Industrial fans play a critical role in manufacturing facilities, and a sudden shutdown of critical fans can cause significant disruptions. Ensuring early, effective, and accurate detection of fan malfunctions first requires confirming the characteristics of anomalies resulting from initial damage to rotating machinery. In addition, sensing and detection must rely on the use of sensors and sensing characteristics appropriate to various operational abnormalities. This research proposes an online industrial fan monitoring and fault detection technique based on acoustic signals as a physical sensing index. The proposed system detects and assesses anomalies resulting from preliminary damage to rotating machinery, along with improved sensing resolution bandwidth features for microphone sensors as compared to accelerometer sensors. The resulting Intelligent Prediction Integration System with Internet (IPII) is built to analyze rotation performance and predict malfunctions in industrial fans. The system uses an NI cRIO-9065 embedded controller and a real-time signal sensing module. The kernel algorithm is based on an acoustic signal enhancement filter (ASEF) as well as an adaptive Kalman filter (AKF). The proposed scheme uses acoustic signals with adaptive order-tracking technology to perform algorithm analysis and anomaly detection. Experimental results showed that the acoustic signal and adaptive order analysis method could effectively perform real-time early fault detection and prediction in industrial fans.

1. Introduction

Industrial fans are critical components in industrial facilities and are used to remove exhaust emissions, ventilate, compress air, and drive air-conditioning systems. Such fans typically operate continuously for long durations, and improper assembly or maintenance can result in malfunctions, including vibrations and audible noise. Common mechanical faults are mostly caused by bearing failures, looseness, shaft cracks, poor balance, misalignment, and resonance. Figure 1 shows general maintenance procedures for rotating machinery. Predictive maintenance of such machinery minimizes unexpected failure and potential downtime but relies on accurate performance and measurement techniques through the integration of various sensors to provide real-time data on equipment conditions.

Problems in rotating machinery are typically indicated by the presence of abnormal vibration, noise, or temperature. Accelerometers are typically used to perform vibration signal analysis [1], with portable and online real-time monitoring methods based on the ISO 13373 standards [2].

The traditional signal processing involves transforming the measured vibration time-domain signal by fast Fourier transform (FFT) into frequency domain feature values, a process also known as Frequency Spectrum Analysis [3]. However, the shaft rotational speed of industrial fans changes with factory production conditions, loading demands, and many external environmental factors. Therefore, the measured vibration signals of bearings produce a nonlinear, nonstationary time-domain signal, while the traditional Fourier transform is only appropriate for linear and stationary time-domain signals and cannot accurately reflect time-variant factors to respond to the nonstationary vibration behavior and complex vibration signals of industrial fans. In response to vibration signals indicating early malfunctions, the common spectrum merely presents the concentrated energy at a certain frequency peak value. Spectrum analyses do not show or resolve the vibration energy distributions of the harmonics when other harmonics change with the rotating speed.

Order-tracking analysis is an effective method for the signal analysis of nonstationary vibration signals from rotating machinery. However, the use of order-tracking methods alone while performing dynamic signal analysis results in inadequate signal resolution following Fourier transform spectrum analysis. Due to equipment complexity and large speed variation in rotating machinery, component interrelations raise a nonstationary cross-correlation function problem. Thus, the traditional Fourier transform frequency domain signal analysis can easily cause undesirable smearing while traditional order analysis cannot improve the resolution of each harmonic order. Hence, order tracking integrates the recursive least-square method and the recursive Kalman filter [4, 5] to provide adaptive order tracking. Using the adaptive algorithm overcomes the frequency smearing effect in adjacent and cross orders of traditional order analysis. The adaptive algorithm’s high-resolution signal and order curve change with rotation speed characteristics make it very suitable for signal analysis in conditions of high rotating speed or variable speeds. Adaptive order-tracking analysis provides good precision for the real-time vibration signal fault diagnosis of rotating machinery.

Wavelet transform (WT) is another method of analyzing nonstationary signals from machinery rotating at variable speeds. WT provides good time and frequency resolution, overcoming the limitations of time-domain and frequency-domain analysis in Fourier transform. The WT-tracking algorithm also provides effective vibration signal tracking in conditions of variable speed. Therefore, WT is often used to diagnose faults of rotating machinery based on vibration signal analysis [69]. It can also be used for spectrum analysis of acoustic signals and to identify signal signatures as a means of diagnosing faults from acoustic signals [10]. However, the major disadvantage of WT is that it applies a fixed wavelet base to decompose signals. Thus, an inappropriate selection of the basis function and a failure due to mismatched signal characteristics can significantly reduce signal analysis performance [11, 12]. In other words, WT is not generally adaptive, and the transform efficiency is highly dependent on the basis function selection.

We have surveyed similar research regarding fault diagnostic methods for various types of rotating equipment, each with their own theoretical foundations, advantages, and solutions. These methods and theories are largely similar and are based largely on vibration sensors [13, 14]. Each approach selects diagnostic subjects and algorithms based on specific research purposes and considerations. This paper starts over from scratch with a reconsideration of sensor type, algorithm simulation, and continuous experimental adjustment in an attempt to improve early anomaly detection in factory-based industrial fans. Our review of the previous three years’ studies in related fields found no mention of the use of acoustic sensors and similar research methods (i.e., algorithms) for use in industrial fan fault diagnosis. Using vibration signals to perform fault prediction in rotating machinery can result in weak extraction of early fault features. The weak abnormality features of early faults are generally masked by background noise or other interference, making them difficult to detect. However, early detection is critical to prevent continuous deterioration or sudden failure. In [15, 16], the authors focused on the early identification of fault signals. Industrial fans most commonly suffer from bearing failures which typically proceed through four failure stages [1719]. In the first stage, slight defects and cracks appear with abnormalities manifesting at ultrasound frequencies (20~60 kHz). The faults will gradually deteriorate until the device fails. However, vibration sensors typically are more sensitive to low frequencies (below 2 kHz roughly) [20], limiting their utility in early fault detection for rotating machinery.

In addition to vibration signals, rotating machinery produces acoustic signals. Certain equipment faults produce specific vibration signals along with corresponding acoustic signals. The acoustic signal from rotating machinery includes considerable meaningful information. Unlike vibration signals, acoustic signals are usually measured by easily applied noncontact methods. Furthermore, acoustic signals have better frequency resolution at medium and high frequencies [20], and acoustic sensors provide advantages for preliminary fault detection in rotating machinery [18].

An industrial fan involves complex acoustic signals in operation. An acoustic sensor possesses superior high-frequency range sensing characteristics to match the signal characteristics of early failure in industrial fans and process fault detection through the order-tracking analysis and adaptive Kalman filter algorithms with good detection results. For the nonstationary signal of the rotation apparatus, the commonly used FFT spectrum analysis cannot be used effectively to distinguish the frequency composition. With regard to the wavelet transformation, its applications are limited to its nonadaptivity. The Kalman filter has been widely applied to various technical fields, featuring its fast signal processing and high performance for big data. The studies in [2123] are good examples with it, where the Kalman filter has been used to deal with equipment failures. Despite its fast convergence speed, the Kalman filter suffers a little bit from poor stability and a higher filtering threshold in the target. As a result, we have added the tachometer signal of the industrial fan into the proposed system designed to be used in conjunction with the order-tracking technique, where the model-based adaptive Kalman filter inspired from [2426] has also been utilized to estimate the best status of the dynamic system.

We use high-resolution-order-analysis technology to analyze the characteristics of acoustic and vibration signals in terms of rotation speed to capture information related to the operation of rotating equipment. This information is then combined with the adaptive Kalman filter algorithm to produce an early fault detection method for rotating equipment. The proposed method has particular advantages for fault detection in variable speed conditions. The key is the integration of the industrial fan’s tachometer signal and to use the motor shaft angle for resampling of nonstationary acoustic signals. We then use the adaptive Kalman filter in a recursive manner to eliminate background noise and various interference signals to estimate the dynamic system’s optimal state. This structure is based on variable speeds, and the proposed algorithm provides advantages for noise elimination.

The ASEF and AKF are both excellent kernel algorithms for rotating machinery, and order-tracking techniques can be applied to acoustic signals to differentiate various fault conditions. The most significant thing of this study is to propose an idea using an acoustic sensor to carry out signal detection. By taking the advantage of the superior high-frequency response and combining it with the ASEF and AKF high-resolution algorithm, the built IPII is capable of predicting early failure of the industrial fan. The user can therefore be warned early to use contingency plans. For the periodic dynamic signal of the fan, the IPII uses acoustic signal and fiber-optical signal to obtain the phase. By combining both the order-tracking and AKF algorithms, the IPII can recognize the features resulting in system failure early. For the nonperiodic dynamic signal, it relies mainly on the ASEF algorithm along with the acoustic signal to perform abnormal feature analysis for early diagnosis of the failure. It is an innovative method for predictive maintenance in factories. This study can contribute significantly not only to industrial fan systems but also to any other periodic and nonperiodic rotation systems. The IPII is capable of carrying out real-time analysis and prediction of rotating machinery conditions before failure occurs. Unexpected failure of crucial rotating machinery can be prevented through the development of a fault prediction system better able to identify preliminary damage. Adjusting or modifying the parameters during production processes can prevent production shutdowns and reduce maintenance costs.

2. Signal Feature Extraction Theory

2.1. ASEF
2.1.1. Acoustic Signal Enhancement

Nonstationary signals such as online acoustics and vibrations generated by industrial fans are difficult to calculate, and signal enhancement is a crucial preprocessing step to eliminate noise interference. The subspace signal enhancement method removes noise from the extracted signals, thus improving SNR quality.

A Karhunen-Loeve Transform (KLT) is performed on the noisy zero-mean normalized vector . The vector has a symmetric nonnegative autocorrelation matrix . Since is an matrix, it determines a complete set of orthogonal eigenvectors associated with the real, nonnegative eigenvectors. These eigenvalues can be ordered as , and the statistical variance of the data set in the direction of the eigenvector corresponds to the eigenvalue . can be expressed using the following orthogonal transformation [27]: and coefficients, which are called the principal components, can be shown by projecting the data vectors onto each eigenvector as follows:

Noise reduction can be achieved by reconstructing the initial data using only the -weighted eigenvectors of the signal-plus-noise subspace, and the linear estimation of the clean vector is used as an estimation criteria as follows [28]: where is a weighting function. An optimal parameter choice for which results in more aggressive noise suppression is given by where is the noise variance and the value of the parameter is fixed experimentally.

Signal enhancement is achieved by removing the noise subspace and estimating the clean signal from the remaining signal space. Our ASEF features acoustic-sensor-based nonperiodic acoustic signal processing for an industrial fan. It begins with the KLT filtering, followed by the MFCC filtering. The results of the MFCC filtering underwent abnormal feature analysis.

2.1.2. Analysis of Cepstral Acoustic Features

Cepstrum analysis is widely used in automatic speech recognition, acoustic echo cancellation, and acoustic filtering. The cepstrum analysis technique is defined as the inverse Fourier transform of the logarithm of the short-term power spectrum of measured data in the time domain. Cepstrum separates and enhances the periodic signal so the significant correlation function can be identified.

When processing signals [27, 29], we first extracted the sensed signals and then proceeded to signal preprocessing. We then turned the signal into a series of spectrum sequences by short-time FFT. The magnitude spectrum of the signal was then passed through a set of 20 triangular bandpass filters to retrieve the output log-energy of each filter. These triangular filters are spaced on the mel-frequency scale. The relation of mel-frequency and general frequency is as follows: where corresponds to the linear frequency scale.

To transform the logarithm energy value of each filter to discrete cosine transform (DCT), retrieve the mel-scale frequency cepstral coefficient (MFCC) and convert the frequency domain signal to the time domain. The DCT equation is as follows: where is the number of the cepstral coefficients, is the analysis order, and represents the log-energy output of the filter.

2.2. AKF
2.2.1. Resampling Theory

This research applies the order analysis of the resampling theorem to analyze and diagnose the acoustic and vibration signal features generated by industrial fans. Particularly in the case of varying rotating shaft speeds, using this theory to extract the acoustic and vibration signals achieves a high resolution from the sampling intervals. A Fourier transform can be used to calculate the order spectrum generated by the periodic vibration/acoustic time sequences from the rotating information. Thus, the acoustic and vibration signal energy corresponds to each phase angle of the rotating information from the fiber-optical switch. Discrete-time signals are resampled according to equal rotating shaft angle intervals. Converting the signals from time domain to the angle domain is a function of resampling theory [30].

The proposed experiment collects synchronous acoustic, vibration, and optical signals. We apply interpolation and curve-fitting to the impulse signal of the fiber-optical switch to retrieve the relative angular displacement at each time point. We then resample the acoustic and vibration signals according to equal angular displacement intervals. The resampled signals are called order signals, and the spectrogram after FFT or STFT of the order signals is called the order spectrogram, which can be used to provide a precise analysis of the fundamental frequency and its relative harmonics, thus facilitating analysis of the online acoustic and vibration signal features of rotating machinery.

2.2.2. Adaptive Kalman Filter

An industrial fan’s acoustic signal is essentially a frequency-modulated signal that can be expressed as the superposition of sinusoidal signals. The acoustic signal containing orders generated by the rotating shaft can be written as where and respectively express the amplitude and phase of the order. The variable represents the angular displacement of the rotating spindle that can be computed as

A proposed adaptive order-tracking analysis based on Kalman filtering is presented. The Kalman-filtering problem entails two equations [4, 5, 2426]. The first equation is called the process equation and is expressed as where is a known state transition matrix related to the state of the system at times and . The vector represents process noise. The vector is modeled as a zero-mean white-noise whose correlation matrix is defined as

The second equation is called the measurement equation and is expressed as where is a known measurement matrix. The vector is the measurement noise modeled as a zero-mean white-noise process whose correlation matrix is defined as. where the noise vectors and are statistically independent so that we have for all and .

We define the Kalman gain expressed as where the is called the correlation matrix. The matrix is called the predicted state-error correlation matrix.

A filtered estimation error vector is shown in where is called the innovation vector into the filtered estimation error vector .

The aforementioned parameter identification problem can be recursively solved using the Kalman-filtering algorithm. The procedure of the recursive Kalman filtering is summarized as follows [23]:

For the above procedure, all the variables are summarized in Table 1. To initialize the adaptive Kalman-filtering process, the initial conditions are generally ( is the number of parameters ) and ( is a identity matrix). Figure 2 shows the block diagram of the recursive Kalman filter.

Although the adaptive Kalman-filtering algorithm has greater computation and time requirements because of its recursive calculation structure, we chose it for its advantages of high estimation precision, fast convergence, and robustness. All of these features are crucial to the equipment fault prediction system which requires immediate response to effectively and accurately detect abnormalities.

3. Experimental Structure

3.1. Sensor Specifications and Installation

Common fault signs of industrial fans consist of abnormal vibrations, audible noise, and temperature. To build a complete and effective monitoring system, this research proposes an Intelligent Prediction Integration System with Internet (IPII) to perform online monitoring and fault prediction of industrial fans [31]. The system uses an NI cRIO-9065 embedded controller and a real-time synchronizing module. The specifications of the microphone and accelerometer sensors used are shown in Tables 2 and 3. The sensor installation structure is shown in Figure 3.

3.2. Online Monitoring and Fault Prediction of Industrial Fans with IPII

In addition to the main acoustic signal, the IPII monitoring system also uses vibration and temperature signals to assist diagnosis. The system structure includes three layers based on different functions: (1)Sensing layer: microphone, accelerometer, thermocouple, and fiber-optical sensors are used to extract various operating signals from industrial fans. This system also uses a wireless sensor network (WSN) to perform wireless sensing transmission.(2)Smart control layer: it performs computation, analysis, and diagnosis from the extracted data and implements smart controls according to practical requirements, such as emergency measures in response to output alarms or automatically switching to spare machinery.(3)Monitoring layer: monitors system status through a graphical user interface (GUI) which can be operated at a central control room or on mobile devices via wireless networks for remote monitoring. Smart devices can also be used to remotely shut down potentially abnormal machinery. The system structure is shown in Figure 4.

Signal detection relies mainly on the microphone sensor, assisted by the accelerometer with an adaptive order-tracking algorithm to more precisely distinguish certain featured signals. For example, the acoustic signal feature of the blade pass frequency can be analyzed by signal order analysis to obtain a more precise identification. Meanwhile, it can also obtain abnormal feature order energy signals of bearings, blades, or motors according to accelerometer characteristics. The system includes a thermocouple to detect bearing temperature, along with acoustic and vibration signals to serve as the basis for executing smart controls in response to fault signs.

System hardware consists of an embedded NI RIO-9065 dual-core controller, FPGA, real-time module, signal synchronization A/D module, and WSN module. The kernel algorithm is an adaptive Kalman filter. Acoustic and vibration signals were individually combined with tachometer signals for simultaneous algorithm analysis and abnormal feature extraction. Big data techniques are used to process multiple sensed signals from collection, storage, extraction, and analysis to decision making as a looping execution. The monitoring and control flow chart is shown in Figure 5.

Our ASEF features acoustic-sensor-based nonperiodic acoustic signal processing for an industrial fan. It begins with the KLT filtering, followed by the MFCC filtering. The results of the MFCC filtering underwent abnormal feature analysis. Anomaly detection is just one process of the system, and the goal of anomaly detection is to ensure smooth production line operation. Thus, we combine diagnostics with controls to determine an optimal operation strategy based on current equipment status, while using fault diagnostics to prevent future equipment failure. In response to anomalies, the system uses an integrated smart control to maintain normal production and safety through implementing corresponding contingency measures such as reducing operation speed, activating spare machinery, or adjusting production parameters. Effective early fault detection requires preprocessing of the various signals to extract important features. Signal preprocessing seeks to separate the signal of interest from noise to ensure the effective detection of abnormal features.

The first step is to enhance the sensed signal. Karhunen-Loeve transform (KLT) high-pass filtering eliminates the remaining noise. The Hamming window is then applied with a certain bandwidth. Next, following inverse transform, the signal is subjected to mel-frequency cepstral coefficient (MFCC) filtering to process the signal features. When the signal preprocessing procedures are completed, the signal is then subjected to Kalman filtering for online real-time fault prediction of industrial fans.

This process can also be used for spectrum feature filtering to confirm the sensing characteristics of different sensors. Spectrum feature filtering analyzes the sensing resolution frequency bandwidth feature in a free field, using consistent ambient noise levels, along with environmental and measured conditions for different sensors to assess differences in the practical measured sensing resolution bandwidth features for two or more sensors and can be used as the selected reference for this research. The analysis flow chart is shown in Figure 6.

The IPII incorporates tachometer signal along with the order-tracking technique in it. It is also with the model-based adaptive Kalman filter [2426]. The filter possesses a recursive least-square algorithm and recursive Kalman filtering in the adaptive signal processing to seek an optimal solution for the problems, where the background noise and any other interference can be effectively filtered out so that we can come out with the best status of the dynamic system. The setup of the IPII is dedicated to signal processing of the collected data from the acoustic and tachometer sensor. The data first underwent short-time Fourier transform (STFT). The results of STFT are resampled and combined with order tracking to perform recursive Kalman filtering, followed by the order feature analysis to obtain the output (Figure 7).

4. Result and Verification

4.1. Setup

The monitoring system uses a high-resolution order-tracking analysis, combining the big data collected from a microphone, accelerometer, and tachometer with the bearing temperature signals to execute the algorithm to implement fault prediction for industrial fans based on acoustic signals. Figure 8 shows the experimental framework based on an industrial fan with an induction motor. The GUI in Figure 9 can be operated in the central control room or on smart networked devices for remote monitoring.

4.2. Acoustic Signal Preprocessing
4.2.1. Spectrum Feature Filtering of Acoustic Signals

We realize the proposed signal preprocessing and spectrum feature filtering of acoustic signal by means of the filtering algorithm and the experiment structure presented in Section 4.1 on the basis of Section 3. The signal preprocessing procedures and their individual signal waveform variations are described as follows. Figure 10(a) shows the original acoustic time-domain signal from an industrial fan. Figure 10(b) shows the signal following noise reduction using KLT high-pass filtering. We applied the Hamming window with a certain bandwidth and obtained the waveform shown in Figure 10(c). Lastly, the waveform following inverse transform and MFCC signal feature processing is shown in Figure 10(d). The sampling frequency, extraction time, FFT record length, and window of the IPII system were respectively set as 5 kHz, 12 seconds, 2048 points, and the Hamming window.

4.2.2. A Comparison of Acoustic and Vibration Signals in Spectrum Feature Filtering

To choose an appropriate sensor for early fault detection and to confirm the differences in sensing characteristics between acoustic and vibration signals, we compare the different signal preprocessing methods.

The practical resolution bandwidth features of acoustic and vibration sensors after MFCC filtering are shown in Figure 11. The differences in resolution bandwidth features clearly show that the vibration sensor has an excellent resolution within low-frequency fields. In contrast, the acoustic sensor has a greater sensing resolution bandwidth, giving it a higher resolution advantage, especially in medium- and high-frequency fields. That is, the acoustic sensor is better able to distinguish abnormal features for entire frequency fields. Moreover, the typical mechanical properties of rotating machinery suggest that initial component damage would be indicated by abnormalities within medium- and high-frequency fields. The sensing resolution bandwidth feature of the acoustic sensor is matched with the abnormality generated by the initial damage. Therefore, the acoustic sensor provides better early fault detection. The results also confirm the superior effectiveness of acoustic signals in predicting faults online.

4.3. Practical Digital Filter Implementation

Using the adaptive Kalman filter for analysis requires some parameter settings. and in (16) and (20) are the most critical parameters, in that they reflect actual conditions and thus have a direct impact on algorithm performance. is the main correlation convergence factor for the process equation, while is the correlation convergence factor for the measurement equation. The initial parameter settings are derived from experience and adjusted to actual test conditions. This paper selects different convergence parameters for and . The filter eliminates noise, and the system achieves a high convergence rate; thus, the order energy characteristics of the acoustic signal feature can achieve high-resolution performance. is 10−9 and 10−8, and the four parameters in (104, 106, 108, and 1012) are used in the actual test, with Figure 12 showing filter convergence performance. Due to the use of different parameters, the test results for the four parameters show that the two parameter sets can be used. However, when is set at 10−9, it produces better convergence than at 10−8. Thus, 10−9 is selected as the main design parameter for filter implementation. Based on industrial fan rotational characteristics and anomaly detection results, 108 and 1012 are selected as the main filter settings for , where 108 is used for low fan rotation speeds and 1012 is used for high rotation speeds. That is, our example system implementation follows the actual convergence conditions using two parameter sets, and , to implement the digital filter.

4.4. Malfunction Detection

The proposed system focuses on early fault detection. Using an adaptive order-tracking algorithm, the system obtains the order energy at the time of rotation which can be used to clearly identify equipment abnormalities, making it very suitable for early fault detection.

This paper identifies faults based on the order energy distribution for abnormal states and normal operation. That is, the degree of difference in the order energy ranking is used to determine whether abnormalities exist or not. The threshold values used in the experiments take the normal (mean) value plus 30% as the order energy standard. In practical applications, the sensitivity of this threshold value can be adjusted according to the importance of the particular piece of equipment. In practice, a threshold value can be derived from new machinery after running for 48 hours, and if a fan requires replacement or repair, the system will automatically self-tune the key indicators for the replacement parts.

4.5. Industrial Fan Verification
4.5.1. Under Fixed Rotation Speed Conditions

Accumulated maintenance experience suggests that the motor’s inner race is the component most likely to fail first under normal wear, followed by the outer race, ball spin, and fundamental train. Therefore, the experiment simulates an inner race failure on the bearing in an industrial fan after the IPII algorithm computes and observes the signal changes reflecting initial abnormalities. Figure 13 shows the order energy of the acoustic signal. The bold line represents the energy distribution under the broken bearing condition while the thin line represents the energy distribution under normal fan operating conditions. In the 1st- to 4th-order energy charts, the microphone sensor can effectively provide an early diagnosis of the abnormal signal. The system uses synchronizing signal sampling processing; thus, the order energy changes of the vibration signal in Figure 14 show that the accelerometer sensor detected a slight fault sign when the microphone sensor detects an abnormality. However, there was no significant difference between the fault signal and normal signals. As a result, the microphone outperforms the accelerometer sensor in providing early fault diagnosis of industrial fans.

Different malfunction conditions result in different order energy configurations. Figure 13 shows the 1st- to 4th-order energies for the inner race failure in an industrial fan. In the event of an outer race malfunction, the distribution for each order energy shows different trends, such as loose or unbalanced blades. As shown in the order energy diagrams of Figures 13 and 14, large diagnostic samples are stored in the feature data bank (FDB) for application to actual fault analysis, allowing the system to diagnose various abnormalities. Through this type of condition monitoring, and the subsequent collection and analysis of big data, we can create a reasonable component maintenance and replacement schedule and adjust production parameters to ensure equipment reliability, thus improving product quality and reducing maintenance costs.

4.5.2. Under Transient Rotation Condition

The user interface design accounts for various fan types and installation environments, allowing for parameter adjustments for monitoring of a wide variety of industrial fans, detection types, and sensors. The parameter settings are presented in Table 4. The proposed algorithm can be adjusted to account and compensate for these hardware architecture, sensor, and fault signal characteristics to provide for optimal online test performance. The user interface also provides real-time anomaly trend analysis and monitoring, allowing for real-time anomaly diagnosis.

At low rotation speeds, fan anomaly diagnosis is mainly derived from acoustic order energy. In addition to acoustic order energy, under transient or variable speed conditions, energy spectrum trends produced by MFCC filtering can be used to assist in anomaly detection. In addition, the system simultaneously compares temperature changes to relevant bearings to improve the accuracy of anomaly detection and diagnosis. Figure 15 shows the acoustic order energy and energy spectrum trend for a coupling-type fan with an acoustic sensor in the motor housing at low-speed (50% of the motor’s normal speed) and variable-speed (i.e., variable frequency control) conditions. The anomaly signal (bold line) clearly shows that the anomaly can be effectively detected during low-speed operation or under transient or variable speed conditions.

4.6. Smart Control

The proposed system is designed for integration into Industry 4.0-type smart factories and integrates diagnostic, control, and production systems to achieve monitoring, diagnostics, control, and management functions. Figure 16 shows the smart factory architecture based on rotating equipment. The architecture is centered on the proposed Intelligent Prediction Integration System with Internet (IPII), including sensing, intelligent control, MES, and ERP four levels. The sensing level collects operational information, which is then through IPII passed up to the factory’s MES and ERP level for production management and resource integration. The system’s intelligent control level provides overall control planning to implement individualized smart control programs for various rotating machinery.

Diagnosis is combined with control to determine appropriate operating modes or policies according to the current state of specific equipment. The ultimate goal of these diagnostic techniques is early malfunction prediction, thus preventing unexpected failure. IPII integrates acoustic, vibration, temperature, and rotating speed data. In addition to fault prediction and diagnosis, it can also be integrated with smart controls. The system can detect abnormalities and integrates smart controls to initiate appropriate responses. If abnormal vibrations, noise, or temperatures are detected, the system can not only immediately alert engineering supervisors but automatically execute appropriate countermeasures. Smart controls are designed to effectively maintain optimal production conditions while preventing unexpected shutdowns by automatically adjusting rotating speed, bringing backup equipment online, or adjusting production parameters to ensure that fan malfunctions do not negatively impact normal production line operations or affect equipment or personnel safety.

5. Conclusion

The difference between the proposed ideal and the state-of-the-art works is a different choice of the sensor. We believe that sensor selection is the key to diagnostic results. In the present studies demonstrated in the literature, most of them directly select vibration sensors for use in a variety of different diagnostic methods with markedly different results. First, this study uses MFCC filtering, commonly used in automatic speech recognition, to analyze the differences between the sensing characteristics of the acoustic and vibration sensors. Secondly, the study seeks to provide a better understanding of signal characteristics for diagnosing equipment failure and the corresponding characteristics of the selected sensor. That is, we first conduct a comparison analysis of the relative advantages of the sensing characteristics of acoustic and vibration sensors in the high-frequency range, to better understand anomalous signals occurring in the high-frequency range for early failure detection of industrial fans. Based on the results, we select the acoustic sensor to replace the vibration sensor. In factory settings, this approach is highly different from traditional vibration-based diagnostic methods. Experimental results verify that the acoustic sensor can provide earlier and more effective failure diagnosis. A review of the literature regarding failure diagnosis for rotating machinery found no mention of previous efforts to use differences in equipment signal anomalies and sensor-sensing characteristics as the basis for sensor selection. Also, we believe that this approach has reference value for sensor selection for early failure detection in other specific types of equipment.

Although the proposed diagnostic method can be applied to most types of rotating machinery, the approach is mainly designed to diagnose potential failure in industrial fans in factory settings. Failure to detect industrial fan failure early can have a significant impact on production; thus, we combine the knowledge of industrial fan experts to elucidate the operational characteristics and fault signal characteristics for industrial fans and analyze the results to select an appropriate sensor and diagnostic method. Experimental results show that the proposed method is very suitable for the early diagnosis of industrial fan failure. A literature review for the previous three years has also been conducted to discover other examples of useful methods for fault diagnosis for industrial fans. The system allows for the fault diagnosis of various types of industrial fans. The original system design considers various types of industrial fans, installation environments, and operating characteristics, and thus, system monitoring accounts for various fan types (e.g., coupling, belt, gear box, and direct), detection types, and sensor types. The proposed algorithm uses these various hardware structures, sensor characteristics, and fault signal characteristics as conditions for adjusting and compensating for the corresponding parameters to optimize actual online test performance.

To summarize, in the technology development of the predictive maintenance for the rotating equipment of factories, IPII possesses three major contributions as compared with the ideas demonstrated in the literature. The first is to replace the “vibration” sensor having only better low-frequency sensitivity to the “acoustic” one. The system equipped with the vibration sensor is subject to poor diagnosis in early failure. On the contrary, the acoustic sensor gives a system better high-frequency response advancing the early diagnosis in system failure. The second is the “property” of the detected dynamic signals. In this regard, we have proposed to adopt the order-tracking technology which can be used to analyze the relationship between the order and rotation speed. The most important thing in this design is the distribution in the order map rather than only the changes in the spectrum. In other words, our way is to fetch each of the order energy calculated from the combination of instant rotation speed and its corresponding sound (or vibration signals). This idea has been proven to be quite effective while being applied to low-rotation-speed equipment, especially for those systems requiring signal analysis of the variable rotation speeds. Last but not least, combining the ASK with the ASEF to deal with, respectively, the periodic and nonperiodic signals is capable of predicting the features causing system failure, where filtering and abnormal feature analysis are carried out. In addition to fault prediction, IPII can also allow for the use of smart controls and remote monitoring via networked devices. The control mode and process parameters of the critical rotating machinery can be instantly adjusted in response to fault signs to prevent production downtime and to reduce maintenance costs. The proposed system can be practically applied in Industry 4.0 smart factories.

Data Availability

Data are provided in the figures and table within the article.

Conflicts of Interest

The authors declare no conflict of interest.

Authors’ Contributions

Cihun-Siyong Alex Gong generated the research ideas, supervised the entire work, polished the article, and proofread the technical details. Huang-Chang Lee assisted in the design. Yu-Chieh Chuang performed the experiments and wrote the manuscript based on his master’s thesis archived domestically in Chinese in Taiwan. Tien-Hua Li, Chih-Hui Su, Lung-Hsien Huang, and Chih-Hsiung Chang technically supported the project. Yih-Shiou Hwang served as technical consultant for signal processing and application. Jiann-Der Lee supervised this work and provided funding support.


The authors appreciate the support from the National Science Council (NSC), Taiwan, and Ministry of Science and Technology (MOST), Taiwan, under Grants MOST 106-2221-E-182-005-, MOST 105-2221-E-182-039-, MOST 103-2221-E-182-070-, MOST 103-2815-C-182-012-E, MOST 104-2815-C-182-052-E, MOST 104-2221-E-182-044, and MOST 104-2221-E-182-023-MY2. This work is also supported in part by the Linkou Chang Gung Memorial Hospital (CGMH) under Contracts CMRPD2F0103, CMRPD2G0331, and CMRPD2H0041.