Manufacturing equipment control via predictive "sequence to sequence" models

Abstract

The one or more processors measure values ​​of control and extrinsic parameters of a manufacturing system and generate a feature set describing the evolution of a state space of the manufacturing system from time series data of sensors measuring values ​​of functional parameters of parts produced by the manufacturing system, the one or more processors also generate a predicted value of at least one of the feature parameters from the feature set, and modify at least one of the control parameters in accordance with the feature set and the predicted value to drive the predicted value towards a target value.

Description

[Technical Field]

[0001] The present disclosure relates to controlling manufacturing equipment. [CROSS-REFERENCE TO RELATED APPLICATIONS] This application claims the benefit of U.S. Provisional Application No. 63 / 256,344, filed October 15, 2021, the disclosure of which is incorporated herein by reference in its entirety. [Background technology]

[0002] Manufacturing control systems may respond to input signals and generate output signals that cause the equipment they control to operate in a particular manner. Summary of the Invention [Problem to be solved by the invention]

[0003] The manufacturing system includes one or more processors that generate a feature set describing the evolution of a state space of the manufacturing system in the frequency or time domain from time series data of sensors that measure values ​​of control and extrinsic parameters of the manufacturing system and values ​​of functional parameters of parts produced by the manufacturing system, and the one or more processors further generate a predicted value of at least one of the functional parameters from the feature set via a "sequence to sequence" model of the manufacturing system, and modify at least one of the control parameters according to the feature set and the predicted value via a control agent to drive the predicted value toward a target value.

[0004] The method includes generating a feature set describing the evolution of a state space of the manufacturing system in the frequency or time domain from time series data of sensors measuring values ​​of control and extrinsic parameters of the manufacturing system and values ​​of functional parameters of parts produced by the manufacturing system. The method also includes generating a predicted value of at least one of the functional parameters from the feature set via a "sequence to sequence" model of the manufacturing system, and modifying, via a control agent, at least one of the control parameters in accordance with the feature set and the predicted value to drive the predicted value toward a target value. [Brief explanation of the drawings]

[0005] [Figure 1] FIG. 1 is a block diagram of a manufacturing system. [Figure 2] FIG. 2 is a block diagram of a control system. [Figure 3] FIG. 2 is a block diagram of a control system. [Figure 4] FIG. 4 is a block diagram of the manufacturing and control system of FIGS. 1, 2 and 3. DETAILED DESCRIPTION OF THE INVENTION

[0006] Herein, embodiments are described. However, it should be understood that the disclosed embodiments are merely exemplary, and that other embodiments may take various alternative forms. The figures are not necessarily to scale. Some features may be exaggerated or minimized to show details of specific components. Therefore, specific structural and functional details disclosed herein should not be construed as limiting, but merely as representative basis for teaching those skilled in the art.

[0007] Various features illustrated or described with reference to any one example may be combined with features illustrated or described in one or more other examples to create embodiments not explicitly illustrated or described. The illustrated combinations of features provide representative embodiments for typical applications. However, various combinations and variations of features consistent with the teachings of this disclosure may be desired for particular applications or implementations.

[0008] "Sequence to sequence" models, particularly recurrent neural networks, are typically used in natural language processing, such as machine translation, question answering, and text summarization. Here, the "sequence to sequence" framework is applied to manufacturing control problems, where the myriad manufacturing conditions (e.g., temperature, pressure, current flow, etc.) that affect the values ​​of measurable properties such as stiffness, thickness, and length vary over time, with the goal of producing products with more consistent measurable properties.

[0009] Machines used in mass production often have control parameters that affect the measurable properties of the resulting parts. To give a simple example, a stamping machine may apply a certain amount of pressure for a certain amount of time to form metal into a desired shape. Therefore, the stamping machine's ability to repeatedly produce the same desired shape depends on this pressure and time. If the values ​​of these control parameters change over time, a part made one hour earlier may have a slightly different shape than a part made one hour later, resulting in reduced consistency between parts.

[0010] In this example, the actual applied pressure may be a function of the power supplied to the stamping machine for a given pressure setting. Therefore, variations in the supplied power may lead to variations in the applied pressure, even if the pressure setting remains unchanged. Therefore, variations in the supplied power may be coupled with variations in part shape, albeit with a time lag. That is, considering the processing time associated with the stamping machine, a change in the supplied power at time zero may manifest as a deviation from the desired shape at time 42 seconds. If it is possible to predict the impact of sudden changes in supplied power on part shape, pressure settings can be strategically altered to offset such changes. Specifically, if a decrease in power is anticipated, the pressure setting can be increased accordingly. If an increase in power is anticipated, the pressure setting can be decreased accordingly.

[0011] Statistical techniques, such as statistical process control, are commonly used to monitor and control manufacturing processes with the goal of producing less waste and more conforming products. Complex manufacturing processes can limit the effectiveness of these techniques. Machines used in mass production can have hundreds, if not thousands, of control parameters (and extrinsic parameters) that affect the measurable characteristics of the resulting manufactured part, and the number can be in the tens (e.g., 20). The ability to predict the effect of changes in control and extrinsic parameters on the measurable characteristics of a part is thus a complex endeavor.

[0012] As mentioned above, machine learning techniques commonly used in natural language processing have been found to be well suited to the task of predicting the impact that instantaneous changes to a large number of parameters may have on measurable characteristics of a part. These predictions can be used as feedback to control the process to produce more consistent part outcomes despite changes in input (including exogenous) parameters.

[0013] Roughly speaking, recurrent neural networks store inputs via internal memory, allowing them to handle continuous data, such as time series data representing ambient conditions, control inputs to manufacturing equipment, and measurable characteristics of parts produced by the equipment. This internal memory allows recurrent neural networks to track information about the inputs they receive and predict what will happen next. That is, recurrent neural networks add the immediate past to the present. Thus, recurrent neural networks have two inputs: the present and the recent past. Weights are applied to the present and recent past inputs. These weights may be adjusted for gradient descent or backpropagation. Furthermore, the mapping from input to output does not need to be one-to-one.

[0014] Long-short-term memory networks are an extension of recurrent neural networks. Long-short-term memory allows recurrent neural networks to store inputs for longer periods of time in what is called memory, which can be read, written, and deleted. This memory can decide whether to retain or delete information based on the importance assigned to the information. The importance of certain information can be learned over time by the long-short-term memory. A typical long-short-term memory has sigmoid input, forget, and output gates. These determine whether to accept a new input, delete it, or allow the new input to affect the current time step output.

[0015] A "sequence to sequence" model can be constructed using a recurrent neural network. A common "sequence to sequence" structure is the encoder-decoder structure, which has two main parts: an encoder and a decoder. The encoder and decoder can be, for example, long-short-term memory models, respectively. Other such models, such as the transformer model, are also possible. The encoder reads the input sequence and summarizes the information into an internal state or context vector. The encoder's output is discarded, and the internal state is preserved, which helps the decoder make accurate predictions.

[0016] The initial state of the decoder is initialized to the final state of the encoder, i.e., the internal state vector of the last cell of the encoder is input to the first cell of the decoder. In the initial state, the decoder can start generating an output sequence.

[0017] These and similar concepts have been adapted for use in manufacturing. Long-term and short-term encoder-decoder models, transformers (e.g., bidirectional encoder representations from transformers, pre-trained generated transformer 3s, etc.), or other models may form the basis of a "sequence-to-sequence" model trained to interpret time-series data describing ambient conditions and manufacturing operations and predict corresponding part characteristics. The time-series data, including actual control parameter values ​​(e.g., current, machine revolutions per minute, machine pressure, machine temperature, etc.) and extrinsic parameter values ​​(e.g., ambient temperature, humidity, etc.), the changes in these values ​​over predefined periods, and other related data, may be pre-processed using various digital signal processing techniques (e.g., Fourier analysis, wavelet analysis, etc.) to generate a feature set (the set of all possible configurations) that describes the evolution of the manufacturing equipment's state space in the frequency and / or time domains. The specific set of digital signal processing techniques for a given application can be determined using standard methodologies, including simulation, trial and error, etc.

[0018] Referring to FIG. 1 , a manufacturing system 10 may include a manufacturing device 12 (e.g., an extruder, a press, etc.) that physically or virtually produces (e.g., assembles, creates, etc.) a manufactured part 14 (e.g., a tube, a panel, etc.). The manufacturing system 10 may also include one or more ambient condition (extrinsic) sensors 16, current sensors 18 (e.g., a motor drive current sensor, etc.), voltage sensors 20 (e.g., an internal temperature sensor, etc.), one or more additional sensors 22 (e.g., a conveyor speed sensor, a percent proportional, integral, or derivative output sensor, etc.), one or more property sensors 24 (e.g., a differential pressure sensor, a part dimension sensor, a material velocity sensor, etc.), and a database 26 (e.g., a relational database, a time series database, etc.). The ambient condition sensor 16 measures one or more ambient conditions (e.g., humidity, temperature, etc.) in the vicinity of the manufacturing device 12. The current and voltage sensors 18 and 20 measure the current and voltage supplied to the manufacturing device 12. The additional sensor 22 measures other control parameters of the manufacturing device 12. The property sensors 24 measure various characteristic parameters (eg, length, stiffness, thickness, etc.) of the manufactured part 14 .

[0019] These detected values ​​may be sequentially reported to database 26. That is, at time t0, each of sensors 16, 18, 20, 22, and 24 detects and reports its value to database 26, at time t1, each of sensors 16, 18, 20, 22, and 24 detects and reports its value to database 26, and so on. In this manner, database 26 receives time series data describing ambient conditions and control parameter values ​​associated with the operation of manufacturing equipment 12, as well as characteristic parameter values ​​associated with manufactured parts 14 produced by manufacturing equipment 12. Such an arrangement can be used to collect vast amounts of data for learning purposes.

[0020] Various transformations (e.g., data cleansing, bandpass filtering, convolution, principal component analysis, wavelet transform, etc.) can be performed on the time series data held in database 26 to generate a streaming feature set spanning the relevant state space that describes the evolution of the manufacturing process associated with manufacturing equipment 12. In one example, data cleansing includes backfilling, forward filling, and / or removal of null values ​​to ensure that the time series data is free of missing or poorly filled features. After data cleansing, principal component analysis can be performed to maximize the amount of useful information while minimizing the number of features. If the original data set includes pressure, temperature, and drive power, all of which have identical response information, principal component analysis compacts the data set while preserving the response information, for example, so that the temperature and drive power values ​​are ignored while the pressure value is used for continuing the transformation and training the process. Other transformation operations may be performed, but are not required. Other transformation steps may be performed, but are not required. At any point in time, the combined transformed data represents the maximum amount of state information for manufacturing system 10. The relevant state space can be identified iteratively during model training and evaluation.

[0021] Referring to FIG. 2, one or more processors 28 may implement a long-term / short-term encoder-decoder model 30 (or other suitable model) trained on at least a portion of the streaming feature set from database 26. For example, a recurrent neural network linking one machine to another iterates the model weights until the gradient, which represents the change in the model's loss function (e.g., squared error loss) per change in the model weight, asymptotically approaches zero. The weights of the recurrent neural network may be randomly seeded. The model may have various depths and widths depending on the number of features present in a particular manufacturing line and the complexity of the line's dynamic behavior. An example model may have two layers with 256 memory units. An adaptive moment estimation (Adam) optimizer may be used to perform gradient descent with a variable learning rate. Other optimizers, such as Adamax, are also contemplated.

[0022] For example, 60 minutes, 600 minutes, or 6000 minutes of streaming feature sets can be used to train the long-term / short-term encoder-decoder model 30 to recognize relationships between the detected ambient condition and control parameter values ​​of the sensors 16, 18, 20, 22 and the resulting detected feature parameter values ​​of the characteristic sensor 24. Once properly trained, the model 30 can predict future feature parameter values ​​of the manufactured part 14 from the streaming feature sets.

[0023] 1 and 3 , the one or more processors 28 may further implement a model 30 and a control agent 32 trained on streaming feature sets from the database 26. Prior to training the control agent 32, the model 30 (or other source) may inform the control agent 32 of the control limits of the manufacturing equipment 12 that may be simulated by the model 30. The control limits may include, for example, the operating pressure range of the press (300 psi to 500 psi), the operating temperature range of the drying oven (50°C to 80°C), etc. Additionally, the control agent 32 may receive target feature parameter values ​​for the manufactured part 14 (e.g., target length = 3 cm, target stiffness = 5 N / m, etc.). During training of the control agent 32, the model 30 and the control agent 32 may each synchronously receive the same portion of the streaming feature set from the database 26 to simulate feedback from the sensors 16, 18, 20, 22, and 24 during a manufacturing run. This enables the model 30 to generate predicted feature parameter values ​​for the simulated manufactured part and report them to the control agent 32. The control agent 32 can then instruct the model 30 to take control actions to change the control settings within the control limits. In an initial iteration, assuming the operating pressure of the press simulated by the model 30 is 310 psi and the operating temperature of the drying oven simulated by the model 30 is 62° C., the control agent 32 can increase one by some amount and decrease the other by some amount to learn how such changes affect the predicted feature parameter values ​​from the model 30 relative to the target feature parameter values. The amount of change can be arbitrary or can follow a predetermined rule. The control agent 32 can perform thousands, if not millions, of such iterations in a relatively short period of time, training itself on how to change the control settings of the manufacturing tool 12 to keep the predicted feature parameter values, and therefore the actual feature parameter values, at or near the target feature parameter values ​​as the values ​​from the sensors 16, 18, 20, 22, and 24 change.

[0024] 4 , once the control agent 32 is properly trained (e.g., the error between the predicted and target feature parameter values ​​is within some predetermined range, such as 5%), one or more processors 28 may be deployed within the manufacturing system 10 to receive the live data output by the sensors 16, 18, 20, 22, 24 and preprocess the data using the various transformations described above (e.g., data cleansing and principal component analysis) to generate a live-streaming feature set spanning the relevant state space that describes the evolution of the manufacturing process associated with the manufacturing equipment 12. Similar to the above, the now-trained control agent 32 may then, based on the live-streaming feature set and the corresponding predicted feature parameter values, instruct the manufacturing equipment 12 to take control actions that modify control settings within control limits in order to maintain the predicted feature parameter values, and therefore the actual feature parameter values, at or near the target feature parameter values.

[0025] The algorithms, methods, or processes disclosed herein can be provided to or implemented by a computer, controller, or processing device, which may include a dedicated electronic control unit or a programmable electronic control unit. Similarly, the algorithms, methods, or processes can be stored in many forms, including, but not limited to, as data and computer- or controller-executable instructions, permanently stored on a non-writable storage medium such as a read-only memory device, and reversibly stored on a writable storage medium such as a compact disc, random-access memory device, or other magnetic and optical media. The algorithms, methods, or processes can also be implemented in software-executable objects. Alternatively, the algorithms, methods, or processes can be embodied in whole or in part using suitable hardware components, such as application-specific integrated circuits, field-programmable gate arrays, state machines, or other hardware components or devices, or a combination of firmware, hardware, and software components.

[0026] While exemplary embodiments have been described above, it is not intended that these embodiments describe all possible forms encompassed by the scope of the claims. The words used herein are words of description rather than limitation, and it is understood that various changes can be made without departing from the spirit and scope of the disclosure.

[0027] As noted above, features of various embodiments can be combined to create additional embodiments of the present invention not explicitly described or illustrated. While various embodiments may be described as offering advantages or being preferred over other embodiments or prior art implementations with respect to one or more desired characteristics, those skilled in the art will recognize that one or more features or characteristics may be compromised to achieve desired overall system attributes depending on the particular application and implementation. These attributes include, but are not limited to, cost, strength, durability, life cycle cost, marketability, appearance, packaging, size, serviceability, weight, manufacturability, ease of assembly, and the like. Thus, embodiments described as being less desirable than other embodiments or prior art implementations with respect to one or more characteristics are not outside the scope of this disclosure and may be desirable for particular applications.

Claims

1. 1. A manufacturing system comprising: One or more processors measuring values ​​of control parameters and extrinsic parameters of the manufacturing system that affect characteristic parameters, which are measurable characteristics of parts manufactured by the manufacturing system; and generating a feature set that describes the evolution of a state space of the manufacturing system in the frequency or time domain from time series data of sensors that measure the values ​​of the characteristic parameters; generating a predicted value of at least one of the feature parameters from the feature set via a "sequence to sequence" model of the manufacturing system; a control agent that has learned and trained the effect of changes in the control parameters on the predicted value, and that is programmed to change at least one of the control parameters according to the feature set and the predicted value, thereby driving the predicted value toward a target value.

2. 2. The manufacturing system of claim 1, wherein the one or more processors are further programmed to train the sequence to sequence model against a historical feature set of the manufacturing system.

3. 10. The manufacturing system of claim 1, wherein the one or more processors are further programmed to train the control agent on past feature sets and corresponding predictions from the "sequence to sequence" model.

4. 2. The manufacturing system according to claim 1, wherein the "sequence to sequence" model is an encoder-decoder model.

5. 5. The manufacturing system of claim 4, wherein the encoder-decoder model includes a long-short-term memory model.

6. 1. A method comprising: measuring values ​​of control parameters and extrinsic parameters of the manufacturing system that affect characteristic parameters, which are measurable characteristics of parts manufactured by the manufacturing system, and generating a feature set that describes the evolution of a state space of the manufacturing system in the frequency or time domain from time series data of sensors that measure the values ​​of the characteristic parameters; generating a predicted value of at least one of the feature parameters from the feature set via a "sequence to sequence" model of the manufacturing system; and varying at least one of the control parameters according to the feature set and the predicted value via a control agent that has been trained to learn the effect of changes in the control parameter on the predicted value, thereby driving the predicted value toward a target value.

7. The method of claim 6 , further comprising training the sequence to sequence model on a historical feature set of the manufacturing system.

8. The method of claim 6 , further comprising training the control agent on past feature sets and corresponding predictions from the sequence to sequence model.

Images