^{1}

^{*}

^{1}

^{1}

^{2}

This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (

Blower and exhaust fans consume over 30% of electricity in a thermal power plant, and faults of these fans due to rotation stalls are one of the most frequent reasons for power plant outage failures. To accurately predict the occurrence of fan rotation stalls, we propose a support vector regression machine (SVRM) model that predicts the fan internal pressures during operation, leaving ample time for rotation stall detection. We train the SVRM model using experimental data samples, and perform pressure data prediction using the trained SVRM model. To prove the feasibility of using the SVRM model for rotation stall prediction, we further process the predicted pressure data via wavelet-transform-based stall detection. By comparison of the detection results from the predicted and measured pressure data, we demonstrate that the SVRM model can accurately predict the fan pressure and guarantee reliable stall detection with a time advance of up to 0.0625 s. This superior pressure data prediction capability leaves significant time for effective control and prevention of fan rotation stall faults. This model has great potential for use in intelligent fan systems with stall prevention capability, which will ensure safe operation and improve the energy efficiency of power plants.

The fan is a common type of air flow control component, widely used in many impeller machines, such as ventilators, compressors, and pumps. These machines are employed in almost every field of industry and consume a significant amount of electricity. In thermal power plants, fans are the power sources of smoke exhaust systems, and blowing fans and induced draft fans of boilers account for approximately 30% of the plant electric consumption [

According to the statistics of the Electric Power Research Institute (EPRI), 30%∼50% of all power plant outage failures are caused by rotating machinery, such as steam turbine generators, fans and pumps. Based on the Chinese statistics of power plant operations in 2007, blower fan and induced draft fan failures are the second frequent cause of unscheduled shutdowns of thermal power plants, which causes significant economic losses. Therefore, effective approaches for online diagnosis and active control of common faults in power plant fans are urgently needed to ensure the reliable operation of the plant. Avoiding operations of the fans in their close-to-fault modes will also improve the energy efficiency of the impeller machinery.

Rotation stalls are the most common cause of fan faults. One or several stall cells will form in the fan impeller passage when the flow of impeller machinery reduces to a certain limit. The occurrence of rotation stalls changes the stable flow field in the impeller, generates additional loads, and can even cause fatigue and fracture of the blades. If the rotation stall cannot be effectively contained, a flow surge may occur and lead to decreased unit efficiency and even vibrations of the fan body and connection pipes, which could possibly cause damages. Therefore, particular attention should be paid to the weak-stall mode of fans during operation, and active prediction and subsequent control of early-stage rotation stall in fans should be pursued.

The mechanism and detection of fan rotation stall have been actively studied in the industrial fan and compressor community. The mechanism of rotation stall has been investigated through signal processing methods or experimental approaches. For instance, Longley

For rotation stall detection, Cameron and Morris [

Most of the aforementioned research work targeted rotation stalls of the compressors of aero-engines, studied the unsteady characteristics of rotation stalls, and explored approaches for expanding the steady state operation and detecting stalls online. However, compared with the axial flow compressors, research on the rotation stalls of centrifugal fans is limited. There are significant differences in the rotation stall mechanisms between centrifugal fans and axial flow compressors [_{2} emissions.

This article presents a novel intelligent model, based on support vector regression machine (SVRM), for predicting the rotation stalls of centrifugal fans from pressure data measured online. Many experiments are performed under different running conditions on a centrifugal fan setup to obtain pressure signals of the fan operation process from normal state to rotation stall. The frequency characteristics of the pressure signals before and during rotation stalls are first analyzed using Fourier transform, and the pressure signals are then used to train the SVRM prediction model. The trained prediction model can accurately predict the fan pressure data 0.0625 s in advance, based on which the rotation stalls can be reliably detected. This model allows the active control of rotation stalls before their occurrence, which represents a major advantage over existing stall prediction techniques.

The experiments were conducted on the 4-73No8D centrifugal fan, which is a common model from the 4–73 series widely used in Chinese thermal power plants. The experimental setup, as schematically shown in _{v} = 0.205; accuracy of speed adjustment: 0.3 rotations) attached to an alternating-current (AC) motor, inlet and outlet pipelines connected to the fan, an inlet flow divider between the inlet pipeline and the fan, and a signal conditioning unit connected with a computer. Five piezoresistive pressure sensors (measurement range: 20 kPa; accuracy: 10 Pa) were arranged on the inner surface of the fan casing with an angular spacing of 60° (

For pressure data analysis, the Fourier transform method was first adopted to extract the frequency distribution, the fundamental frequency, and the harmonic frequencies of the pressure data obtained in normal operations and fan rotation stall conditions. These results helped us understand the frequency characteristics of the fan rotation stalls. However, the Fourier transform results cannot reflect the gradual development of fan rotation stall, and thus cannot be used for rotation stall prediction. In recent years, with the advances in unsteady flow research, wavelet analysis has been applied to the time-frequency characteristic analysis of the mechanically unsteady flows of impellers. In this research, we applied the wavelet analysis to decipher the time-frequency characteristics of the pressure signals obtained during the transition process from normal operation to rotation stall, to quantitatively reveal the gradual development of rotation stalls, and to formulate a novel method for predicting the onset of rotation stalls.

The signals from all the five pressure sensors were analyzed using fast Fourier transform (FFT), and the FFT spectra were used to: (i) quantify the fundamental frequencies of the fan rotor and the rotating stall group, and (ii) reveal the graduate development of the stall group. Using the experiment with a full opening (0°) of the guide vane as an example,

From _{v}

Through Fourier transform of the pressure signals, we have clearly observed the gradual development of the rotation stall and its frequency characteristics. However, the Fourier transform cannot reveal the time-domain characteristics of rotation stalls, and thus cannot be used for stall prediction. Differently, wavelet transform can identify signal components with different frequencies and the time associated with these frequencies, and has been widely applied to analysis of non-stationary signals [

In this work, the Daubechies orthogonal wavelet was utilized for multi-resolution decomposition of fan pressure signals and obtaining of the wavelet coefficients of the relevant frequency bands, which were used for detecting the rotation stall from the pressure data. The pressure signal (denoted by ^{th}-level approximation component (denoted by _{N}_{i}_{N}_{N}_{N}_{N}_{2} + _{1}. Changes in the frequency and amplitude of the pressure signal during fan operation were examined by analyzing a specific detail component _{j}

The pressure signals during the gradual development of fan rotation stall were obtained at 1300 RPM, and the opening of the guide vane was set to be 0°, 15°, 30° and 45°. The wavelet transform results of the pressure data from sensor #2 (guide vane opening: 15°), obtained during the rotating stall development, are shown in

Thus, we used the four-layer detail coefficient as a quantitative indicator for detecting the onset of a rotation stall. For the pressure data from sensor #2 (guide vane opening: 15°), we empirically set a threshold of 200 Pa for the fourth-layer detail coefficient, and any pressure data point yielding a four-layer detail coefficient above 200 Pa indicates the presence of the rotation stall. This empirically-chose threshold was proven to be robust for other pressure data sets collected from sensor #2 at the guide vane opening of 15°. In

At the initial stage of fan rotation stall, the formed stall groups are not stable and tend to circularly occur, rupture, disappear, and re-develop with changes of the flow field [

To effectively predict the fan rotation stall, we proposed a support vector regression machine (SVRM) model for multi-step prediction of the pressure data. These predicted data were then used to predict the fan rotation stall via wavelet transform. For initial training of the SVRM model, sample data of the pressure signals were obtained during the transition process from normal operation to rotation stall. Since the SVM model can predict the pressure data of the fan by a certain time advance, the rotation stall can then be

The proper selection of a suitable prediction model is the key to ensure the precision and robustness of the rotation stall prediction. Existing regression models for data prediction mainly include support vector machines [

For the training of SVRMs, the feature extraction and selection are crucial tasks. Feature extraction is difficult in many prediction applications, and feature selection is largely subjective and usually induces lack or redundancy of information. If we simply take the time series as the input vector for SVRM training, the input vector contains less information and will result in low accuracy of prediction. Therefore, we employed a phase space transformation approach based on chaos theory for preprocessing the time series and reconstructing their phase space which was then used as the input vector for SVRM training.

The chaos characteristics of the pressure data were first analyzed using chaos theory [

The evolution of any signal in a system is determined by its interactions with other signal components. Thus, the evolution of one signal includes the information of other related signals. The characteristic of the system can be extracted from a set of time series data of any signal in the system, and this characteristic can be represented by a trajectory in a high-dimensional space. Takens [

Using _{1}, _{2}, …, _{M}

The selection of the

Using the reconstructed phase space of the pressure time series (_{1}, _{2}, …, _{M}_{i}

The function _{i}^{T}ϕ_{i}_{i}_{i}

Here, training vectors _{i}

Accurate determination of the parameters ^{−10}, 2^{−9}, … , 2^{10}) were used for iteration. To reduce the computation complexity, we first performed a coarse grid search to identify a region including the optimal combination of

After being trained using the sample pressure data, the SVRM model is ready for rotation stall prediction. The major advantage of this SVRM prediction model is that it can provide accurately predicted pressure data ahead of a certain time, which enables

With a fixed sampling frequency, the time advance by which the rotation stall can be predicted mainly depends on the number of prediction steps performed by the SVRM model. We performed pressure data prediction in single and multiple steps. In single-step prediction, we fed five measured pressure data points into the SVRM model to predict the next pressure data point. In multi-step prediction, we iterated the predicted pressure data points as part of the input data of the SVRM model for next prediction. The number of predicted data points used for next prediction was defined as the number of prediction steps. Multi-step predictions are desired to increase the time advance of the stall prediction. However, the prediction errors are accumulated in these multiple steps, and we must verify the maximum number of steps by which the SVRM model can predict the pressure data accurately enough to guarantee reliable stall prediction. In the experiments, we predicted the pressure data at 1–6 steps and demonstrate that the pressure data predicted by up to five steps are reliable for wavelet-transform-based stall prediction.

To further increase the time advance of stall prediction, we also increased the time interval of pressure data sampling by decreasing the sampling frequency from 320 Hz to 80 Hz. Since the rotational frequency of the centrifugal fan is 21.7 Hz, the new sampling frequency of 80 Hz still satisfies the sampling frequency theorem. The time-series pressure data were obtained from pressure sensor #2 at 80 Hz when the centrifugal fan was operated at 1300 rpm with guide vane openings of 0°, 15°, 30° and 45°. Two hundred and fisty pressure data points were obtained for each guide vane opening.

The time-series pressure data were first converted into their reconstructed phase space _{1}, _{2}, …, _{M}

The time advance of rotation stall prediction is the most important parameter of any stall prediction technique. We determined the longest time advance of our SVRM prediction model by examining the maximum number of prediction steps the model can perform. Multi-step prediction was performed by iterating the previously predicted pressure data points as part of the input vector of the SVRM model. The experimental results show that, by up to five prediction steps, the SVRM model can predict pressure data accurate enough to yield reliable wavelet-transform-based stall prediction.

^{−3} s) to ^{−1} s) [

This paper presents the development and experimental validation of a SVRM model for intelligent prediction of rotation stalls of centrifugal fans in power plants. The internal pressure of the fan casing was used as the experimental data for rotation stall prediction. The SVRM prediction model was established using chaotic phase space transformation and a SVRM model, and was demonstrated to perform single and multi-step (up to five steps) prediction of the pressure data which can be further processed by real-time stall detection techniques. To validate the feasibility of using the SVRM model for rotation stall prediction, offline wavelet transform was performed for rotation stall prediction using both the measured and predicted pressure data sets. Through comparison of the stall prediction results obtained from the measured and predicted pressure data, we confirmed that the SVRM model is capable of predicting the pressure data by up to five steps and yielding reliable stall prediction. The effective time advance of stall prediction was 0.0625 s, which is significant enough for a stall control mechanism to react and suppress the stall development. The experimental studies were made on a 4-73No8D centrifugal fan, which is one of the most widely used centrifugal fans in Chinese power plants. We believe that this SVRM prediction model will find important applications in experimental rotation stall prediction and control of centrifugal and other types of fans, and therefore improve the power efficiency of combustion-based power plants.

This work is supported by the Fundamental Research Funds for the Central Universities (12MS116).

The authors declare no conflict of interest.

Experimental setup for internal pressure measurements of a centrifugal fan operated at normal conditions and under rotating stalls.

The spectrum diagram of pressure signal (from sensor #2) in the development of a rotation stall. From (_{v}) of 0.172, 0.154, 0.151, 0.147, respectively. Starting from (

Fan casing pressure signal (from sensor #2) and its time-frequency map of wavelet analysis when βbeta; βequals; 15βdeg;. (

Single-step prediction results of the fan rotating stall at guide vane openings of 0°, 15°, 30° and 45° (panels (

Five-step prediction results of the fan rotation stall at guide vane openings of 0°, 15°, 30° and 45° (panels (

Six-step prediction results of the fan rotation stall at guide vane openings of 0°, 15°, 30° and 45° (panels (