@techreport{Agrawal2020,
    Author = "Agrawal, Vipin K and Maurya, Shiv Shankar",
    title = "Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "Autoencoders are a very popular approach in detecting anomalies in a system, where reconstruction error is generally used as an anomaly score. However, the reconstruction errors, generated in such manners, contain external noises of the system, making reconstruction errors as anomaly scores less effective. In this brief, we present an additional hypothesis that autoencoders may introduce additional statistical noise in the reconstruction errors as well. Our proposal includes a design of an autoencoder, lays out a theoretical basis of designing a noise filter for reconstruction errors, and outlines various aggregation methods to reduce the effect of the noise. While further work is still needed, we are able to show the accuracy improvement by using various aggregation methods ."
}

@techreport{Ahmed2020,
    Author = "Ahmed, Faruk and Nguyen, Phong and Courville, Aaron",
    title = "An ensemble approach for detecting machine failure from sound",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "We develop an ensemble-based approach for our submission to the anomaly detection challenge at DCASE 2020. The main members of our ensemble are auto-encoders (with reconstruction error as the signal), classifiers (with negative predictive confidence as the signal), mismatch of the time-shifted signal with its Fourier-phase-shifted version, and a Gaussian mixture model on a set of common short-term features extracted from the waveform. The scores are passed through an exponential non-linearity and weighted to provide the final score, where the weighting and scaling hyper-parameters are learned on the development set. Our ensemble improves over the baseline on the development set."
}

@techreport{Alam2020,
    Author = "Alam, Jahangir and Boulianne, Gilles and Gupta, Vishwa and Fathan, Abderrahim",
    title = "An Ensemble Approach to Unsupervised Anomalous Sound Detection",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "The task of anomalous sound detection (ASD) is to determine whether an observed sound is anomalous or normal. Both supervised and unsupervised approach can be adopted for the ASD task. In supervised approach anomalous and normal data are used in training whereas in unsupervised approach only normal data are used for training. In this work, we provide an overview of the systems developed for the task 2 i.e., unsupervised detection of anomalous sounds for machine condition monitoring, of the DCASE 2020 challenge. We employ various handcrafted local representations from the short-time spectral analysis of sounds. We also use fisher vector encoding -a learned global representations obtained from local representations of sound. Autoencoder variants and copy detection approaches are applied on the top of local representations and a standard GMM classifier is used with fisher vector encodings for unsupervised detection of anomalous sounds."
}

@techreport{Bai2020,
    Author = "Bai, Jisheng and Chen, Chen and Chen, Jianfeng",
    title = "Bai\_LFXS\_NWPU\_dcase2020\_submission",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "task 2 and task 5"
}

@techreport{Chaudhary2020,
    Author = "Chaudhary, Nitesh Kumar and Mathew, Josey and Sivadas, Sunil",
    title = "CONVOLUTIONAL AUTO ENCODER FOR MACHINE CONDITION MONITORING FROM ACOUSTIC SIGNATURES",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "Condition monitoring of machinery is critical for early detection and prevention of failures in factories. Recent advancements in machine learning is driving the development of data driven tools like monitoring acoustic signatures from microphones. This report presents a convolutional autoencoder (CAE) trained to minimize the acoustic spectrogram reconstruction error during normal operation of the machine. The model is pre-trained on machines of similar type and then fine-tuned on a specific machine. The reconstruction error is used as the anomaly score for an unseen acoustic sample. The proposed model improves performance compared to baseline system for the DCASE2020 challenge task2"
}

@techreport{Chen2020,
    Author = "Chen, Yunqi and Song, Yuheng and Cheng, Ting",
    title = "Anomalous Sounds Detection Using A New Type of Autoencoder based on Residual Connection",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "This report describes our submission for task2 (Unsupervised Detection of Anomalous Sounds for Machine Condition Mon-itoring) of the DCASE 2020. In this report we propose net-works of fully connected autoencoder based on residual con-nections, which can increase the accuracy of anomaly sound detection. As for data preprocessing, we use data augmenta-tion methods to generate more data from existing data. Our feature extraction is still carried out with log mel spectrogram. Finally, our method has achieved average AUC of 0.7912 and average pAUC of 0.6105 on the development dataset."
}

@techreport{Daniluk2020,
    Author = "Daniluk, Pawel and Gozdziewski, Marcin and Kapka, Slawomir and Kosmider, Michal",
    title = "ENSEMBLE OF AUTO-ENCODER BASED SYSTEMS FOR ANOMALY DETECTION",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "In this paper we report an ensemble of models used to perform anomaly detections for DCASE Challenge 2020 Task 2. Our solu- tion comprises three families of models: Variational Heteroskedas- tic Auto-encoders, Conditioned Auto-encoders and a WaveNet- based network. Noisy recordings are preprocessed using a U-Net trained for noise removal on training samples augmented with noised obtained from the AudioSet. Models operate either on OpenL3 embeddings or log-mel power spectra. Heteroskedastic VAEs have a non-standard loss function which uses model's own error estimation to weigh typical MSE loss. Model ar- chitecture i.e. sizes of layers, dimension of latent space and size of an error estimating network are independently selected for each machine type. ID Conditined AEs are an adaptation of the class conditioned auto- encoder approach designed for open set recognition. Assuming that non-anomalous samples constitute distinct IDs, we apply the class conditioned auto-encoder with machine IDs as labels. Our approach omits the classification subtask and reduces the learning process to a single run. We simplify the learning process further by fixing a target for non-matching labels. Anomalies are predicted either by poor reconstruction or attibution of samples to the wrong machine ID. The third solution is based on a convolutional neural network and a simple noise reduction method. The architecture of the model is inspired by the WaveNet and uses causal convolutional layers with growing dilation rates. It works by predicting the next frame in the spectrogram of a given recording. Anomaly score is derived from the reconstruction error. We present results obtained by each kind of models separately, as well as, a result of an ensemble obtained by averaging anomaly scores computed by individual models."
}

@techreport{Durkota2020,
    Author = "Durkota, Karel and Michael, Linda and Martin, Ludvik and Jan, Tozicka",
    title = "NEURON-NET: SIAMESE NETWORK FOR ANOMALY DETECTION",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = {This paper describes our submission to the DCASE 2020 challenge Task 2 "Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring." Acoustic-based machine condition monitoring is a challenging task with a very unbalanced training dataset. In this submission, we combine the Siamese Network feature extractor with KNN anomaly detection algorithm. Experiment results prove it to be a viable approach with an average AUC 85.85 and pAUC of 77.1. This novel approach have not been used by NeuronSW SE so far.}
}

@techreport{Giri2020,
    Author = "Giri, Ritwik and Tenneti, Srikanth V. and Helwani, Karim and Cheng, Fangzhou and Isik, Umut and Krishnaswamy, Arvindh",
    title = "Unsupervised Anomalous Sound Detection Using Self-Supervised Classification and Group Masked Autoencoder for Density Estimation",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "This technical report outlines our solutions to Task 2 of the DCASE 2020 challenge, \emph{Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring}. The objective is to detect audio recordings containing anomalous machine sounds in a test set, when the training dataset itself does not contain any examples of anomalies. Our approaches are based on an ensemble of a novel density estimation based anomaly detector (Group Masked Autoencoder for Density Estimation (GMADE)) and self-supervised classification based anomaly detector."
}

@techreport{Grollmisch2020,
    Author = "Grollmisch, Sascha and Johnson, David and AbeBer, Jakob and Lukashevich, Hanna",
    title = "IAEO3 - Combining OpenL3 Embeddings and Interpolation Autoencoder for Anomalous Sound Detection",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "In this technical report, we present our system for task 2 of the IEEE AASP Challenge on Detection and Classification of Acoustic Scenes and Events (DCASE2020 Challenge): Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring. The focus of this task is to detect anomalous industrial machine sounds using an acoustic quality control system, which is only trained with sound samples from the normal (machine) condition. The dataset covers a variety of machines ranging from stable sound sources such as car engines, to transient sounds such as opening and closing valves. Our proposed method combines pre-trained OpenL3 embeddings with the reconstruction error of an interpolation autoencoder using a gaussian mixture model as the final predictor. The optimized model achieved 88.5\% AUC and 76.8\% pAUC on average over all machines and types provided with the development dataset, and outperformed the published baseline by 14.9\% AUC and 17.2\% pAUC."
}

@techreport{Haunschmid2020,
    Author = "Haunschmid, Verena and Praher, Patrick",
    title = "Anomalous Sound Detection with Masked Autoregressive Flows and Machine Type Dependent Postprocessing",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "This technical report describes the submission from the CP JKU/SCCH team for Task 2 of the DCASE2020 challenge - Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring. Our approach uses a Masked Autoregressive Flow (MAF) model for density estimation trained solely on normal samples. Anomaly scores per input snippet are computed using the negative log likelihood of new samples. The anomaly scores per input audio are summarised using different metrics depending on the machine type instead of simply averaging them."
}

@techreport{Hayashi2020,
    Author = "Hayashi, Tomoki and Yoshimura, Takenori and Adachi, Yusuke",
    title = "CONFORMER-BASED ID-AWARE AUTOENCODER FOR UNSUPERVISED ANOMALOUS SOUND DETECTION",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "This paper presents an autoencoder-based unsupervised anomalous sound detection (ASD) method for the DCASE 2020 Challenge Task 2. Inspired by the great successes of the self-attention architecture in various fields such as speech recognition, we propose Transformer- and Conformer-based autoencoder for ASD, enabling us to perform sequence-by-sequence processing. As opposed to the standard autoencoder, they can extract sequence-level information from whole audio inputs. Furthermore, we propose two simple methods for exploiting machine ID information: machine ID embedding and machine ID regression. The two methods enable the proposed models to avoid the confusion of anomalous and normal sounds among the different machine IDs. The experimental evaluation demonstrates that the proposed autoencoders outperform the conventional frame-level autoencoder, and the explicit use of machine ID information significantly improves the ASD performance. We achieved an averaged area under the curve (AUC) of 91.33\% and averaged partial AUC of 83.34\% on the development set."
}

@techreport{Hoang2020,
    Author = "Hoang, Truong and Nguyen, Hieu and Pham, Giao",
    title = "Unsupervised Detection Of Anomalous Sound For Machine Condition Monitoring Using Different Auto-encoder Methods",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "Anomaly detection from the sound of machines is an important task for monitoring machines. This paper presents four auto-encoder methods to detect anomalous sound for machine condition monitoring using Long-short term memory auto-encoder, U-Net auto-encoder, Interpolation deep neural network, and Fully-connected auto-encoder. With experiments on the same dataset with the baseline system, experimental results show that our methods out-perform the baseline system in terms of AUC and pAUC evaluation metrics."
}

@techreport{Jalali2020,
    Author = "Jalali, Anahid and Schindler, Alexander and Bernhard, Haslhofer",
    title = "DCASE Challenge 2020: Unsupervised Anomalous Sound Detection of Machinery with Deep Autoencoders",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "Inourwork,wepresentanunsupervisedanomaloussounddetection framework trained on DCASE2020 audio dataset. This dataset is a subset of two datasets ToyADMOS and MIMII. We use the state of the art anomaly detection approach, deep autoencoder architecture trained on Mel-Spectrograms. This architecture uses LSTM-RNN units to learn the normal condition of the machine, and is proven efficientatdetectingdiversemachineanomalies. Ourtrainedmodel on MIMII dataset achieves average result of 73.51\% AUC and 57.90\% pAUC, resulting in a slight improvement compared to the baseline system with the average results of 72.44\% AUC and 57.48 pAUC. The average performance of the baseline system on ToyADMOSdatasetis75.65\%AUCand64\%pAUC,whereourmodel reaches to average of 73.51\% AUC and 57.90\% pAUC. Our system reaches overall average of 73.41\% AUC and 59.27\% pAUC on the development data set, with overall similar performance to the baseline system with average of 73.51\% AUC and 59.66\% pAUC."
}

@techreport{Jiang2020,
    Author = "Jiang, Huitian and Lan, Bo and Li, Huiyong",
    title = "ABNORMAL SOUND DETECTION SYSTEM BASED ON AUTOENCODER",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "This report describes our submissions with an autoencoder (AE) to solve the DCASE 2020 challenge task 2 (unsupervised detec-tion of anomalous sounds for machine condition monitoring). Previous research results show that AE is a very effective solution to abnormal sound detection (ASD). This design continues previ-ous research, using AE to implement unsupervised ASD. To decrease the false positive rate (FPR), the AE is trained to mini-mize the reconstruction error of normal sound. In addition, the design uses variational autoencoder (VAE) to generate normal sound samples. The generated sound samples are used to enhance AE's ability to reconstruct normal sound samples."
}

@techreport{Kaltampanidis2020,
    Author = "Kaltampanidis, Yannis and Thoidis, Iordanis and Vrysis, Lazaros and Dimoulas, Charalampos",
    title = "Unsupervised Detection of Anomalous Sounds via ProtoPNet",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "Prototypical part network (ProtoPNet) is a novel method proposed for the task of image classification, offering the ability to interpret the network's reasoning process during classification. The subject of this work is the examination of ProtoPNet as an unsupervised anomaly detection method, through its application at the Detection and Classification of Acoustic Scenes and Events (DCASE) 2020 task 2 challenge. It is also showed that ProtoPNet shares common grounds with Deep One-Class Support Vector Data Descriptor (DOCSVDD)."
}

@techreport{Koizumi2020,
    Author = "Koizumi, Yuma and Kawaguchi, Yohei and Imoto, Keisuke",
    title = "Description and discussion on DCASE2020 challenge task2: unsupervised anomalous sound detection for machine condition monitoring",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "This paper presents the details of the DCASE 2020 Challenge Task 2; Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring. The goal of anomalous sound detection (ASD) is to identify whether the sound emitted from a target machine is normal or anomalous. The main challenge of this task is to detect unknown anomalous sounds under the condition that only normal sound samples have been provided as training data. We have designed a DCASE challenge task which contributes as a starting point and a benchmark of ASD research; the dataset, evaluation metrics, a simple baseline system, and other detailed rules. After the challenge submission deadline, challenge results and analysis of the submissions will be added."
}

@techreport{Lapin2020,
    Author = "Lapin, Dmitry and Khubbatulin, Mark and Klychnikov, Vladimir",
    title = "THE STUDY OF ANOMALOUS MACHINE SOUND DETECTION BASED ON CYCLOSTATIONARITY MODEL",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "In industrial predictive maintenance, one of the most important direction in Industry 4.0, machine monitoring and diagnostics is critical part of its operation. Non-contact acoustic data gathering is particular interest because of high ergonomics and low costs. The general method for processing such type of data is anomalous sound detection. This method allows express diagnostics of machines and units with mini-mum integrations. Based on DCASE 2020 Challenge, the study of the proposed method was presented. Problem description with physical interpretation and model elements review was conducted. Model based on Winger-Ville trans-form with architecture improvement and ensemble score calculation was developed. Model results on provided development dataset were calculated. Discussion of model results with assumptions for further research and development was shown. Conclusion about present study and future work was received."
}

@techreport{Lopez2020,
    Author = "Lopez, Jose and Hong, Lu and Lopez-Meyer, Paulo and Nachman, Lama and Stemmer, Georg and Huang, Jonathan",
    title = "A Speaker Recognition Approach To Anomaly Detection",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "We discuss our unsupervised speaker-recognition-based submission to the DCASE 2020 Challenge Task 2. We found that a speaker-recognition approach enables the use of all the training data, even from different machine types, to detect anomalies in specific machines. Using this approach, we obtained AUCs close to, or greater than, 0.9 for 5 out of 6 machines. We also discuss the modifications needed to surpass the baseline score for the ToyConveyor data."
}

@techreport{Morita2020,
    Author = "Morita, Kazuki and Yano, Tomohiko and Tran, Khai Q",
    title = "ANOMALOUS SOUND DETECTION BY USING LOCAL OUTLIER FACTOR AND GAUSSIAN MIXTURE MODEL",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "In this report, we introduce our methods and results of the anomalous sound detection in DCASE2020 task2. We attempted to detect anomalous sound without using deep learning methods. Precisely, we first extracted features by applying principal component analysis (PCA) to the log-mel spectrogram of the sound signal. Then we used Local Outlier Factor (LOF) and Gaussian Mixture Model (GMM) as the anomaly detection method. Our experiment showed the proposed method improved the Area Under Curve (AUC) to 0.8706 and the partial Area Under Curve(pAUC) to 0.7403 compared to the baseline system on development dataset."
}

@techreport{Naranjo-Alcazar2020,
    Author = "Naranjo-Alcazar, Javier and Perez-Castanos, Sergi and Zuccarello, Pedro and Cobos, Maximo",
    title = "TASK 2 DCASE 2020: ANOMALOUS SOUND DETECTION USING UNSUPERVISED AND SEMI-SUPERVISED AUTOENCODERS AND GAMMTONE AUDIO REPRESENTATION",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "Anomalous sound detection (ASD) is one of the fields of machine listening that is attracting most attention among the scientific community. Unsupervised detection is attracting a lot of interest due to its immediate applicability in many fields. For example, related to industrial processes, the early detection of malfunctions or damage in machines can mean great savings and an improvement in the efficiency of industrial processes. This problem can be solved with an unsupervised ASD solution since industrial machines will not be damaged simply by having this audio data in the training stage. This paper proposes a novel framework based on convolutional autoencoders (both unsupervised and semi-supervised) and a Gammatone-based representation of the audio. The results obtained by these architectures substantially exceed the results presented as a baseline."
}

@techreport{Park2020,
    Author = "Park, Jihwan and Sooyeon, Yoo and Yoo, Sooyeon",
    title = "Unsupervised detection of anomalous machine sound using various spectral features and focused hypothesis test in the reverberant and noisy environment",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "In this technical report, we describe our anomalous sound detection (ASD) systems submitted in DCASE 2020 Task2. To improve the ASD performance in the reverberant and noisy condition, normal machine sound augmentation, focused hypothesis test, and selecting the distinctive spectral features is applied to deep neural network (DNN)-based autoencoder (AE). In the experiments, we found that our approaches outperform baseline methods under the condition that only reverberant and noisy normal sound samples have been provided as training data."
}

@techreport{Phan2020,
    Author = "Phan, Duc and Jones, Douglas",
    title = "DCASE 2020 TASK 2: UNSUPERVISED DETECTION OF ANOMALOUS SOUNDS FOR MACHINE CONDITION MONITORING",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "A multiple layer neural network predictor is proposed for anomalous sound detection instead of a traditional auto-encoder approach. The network operates on the log-mel-spectrogram, predicting the log-mel feature vector given the previous and future feature vectors. The prediction error is used as the anomaly score measure. The proposed system outperforms the baseline system [1] on Detection and Classification of Acoustic Scenes and Events 2020 (DCASE2020) Task 2 development data set [2, 3]."
}

@techreport{Primus2020,
    Author = "Primus, Paul",
    title = "Reframing Unsupervised Machine Condition Monitoring as a Supervised Classification Task with Outlier-Exposed Classifiers",
    institution = "DCASE2020 Challenge",
    year = "2020",
    month = "July",
    abstract = "This technical report contains a detailed summary of our submissions to the Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring (MCM) Task of the IEEE AASP Challenge on Detection and Classification of Acoustic Scenes and Events 2020 (DCASE). The goal of acoustic MCM is to identify whether a sound emitted from a machine is normal or anomalous. In contrast to the task coordinator's claim that 'this task cannot be solved as a simple classification problem,' we show that a simple binary classifier substantially outperforms the provided unsupervised Autoencoder baseline across all machine types and instances, if outliers i.e., various other recordings, are available. In addition to this technical description, we release our complete source code to make our submission fully reproducible."
}