1. Introduction The recent expansion of the electric vehicle market has coincided with an increase in the circulation of second-hand lithium-ion batteries. These batteries incorporate rare metals such as cobalt, making repurposing or reusing second-hand batteries a subject of consideration from a resource efficiency perspective. As battery degradation is accompanied by changes in internal impedance, electrochemical impedance spectroscopy (EIS) is a valuable non-destructive technique for assessing whether second-hand batteries are suitable for reuse or repurposing based on their internal impedance information. However, since internal impedance also varies with the state-of-charge (SoC), knowing changes due to degradation alone necessitates prior SoC adjustment procedures, such as charging used batteries to 100% SoC. This adjustment process is time-consuming and reduces the efficiency of inspecting batteries for potential reuse or repurposing. Therefore, this study aimed to develop a method that can instantly determine the reusability of second-hand batteries regardless of their SoC without requiring time-consuming pre-adjustment. 2. Experimental 2-1. Preparation of 300 training data In this study, ten normal (new) batteries, US18650FTC1 (LFP/Graphite), were prepared and adjusted sequentially to ten levels ranging from 100% to 10% SoC by decrements of 10%. Then EIS measurements were performed three times on each level, and all Nyquist plots obtained were fitted into an equivalent circuit to obtain plural impedance parameters. By plotting these results (300 data points in total) within a space with dimensions equal to the number of parameters obtained by the fitting, we defined the area where there are normal batteries of any SoC. 2-2. Preparation of 280 test data Another set of twelve normal batteries was prepared; eight were used to obtain 240 data points using the same procedure as described above. The remaining four batteries were stored at temperatures of 60 °C, 70 °C, 80 °C, and 90 °C for 25 days to induce thermal degradation; then EIS measurements were performed once on each to obtain a total of 40 data points. 2-3. Calculation of degree of anomalies For each test datum among the 280 test data points, we calculated Mahalanobis distance from it to its k-nearest neighbors within the training dataset using k-nearest neighbor method; this distance was considered as a degree of anomalies. For each datum where the degree of anomalies exceeded a threshold (thus being classified as abnormal), we verified whether it was a thermally degraded battery (TP: true positive) or a new battery (FP: false positive). Similarly, for each datum where the degree of anomalies did not exceed the threshold (thus being classified as normal), we checked whether it was a new battery (TN: true negative) or a thermally degraded battery (FN: false negative). The numbers of TP, FP, TN, and FN were counted, then the proportion of data that were correctly identified as abnormal out of all data labeled as abnormal, termed Precision, was calculated as TP / (TP + FP). Additionally, the proportion of actual abnormal that were correctly identified, known as Recall, was calculated as TP / (TP + FN). Furthermore, the harmonic mean of Precision and Recall, known as the F1 score, was calculated. 3. Results and discussion The threshold of the degree of anomalies was determined based on the training data themselves; Threshold A was set at the degree of anomalies for which 99% of the training data were deemed normal, while Threshold B corresponded to the degree of anomalies classifying 95% as normal. The F1 scores obtained using thresholds A and B were 0.873 and 0.842, respectively, both exceeding 0.8. However, it should be noted that F1 scores are relative metric and without an external comparator, it is difficult to conclusively assess the merits of obtaining this high F1 score level. Nonetheless, surpassing the F1 score of 0.8 requires high Precision and Recall values, indicating that a well-balanced and robust classification has been achieved. 4. Conclusion In this study, we developed an anomaly detection method that utilizes a Mahalanobis distance from own data point, with its impedance parameters serving as coordinates, to a predefined normal area within a multidimensional space. By defining the normal area using new batteries of various SoC, the method attained F1 scores greater than 0.8 upon evaluation with batteries across a range of SoCs. This indicates that it is possible to distinguish between normal and abnormal batteries with high accuracy without the need for prior SoC adjustment. Applying this method to the diagnosis of used batteries could eliminate the time-consuming pre-adjustment of SoC previously required. This will be considered valuable in advancing battery reuse and repurposing efforts. Figure 1
Read full abstract
Nov 23, 2024
Muhammad Saqib + 6
Just Published
Open Access