Intelligent process control through prediction

Abstract

The present invention proposes a method for controlling a process with at least pre- and post-measuring stations, comprising the steps of: collecting pre-measurement data prior to the process on a workpiece at a pre-measuring station, which is a first station; collecting process data after processing the workpiece at a process station, which is a second station; and collecting post-measurement data as test values ​​at a post-measurement station, which is a fourth station; storing the data in a database in accordance with the type of workpiece; defining the output as “Δ”, where “Δ = post-measurement test value - target adjustment value”, and defining the inputs as data collected from the pre-measurement station and the process station; AI modeling of “Δ” using machine learning algorithms with a software program that uses the input data; and adjusting the target adjustment value of the process station.

Description

Technical field of the invention

[0001] The present invention relates to a method for controlling a process of a machine in order to obtain a target value of the process for a workpiece using machine learning, in accordance with the preamble of claim 1. State of the art

[0002] The invention aims to control the test results rather than the process target. This is applied to a hydroerosive rounding machine with integrated measuring stations before and after the process. The critical output of the hydroerosion process is the hydraulic flow rate, which is checked at a fourth station after the hydroerosive rounding process. This type of process control can also be applied to other processes where part data can be collected before and after the process.

[0003] In the prior art, there is a constant target value "X" for production at a process station, which is the second station of the HE machine. The result is only checked at a subsequent measurement station, which is the fourth station of the HE machine, and the stability of the process can be derived at this step.

[0004] A hydro-erosive rounding process (HE process) is applied to any workpiece with holes or channels. An example is a fuel injection nozzle with injection channels.

[0005] A nozzle is part of a fuel injection valve through which the fluid is injected into a combustion chamber of an internal combustion engine.

[0006] To complete the injection hole production of nozzle bodies, parts are fed into the HE machine after the injection holes have been drilled. This machine operates with a round table with four stations.

[0007] Fig. Figure 3 illustrates the stages of a HE machine according to the state of the art: The first station is a test station, referred to as a "pre-measurement station," for measuring the hydraulic flow rate in a hole of the workpiece before the high-exposure (HE) process or any other machining operation. A test oil (a fluid) is passed through the injection hole, and the flow rate is measured.

[0008] The second station is a "process station," which is a high-efficiency (HE) process where a mixture of oil and abrasive particles flows through the injection holes to round off the inlet. This rounding is desired to occur up to a defined target value "X" for the hydraulic flow rate. An "in-process measurement" is taken here to measure the hydraulic flow rate of the abrasive mixture. Once the target value is reached, the valve for the abrasive medium is closed.

[0009] The third station is a rinsing station to remove abrasive particles from the nozzle body. No input or output data is measured.

[0010] The fourth station is a hydraulic flow rate testing station, referred to as the "post-measurement station." At this station, the hydraulic flow rate is tested after the HE process. The distribution of measurements at this station is a measure of HE process stability, as mentioned throughout this document.

[0011] The deviations of an individual part from the target value are calculated by subtracting the remeasurement test value (hydraulic flow rate (Qclose)) from the test value (Qtest) at the fourth station from the target adjustment value of the hydraulic flow rate (Qclose) at the second station, which is formulated as follows: “Δ = Qtest - Qclose”, as in Fig. 3 shown.

[0012] Fig.Figure 2 illustrates a graph showing the distribution of hydraulic flow measurements at a fourth station, consistent with test results at a fourth station in each actual and corrected operating state of the machine. Although the HE machine exhibits very accurate completion behavior at the second station, we can observe an "initial (i)" distribution at the fourth station in the actual operating state of the machine. Here, if a partial flow rate exceeds an upper limit "UL," the product is defined as reject "S." If the flow rate is lower than a defined lower limit "LL," which is a recursion limit "R," then the product has a single additional chance to undergo HE rounding again. Between these limits, the part is OK; the HE rounding process is within the desired limits.The reject rate "S" is a direct loss, and the recursion rate "R" is also a loss, as it consumes the service life of the abrasive medium and the capacity of the HE machine. Qhyd reject rates (hydraulic flow rate reject rates) are highlighted in the Pareto defect cost of nozzle production and amount to approximately 8% of the defect cost.

[0013] Reference may be made to document US2005096757 (A1) as a prior art publication in a narrower technical field related to the invention. The document discloses a method and device for detecting faults in components of a continuous process, such as a steam generator. A model of the process is developed using a modeling technique, such as advanced pattern recognition, and the model is used to generate predicted values ​​for a predetermined number of process operating parameters. Statistical process control methods are used to determine whether the difference between the predicted and actual measured values ​​for one or more of the parameters exceeds a configured statistical limit. A set of rules is used to indicate an actual or probable fault in a component of the continuous process. Description of the invention

[0014] The purpose of the invention is to improve process stability, i.e., to reduce the standard deviation of test results during the remeasurement process of the products.

[0015] The present invention proposes a method for controlling a process with at least upstream and downstream measuring stations. The method comprises the following steps: - Collection; of pre-measurement data prior to the process on a workpiece at a pre-measurement station, which is a first station; of process data after processing the workpiece at a process station, which is a second station; and of post-measurement data as test values ​​at a post-measurement station, which is a fourth station; and storage of the data in a database in accordance with the type of workpiece. - Defining the output as "Δ", where "Δ = post-measurement test value - target adjustment value", and defining the inputs as collected data from the pre-measurement station and the process station, - AI modeling of “Δ” using machine learning algorithms with a software program that uses the input data, - Adjusting the target adjustment value of the process station by applying the steps for: o Sending input data of the workpiece to a central computer while the workpiece is in the process station, o Forming a value “Δ Vorhers “per workpiece within a process time at the process station (the second station) using the input data in a selected machine learning model, o Sending the value “Δ Vorhers “to the machine, Applying the target adjustment value as "X - Δ" Vorhers“ at the process station to achieve the desired “X” value at the downstream measuring station, in a corrected operating state of the machine. ◯

[0016] Therefore, achieving the target "X" value is ensured in the corrected operating state of the machine. Process stability is therefore achieved by controlling the test results rather than the process target. A significant reduction in the deviation value is achieved, which leads to a reduction in the scrap rate and a reduction in the recursion rate.

[0017] In one possible method of the present invention, the method comprises the step of: measuring via sensors at upstream and downstream measuring stations and at the process station, where measured values ​​are acquired by these sensors. Thus, the measurement process is carried out simply and with sensitivity, and the measured values ​​are used as input and output data for calculating the predicted delta "Δ". Vorhers " used.

[0018] In another possible method of the present invention, the method comprises the step of separately applying the target adjustment value as “X - Δ Vorhers “For each product type, i.e., a fuel injection nozzle. Therefore, it is possible to implement the process control method for different nozzle types.”

[0019] In another possible method of the present invention, the method comprises the step of: processing, using a single type of machine learning algorithm (ML) suitable for each product type, and applying a single ML algorithm to each product type, i.e., a fuel injection nozzle. Therefore, the use of only a single type of ML algorithm for the respective product type is provided.

[0020] In another possible method of the present invention, the inputs are the test results from the first station before the machining process, i.e., before the HE process, and data from the second station at the start of the process, i.e., flow rate, pressure, and temperature, in a real serial operating state of the machine. Therefore, the accuracy of the model can be calculated.

[0021] In another possible method of the present invention, the outputs are a deviation from the target, defined as "Δ", which is determined at the fourth station after the test value has been measured in a real serial operating state of the machine. Therefore, the accuracy of the model can be calculated.

[0022] In another possible method of the present invention, the method comprises the step of: applying the HE process to a hole of a product at a second station, which is the process station, with a target adjustment value in accordance with the value “X - Δ Vorhers “, in a corrected operating state of the machine. Therefore, the desired ideal value can be measured at the fourth station, at the end of the HE rounding process, or any process that includes pre- and post-measurements.” Brief description of the drawings Fig. Figure 1 illustrates stations of an HE machine according to the present invention. Fig. Figure 2 illustrates a graph showing a distribution in accordance with test results at a fourth station in each actual and corrected operating state of the machine. Fig. Figure 3 illustrates the stages of a HE machine according to the state of the art.

[0023] Table 1 shows ideal work values, actual work values ​​and corrected work values ​​at each upstream and downstream measuring station. Detailed description of the invention

[0024] The present invention proposes a method comprising machine learning algorithms that are applied to control the test results rather than the process goal. This is applied to a hydroerosive rounding machine with integrated pre- and post-measuring stations, which are a first station (1) and a fourth station (4), as shown in Fig. Figure 1 illustrates this. This type of process control can also be applied to other processes where part / workpiece data can be collected on the workpiece before and after the process.

[0025] Table 1 shows how the present invention works. Ideally, the goal is to obtain a flow rate (Qhyd, Qtest “Qt” value) equal to the final flow rate (Qschliessen “Qc” value), shown as the “X” value in the first row of Table 1. However, in an actual serial operating state of the HE machine, there is a deviation, “Δ”. “Δ” is defined as a difference between the closing value (Qschliessen (Qc)) and the test result value (Qtest (Qt)). The present invention aims to control the output at the fourth station (4) by adjusting at the second station (2) through predictions of “Δ” in step A and subsequent adjustment of the target adjustment value (Qschliessen values) in step B while the workpiece is still in the second station (2).

[0026] The invention comprises a method for predicting this deviation value, which is predicted as “Δ before the end of the process: Δ Vorhers“ is shown. For this reason, the value “Δ Vorhers “ adjusted at the second station (2) to obtain the ideal value “X - Δ Vorhers + Δ“ at the fourth station (4). In other words, a standard HE machine controls the process at the second station (2), which is the process station, according to the prior art. But an intelligent (HE) machine controls the output at the fourth station (4), which is the post-measurement station, by predicting the result and implementing an adjustment of the process within the machining time at the second station (2), in other words, at the process station, according to the present invention. This invention uses data from the beginning of the process and is much more part-specific and challenging. To understand how this process works under serial conditions, this infrastructure is tested and built:

[0027] The HE machines are data source zones, and transmission information is extracted from the machine and converted into a message containing the part's identity, machine number, and type number. This data reaches edge computing via a computing program at a landing zone, i.e., the "Kafka system." On an artificial intelligence (AI) platform, the predicted value (Δ Vorhers ) calculated and sent back to the machine in a message format with machine number and part ID.

[0028] Historical data is collected by the remeasuring station, which is the fourth station (4), and stored for each workpiece type, i.e., nozzle type. For example, these nozzle types have codings such as A, B, C, D, etc. Historical data for each nozzle type is stored on the computer server by coding the nozzle types. Sensors are used in the data source zone of the fourth station (4). This is referred to as "big data collection" because a huge number of test results are stored in the databases. In fact, the HE machines were already connected to servers and data was being collected, but there was still a need to organize the data; some data was collected but not stored in the database, or some additional possibilities were being evaluated, and so on.

[0029] In the second step, in the integration zone, a data cleansing and preparation phase is implemented before the machine learning (ML) algorithms are executed. This is a critical step before the ML algorithms begin training. However, expertise in the state-of-the-art HE rounding process, within the actual operating conditions of the HE rounding machine, proved to be a real strength in this step.

[0030] Following this step, the next step is "AI modeling," which employs several machine learning algorithms. Improvements are needed in two main areas: firstly, the coding language, and secondly, machine learning and statistical aspects. Afterward, it was necessary to develop the infrastructure and update the HE machine's software to adapt it to this new operating mode.

[0031] Referencing historical data for each nozzle type, machine learning algorithms (ML) are executed to define the optimal model representing the relationship between process inputs and output. The output is the deviation from the target value "X," defined as "Δ"; and the inputs are the test results from the first station (1) and the data from the second station (2) at the start of the process (flow rate, pressure, temperature, etc.). Once the optimal model for a nozzle type (for a workpiece) is defined, it is used in serial production to predict the deviation from the target within the process time and the final target; in other words, a target adjustment value (X - Δ) is determined. Vorhers ) determined with regard to the prediction for each individual nozzle.

[0032] In the diagram in Fig.Figure 2 shows actual distributions for a nozzle type. The initial distribution (i) shows the state with standard HE, and the target distribution (t) shows the state with the intelligent HE process. Here, in accordance with a nozzle test result, we have a reduction in the standard deviation of the hydraulic flow rate measured at the fourth station (4) from 5.8 to 3.9 mm. 3 / minute, which corresponds to a 30 percent improvement in standard deviation, a 40 percent reduction in rejection rate and a 50 percent reduction in recursion rate. Table 1 Operating state Target adjustment value (Q close) Test value (Qtest) State of the art ideal X X real X X+Δ invention corrected X-D Vorhers X-D Vorhers +D

[0033] The adjustment of “Δ VorhersThis can be explained by an example according to Table 1. In the actual operating state of the HE machine, the inputs from the first station (1) and the second station (2) are collected, and the result Qtest (Qt) (test values) from the fourth station (4) is collected. The “AI” is formed in accordance with the defined machine learning algorithm, the process is implemented, and the correct value “Δ” is determined. Vorhers “ is found for the defined nozzle type. The HE rounding machine is used in accordance with the formulation for subtracting the value “Δ Vorhers “Adjusted from the desired target value “X” as shown in Table 1, in the corrected operating state of the machine. In the example, “X” is 5 mm 3 / Minute Qhyd value and the value “Δ Vorhers “ is 3; then Qclose (Qc) at the second station (2) will be “X - Δ Vorhers “adjusted. And the test result at the fourth station (4) “X - Δ Vorhers+ Δ" is achieved as in the formula; 5 - 3 + 3 = 5. Therefore, the desired ideal value "5" is achieved at the end of the process. The same idea can be applied to any other process that includes before and after measurements. List of reference symbols 1 First stop 2 Second Station 3 Third Station 4 Fourth Station R Recursion S Committee i initially t goal LL lower limit UL upper limit Qt Qtest (test value): Test result for the hydraulic flow rate at the fourth station Qc Qclose (target adjustment value): hydraulic flow rate when closing the valve at the second station Step A: Prediction of Δ Step B Adjusting QC QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] US 2005096757 (A1

[0013]

Claims

[1] Method for controlling a process with at least upstream and downstream measuring stations, wherein the method comprises steps for: - Collecting pre-process measurement data on a workpiece at a pre-measurement station, which is a first station (1), process data after processing the workpiece at a process station, which is a second station (2), and post-measurement data as test values ​​at a post-measurement station, which is a fourth station (4), and storing the data in a database in accordance with the type of workpiece, - Defining the output as "Δ", where "Δ = post-measurement test value - target adjustment value", and defining the inputs as collected data from the pre-measurement station and the process station, - AI modeling of “Δ” using machine learning algorithms with a software program that uses the input data, - Adjusting the target adjustment value of the process station by applying the steps for: o Sending input data of the workpiece to a central computer while the workpiece is in the process station, o Forming a value “Δ Vorhers “ per workpiece within a process time at the process station using the input data in a selected machine learning model, o Sending the value “Δ Vorhers “to the machine, Applying the target adjustment value as "X - Δ" Vorhers “ at the process station to achieve the desired “X” value at the post-measurement station, in a corrected operating state of the machine. [2] The method of claim 1, wherein the method comprises the step of: Measurements are taken via sensors at upstream and downstream measuring stations and at the process station, where measured values ​​are recorded by these sensors. [3] Device according to claim 1 or 2, wherein the method comprises the step of: separately applying the target adjustment value as “X - Δ Vorhers “for each product type, i.e., nozzle of a fuel injection system.” [4] The method of claim 3, wherein the method comprises the step of: Processing, using a single type of machine learning algorithm (ML) suitable for each product type, and Applying a single ML algorithm to each product type, i.e., a fuel injection nozzle. [5] Method according to one of the claims, wherein the inputs are the test results from the first station (1) before the machining process, i.e. before an HE process, and data from the second station (2) at the start of the process, i.e. flow rate, pressure and temperature, in a real serial operating state of the machine. [6] Method according to one of the claims, wherein the outputs are a deviation from the target, which are defined as “Δ” and are defined at the fourth station (4) after the test value has been measured in a real serial operating state of the machine. [7] Method according to one of the claims, wherein the method comprises the step of: applying the HE process to a hole of a product at a second station (2), which is the process station, with a target adjustment value in accordance with the value “X - Δ Vorhers “, in a corrected operating state of the machine.

Images