A deep learning-based intelligent online warning method for gear wear damage based on oil vibration

Gearboxes, as essential connecting and transmission components in mechanical equipment, have been widely used in modern industrial development. Gearboxes are prone to malfunction or even failure due to complex structures and harsh working environments. This article takes online monitoring of gear wear and damage as the research object and studies the fault diagnosis method of gear multi-source heterogeneous parameters for oil monitoring and vibration monitoring. The Yolov5 model is used to identify multi-objective wear particles. The experimental outcomes suggest that the optimized detection method can sensitively reflect the evolution process of gear wear.


Introduction
Oil monitoring and vibration monitoring have long been regarded as two technical categories with vastly different focuses.A single fault diagnosis method has limitations in terms of applicability and diagnostic accuracy, often unable to fully diagnose the fault conditions of complex mechanical equipment such as gearboxes [1].The actual operation of the equipment is complex, and there may be multiple types of wear and tear simultaneously [2].Related studies have shown that there is a close correlation between vibration monitoring and oil analysis.The comprehensive application of the two fault diagnosis methods for equipment condition monitoring and fault diagnosis is not only feasible due to their effectiveness but also theoretically inevitable.This project aims to optimize the fault diagnosis method for the oil vibration gearbox.By simulating the full life cycle fault test of the gearbox, the oil information and vibration signals during gearbox wear are obtained, and a gear wear damage oil vibration fault diagnosis model based on deep learning is constructed.This can achieve a safe environment and reliable performance of the gearbox, preventing economic losses and casualties caused by gearbox failures, which is of great significance.

2LSTM-based vibration analysis method for gearbox
This section introduces the long and STM-based neural network, which is utilized to extract features from the vibration signal raw data collected from the gear full-life acceleration experiment [3].In response to the traditional feature extraction method that relies on personnel's professional knowledge and manually extracts features from the raw data, which is difficult to handle complex fault information, this article uses CNN to extract features from the vibration raw data and input them into LSTM for calculation [4].From a data perspective, automatic learning of data features eliminates the impact of noise on the data and improves the impact of manual feature extraction on the results in traditional feature extraction methods.

Data preprocessing
Due to various factors interfering with the testing system and testing process, the aboriginal vibration signal monitored by the vibration transducer in data preprocessing must be mixed with a lot of noise during output.Therefore, it is necessary to preprocess the vibration signal, correct waveform distortion, reduce excess constituents in the original vibration signal, remove noise and interference, highlight the required data, and make the vibration data more realistic and closer to the original vibration data [5].
The vibration signal and precise data collected in vibration measurement testing often make zero drift of the amplifier, leading to a departure from the baseline in the vibration signal data caused by temperature changes, the instability of low-frequency performance outside the frequency range of the sensor, and interference from the surrounding environment, even the magnitude of the deviation from the reference can change as time goes on [6].This article uses the least squares method to fit the curve, remove zero drift of vibration data, and eliminate polynomial trend terms.As shown in Figure 1, it is a multi-trend term model for eliminating vibration signals.
. The five-point cubic smoothing algorithm is a type of smoothing filtering algorithm commonly used in signal processing.Other commonly used methods for smoothing filtering include spline difference function smoothing and average calculation smoothing.The spline difference function smoothing method utilizes spline interpolation to approximate the sampling points, resulting in a smoothing effect.This method is flexible and effective but has a complex calculation process and may not perform well in terms of smoothing amplitude [7].On the other hand, the average calculation smoothing method is relatively simple, but it may not yield satisfactory filtering results.In contrast, the five-point cubic smoothing method achieves smooth filtering by utilizing the polynomial least squares method to approximate the sampling points.This method produces good filtering results [8].

Dataset production
This article divides the vibration data into four datasets, corresponding to four stages, normal wear, slight wear, abnormal wear, and severe wear in gear wear faults [9].The time-domain signals of each stage of wear are shown in Figure 2.Each type of fault is divided into 1000 samples, and 70% of these 4000 samples are divided into a test set and 30% into a training set.

Experimental result
Vibration data is trained by using the building of complex networks from the convolutional neural network.The given network initialization training step size is steps=2400, and the selected learning rate is lr=0.001.It can be seen that the final loss value can reach 0.05, with an accuracy rate of 96.88%.
After the algorithm training is completed, the full life cycle vibration data used in this article is input into the model for operation, and the wear status results of the gearbox are obtained.It can be seen that 0-50 hours is the normal wear stage, 50-80 hours is the mild wear stage, 80-110 hours is the abnormal wear stage, and after 110 hours is the severe wear stage.
Time domain waveform analysis is usually the most intuitive diagnostic method.For some faults with obvious characteristics, time domain waveform can be used for preliminary and intuitive judgment.Corresponding to different faults, their vibration waveforms exhibit different shapes.If pitting failure occurs, it will cause significant amplitude modulation in the waveform [10].Cracks or tooth breakage faults will cause significant periodic impacts in the waveform.
Traditional time-domain statistical feature parameters can be divided into two categories.One type is dimensional feature parameters that are closely related to the structure, working environment, and operating conditions of the gearbox, and there is no comparability between different gearboxes.The other type is dimensionless feature parameters, which are comparable for different gearboxes in certain situations.The dimensional feature parameters mainly include peak, mean, variance, mean square amplitude (effective value), and average amplitude.This section will extract features of the entire life cycle vibration data used, selecting 6 timedomain feature parameters.As shown in Figure 3, the time-domain feature parameter variation diagram of the entire life cycle vibration data is presented.By extracting the feature parameters of the time-domain features, wear damage can be effectively diagnosed.

Conclusion
This article takes the gearbox as the research object and conducts research on multi-source heterogeneous fault diagnosis of the gearbox.This article proposes two algorithms for diagnosing oil monitoring and vibration data analysis and finally fits the abnormal indicators of the two diagnostic methods.By collecting oil image data and vibration data throughout the gearbox's entire life cycle, the oil image features and vibration data features were identified and diagnosed, realizing the simultaneous recognition of multi-objective wear particles [11].The test accuracy reached 96.88%, with a good accuracy.The experimental process was validated through oil monitoring and vibration data analysis [12].By comparing the surface wear of gears observed during the experiment, it was found that the proposed abnormal indicator of oil vibration gear wear damage can sensitively reflect the actual wear of gears.

Figure 2 .
Figure 2. Data Smoothing(X:position in data Y:noise reduction effect).