Predictive Modeling Tools

Abstract

The present invention discloses a system and method for simulating and optimizing industrial and other processes, the system and method including a computer that performs multivariate analysis on input variables and output variables to generate a data model of the operation of the process. For industrial applications, the input variables include process variables, and the output variables include result variables from the operation of the industrial process. The data model determines the contribution of the corresponding input variables or process variables to the change of the output variable or result variable, and is provided to a prediction algorithm to identify the parameter values ​​of the input variables or process variables that are expected to have the most significant impact on the output variable or result variable during the execution of the process. The output of the prediction algorithm is a parameter value, which is provided to the industrial process as an input variable or process variable for simulation or performance optimization or product recommendation / optimization.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This patent application claims the benefit of priority to U.S. Provisional Patent Application 62 / 713,025, filed by Maiken Givot et al. on August 1, 2018, and entitled “Predictive Modeling Tool,” which is hereby incorporated by reference in its entirety. Technical Field

[0003] The present application relates to a predictive modeling tool that uses multivariate statistical methods and predictive models to drive iterative machine learning applications for simulation and continuous optimization of industrial engineering and other applications. Background Art

[0004] For more than 25 years, spreadsheet-based modeling has been used to describe and report results in industrial applications. Such spreadsheets are often used to record actual performance results and then calculate cost / value estimates based on cost increases or decreases. Reports are generated by individuals and are often stored on local servers. However, it has been found that such methods have low predictive or optimization capabilities and are not particularly useful in obtaining reliable conclusions.

[0005] For example, in an exemplary industrial engineering application for roller mills, technical service support for roller mill customers was done in spreadsheet files and stored in a database. It was difficult to get an objective overview of the collected data and results. Often, reports were not presented until after the trial via email.

[0006] This limitation is unfortunate for the abrasive bonded products industry, which includes, for example, custom engineered products where a grinding wheel with 10% higher performance provides high value in a production line. The limiting factor in such industrial systems is processing and forming a logical structure that can be understood and recognized by the product manufacturer and customer. In addition, it takes three to five years to become an experienced industrial application engineer. If a key industrial application engineer leaves the company, the performance information and operational knowledge learned on the job is usually lost as well.

[0007] Traditionally, physical simulations or models have been used to simulate products. By fine-tuning the formula coefficients, key performance indicators (KPIs) and settings of these models, data models that describe the current knowledge in detail can be created. However, physical modeling requires users to define formulas that usually describe core process relationships via engineering indicators and / or key performance indicators. The negative aspect of this method is that only known relationships can be predicted / simulated. It is also difficult to separate outliers from new "learning" points (i.e., containing new information). This method requires companies to have formula and coefficient experts who actively maintain the generated physical models. In addition, although such physical models often allow the available knowledge to be spread from knowledge personnel to a wider group, such models do not have predictive capabilities.

[0008] Statistical methods have been used in the pharmaceutical industry since 1978 (Astra Sweden) and is currently approved and supported by the FDA. The chemical process industry in Sweden began using statistical selection methods in 1985. For industrial applications of centerless, center-type and internal grinding, Tyrolit reportedly has tools for supporting optimization at customer sites for recommended grinding tools for cylindrical grinding applications and recommended machine settings. Tyrolit has a product selector that uses a logical "tree" to select its products. However, to date, statistical machine learning type applications have not been used for such industrial applications that allow industrial applications to be simulated and continuously optimized.

[0009] A technical solution is desired that enables companies to maintain and protect their performance data while using the performance data to provide reliable predictive solutions to the world, independent of the presence of application engineers or sales representatives. It is also desirable to create a logical structure that captures the performance data created during the industrial process to enable simulation and continuous optimization of the industrial process at each customer site. Summary of the invention

[0010] Various examples will now be described to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.

[0011] The systems and methods described herein combine modern statistical analysis with experience in development, application engineering and sales of industrial products to develop optimization methods and web tools to which new models or applications can be added in a very simple manner without changing the web interface coding. Multivariate statistical methods are used to increase the understanding of the underlying industrial processes and facilitate communication across functional boundaries. In an exemplary embodiment, the systems and methods described herein support any iterative / machine learning application to provide recommended machine parameters.

[0012] In an exemplary embodiment, a computer-implemented method for implementing predictive modeling of industrial applications and related processing devices is provided. The method includes a computer performing multivariate analysis on input variables and output variables generated during the operation of an industrial process to generate a data model of the operation of the industrial process. The input variables may include process variables, and the output variables may include results from the operation of the industrial process. The resulting data model represents the contribution of the corresponding input variables or process variables to the change in the output variable. The data model is provided to a prediction algorithm to identify parameter values ​​of the input variables or process variables that are expected to have the most significant impact on the selected output variable during the execution of the industrial process. The prediction algorithm outputs parameter values, which are provided to the industrial process as input variables or process variables to optimize the selected output variable.

[0013] In an exemplary embodiment, a computer generates a data model by automatically generating a design of experiments (DoE) for the input variable space of the industrial process, wherein the design of experiments is processed by the data model and results are stored based on preferred results. A prediction algorithm predicts a predetermined number of best next trials for the DoE, and outputs parameter values ​​for the predetermined number of best next trials and the probability of an improved result using the parameter values.

[0014] In other exemplary embodiments, the parameter values ​​and the output variables or result variables generated by the industrial process in response to the parameter values ​​are provided as feedback data to the computer for adjustment of the data model. In an exemplary embodiment, the parameter values ​​can be output to a display as a simulation of the output of the industrial process in response to the specified input variables. In addition, the method may include recommending products with parameter values ​​that optimize the selected output variables in the industrial process. The recommended products may optimize the selected output variables in the industrial process for at least one of a specified machine, a specified substrate, and a specified application. In other exemplary embodiments, the method may include identifying products in the product portfolio that have overlapping parameter values ​​that optimize the same selected output variables, and / or identifying gaps in the product portfolio by identifying parameter values ​​for optimizing the selected output variables that are not represented in the product portfolio.

[0015] The statistical methods described herein can be applied to a plurality of industrial processes that can be modeled using the DoE approach. For example, the industrial process can be a grinding machine operation, wherein the input variables include machine settings of the grinding machine, and the output variables depend on the type of grinding process and include at least one of G-ratio, material removal rate of the grinding wheel, chip thickness, number of workpieces per dressing cycle, and surface finish. In such an embodiment, the machine settings include at least one of grinding wheel speed, roller speed, traverse speed, continuous and end infeed, grinding time, feed rate, shift, dressing, dressing infeed, infeed, and overlap ratio.

[0016] In another example, industrial process can be a product selection process, such as adhesive selection process. In this example, input variables include selection variables for adhesive or tape, and output variables include selection or recommendation to at least one adhesive or tape. In an exemplary embodiment, input variables include adhesive physical properties, adhesive thermal properties, adhesive electrical properties, adhesive curing properties, adhesive performance properties, adhesive durability properties, adhesive chemical resistance properties, adhesive rheological properties, adhesive viscosity, adhesive setting time, adhesive elastic modulus, adhesive solvent resistance, adhesive composition, adhesive distribution properties, adhesive use requirements, standardized testing or certification, environmental parameters, health parameters, safety parameters, carrier properties, backing properties, liner properties and at least one of the materials to be bonded by adhesive or tape. On the other hand, output variables include at least one of tensile strength, peel strength value, adhesive name, adhesive structural properties, adhesive performance properties, quantification of lamination quality and purchase information. Adhesive can be a pressure-sensitive adhesive with or without additional adhesive layer or non-adhesive layer.

[0017] In other exemplary embodiments, the industrial process may be a grinding operation that includes recommending a product that optimizes selected output variables in the industrial process for at least one of a specified machine, a specified substrate, and a specified application for an abrasive belt.

[0018] In other exemplary embodiments, a method for implementing predictive modeling is provided, wherein a computer performs multivariate analysis on input variables related to the characteristics of a product and output variables related to the performance of the product to generate a data model of the product. In an exemplary embodiment, the data model represents the contribution of the corresponding input variables to the change in the output variable. The data model is provided to a predictive algorithm to identify parameter values ​​of the input variables that are expected to have the most significant impact on the selected output variable. The predictive algorithm outputs the parameter values, and the parameter values ​​are provided as input variables to optimize the selected output variable. Products that optimize the selected output variable for the specified input variables are recommended.

[0019] In further exemplary embodiments, the method includes identifying products in the product portfolio having overlapping parameter values ​​that optimize the same selected output variables and / or identifying gaps in the product portfolio by identifying parameter values ​​for optimizing selected output variables that are not represented in the product portfolio.

[0020] In other exemplary embodiments, input variables include characteristics of a grinder, a bonded abrasive grinding wheel, an abrasive belt, a coated abrasive belt or disc, a nonwoven abrasive, a bristle brush, an abrasive article mounted on a robot, an adhesive, and a safety harness. The input variables may also include process variables of an industrial process, and the output variables include results from the operation of the industrial process. In such embodiments, the data model represents the contribution of the corresponding input variables or process variables to the change in the output variables. Parameter values ​​are provided to the industrial process as input variables or process variables to optimize the selected output variables.

[0021] Additional embodiments include a system for implementing predictive modeling to optimize an industrial process, the system comprising at least one processor and a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform operations including: performing multivariate analysis on input variables and output variables generated during operation of the industrial process to generate a data model of the operation of the industrial process, the input variables comprising process variables, and the output variables comprising results from the operation of the industrial process, and the data model representing contributions of the respective input variables to changes in the output variables; automatically generating a design of experiments (DoE) for an input variable space for the industrial process; processing the design of experiments using the data model; providing results of processing the design of experiments to a predictive algorithm based on preferred outcomes; the predictive algorithm identifying parameter values ​​of the input variables that are expected to have the most significant impact on selected output variables during execution of the industrial process, making predictions for a predetermined number of best next trials for the DoE, and outputting parameter values ​​for a predetermined number of best next trials and a probability of improved results using the parameter values; and providing the parameter values ​​as input variables to the industrial process to optimize the selected output variables.

[0022] In such embodiments, the at least one processor further performs operations including feeding the parameter values ​​and output variables generated by the industrial process in response to the parameter values ​​as feedback data to the at least one processor and adjusting the data model using the feedback data. The at least one processor may also perform operations including outputting the parameter values ​​to a display as a simulation of the output of the industrial process in response to the specified input variables.

[0023] In other embodiments, the at least one processor may also perform operations including recommending products having parameter values ​​that optimize selected output variables in the industrial process and recommending products that optimize the selected output variables in the industrial process for at least one of a specified machine, a specified substrate, and a specified application. The at least one processor may also perform operations including identifying products in the product portfolio having overlapping parameter values ​​that optimize the same selected output variables and / or identifying gaps in the product portfolio by identifying parameter values ​​for optimizing the selected output variables that are not represented in the product portfolio.

[0024] Any of the foregoing examples may be combined with any one or more of the other foregoing examples to form new implementations within the scope of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In the drawings, which are not necessarily drawn to scale, like numerals may describe similar components in different views.The drawings are generally shown by way of example, but not limitation, of the various embodiments discussed in this document.

[0026] Figure 1 An exemplary embodiment of a grinding application is shown in which sensors provide feedback data that is processed by a machine learning tool to generate simulation and continuous optimization data.

[0027] Figure 2 A DModX plot showing the distances of observations in the training set from the input variables in the X model plane or hyperplane.

[0028] Figure 3A Graph showing continuous infeed versus tip infeed for grinding wheel L8, wherein the labeled area represents a G-Ratio of 4.0, wherein the X-axis is continuous infeed and the Y-axis is tip infeed.

[0029] Figure 3B Graph showing continuous infeed versus end infeed for grinding wheel M10, where the labeled area represents a G-ratio of 3.2, with the X-axis being continuous infeed and the Y-axis being end infeed.

[0030] Figure 4A A graph showing how the characteristics of two products, LB-UW and EXL-Pro 8, differ due to the shapes of the two sets of product characteristics.

[0031] Figure 4B To show Figure 4A A diagram showing a view of the variables of the product.

[0032] FIG. 4C to FIG. 4E is a table showing test data for a grinding application.

[0033] Figure 5is a flow chart illustrating a predictive modeling process in an exemplary embodiment.

[0034] Figure 6 An exemplary output display showing predicted values ​​for parameters of interest in a particular industrial application is shown.

[0035] Fig. 7A Graphs showing PCA analysis of product properties of adhesives and tapes.

[0036] Figure 7B A score graph showing the relationship between different rows or products.

[0037] Figure 8 A score plot that shows the possible presence of outliers, groups, similarities, and other patterns in the data.

[0038] Fig. 9 It is a pq-graph, which is the superposition of the p-graph and q-graph of the first prediction component and the second prediction component of the orthogonal projection to latent structure (OPLS) / bidirectional orthogonal PLS (O2PLS) model.

[0039] Fig.10 is a normal probability plot of the residuals, where the residuals are standardized on a double logarithmic scale.

[0040] Fig.11 Shown is a handheld device for measuring the surface free energy of a surface of a substrate.

[0041] Fig.12 Shown are the FTIR spectra of two polypropylene samples.

[0042] Fig.13 A workflow for applying predictive modeling techniques to an adhesive product selection process in an exemplary embodiment is shown.

[0043] Fig.14 is a graph showing the average measured peel force versus the average predicted peel force for various substrate / adhesive combinations.

[0044] Fig.15 A diagram showing an exemplary abrasive belt by application—substrate—machine in an exemplary embodiment.

[0045] FIG. 16A to FIG. 16B To show Fig.15 The information collected in the load diagram is converted into a load diagram of recommended products for different machines / substrates, where Fig. 16B for Fig.16A The enlarged format.

[0046] FIG. 17A to FIG. 17B To show the Fig.15A graph of qualitative information that is converted into a numerical "grade" indicating, on a scale of 1-12, the suitability of a given machine / substrate for a particular application.

[0047] Fig.18 is a graph showing observed versus predicted belt application for a robotic application.

[0048] Fig.19 is a graph showing observed versus predicted belt application by machine (and pressure).

[0049] FIG. 20A to FIG. 20B is a graph showing the refined numerical ranking of the refined data (taking pressure into account).

[0050] Fig.21 A table showing performance ratings of robotic machine applications and products.

[0051] Fig. 22 A graph showing observed results versus predicted results for a specific product-application-machine / pressure to provide a corresponding performance rating.

[0052] Fig.23 A table showing modeled ratings based on numerical ratings is shown, and that information from each predicted data point can be viewed.

[0053] Fig.24 A table containing a product list of harness data containing potentially hundreds of products with potentially thousands of stock keeping units (SKUs) worldwide for provision to a predictive modeling tool to provide product portfolio rationalization.

[0054] Fig.25 is a graph generated using multivariate statistical analysis tools. Fig.24 Figure 2. A graph of a preliminary predictive model of the beam data. The multivariate statistical analysis tool plots the correlations between the collected parameters.

[0055] Fig.26 A general purpose computer is shown that can be programmed as a special purpose computer suitable for implementing one or more embodiments of the system described in the exemplary embodiments. DETAILED DESCRIPTION

[0056] It should be understood from the outset that although the following provides an exemplary implementation of one or more embodiments, reference is made to Figures 1 to 26The disclosed systems and / or methods of the described invention may be implemented using any number of techniques, whether currently known or in existence. The present disclosure should in no way be limited to the exemplary implementations, drawings, and techniques shown below, including the exemplary designs and implementations shown and described herein, but may be modified within the scope of the appended claims along with their full scope of equivalents.

[0057] The methods described below can be used for a variety of industrial processes or industrial engineering applications, rapid pricing of complex engineered products used in such industrial processes or industrial engineering applications, medical treatments, etc. As used herein, "industrial processes" or "industrial engineering applications" are intended to cover a variety of technical field applications, including resin bonded wheels (such as roller grinding, tool or groove grinding, end grinding, orbital grinding wheels, hot press wheels, portable or mounted cutting wheels, center recessed grinding wheels, cutting and grinding wheels, flexible wheels, cup wheels, coarse grinding wheels), vitrified bonded grinding wheels (such as gear grinding, cylindrical (ID / OD / centerless, cam / crank grinding wheels, surface / creep feed grinding wheels, end grinding wheels), superabrasive grinding wheels (for tool grinding, cam / crank grinding, surface and creep feed grinding, cylindrical ID / OD, profile, turbine and gear grinding), coated abrasive belts or discs, nonwoven abrasives, bristle brushes, abrasive articles mounted on robots, adhesive options, and even for more diverse industrial products, such as safety harnesses Products, smart filters, product recommenders and selectors, product portfolio rationalization systems for determining which overlapping product offerings to merge, applications for determining research and development in the areas of targeted research, product price calculators (including value calculators), training, etc. The samples described herein include simulation and / or optimization of grinding applications and an interactive modeling process for selecting the correct adhesive for a specific application by collecting test data related to input variables and materials to be bonded by the adhesive and measurable output variables. Other samples described herein include a product recommender for abrasive belts, and a product portfolio optimizer and product recommender for safety harness products. The simulation, optimization, product selection, product recommendation, and product portfolio optimization features can be applied to any suitable industrial process within the scope of the present disclosure.

[0058] As will be apparent from the following detailed description, the predictive modeling tools described herein address the above and other technical shortcomings in the prior art by providing application engineers with a simple software application on a smart phone or handheld computer (e.g., iPad) that enables them to simultaneously change multiple machine parameters and work with customers to achieve desired results (e.g., performance optimization). The predictive modeling tools described herein also enable equipment suppliers to build and maintain knowledge about the operation of their equipment and achieve maximum product performance.

[0059] Industrial Engineering Application Example – Grinding Tools

[0060] Figure 1 An exemplary embodiment of a grinding application is shown in which sensors provide feedback data that is processed by a machine learning tool to generate simulation and continuous optimization data. In this embodiment, a substrate 10 is ground by a grinding wheel 12 under the control of a servo motor 14. In turn, the servo motor 14 is controlled by a main controller 16, which automates the operation of the servo motor 14 and the grinding wheel 12. During operation, a sensor 18 (e.g., speed, torque, environment, temperature, pressure, rotation, vibration, imaging sensor, etc.) provides feedback about the operation of the grinding wheel 12 to an auxiliary controller 20, which can provide feedback control signals to the servo motor 14 and / or the main controller 16 as appropriate to adjust the operation of the grinding operation based on the sensor feedback. In an exemplary embodiment, the feedback sensor data is stored in a cloud storage device 22 during operation and provided to a machine learning tool 24, which learns the operation of the grinding wheel 12 under different conditions. As will be explained in more detail below, the machine learning tool 24 learns the optimal operating conditions of the grinding wheel 12 for different operating conditions and wheel specifications, and provides corresponding parameters to the auxiliary controller 20 and / or the main controller 16 to provide continuously optimized control during operation of the grinding wheel 12. In addition, the stored operating parameters can be used to simulate the operation of the grinding wheel 12 on the display 26 in response to specified inputs. The operating parameters can also be provided to an application or web page that can be accessed using a user's (e.g., an application engineer) smartphone or handheld computer 28 to help the user set the control parameters of the grinding wheel 12.

[0061] Those skilled in the art of machine learning and artificial intelligence (AI) applications will know that the machine learning tool learns the optimized settings that evolve over time and continually improve grinding performance. Secure application data from all completed tests and operations is stored in the cloud data storage device 22 to keep a continuous record of the continuous optimization performed by the machine learning tool 24. It should also be understood that the stored simulations make product training much easier and significantly shorter for new sales representatives and application engineers, as well as machine builders or end-user customers, for their own product setup and use. Such simulations also help application engineers and / or product developers develop grinding tools into finished, usable applications based on the simulations.

[0062] The test data used to “train” the machine learning tool 24 was obtained by performing multiple tests of the grinding equipment. In the test, a master data file containing 98 grinding tests was used to generate a predictive data model, and the grinding tests were used in 29 different test occasions using a 3M TM 3M Precision Shaped Granules (PSG) TM Cubitron TMII 19 Examples of Resin Bonded Roller Grinding Wheels 83 data points and 19 variables were used in the data model with a predictive ability of Q' = 0.63, indicating a good level of prediction.

[0063] A web application with an adaptable / modifiable interface has also been created to load new grinding models into the user's smartphone 28, providing a very quick and easy process to add new grinding applications. Although the model only contains 3M based TM Cubitron TM II resin bonded roller grinding wheel, but the validation of the model results was performed for any wheel containing crushed abrasive. Five test cases showed improvements of + / -0, +20% and 3x>30% compared to the existing settings, illustrating the robustness and flexibility of the modeling tool. In addition, as described below, an optimization tool has been developed that supports changing multiple variables simultaneously. The optimization tool significantly speeds up the optimization process. Such rapid optimization allows moving from the operating point in the direction of the optimal machine settings defined by the model and calibration data set, which reduces the risk of damaging the product or machine, increases customer value with each step, and provides a safer and more stable working process. Those skilled in the art will know that the optimal point is personalized for each machine / customer and may change over time. The optimal point is reached by gradually approaching or passing the limits of the parameter settings in the machine.

[0064] The collected data can also be used to help users of the grinding wheel 12 verify their generated data by creating a score map to locate the user relative to the user database in the cloud data storage device 22 and identify the most relevant test reports. The machine learning tool 24 or other processing element in the cloud data storage device 22 can then verify the correctness of the maximum or best possible performance of the user-provided data or settings based on the user-provided data, and use the distance to the model map 30 to evaluate the degree of "stretch" in the user data relative to the user database, such as Figure 2 shown. Figure 2 A DModX plot showing the distance of observations in the training set from the input variables in the X model plane or hyperplane. DModX is proportional to the residual standard deviation (RSD) of the observations of the input variables (X). By default, DModX is displayed in normalized units, i.e., the absolute DModX divided by the pooled residual standard deviation of the model. The critical value of DModX (denoted as Dcrit) is calculated from the F distribution, also known as Snedecor's F distribution or Fisher–Snedecor distribution, which is a continuous probability distribution in probability theory and statistics. Dcrit regulates the size of the envelope around the data points of the training set.

[0065] Additionally, based on user-supplied data, performance can be simulated and test wheel specifications (i.e., hardness, structure, and 3M TM Precision Shaped Grains (PSG) percentage, G ratio). For example, Figure 3A is a diagram showing grinding wheel L8, where region 32 represents a G ratio of 4.0, where the X axis is continuous infeed and the Y axis is end infeed, and Figure 3B A graph is shown for grinding wheel M10, where region 34 represents a G ratio of 3.2. The output thus provides a defined test wheel specification. It will be appreciated by those skilled in the art that the G ratio target is ideally as high as possible to obtain maximum performance.

[0066] Once the grinding wheel parameters are selected, a user reference run may be run with the test grinding wheel. The generated data is saved to the global database in the temporary test space and cloud data storage device 22 by selecting "Save Data Points" and entering the X (input) and Y (result) parameters. Data points for multiple competing grinding wheels may be saved to generate a global database by selecting "Save Data Points" and manually entering the X and Y parameters. In an exemplary embodiment, the variables that may be varied in the optimization are selected (i.e., wheel speed, roller speed, traverse speed, continuous infeed, and end infeed). The result variables to be optimized are also selected (i.e., G ratio and removal rate of the grinding wheel). The selected variables are then provided to the machine learning tool 24 to generate performance predictions. In an exemplary embodiment, the machine learning tool 24 automatically generates a face-centered cubic (FCC) design of experiments (DoE) (where experiments are not just in the corners and the middle, but also in the middle of the face), makes predictions, and presents three best next trials based on the results and the probability of improved results using the parameter values ​​in the results. A test is then selected and the test is rerun. Data points saved during a rerun are saved and transferred to a Value Calculation Template, and the results of the value calculation are viewed, edited, and submitted to the user as a test report.

[0067] Typically, DoE is used to minimize the number of experiments and maximize information output in order to minimize the workload. Under ideal conditions, the contribution of 256 variables from 16 experiments can be understood. In an exemplary embodiment, the experiment is not performed, but a simulation is run to calculate the results in different "experimental points". DoE is used to maximize the efficiency of searching for the best point in the n-dimensional operating range. The more points and the better the design, the higher the accuracy. In an exemplary embodiment, a 3-level full factor is used, but other experimental designs may also be used. In an even more extensive design, more data points than a 3-level full factor may be used, wherein experiments are also added in the middle of all edges.

[0068] By collecting grinding data in this manner, the machine learning tool 24 can use statistical tools such as multivariate analysis and partial least squares (PLS) / orthogonal projections to latent structures (OPLS) analysis to extract information from historical grinding data (or design of experiments (DoE) - test grinding) to create a model that describes the variable contributions to performance with good predictive power. For example, in the case of roller grinding, there are 63 variables available in the original data, and 19 of these variables carry information that helps predict performance, in this case, the grinding ratio of the grinding wheel (G ratio = grinding ratio, defined as the volume of material removed per volume of wheel wear. G therefore increases with decreasing wheel wear and / or increasing material removal). From the 3M referenced above TM Cubitron TM The data of 98 grinding results of II resin bonded roller grinding wheel are the data in the data set. 83 of these points are qualified for use in creating a data model. The key elements in multivariate analysis are the ability to extract / separate information and noise and the treatment of qualitative variables.

[0069] To optimize the machine parameters of a roller mill, 14 variables were selected to describe the machine, roller or workpiece, and grinding wheel. 5 variables were selected to describe the machine setup. A design window was defined using a 3-level full factorial DoE for the 5 selected machine parameters. The model was then used to predict the performance of the 243 resulting machine setup alternatives. The top 10-20 were then presented to the user to select the setup to use.

[0070] By storing the selected settings and including a step in which the user (e.g., application engineer) reports back actual results, new data is generated which can then be used to improve the underlying models and predictions. In this way, the machine learning tool 24 continuously improves its data and its predictive capabilities.

[0071] In conventional manual and physical optimization, a very common strategy is to change one parameter at a time and sometimes iterate to identify the next setting to evaluate. However, when changing one parameter at a time, there is a risk that local limitations may force the user to stop earlier than in the case of changing multiple parameters simultaneously. In contrast, the method described in this article has the following advantages: the model exploits all the information present in the historical data and allows or even encourages changing multiple parameters simultaneously to point to the optimal parameter set. In this way, the probability of finding an operating point with higher performance than iterative changes is significantly higher.

[0072] Already used based on 3M TM Cubitron TMThe data model of II resin bonded roller grinding wheel data was tested five times. In machine #1, no improvement could be made because the grinding wheel was too soft. Machine #2 was the subject of optimization for 5 years with no progress. When three parameters were varied simultaneously, using the machine learning tool provided a 20% improvement in G-Ratio. Machines #3, 4, and 5 all showed +30% G-Ratio improvement as well as improved cycle times.

[0073] The advantage of using machine learning tool 24 to simulate and optimize applications as described herein is that it can predict improved product use and pricing based on historical business operations, and then subsequent predictions are made to non-existing materials / variants. In many cases, industrial products such as bonded abrasive products are classified according to different safety standards such as fracture speed. Fracture tests can be used to provide data to subsequently predict the fracture speed of bonded abrasives with successful results. This is likely to reduce the need for destructive testing and automate the allocation of the maximum operating speed of bonded abrasive products. Fracture tests are usually carried out using automatic fracture test equipment, and the wheel selection for fracture tests follows regional standards, such as ANSI, EN and JIS. In order to automate such a process, it is necessary to understand the understanding of different product groups and their relationship to each other based on the difference in characteristics. For example, in bonded abrasives, such characteristics may include abrasive groups and specifications. A very simple example of a product map may include data such as existing product names, sizes, accessories, system components, forces, speeds and grinding processes used.

[0074] The resulting data can be plotted to show how products differ from one another in characteristics. For example, Figure 4A Graph showing an analysis of how the characteristics of two grinding wheel products LB-UW and EXL-Pro 8 differ due to the shapes of the two groups (LB-UW is shown at 36 and EXL-Pro 8 is shown at 38). Figure 4B To show Figure 4A The diagram shows a view of a variant 40 of the product. If considered together Figure 4A and Figure 4B From the graph of the LB-UW product, it is evident that the variable wheel speed and travel speed are higher, while the variable force is lower, for the LB-UW product. This data was found to correlate with a longer life expectancy for the LB-UW product compared to the EXL-Pro 8 product. Figure 4A and Figure 4B Together it is shown that constructive discussion of different product characteristics is greatly facilitated by utilizing the type of analysis described herein.

[0075] As another example, when a large number of products are used in the same application but on different materials or for different machine settings, the analysis described herein can be used to enable visualization to see where there are overlapping products or gaps to be filled. When launching a new product and utilizing the machine learning prediction tool described herein, it will be possible to describe very quickly which products can be replaced in the existing product portfolio and which features are truly unique to the new product. With respect to the roller mill application engineer tool data, the data can also be used to analyze and understand the customer base. For example, a score chart can be used to show customers the "mode" in their machine settings and the performance results achieved. From a variable perspective, data can be used to separate G ratios and materials that are difficult to grind or easy to grind. In addition, if a new product can be tested with selected customers in a scoring group, the new product can be launched faster and in a structured manner, it is defined in which group the product will provide the greatest benefit. As the knowledge base develops and the model improves, the machine learning tool 24 can be used to more accurately provide the correct specifications for customer needs and provide the correct and relevant machine settings. Over time, the knowledge base developed in this way will become the key to developing and improving industrial tools.

[0076] The statistical method described herein creates a general data model from the data and variables that contribute to the predictive power. As described in the above examples, the statistical model method can be used to predict grinding performance. The advantage of this method is that it automatically includes statistical analysis of values ​​and model capabilities (i.e., model confidence, error sources for outliers, identification of new objects with different characteristics, effective visualization of data, etc.). However, those skilled in the art will find it difficult to understand the above examples and the following discussion of the statistical model method. Figure 5 It should be understood that the techniques described herein can be used to generate corresponding data models from other types of industrial processes.

[0077] Prediction Model

[0078] Figure 5 A generalized predictive modeling process 42 in an exemplary embodiment is shown. As shown, the predictive modeling process 42 first enters test data into a spreadsheet such as Xcel at 44. Such test data depends on the industrial application. For example, in a grinding application, the test data may include machine type, machine age, machine power, and grinding parameters such as infeed, wheel speed, roller speed, etc. FIG. 4C to FIG. 4EAs shown. The more test data provided, the more accurate the prediction obtained can be expected. Any data of confidential nature is only evaluated internally for the model when necessary. In an exemplary embodiment, up to 30 experimental design (DoE) tests can be used to generate test data. As known to those skilled in the art of statistics, DoE is a method of changing variables in a structured manner to quantify the variable contribution to the response, and by carefully designing / selecting experiments, variable interactions, exponential relationships, and nonlinear relationships can be quantified and used to generate models with predictive power, usually using a small number of experiments. DoE tests not only involve the selection of suitable independent variables, dependent variables, and control variables, but also involve the delivery of planned experiments under statistically optimal conditions given available resource constraints. In an exemplary embodiment, such experimental settings are provided to enable the generation of suitable input variables and output variables, for generating a data model for the operation of an industrial process of interest. Typically, input variables include process variables, and output variables include result variables from the operation of an industrial process, and the resulting data model represents the contribution of the output variable or result variable to the corresponding input variable or process variable according to the experimental arrangement.

[0079] At 46, a data model is created from the collected statistical test data using the statistical techniques described in more detail below. Then at 48, the data model is provided to the model development software and model runtime engine, such as the SIMCA Q.DLL mathematical engine. The model development software and model runtime engine include a prediction algorithm that generates predictions for the specified input variables that are expected to have the most significant impact on the output variables that affect the performance of the grinding application. In operation, the model development software and model runtime engine 48 perform a pre-processing process for centering and scaling the variables including the Y space score vector and the X space score vector. By calculating the transfer vector between the X space vector and the Y space vector, the prediction model is defined. The post-processing process (that is, the reverse centering and scaling of the variables) includes running sample data or DoE data through the model and presenting variables and results. The prediction algorithm is then a function of the model, DoE and the tests, Q2, R2, scores, loads, Hotellings and DmodX around it, wherein the tests provide a measure of the data structure and model capabilities.

[0080] A graphical interface is generated and the prediction results are provided graphically in a web interface at 50. In an exemplary embodiment, the web interface display includes a graphical display (Windows 10 TM Several variables / machine parameters and output performance values ​​such as G ratio or surface finish ( Figure 1 ).

[0081] An example output display is shown in Figure 6 , which shows values ​​of predicted parameters 54 for a machine setting of interest and best possible performance values ​​56 for the machine setting in a particular industrial application (e.g., G ratio). The user then uses these values ​​to manually update the machine settings at 52 and provide the machine settings as input variables or process variables to the industrial process to optimize the selected output variables. In the illustrated example of a grinding wheel application, such machine settings may include G ratio, material removal rate of the grinding wheel, chip thickness, number of workpieces per dressing cycle, surface finish, speed of the grinding wheel, roller speed, lateral speed of the grinding wheel relative to the roller, continuous infeed, end infeed, grinding time, feed rate, shift, dressing, dressing infeed, infeed, and overlap ratio, etc. Alternatively, the adjustments may be automatic whereby the variables / machine parameters are provided directly to the processing unit. Figure 1 The auxiliary controller 20 or the main controller 16 in the embodiment of the invention can automatically control the operation of the servo motor 14. Then at 48, the updated machine settings can be fed back to update the prediction model. Alternatively, as will be described below with reference to other embodiments, the prediction can optionally be used to provide product recommendations and / or optimize product combinations by identifying holes and / or overlaps in product lines at 53.

[0082] therefore, Figure 5 The predictive model can be used to process inputs and outputs associated with an industrial process, create a data model of the data, generate predictions from the data model, and provide predictions of the best possible results (e.g., optimal machine settings) based on the test data. The user can change multiple parameters (e.g., wheel specifications, speeds, etc.) at the same time and make adjustments until the desired output is developed (e.g., the best possible G ratio while eliminating redundant iterations that do not result in improved performance). In the above roller mill example, general test data is collected (e.g., collected into a spreadsheet), and data related to rollers (e.g., roller diameter, roller width), wheels (e.g., wheel diameter, wheel width, abrasive hardness, hardness number, structure, bonding, etc.) and grinding machine parameters (e.g., wheel speed, roller speed, table speed, continuous infeed, end infeed, number of passes, chip thickness, special material removal rate, wheel / roller speed ratio, overlap ratio, machine power used, machine pressure, machine current, etc.) and test results (e.g., G ratio, surface roughness, grinding time) are collected. A data model is created from the statistical data using multivariate analysis and processed using a mathematical engine, and the data model is used to generate forecasts and graphs for display in a web-based application. The forecast results are displayed on a monitor ( Figure 6 ). Allows multiple variables and machine parameters to be varied simultaneously to speed up predictions of best possible outcomes and recommended machine settings based on test data. New models can be loaded to address new applications and new industrial processes. Additional examples are provided below.

[0083] Data model creation

[0084] exist Figure 5 46 in the creation of a predictive model can use any of a variety of methods known to those with a background in statistical analysis and / or linear algebra. For example, the model builder first sorts out the test objectives and looks at the data set. In the above roller mill example, 98 rows of grinding test data from 29 different test occasions were used. Then the variables supporting the objectives are grouped and an initial principal component analysis is run on the data. View the resulting scores, loads, DModX, Hotellings, and residuals (described below). Also view the variables to identify information and patterns not described in the data set. Generate new variables as appropriate. Perform principal component analysis on the updated data set. Keep any changes that improve the model fit R2. R2 is the percentage change of the training set (Y with orthogonal projection of potential structure (OPLS)) explained by the Y prediction component. R2 is a measure of fit, that is, the degree of fit of the model to the data. A large R2 (close to 1) indicates a good model, but this is not enough because a poor prediction model may still have a large R2. Poor R2 can also be obtained when there is poor reproducibility in the training set (lots of noise), or when X does not explain Y for other reasons.

[0085] Output or result (Y variable) is introduced as appropriate, and the model is run according to the instruction, and the algorithm of partial least squares (PLS) or OPLS analysis is usually used to generate the model. The contrast prediction graph is checked again and the obtained score, load, DModX, Hotellings and residual are observed. Generate and view the variable importance graph. Add the cross / square variable of extension and the conversion of the recommended variable, and gradually generate a new model and keep the change of the model fitting R2 and / or the prediction ability Q2 of the improved data set. Q2 is the percentage change of the training set (Y with OPLS) predicted by the model according to cross validation. Q2 indicates the degree of model prediction of new data. Large Q2 (Q2>0.5) indicates good predictability. When the data has a larger noise, or when the relationship between X and Y is poor, or when the model is dominated by some scattered outliers, poor Q2 can be obtained. In the process described herein, the outliers (observations) in the data are usually removed, and the variables that are not uniquely helpful to improve the prediction ability are removed to generate the model by an iterative process. Check the variable graph again to confirm the contribution of the variable. When the improvements in R2 and Q2 give sufficiently small changes (eg, one hundredth or one thousandth of a percentage point), the process stops.

[0086] Those skilled in the art will know that PLS regression is a statistical method that has some relationship to principal component regression. However, instead of finding the hyperplane of maximum variance between the response variable and the independent variables, a linear regression model is found by projecting the predictor and observable variables into a new space. Since both the X and Y data are projected into a new space, the PLS family of methods is called a bilinear factor model.

[0087] PLS is used to find the underlying relationship between two matrices (X and Y), that is, a latent variable method for modeling the covariance structure in these two spaces. The PLS model attempts to find the multidimensional direction in the X space that explains the maximum multidimensional variance direction in the Y space. PLS regression is particularly appropriate when the predictor matrix has more variables than observations and when there is multicollinearity between the X values.

[0088] More information on exemplary prediction procedures that may be incorporated herein can be found in the book entitled "Multi and Megavariate Data Analysis" (3rd revised edition, L. Eriksson, T Byrne, E Johansson, J Trygg, C ), so it will not be described further in this article.

[0089] Prediction and Optimization

[0090] Once R2 and Q2 are calculated, the working file is saved. All working models are removed and only the last data model is saved as a model description file. The resulting data model can be used to predict and / or simulate the system from which the data originated (in this case, the grinding model). For example, once the data model is created at 46, the model development software and model runtime engine call the model description file containing the data model for application at 48. In an exemplary embodiment, the code automatically sends the test / machine settings to the model development software and model runtime engine, which performs predictions by running the data model using the prediction algorithm as described above.

[0091] The optimization is then performed by defining the variables to be optimized (usually machine variables). Ranges are defined for each variable. High / low settings are used where it is known that the process is stable and the operators of the machine being optimized are comfortable. In the above example, a DoE matrix is ​​generated to run a 3-level full factorial (3 levels and 5 variables giving 3 5 (243) different combinations) and DoE matrix. Rank the variable combinations based on performance / value / result in the response variable as appropriate. Select the preferred variable settings and run to confirm the results.

[0092] The machine learning tool 24 also implements a "self-learning process" that provides continuous optimization. By saving the actual results after optimizing the initial model / calibration / data file, a new model can be generated that includes learning and information from the experiment just performed. If the new data points improve the model fit and / or predictive ability, these data points can be added and a new model generated. Initially, this step is manual, but since the variable contributions and transformations may not change, this step can be automated and greatly simplified compared to the above process. If the model fit and / or predictive ability decreases, it is necessary to review the variable contributions to identify new variables / interactions / transformations to determine whether the observation is an outlier or an observation that contains new information.

[0093] simulation

[0094] Once the data model is created, simulations of different operating conditions of the industrial process can be generated to show the predicted results of specified input machine parameters without actually running the industrial process. The simulation results so generated can be displayed on the display 26 or sent to the operator's smartphone 28 for display (see Figure 1 ).

[0095] Statistical analysis

[0096] Statistical analysis is used to generate test data used to generate a data model and predict outcomes based on changes in different variables. Brief definitions of terms useful for understanding the statistical analysis applied in the exemplary embodiments are described below.

[0097] Partial Least Squares Projection of Latent Structure

[0098] Partial least squares (PLS) is used to relate the outcome variable to the input and process variables in the exemplary embodiment to solve the problem of identifying those input and process variables (X) that are "responsible" for the variation in the output or outcome variable (Y). To this end, multiple regression can be used. However, multiple regression leads to great difficulties because process data usually do not have the right characteristics for regression modeling because regression treats each outcome variable (y) separately. m ). Thus, a single regression ends up with a set of models, one for each output of interest. This makes interpretation and optimization difficult or impossible. To be able to rigorously interpret cause and effect, use a Software that uses statistical DoE to collect data in a careful experimental process. To search for relationships between inputs and outputs, process logs are risky and generally less successful because the process does not provide good information content for the data when important factors are well controlled within small control intervals.

[0099] Partial Least Squares - Score

[0100] PLS modeling has been developed for situations with many often correlated input and process variables and several to many outcome variables. To use PLS, specify which variables in the database are predictors (X) and which variables are dependent variables (Y). PLS then finds the relationship between variables X and Y. The PLS model is represented as a set of X score vectors, Y score vectors, X weight vectors, and Y weight vectors for a set of PLS ​​model dimensions. Each dimension (index a) represents an x ​​score vector (t a ) and Y score vector (u a ). The weight vector for each model dimension indicates how the X variables are combined to form t a and how the Y variables are combined to form u a In this way, the data is modeled as a set of factors in X and Y and their relationships. Plots of the scores and weights facilitate the interpretation of the model.

[0101] Partial Least Squares - Loadings

[0102] PLS analysis obtains model coefficients of variables, called PLS weights or loadings. The loadings of X variables, represented as w, indicate the importance of these variables, that is, the amount to which they participate in the modeling of Y in a relative sense. The loadings of Y variables, represented as c, indicate which Y variables are modeled in the corresponding PLS model dimensions. When these coefficients are plotted in a w*c diagram, a picture showing the relationship between X and Y, those X variables that are important, and those Y variables that are related to which X variable, etc. is obtained.

[0103] Partial Least Squares - Residuals

[0104] PLS provides residuals on both the result (Y side) and the input (X side). The standard deviation of these residual distances can be plotted as in principal component analysis to provide a third statistical process control chart that shows whether the process is behaving normally in the DModX and DModY charts.

[0105] Principal component analysis

[0106] The principal component analysis (PCA) of the data table gives the values ​​T iaA score vector is calculated that summarizes all the variables that entered the analysis. Two or three score vectors are usually calculated and then plotted against each other in a tt plot. The tt plot gives a picture that is the best summary of the process behavior over time. The tt plot allows trends, unusual behavior, and other things of interest to be seen. The number of components is determined by the component contribution to the predictive power of the model or the individual component contribution to a single outcome variable. When the contribution falls below a threshold, the generation of components stops. As a rule of thumb, areas of the PCA score plot where the process remains under "normal" operation can be identified to provide multivariate control charts. Score Plots and Loadings Figure 1 A value that indicates the cause of a deviation from normal operation.

[0107] For example, Fig. 7A and Figure 7B Loading plots ( ) showing PCA analysis of product properties of adhesives and tapes (e.g., elastic modulus, tensile strength, viscosity, toughening, acrylic, abrasion, elongation at break, FRP, PVC, open time, etc.) 58 Fig. 7A ),in Figure 7B The score graph shows the relationship between different rows or products (e.g., DP100 Clear, DP100FR White, DP190Gray, DP270 Black, DP105 Clear, DP 100Plus Clear, DP110 Gray, DP405 Black, DP100FRWhite, etc.) 60. Each product is positioned according to its unique combination of characteristics. Fig. 7A and Figure 7B The load plot shows how the features / columns are related to each other and provides a basis for a product selector or conversion table. PCA also provides residuals, deviations between the data and the PCA model called DModX. When these residuals are large, this indicates abnormal behavior in the process. For this reason, a graph of the residual standard deviation DModX (residual distance, root mean square) is provided. The DModX graph can be used to identify data outliers. This shows that these observations are different from normal observations about the correlation structure of the variables.

[0108] In the case of principal component analysis, it will be appreciated that variables with the same pattern of correlations are grouped. The variable group with the highest correlation to the outcome parameter (price or performance) is the first principal component, and the second highest correlation is the second principal component. Principal components are added until the contribution criteria to predictive power or model fit are no longer met. On the other hand, in the case of application mapping or product selector applications, there are no outcome parameters. Variables, characteristics, uses, and applications are plotted as principal components, and products and applications are plotted in positions corresponding to the combinations of variables and characteristics that constitute a particular observation. These and other variations will become apparent to those skilled in the art of statistical analysis.

[0109] Orthogonal Projection to Latent Structure (OPLS)

[0110] OPLS is an extension of PLS ​​and solves regression problems. OPLS separates the systematic variation of X into two components, one that is correlated with Y (predicted) and the other that is uncorrelated with Y (orthogonal). This provides improved model interpretability. In the single Y case, there is only one predicted component, and all components except the first reflect orthogonal variation. However, for multiple Y variables, there may be more than one predicted OPLS component.

[0111] Bidirectional Orthogonal PLS (O2PLS) is another extension of PLS ​​that solves data integration problems. In a two-block (X / Y) context, O2PLS examines which information overlaps between the two data tables and which information is unique to a specific data table (X or Y). O2PLS accomplishes this task through a flexible model structure that combines three types of components:

[0112] (i) Component indicating the overlap of joint X / Y information;

[0113] (ii) a component representing a characteristic unique to X; and

[0114] (iii) A component representing characteristics unique to Y.

[0115] For both OPLS and O2PLS, the different components can be interpreted in the usual way, as scores, loadings, and residual-based parameters with familiar meanings are retained.

[0116] Score scatter plot t1 vs. t2

[0117] Scores t1, t2, etc. are new variables summarizing the X variables. The scores are orthogonal, i.e., completely independent of each other. There are as many score vectors as there are components in the model. Score t1 (the first component) accounts for the larger changes in X space, then t2, etc. Therefore, a scatter plot of t1 vs. t2 is a window in X space that shows how the X observations are positioned relative to each other. Figure 8 The score plot in shows the possible presence of outliers, groups, similarities, and other patterns in the data. The score plot 58 is a mapping of the observations. Using the two-dimensional score plot 58, the model development software and model runtime engine draw a tolerance ellipse 60 based on Hotelling's T2. Observations outside the ellipse are outliers, e.g. Figure 8 Observation 39 (rightmost point) in .

[0118] DModX observations that are twice as large as the DModX critical value (Dcrit) are moderate outliers. This indicates that these observations differ from normal observations with respect to the correlation structure of the variables. Moderate outliers can be interpreted by selecting it and opening the corresponding contribution plot showing the residuals of all X variables. Variables with large positive or negative residuals are those that differ with respect to the system structure captured by the model.

[0119] Hotelling's T2 map

[0120] Hotelling's T2 (column or line) graph shows the distance from the origin in the model plane (score space) for each selected observation. The figure shows the T2 calculated for a range of selected components (i.e., 1 to 7 or 3 to 6). Values ​​greater than the 95% confidence limit are suspicious, and values ​​greater than the 99% confidence limit can be considered serious. A large T2 range value for a given observation, i.e., a value far above the critical limit, indicates that the observation is far away from other observations in the selected component range in the score space. Therefore, it may be an abnormal observation, which may pull the model in an unfavorable way if it is in the training set. T2 is essentially calculated as the sum of the scores (in squares) for the selected component range divided by its standard deviation (in squares). Therefore, within a specified component range, T2 is the distance from the origin in the model plane (score space).

[0121] When seeing one or several training set observations with large T2 range values, the operator can also view the score, DModX, and contribution plots for the same model and the same component range to help understand why the observation has a high T2 range value. Note that a group of points marked in the T2 plot will also be marked in the score and DModX plots.

[0122] Load scatter plot pq1 vs. pq2 OPLS / O2PLS

[0123] The OPLS / O2PLS loading plot shows the relationship between the X and Y variables of the predictor components. The loadings p correspond to the X part of the model and the loadings q correspond to the Y part of the model. For ease of interpretation, the plot is color-coded according to the model terms for X and Y. Fig. 9 The pq plot is a superposition of the p and q plots of the OPLS / O2PLS model for the first and second prediction components. The loadings p correspond to the covariance between the X variables and the prediction score vector t, while the loadings q correspond to the covariance between the Y variables and the prediction score vector u. X and Y variables with large p or q contribute a lot to the model. Therefore, you can see how the Y variables change relative to each other, which Y variables provide similar information, and how they relate to the X variables in the model.

[0124] Normal probability plot of the residuals

[0125] like Fig.10 The normal probability plot of the residuals shows the residuals standardized on a log-log scale. The standardized residuals are the raw residuals divided by the residual standard deviation (RSD). Fig.10 The plot enables one to detect outliers and assess the normality of the residuals. If the residuals are random and normally distributed, the normal probability plot of the residuals has all points lying on the straight line 62 between -4 and +4 standardized standard deviations. Experimental runs that lie outside of -4 or +4 standard deviations are outliers.

[0126] As explained further below, the predictive models and associated statistical analysis may be applied to data generated by any of a number of industrial processes to provide product selection tools, product recommendation tools for specific applications, and product portfolio optimizers for product portfolio rationalization.

[0127] Industrial Engineering Application Examples – Adhesive Selection Tool

[0128] Currently, most of the adhesive product selection process for customer applications is performed based on the experience of application engineers. The performance of the bond will depend on the interaction between the adhesive and the substrate. Whether the adhesive is a pressure sensitive adhesive with or without an additional adhesive layer or non-adhesive layer, the adhesive behavior is generally well characterized; however, the substrate properties can vary greatly depending on a large number of parameters such as surface finish, bulk polymer additives, surface treatment (corona, primer, etc.), aging, storage conditions, etc. Therefore, adhesive selection provides another industrial environment where the predictive modeling tools described herein can process the collected data to predict adhesive performance, thereby helping to speed up and improve the adhesive selection process.

[0129] In an exemplary embodiment, a handheld instrument can be used to collect the desired data from the substrate surface. This data collection can be performed at the customer site or any other convenient location. For example, the surface free energy of a surface is a parameter that determines the performance of an adhesive / substrate. The surface free energy is calculated using the contact angles of two known liquids and by measuring the contact angle of a droplet on the substrate surface. For example, Fig.11 A handheld device 1100 is shown that measures the surface free energy of a surface of a substrate 1110 . Fig.11 The device uses a liquid drop on a solid surface to measure the contact angle.

[0130] Fourier transform infrared (FTIR) spectroscopy can also be used to measure the parameters of the adhesive / substrate in the exemplary embodiment, because FTIR spectroscopy is ideally suited to determine the type of polymeric material. FTIR spectroscopy is a technique for obtaining the far infrared spectrum of absorption or emission of a solid, liquid, or gas by measuring how much light the sample absorbs at each wavelength. FTIR spectrometers collect high spectral resolution data over a wide spectral range simultaneously by pressing the sample against a diamond crystal to collect absorption data. For example, Fig.12 FTIR spectra of two polypropylene samples are shown 1200 and 1210. In FTIR spectroscopy, Fourier transform is used to convert the raw data into Fig.12 Actual spectra of the types shown.

[0131] Other measurable input variables for the adhesive / substrate may include one or more of the following: adhesive physical properties, adhesive thermal properties, adhesive electrical properties, adhesive cure properties, adhesive performance properties, adhesive durability properties, adhesive chemical resistance properties, adhesive rheological properties, adhesive composition, adhesive dispensing properties, adhesive usage requirements, standardized testing or certification, environmental parameters, health parameters, safety parameters, carrier properties, backing properties, liner properties, and materials to be bonded by the adhesive.

[0132] As in the grinding example above, multivariate analysis can also be used to understand how different factors affect each other and to determine the main causes of variation in the data set. In addition, as in the grinding example above, partial least squares (PLS) can be used to process data from analytical instruments (e.g., FTIR, nuclear magnetic resonance (NMR), etc.) to build predictive models.

[0133] The workflow for applying the predictive modeling techniques described in this article to the adhesive product selection process is summarized in Fig.13In. As shown, a variety of measurement techniques such as FTIR spectroscopy are used to collect X variables representing substrate parameters, thereby generating a database 1300, which includes parameters of corresponding substrates such as polypropylene, polyethylene, polycarbonate, stainless steel, aluminum, red paint, nylon, glass, polymethyl methacrylate (PMMA), black acrylonitrile butadiene styrene (ABS), polyvinyl chloride (PVC), etc. and corresponding adhesive products such as adhesive tapes (e.g., according to part numbers). Y variables can be collected, for example, by performing a peel test (e.g., 3M Industrial Adhesives and Tapes Division (3M Industrial Adhesives and Tapes Division, IATD) 90° peel test) at 1310 to establish the main response of the adhesive tape on the corresponding substrate. This type of analysis can be used to create a prediction model using statistical packages such as Unscrambler, Umetrics and R based on a set of variables that affect measurable output parameters at 1320. For example, the performance of the tape / substrate combination evaluated using a standard peel strength test can be used as the main response. The resulting output variables may include the name of the recommended adhesive, adhesive structural properties, adhesive performance properties, quantification of the quality of the fit, and / or purchasing information. The two surfaces of interest are analyzed using the selected techniques (surface energy and FTIR spectroscopy) and used as variables in a statistical package to create a predictive model 1320, as described in reference Figure 5 The prediction model 1320 is then tested at 1330 by testing the substrate, and the process is iterated using additional samples to improve the prediction model 1320 as an iterative process over time.

[0134] Fig.14 1 is a graph showing the average measured peel force versus the average predicted peel force for a variety of substrate / adhesive combinations. As shown, the prediction model is quite accurate using the developed data set. However, Lexan sample 1400 shows that the prediction model 1320 is confused by similar types of chemistry when compared to the reference. This inaccuracy in the prediction model 1320 can be improved over time by using more samples to build the prediction model 1320. In this example, an adhesive product with a desired peel strength value for the substrate will be selected.

[0135] Industrial Engineering Application Examples – Product Recommender for Abrasive Belts

[0136] Abrasive belts are offered in a variety of configurations (e.g., backing, mineral type, size, etc.) for different applications and pressure scenarios. To provide tools to global application engineers, sales, and customers, the predictive modeling techniques described herein can be used to develop tables to group "process-substrate-machine" and match them to agreed product recommendations. As with the other examples, the predictive modeling process begins with building a predictive model to predict the best product recommendation for a specific customer situation (process-substrate-machine).

[0137] The product finder / recommender logic is based on multivariate analysis to provide continuous refinement of predictions as new user data is added to the base data set. As in the above example, a predictive model is built based on existing user data from product developers and application engineers (as well as customers). The predictive model determines and recommends the next best (best match) product based on the "distance" between customer demand and the existing product portfolio offering. In an exemplary embodiment, the product finder is web-based and can be accessed by computers and other handheld processing devices such as tablets, laptops, smart phones, etc. In an exemplary embodiment, the product finder is implemented on a common architecture for recommender systems to provide standardized functionality.

[0138] Fig.15 A graph 1500 of sample abrasive belts by application—substrate—machine is shown. Fig.15 Applications 1510 and typical levels 1520 are shown along the left side of the chart 1500, and a base 1530 and machines 1540 determined by the predictive model are shown along the top of the chart 1500. Fig.15 , cells 1540 in chart 1500 are displayed in different colors to identify the best value options based on experimentation, the best options for applications where it may be difficult to demonstrate value, applications that are considered unlikely applications for the indicated machine / substrate, and applications that are unlikely and therefore not recommended. Thus, the ranking shows the tool with the highest performance for each combination of application and substrate. FIG. 16A to FIG. 16B To show Fig.15 The information collected in the load diagram is converted into a load diagram of recommended products 1600 for different machines / substrates, where Fig. 16B for Fig.16A The enlarged format.

[0139] FIG. 17A to FIG. 17B To show the Fig.15 A graph of qualitative information that is converted into a numerical "grade" indicating, on a scale of 1-12, the suitability of a given machine / substrate for a particular application.

[0140] Fig.18A graph showing observed versus predicted belt application for a particular application (in this case a robotics application) is shown. Fig.18 As shown, for robotic applications, optimal observations and predictions are provided for the machine / substrate / application that falls on line 1800.

[0141] As a next step, consider the pressure (high-medium-low) and the contact wheel, where the pressure is related to the machine type. Fig.19 A graph of observed versus predicted tape application by machine (and pressure) is shown, with the best observed versus predicted results again provided for the machine / substrate / application that falls on line 1900. FIG. 20A to FIG. 20B is a graph showing the refined numerical ranking of the refined data (taking pressure into account).

[0142] Fig.21 21 is a graph showing performance ratings 2100 of robotic machine applications 2110 and products 2120. Fig.21 In the case of pressure, it is also related to the machine type.

[0143] Fig. 22 A graph showing observed results versus predicted results for a specific product-application-machine / pressure is shown to provide a corresponding performance rating. For example, the highest rating for weld cleanliness and feed-through is predicted for the circled example 2200 provided on the right side of the graph. To confirm that the predictions are consistent with actual experience, the Fig. 22 The information is converted into a table for application engineers to view. For example, Fig.23 A chart showing modeled ratings based on numerical ratings is shown, and it is shown that information from each predicted data point can be viewed. For example, arrow 2300 shows the predicted grade. This use of data from consumers will increase productivity by helping consumers analyze available data.

[0144] It should be understood that the product recommender can be divided into three stages: product finder / recommender, R&D design tool, and product selector and sales guidance. In an exemplary embodiment, an application engineer uses a product recommender to identify the appropriate abrasive belt for a specific application. For different markets, changes in specific terms that are not used in the market can be identified to appropriately adjust the product recommender. On the other hand, customers can use the product recommender to identify which product to choose based on the machine, substrate, and process that the customer is operating. In the R&D design tool stage, test data from the abrasive belt manufacturing process can be collected and used to predict the product that will provide the best results. The product selector can also be applied to sales data collected and / or reported during the sales process to recommend the best product based on, for example, Finish-Application-Substrate-Tool (FAST).

[0145] Product portfolio rationalization example - portfolio optimizer and product recommender for safety harness products

[0146] The following embodiments provide examples of how the predictive modeling tools described herein can be used as a tool for product portfolio rationalization, white space identification, and product recommendations. As in the above embodiments, multivariate analysis tools are used to analyze data generated by the safety harness product offering and are used to help business organizations rationalize their product portfolios by visualizing groups of products with similar performance or characteristics. Although described with respect to safety harness products, it should be understood that the techniques described herein can be used for any of a variety of products, where there are multiple products with similar and potentially overlapping characteristics. The same techniques can be used to simplify product portfolios and can also be used to develop the same product selector / recommender as in the above abrasive belt embodiments.

[0147] This embodiment addresses situations where, for a variety of reasons, a company may offer many products (e.g., fall protection products) available under multiple brands and legacy businesses that offer products with overlapping features. In such situations, it is often difficult for application engineers and customers to know which product is best suited for a specific application requirement. On the other hand, it would be very beneficial for a business organization to understand which products can be grouped together in an effort to address global product portfolio rationalization.

[0148] As in other embodiments, the process begins by collecting product data for use in the predictive modeling process. Fig.24 As shown, the product list may contain potentially hundreds of products with potentially thousands of stock keeping units (SKUs) throughout the world. It should be understood that the more SKUs there are, the more helpful the predictive modeling tool will be in providing product portfolio rationalization. Fig.24 In the example of , qualitative harness data 2400 is provided. In this example, there is no quantitative data to allow products to be grouped based on criteria such as application, harness type, D-ring location and buckle type or other criteria that are qualitatively significant to distinguishing the products.

[0149] Fig.25 The following is a graph generated using Simca, a multivariate statistical analysis tool from Umetrics. Fig.24 Figure 2. Graph of preliminary prediction model for harness data. This multivariate statistical analysis tool plots the correlations between the collected parameters 2500. To provide visualization of product groupings, 14 settings were defined using a random generator. Fig.25The results are shown as visualizations of groupings of products with similar features and characteristics. Using a "bundle score", the model scores the closest product model that fits the predictive model for the identified settings. If a product does not have a 100% fit, the predictive model will display the closest fit. Then, once several product groups are clustered together, the cluster data can be evaluated to identify the top N closest bundles. These product groupings can be used to identify clusters of products with similar characteristics, thereby achieving meaningful product portfolio rationalization. On the other hand, for a product recommender, it may be more important to limit the selection criteria to a manageable number, because if customers have too many options, they may stop looking for products.

[0150] Thus, the embodiment uses predictive modeling tools to identify gaps and overlaps in product lines by monitoring the fit of the output of the predictive model with the input feature parameters.Analyzing the order of variables and customer choice variables can be used to understand customer preferences / choices.

[0151] Enabling customers to select safety products also brings up other considerations. For example, regulations and the like may require warning the user that the selected safety product may not be a perfect match for a particular application. In this case, key performance criteria may need to be the key consideration for the safety product before other features are considered. It may be necessary to add a mandatory field to the product data to indicate which standard the product is approved for, and limit the selection of that product to applications based on the identified standard. Additionally, the product recommender may also highlight the larger area of ​​concern for a product's specific application. For example, product color may not be as important as arc flash rating / high energy discharge and other features that are critical to safety that trigger a warning. For example, selecting a specific product for a specific application may pop up a notification: "This selection is only a 90% match to the selected safety critical variables, not a 100% match, less than a perfect match." Thus, the product recommender enables suppliers to identify and fill gaps in their product portfolio and provide customized products while reducing the burden on their customers or suppliers.

[0152] Thus, the systems and methods described herein implement multivariate statistical analysis and machine learning techniques to predict the performance of industrial products from simple to increasingly complex data files, thereby identifying product characteristics and desired results or features for specific applications of industrial products. Predictive modeling tools can be used by customers to find appropriate products for specific applications, and by companies to recommend products for specific applications, as well as to rationalize product portfolios to find holes / overlaps in product offerings. These and other applications will be apparent to those skilled in the art from the above description of exemplary embodiments.

[0153] Computer Implementation

[0154] Fig.26 A typical general purpose computer that can be programmed as a special purpose computer suitable for implementing one or more embodiments of the systems disclosed herein is shown. The controllers 16 and 20 described above and the machine learning tool 24 can be implemented on any general purpose processing component, such as a computer with sufficient processing power, memory resources, and communication throughput to handle the necessary workload placed on it. The processing component 70 includes a processor 72 (which may be referred to as a central processor unit or CPU) in communication with a memory device, which includes a secondary storage device 74, a read-only memory (ROM) 76, a random access memory (RAM) 78, an input / output (I / O) device 80, and a network connectivity device 82. The processor 72 may be implemented as one or more CPU chips, or may be part of one or more application specific integrated circuits (ASICs).

[0155] The secondary storage 74 is typically made up of one or more disk drives or tape drives and is used for non-volatile storage of data and as an overflow data storage device if the RAM 78 is not large enough to hold all the working data. The secondary storage 74 can be used to store programs that are loaded into the RAM 78 when a program is selected for execution. The ROM 76 is used to store instructions and possibly data that are read during program execution. The ROM 76 is a non-volatile memory device that typically has a small storage capacity relative to the larger storage capacity of the secondary storage 74. The RAM 78 is used to store volatile data and possibly to store instructions. Access to both the ROM 76 and the RAM 148 is typically faster than access to the secondary storage 74.

[0156] The apparatus described herein may be configured to include a computer-readable non-transitory medium storing computer-readable instructions and one or more processors coupled to the memory, and when the computer-readable instructions are executed, the processing component 70 is configured to perform the above-referenced Figures 1 to 25 The method steps and operations described herein. Computer-readable non-transitory media include all types of computer-readable media, including magnetic storage media, optical storage media, flash memory media, and solid-state storage media.

[0157] It should also be understood that software including one or more computer executable instructions that facilitate processing and operation as described above with reference to any or all steps of the present disclosure can be installed in one or more servers and / or one or more routers and / or one or more devices in the field of consumers and / or producers in accordance with the present disclosure and sold together with the one or more servers and / or one or more routers and / or one or more devices. Alternatively, the software can be acquired and loaded into one or more servers and / or one or more routers and / or one or more devices in the field of consumers and / or producers in accordance with the present disclosure, including acquiring the software through a physical medium or distribution system, including, for example, acquiring from a server owned by the creator of the software or from a server that the creator of the software does not own but uses. For example, the software can be stored on a server to be distributed through, for example, the Internet.

[0158] In addition, it should be understood by those skilled in the art that the present disclosure is not limited to its application to the specific constructions and component arrangements mentioned in the following description or shown in the accompanying drawings. The embodiments herein can have other embodiments and can be practiced or implemented in various ways. Moreover, it should be understood that the words and terms used herein are for illustrative purposes and should not be considered restrictive. The use of "including", "comprising" or "having" and its variations herein is intended to cover the items listed thereafter and their equivalents and additional items. Unless otherwise specified, the terms "connect", "couple" and "install" and their variations herein are used broadly to cover direct and indirect connections, couplings and installations. In addition, the terms "connect" and "couple" and their variations are not limited to physical or mechanical connections or couplings. In addition, terms such as upward, downward, bottom and top are relative and are used to assist in explanation, but are not restrictive.

[0159] The components of the exemplary apparatus, systems and methods employed in accordance with the illustrated embodiments of the invention may be implemented at least in part in digital electronic circuitry, analog electronic circuitry, or in computer hardware, firmware, software, or in a combination thereof. These components may be implemented, for example, as a computer program product, such as a computer program, program code, or computer instructions tangibly embodied in an information carrier or in a machine-readable storage device, for execution by a data processing device, such as a programmable processor, one or more computers, or to control the operation of the data processing device.

[0160] The computer program can be written in any form of programming language (including compiled or interpreted languages), and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for a computing environment. The computer program can be deployed to be executed on a computer or multiple computers located at a site or distributed at multiple sites and interconnected by a communication network. In addition, the functional program, code and code segment used to implement the present invention can be easily understood by programmers in the field of the present invention as being within the scope of the present invention. The method steps associated with the exemplary embodiments of the present invention can be performed by one or more programmable processors that execute computer programs, codes or instructions to perform functions (e.g., by operating on input data and / or generating output). The method steps can also be performed by a dedicated logic circuit (e.g., FPGA (field programmable gate array) or ASIC (application-specific integrated circuit)), and the device of the present invention can also be implemented as a dedicated logic circuit.

[0161] The various illustrative logic blocks, modules and circuits described in conjunction with the embodiments disclosed herein may be implemented or executed with a general purpose processor, a digital signal processor (DSP), an ASIC, an FPGA or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in an alternative form, the processor may be any conventional processor, controller, microcontroller or state machine. The processor may also be implemented as a combination of computing devices, for example, a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.

[0162] Processors suitable for executing computer programs include, by way of example, general and special purpose microprocessors, and any one or more processors in any type of digital computer. Generally speaking, the processor will receive instructions and data from a read-only memory or a random access memory or both. The basic elements of a computer are a processor for executing instructions and one or more storage devices for storing instructions and data. Generally speaking, a computer will also include one or more mass storage devices (e.g., magnetic disks, magneto-optical disks, or optical disks) for storing data, or will be operatively connected to receive data from a mass storage device, or transfer data to a mass storage device, or both. Information carriers suitable for embodying computer program instructions and data include all forms of non-volatile memory, including, for example, semiconductor memory devices, such as electrically programmable read-only memory or ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory devices, and data storage disks (e.g., magnetic disks, internal hard disks or removable disks, magneto-optical disks, and CD-ROM and DVD-ROM disks). The processor and memory may be supplemented by or incorporated into a dedicated logic circuit.

[0163] It will be appreciated by those skilled in the art that any of a variety of different techniques may be used to represent information and signals. For example, data, instructions, commands, information, signals, bits, symbols, and chips that may be referenced in the above description may be represented by voltage, current, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.

[0164] It will also be appreciated by those skilled in the art that the various exemplary logic blocks, modules, circuits and algorithm steps described in conjunction with the embodiments disclosed herein may be implemented as electronic hardware, computer software or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, various exemplary components, blocks, modules, circuits and steps have been described above generally in terms of their functions. Whether such functions are specifically implemented as hardware or software depends on the specific application and the design limitations imposed on the overall system. The technician may implement the functions in a varying manner for each specific application, but the decision of such implementation should not be understood as causing a departure from the scope of the present invention. The software module may reside in a random access memory (RAM), a flash memory, a ROM, an EPROM, an EEPROM, a register, a hard disk, a removable disk, a CD-ROM or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. In an alternative form, the storage medium may be integral with the processor. In other words, the processor and the storage medium may reside in an integrated circuit or be implemented as discrete components.

[0165] As used herein, "machine-readable medium" means a device capable of temporarily or permanently storing instructions and data, and may include, but is not limited to, random access memory (RAM), read-only memory (ROM), buffer memory, flash memory, optical media, magnetic media, cache memory, other types of storage devices (e.g., erasable programmable read-only memory (EEPROM)), and / or any suitable combination thereof. The term "machine-readable medium" should be deemed to include a single medium or multiple media (e.g., a centralized or distributed database, or associated caches and servers) that can store processor instructions. The term "machine-readable medium" should also be deemed to include any medium or combination of multiple media that can store instructions for execution by one or more processors, such that when the instructions are executed by the one or more processors, the one or more processors perform any one or more of the methods described herein. Therefore, "machine-readable medium" refers to a single storage device or device, as well as a "cloud-based" storage system or storage network that includes multiple storage devices or devices. As used herein, the term "machine-readable medium" does not include the signal itself.

[0166] The above description and accompanying drawings are given by way of example only and are not intended to limit the exemplary embodiments in any way, except as shown in the appended claims. It should be noted that the various technical aspects of the various elements of the various exemplary embodiments described above can be combined in many other ways, all of which are considered to be within the scope of the present disclosure.

[0167] Therefore, although exemplary embodiments have been disclosed for illustrative purposes, those skilled in the art will appreciate that various modifications, additions and substitutions are possible. Therefore, the present disclosure is not limited to the above embodiments, but may be modified within the scope of the attached claims together with the full scope of their equivalents.

Claims

1. A computer-implemented method for predictive modeling, comprising: a computer performing a multivariate analysis on input variables and output variables generated during operation of an industrial process to generate a data model of the operation of the industrial process, the input variables comprising process variables and the output variables comprising results from the operation of the industrial process, the data model representing contributions of the respective input variables or process variables to changes in the output variables; providing the data model to a predictive algorithm to identify parameter values ​​of input variables or process variables that are expected to have the most significant impact on a selected output variable during performance of the industrial process, the predictive algorithm outputting the parameter values; as well as providing the parameter value as the input variable or the process variable to the industrial process to optimize the selected output variable, wherein generating the data model comprises the computer automatically generating an experimental design DoE for the input variable space of the industrial process, processing the experimental design using the data model and storing the results based on the preferred results, and further comprises the prediction algorithm predicting a predetermined number of best next trials of the DoE and outputting parameter values ​​of the predetermined number of best next trials and the probability of an improved result using the parameter values, and Wherein, the industrial process includes at least one of a grinder operation, a binder selection process, and a grinding operation.

2. The method of claim 1 further comprising feeding the parameter values ​​and output variables generated by the industrial process in response to the parameter values ​​as feedback data to the computer, and the computer adjusting the data model using the feedback data.

3. The method of claim 1, further comprising outputting the parameter value to a display as a simulation of an output of the industrial process in response to a specified input variable.

4. The method of claim 1, further comprising recommending a product having parameter values ​​that optimize the selected output variable in the industrial process. 5 . The method of claim 4 , wherein recommending the product comprises recommending the product that optimizes the selected output variable in the industrial process for at least one of a specified machine, a specified substrate, and a specified application.

6. The method of claim 1 further comprising identifying products in the product portfolio having overlapping parameter values ​​that optimize the same selected output variables.

7. The method of claim 1 further comprising identifying gaps in the product portfolio by identifying parameter values ​​for optimizing the selected output variables that are not represented in the product portfolio.

8. The method of claim 1 , wherein the industrial process is a grinding machine operation, the input variables include machine settings of the grinding machine, and the output variables depend on the type of grinding process and include at least one of G-ratio, material removal rate of the grinding wheel, chip thickness, number of workpieces per dressing cycle, and surface finish.

9. The method of claim 8, wherein the machine settings include at least one of grinding wheel speed, roller speed, infeed speed, continuous infeed, end infeed, grinding time, feed rate, shift, dressing, dressing infeed, infeed, and overlap ratio.

10. The method of claim 1, wherein the industrial process is an adhesive selection process, the input variables include adhesive or tape selection variables, and the output variables include at least one adhesive or tape selection or recommendation.

11. The method of claim 10, wherein the input variables include at least one of adhesive physical properties, adhesive thermal properties, adhesive electrical properties, adhesive curing properties, adhesive performance properties, adhesive durability properties, adhesive chemical resistance properties, adhesive rheological properties, adhesive viscosity, adhesive setting time, adhesive elastic modulus, adhesive solvent resistance, adhesive composition, adhesive dispensing properties, adhesive usage requirements, standardized testing or certification, environmental parameters, health parameters, safety parameters, carrier properties, backing properties, liner properties, and materials to be bonded by the adhesive or tape.

12. The method of claim 10, wherein the output variable comprises at least one of tensile strength, peel strength value, adhesive name, adhesive structural properties, adhesive performance properties, quantification of fit quality, and purchasing information.

13. The method of claim 10, wherein the adhesive is a pressure sensitive adhesive with or without an additional adhesive or non-adhesive layer.

14. The method of claim 1, wherein the industrial process is a grinding operation, the method further comprising recommending a product that optimizes the selected output variable in the industrial process for at least one of a specified machine, a specified substrate, and a specified application for an abrasive belt.

15. A computer-implemented method for performing predictive modeling, comprising: The computer performs a multivariate analysis on input variables related to characteristics of a product and output variables related to performance of the product to generate a data model of the product, the data model representing contributions of the respective input variables to variations of the output variables; providing the data model to a predictive algorithm to identify parameter values ​​of input variables that are expected to have the most significant impact on a selected output variable, the predictive algorithm outputting the parameter values; providing the parameter value as the input variable to optimize the selected output variable; and Recommending a product that optimizes the selected output variables for specified input variables, wherein generating the data model includes the computer automatically generating an experimental design DoE for the input variable space, using the data model to process the experimental design and storing the results based on the preferred results, and also includes the prediction algorithm predicting a predetermined number of best next trials for the DoE and outputting parameter values ​​of the predetermined number of best next trials and the probability of improved results using the parameter values.

16. The method of claim 15, further comprising identifying products in the product portfolio having overlapping parameter values ​​that optimize the same selected output variable.

17. The method of claim 15, further comprising identifying gaps in the product portfolio by identifying parameter values ​​for optimizing the selected output variables that are not represented in the product portfolio.

18. The method of claim 15, wherein the input variables include characteristics of at least one of a grinder, a bonded abrasive grinding wheel, an abrasive belt, a coated abrasive belt or disc, a nonwoven abrasive, a bristle brush, an abrasive article mounted on a robot, an adhesive, and a safety harness.

19. The method of claim 15, wherein the input variables comprise process variables of an industrial process and the output variables comprise results from the operation of the industrial process, and the data model represents the contribution of the corresponding input variables or process variables to the change of the output variables, wherein the parameter values ​​are provided to the industrial process as the input variables or the process variables to optimize the selected output variables.

20. A system for implementing predictive modeling to optimize an industrial process, the system comprising: at least one processor; as well as a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform operations including: performing a multivariate analysis on input variables and output variables generated during operation of the industrial process to generate a data model of the operation of the industrial process, the input variables comprising process variables and the output variables comprising results from the operation of the industrial process, and the data model representing contributions of respective input variables to changes in the output variables; Automatically generate a design of experiments (DoE) for the input variable space of the industrial process; Processing experimental design using the data model; providing the results of processing the experimental design to a prediction algorithm based on the preferred outcome; The prediction algorithm identifies parameter values ​​of input variables that are expected to have the most significant impact on a selected output variable during performance of the industrial process, makes predictions for a predetermined number of best next trials for the DoE, and outputs parameter values ​​for the predetermined number of best next trials and probabilities of improved outcomes using the parameter values; as well as providing the parameter value as the input variable to the industrial process to optimize the selected output variable, Wherein, the industrial process includes at least one of a grinder operation, a binder selection process, and a grinding operation.

21. The system of claim 20, wherein the at least one processor further performs operations including feeding the parameter value and the output variable generated by the industrial process in response to the parameter value as feedback data to the at least one processor and adjusting the data model using the feedback data.

22. The system of claim 20, wherein the at least one processor further performs operations including outputting the parameter value to a display as a simulation of an output of the industrial process in response to a specified input variable.

23. The system of claim 20, wherein the at least one processor further performs operations including recommending a product having parameter values ​​that optimize the selected output variable in the industrial process.

24. The system of claim 23, wherein the at least one processor recommending the product further comprises the at least one processor performing operations including recommending the product that optimizes the selected output variable in the industrial process for at least one of a specified machine, a specified substrate, and a specified application.

25. The system of claim 20, wherein the at least one processor further performs operations including identifying products in the product portfolio having overlapping parameter values ​​that optimize the same selected output variables.

26. The system of claim 20, wherein the at least one processor further performs operations including identifying gaps in the product portfolio by identifying parameter values ​​for optimizing the selected output variables that are not represented in the product portfolio.

Images