Intelligent planetary gear fault diagnosis system based on MEMS acoustic emission sensor | Microsystems & Nanoengineering

Microsystems & Nanoengineering volume 11, Article number: 126 (2025) Cite this article

609 Accesses

Metrics details

Early equipment fault diagnosis can identify potential risks, significantly reduce maintenance costs, and minimize property damage. However, vibration, strain, and force sensors operating at low frequencies with narrow bandwidths are insufficiently sensitive to fault information, making early fault prediction challenging. Here, we introduce a high-performance, cost-effective, and tiny-sized micro-electromechanical system (MEMS) acoustic emission sensor. This sensor utilizes a 10 × 11 hexagonal array of piezoelectric micromachined ultrasonic transducers with a chip size of 4 mm × 4 mm × 0.4 mm. The sensor is encapsulated using an epoxy/Al2O3 composite for acoustic impedance matching, and its overall size is Φ 16 mm × H 5.5 mm, with a weight of approximately 3 g. This acoustic emission sensor achieves a peak sensitivity of 88.4 dB (ref. V/(m/s)) at 335 kHz, and its sensitivity remains above 60 dB across the frequency range from 15 kHz to 620 kHz. In addition, combined with the residual neural networks, an intelligent fault diagnosis of the planetary gear is realized. This MEMS acoustic emission sensor can provide a promising approach for in-situ fault monitoring of highly integrated and miniaturized industrial equipment.

The development of industrial automation technology provides new driving forces for intelligent manufacturing, improving product quality and production efficiency while reducing labor intensity1. As key tools for automated production, the operational status of equipment is vital to product quality and production efficiency. The degradation path of equipment can be divided into five stages: regular operation, initial degradation, noticeable degradation, functional fault, and catastrophic failure2. If maintenance can be performed before the equipment enters the functional fault stage, it will significantly reduce maintenance costs and minimize property damage. Therefore, monitoring the operational status of equipment, promptly detecting and diagnosing potential faults is of utmost importance.

Sensors capture signals with various characteristics generated during equipment operation, which is a critical part of intelligent fault diagnosis systems. Numerous sensors have been applied in intelligent fault diagnosis systems. Vibration sensors monitor signals such as displacement, velocity, and acceleration of equipment3,4,5. They operate at low frequencies with narrow bandwidth, making them susceptible to mechanical resonance and noise interference. Strain6,7 and force sensors8 are used to measure changes in structural deformation or force. Their installation is restricted, and they are prone to fatigue and wear. The captured signals are typically slow-varying with low frequencies and few characteristics. In addition, these sensors only work when the equipment has apparent degradation, making it challenging to detect early-stage fault signals. Fortunately, the appearance of acoustic emission (AE) sensors offers a potential solution to these challenges. AE is a phenomenon where materials generate transient elastic waves due to the rapid release of energy9. The identical signal frequency spans a wide range from several kHz to several MHz10,11. Minor structural defects often trigger markable AE activity, promoting AE sensors more sensitive to abnormal information. Therefore, researchers have introduced AE technology into equipment fault diagnosis systems12,13, hoping that it could diagnose faults earlier than major equipment failures occur. However, the AE sensors used so far are traditional ceramic-based, which are large, heavy, and costly, making it challenging to meet the requirements for sensor miniaturization, low power consumption, and low-cost in integrated industrial equipment.

Micro-electromechanical system (MEMS) technology inherently delivers cost-effectiveness and mass-producibility of compact, low-power devices, positioning it as a promising solution for the above demands14,15,16. This has prompted growing research efforts on MEMS AE sensors. Didem Ozevin et al. designed various AE sensors that achieved high sensitivity near the resonant frequency, but with a very narrow bandwidth17,18,19. Subsequently, they expanded the bandwidth to 100 kHz–700 kHz by paralleling multiple units with different resonate frequencies20. However, the issue of uneven sensitivity remained. Butaud et al. produced a multi-frequency and wide-bandwidth AE sensor based on a capacitive micromachined ultrasonic transducer (CMUT), realizing wide-band AE signal detection21. However, the high bias voltage leads to high power consumption, and the harsh industrial environment also poses serious challenges to its reliability. In general, MEMS-based AE sensor technology still faces several challenges: (1) Narrow bandwidth and low sensitivity make it prone to signal loss; (2) The absence of proper acoustic impedance matching results in low sensitivity; (3) Poor encapsulation design and signal conditioning make the sensor susceptible to interference. Therefore, further improvement of MEMS AE sensors is required for applications in integrated intelligent industrial equipment.

This study develops a high-performance, cost-effective, compact MEMS AE sensor, which is validated through application in an intelligent fault-diagnosis system for planetary gears (PG). To address the narrow bandwidth and low sensitivity, we designed an AE sensor based on a piezoelectric micromachined ultrasonic transducer (PMUT). Subsequently, an epoxy/Al2O3 composite encapsulation method was proposed to achieve efficient acoustic impedance matching and compact device integration. Based on the encapsulated MEMS AE sensor, we established an intelligent fault diagnosis test rig for PG. Different gear fault states were tested through this system, integrated with residual neural networks (ResNet), thereby achieving multi-fault diagnosis of the PG speed reducer under various operational conditions.

Functioning as a core transmission element in industrial systems, the operational integrity of speed reducers critically governs the positioning accuracy of mechanical assemblies. Since gears are the core components of reducers, their fault condition monitoring is important. Figure 1 illustrates the working mechanism of the AE sensor for PG fault diagnosis. The meshing dynamics of PG systems inherently generate elastic stress waves during operation, propagating as structure-borne AE through transmission components including the planetary carrier, output shaft, and robotic arm. Progressive gear wear or localized failures induce detectable modulations in the spectral signatures of these AE signals, with failure severity correlating to specific frequency-domain features. By capturing these signals using an AE sensor mounted on the exterior and analyzing the signals, intelligent fault diagnosis of PG can be achieved.

PG fault diagnosis system based on MEMS AE sensor

The MEMS AE sensor developed in this study employs a PMUT-based sensing mechanism for acoustic detection, departing from conventional piezoelectric ceramic transducers that typically operate in bulk vibration modes. When incident acoustic waves interact with the surface of the sensor, the thin-film structure undergoes out-of-plane bending, enabling piezoelectric transduction of mechanical vibrations into electrical signals through the piezoelectric effect14,22,23. In this design, 20% scandium-doped aluminum nitride (ScAlN) is adopted as the piezoelectric sensitive layer for PMUT. Compared with aluminum nitride (AlN), ScAlN exhibits a significantly higher piezoelectric coefficient, which enhances sensitivity in acoustic sensing applications24. We employ hexagonal configurations for PMUT structures due to their superior receiving sensitivity compared to square and circular designs with consistent resonant frequency. Furthermore, hexagonal structures offer a better filling ratio than circular designs, enhancing area utilization in array arrangements, as detailed in our prior work16,25. Figure 2a illustrates the schematic of the hexagonal thin-film structure of the chip, the layers from bottom to top are silicon dioxide (SiO2), silicon (Si), molybdenum (Mo), ScAlN, and Mo, the detailed dimensional parameters are shown in Supplementary Table S1. The resonant frequency \(f\) of this structure in air is given by26:

where \(d\) is the side length of the hexagonal structure, and the parameter \({\lambda }_{0}^{2}\) represents the correction factor. For the designed regular hexagonal diaphragm, \({\lambda }_{0}^{2}\) = 12.81. \(h\), \(E\), \(\sigma\), and \(\upsilon\) are the diaphragm thickness, equivalent elastic modulus, equivalent surface density, and equivalent Poisson’s ratio of the diaphragm, respectively. To improve the reliability of the device and prevent diaphragm wear, a layer of protective material needs to be applied on the surface of the sensitive membrane, and the quality factor (Q) and resonant frequency will be reduced due to the loading effect by using an encapsulating medium24,27.

Fabrication of MEMS AE chips. a Structural schematic of the chips. b Detailed fabrication process flow of the ScAlN-based chips: (I) the bottom electrode (Mo), piezoelectric layer (ScAlN) and top electrode (Mo) deposited; (II) top electrode patterned; (III) the insulation layer (SiO2) deposited; (IV) insulation layer and piezoelectric layer patterned; (V) wire (Au) deposited and patterned; (VI) back cavities etched. c 10 × 11 device with chip size 4 mm × 4 mm and zoomed-in view showing the cell details. d SEM image showing the excellent crystal orientation. e The measured impedance characteristics of chips

The AE sensor chips were fabricated based on the ScAlN-Silicon-on-Insulator (SOI) platform, and its detailed fabrication process is shown in Fig. 2b. Firstly, a 300 nm/1 μm /300 nm Mo/ScAlN/Mo stack was deposited on the SOI by magnetron sputtering. Secondly, the top electrode was patterned using reactive ion etching (RIE), and the size of the top electrode is designed to be 70% of the cavity size to maximize the output charge of each cell. Thirdly, plasma-enhanced chemical vapor deposition (PECVD) was used to deposit a 200 nm layer of SiO2 on the top electrode, which serves to protect the top electrode and prevent electrical short circuits. Fourthly, a combination of dry etching and wet etching processes was employed to pattern the deposited oxide layer and the ScAlN piezoelectric layer, exposing the bottom electrode and forming via holes for wiring. Fifthly, a 100 nm gold electrode and wire were deposited and patterned to achieve electrical parallel connections between each cell and enable signal output. Finally, deep reactive ion etching (DRIE) was used to pattern deep cavities on the backside of the SOI substrate, releasing the films.

The morphology and impedance characteristics of the chips were characterized by using a scanning electron microscope (SEM) and a network analyzer, confirming consistency with the design specifications. Figure 2c presents the SEM images of the fabricated chip array and its sensitive cell details. The chip array consists of 10 × 11 diaphragms connected in parallel. This parallel configuration of multiple sensitive cells enhances the output charge of the sensor, thereby increasing the reception sensitivity. The chip size is 4 mm × 4 mm, with a thickness of only 0.4 mm, significantly smaller than the piezoelectric ceramics used in conventional AE sensors. Figure 2d shows the SEM of the chip cross-section. The thickness of each layer of the processed sensitive film is 300 nm/974 nm/310 nm, which is consistent with the design. Subsequently, the impedance characteristics of the sensor array were characterized using a network analyzer (E5080A, Keysight). X-ray diffraction (XRD) measurements revealed that the (002) peak of the ScAlN film appears near 36°, and the full width at half maxima (FWHM) of ScAlN film is 1.359°, which indicates that the c axis of ScAlN film is well arranged (Supplementary Fig. S1). As shown in Fig. 2e, the resonant frequency of the array is 1.047 MHz, which is 3.7% higher than the predicted result. This discrepancy is primarily attributed to variations in the thickness of each layer’s material and slight differences in the cavities’ diameter.

Acoustic impedance matching is a key factor in determining the energy transfer efficiency between the sensor and the acoustic wave transmission medium. Serious acoustic impedance mismatch will lead to energy reflection, thereby reducing the transmission coefficient of the acoustic wave energy. As shown in Fig. 3a, when an acoustic wave propagates through multiple material interfaces, the transmission coefficient \({T}_{13}\) for energy transfer from medium 1 to medium 3 is given by the following formula28:

where \({Z}_{1}\), \({Z}_{2}\), and \({Z}_{3}\) are the acoustic impedance of the mediums in each layer, 2 represents the matching layer, and \({k}_{2}\) represents the wave number of the matching layer. When the thickness of the matching layer \({l}_{2}=(2n+1)\lambda /4\):

Encapsulation and test of MEMS AE sensor. a Impedance matching principle. b Structure of the sensor. c Schematic diagram of the signal conditioning circuit. d Encapsulation process flow chart. e Optical topography of MEMS AE sensor

Equation (3) shows that when the thickness of the matching layer is an odd multiple of a quarter wavelength and the acoustic impedance of the matching layer is \({Z}_{2}=\sqrt{{Z}_{1}{Z}_{3}}\), the transmission coefficient \({T}_{13}\) is equal to 1, and the energy in the acoustic wave can be transmitted from medium 1 to medium 3 without loss. In this work, acoustic impedance matching between high-impedance metals and low-impedance MEMS AE sensors is essential to ensure efficient signal transmission and minimize energy loss. The acoustic impedance of the sensor is given by29:

where \({A}_{eff}\) is the effective area of the film vibration, which is about 1/3 of the diaphragm area \(A\), \({A}_{pitch}\) is the total area of the vibration cells, including the gap between the vibration cells, and \({c}_{media}\) is the medium sound velocity.

Single-phase materials inherently exhibit high acoustic impedance and solid-state properties, which pose challenges for direct integration with delicate microelectronic structures. Direct adhesion could potentially destroy the surface features of the chip through mechanical stress. To mitigate this issue while forming an impedance-matching layer, a hybrid composite material was synthesized using epoxy resin as the matrix and high acoustic impedance alumina (Al2O3) nanoparticles as functional fillers. This approach ensures structural protection of the chip and achieves the desired acoustic impedance characteristics through controlled dispersion of the nanoparticles within the polymer matrix, while simultaneously reducing the quality factor of the sensor to expand its bandwidth. The acoustic impedance \({Z}_{c}\) of the composite can be calculated from the volume fraction and elastic parameters of the base material and Al2O3 powder30, that is:

where \(\rho\),\(v\), \(K\),\(G\),\(c\) represent the density, volume fraction, elastic bulk modulus, elastic shear modulus and sound velocity of the material, respectively. Subscripts c, 1, and 2 represent epoxy/Al2O3 composite, base material (epoxy resin) and fillers (Al2O3), respectively.

To achieve acoustic impedance matching, improve reliability, and reduce electromagnetic interference, the sensor structure was carefully designed. As illustrated in Fig. 3b, the structure is composed of several layers arranged sequentially from bottom to top: a metal shell for electromagnetic shielding, a signal conditioning circuit, an AE sensor chip, and a metal lid. The adoption of Al shell is primarily due to its electromagnetic shielding capability, which is comparable to that of copper. Additionally, its acoustic impedance lies between the steel and the matching layer, effectively reducing acoustic reflection. Moreover, Al is lightweight and easy to manufacture. A customized signal conditioning circuit was designed to capture the charge signals generated by the AE sensor chip during operation. Figure 3c depicts the schematic diagram of the implemented signal amplification circuit, which employs Analog Devices’ AD8066 as the core component for voltage gain and signal conditioning. The first-stage charge amplifier provides a gain of 10 mV/pC, while the second-stage voltage amplifier achieves a gain of 40 dB. The chip is bonded to the predefined pad locations on the circuit board and electrically connected to the circuitry via wire bonding. After securing the chip and circuit within the metal shell, the epoxy/Al2O3 paste fills the entire shell interior, achieving acoustic impedance matching while providing effective protection for both the chip and circuit.

The detailed encapsulating process is illustrated in Fig. 3d. Firstly, epoxy resin and hardener are mixed in a 5:1 weight ratio and stirred thoroughly to form an epoxy resin prepolymer. Secondly, Al2O3 powder is weighed, dispersed, and added to the prepolymer. Thirdly, the Al2O3 powder and prepolymer are thoroughly stirred to ensure a uniform paste. Fourthly, the paste is transferred to a vacuum chamber for degassing, removing air bubbles to prevent attenuation of acoustic signals. Fifthly, slowly fill the vacuum-degassed paste into the metal shell, and cover it with a metal lid. The metal lid shall fit the shell well to reduce contact resistance and improve electromagnetic shielding capability. Lastly, cure at room temperature for 24 h. The encapsulated sensor, as shown in Fig. 3e, has a dimension of Φ 16 mm × H 5.5 mm and weighs approximately 3 g, which is smaller than conventional AE sensors, making it suitable for portable, implantable, and in-situ applications.

To verify the performance of the developed MEMS AE sensor under real working conditions, a PG fault diagnosis test rig was set up; the platform mainly comprises a speed control module, servo motor, planetary speed reducer, the developed sensors, and an oscilloscope (Fig. 4a). Figure 4b shows the overall composition of the test rig, with the planetary speed reducer comprising an 18-tooth sun gear, three 36-tooth PGs, a fixed carrier frame, and an outer 90-tooth ring gear (Fig. 4c). Its overall gear reduction ratio is 5:1. Figure 4d–h illustrate five common gear faults31,32: planetary gear tooth surface wear (PGTSW), planetary gear minor breakage (PGMB), planetary gear severe breakage (PGSB), planetary gear tooth root fracture (PGTRF), and sun gear breakage (SGB).

PG test rig setting. a Schematic. b Test rig. c Schematic diagram of PG. d Planetary gear tooth surface wear (PGTSW). e Planetary gear minor breakage (PGMB). f Planetary gear severe breakage (PGSB). g Planetary gear tooth root fracture (PGTRF). h Sun gear breakage (SGB)

During the experiment, the servo motor was connected to the sun gear as a power source, driving the sun gear to engage with the PG, thereby driving the outer ring gear to rotate. The operation of the servo motor was regulated by the control module. For each fault type, the servo motor ran at five different rotation speeds: 200, 400, 600, 800, and 1000 rpm. The entire experiment was conducted under zero-load conditions, with all test parameters remaining consistent except for the different fault gear. The acoustic signals received by the AE sensor were recorded by a digital oscilloscope (MSO44, Tektronix) at a sampling rate of 1 MHz. Six samples were collected at each speed for different fault types, and each sample was 10 s long.

The sensor was characterized using the pencil lead break (PLB) experiment33 and the comparative method. Pencils with a hardness of 2H, a lead diameter of 2.5 mm, and an exposed end length of 10 mm were selected for the test. During the test, the angle between the pencil and the rigid plane was maintained at 45° (Supplementary Fig. S2). The results of the PLB test are shown in Fig. 5a. The designed sensor can accurately capture the AE signals generated by the PLB. The output signal amplitude of the sensor is comparable to that of the commercial AE sensor PXR15 and exhibits a high signal-to-noise ratio. The spectrogram characteristics demonstrate that the PLB signal spans a frequency range from approximately 40 kHz to 300 kHz, confirming the capability of the MEMS AE sensor to capture broadband signals. Meanwhile, we further investigated the electromagnetic shielding performance of the aluminum shell through experimental verification. Supplementary Fig. S3a and c show that the testing platform employed piezoelectric ceramic transducers as the AE source to emit five 335 kHz/10 V acoustic pulses. The acquired data in Supplementary Fig. S3b and S3d demonstrate a remarkable 53.0 mV to 0.2 mV reduction in interference voltage induced by electromagnetic interference (EMI), corresponding to a 48.5 dB enhancement in shielding effectiveness (SE). This substantial suppression confirms that the aluminum-encased configuration effectively isolates EMI while maintaining signal integrity. Subsequently, the sensitivity of the sensor was tested on a standard testing platform. As shown in Fig. 5b, the sensor demonstrated a sensitivity more significant than 60 dB (ref. V/(m/s)) from 15 kHz to 620 kHz, with a peak sensitivity of 88.4 dB at 335 kHz. The performance of the developed MEMS AE sensor was compared with several commercial sensors, as shown in Table 1. The sensor has the advantages of wide bandwidth, high sensitivity, small size and light weight, which proves its competitive performance.

Performance of the MEMS AE sensor. a Comparison of PLB signals between commercial (Top) and manufactured MEMS AE sensor (Middle); spectrogram of PLB signals in MEMS sensor (Bottom). b Sensitivity characteristic of MEMS AE sensor

The acoustic signals generated by the speed reducer during operation are non-stationary and contain significant background noise. We employed wavelet denoising to reduce the interference from background noise and extract the characteristic signals related to hidden faults. Specifically, we used Beylkin as the wavelet basis and performed an 8-level wavelet decomposition on the measured acoustic signals. Threshold denoising was applied, with a particular focus on the wavelet sub-bands from the 4th to the 7th level, which contain the primary information. Finally, signal reconstruction was executed to obtain the denoised signals. Notably, due to the cyclical operation of the reducer, the gear operation process exhibits periodicity, which in turn causes AE signals generated by collisions or friction to display periodic characteristics. Figure 6a demonstrates that under varying fault conditions, the AE signals exhibit distinct variations in amplitude and temporal patterns, and the signal period demonstrates a strong temporal alignment with the gear engagement cycle, whereas the signal amplitude progressively increases in accordance with the severity of the fault, accompanied by increasingly complex waveform characteristics. These observations collectively characterize the AE signatures associated with gear fault mechanisms during operational loading. To further reveal the characteristics of the signal, we conducted a time-domain analysis on the acquired AE signals from the PG, which mainly includes variance, root mean square (RMS), entropy, and kurtosis, as shown in Fig. 6b. Here, the signals were peak-normalized to facilitate comparison. With the increasing severity of the fault, the variance, RMS, and entropy of the signal exhibit significant growth, indicating that signal changes caused by damage can be effectively captured in terms of statistical dispersion, energy intensity, and signal complexity. The comparatively lower PGTRF values, relative to those of SGB and PGSB, are attributed to the fact that the tooth root failure in the gear does not directly affect the AE signals generated by gear meshing during the gearbox’s operation. Kurtosis, which characterizes the statistical symmetry of a signal distribution, remains consistently positive across the measured gear signals. This indicates a positively skewed distribution relative to the normal distribution, suggesting the presence of heavy-tailed behavior. Such skewness is primarily caused by the impulsive components inherent to the gear operation, especially under fault conditions.

Fault diagnosis of PG. a Time-domain AE signals under different gear faults. b Time-domain analysis of AE signals under different gear faults. c PG fault diagnosis architecture based on ResNet-18. d t-SNE, and e confusion matrix of diagnostic results

To realize the fault diagnosis of PG, we used ResNet to classify and diagnose the PG fault data under various working conditions. The ResNet was first proposed by He et al. in 201534, by introducing residual blocks on the Convolutional Neural Networks (CNN), it solved the problem of gradient vanishing caused by increasing network depth and achieved a good trade-off between computational speed and classification accuracy35,36. Due to its ability to get subtle characteristic variations, ResNet shows superior performance in fault feature identification. The gear fault diagnosis algorithm based on the ResNet is outlined in Fig. 6b. Firstly, all the collected data were preprocessed, including filtering, data slicing, and continuous wavelet transform (CWT). The filtering process has been described earlier. The filtered data was evenly cut and divided according to the maximum rotation period of the speed reducer to ensure that each data segment contains at least one signal feature within a rotation period. Then, the one-dimensional time signals were converted into a 224 × 224 time-frequency diagram using CWT to generate a dataset for the fault diagnosis system, increasing the signals’ feature dimension while compressing the data volume. The preprocessed dataset was fed into ResNet for training, ultimately obtaining a PG fault diagnosis model. The network architecture used in this work is ResNet-18, which consists of 17 convolutional layers, 1 maximum pooling layer, 1 fully connected layer, and 1 SoftMax classifier. In order to achieve an accurate diagnosis of different gear faults, we adjusted the network parameters, as shown in Supplementary Table S2. A total of 2700 time-frequency spectrograms were used in this work, with 1890 images (70%) allocated for training and 810 images (30%) used for validating.

To demonstrate the clustering effect of the model, the t-stochastic neighbor embedding (t-SNE) algorithm was used to compress the high-dimensional features extracted by ResNet-18 into a two-dimensional plane for visualization. As shown in Fig. 6c, the diagnosis results of the gears show that different fault types can be effectively clustered together, indicating that the network can well learn the decision boundary on the training dataset. The confusion matrix in Fig. 6d shows the ability of the used model to distinguish different faults. The adopted ResNet-18 model has an accuracy rate of over 96% for classifying different types of faults, and the overall classification accuracy can reach 98.63%, indicating that the fault diagnosis model has excellent performance and can realize the classification and identification of different PG faults.

This paper presents a tiny-sized, cost-effective MEMS AE sensor with high performance. The full process from design to fabrication, encapsulation, and performance characterization of the ScAlN-based PMUT chip is systematically described. By integrating the chip with a signal conditioning circuit and encapsulating it within an epoxy/Al2O3 composite inside a metal shell, acoustic impedance matching and electromagnetic shielding were achieved. This approach not only ensures long-term operational reliability but also enables high-sensitivity detection of high-frequency AE signals across a wide bandwidth. The encapsulated MEMS AE sensor features a compact dimension of Φ 16 mm × H 5.5 mm and a mass of approximately 3 g, making it suitable for portable, implantable, and in-situ applications. It exhibits a peak sensitivity of 88.4 dB at 335 kHz, with a frequency response range from 15 kHz to 620 kHz and maintaining sensitivity levels above 60 dB throughout. This design effectively addresses the critical limitations of conventional AE sensors, including compact size optimization and lightweight mass reduction.

Using the developed sensor, we established a PG fault diagnosis test rig to verify its capability to detect subtle fault signals. Under various fault conditions, the captured AE signals exhibited distinct differences, with signal amplitude significantly increasing as the fault severity progressed. Subsequently, a ResNet-18 model was employed to classify different types of PG faults. The model achieved an accuracy exceeding 96% for individual fault types and an overall fault classification accuracy of 98.63%, demonstrating its excellent performance and capability to distinguish between diverse PG failure modes. In general, this MEMS AE sensor can provide a promising approach for in-situ fault monitoring of highly integrated and miniaturized industrial equipment, and injects new impetus into the developing industrial automation technology.

Raja Santhi, A. & Muthuswamy, P. Industry 5.0 or industry 4.0 S? Introduction to industry 4.0 and a peek into the prospective industry 5.0 technologies. Int. J. Interact. Des. Manuf. 17, 947–979, https://doi.org/10.1007/s12008-023-01217-8 (2023).

Article Google Scholar

Jaw, L. C. Recent advancements in aircraft engine health management (EHM) technologies and recommendations for the next step. Turbo Expo: Power Land Sea, Air 46997, 683–695 (2005).

Google Scholar

Gawde, S. et al. Multi-fault diagnosis of industrial rotating machines using data-driven approach: A review of two decades of research. Eng. Appl. Artif. Intell. 123, 106139, https://doi.org/10.1016/j.engappai.2023.106139 (2023).

Article Google Scholar

Wu, P. et al. A highly sensitive triboelectric quasi-zero stiffness vibration sensor with ultrawide frequency response. Adv. Sci. 10, 2301199, https://doi.org/10.1002/advs.202301199 (2023).

Article Google Scholar

Pang, Y., He, T., Liu, S., Zhu, X. & Lee, C. Triboelectric nanogenerator-enabled digital twins in civil engineering infrastructure 4.0: a comprehensive review. Adv. Sci. 11, 2306574 (2024).

Article Google Scholar

Hamed, Y., O’Donnell, G., Lishchenko, N. & Munina, I. Strain sensing technology to enable next-generation industry and smart machines for the factories of the future: A review. IEEE Sens. J. https://doi.org/10.1109/JSEN.2023.3313013 (2023).

Yoon, J., He, D. & Van Hecke, B. On the use of a single piezoelectric strain sensor for wind turbine planetary gearbox fault diagnosis. IEEE Trans. Ind. Electron. 62, 6585–6593, https://doi.org/10.1109/TIE.2015.2442216 (2015).

Article Google Scholar

Gao, W. et al. A high-resolution MEMS capacitive force sensor with bionic swallow comb arrays for ultralow multiphysics measurement. IEEE Trans. Ind. Electron. 70, 7467–7477, https://doi.org/10.1109/TIE.2022.3203756 (2022).

Article Google Scholar

Hase, A., Mishina, H. & Wada, M. Correlation between features of acoustic emission signals and mechanical wear mechanisms. Wear 292, 144–150, https://doi.org/10.1016/j.wear.2012.05.019 (2012).

Article Google Scholar

He, K. & Li, X. Time–frequency feature extraction of acoustic emission signals in aluminum alloy MIG welding process based on SST and PCA. IEEE Access 7, 113988–113998, https://doi.org/10.1109/ACCESS.2019.2935117 (2019).

Article Google Scholar

Kundu, P. Review of rotating machinery elements condition monitoring using acoustic emission signal. Expert Syst. Appl. 252, 124169, https://doi.org/10.1016/j.eswa.2024.124169 (2024).

Article Google Scholar

Liu, L., Zhi, Z., Yang, Y., Shirmohammadi, S. & Liu, D. Harmonic reducer fault detection with acoustic emission. IEEE Trans. Instrum. Meas. https://doi.org/10.1109/TIM.2023.3291747 (2023).

Feng, P. et al. Monitoring gear surface degradation using cyclostationarity of acoustic emission. Mech. Syst. Signal Process. 131, 199–221, https://doi.org/10.1016/j.ymssp.2019.05.055 (2019).

Article Google Scholar

Liu, T. et al. Airborne rangefinding with pMUTs array using differential structure. IEEE Sens. J. https://doi.org/10.1109/JSEN.2023.3298671 (2023).

Gao, Y. et al. A miniaturized transit-time ultrasonic flowmeter based on ScAlN piezoelectric micromachined ultrasonic transducers for small-diameter applications. Microsyst. Nanoeng. 9, 49, https://doi.org/10.1038/s41378-023-00518-y (2023).

Article Google Scholar

Liu, T. et al. A wearable acoustic sensor for identification in harsh noisy environments. 2024 IEEE 19th International Conference on Nano/Micro Engineered and Molecular Systems (NEMS), 1–4 (2024).

Saboonchi, H., Ozevin, D. & Kabir, M. MEMS sensor fusion: Acoustic emission and strain. Sens. Actuator A Phys. 247, 566–578, https://doi.org/10.1016/j.sna.2016.05.014 (2016).

Article Google Scholar

Ozevin, D. MEMS acoustic emission sensors. Appl. Sci. 10, 8966, https://doi.org/10.3390/app10248966 (2020).

Article Google Scholar

Kabir, M., Kazari, H. & Ozevin, D. Piezoelectric MEMS acoustic emission sensors. Sens. Actuator A Phys. 279, 53–64, https://doi.org/10.1016/j.sna.2018.05.044 (2018).

Article Google Scholar

Khan, T. M., Taha, R., Zhang, T. & Ozevin, D. Multi-frequency MEMS acoustic emission sensor. Sens. Actuator A Phys. 362, 114648, https://doi.org/10.1016/j.sna.2023.114648 (2023).

Article Google Scholar

Butaud, P. et al. Towards a better understanding of the CMUTs potential for SHM applications. Sens. Actuator A Phys. 313, 112212, https://doi.org/10.1016/j.sna.2020.112212 (2020).

Article Google Scholar

Liu, T. et al. Emerging Wearable Acoustic Sensing Technologies. Adv. Sci. 2408653. https://doi.org/10.1002/advs.202408653 (2025).

Liu, T. et al. A gas flow measurement system based on lead zirconate titanate piezoelectric micromachined ultrasonic transducer. Micromachines 15, 45, https://doi.org/10.3390/mi15010045 (2023).

Article Google Scholar

Yao, Y. et al. A transceiver integrated piezoelectric micromachined ultrasound transducer array for underwater imaging. Sens. Actuator A Phys. 359, 114476, https://doi.org/10.1016/j.sna.2023.114476 (2023).

Article Google Scholar

Liu, T. et al. Machine learning-assisted wearable sensing systems for speech recognition and interaction. Nat. Commun. 16, 2363 (2025).

Article Google Scholar

Blevins, R. D. & Plunkett, R. Formulas for natural frequency and mode shape. J. Appl. Mech. 47, 461 (1980).

Article Google Scholar

Zhou, J. et al. Continuous and non-invasive monitoring of blood pressure based on wearable piezoelectric micromachined ultrasonic transducers array. J. Microelectromech. Syst. 32, 437–444 (2023).

Article Google Scholar

Rathod, V. T. A review of acoustic impedance matching techniques for piezoelectric sensors and transducers. Sensors 20, 4051, https://doi.org/10.3390/s20144051 (2020).

Article Google Scholar

Lohfink, A. & Eccardt, P.-C. Linear and nonlinear equivalent circuit modeling of CMUTs. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 52, 2163–2172, https://doi.org/10.1109/TUFFC.2005.1563260 (2005).

Article Google Scholar

Devaney, A. & Levine, H. Effective elastic parameters of random composites. Appl. Phys. Lett. 37, 377–379, https://doi.org/10.1063/1.91949 (1980).

Article Google Scholar

Yang, Y., Hu, N., Li, Y., Cheng, Z. & Shen, G. Dynamic modeling and analysis of planetary gear system for tooth fault diagnosis. Mech. Syst. Signal Process. 207, 110946, https://doi.org/10.1016/j.ymssp.2023.110946 (2024).

Article Google Scholar

Zeng, Q., Feng, G., Shao, Y., Gu, F. & Ball, A. D. Planetary gear fault diagnosis based on an instantaneous angular speed measurement system with a dual detector setup. IEEE Access 8, 66228–66242, https://doi.org/10.1109/ACCESS.2020.2985170 (2020).

Article Google Scholar

FR, B. Characterization and calibration of acoustic emission sensors. Mater. Eval. 39, 60–68 (1981).

He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 770–778 (2016).

Wei, D., Liu, K., Wang, J., Zhou, S. & Li, K. ResNet-18-based interturn short-circuit fault diagnosis of PMSMs with consideration of speed and current loop bandwidths. IEEE Trans. Transp. Electrif. 10, 5805–5818, https://doi.org/10.1109/TTE.2023.3319157 (2024).

Article Google Scholar

Gao, M. et al. AI-enabled metal-polymer plain bearing based on the triboelectric principle. Adv. Funct. Mater. 33, 2304070 (2023).

Article Google Scholar

Download references

This work was supported in part by the National Key Research and Development Program of China (Grant No. 2022YFB3205400), in part by the Fundamental Research Funds for the Central Universities (Grant No. 2024CDJGF-005), and in part by the Science Fund for Distinguished Young Scholars of Chongqing (Grant No. CSTB2022 NSCQJQX0006).

Key Laboratory of Optoelectronic Technology and Systems, Ministry of Education and International Research and Development Center of Micro-Nano Systems and New Materials Technology, Chongqing University, Chongqing, China

Hanjie Dou, Tao Liu, Zhihao Li, Jixuan Zhang, Jiaqian Yang, Yuchen Mao, Wanyu Xu & Xiaojing Mu

Search author on:PubMed Google Scholar

The authors of this work made the following contributions: H.D. contributed to conceptualization, writing (original draft), validation, algorithm, experiment, visualization, and graphics designing; T.L. provided methodology support, chip fabrication, and manuscript modification; Z.L. provided original algorithm support; J.Z. and W.X. contributed to the design of the signal conditioning circuit; J.Y. and Y.M. contributed to the encapsulation and characterization of the sensors; X.M. conducted formal analysis, editing, investigation, review, supervision, project administration, provided resources, and funding acquisition.

Correspondence to Xiaojing Mu.

The authors declare no competing interests.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

Dou, H., Liu, T., Li, Z. et al. Intelligent planetary gear fault diagnosis system based on MEMS acoustic emission sensor. Microsyst Nanoeng 11, 126 (2025). https://doi.org/10.1038/s41378-025-00961-z

Download citation

Received: 26 February 2025

Revised: 14 April 2025

Accepted: 22 April 2025

Published: 18 June 2025

DOI: https://doi.org/10.1038/s41378-025-00961-z

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative