Machine Condition Monitoring

In this study, a deep learning model is presented for detecting bearing and induction motor faults. Vibration and acoustic data are used to encourage the model to learn representational and sequential features. In the model, convolutional neural networks (CNNs) are used to evaluate accelerometer and microphone data for bearing and induction motor diagnosis. A long short-term memory (LSTM) recurrent neural network is used to combine sensor information effectively, highlighting the benefits of data fusion. This approach encourages researchers to focus on multi-model diagnosis for constant speed data collection by proposing a comprehensive way to use deep learning and sensor fusion, and encourages data scientists to collect more multi-sensor data, including acoustic and accelerometer datasets.

Although they are crucial parts of heavy machinery, bearings are also prone to failure; between 50 and 60 percent of rotating machinery failures are caused by bearings. Data is gathered for the diagnosis and prognosis of bearing faults to help remedy this. Diagnosis relates to identifying a fault situation using recognized symptoms, whereas prognosis focuses on forecasting how the fault will turn out. To determine the states of machines, machine learning techniques are utilized. Deep learning is a branch of machine learning that focuses on training neural networks with several layers. Traditional (shallow) learning, which uses pre-defined features in data to identify machine conditions, and deep learning, which employs complex data types, are two methods of machine learning. In recent years, data fusion and transfer learning have become interesting and important subfields of deep learning. Transfer learning uses previously trained models as a starting point for new tasks by applying learned knowledge from the previous task on the related new task, whereas data fusion mixes many data types to improve comprehension. In this study, the need for large quantities of data fusion and transfer learning data is used to design a test rig for roller bearing fault identification and prognosis. Historically, academic researchers have concentrated on using the data they have gathered with their test rigs to undertake data analysis, rather than concentrating on the design of the test rig itself. Yet, there is a need for a test rig design that appropriately gathers clean multi-sensor data from a variety of bearing sizes to enhance deep machine learning algorithm research for bearing health diagnosis and prognosis via transfer learning. By offering useful methods that can be utilized to perform machine health diagnosis and prognosis on bearings, this work aims to close the gap between the needs of industry and the research community. This can aid in the early detection of faults, offer time frames for repair or replacement, and prevent catastrophic failures.

Reliability predictions on industrial equipment are becoming increasingly accessible as more devices become connected to sensors. Large strides in the field of predictive maintenance have been taken towards reducing downtime and optimizing operating schedules. However, extracting actionab le in-sight from vibration data coming from rotating machinery working with variable external loads re- mains challenging. Due to the realities of machinery at work in real-world environments, many factors can contribute to abnormal and misleading signals: quality of the connection, infiltration of noise, and the natural task (often varying) of the machine at hand. A method is proposed which uses vibration data to isolate the various states of a machine, using time-frequency audio processing techniques and neighbour search image processing. A 2D graphic image is generated which represents a fingerprint of that machine's current state. The fingerprint relies primarily on harmonic peaks, and therefore, features a robust signal to noise ratio. The 2D images are then stacked and compared to a 3D mapping surface, which is generated by a Gaussian Mixture Model. This helps define the relationship between states, and whether the data should be grouped together further. The method identifies different tasks performed by the machine and associates these tasks with a specific machine state. In the case of new tasks, the method creates a new parallel channel. By comparing the results of these isolated parallel channels, conclusions can be drawn related to machines operating under semi-variable conditions, performing different tasks that would otherwise disrupt typical vibration data sets. Tests were run on 2 different data sets and demonstrated the ability to properly separate different mechanical compo-nents mounted on the same shaft, as well as different rotational operating speeds of the machine.

The act of leveraging sensor data, such as accelerometers on connected industrial equipment, to diagnose and predict failures in roller element bearings is of growing interest in industrial and commercial applications. To do this, a large effort must be made to gather and process data, then apply and manage algorithms to make predictions and derive value. For these reasons, knowledge from previous models and tools to help in the evaluation and adaptation of new data sources using methods such as transfer learning and meta-learning have received significant attention in recent years. This paper aims to provide a methodology for combining such tools and presents a framework for an automated machine learning (AutoML) process by using a combination of pre-existing AutoML tools and new transfer learning methodologies. The framework uses recent observations in vibration analysis and deep learning architectures to abstract away many of the small, but intricate, decisions required for data preparation and model building. This allows for a unique data science strategy and dramatic reduction in the development time required to achieve competitive baseline models when compared to industry averages, while simultaneously improving the ability to address unseen environments. First, a signal processing engine is used to standardize the data, then fingerprinting and unsupervised learning is used to separate data into groups based on a machine's operating state. Meta-learning is then used to evaluate these states as tasks to be associated with a model architecture from a selection of two pretrained models for transfer learning, as well as one custom built model, based on the tasks themselves. In parallel, baseline models are generated using an existing AutoML library. Finally, the top performers from the existing AutoML models are combined with the transferred or built models in an ensemble method regime that uses a Support Vector Machine (SVM) to yield a final classification output...

Rolling element bearings are critical components of rotating machinery. For this reason, their condi- tion should be monitored and maintained for continuous and safe operation. To monitor the health of the bearing, sensors of various types can be used during operation to observe key characteristics of a bearing’s condition. However, to correctly identify a possible fault in a bearing, the sensor data must be recorded and further analyzed using a range of algorithms. To improve the performance of this analysis, additional measures will be taken to classify potential faults more accurately. First, the sensors installed on the bearings will monitor three operational characteristics: vibration, tempera-ture, and current draw. Vibration is generally considered to be the most effective method for fault diagnosis as the analog frequency signals can be plotted and compared with standard critical fre-quencies for common bearing faults. Meanwhile, temperature and current draw are both characteristics that will increase as these common faults progress towards failure of the component. Together, these sensors will monitor bearings with three different faults occurring on the outer race, innerrace, and ball, and will be compared to results of a healthy bearing. The second measure that will be used to improve the analysis involves the fusion of these three distinct sensor data types using a convolutional neural network to detect the fault type of a tested bearing. To reduce the amount of training data required, a unique implementation of GoogLeNet will be used to fuse the three different sensor data sets via GoogLeNet’s three RGB channels. Results demonstrate that this process can have strong fault type prediction certainty depending on algorithm training approaches and shows promise for prognosis health indication in future work.

This paper demonstrates various data augmentation techniques that can be used when working with limited run-to-failure data to estimate health indicators related to the remaining useful life of roller bearings. The PRONOSTIA bearing prognosis dataset is used for benchmarking data augmentation techniques. The input to the networks are multi-dimensional frequency representations obtained by combining the spectra taken from two accelerometers. Data augmentation techniques are adapted from other machine learning fields and include adding Gaussian noise, region masking, masking noise, and pitch shifting. Augmented datasets are used in training a conventional CNN architecture comprising two convolutional and pooling layer sequences with batch normalization. Results from individually separating each bearing’s data for the purpose of validation shows that all methods, except pitch shifting, give improved validation accuracy on average. Masking noise and region masking both show the added benefit of dataset regularization by giving results that are more consistent after repeatedly training each configuration with new randomly generated augmented datasets. It is shown that gradually deteriorating bearings and bearings with abrupt failure are not treated significantly differently by the augmentation techniques.

Through the intelligent classification of bearing faults, predictive maintenance provides for the possibility of service schedule, inventory, maintenance, and safety optimization. However, real-world rotating machinery undergo a variety of operating conditions, fault conditions, and noise. Due to these factors, it is often required that a fault detection algorithm perform accurately even on data outside its trained domain. Although open-source datasets offer an incredible opportunity to advance the performance of predictive maintenance technology and methods, more research is required to develop algorithms capable of generalized intelligent fault detection across domains and discrepancies. In this study, current benchmarks on source–target domain discrepancy challenges are reviewed using the Case Western Reserve University (CWRU) and the Paderborn University (PbU) datasets. A convolutional neural network (CNN) architecture and data augmentation technique more suitable for generalization tasks is proposed and tested against existing benchmarks on the Pb U dataset by training on artificial faults and testing on real faults. The proposed method improves fault classification by 13.35%, with less than half the standard deviation of the compared benchmark. Transfer learning is then used to leverage the larger PbU dataset in order to make predictions on the CWRU dataset under a challenging source-target domain discrepancy in which there is minimal training data to adequately represent unseen bearing faults. The transfer learning-based CNN is found to be capable of generalizing across two open-source datasets, resulting in an improvement in accuracy from 53.1% to 68.3%.

Vibration-based condition monitoring of rotating machinery represents the most widely adopted technique for the early detection and prevention of machine failures. Due to the high dimensionality and complexity of vibration data measurements, researchers have developed ways to reduce dependency on human expertise when analyzing complex signals. Machine learning algorithms such as convolutional neural networks are increasingly being developed to meet this challenge but are often trained with data that is not preprocessed to efficiently preserve physically meaningful information at the network input. Therefore, a novel algorithm is proposed using fused data from multiple sensors to present the network with rich time-frequency and spatial information to enhance prognosis. Spectrograms are computed from signals taken at multiple locations on the machine and layered to create a 3D input matrix. This allows the network to learn patterns in the time-frequency plane, as well as patterns relating the two signals through cuboidal convolution. The network’s performance is demonstrated with data gathered from accelerated lifetime tests. The inherent noise-canceling capabilities of the algorithm are evaluated. This work introduces new implementations of machine learning to machine health prog