual data source by means of TwinCAT Database Server functions. The same sampling frequency as with the on-site sensors is adopted, therefore the scenario at the end customer’s site can be optimally simulated. Similarly, new datasets collected on the production machine can be used in the test environment to explore a wide variety of situations. Ultimately, the test data serve to validate the entire ML application on the Embedded PC, to assess it and ensure safe operation.
The development, verification and validation using the open TwinCAT platform eliminates the need for testing the ML algorithms on industrial plants and considerably accelerates the implementation phase. According to Tianjin FengYu, this contributed to the fact that the progress of the joint project at this food company was barely affected by the corona pandemic.
Multi-tasking and multi-core capabilities secure ML in real time
The ML algorithm is executed on the CX51xx in three steps:
- acquisition of sensor data
- preprocessing of the data
- execution of ML models for the detection of faulty products
On the one hand, it should be noted that the product inspection requires a high sampling frequency in order to acquire sensor data during the entire cutting process. On the other hand, a lower execution frequency is necessary in order to be able to process the acquired data and to execute the ML model. However, this apparent contradiction for two real-time sequences in one PLC task can be solved very well through the multi-tasking and multi-core capabilities of TwinCAT, because this ensures both the reliable execution of multiple tasks on different processor cores and the error-free exchange of data between multiple PLC tasks. Another advantage is that these functions can be implemented through simple configuration and via ready-made PLC function blocks with little development effort. In the project described for the instant noodle production, the two-step execution of the ML algorithm through the use of two PLC tasks and two processor cores was realized reliably.
Machine learning and PC-based control minimize effort
Through the approaches of machine learning and with the help of data mining, production problems can be solved faster and more efficiently, thus saving R&D costs. The experiences of Tianjin FengYu with the instant noodle production line have shown that TwinCAT Machine Learning is superior to the traditional engineering methods in the detection of anomalies. With the open TwinCAT platform, it was possible to automate the complete workflow for data acquisition, training and ML inference. As of the end of 2020, the new quality inspection system has already been running fully automatically on several of the end customer's production lines for about four months. The customer's conclusion: Due to the open Beckhoff solution it was possible to implement the inspection system without modifying the existing main control system; faulty products are detected promptly and reliably, which effectively reduces customer complaints.
System integrator Tianjin FengYu pointed out that the use of the machine has become much simpler and more flexible through the Beckhoff products CX51xx, EtherCAT I/Os and TwinCAT. Despite the complex mechanisms of the production line on site, commissioning and maintenance was very simple with the great support of the Beckhoff engineers. Moreover, the open PC-based control technology has solved the problem of the quality inspection of packaging, also by using ML methods to acquire the data from existing noodle machines, irrespective of their make. Tianjin FengYu is convinced that more and more end users will benefit from TwinCAT Machine Learning in the near future.