Adaptive Mode Decomposition Research Articles

A novel integrated prediction method using adaptive mode decomposition, attention mechanism and deep learning for coking products prices

Accurate prediction of coking product prices is crucial for enhancing production efficiency, cost optimization, and profit maximization in smart coking facilities. To address the volatility caused by nonlinear factors such as raw material costs, substitutes, macroeconomic indicators, sudden events, policy changes, and market behaviors, we propose a novel integrated prediction method for coking product price prediction. This method combines Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) for signal decomposition, Bidirectional Encoder Representations from Transformers (BERT) for natural language processing, attention mechanisms (AT) to weigh feature importance, and an ensemble of Bidirectional Gated Recurrent Unit, Bidirectional Long Short-Term Memory, and Gated Recurrent Unit, abbreviated BBG, for robust feature extraction. We design a feature selection strategy to avoid data leakage and improve the predictive ability of the model, and describe a method to maintain textual data information integrity when combining data from different sources. Experimental results on coke and methanol datasets show that our approach retains multi-source text richness improves predictive capability, and outperforms other state-of-the-art methods, providing an effective tool for developing smart coke plants.

Adaptive mode decomposition method based on fault feature orientation and its application to compound fault diagnosis of planetary gearboxes

Abstract Planetary gearboxes often experience multiple component failures during service, which can accelerate the degradation and failure of industrial equipment. Accurate separation and identification of multiple faults is an important means of ensuring the safe and stable operation of equipment. However, different faults can interact with each other, along with the influence of background noise, making it challenging to accurately extract faults with relatively weak energy among multiple faults. This difficulty leads to the problems of potential misdiagnosis and underdiagnosis. To address this issue, an adaptive mode decomposition method based on fault feature orientation (AMD-FF) is proposed in this paper. Initially, a fault impact indicator (FII) is constructed based on period-weighted kurtosis of envelope spectral and correlated combination negentropy to effectively characterize the impulsiveness and periodicity of fault features. Furthermore, with the objective of maximizing the FII, an adaptive decomposition of the original signal is designed based on blind convolution theory using a finite-impulse response filter group. Subsequently, a variable weight particle swarm optimization is employed to adaptively optimize the key decomposition parameters. Finally, the data of industrial-grade planetary gear transmission test rig are collected to validate the proposed method for compound fault diagnosis of planetary gearboxes. The results indicate that the AFMD-FF can effectively separate and extract compound faults in planetary gearboxes, demonstrating superior fault separation and diagnostic performance compared to the fault mode decomposition (FMD) and adaptive FMD. This method offers a novel approach to diagnosing compound faults in rotating equipment in industrial scenarios.

Solar photovoltaic power forecasting system with online manner based on adaptive mode decomposition and multi-objective optimization

Constructing an accurate and reliable solar photovoltaic (PV) power forecasting system is crucial for smart grid management and dispatch. However, due to the intermittent, non-stationary and random nature of solar energy, current methods cannot effectively capture the dynamic change patterns of PV data, resulting in a forecasting accuracy that cannot satisfy the requirement for stable operation of smart grids. To address this issue, in this paper, a new solar PV power forecasting system is proposed based on a new adaptive mode decomposition method that combines machine learning models. First an adaptive mode decomposition method is proposed for adaptively eliminating high-frequency information of PV data, mitigating the impact of high noise on the forecasting performance. Then the powerful mapping ability of machine learning model is utilized to construct a solar PV power generation combination forecasting system. The system adopts a multi-objective optimization strategy to determine the optimal weights of the combination model, which reduces the volatility of the forecasting results and enhances the stability of the forecasting system. Experimental results demonstrate that the proposed system achieves the best overall performance with online manner on 10 state-of-the-art baselines. Furthermore, the evaluation analysis by Diebold–Mariano test, improvement ratio and sensitivity analysis further show that the stability and accuracy of the proposed system outperforms the comparative baselines.

Prognostics of lithium-ion batteries health state based on adaptive mode decomposition and long short-term memory neural network

In the field of battery health state prognostics, the inaccurate lithium-ion battery's health status prediction is usually caused by the capacity regeneration (CR) phenomenon triggered by relaxation effects during the degradation process. To address this issue, we pay more attention to the main rapid degradation trend of battery capacity instead of many researches focusing on only CR phenomenon. This paper presents a prognostic framework called CEEMDAN-LSTM to decouple the normal capacity degradation process while eliminating local regenerative capacity, capturing degradation characteristics for battery state-of-health (SOH) prognostics. It introduces the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to decompose the original battery capacity degradation curve into principal trend sequence and other high-frequency subsequences, and Pearson correlation coefficient (PCC) is calculated to remove irrelevant high-frequency subsequences. Predictions task is completed by bidirectional long short-term memory network (Bi-LSTM) and long short-term memory network (LSTM) groups. Experimental validations are conducted on two lithium-ion battery datasets from NASA Ames Research Center and the Advanced Life-Cycle Engineering Center of the University of Maryland. The results demonstrate that the proposed framework achieves more accurate SOH prediction than many previous mainstream methods.

Adaptive mode decomposition based on the instantaneous frequency basis space

Signal decomposition has attracted increasing interest in various applications in recent years. This paper proposes a new mode decomposition algorithm, called “mode decomposition based on the instantaneous frequency basis space” (IFBMD), to decompose complex chirp signals superposed with multiple-mode signals with a low-frequency instantaneous frequency (IF). In IFBMD, a global optimization algorithm is combined with an alternating iterative computation method to extract the optimal IF of the modes from the low-frequency IF basis space. In addition, to avoid the limitation of a fixed number of modes, an adaptive IFBMD is designed based on a scale-factor iteration approach that can extract modes successively. This new scaling-factor updating method can be widely usable to improve other mode-decomposition algorithms. Compared with other well-known algorithms, IFBMD has been proved to be superior in performance.

Bandwidth-aware adaptive chirp mode decomposition for railway bearing fault diagnosis

It is a challenging task to accurately diagnose a railway bearing fault since bearing vibration signals are under strong interferences from wheel–rail excitations. The commonly used Kurtogram-based methods are often trapped in components induced by the wheel–rail excitations while adaptive mode decomposition methods are sensitive to input control parameters. To address these issues, based on a recently developed powerful signal decomposition method, that is, adaptive chirp mode decomposition (ACMD), a novel method called bandwidth-aware ACMD (BA-ACMD) is proposed in this article. First, the filter bank property of ACMD is thoroughly analyzed based on Monte-Carlo simulation and then a bandwidth expression with respect to the penalty parameter is first obtained by fitting a power law model. Then, a weighted spectrum trend (WST) method is proposed to partition frequency bands and then guide the parameter determination of ACMD through the integration of the obtained bandwidth expression. In addition, according to the order of magnitude of the WST in each band, the BA-ACMD adopts a recursive framework to extract signal modes one by one. In this way, dominating signal modes related to wheel–rail excitations can be extracted and then subtracted from the vibration signal in advance so that the bearing faults induced signal modes can be successfully identified. Both simulation and experimental validations are conducted showing that BA-ACMD can effectively detect single and compound faults of railway bearings under strong wheel–rail excitations.

Deep learning-based adaptive mode decomposition and instantaneous frequency estimation for vibration signal

Under time-varying environments and loads, vibration monitoring data of structures exhibit nonstationarity with multiple frequency components, and may have obvious nonlinearity under some extreme loads. The time–frequency analysis method is ideal for analyzing these nonlinear and nonstationary signals. However, the vibration signals of large-scale structures typically contain a variety of frequency components, making it challenging for effective mode decomposition without the interference of false mono-components by existing adaptive mode decomposition methods. False mono-components will severely decrease the accuracy of instantaneous frequency. To address the mode mixing and false mono-components in mode decomposition, a data-driven approach based on deep learning for signal adaptive mode decomposition and instantaneous frequency estimation is proposed in this paper. By transforming the time–frequency analysis problem into an optimization task of a deep neural network, built-in redundant adaptive filters with sparse regularization are designed, then the accurate intrinsic mode functions and time-varying frequencies can be estimated. The proposed method is verified by the numerical nonstationary signal and vibration signal of an experimental long cable model. The results demonstrate the superior capability of the proposed method in decomposing a signal into mono-components adaptively without false mono-components and achieving high accuracy of corresponding instantaneous frequencies. From the examples, an accuracy of 98.52% is achieved for a noise-free signal with spectral overlap. The proposed method is more robust to noise, and the instantaneous frequency identification error is 4.90%, which is much lower than EEMD and VMD under a signal-to-noise ratio of 5 dB.

Difference mode decomposition for adaptive signal decomposition

Adaptive extraction of concerned components (CC) from mixed frequency components remains to be a challenging topic in various research domains. Most existing adaptive mode decomposition algorithms for extracting CC, such as wavelet transform, wavelet packet transform, singular value decomposition, empirical mode decomposition, empirical wavelet transform, and variational mode decomposition, are intrinsically adaptive band-pass filter banks and they face the following two tough problems. The first problem is that they cannot separate CC from interferential components in same frequency bands. The second problem is that it is not fully effective in automatically selecting CC distributed in different frequency bands by using some designed criteria. In this paper, a new decomposition approach called difference mode decomposition (DMD) is proposed to adaptively decompose a mixed signal into CC, reference components, and noise, and enrich the domain of adaptive mode decomposition. The proposed DMD relies on convex optimization and Fourier transform, and its decomposition is mathematically justified and composes physical interpretations. Analyses of simulated and real-world bearing and gear vibration signals are used to verify the effectiveness and superiority of the proposed DMD over existing adaptive mode decomposition algorithms. It is demonstrated that the proposed DMD can effectively extract CC such as repetitive transients caused by bearing and gear faults. Moreover, since the proposed DMD is based on Fourier transform expanded by trigonometric basis functions, the proposed DMD can be easily extended to other domains by expanding basis functions besides the frequency domain.

Identification and separation of coupled vibration sources in multi-rotor gas turbines under time-varying speed conditions

Vibration sources in multi-rotor gas turbines under time-varying speed conditions are strongly coupled in the time–frequency domain, which prevents further mining of dynamic information from measured signals. Hence, a novel scheme of vibration source identification and separation is proposed to obtain individual source information. Firstly, the order and resonance components are identified by order analysis, and precise order numbers are further determined by computed order tracking. Then, in four typical vibration source coupling scenarios, the performances of Vold-Kalman filter (VKF), several typical adaptive mode decomposition methods and derivative methods are comparatively studied using Jeffcott rotor simulation system. Finally, the optimal source separation based on VKF and variational mode decomposition/extraction (VMD/VME) is determined. Furthermore, the settings of VKF filter bandwidth are studied quantitatively, and the effects of VMD and VME penalty factors are analyzed. In conclusion, the proposed scheme effectively identifies and accurately separates vibration sources from the test bed and gas turbine datasets, which provides important evidences for vibration monitoring and control as well as vibration transmission analysis of gas turbines.

State of the art on vibration signal processing towards data‐driven gear fault diagnosis

Gear fault diagnosis (GFD) based on vibration signals is a popular research topic in industry and academia. This paper provides a comprehensive summary and systematic review of vibration signal-based GFD methods in recent years, thereby providing insights for relevant researchers. The authors first introduce the common gear faults and their vibration signal characteristics. The authors overview and compare the common feature extraction methods, such as adaptive mode decomposition, deconvolution, mathematical morphological filtering, and entropy. For each method, this paper introduces its idea, analyses its advantages and disadvantages, and reviews its application in GFD. Then the authors present machine learning-based methods for gear fault recognition and emphasise deep learning-based methods. Moreover, the authors compare different fault recognition methods. Finally, the authors discuss the challenges and opportunities towards data-driven GFD.

Synchro-Reassigning Scaling Chirplet Transform for Planetary Gearbox Fault Diagnosis

It is difficult for classic time-frequency analysis (TFA) methods to characterize close-spaced nonlinear frequency curves with satisfactory time-frequency concentration and adaptively decompose above frequency components. As such, a novel TFA technique, termed synchro-reassigning scaling Chirplet transform (SRSCT) is developed in this paper. The proposed SRSCT is inspired by the scaling-basis Chirplet transform (SBCT) and reassignment theory. The novelties of the SRSCT are concluded as (a) a synchro-reassigning operator (SRO) is constructed for adaptively calculating the ideal time-frequency amplitudes on the instantaneous frequency (IF) ridges from the SBCT result and removing other smeared time-frequency coefficients, as a result, time-frequency resolution of the novel time-frequency representation (TFR) is greatly improved, close-spaced frequency curves are clearly characterized and noise interference is eliminated; and (b) perfect adaptive mode decomposition ability is achieved by combining the SRO, ridge detection method and vold-Kalman filter (VKF). Analysis results of the simulated signal show the effectiveness of the proposed SRSCT. Experimental analysis on faulty planetary gearbox signal shows that the proposed method can clearly characterize fault features. Compared with eight classic TFA methods, the Rényi entropies of the TFRs calculated by the proposed method are smallest in two cases, and they are 2.75 and 3.03, respectively, which means that the SRSCT can generate the TFR with the highest energy concentration.

Improved shock tube method for dynamic calibration of the sensitivity characteristic of piezoresistive pressure sensors

The dynamic calibration of pressure sensors with shock tube system are inevitably affected by the non-ideal shock wave characteristics and complex noise interferences, which limits the achievable accuracy of the dynamic pressure measurements. This paper proposes an improved shock tube (IST) method for identifying the sensitivity characteristic of piezoresistive pressure sensors. The incident shock wave (ISW) velocity attenuation model is first established by a distributed ISW velocity measurement combined with least squares fitting to compensate the shock wave pressure in shock tube. Besides, a signal correction method based on the hybrid adaptive mode decomposition is presented to eliminate the effects of complex noise on the response signals of pressure sensor. An accurate dynamic calibration result of the sensitivity characteristic is finally achieved with the compensated shock wave pressure and corrected response signal. A series of dynamic calibration experiments for a piezoresistive pressure sensor are carried out to verify the performance of the IST method. Results show that the IST method is able to effectively reduce the effects of the non-ideal shock wave and complex noises on the sensitivity calibration results. The mean relative error of sensitivity calibration results with the IST method is only 0.55%, which is about 1/7 times of the results obtained by the usual shock tube method. Furthermore, the sensitivity calibration uncertainty is evaluated and the smaller uncertainty results further verify the superiority of the IST method in high-accuracy dynamic calibration of piezoresistive pressure sensors.

Rolling Bearing Fault Feature Extraction Using Local Maximum Synchrosqueezing Transform and Global Fuzzy Entropy

To achieve good performance of fault feature extraction for a rolling bearing, a new feature extraction method is presented in this paper based on local maximum synchrosqueezing transform (LMSST) and global fuzzy entropy (GFuzzyEn). First, targeting the time-varying features of the vibration signals of the rolling bearing, the LMSST algorithm, which is a newly developed time-frequency method and allows for adaptive mode decomposition, is used to preprocess the vibration signals into a number of mode components. Then, as a modification of FuzzyEn, GFuzzyEn is adopted to evaluate the complexity of these mode components. Compared to FuzzyEn, which focuses mainly on the local characteristics of the short-term physiological time series, the GFuzzyEn emphasizes the global characteristics of the signal considering that the bearing vibration signals' global fluctuation may change as the bearing works under various conditions. Finally, the fault features of the bearing vibration signals are extracted by combining the LMSST and the GFuzzyEn. The experimental analysis shows that the proposed LMSST-GFuzzyEn method can extract rich fault-related information from the bearing vibration data and can achieve good classification performance for rolling bearing fault diagnosis.

Integrated early fault diagnosis method based on direct fast iterative filtering decomposition and effective weighted sparseness kurtosis to rolling bearings

The fault vibration signal of rolling bearing is typical non-stationary and non-linear signals. It usually exhibits the characteristics of amplitude modulation and frequency modulation (AM-FM) due to the interaction between the various components within the bearing and modulation of the rotor speed. In addition, the incipient weak fault feature information is inevitably masked by the strong background noise and other vibration interference because of the influence of the working environment of mechanical equipment. Therefore, new high efficiency and effective processing methods are indispensable for feature extraction of bearing fault diagnosis signals. Direct fast iterative filtering (dFIF) is a new kind of adaptive mode decomposition method proposed by Antonio Cicone based on iterative filtering (IF) and fast iterative filtering (FIF). dFIF is used to quickly decompose multi-component signals into a set of intrinsic mode functions (IMFs) by means of fast Fourier transform (FFT). In order to study the decomposition performance and examine the feature extraction ability of the dFIF method, a group of harmonic signals with simple linear superposition, multi-components AM-FM signals with different data length and different signal-to-noise ratio (SNR) are simulated and analyzed. For comparison, other mode decomposition methods were also applied to the simulated signals, such as adaptive local iterative filtering (ALIF), local mean decomposition (LMD), and empirical mode decomposition (EMD). The results show that dFIF decomposition methods have faster decomposition speed and better accuracy than other methods. Furthermore, the results also demonstrate excellent performance in AM-FM feature extraction and anti-mode mixing. On the other hand, the effective weighted sparseness kurtosis (EWSK) index integrates the periodicity and intensity of impact in each mode, which can effectively identify the effective mode. Therefore, considering the above merits, EWSK is integrated on the dFIF to effectively accomplish the fault diagnosis for rolling bearing under strong background noise. Firstly, the raw vibration signals of rolling bearing fault are decomposed into a set of IMFs by using dFIF. The EWSK indicator is then utilized to select IMFs with rich bearing fault impulse information for reconstruction. Finally, the Hilbert envelope demodulation analysis is used to analyze the reconstructed signal, extract the bearing fault feature and then judge the fault type. This proposed method is applied to analyze the simulated and field measured vibration signals of rolling bearing faults. Simultaneously, the other above-mentioned adaptive TFA methods are also adopted to analyze the signals. The results compared show that the proposed method has excellent performance in the aspects of decomposition speed and accuracy under the influence of strong background noise. It demonstrates the effectiveness and applicability of the proposed method for signal processing and fault pattern recognition of rolling bearings.

Research on Fault Feature Extraction of Rolling Bearing Based on dFIF

Aiming at the problems that the traditional adaptive mode decomposition method has low decomposition efficiency, poor ability of anti-mode mixing and decomposition accuracy, a rolling bearing fault feature recognition method based on direct fast iterative filtering (dFIF) and the crest factor of envelope spectrum is proposed. dFIF first decomposes the raw rolling bearing vibration signal into a set of intrinsic mode function (IMFs), then the optimal component is chosen based on the maximal crest factor of envelope spectrum, and finally the bearing fault characteristic frequency and its frequency multiplication are extracted using envelope demodulation analysis. It is demonstrated by relevant experimental data that the proposed method has efficient decomposition speed, accurate decomposition accuracy, and the accurate selection of effective components, and can effectively realize rolling bearing fault detection.

Enhancement of adaptive mode decomposition via angular resampling for nonstationary signal analysis of rotating machinery: Principle and applications

Vibration signal analysis provides an effective approach for condition monitoring and fault diagnosis of rotating machines. Under time-varying conditions, vibration signals feature nonstationarity, consist of multiple frequency components, and usually have spectral overlaps. Adaptive mode decomposition methods can extract mono-components from a given multi-component signal to meet the requirement for estimating instantaneous frequency. Among them, the recently proposed methods, including empirical wavelet transform, variational mode decomposition and Fourier decomposition method, outperform the classic empirical mode decomposition in terms of rigorous mathematical formulation. Nevertheless, for multi-component signals with spectral overlaps, these three methods are subject to mode mixing and/or integrity issues, and fail to extract true mono-components, because they essentially separate mono-components based on spectral segmentation. In this paper, we propose a framework by exploiting the capability of angular resampling to address the mono-component overlapping issue. This methodology can separate true mono-components, thus facilitating accurate estimation of instantaneous frequency through Hilbert transform and generating perfect time–frequency representations (TFRs). Firstly, angular resampling is employed to make the constituent mono-components well separable in the frequency domain. Then, true mono-components are separated through adaptive mode decomposition. Next, informative mono-components are selected for further processing based on the prominent order information, and the selected mono-components are mapped into the time domain according to the relationship between the equal time and equal angle sampling. Finally, the instantaneous frequency and amplitude envelope of the recovered mono-components are calculated via Hilbert transform, and the TFR of raw signal is obtained by superposing the TFRs of all recovered mono-components. By doing so, the TFR achieves a fine time–frequency resolution and is free of both outer and inner interferences. The proposed methodology is demonstrated by simulated signal analysis, and further validated using the vibration data sets of three typical rotating machines (including a planetary gearbox in a wind turbine drivetrain, a civil aircraft engine and a hydraulic turbine rotor). The analysis results show its excellent capability to reveal the time-varying features of rotating machinery nonstationary signals.

A Comparative Analysis of Signal Decomposition Techniques for Structural Health Monitoring on an Experimental Benchmark.

Signal Processing is, arguably, the fundamental enabling technology for vibration-based Structural Health Monitoring (SHM), which includes damage detection and more advanced tasks. However, the investigation of real-life vibration measurements is quite compelling. For a better understanding of its dynamic behaviour, a multi-degree-of-freedom system should be efficiently decomposed into its independent components. However, the target structure may be affected by (damage-related or not) nonlinearities, which appear as noise-like distortions in its vibrational response. This response can be nonstationary as well and thus requires a time-frequency analysis. Adaptive mode decomposition methods are the most apt strategy under these circumstances. Here, a shortlist of three well-established algorithms has been selected for an in-depth analysis. These signal decomposition approaches—namely, the Empirical Mode Decomposition (EMD), the Hilbert Vibration Decomposition (HVD), and the Variational Mode Decomposition (VMD)—are deemed to be the most representative ones because of their extensive use and favourable reception from the research community. The main aspects and properties of these data-adaptive methods, as well as their advantages, limitations, and drawbacks, are discussed and compared. Then, the potentialities of the three algorithms are assessed firstly on a numerical case study and then on a well-known experimental benchmark, including nonlinear cases and nonstationary signals.

A Methodology to Handle Spectral Smearing in Gearboxes Using Adaptive Mode Decomposition and Dynamic Time Warping

Tacho-less order tracking is an effective tool for fault detection in gearboxes operating under speed fluctuations. This technique's performance depends on extracting instantaneous shaft speed information from a complex multicomponent gearbox signal, which is first preprocessed by a bandpass filter. However, spectral overlap resulting from the time-varying frequency components makes it challenging to isolate the shaft speed harmonic mono-component using this approach. In order to overcome such issues, an adaptive tacho-less order-tracking (OT) method combining the variational mode decomposition (VMD) and the fast dynamic time warping (FDTW) is proposed in this article. The proposed method first decomposes the measured gearbox vibration signal using VMD to estimate the instantaneous shaft speed profile to construct the shaft vibration signal. The gearbox vibration signal is then resampled based on the optimal warping path obtained by FDTW, which performs an “elastic” stretching and compression along the time axis of the extracted shaft vibration signal with respect to a sinusoidal reference signal with a constant shaft rotational frequency. Finally, the presence of gear fault is detected by constructing the order spectrum of the resampled vibration signal. The effectiveness of the proposed algorithm is demonstrated using both simulation analysis and experimental validation using measurements from two different wind-turbine gearboxes collected under test and field conditions, respectively.

A Novel Adaptive Mode Decomposition Method Based on Reassignment Vector and Its Application to Fault Diagnosis of Rolling Bearing

To solve the problem that the random distribution of noise in the time-frequency (TF) plane largely affects the readability of TF representations, a novel signal adaptive decomposition algorithm processed in TF domain, which provides adequate information about the time-varying instantaneous frequency, is presented in this paper. The theoretical basis of this algorithm is short-time Fourier transform (STFT). The research into the algorithm comprises two steps: the TF plane denoising takes sparse low-rank matrix estimation as a priority and then achieves signal decomposition based on reassignment vector (RV). A low-rank matrix approximation scheme, which exploits the sparse properties of the TF transformation coefficient and uses non-convex penalty, is put forward to obtain clean STFT. Then, a new approach called RV, which is different from the traditional mode decomposition methods such as Empirical Mode Decomposition (EMD), is used to estimate the characteristic curve corresponding to the TF ridges of the interested modes. Based on the classical reassignment method, RV has a solid theory foundation. Moreover, it can identify different signal components such as stationary signal, modulating signal and impulse characteristic. Combining the advantages of low-rank matrix approximation approach and those of RV defined in TF plane, a novel signal adaptive decomposition method is proposed in this paper to identify fault characteristics. To illustrate the effectiveness of the method, fault signals of rolling bearing under stationary condition and time-varying speed are respectively analyzed.

End-to-End Noisy Speech Recognition Using Fourier and Hilbert Spectrum Features

Despite the progress of deep neural networks over the last decade, the state-of-the-art speech recognizers in noisy environment conditions are still far from reaching satisfactory performance. Methods to improve noise robustness usually include adding components to the recognition system that often need optimization. For this reason, data augmentation of the input features derived from the Short-Time Fourier Transform (STFT) has become a popular approach. However, for many speech processing tasks, there is an evidence that the combination of STFT-based and Hilbert–Huang transform (HHT)-based features improves the overall performance. The Hilbert spectrum can be obtained using adaptive mode decomposition (AMD) techniques, which are noise-robust and suitable for non-linear and non-stationary signal analysis. In this study, we developed a DeepSpeech2-based recognition system by adding a combination of STFT and HHT spectrum-based features. We propose several ways to combine those features at different levels of the neural network. All evaluations were performed using the WSJ and CHiME-4 databases. Experimental results show that combining STFT and HHT spectra leads to a 5–7% relative improvement in noisy speech recognition.