Hidden Markov Model Model Research Articles

Evaluation of hidden Markov models using deep CNN features in isolated sign recognition

Isolated sign recognition from video streams is a challenging problem due to the multi-modal nature of the signs, where both local and global hand features and face gestures needs to be attended simultaneously. This problem has recently been studied widely using deep Convolutional Neural Network (CNN) based features and Long Short-Term Memory (LSTM) based deep sequence models. However, the current literature is lack of providing empirical analysis using Hidden Markov Models (HMMs) with deep features. In this study, we provide a framework that is composed of three modules to solve isolated sign recognition problem using different sequence models. The dimensions of deep features are usually too large to work with HMM models. To solve this problem, we propose two alternative CNN based architectures as the second module in our framework, to reduce deep feature dimensions effectively. After extensive experiments, we show that using pretrained Resnet50 features and one of our CNN based dimension reduction models, HMMs can classify isolated signs with 90.15% accuracy in Montalbano dataset using RGB and Skeletal data. This performance is comparable with the current LSTM based models. HMMs have fewer parameters and can be trained and run on commodity computers fast, without requiring GPUs. Therefore, our analysis with deep features show that HMMs could also be utilized as well as deep sequence models in challenging isolated sign recognition problem.

Modeling the received signal strength intensity of Wi-Fi signal using Hidden Markov Models

Wi-Fi fingerprinting is one of the methods that are widely used to provide Location Based Services (LBS). Gaussian, or a mixture of Gaussians, is the preferred model used by Wi-Fi fingerprinting for LBS. Nevertheless, Received Signal Strength Intensity (RSSI) Wi-Fi histograms are skewed, and a Gaussian model is not well suited for modeling data when their histogram is skewed. In addition, another important characteristic present in the RSSI Wi-Fi temporal series is autocorrelation, which cannot be modeled using a Gaussian model. In this paper, we explore the feasibility of using Hidden Markov Models (HMM) to model RSSI Wi-Fi signals. The mathematical derivation of formulas to calculate autocorrelation based on the HMM parameters is presented. Exhaustive experimentation, using data sampled in a real scenario, was performed to test the dependency of the autocorrelation coefficients on the number of hidden states, and the number of iterations used when creating the HMM. The results are compared with autocorrelation coefficients calculated using the real data. Kullback–Leibler (KL) divergence was used to compare the similarity of the real histograms and those provided by a mixture of Gaussians and by an HMM. HMM models reported more accurate results than a mixture of Gaussians model in both cases.

A grouped beta process model for multivariate resting-state EEG microstate analysis on twins.

EEG microstate analysis investigates the collection of distinct temporal blocks that characterize the electrical activity of the brain. Brain activity within each microstate is stable, but activity switches rapidly between different microstates in a nonrandom way. We propose a Bayesian nonparametric model that concurrently estimates the number of microstates and their underlying behaviour. We use a Markov switching vector autoregressive (VAR) framework, where a hidden Markov model (HMM) controls the nonrandom state switching dynamics of the EEG activity and a VAR model defines the behaviour of all time points within a given state. We analyze the resting-state EEG data from twin pairs collected through the Minnesota Twin Family Study, consisting of 70 epochs per participant, where each epoch corresponds to 2 s of EEG data. We fit our model at the twin pair level, sharing information within epochs from the same participant and within epochs from the same twin pair. We capture within twin-pair similarity, using an Indian buffet process, to consider an infinite library of microstates, allowing each participant to select a finite number of states from this library. The state spaces of highly similar twins may completely overlap while dissimilar twins could select distinct state spaces. In this way, our Bayesian nonparametric model defines a sparse set of states that describe the EEG data. All epochs from a single participant use the same set of states and are assumed to adhere to the same state switching dynamics in the HMM model, enforcing within-participant similarity.

Application of Hidden Markov Models to Analyze, Group and Visualize Spatio-Temporal COVID-19 Data

The coronavirus epidemic (COVID-19) is a public health challenge due to its rapid global spread. Its unprecedented speed and pervasiveness have led many governments to implement a series of countermeasures, such as lock-downs, stopping/restricting travels, and mandating social distancing. To control and prevent the spread of COVID-19, it is essential to understand the latent dynamics of the disease’s evolution and the effectiveness of the intervention policies. Hidden Markov models (HMMs) capture both randomnesses in spatio-temporal dynamics and uncertainty in observations. In this paper, we apply an overall HMM that, based on multiple nations’ COVID-19 data including the USA, several European countries, and countries that have strict control policies, explore different types of observations, and we use it to infer the severity state on small geographical states or regions in the USA and Italy as test cases. Further, we aggregate the severity level of each region over a fixed time period to visualize the time evolution and propagation across regions. Such an analysis and visualization provide suggestions for interventions and responses in a calibrated manner. Results from HMM modeling are consistent with what is observed in Italy and the USA and these models can serve as visualization and proactive decision support tools to policymakers.

Towards Constructing HMM Structure for Speech Recognition With Deep Neural Fenonic Baseform Growing

For decades, acoustic models in speech recognition systems pivot on Hidden Markov Models (HMMs), e.g., Gaussian Mixture Model-HMM system, Deep Neural Network-HMM system, etc., and achieve remarkable results. However, the popular HMM model is the three-state left-to-right structure, without the superiority certainty. There are multiple studies on the HMM structure’s optimization, but none of them addresses this problem leveraging deep learning algorithms. For the first time, this paper proposes a new training method based on Deep Neural Fenonic Baseform Growing to optimize the HMM structure, which is concisely designed and computationally cheap. Moreover, this data-driven method customizes the HMM structure for each phone precisely without external assumptions concerning the number of states or transition patterns. Experimental results on both TIMIT and TEDliumv2 corpora indicate that the proposed HMM structure improves both the monophone system and the triphone system substantially. Besides, its adoption further improves state-of-the-art speech recognition systems with remarkably reduced parameters.

An Automatic Emergency Braking Model considering Driver’s Intention Recognition of the Front Vehicle

Driver’s intention of the front vehicle plays an important role in the automatic emergency braking (AEB) system. If the front vehicle brakes suddenly, there is potential collision risk for following vehicle. Therefore, we propose a driver’s intention recognition model for the front vehicle, which is based on the backpropagation (BP) neural network and hidden Markov model (HMM). The brake pedal, accelerator pedal, and vehicle speed data are used as the input of the proposed BP-HMM model to recognize the driver’s intention, which includes uniform driving, normal braking, and emergency braking. According to the recognized driver’s intention transmitted by Internet of vehicles, an AEB model for the following vehicle is proposed, which can dynamically change the critical braking distance under different driving conditions to avoid rear-end collision. In order to verify the performance of the proposed models, we conducted driver’s intention recognition and AEB simulation tests in the cosimulation environment of Simulink and PreScan. The simulation test results show that the average recognition accuracy of the proposed BP-HMM model was 98%, which was better than that of the BP and HMM models. In the Car to Car Rear moving (CCRm) and Car to Car Rear braking (CCRb) tests, the minimum relative distance between the following vehicle and the front vehicle was within the range of 1.5 m–2.7 m and 2.63 m–5.28 m, respectively. The proposed AEB model has better collision avoidance performance than the traditional AEB model and can adapt to individual drivers.

Disrupted brain dynamics across the Alzheimer’s disease spectrum is related to tau accumulation

AbstractBackgroundRecent evidence has implicated brain networks in the pathophysiology of Alzheimer’s disease (AD). However, the phenotypic variability in AD, inter‐individual differences in tau PET uptake patterns, and dynamic nature of functional connectivity have been relatively neglected. In this talk, I will discuss our recent work linking measures of dynamic brain organization to spatially independent tau PET patterns defined in a phenotypically diverse cohort.MethodsWe identified all Mayo Clinic participants with concurrent amyloid PET (PiB), tau PET (AV1451) and task‐free fMRI (impaired participants were all Aβ+). We used independent component analysis (ICA) to identify spatially independent patterns of tau PET uptake. For the fMRI data, we used a Hidden‐Markov Model (HMM) design to identify distinct patterns of activity/connectivity (‘states’). The ICA and HMM parameters were used to project/segment an independent cohort (ADNI) and subsequent longitudinal timepoints for the Mayo cohort. Multilevel Bayesian regression models were fit in Stan to explore the association between phenotype, component load and brain state.ResultsThe optimal HMM model identified 10 brain states (Figure 1). The optimal ICA decomposition identified 42 tau PET components, of which 13 were potentially related to AD pathobiology (Figure 1). Component loads and brain state dwell time differed between diagnostic groups in both Mayo and ADNI cohorts. One state was much more likely to occur in AD in both cohorts (State 8). Cross‐sectionally, State 8 dwell time was associated with higher tau loads, while higher baseline dwell‐time was predictive of ∆tau in participants with at least 2 timepoints (Figure 2).ConclusionWe found links between dynamics of functional connectivity, tau deposition, and cognitive impairment across the AD spectrum using two different cohorts and longitudinal data. Specifically: 1) our method for characterizing tau PET uptake was able to capture the magnitude and spatial pattern of binding across diverse phenotypes; 2) Our data supports a role for dynamic brain connectivity in cognitive impairment due to Alzheimer’s disease; 3) Dynamic connectivity appears to be directly related to AD pathobiology as measured by tau PET. Incorporating the time‐varying nature of connectivity will be important in future studies of connectome‐based disease progression in AD.

Detection of influenza epidemics using hidden Markov and Serfling approaches.

Detection of epidemics is a critical issue in epidemiology of infectious diseases which enable healthcare system to better control it. This study is devoted to investigating the 5-year trend in influenza and severe acute respiratory infection cases in Iran. The epidemics were also detected using the hidden Markov model (HMM) and Serfling model. In this study, we used SARI data reported in the World Health Organization (WHO) FluNet web-based tool from August 2011 to August 2016. SARI data in Iran from August 2011 to August 2016 were used. We applied the HMM and Serfling model for indicating the two epidemic and non-epidemic phases. The registered outbreak activity recorded on the WHO website was used as the gold standard. The coefficient of determination was reported to compare the goodness of fit of the models. Serfling models modified by 30% and 35% of the data had a sensitivity of 91.67% and 95.83%, while for 15%, 20% and 25% were 70.83%, 79.17% and 83.33%, respectively. Sensitivity of HMM and autoregressive HMM (AHMM) was 66.67% and 92.86%. All fitted models have a specificity of over 96%. The R2 for HMM and AHMM was calculated 0.73 and 0.85, respectively, showing better fitness of these models, while R2 was around 50% for different types of Serfling models. Both modified Serfling and HMM were acceptable models in determining the epidemic points for the detection of weekly SARI. The AHMM had better fitness, higher detection power and more accurate detection of the incidence of epidemics than Serfling model and high sensitivity and specificity. In addition to AHMM, Serfling models with 30% and 35% modification can be used to detect epidemics due to approximately the same accuracy but the simplicity of the calculations.

Condition Based Maintenance of Low Speed Rolling Element Bearings using Hidden Markov Model

This paper presents an integrated hidden Markov model (HMM) approach to undertake fault diagnosis and maintenance planning for low-speed roller element bearings in a conveyor system. The components studied are relatively long-life components for which run-to-failure data is not available. Furthermore, the large number of these components in a conveyor system makes the individual monitoring of each bearing impractical. In this paper, HMM is employed to overcome both these challenges. For fault diagnosis, a number of bearings varying in age and usage were extracted from the system and tested to develop a baseline HMM model. This data was then used to calculate likelihood probabilities, which were subsequently used to determine the health state of an unknown bearing. For maintenance planning, experimentally determined thresholds from faulty bearings were used in conjunction with simulated degradation paths to parametrize a HMM. This HMM is then used to determine the state duration statistics and subsequently the calculation of residual useful life (RUL) based on bearing vibration data. The RUL distribution is then used for maintenance planning by optimizing the expected cost rate and the results so obtained are compared with the results obtained from a traditional age based replacement policy.

802.11 wireless simulation and anomaly detection using HMM and UBM

Despite the growing popularity of 802.11 wireless networks, users often suffer from connectivity problems and performance issues due to unstable radio conditions and dynamic user behavior, among other reasons. Anomaly detection and distinction are in the thick of major challenges that network managers encounter. The difficulty of monitoring broad and complex Wireless Local Area Networks, that often requires heavy instrumentation of the user devices, makes anomaly detection analysis even harder. In this paper we exploit 802.11 access point usage data and propose an anomaly detection technique based on Hidden Markov Model (HMM) and Universal Background Model (UBM) on data that is inexpensive to obtain. We then generate a number of network anomalous scenarios in OMNeT++/INET network simulator and compare the detection outcomes with those in baseline approaches—RawData and Principal Component Analysis. The experimental results show the superiority of HMM and HMM-UBM models in detection precision and sensitivity.

Fuzzy‐HMM modeling for emotion detection using electrocardiogram signals

AbstractIn this paper, a Fuzzy Hidden Markov Model (FHMM) for electrocardiogram (ECG)‐based emotion recognition is proposed. The FHMM model is modeled using the fuzzy membership of each class of feature vectors to compute the elements of the matric in the model. Each element in the matric is determined by the two nearest classes of feature vectors with fuzzy classification to avoid the winner‐takes‐all situation that usually happens in tradition discrete Hidden Markov Models (HMM) and that reduces the accuracy of the modeling results. The FHMM modeling proposed in this paper can improve the precision of traditional discrete HMM modeling. Also, the features for emotion recognition, which are selected according to those used in prior research, are calculated from ECG signals. The selected features form a feature vector and these are calculated for all ECG signals. Some feature vectors are used to train the FHMM model and the remaining are used for testing. Experimental results show that the proposed FHMM model can achieve impressive improvements on the three indices, sensitivities (Se%), positive predictive value (PPV%), and total classification accuracy (TCA%), compared to traditional HMM. Moreover, compared to some existing studies, the recognition rates of the proposed method are higher. These results verify the efficiency of the proposed method for emotion recognition from ECG signals.

An improved feature extraction algorithm for insider threat using hidden Markov model on user behavior detection

PurposeBy using a new feature extraction method on the Cert data set and using a hidden Markov model (HMM) to model and analyze the behavior of users to distinguish whether the behavior is normal within a continuous period.Design/methodology/approachFeature extraction of five parts of the time series by rules and sorting in chronological order. Use the obtained features to calculate the probability parameters required by the HMM model and establish a behavior model for each user. When the user has abnormal behavior, the model will return a very low probability value to distinguish between normal and abnormal information.FindingsGenerally, HMM parameters are obtained by supervised learning and unsupervised learning, but the hidden state cannot be clearly defined. When the hidden state is determined according to the data set, the accuracy of the model will be improved.Originality/valueThis paper proposes a new feature extraction method and analysis mode, which determines the shape of the hidden state according to the situation of the data set, making subsequent HMM modeling simple and efficient and in turn improving the accuracy of user behavior detection.

Significance of spectral cues in automatic speech segmentation for Indian language speech synthesizers

Building speech synthesis systems for Indian languages is challenging owing to the fact that digital resources for these languages are hardly available. Vocabulary independent speech synthesis requires that a given text is split at the level of the smallest sound unit, namely, phone. The waveforms or models of phones are concatenated to produce speech. The waveforms corresponding to that of the phones are obtained manual (listening and marking) when digital resources are scarce. But the manual labeling of speech data (also known as speech segmentation) can lead to inconsistencies as the duration of phones can be as short as 10ms.The most common approach to automatic segmentation of speech is to perform forced alignment using monophone hidden Markov models (HMMs) that have been obtained using embedded re-estimation after flat start initialization. These results are then used in neural network frameworks to build better acoustic models for speech synthesis/recognition. Segmentation using this approach requires large amounts of data and does not work very well for low resource languages. To address the issue of paucity of data, signal processing cues like short-term energy (STE) and sub-band spectral flux (SBSF) are used in tandem with HMM based forced alignment for automatic speech segmentation.STE and SBSF are computed on the speech waveforms. STE yields syllable boundaries, while SBSF provides locations of significant change in spectral flux that are indicative of fricatives, affricates, and nasals. STE and SBSF cannot be used directly to segment an utterance. Minimum phase group delay based smoothing is performed to preserve these landmarks, while at the same time reducing the local fluctuations. The boundaries obtained with HMMs are corrected at the syllable level, wherever it is known that the syllable boundaries are correct. Embedded re-estimation of monophone HMM models is again performed using the corrected alignment. Thus, using signal processing cues and HMM re-estimation in tandem, robust monophone HMM models are built. These models are then used in Gaussian mixture model (GMM), deep neural network (DNN) and convolutional neural network (CNN) frameworks to obtain state-level frame posteriors. The boundaries are again iteratively corrected and re-estimated.Text-to-speech (TTS) systems are built for different Indian languages using phone alignments obtained with and without the use of signal processing based boundary corrections. Unit selection based and statistical parametric based TTS systems are built. The result of the listening tests showed a significant improvement in the quality of synthesis with the use of signal processing based boundary correction.

Estimating required ‘lockdown’ cycles before immunity to SARS-CoV-2: model-based analyses of susceptible population sizes, ‘S0’, in seven European countries, including the UK and Ireland

Background: Following stringent social distancing measures, some European countries are beginning to report a slowed or negative rate of growth of daily case numbers testing positive for the novel coronavirus. The notion that the first wave of infection is close to its peak begs the question of whether future peaks or ‘second waves’ are likely. We sought to determine the current size of the effective (i.e. susceptible) population for seven European countries—to estimate immunity levels following this first wave. Methods: We used Bayesian model inversion to estimate epidemic parameters from the reported case and death rates from seven countries using data from late January 2020 to April 5th 2020. Two distinct generative model types were employed: first a continuous time dynamical-systems implementation of a Susceptible-Exposed-Infectious-Recovered (SEIR) model, and second a partially observable Markov Decision Process or hidden Markov model (HMM) implementation of an SEIR model. Both models parameterise the size of the initial susceptible population (‘S0’), as well as epidemic parameters. Results: Both models recapitulated the dynamics of transmissions and disease as given by case and death rates. Crucially, maximum a posteriori estimates of S0 for each country indicated effective population sizes of below 20% (of total population size), under both the continuous time and HMM models. Using a Bayesian weighted average across all seven countries and both models, we estimated that 6.4% of the total population would be immune. From the two models, the maximum percentage of the effective population was estimated at 19.6% of the total population for the UK, 16.7% for Ireland, 11.4% for Italy, 12.8% for Spain, 18.8% for France, 4.7% for Germany and 12.9% for Switzerland. Conclusion: Our results indicate that after the current wave, a large proportion of the total population will remain without immunity.

Comparison of HMM and RNN models for network traffic modeling

Major applications for statistical modeling of network traffic flows can be found in network testing and imitating of unavailable devices. Since packet‐level modeling is considered, packet size (PS) and inter‐arrival time (IAT) features are sufficient for accurate statistics. Two models are compared based on the hidden Markov model (HMM) framework and a recurrent neural network (RNN). In the RNN model, the feature space is encoded with latent components of a Gaussian mixture model (GMM). The comparison is carried out with a voice Skype call and traffic of an IoT device, and evaluated with the rolling entropy and Kulback‐Leibler divergence (KLD) metrics that are derived from the generated PS and IAT parameters. The results show that the RNN is applicable for the packet‐level modeling task, but it underperforms the HMM.

Confusion analysis in phoneme based speech recognition in Hindi

Phoneme recognition is an essential step in the development of a speech recognition system (SRS), as phonemes are fundamental building blocks in a spoken language. This research work aimed to present phoneme recognition with systematic confusion analysis for the Hindi language. The accuracy of phoneme recognition is the foundation for developing an efficient SRS. Therefore, the systematic confusion analysis for phoneme recognition is essential to improve speech recognition performance. Experiments conducted on Continuous Hindi speech corpus for phoneme recognition with speaker-dependent mode using Hidden Markov Model (HMM) based tool kit HTK. Feature extraction technique Perceptual Linear Predictive Coefficient (PLP) was used with five states Monophones HMM model. Tests were performed for exploring the recognition of Hindi vowels and consonants. Confusion matrices were presented for both vowels and consonants with analysis and possible solutions. During systematic analysis, the vowels were divided into front, middle, and back vowels while consonants were categorized based on place of articulation and manner of articulation. Research findings show that some Hindi phonemes have significant effects on speech recognition. The investigations also reveal that some Hindi phonemes are mostly confused, and some phonemes have more deletions and insertions. The research further demonstrates that the words made of less number of phonemes show more insertion errors. It was also found that most of the Hindi sentences end with some specific words. These particular words can be used to reduce the search place in language modeling for improving speech recognition. The research findings can be utilized to enhance the performance of the speech recognition system by selecting suitable feature extraction techniques and classification techniques for phonemes. The outcome of the research can also be used to develop improved pronunciation dictionaries and designing the text for developing phonetically balanced speech corpus for improvement in speech recognition. Experimental results show an average corrected recognition score of 70% for vowel class and consonant categories, the maximum average corrected recognition score of 94% was obtained with palatal sounds, and the lowest average corrected recognition score of 54% was achieved with liquid sounds. The comparative analysis of the presented work was made to similar existing works.

Grinding wheel condition prediction with discrete hidden Markov model using acoustic emission signature

Tool Condition Monitoring (TCM) is a vital activity to monitor and maintain the quality of products manufactured in any machining process without any manual intervention. Hidden Markov Models (HMM) are developed in this study for predicting grinding wheel conditions in the auto-feed surface grinding process using Acoustic Emission (AE) signatures captured during the grinding operation. A correlation between AE features and tool conditions are established using Baum-Welch and Viterbi algorithms. HMMs proposed in this study are integrated with the K-means clustering algorithm. The clustered data are represented as an integer sequence and are divided into three tool states such as ‘sharp’, ‘intermediate’ and ‘worn out’. HMM, models are created for each state of the grinding wheel. The performances of the HMMs are evaluated using log-likelihood measure. It is observed that HMM models build with AE–RMS feature are predicting the grinding wheel conditions with good accuracy.

Identification of AIDS-Associated Kaposi Sarcoma: A Functional Genomics Approach.

BackgroundKaposi sarcoma-associated herpes virus (KSHV) is one of the most common causal agents of Kaposi Sarcoma (KS) in individuals with HIV-infections. The virus has gained attention over the past few decades due to its remarkable pathogenic mechanisms. A group of genes, ORF71, ORF72, and ORF73, are expressed as polycistronic mRNAs and the functions of ORF71 and ORF72 in KSHV are already reported in the literature. However, the function of ORF73 has remained a mystery. The aim of this study is to conduct comprehensive exploratory experiments to clarify the role of ORF73 in KSHV pathology and discover markers of AIDS-associated KSHV-induced KS by bioinformatic approaches.Methods and ResultsWe searched for homologues of ORF-73 and attempted to predict protein-protein interactions (PPI) based on GeneCards and UniProtKB, utilizing Position-Specific Iterated BLAST (PSI-BLAST). We applied Gene Ontology (GO) and KEGG pathway analyses to identify highly conserved regions between ORF-73 and p53to help us identify potential markers with predominant hits and interactions in the KEGG pathway associated with host apoptosis and cell arrest. The protein p53 is selected because it is an important tumor suppressor antigen. To identify the potential roles of the candidate markers at the molecular level, we used PSIPRED keeping the conserved domains as the major parameters to predict secondary structures. We based the FUGE interpretation consolidations of the sequence-structure comparisons on distance homology, where the score for the amino acids matching the insertion/deletion (indels) detected were based on structures compared to the FUGE database of structural profiles. We also calculated the compatibility scores of sequence alignments accordingly. Based on the PSI-BLAST homologues, we checked the disordered structures predicted using PSI-Pred and DISO-Pred for developing a hidden Markov model (HMM). We further applied these HMMs models based on the alignment of constructed 3D models between the known structure and the HMM of our sequence. Moreover, stable homology and structurally conserved domains confirmed that ORF-73 maybe an important prognostic marker for AIDS-associated KS.ConclusionCollectively, similar variants of ORF-73 markers involved in the immune response may interact with targeted host proteins as predicted by our computational analysis. This work also suggests the existence of potential conformational changes that need to be further explored to help elucidate the role of immune signaling during KS towards the development of therapeutic applications.

The First Azeri (Azerbaijani) Language Next Word Predictor

Azeri (Azerbaijani) language is one of the more than 50 Turkic languages which it is a little studied language in terms of using the modern signal processing algorithms. This paper tackles the problem of Hidden Markov Models (HMMs) based next word prediction for this language based on Natural Language Processing (NLP) principles using Python high-level programming language. The software is included a small Azeri vocabulary database, the various Python libraries, a HMM model and a Web based interface. In this research, the database was constructed by a predictor parser which it was implemented for the first time for Azeri language. The database was concluded by the most general Azeri language words to introduce HMMs based generated word pairs. The Model was trained by 90% of the database, hence, predicting the next 5 words on the test data resulted 54% accuracy.

Hidden Markov model based rotate vector reducer fault detection using acoustic emissions

This paper proposes a hidden Markov model (HMM) based RV reducer fault detection using acoustic emission (AE) measurements. Compared with the conventional faults from the common rotating machinery (such as bearings and gears), faults from RV reducer are more complicated and undetectable due to its inherent inline and two-stage meshing structure. To this end, this work modifies the HMM model by taking into account not only the current observations and previous states, but the subsequent series of observations within posteriori probability framework. Through this way, the random and unknown disturbance could be suppressed. Besides, HMM is also applied to separate AE signal bulks within one cycle that has 39 subcycles. The proposed method has been evaluated on our collected AE signal dataset from the RV reducer in the industrial robotic platform. The experimental results and analysis validate the effectiveness and accuracy of our RV reducer fault detection model.