Source-linked AI summary
Intelligent Bearing Fault Diagnosis Method Combining Mixed Input and Hybrid CNN-MLP model
V. Sinitsin, O. Ibryaeva, V. Sakovskaya, V. Eremeeva
TL;DR
Rolling-bearing failures threaten rotating machinery, while diagnostic information depends on sensor placement and conventional methods often use separate data representations. The paper proposes a hybrid CNN-MLP model combining HHT images with numerical signal-power inputs from a shaft-mounted wireless sensor. The hybrid model outperforms separate CNN and MLP models, achieving 99,6% detection accuracy versus 98% for CNN and 81% for MLP, although experiments used one shaft rotation frequency.
Problem
Rolling-bearing failures can affect rotating machinery, and diagnostic information is collected using sensors whose placement influences the available acceleration data.
Method
The proposed hybrid CNN-MLP model simultaneously processes HHT images and numerical signal-power inputs from a sensor mounted directly on the rotating shaft.
Results
99,6% detection accuracy was achieved for bearing faults, compared to 98% for CNN and 81% for MLP models, and the hybrid model was superior to both in isolation.
Takeaways & Limitations
The study provides a mixed-input diagnostic method and a shaft-mounted-sensor dataset for bearing-fault detection.
Takeaways & Limitations
Experiments were carried out for only one shaft rotation frequency; future work will test signals at different shaft speeds.
Abstract
from arXiv · showhide
Rolling bearings are one of the most widely used bearings in industrial machines. Deterioration in the condition of rolling bearings can result in the total failure of rotating machinery. AI-based methods are widely applied in the diagnosis of rolling bearings. Hybrid NN-based methods have been shown to achieve the best diagnosis results. Typically, raw data is generated from accelerometers mounted on the machine housing. However, the diagnostic utility of each signal is highly dependent on the location of the corresponding accelerometer. This paper proposes a novel hybrid CNN-MLP model-based diagnostic method which combines mixed input to perform rolling bearing diagnostics. The method successfully detects and localizes bearing defects using acceleration data from a shaft-mounted wireless acceleration sensor. The experimental results show that the hybrid model is superior to the CNN and MLP models operating separately, and can deliver a high detection accuracy of 99,6% for the bearing faults compared to 98% for CNN and 81% for MLP models.
1 Introduction
Rolling-bearing condition is important for rotating machinery, motivating extensive signal-processing and machine-learning approaches to fault diagnosis. This paper addresses limitations of conventional housing-mounted sensing with a mixed-input hybrid CNN-MLP method using shaft-mounted acceleration data.
- Rolling bearing failures account for 30% or more of all failures in rotating equipment.
- Bearing condition monitoring and intelligent diagnosis are critical aspects of rotating-system design and maintenance.
- Conventional vibration diagnosis typically extracts time-, frequency-, or time-frequency-domain features before machine-learning-based fault recognition.Examples include RMS and kurtosis, FFT-based analysis, short-time Fourier transform, wavelet transform, and HHT-related methods.
- Machine-learning methods used for bearing and related fault diagnosis include k-NN, naive Bayes, Bayes networks, support vector machines, and Random Forests.
- CNNs have been widely applied to bearing diagnosis, often after vibration signals are converted into time-frequency or vibration images.
- The study uses a dataset collected by a wireless acceleration sensor mounted directly on the rotating shaft.The paper presents this dataset as distinct from commonly used housing-mounted CWRU accelerometer data.
- The proposed method combines CNN processing of HHT images with MLP processing of signal power at resonant frequencies.It therefore uses image and numerical inputs within one hybrid model.
- Experiments demonstrate that the hybrid model is superior to CNN and MLP models applied separately.
2 Proposed intelligent fault diagnosis method
The proposed method converts linear and angular acceleration signals into complementary inputs for a hybrid CNN-MLP model. Linear signals become HHT time-frequency images, while angular signals provide frequency-band power values for MLP processing before feature concatenation and classification.
- Angular signal-to-number conversion: Angular acceleration windows are processed with FFT to obtain signal-power values around the shaft’s first and second torsional frequencies, N1 and N2, for MLP input.The frequency-band values are computed over repeated overlapping windows for training and testing data.
- Linear signal-to-image conversion: Linear acceleration windows are transformed with HHT into 32 x 32 time-frequency images for CNN input.The input windows last 125 ms and overlap with a 50 ms offset.
- Linear signal-to-image conversion: The first three IMFs are used to produce the Hilbert spectrum, based on experiments reporting effective results with reduced computational cost.EMD decomposes the signal into intrinsic mode functions, after which the Hilbert transform provides time-varying amplitude and frequency information.
- Angular signal-to-number conversion: The angular features N1 and N2 are computed around approximately 240 Hz and 820 Hz, the first and second torsional natural frequencies of the 3/4-inch shaft.These frequencies are identified using shock response spectrum analysis, and FFT components around the peaks are summed.
- Method overview: The method processes mixed acceleration data through separate CNN and MLP branches before concatenating their learned features for bearing-fault classification.CNN extracts hierarchical features from structural image data, while MLP handles numerical inputs.
- Hybrid model: All hybrid-model inputs, including N1, N2, and image-pixel intensities, are normalized to the range [0,1].The resulting mixed inputs are image-number pairs derived from linear and angular acceleration measurements in the same time interval.
3 Case study
The case study evaluates bearing-fault diagnosis using shaft-mounted wireless acceleration data, transformed into HHT images and FFT-derived inputs for a Hybrid CNN-MLP model. Preliminary visualization identified expected class separability and confusion patterns, while experiments showed complementary CNN and MLP behavior in the hybrid model.
- Experimental rig: The experimental rig used a shaft-mounted WAS prototype with three one-axis MEMS accelerometers near the test bearing.The shaft was driven by an AC motor, and a revolution indicator measured shaft frequency.
- Experimental rig: Five bearings represented normal operation, three single faults, and one combined fault involving inner race, outer race, and ball damage.The tested fault classes were inner race, outer race, ball, and combined defects.
- Data preparation: Each 3,897-point window covered 2.5 shaft revolutions; HHT produced 32 x 32 time–frequency images, while FFT calculated N1 and N2 from angular signals.The window shift was 1,559 samples, corresponding to one shaft revolution.
- Preliminary analysis: The unlabeled visualization showed three clusters: normal, combined fault, and the other defects, with inner-race and ball signals nearly overlapping.This suggested easier recognition of normal and combined signals and more errors separating inner-race, outer-race, and ball faults.
4 Conclusions and future work
The paper presents a hybrid CNN-MLP bearing-diagnosis method that combines mixed inputs from a shaft-mounted sensor and outperforms isolated CNN and MLP models. Its current evaluation is limited to one shaft rotation frequency, motivating tests at different speeds.
- The hybrid CNN-MLP model simultaneously processes numerical inputs with an MLP and HHT images with a CNN.
- The dataset uses a sensor mounted directly on the shaft, and this acquisition method is reported to be more sensitive to bearing defects.
- The dataset is made openly accessible for use by other scientists.
- The hybrid model is superior to CNN and MLP models operating in isolation.
- 99,6% detection accuracy is reported for bearing faults, compared with 98% for CNN and 81% for MLP models.
- The experiments were conducted at one shaft rotation frequency, with future work planned for signals recorded at different shaft speeds.