Predictive analysis in industry

Predictive analytics is a field of knowledge that allows you to make informed decisions, prepare for unforeseen situations and anticipate all kinds of emergencies. Recently, predictive analysis has been actively used in industry: based on historical data, the model makes a probabilistic forecast of the device’s behavior in the near future. This paper provides a comparative analysis of two predictive models, which both could self-learn and had the property of self-correction. The accuracy of predicting the development of a defect in industrial equipment, as well as the prediction horizon, were evaluated. Particular attention is paid to the peculiarity of working with data obtained from production sensors.


Applications and results
Digitalization of equipment -the connection of digital sensors, interfaces, and data acquisition systems -enables the use of automated process control systems.It monitors equipment performance, warns of critical deviations, and protects units from operation in prohibitive modes.But despite their complexity, they work only in real time, signalling danger or shutting down equipment at the moment.
Predictive analytics (PA) is used in predicting future events.PA analyses current and historical data using techniques from statistics, data mining, machine learning, and artificial intelligence to make predictions about future condition.PA is a mix of several branches of science: mathematics, information technology, business processes of production and management.
Among the examples that support the relevance of predictive analytics, the main one is intelligent machine maintenance, which is aimed at:  Reduction of downtime of process equipment. Reduced maintenance and repair costs. Reduction of maintenance staff workload and increased efficiency with existing equipment. Fewer rejects. Improved quality of management decisions. Increased energy efficiency of production. Reduction of infrastructure development costs during system deployment. Optimization of the procurement system for missing repair components. Increased reliability of production systems.

Stages of developing a predictive model
There are many predictive modelling platforms on the market right now.However, let's look at the basic requirements for these platforms and their results.Let's consider the stages of building a predictive model: Let's look at where most of the PA's difficulties lie.

Pumping equipment
On the schematic representation of a pump equipment you can see a number of points highlighted.There are set of vibration sensors, part of them piezo-crystals accelerometers and others are current-vortex proximeters.Sensors measure vibration in three directions: vertically, horizontally and axial.After receiving raw data, transforming force or induction to voltage, digitizing and additional calculations we get a familiar acceleration / or movement in µm.This movement characteristic is often calculated as a squared mean and called overall level.


The work uses data upload from an enterprise where industrial pumping equipment is installed.
The units are equipped with sensors (stationary and mobile) that measure various indicators of temperature and vibration in real time.


The task used an unload from the database and divided into 4 tables:  The Trains table contains information about pumping equipment.Each mechanism has a unique idTrain number, which is a surrogate key in the database.

Conclusion
To estimate the model, we considered the f1 metric.F-score or F-measure is a measure of a test's It is calculated from the precision and recall of the test, where the precision is the number of true positive results divided by the number of all positive results, including those not identified correctly, and the recall is the number of true positive results divided by the number of all samples that should have been identified as positive.Precision is also known as positive predictive value, and recall is also known as sensitivity in diagnostic binary classification.
The LSTM-based model gives an accuracy of f1 = 0.9 -0.8, which is a good enough result.However, if we look at the graphs, we can see that the model does not predict sharp outliers well enough, so we need to combine several models.This is a starting point for further work.

Figure 4 .Figure 5 .
Figure 4. Comparison of predicted and real value of vibration velocity at point 3(Horizontal direction) Gathering requirements-what we want to know. Collecting data. Preparation and analysis of data.Unstructured data are converted into a structured form.Checking data quality, eliminating errors and omissions.Formation of training and test samples. Statistics, machine learning.All predictive analytics models are based on statistical and/or machine learning techniques.Machine learning often has an advantage over traditional statistical methods, but statistical methods are usually always involved in the development of any predictive model. Predictive model.Model development and testing. Prediction and monitoring.After successful testing, the model is deployed for daily forecasts and decision-making processes.Results and reports are generated by the model for the management process.The model is monitored regularly to make sure it is producing the correct results and making accurate predictions.