DE202011110730U1 - Cognitive machining head for machining workpieces - Google Patents

Abstract

A cognitive machining head for machining workpieces, comprising a control unit adapted to control the steps of: - performing an initial machining process of an initial workpiece type by acting based on learned knowledge; and - making an autonomous decision in a previously unknown process situation for adapting the learned knowledge to a secondary machining process of a secondary workpiece type.

Description

The present invention relates to a cognitive machining head for machining workpieces. In particular, the present invention is directed to transferring stimuli from natural cognition into real time process control for industrial production systems.

Industrial production uses a large number of robots and complex technical systems to produce various products. In addition, a manufacturing process has to meet two requirements at once, namely to increase mass production and to provide increased flexibility in individualized versions of complex products, which is referred to as mass customization. This requirement can be met in part by highly adaptive production systems that can learn how to accomplish manufacturing tasks or how to monitor and control them using machine learning. Such adaptive and autonomous production systems would save costs and labor by facilitating easier replacement of production lines, improving product processing quality, and preserving environmental resources. Laser material processing (and particularly laser beam welding) is one example of a complex and difficult to control manufacturing process that is often used in, for example, automotive or aircraft assembly lines. However, the execution times of many machine learning algorithms are too slow to control the production processes in real time.

It is an object of the invention to provide a cognitive machining head for machining workpieces which provides an architecture or design of technically cognitive methods for controlling at least one parameter in industrial production processing tasks using sensors and actuators in a reasonable time frame.

This object is achieved by a cognitive machining head for machining workpieces according to claim 1. Further advantageous developments and refinements of the invention are specified in the corresponding subclaims.

In many cases, when a production system is installed or reconfigured, qualified professionals are present. In this case, the technical cognitive architecture according to the invention is given the opportunity to first learn from a human expert and then to learn without supervision or to make independent decisions.

Moreover, the architecture of the present invention can also be applied to problems of laser material processing or robot welding with laser beams. Laser beam welding is a suitable model case for production processes that require a high degree of precision, using many individually weak sensors for data acquisition, and requiring rapid adaptation to complex system configurations and real-time process control. The present invention aims to provide a data analysis concept which satisfies the needs for various production scenarios, particularly in relation to adaptive and autonomous laser material processing. Many other architectures and structures towards cognitive engineering systems and machine learning approaches suitable for the control of production systems have been demonstrated.

In order to achieve the above object, the present invention provides a cognitive machining head for machining workpieces having a control unit adapted to control the steps of: performing an initial machining process on a workpiece of an initial workpiece type by acting based on learned Knowledge; and making an autonomous decision in a previously unknown machining situation to adapt the learned knowledge to a secondary machining process on a workpiece of a secondary workpiece type.

In another aspect, the present invention provides a cognitive machining head for machining workpieces having a control unit adapted to control the following steps: abstracting relevant sensor information relating to an initial machining process on a workpiece of an initial workpiece type; Learning from a human expert how to control the initial machining process on a workpiece of the initial workpiece type; Performing an initial machining process on a workpiece of the initial workpiece type by acting based on learned knowledge; and making an autonomous decision in a previously unknown machining situation to adapt the learned knowledge to a secondary machining process on a workpiece of a secondary workpiece type.

The machining of the workpieces preferably includes a laser welding or laser cutting process.

According to a preferred embodiment of the invention, abstraction involves more relevant Sensor information an initial trial run in which an actuator parameter is changed and the sensor information is processed by a feature extraction technique to embed the sensor data input into a learned feature space.

Preferably, the relevant sensor information comprises information from sensors comprising a high-speed camera, solid-state and airborne sound sensors and three photodiodes which receive different process emissions at different wavelengths.

According to an advantageous development of the invention, the feature extraction technique comprises a linear or non-linear dimensional reduction processing of the sensor data input.

It is preferred that the learning be controlled by a human expert, such as an initial machining process, to control actuators including a linear drive or robotic positioning actuator or an actuator to adjust the laser power.

According to one embodiment of the invention, learning from a human expert includes classification techniques such as support vector machines, artificial neural networks and fuzzy k nearest neighbor technique.

Preferably, performing an initial machining process includes combining PID control with classification results.

In another embodiment of the present invention, performing an initial machining process involves performing laser power controlled lap welds in which excessive gaps or gaps between two sheets to be joined are monitored.

In an advantageous embodiment of the present invention, the meeting of an autonomous decision comprises a novelty check based on trained data.

Preferably, if the novelty check is positive, another test action is performed to classify the secondary workpiece type with the previously trained features using a supervised or unmonitored learning procedure.

In one embodiment of the present invention, the secondary workpiece type is preferably based on a different material thickness of the workpiece.

Preferably, the supervised learning procedure includes reinforcement learning.

Preferably, the machining head of the present invention comprises a laser optics for laser machining of workpieces, actuators including a linear actuator or robotic positioning actuators, a laser power adjusting actuator, and sensors including a high speed camera, solid state and airborne sounding sensors, and three photodiodes representing different process emissions in different ones Record wavelengths.

The accompanying drawings, which serve to further understand the invention and form a part of this application, illustrate embodiments of the invention together with the description and serve to explain the principle of the invention. Hereby show:

1 an architecture for adaptive real-time process control of actuators in industrial production systems according to the invention;

2 an embodiment of an architecture for an adaptive real-time process control of a laser welding process according to the invention;

3A an in-process image taken with a coaxial camera;

3B an image of a laser welding process with acoustic sensors;

4 a plot of features and a probability related to a gap between two sheets of stainless steel;

5 a graph related to a probability of a gap between two sheets of stainless steel for a workpiece number 151, wherein the real gap is marked; and

6 a diagram related to the mapping of features of a reference workpiece on workpieces with different thicknesses, wherein the gray area marks the usual laser welding process processing area.

First, inspiration from natural cognition is presented in a first section. Thereafter, an architecture for real-time process control according to the invention is introduced. This concept is transferred to laser beam welding in the last section and confirmed experimentally.

Natural cognition is in their complex variations across different species and individuals not fully understood until now. However, recent results may be an inspiration for the design of cognitive technical systems.

Memory coding units were identified in the hippocampus of mouse brains. Events of cognitive relevance are not memorized in all details, but in a selective representation. The brain seems to abstract information by decoding it with active patterns. Sensory perceptions then generate such activation patterns that enable the mouse to identify previously learned situations. Moreover, these characteristics not only seem to exist and are likely to be between individual mice, but also between different species. It has been demonstrated in human brain activity that patterns of neuronal activation are related to thinking in different semantic categories.

A functional magnetic resonance imaging combination with a computation model involving dimensional reduction and classification techniques will be described, which may indicate when a subject is thinking of previously memorized objects, such as tools. Once a subject's neural activation pattern is learned, this method can visualize when another subject is thinking of a hammer or screwdriver. Groups of nerve cells in different combinations appear to respond to specific events or thoughts encoded by specific patterns. It seems that natural cognition is organized with abstracted information or patterns that represent real events in combination with categorization. Information from the real world is reduced to signals with patterns that allow the brain to distinguish and recognize events. These cognitive abilities enable people and other natural cognitive organisms not only to learn, but also to respond quickly. This is also very desirable in technical systems.

If natural cognitive abilities can be modeled using mathematical data reduction techniques coupled with classifiers, it should also be possible to realize similar abilities by computation. If neuroscientists can only move a cursor on a screen through thought, the computational methods may also be useful to improve complex control tasks in industrial production.

In the following, the technical cognitive architecture for real-time process control of the present invention will be presented.

A natural cognitive ability is to abstract abstract relevant information from a larger set and to differentiate categories within that information. Applying this concept of natural cognition to the world of mathematical data analysis may allow a combination of data reduction techniques and classification techniques to be used to accomplish something with similar behavior. In industrial production, many manufacturing processes can be thought of as a black box model that focuses on the box's inputs and outputs rather than paying attention to what happens inside that box. The connections to the black box for production systems are often sensors and actuators. Sensors such as cameras, microphones, touch sensors, and many more are provided to monitor production processes.

In addition, the systems require actuators such as linear drives or robotic positioning to interact with the real world environment. For each production process, these actuators must be parameterized. In order to learn how to adaptively control at least one parameter of such a production system, many combinations of self-learning algorithms, classification techniques, knowledge stores, feature extraction, dimensional reduction, and manifold learning techniques need to be tested. Furthermore, various control techniques, open and closed-loop, with many different sensors and actuators, must be investigated. After many simulations and experiments, a simple architecture according to the present invention has finally been found which shows how these techniques must be combined, which has been shown to be successful and reliable at least for laser welding. However, the laser welding process can be interpreted as a black box, which may be equally applicable to other types of production processes.

In 1 FIG. 10 is a diagram of an architecture or method of the present invention suitable for adaptive process control. FIG. The diagram describes the unit communication and information processing steps. The architecture or method of the present invention may be configured for four different types of use or method steps: first, abstracting relevant information; Second, receiving feedback from a human expert on how to control processes are, in other words, supervised learning; third, action based on learned knowledge; and fourth, autonomous control of the processes in previously unknown situations.

First, the step of abstracting relevant information will be discussed in detail. In natural human cognition we abstract or take information from everything we hear, feel and see. So, in general, we only remember the most interesting things. Inspired by this, a technical cognitive system can also abstract relevant information from a production process. Working with abstracted features rather than raw sensor data has some advantages, such as the following: many weak sensor signals can be translated into fewer and more reliable features. In addition, in order to realize real-time process control, it is necessary to reduce the amount of incoming sensor data, because larger amounts of data have a significant influence on the generation of longer processing times of the entire system. The architecture requires a test run to abstract information where the parameter range of the actuator to be controlled is changed. To determine which information is more relevant than others, the production system should explore its own range of actions. After this test run, the system analyzes the recorded sensor data to find representative features. It may solve a feature calculation separately for different types of sensors, but ideally the sensor units should be trained to map or map the sensor input to the learned feature space. Finding a usable feature space representation is critical because the system can only detect or respond to changes in the features. The type of feature selection may vary for different production processes, with the last section presenting a design for laser beam welding with the best practical results.

Second, the supervised learning step is discussed. In natural human cognition, such as childhood, we often learn from our parents to handle complex tasks. A machine should have the opportunity to initially learn its tasks from a human expert. In industrial production, a qualified human supervisor is usually present when the production system is installed or configured. The architecture of the present invention uses human-machine communication to obtain feedback or feedback from an expert, for example via a graphical user interface. As mentioned above, this architecture requires at least one test action per actuator or a test run in an initial learning mode. During these tests, the robot executes an actuator from a minimum to a maximum output and the sensor data input is stored. After this run, an expert gives feedback on whether the robot has operated the actuator correctly or whether its action was unsuccessful or undesirable. The feedback can have different categories, so types of errors and termination strategies can be defined. A classification technique may then collect the features with the corresponding supervisor feedback. Together with lookup tables, the classifier can serve as a knowledge and scheduling memory and for classifying the current system state. Support Vector Machines, Fuzzy k-Nearest Neighbor, and Artificial Neural Networks are discussed later as classification techniques. The more the human expert learns the machine, the more likely the system will be to reach the set goal. For cost-saving reasons, the necessary time of the human supervisor should be minimized. Thus, if no feedback is available from an expert, the architecture after the first test run can use its learned knowledge to provide feedback to itself, as explained in the Unmonitored Learning section.

Third, the step of adaptive process control will be discussed. In order to realize a robust process control in an industrial production process, a faster closed loop is often needed. When thinking of a joining or cutting process, a loop should be completed at least once before the interaction zone has left the processing area. For example, in a laser powered welding process, real-time process control implies that the laser focus should not have left the previous interaction zone from where the control loop was started. The advantage of the architecture of the present invention is that the use of features rather than raw sensor data allows the control loops to be completed more quickly while minimizing the loss of information. Any type of control design that matches the classification output can be implemented in this architecture. A simple version would be to have two possible classification output values: too much and too little. A PID controller can adjust a parameter of the system actuator, as will be described in the last section.

Fourth, the step of autonomous process control and unmonitored learning is discussed. While a production system has an adaptive process control with the discussed Architecture, it can happen that the system experiences something new in the sense of what was not previously learned. This can be the case with assembly lines when the workload changes or any other process parameters are changed that are not recognized by the system. A novelty check or a novelty check based on the training data may detect such differences. In this architecture, there would be a change in system mode, either to supervised learning when a human expert is present, or to autonomous process control. Thus, the system can also try to solve the problem with a self-learning mechanism. In the remainder of this description, the mapping or mapping of features will be described as a possibility. Another possibility would be the integration of reinforcement learning.

As described above, the architecture abstracts information that is reduced in its amount of data. What has been referred to as the 'activation pattern' in the section on inspiration from natural cognition can also be understood as features that represent sensor events. For example, using dimension reduction, a low-dimensional feature calculated from the training events would indicate if the system has experienced a particular event. Classification with support vector machines can categorize and distinguish these events. After an initial feature calculation, the presence of features in sensor data inputs can be processed much faster than with raw sensor data. A combination of data reduction and classification techniques is promising for process control, which allows it to act quickly.

The architecture of the present invention proposed for cognitive technical systems allows for various cognitive abilities, such as obtaining relevant information, learning from a human expert, and responding to new situations based on previously learned knowledge. This architecture may be used for various types of systems that require control of one or more actuators based on the input of a large amount of sensor data. Compared to some other approaches, learning and response skills seem limited; however, the architecture is very robust in terms of data acquisition; it is easy to use and can be implemented for fast computation up to a closed loop in real time of complex systems, such as the application in the following section shows.

The last section describes an illustrative embodiment of the present invention related to technical cognition and adaptive laser welding process control.

A typical procedure for production systems is that a unit of a production line must first be configured and then monitored for quality assurance. This is also the case for laser welding. When materials are processed using laser light, a high degree of precision is required. On the other hand, welding with laser beams is difficult to observe because of strong radiation and process emissions. For these reasons, different sensors are used to monitor the activities. Even then, it is challenging for human experts to evaluate, by evaluating the monitoring results, whether a welding operation has been successful and results in stable welds. In industrial production, these processes are initially configured with many manual trials, resulting in high labor and machinery costs. All process parameters are kept constant as a change in high recalibration costs would result and produce a production stop. A cognitive laser material processing system capable of responding adequately to change would be of great help and economic importance.

An approach to a cognitive system for laser material processing which follows the architecture of the present invention is disclosed in U.S. Pat 2 shown. The data processing is structured within this architecture. The following sensors are used: a high-speed camera, solid-borne and airborne sound sensors and three photodiodes that record different process emissions at different wavelengths, such as the 3A and 3B is shown. As mentioned above, the assessed cognitive abilities can be divided into four categories: first, abstracting relevant information; second, learning from a human expert how to control; third, action based on learned knowledge; and fourth, autonomous decision-making in previously unknown situations.

First, the step of abstracting relevant information should be discussed. The cognitive laser material processing system uses cameras, photodiodes and sensors for solid-state and airborne sound, which provide valuable process information. These sensor data must be recorded in an initial test run in which the applied laser power is changed. Other actuators can also be changed, in which case laser power changes have negligible reaction times, but greatly affect the result of the weld. Dimensional reduction serves primarily as a feature extraction technique. Linear and non-linear dimensional reduction allows mapping of a high-dimensional observation space Y to a low-dimensional feature space X. All data from the on-line sensor system used can be represented by a 26,603-dimensional vector which can be reduced to only 33 dimensions, where he represents the relevant process information by connecting various dimension reduction techniques. Thereafter, it is possible to embed the sensor data input in the learned feature space and to gain the current system process status by "out-of-sample extension". Experiments have shown that ideally each sensor has an optimized combination of feature extraction.

Second, the supervised learning step is discussed. Laser material processing systems are usually set up by human experts and their processes configured. The discussed architecture can simplify and accelerate this process. When the system is performing a test action such as a laser power ramp for welding, a human expert will present, via a graphical user interface representing the workpiece, how to classify the processing result for different workpiece areas. For example, the expert may mark bad welding with insufficient laser power, good welding, and poor welding with too much laser power. Our system stores this information, along with the extracted feature characteristics described above, with a classification technique. Different classification techniques have been evaluated and found to be suitable embodiments of the invention, such as Support Vector Machines, Artificial Neural Networks, and Fuzzy k Nearest Neighbor. All mentioned classifiers achieved good results; the extracted features appear to be separable from one another for many different process configurations. The combination of feature extraction and classification has proven to be a reliable stainless steel welding monitoring tool, such as contaminated areas on workpieces or variation in material thickness can be monitored herewith. Even previously difficult to detect errors can be observed, such as too large a gap between two sheets to be joined, as in the 4 and 5 shown.

Third, the step of adaptive process control will be described. Various closed-loop experiments were performed on the basis of classification probabilities. Combining a PID controller with the classification results allowed laser power controlled overlap welds of stainless steel workpieces to compensate for focus shift in process optics by varying the z axis and reducing the influence of the gap between the parts to be joined. Using features with less data and simple class structures allowed real-time laser welding process control to be performed. In accordance with the present invention, closed loop cycle frequencies greater than 500 Hz could be achieved, including feature mapping or mapping and classification of sensor data from camera images and photodiodes.

Fourth, the step of autonomous process control and unmonitored learning is discussed. Although every attempt has been made to keep all processing parameters constant for a configured process, influences can occur on varying workpieces, such as changes in assembly and workload. According to the invention, a self-learning mechanism is integrated. A novelty check based on trained features may detect new or previously unknown situations. In this case, the system performs another test action and classifies the new workpiece with the previously trained features. This time it does not have to consult a human expert; It can autonomously map the knowledge gained to the new workpiece ("mapping") and adjust the process control accordingly. This cognitive ability was tested for varying material thickness of ± 15%; the results are in 6 demonstrated. The usual work area is indicated in this figure, with good and bad welds being created. Especially within this range, the trained feature sets are now identified sooner or later on the workpiece with different material thicknesses. The results show that the mapping of characteristics seems to work reliably, since the mapping shows the necessary laser power adjustments. The experimental results show that the system of the present invention is capable of performing a judgment and correspondingly adjusting the applied laser power to achieve desired welding results. Other techniques for this step can be used, such as reinforcement learning, to increase the level of adaptivity at a high level.

In summary, an architecture of the present invention is presented which is applicable to various process control requirements within industrial production. The natural cognition-inspired design is applied to laser welding process control problems. The method and apparatus of the present invention provide cognitive capabilities for handling many sensors and actuators. The system is able to learn from a human expert how to weld materials, make decisions in previously untrained situations, and improve monitoring and process control at 500 Hz and more.

In order to carry out the present method, it is necessary to store the dimensionally reduced sensor data from the initial processing processes. These data entries are tagged or labeled with the feedback of the human expert and thus grouped within classes. In performing the process control, the system classifies the incoming dimensionally-reduced sensor data by calculating the class membership probability of the existing data entries or features. This step creates a similarity probability within the feature space. It therefore determines the process status and its probability of being in a desired state. The system optimizes this probability of being in a desired state. The probability of belonging in a good class can be used as a manipulated variable for a controller, such as a PID controller, which tries to achieve 100%. Further, feedback from the human expert may augment the tagged data stored within the classification. Within this step, the current processing process is tagged or labeled with human expert feedback and is additionally stored within the classifier. This allows the machine to be supported when it can not get into a good system class state or fine-tune its performance.

A number of different methods, such as kernel principal component analysis (PCA) or artificial neural networks or support vector machines and similar classifiers, allow the machine to perform a novelty check. With the novelty check, the system recognizes whether the system status is within the learned data entries. It can be realized with a further classification. He is currently able to determine if the system status is similar to the previously learned reduced-dimensional sensor data. For example, all tagged data entries within the feature space may be tagged or labeled for a different classification than the known data entries. With a support vector machine classification, the known data entries may be one class and the feature space origin another class. The support vector engine then calculates the probability of whether the system status is known or similar to the previously learned data entries. If this probability is below a predetermined threshold, the system status may be considered new or unknown. The novelty check creates security, since the system performs the desired action of a human expert only on the basis of learned knowledge. In unknown states, the system acts based on similarities that may not be appropriate for the desired machining process.

This novelty check can for example also be used as a further monitoring signal. The human expert can trust in all machine processes where the novelty check is not positive. In the case of a positive novelty check, the system may either require additional human expert feedback or it may also initiate a self-learning process, such as reinforcement learning or workpiece mapping. Other sensor data feedback can be used as a reward function for a reinforcement learning agent. Workpiece mapping describes a process in which dimensionally reduced sensor data is obtained from a new machine process and marked or labeled by similarity of a previous feature.

Despite other process control procedures, the present invention allows an adaptive process control to be performed based on a black box model. It is not necessary to assume a process control model in which the parameters can be adjusted. The system can also monitor itself using the novelty check and then perform an action. It is particularly suitable for processes with noisy and complex sensor data such as laser welding or laser cutting.

Claims (14)

A cognitive machining head for machining workpieces, comprising a control unit adapted to control the steps of: - performing an initial machining process of an initial workpiece type by acting based on learned knowledge; and - Making an autonomous decision in a previously unknown process situation for adapting the learned knowledge to a secondary processing process of a secondary workpiece type. The cognitive processing head of claim 1, wherein the control unit is further adapted to control the following steps: - abstracting relevant sensor information concerning the initial machining process of the initial workpiece type; and Learning from a human expert how to control the initial machining process of the initial workpiece type. The cognitive processing head of claim 1 or 2, wherein the machining of the workpieces comprises a laser welding or laser cutting process. The cognitive processing head of claim 2 or 3, wherein abstracting relevant sensor information comprises an initial test run in which an actuator parameter is changed and the sensor information is processed by a feature extraction technique to embed the sensor data input in a learned feature space. The cognitive processing head of claim 4, wherein the relevant sensor information comprises information from sensors including a high-speed camera, solid-state and airborne sound sensors, and three photodiodes that receive different process emissions at different wavelengths. The cognitive processing head of claim 4 or 5, wherein the feature extraction technique comprises linear or non-linear dimensional reduction processing of the sensor data input. The cognitive processing head of any one of claims 2 to 6, wherein the learning is to be controlled by a human expert such as an initial machining process, controlling an actuator comprising a linear actuator or a robotic positioning actuator or an actuator for adjusting a laser power. The cognitive processing head of any one of claims 2 to 7, wherein learning from a human expert includes classification techniques such as support vector machines, artificial neural networks, and fuzzy k nearest neighbor. The cognitive processing head of any one of the preceding claims, wherein performing an initial processing process comprises combining a PID controller with classification results. The cognitive processing head of any one of the preceding claims, wherein performing an initial machining process comprises performing laser power controlled lap welds in which too large gaps between two sheets to be joined are monitored. The cognitive processing head of claim 1, wherein the making of an autonomous decision comprises a novelty check based on the trained data. The cognitive processing head of claim 11, wherein if the novelty check is positive, another test action is performed to classify the secondary workpiece type with the previously trained features by a supervised or unmonitored learning procedure. The cognitive processing head of claim 12, wherein the secondary workpiece type is related to a different material thickness of the workpiece. The cognitive processing head of claim 12 or 13, wherein the supervised learning procedure comprises reinforcement learning.

Images