WO2023121581A1 - A machine learning based method to estimate costs of a third party - Google Patents

Abstract

The invention disclosed is a machine learning method to estimate costs of a production, particularly in manufacturing. As costs and cost structures vary from company to company, estimating the cost of a procurement in between inconsistent bids remains as an unsolved problem. The invention disclosed proposes a machine learning model based solution to this problem.

Description

A Machine Learning Based Method to Estimate Costs of a Third Party

The invention disclosed is a machine learning method to estimate costs of a production, particularly in manufacturing. As costs and cost structures vary from company to company, estimating the cost of a procurement in between inconsistent bids remains as an unsolved problem. The invention disclosed proposes a machine learning model based solution to this problem.

In the state of the art, there are multiple solutions proposed for the cost estimation problem. These vary from statistical approaches to financial approaches but there is no direct and unambiguous solution in complex environments. Some prior solutions are as follows;

CA3133800 generates a third party cost estimation for a specific design; where the metrics of a specific design is given. But not every offer is requested for the single type of product; products and services procured are practically infinite. A mere generalization of the disclosed method can not overcome difficulties in this variance of goods and services procured. Besides, the method disclosed does not enjoy capabilities conferred by artificial intelligence; thus renders a mechanical method for estimating a single product type.

CN113129059 discloses an artificial supply chain for a given time constraint and predicts the cost via sole abstraction. On contrary, supply chains are tough to estimate or re-generate; they are multi dimensional spaces where a single way to connect in between the two points. Besides, in the real world, organizations shall not have to make a digital twin of the whole supply chain; they usually lack of information of the supply chain nor they have an interest on. Organizations procure some goods and services and they have information on the supplier solely. So, the problem should be solved with this limited data where the method in the document is not capable of.

CN111415220 is an example of deterministic methods for cost estimation. The method proposed includes calculation of costs for the production and machining to be made and processes them by simple math; a direct calculation of the data by pen and paper will give the user a result. Such a method would fail at varying cost structures and varying capacity utilizations besides other factors.

US2021357965 discloses a method to gather user information along websites and make a cost estimation by comparison. On the real world application, not every data can be found on websites, so chances of such a system for accurately estimating costs onto a web-based comparison. But in industry conditions, not all variables are present in websites; some may be varying by time (i.e. labor costs)

JP2001265825 discloses a method to obtain geometrical parameters and attribute data of the production or machining to estimate costs based on the geometric and attribute data. To do so, the document discloses use of a computer to calculate a three dimensional simulation of the production and base costs on this 3D data. But on industry conditions, a cost has much more than that; only geometrical analysis of the part can not be adequate for cost estimation. Real life costs happen to be much detailed than that; number of variables may increase rapidly by time. So this document is not considered to be similar in this regard.

CN113065778 discloses a method for optimizing a production, thus reducing costs. The method disclosed is a result of direct hereustic object oriented programming and has not forecast on costs, but only an abstract will to reduce the cost of production. It does not employ any means for a specific production, but general means to organize the workplace. The invention disclosed, on contrary, predicts costs for each and every production in the workplace and estimates the true cost of production even if the production made is statistically imperfect.

CN113127498 discloses a method to generate a database of previously made productions and generates trees of production starting from raw material to finished product. In the case of an update of a raw material, it updates the overall cost estimation, as to the weight in the production tree suggests. This method is proposed for shipbuilding industry and it may have some impact on raw-material based industries. On the other hand, a general solution fo cost estimation can not be solely based on raw material per se. Instead, many variances at cost structure has to be taken into account; from machine prices to inventory costs; from cutting inserts to labor costs. We hereby propose a method to cover all in a single context.

CN113240475 discloses a cost estimation method based on likelihood of the component composition. The said document aims raw material fluctuations and deviations within a model. This model is a direct arithmetic function of the costs associated with previous productions as far as understood. Such a model neglects plenty of data in the production for accurately estimate the cost of a production. To do so, one should take into account many variations, taking a single variation is far from resulting with a realistic estimation. When the number of variations increase, it becomes nearly impossible to construct a polynomial model as described in the said document. So, an accurate estimation of this complex problem is simply impossible by such a method.

US2021350037 discloses a method for cost estimation, based on a CAD file; generating a non-CAD unfolded model to estimate costs. This may have other applications, but a totally incompetent method for accurate cost estimation. In the state of the art, there are accurate models to predict a timing for a production based on a CAD model. One can make a time estimation and machine allocation time easily, without even employing the method at the said document. Even in that case, an estimation of total cost can not be achieved, as cost structure is a complex multi-body problem rather than a solely machine related variable. Each supplier of production usually has its own set of variables for accurate determination of cost structure where in this document these variations are neglected.

US2001023418 defines a cost estimation method where data is obtained from the CAD machine directly to calculate a production variable set and estimate a cost based thereon. But in real life conditions, not all costs are machine based, nor machine working is steady as it is modeled in the document. To have an accurate estimation of the total production cost, one should take much more into consideration.

WO2021015295 proposes a cost estimation model based on resource consumption. Although this is a more robust way of estimating costs, generating a single model from production parameters and concluding a cost estimation is by no means effective, as a single model can not be inclusive for plurality of suppliers, cost structures, methods of suppliers and the like. At least financial costs are excluded from the cost estimation in this figure, which are usually a determining part in severe competitive markets. There are many other cost parameters excluded in the model of this document.

CN111860680 defines a clustering method for each and every piece by means of K-means algorithm, suitable for many body production lines. Via such manner, the document proposes a cost estimation method for complex productions, such as automobile industry. But the cost estimation of each part is still a sustaining problem where the document proposes a solution on non-existing data on parts. The method defined in the document estimates costs of an assembly line, where part costs can not be estimated accurately.

CN110766477 discloses a process library and production metrics for a defined production of a part where the library is based on work hours. 
This document proposes a method on K-means algorithm and gives an “average” cost for production. In the real world conditions, an average of a mess-up of previous data is never sufficient to make an accurate estimation. Any indelicate estimation of costs result either in loss of contract due to high offering or loss remnant at the supplier due to low bidding. This level of accuracy is a direct result of the single algorithm use and is not able to solve problem of manufacturing in real world conditions.

The invention disclosed is a method to estimate costs for a defined supplier, based on the cost occurrence of operations of the same class that have been made with the said supplier up to that date. The method of invention uses the data of past operations to estimate costs of future operations. To do so, we propose a method based on artificial intelligence to avoid both making low offers and making losses and high offers to lose the tender.

The main problem of the invention is forecasting the total cost of a unit of operation (i.e. machining) to be done by a specific supplier in a precise way. The main question to define this problem is the cost structure of an supplier in its all complexity. As cost structure of each supplier is unique on its own, determining actual cost before operation is a suspending problem. Without a solution to this problem, suppliers either give high offers to lose some job or low offers to result some losses.

Another problem addressed by this invention is offer/bid estimation problem. For suppliers; preparing an offer takes a considerable amount of time, resulting in extra labor and loss of time. Especially in an era of automatic production and interacting machines, offering becomes the manual thus slowest part of the process.

The invention disclosed is a method of cost estimation, working on the data of previous transactions by a supplier. The supplier mentioned herein may be of any kind; a manufacturing company, a logistics company or any service company, etc. For ease of explanation, a manufacturing company will be given as the example for clearer understanding.

The method described herein is a method to run on a computer, the said computer being a desktop, server, laptop, embedded, mini, cloud, distributed, virtual or any other type of computer capable of handling logical and mathematical tasks and run algorithms as instructed. The said computer (100) has at least a processor, memory element and input and output gateways to communicate. The characteristics of the said computer (100) changes the speed and capacity of operations made, but functional algorithm of method steps do not vary upon specifications of the computer (100) thereon.

The method consists of three connected steps of operation. In the first step, a web interface (101) takes effect and helps user to classify and upload data to a computer (100). In this step, only user data is classified and taken into memory element of the computer (100) and transcribed into a statistically significant database, where each row and column represent previously decided characteristics of the operation. For instance, for a machining company, one column represents processing time, other column represents total volume milled, other column represents equipment used and the like.

The basics of this database (102) needs to be defined before the method isi initiated; key parameters of the actions of supplier or the data at hand is required to be set in a clear and concise manner within a database. A person skilled in the art can easily vary these where a totally new set of variables can be introduced with ease and speed. The only important aspect of this database is that it shall cover costs of previous actions. At least, it should cover prices of previous actions and total loss/profit of the supplier.

In the second step, a cost estimation model is built. To build a cost estimation model, a model library (103) is used within the method. Previously having loaded, a plurality of artificial intelligence modules are present in the memory of the computer. At this point, the data is split into two by previously defined proportions; for example 85%-15% could be an advisable split. The first portion of the data is called training data (201) and second portion of the data is called verification data (202).

The artificial intelligence models previously loaded into the computer are consecutively trained by the training data to obtain a model for each method. Here we take the first method and train it with the training data to obtain a first surrogate model (203) for the supplier and save this surrogate model as the first surrogate model. Then comes the second model; we train the second model with the same training data and obtain a second model for the supplier. This repeats until all models on the model library to obtain as many surrogate models (203) as the number of models in the model library. To run the method disclosed effectively, the said model library shall have at least two independent AI models enclosed.

After having a set of surrogate models (203) then comes the verification process. Having plenty of models at hand, each model is verified according to accuracy relative to the verification data. Models having a pre-defined accuracy are presumed to be valid models for cost estimation. In our studies, we have observed that 90% accuracy can easily be achieved through this method where number of models in the model library is 15.

Model having the highest accuracy score is presumed to be the best possible model to predict future costs and further generate further offers automatically. If all scores of accuracy is bellow the desired score of accuracy, the number and variance of models in the model library should be increased, the data should be accumulated more and inconsistent data generation should be avoided.

Upon selection of the method, the surrogate model (203) selected becomes the prediction model (204) and is the prior model to make estimations for that specific supplier. By this selection, we take into account that each supplier has a distinct cost structure, different set of data and unequal numbers of dimensions of cost. Therefore the invention disclosed generates a model taking into account of specific variables and factors of cost.

Besides, a relational vector space is generated from the previously made procurements at hand. When a new request is made, the method disclosed pins the point of the request and calculates a path in between performed goods and services and the request. Via such a calculation, a clustering for each supplier is done by grouping the similar experiences at around the request. The low the distance calculated, the high experience the supplier has. This also contributes to decision support at procurement level.

In the third step, the selected model is used to generate cost estimations based on the data of the new coming action data. The prediction model is now executed to make predictions of cost or any equivalent (i.e. offer, bid) for the supplier. For example for machining company, the company can input the 3D model, and specs of the part to be produced (material type, amount milled, time constraint, coating, etc.) to the web interface and get an instant result on a model trained and selected with regards to parameters and physical realities of the specific supplier.

The invention disclosed calculates supplier-specific “should” price and provides supplier experience assessment for the products to be manufactured. Highly accurate (>90%) supplier specific machine learning models that reflect the cost characteristics of suppliers are created with supplier’s manufacturing data rather than using generic market average models. With this approach, companies prevent cost losses during the supplier selection, while minimizing the risk of wrong supplier selection.

Fig.1

is a brief visualization of the invention disclosed.

Examples

PTL1:

NPL1:

Claims (4)

1- A method for estimating the cost of a third party comprising steps of;

1. gathering of previous data of jobs completed by the third party and save it on the memory element of the said computer;
2. splitting the data into two parts; the first part being training data and second part being verification data;
3. plurality of models of artificial intelligence in the model library being present on the memory element;
4. training first artificial intelligence model with the training data to obtain a first surrogate model;
5. repeating (d) for each model to obtain surrogate models until each model (104) in the model library (103) is trained to generate a corresponding surrogate model;
6. running each consecutive surrogate model (203) on verification data (202) to obtain a result;
7. comparing the results obtained at (f) with actual results of the verification data; marking based on accuracy;
8. ranking surrogate models (203) based on accuracy; picking the highest ranking model to save as prediction model (204);
9. mapping previous items procured from the said third party to generate a space
10. getting input data on an action to estimate third party costs on;
11. calculating a multi-dimensional distance from the request to nearest item in the space generated at (i); getting the inverse of the calculated distance;
12. running the prediction model (204) on the data inputted at (g) to obtain a cost estimation for the defined production.

The method in claim 1 wherein there are at least 2 models in the model library.

A computer programmed to implement steps of;

1. gathering of previous data of jobs completed by the third party and save it on the memory element of the said computer;
2. splitting the data into two parts; the first part being training data and second part being verification data;
3. plurality of models of artificial intelligence in the model library being present on the memory element;
4. training first artificial intelligence model with the training data to obtain a first surrogate model;
5. repeating (d) for each model to obtain surrogate models until each model (104) in the model library (103) is trained to generate a corresponding surrogate model;
6. running each consecutive surrogate model (203) on verification data (202) to obtain a result;
7. comparing the results obtained at (f) with actual results of the verification data; marking based on accuracy;
8. ranking surrogate models (203) based on accuracy; picking the highest ranking model to save as prediction model (204);
9. mapping previous items procured from the said third party to generate a space
10. getting input data on an action to estimate third party costs on;
11. calculating a multi-dimensional distance from the request to nearest item in the space generated at (i); getting the inverse of the calculated distance;
12. running the prediction model (204) on the data inputted at (g) to obtain a cost estimation for the defined production.

A memory element comprising instructions for a computer comprising steps;

1. gathering of previous data of jobs completed by the third party and save it on the memory element of the said computer;
2. splitting the data into two parts; the first part being training data and second part being verification data;
3. plurality of models of artificial intelligence in the model library being present on the memory element;
4. training first artificial intelligence model with the training data to obtain a first surrogate model;
5. repeating (d) for each model to obtain surrogate models until each model (104) in the model library (103) is trained to generate a corresponding surrogate model;
6. running each consecutive surrogate model (203) on verification data (202) to obtain a result;
7. comparing the results obtained at (f) with actual results of the verification data; marking based on accuracy;
8. ranking surrogate models (203) based on accuracy; picking the highest ranking model to save as prediction model (204);
9. mapping previous items procured from the said third party to generate a space
10. getting input data on an action to estimate third party costs on;
11. calculating a multi-dimensional distance from the request to nearest item in the space generated at (i); getting the inverse of the calculated distance;
12. running the prediction model (204) on the data inputted at (g) to obtain a cost estimation for the defined production.

Images