Dual-scale Spatiotemporal Learning Based Multiscale Detection for BMS under Edge Sensor Network

Project: Research

View graph of relations



Carbon neutral policies require clean energy, in which battery systems will play a major role. The battery should operate in proper temperature range for optimum performance and longevity. The battery thermal process belongs to a complexdistributed parameter system(DPS) described in the nonlinear partial differential equation. This type of process isdifficult to model, especially for sensor-less online operations, let alone theabnormal detectionand localization.PassiveRFID tags are widely used in production to collect information about the location and movement of products due to their low cost and no power supply, but are usually not used after production. It is a waste and does not fit the carbon neutral philosophy. This proposal aims to study thedual-scale iterative learning based optimal DPS modellingmethod andmultiscale abnormal diagnosisfor the battery system underedge sensor network. • First, the passive RFID tags attached to battery cells for production will be designed to have a“second life”for temperature sensing, which actually leads to a new functional“edge sensor”for additional information measurement. All these tags on battery cells can then form a low-cost “edge sensor” based network that continues to serve while the battery system is running. This will help improve the spatial measurement of temperature distribution inside the battery system, while also meeting the carbon neutrality requirements - waste utilization for edge use. • Second, dual-scale iterative learning will be studied to obtainoptimal space/time separation,and base on this, an optimal spatiotemporal model will be developed for online performance prediction of DPS. This spatiotemporal learning includes slow-scale spatial adaptation and fast-scale temporal learning. Thisiterative learningbetweenfast/slow scalesis a long-standing challenge in allmulti-scaleprocesses in manufacturing. • Third,sliding windowbased multiscale information will be studied for abnormal detection and localization.Multiscalestatistic assessing approaches will be designed, one is thetemporal dissimilarityof the subspace for early detection, and the other is thespatiotemporal entropyfor global scale evaluation. Combining assessments at these different scales will lead to a more reliable diagnosis. By analyzing thespatial entropyof normal and abnormal conditions, the location of the abnormality can be determined with the greatest probability. The results of the proposed research, if successful, will produce pioneering work in smart sensing, online spatiotemporal learning and abnormal diagnosis in battery systems. The successful implementation can be applied to existing battery management systems to improve its operational safety. 


Project number9043524
Grant typeGRF
Effective start/end date1/09/23 → …