MODERN ML METHODS FOR FAILURE

ANALYSIS AUTOMATION ON CLASSICAL ERP

SYSTEMS

DOI: https://doi.org/10.37811/cl_rcm.v8i5.14782

Modern ML methods for failure analysis automation on classical ERP

systems

Typically IT teams working with ERP systems have little or no knowledge of artificial intelligence and

more specifically Machine Learning (ML) and Natural Language Processing (NLP) models because

their working environment is mostly focused in supporting usually commercial ERP systems like SAP,

Oracle ERP, Microsoft Dynamics, etc., so they focus on the functional aspects and sometimes in some

therefore to achieve significant savings in time and IT human resources consumed in failure analysis.

Another objective is to share with the reader several lessons learned by the researchers while

investigating the available literature and while experimenting and testing with several of the existing

models available in Python and C# languages when comparing different technology platforms. This

should provide valuable information to IT managers, project managers, developers, and testers who

normally work with ERP systems and not with AI, so they are not so familiar with it.

Normalmente, los equipos de TI que trabajan con sistemas ERP tienen poco o casi ningún conocimiento

de inteligencia artificial y, más concretamente, de modelos de Aprendizaje Automático (ML) y

Procesamiento del Lenguaje Natural (NLP) porque su entorno de trabajo se centra principalmente en

dar soporte a sistemas ERP normalmente comerciales como SAP, Oracle ERP, Microsoft Dynamics,

etc., por lo que se enfocan en los aspectos funcionales y, a veces, en algunos entornos de desarrollo

propietarios como el lenguaje ABAP para sistemas SAP, para realizar personalizaciones muy

específicas. El presente trabajo de investigación pretende dar una visión detallada del estado del arte de

los modelos ML y NLP que podrían ser utilizados en un entorno ERP clásico y además persigue el

objetivo de investigar la viabilidad técnica y la facilidad o dificultad de dotar de inteligencia artificial a

un sistema ERP clásico que no disponga de ella con la intención de automatizar el análisis de errores y

fallos en el sistema ERP que por su volumen es difícil de gestionar por operadores humanos. Se persigue

están tan familiarizados con ella.

Palabras clave: artificial intelligence, machine learning, automation, supervised learning, erp

The goal of current research work is to provide details of the state-of-the-art ML and NLP models that

knowledge foundation. Later, in the methodology section, it is explained all the experimentation steps

done for this research.

The industrial revolution brought with it the emergence of large corporations in the United States and

Europe. These large scale organizations, many of them dedicated to the manufacture of a diversity of

products, had the need to implement inventory and manufacturing controls for which it was necessary

to have adequate inventories of raw materials. Thus it appeared the planning of material requirements

and the need to have information systems for inventory control, material purchase and production

planning. Information systems quickly evolve to MRP (materials requirements planning) and later to

ERP (enterprise resource planning) systems around the year of 1960. Also, about the same time, the

Artificial Intelligence (AI) has been a field of research for several decades, but in the twenty-first century

Questions), “Intelligence is the computational part of the ability to achieve goals in the world. Varying

Applicaons of AI) defines AI as follows: ‘It is the science and engineering of making intelligent

the 21st century work with information technologies, software development, data science, artificial

intelligence, automatic learning and other disciplines similar to data processing with computers, may

think that artificial intelligence is something that was born in this 21st century, but it is not so. In fact, it

is considered that the discipline and field of knowledge currently known as 'artificial intelligence' was

born in 1956 in a workshop held at Dartmouth University(Chow, 2021), in the city of Hanover, in the

took place, they were somewhat disappointed with the research papers submitted to the journal Annals

of Mathematical Studies. It was thought that the contributors to the journal for some reason did not focus

was that any feature or aspect of both learning and human intelligence could be described in a simple

but very precise way as if they were instructions to be followed step by step or as if they were a procedure

or an algorithm that a computer could then simulate with a program in some programming language.

In 1956, what many consider to be the first AI computer program was deveZ|loped, which was intended

to mimic humans in their problem-solving ability. This program was called the logic theorist and the

code was written by programmers Allen Newell and Herbert A. Simon.

Later in the 1960s to 1970s, "expert systems" were developed, which to this day use a series of rules

and knowledge bases to solve specific problems in various fields, but ultimately use deterministic logic

with classical programming based on typical concepts such as variables, iterative structures and single

and multiple conditionals to translate the logic of business rules to a series of conditionals and result

as the previous decade created great expectations for AI, possibly exaggerating its capabilities in order

to achieve attractive economic profits from the sales of AI software in the market. However, in practice

there were really poor advances in trying to use AI in real business applications and this led to a loss of

credibility and interest in investing money and resources in AI software.

In 1980, despite "The Winter of AI," a new idea or invention emerged that rekindled interest in artificial

neural networks. This idea was called "back-propagation". Neural networks usually have multiple

layers, which are called "hidden layers" and work as in mathematical differential calculus and the "chain

rule" to derive, there may be a function that in turn calls another function, i.e. they are nested functions

and the innermost function is calculated first and delivers its results to the outermost function that called

learning and neural networks of all types (convolutional or CNN, recurrent or RNN, etc.) that made it

possible to apply AI to image recognition and natural language processing.

It is from 2010 onwards that AI becomes more general use as it is introduced in industry and in everyday

life thanks to social media and mobile devices that are now accessible to everyone. In the year 2024, the

time this paper is written, use cases are so common in situations like whatsapp chatbots for natural

language processing in all kinds of commercial and government internet sites, computer vision with

applications like facial recognition used to unlock cell phones, robotics which is mostly used in industrial

environments, and other sectors such as healthcare, finance, transportation and entertainment, the latter

where it is very common to see entertainment platforms such as Netflix and Spotify analyzing customer

preferences and consumption habits and making suggestions based on these consumption patterns. Large

Machine Learning (ML) is a sub-field derived from AI that specializes in applying statistical methods

to make classifications or predictions. ML works with data sets and there are several dozen different

computer programs that have been developed and tested over the decades that are called models. These

models perform statistical calculations by applying various methods that statisticians and

mathematicians have developed since at least the middle of the 20th century. The logic or procedure

used to perform these calculations is called an algorithm and each model is based on a given set of

parameters. When you want to work with a model, you feed it with data sets. Before a dataset can be

used, it is necessary to analyze its columns of information and perform data cleaning or data preparation.

For example, if there are columns containing date data, it is necessary to put them in a format that the

are missing values. This happens frequently, for example, in the field of medicine, where there is no

consistency of data because each hospital or medical office, public or private, may have different

processes and procedures for capturing and processing data on patients and their diseases, diagnoses,

treatments, laboratory studies, follow-up information, etc. For all these reasons, it is necessary to

standardize the datasets in order to curate them and have cleaner information to feed the ML models.

Once the data is clean, the next step is to feed the model waiting for it to perform either a classification

task or a prediction or prognosis task. In ML, the model is said to learn based on its experience with the

data without the need to explicitly program it (write specific code) to recognize every possible scenario.

Therefore, it is said that the model learns and then it could be stated that there are different types of

learning. There are many authors who claim that there are at least three types of learning, so it can be

2. Unsupervised learning

3. Reinforced learning

use case

follows: ‘Reinforcement learning is a learning paradigm that learns to optimize sequential decisions,

Reinforcement learning addresses sequential decision-making problems that are often subject to

uncertainty, for example: multi-tier, multi-supplier inventory management with lead times under demand

uncertainty; control issues such as autonomous manufacturing operations or production plan control;

and resource allocation issues in finance or operations.

Reinforcement learning interacts with an environment that is a simulator of the stochastic dynamic

system.

Deep learning (DL) is a subset of machine learning and has demonstrated significantly superior

performance to some traditional machine learning approaches. DL uses a combination of multi-layered

artificial neural networks, processing- and data-intensive training, inspired by our latest understanding

of human brain behavior. This approach has become so effective that it has even begun to surpass human

capabilities in many areas, such as image and speech recognition and natural language processing. DL

models process large amounts of data and are generally unsupervised or semi-supervised.

natural. A conclusion that seems obvious, but that is very important to emphasize because within single-

layer ML there is the "artificial neural network" or ANN model and in the computational or information

technology profession it is ignored the fact that all computational neural networks (recurrent,

convolutional, adversarial) used in DL (multilayer) are also artificial even if their name does not include

that word.

DL learns features and tasks directly from data which can be images, text or sound. In medicine, for

example, deep learning models are used for image interpretation in imaging or radiology, deep learning

models learn to classify input images into appropriate categories.

According to Beysolow (Beysolow II, 2018), ‘natural language processing (NLP) is a subfield of

techniques of conversion of words to vectors (vectorization) it is also considered a part of ML focused

on human-computer communication. NLP addresses the inherent problem that while human

communications are often ambiguous and imprecise, computers require unambiguous and precise

messages to be able to communicate between machines.

Several NLP models that could be used in a classical ERP system to investigate the technical feasibility

and the ease or difficulty to provide artificial intelligence to a classic ERP system were explored.

Two software development platforms, Python and Microsoft ML.Net, were investigated. The

corresponding IDEs were installed and various packages or extensions required to work with ML and

Also, in the part of ERP system, several ABAP programs were written to emulate a production

environment in SAP ERP where multiple job failures would occur. The job logs of the failures were

downloaded to local files and were fed as input to Python and C# programs executing NLP methods to

experiment with.

In the following paragraphs the steps of the experimentation are described from selecting the

programming language and references to where to download and install integrated development

environments (IDE) up to ERP software development, integration and execution of the components. It

is provided also a discussion of the software development in Python, and C#.

According to (Ghosh, T., & Kumar, S., 2022), the mathematical model for Bag of Words is given by

Term frequency.

Term frequency, is the relative frequency of term “t” within document “d”.

A different notation can be:

community

random messages sent to job log when forcing the program to terminate.

Jobs with names ZJOB_DTI_JLG_NLPGEN1 to ZJOB_DTI_JLG_NLPGEN5 were created and

released (Figure 10) to run every five minutes starting at arbitrary date and time for our experiment.

Transaction code SM36 is used to create new jobs. Transaction code SM37 is used to monitor existing

jobs. It’s also possible to repeat scheduling of existing jobs. The reader should notice that all job and

program names start with letter ‘Z’. They could also start with letter ‘Y’. Name spaces starting with

either ‘Z’ or ‘Y’ are for custom objects created in SAP. This applies also to other objects like function

modules, include programs, tables, views, data elements, domains, even classes can be named following

this naming convention. Since SAP is a proprietary software, all standard objects have names that start

TBTCO is the Job Status Overview. It is necessary to test condition TBTCO-STATUS = ‘A’ to read

failed jobs (Figure 11). To read job status for any failed jobs in an ABAP program we can use conditions

like the following:

The FM ‘BP_JOBLOG_READ’ is used to retrieve job logs. This FM call was put on program

Z_DTI_JLG_DWNLD (Figure 12). The table TBTCO is used as driver table to know which are the

jobs that failed for a specific date and provides key search fields like failed jobname and joblog (TemSe

object name) field which is a search key with values like ‘JOBLGX01000200X43494’ that is also used

by the referred FM to search for failed job logs, finally the method ‘FILE_SAVE_DIALOG’ is used to

save or download job logs to a local file i.e.: the output file ‘ZDOWNLOAD_JOBLOG.TXT’ to be

loaded in Python and C# NLP models (Figure 13). The output path in our experiment was arbitrarily

defined as "E:\DATAMV\00.PhD\Python-NLP\PycharmProjects\Bag-of-

environment but this could easily be implemented in production to retriev and download real job logs

for real failed jobs. Normally, when a job fails it produces an ABAP dump and it can be displayed with

SAP transaction code ST22 (Figure 14). Such information could also be retrieved for NLP and not only

the job log.

An experimental python program was writen in Pycharm importing NLTK library which supports the

NLP models in Python (Figure 15). Then, Python program reads file ‘ZDOWNLOAD_JOBLOG.TXT’

from local path, in this case "E:\DATAMV\00.PhD\Python-NLP\PycharmProjects\Bag-of-Words2\"

An experimental C# program was writen in Visual Studio 2022 importing Micrsoft.ML library which

supports the NLP models in C# (Figure 18). Then, C# program reads file

‘ZDOWNLOAD_JOBLOG.TXT’ from local path, in this case