Daniel V. Smith

Dynamic cattle behavioural classification using supervised ensemble classifiers

Cattle behavioural classification using cattle tag and supervised ensemble classifiers.Unsupervised hybrid clustering used to study inherent natural grouping in data set.Best classification accuracy was 96% using the bagging ensemble with Tree learner.A mechanism for the early detection and quantitative assessment of animal health. In this paper various supervised machine learning techniques were applied to classify cattle behaviour patterns recorded using collar systems with 3-axis accelerometer and magnetometer, fitted to individual dairy cows to infer their physical behaviours. Cattle collar data was collected at the Tasmanian Institute of Agriculture (TIA) Dairy Research Facility in Tasmania. In the first stage of analysis a novel hybrid unsupervised clustering framework, comprised of probabilistic principal component analysis, Fuzzy C Means, and Self Organizing Map network algorithms was developed and used to study the natural structure of the sensor data. Findings from this unsupervised clustering were used to guide the next stage of supervised machine learning. Five major behaviour classes, namely, Grazing, Ruminating, Resting, Walking, and other behaviour were identified for the classification trials. An ensemble of classifiers approach was used to learn models of cow behaviour using sensor data and ground truth behaviour observations acquired from the field. Ensemble classification using bagging, Random Subspace and AdaBoost methods along with conventional supervised classification methods, namely, Binary Tree, Linear Discriminant Analysis classifier, Naive Bayes classifier, k-Nearest Neighbour classifier, and Adaptive Neuro Fuzzy Inference System classifier were compared. The highest average correct classification accuracy of 96% was achieved using the bagging ensemble classification with Tree learner, which had 97% sensitivity, 89% specificity, 89% F1 score and 9% false discovery rate. This study has shown that cattle behaviours can be classified with a high accuracy using supervised machine learning technique. As dairy and beef systems become more intensive, the ability to identify the changes in the behaviours of individual livestock becomes increasingly difficult. Accurate behavioural monitoring through sensors provides a significant potential in providing a mechanism for the early detection and quantitative assessment of animal health issues such a lameness, informing key management events such as the identification of oestrus, or informing changes in supplementary feeding requirements.

A Bayesian Framework for the Automated Online Assessment of Sensor Data Quality

Online automated quality assessment is critical to determine a sensors fitness for purpose in real-time applications. A Dynamic Bayesian Network (DBN) framework is proposed to produce probabilistic quality assessments and represent the uncertainty of sequentially correlated sensor readings. This is a novel framework to represent the causes, quality state and observed effects of individual sensor errors without imposing any constraints upon the physical deployment or measured phenomenon. It represents the casual relationship between quality tests and combines them in a way to generate uncertainty estimates of samples. The DBN was implemented for a particular marine deployment of temperature and conductivity sensors in Hobart, Australia. The DBN was shown to offer a substantial average improvement (34%) in replicating the error bars that were generated by experts when compared to a fuzzy logic approach.

A Novel Machine Learning Approach Toward Quality Assessment of Sensor Data

A novel machine learning approach to assess the quality of sensor data using an ensemble classification framework is presented in this paper. The quality of sensor data is indicated by discrete quality flags that indicate the level of uncertainty associated with a sensor reading. Depending on the domain and the problem under consideration, the level of uncertainty is different and thus unsupervised methods like outlier detection fails to match the expectation. The quality flags are normally assigned by domain experts. Considering the volume of sensor data, manual assignment is a laborious task and subject to human error. Given a representative set of labelled data, a supervised classification approach is thus a feasible alternative. The nature of sensor data, however, poses some challenges to the classification task. Data of dubious quality exists in such data sets with very small frequency leading to the class imbalance problem. We thus adopt a cluster oriented sampling approach to address the imbalance issue. In addition, it is beneficial to train multiple classifiers to improve the overall classification accuracy. We thus produce multiple under-sampled training sets using cluster oriented sampling and train base classifiers on each of them. Decisions produced by the base classifiers are fused into a single decision using majority voting. We have evaluated the proposed ensemble classification framework by assessing the quality of marine sensor data obtained from sensors situated at Sullivans Cove, Hobart, Australia. Experimental results reveal that the proposed framework agrees with expert judgement with high accuracy and achieves superior classification performance than other state-of-the-art approaches.

Automated Data Quality Assessment of Marine Sensors

The automated collection of data (e.g., through sensor networks) has led to a massive increase in the quantity of environmental and other data available. The sheer quantity of data and growing need for real-time ingestion of sensor data (e.g., alerts and forecasts from physical models) means that automated Quality Assurance/Quality Control (QA/QC) is necessary to ensure that the data collected is fit for purpose. Current automated QA/QC approaches provide assessments based upon hard classifications of the gathered data; often as a binary decision of good or bad data that fails to quantify our confidence in the data for use in different applications. We propose a novel framework for automated data quality assessments that uses Fuzzy Logic to provide a continuous scale of data quality. This continuous quality scale is then used to compute error bars upon the data, which quantify the data uncertainty and provide a more meaningful measure of the data’s fitness for purpose in a particular application compared with hard quality classifications. The design principles of the framework are presented and enable both data statistics and expert knowledge to be incorporated into the uncertainty assessment. We have implemented and tested the framework upon a real time platform of temperature and conductivity sensors that have been deployed to monitor the Derwent Estuary in Hobart, Australia. Results indicate that the error bars generated from the Fuzzy QA/QC implementation are in good agreement with the error bars manually encoded by a domain expert.

Hydrological Sensor Web for the South Esk Catchment in the Tasmanian state of Australia

The sensor Web is a distributed sensing system in which information is shared globally. The emergence of this technology will enable the integration of different sensing platforms with temporal and spatial variability. This has a potential to revolutionise hydrological monitoring and forecasting. Our project will establish a sensor Web test bed in the South Esk river catchment, which is located in the North East of Tasmania. The test bed will allow us to evaluate the emerging open geospatial consortium standards and specifications for sensor Web enablement (SWE) and provide a research platform for developing next-generation hydrological and water resource management tools. We intend to use short-term river flow forecasting as our use case for the SWE test bed.

Bag of Class Posteriors, a new multivariate time series classifier applied to animal behaviour identification

A new multi-scale time series classifier is proposed using class posterior estimates.The classifier infers a large set of animal behaviour using motion based time series.The proposed classifier outperforms benchmark classifiers by between 43% and 77%.The proposed classifier is found to be more efficient than the Bag of Features model. In this paper, two new multivariate time series classifiers are introduced as the Bag of Class Posteriors (BOCP) and the Bag of Class Posterior with Ordering (BOCPO). The models propose a new multi-scale feature representation where the class posterior estimates of contiguous local patterns are aggregated over longer time scales. The models are employed as part of an animal behaviour monitoring system that are comprised of sensors, which are fitted to the animals, and a classifier that translates sensor data into knowledge of the animals behaviour.Animal monitoring systems are commonly developed to infer a small number of behaviours with relevance to a specific application. To investigate if a standard monitoring system with an Inertial Measurement Unit (IMU) can be reused for different management applications, a set of ten cattle behaviours relevant to different management applications were classified with the proposed models. Results indicate that the multi-scale BOCP and BOCPO models were far more capable of classifying the cow behaviours offering a 43% to 77% improvement over benchmark time interval classifiers with fixed time resolution. In addition, the BOCPO model was shown to offer a far more efficient feature representation than the related multi-scale Bag of Features (BOF) classifier (up to 200 times smaller) making it better suited to deploy upon monitoring devices fitted to animals in the field.

Multiple classifier system for automated quality assessment of marine sensor data

Numerous sources of uncertainty are associated with the data acquisition process in marine sensor networks. It is thus required to assure that the data quality of sensors is fit for the intended purpose. We propose a supervised learning framework to infer the quality of sensor observations online. A problem with using supervised classification in quality assessment is that sensor observations from the class of uncertain data will be far out-weighed by class instances of good data quality. This leads to an imbalanced data set, which can potentially reduce the classification accuracy of uncertain data. A multiple classifier (or ensemble classifier) system is proposed to deal with this problem. Training sets are randomly undersampled to develop training subsets with balanced class membership. The process is repeated to produce multiple balanced training subsets. Individual classifiers are then trained upon each of these balanced data sets. The quality classifications from the individual classifiers are then combined using majority voting. We evaluated the ensemble classifier system using conductivity and temperature sensors from the Tasmanian Marine Analysis Network (TasMAN). Experiments demonstrate that the ensemble classifier balances the classification accuracy of the majority and minority classes, achieving a higher overall classification accuracy than its constituent classifiers.

Detecting heat events in dairy cows using accelerometers and unsupervised learning

We developed a heat detection algorithm for pasture-based dairy cows.Our algorithm uses data from accelerometer attached to the cow collars.We present the overall accuracy of 82-100% with 100% sensitivity. This study was conducted to investigate the detection of heat events in pasture-based dairy cows fitted with on-animal sensors using unsupervised learning. Accelerometer data from the cow collars were used to identify increased activity levels in cows associated with recorded heat events. Time series data from the accelerometers were first segmented into windows before features were extracted. K-means clustering algorithm was then applied across the windows for grouping. The groups were labelled in terms of their activity intensity: high, medium and low. An activity index level (AIxL) was then derived from a count of activity intensity labels over time. Change detection techniques were then applied on AIxL to find very high activity events. Detected events in AIxL were compared with recorded heat events and observed significant associations between the increased activities through high AIxL values and the observed heat events. We achieved overall accuracy of 82-100% with 100% sensitivity when change detection technique is applied to activity index level.

Behavior classification of cows fitted with motion collars: Decomposing multi-class classification into a set of binary problems

Precision management systems for livestock offer the potential to monitor and manage animals on an individual basis. A key component of these sensor based systems are the analytical models that automatically translate sensor data into different behavioral categories. A new methodology was proposed for multi-class behavior modeling based upon the “one-vs-all” framework. This methodology differs from the standard approach to behavior classification where a single classifier is trained to discriminate between multiple behaviors. Instead a set of binary classifiers are trained to each discriminate one of the behavior classes against a combined class of all the remaining behaviors. The confidence scores from the set of binary classifiers are then combined to generate a behavioral estimate. The performance of this new modeling approach is validated across a study involving 24 Holstein-Friesian dairy cows that were each fitted with an Inertial Measurement Unit (IMU) on a collar upon their neck. Five general classes of cow behavior grazing, walking, ruminating, resting and “Other” were classified. Binary time series classifiers were tailored to each of the five behaviors through training and validation of each model across a range of configurations including nine window sizes, five classifiers and a feature selection process of 84 candidate input features. Results revealed that the proposed model classified grazing behavior with an extremely high classification accuracy (F-score of 0.98), whilst ruminating and resting behaviors were also classified with a high accuracy (F-scores of 0.87 and 0.85, respectively), and walking was classified with a lower accuracy (F-score of 0.73). The proposed model offered a 5% performance improvement over standard multi-class time series classifiers that was attributed to the “one-vs-all” framework training a classifier for each behavior independently; this created diversity in the behavior model. The feature selection process used in developing each of the binary classifiers found that features representing the motion intensity and pitch of the cow’s head were most important to each of behavior’s classification. Whilst a minor performance improvement was obtained using the proposed methodology, it is suggested that further performance improvements could be obtained by increasing the diversity of the classifier’s inputs. Diversity could be created by fusing the data of other sensors that can be fitted to the cows i.e. GPS tracking unit, pressure sensor and microphone.

Dynamic annotation and visualisation of the South Esk hydrological sensor web

This research study focused on automatic sensor data annotation and visualisation of dynamic weather data acquired from a large sensor network. The aim was to develop a data visualisation method for CSIROs South Esk hydrological sensor web to evaluate the overall network performance and provide visual data quality assessment. The visual data quality technique developed in this study could be applied to quality assurance of any sensor network.