Revelis has always been involved in IoT Analytics and Industry 4.0 predictive maintenance and beyond, implementing tools for monitoring production processes and advanced data analysis techniques aimed at predicting and preventing failures. In this article, we will describe an industrial project in which these techniques have been applied.
Industry 4.0 predictive maintenance: IoT Components in Industrial Textile Production
The textile sector has historically been one of the most important industries in Italy. It is the third manufacturing sector (after mechanical engineering and automotive) and according to the data from the Italian Trade Agency, it comprises nearly 45,000 active companies in the territory, employing 393.7 thousand people.
Our country is known not only for the production of clothing but also for its significant production of technologically advanced machinery for industrial weaving.
These looms are electromechanical machines that must guarantee high performance in terms of consecutive working hours and production capacity. Any failure in the production line could result in the loss of the entire production or the need for long and difficult repairs, leading to a slowdown in the production cycle.
In this context, Revelis has developed an application case.
The client is a leading company in the production of electronic components for textile looms. These components are IoT sensors installed on the feeders of textile looms. A feeder is an electromechanical device through which the warp thread passes.
It is one of the most important components of a loom, as it is responsible for maintaining the warp thread at the correct tension.
Problem and Application Context
The looms are designed to work continuously, as if the coils that feed the warp were infinite: there is a sort of buffer area where the coils that run out are replaced without stopping the loom. Each loom has several tens of feeders, and as you can imagine, if a thread breaks, the entire loom stops.
Therefore, it becomes essential to be able to predict warp thread break events, also known as “Yarn Break.”
Below you can see a schematic representation of the data communication.
Each feeder is equipped with various sensors, each of which measures a specific quantity with a sampling time of one minute. These measurements are transmitted by the control unit on an MQTT queue.
The values transmitted by each feeder are of different types and can be summarized as follows in the following table.
The values that the “status” variable can assume are:
- “OK”;
- “Yarn Break”;
- other values not used in this analysis.
The industry 4.0 predictive maintenance task, conducted in this case, aims to predict how much time will pass before the variable “status” assumes the value “Yarn Break,” in other words, how much time will elapse before the next yarn break.
Industry 4.0 predictive maintenance: the solution adopted
To address the problem just described, the data from sensors have been treated as multi-dimensional time sequences, and in particular, a mechanism called “sliding window” has been adopted. This approach allows grouping the sensor data into “windows” that “slide” over time.
Given a data window, it is possible to analyze a certain number of points beyond the window itself, checking if there will be a wire break event.
This procedure, essentially, allows to convert the predictive maintenance problem into a binary classification task, whose most relevant phases are:
- Data preparation: the most important processing steps were normalization and padding. Normalization is important to standardize the values and avoid undesired effects during the modeling phase, related to numerically different scales. Padding was necessary to make the sequences uniform.
-
Creation of the training dataset: The acquired data, organized into sliding windows, was divided into two portions:
- Training set: obtained by considering 70% (randomly chosen) of the entire dataset. This is the portion on which the learning phase of the model is performed.
- Test set: obtained by considering the remaining 30% of the dataset. This is the portion on which the model validation is performed.
-
Training of a neural network for predicting the wire break event. A Recurrent Neural Network (RNN) was used, constructed using TensorFlow and Keras. In particular, the network architecture consists of two layers:
- A masking layer, which instructs the network to ignore the values generated by the previously applied padding. This allows for efficient handling of sequences of variable length, focusing only on relevant data.
- A GRU (Gated Recurrent Unit) layer, which is similar to a Long Short-Term Memory (LSTM) layer but has a fundamental difference in its structure. While LSTM uses three gates (input, output, and forget gates) to manage the flow of information, GRU uses only two.
- Performance evaluation: Since it is a binary classification problem (breakage event in X minutes: YES/NO), the performance of the model was evaluated using the so-called “confusion matrix”.
Industry 4.0 predictive maintenance: the results obtained
The project took approximately two months for data understanding, preparation, and modeling. Twenty training sessions were conducted, resulting in an accuracy of 70% with an Area Under the Curve (AUC) of approximately 0.67.
The AUC score was calculated by measuring the area under the Receiver Operating Characteristic (ROC) curve. The analysis of the ROC curve, depicted in the following image, was used to determine the optimal cutoff value.
The operating process
The model trained during the learning phase can be saved to a file, and its execution can be automated to make predictions upon the arrival of new data and determine whether the wire is about to break or not.
From a technological standpoint, the model has been “encapsulated” into a containerized REST service and deployed within the client’s cloud infrastructure.
To make the prediction process fully automatic, an “orchestrator” application has been implemented (using Spring Boot technology). The orchestrator listens on the MQTT queue, where feeder signals arrive, and for each received signal, it invokes the prediction REST service.
If the model predicts a wire break, the orchestrator generates an alarm and sends it to an appropriate topic on the MQTT queue.
Below is a diagram illustrating the system’s operational process.
Benefits for the Customer
Thanks to industry 4.0 predictive maintenance solutions, organizations can obtain multiple benefits. In the case of the discussed application, the presence of this service allows the customer to display appropriate alarms on the machine’s display.
The visualization of these alarms enables the operator to promptly intervene in the replacement of the wire coil (in the buffer area) and thus avoid the shutdown of the entire frame. Lastly, this functionality has helped make the customer even more competitive in the market as the only one offering this type of solution.
Author: Massimiliano Ruffolo
If you also want to improve the efficiency of your company through predictive maintenance solutions in the Industry 4.0, contact us. One of our consultants will respond to you promptly to offer you personalized advice.
