Quantitative Abnormality Diagnosis Method for Power Batteries Based on 1D CNN-LSTM

1. Research Background

The real-time voltage of each cell is one of the main parameters of the battery and is widely used in data-driven methods combined with signal processing technology. The evaluation of battery inconsistency is often based on comparisons between individual cells. Considering the inherent differences in the battery itself and the performance variations under different operating conditions, the accuracy of fault diagnosis algorithms is thus limited. Additionally, in practical applications, the complexity of vehicle operating conditions and the nonlinear relationships between battery electrical signals caused by electrochemical performance, along with the coupling relationships among multiple features, mean that only considering the individual voltage signals cannot adequately summarize the battery’s state, necessitating the expansion of relevant data items.

2. Research Content

1. Extraction and Preprocessing of Real Vehicle Operating Data: The discharge voltage of the battery represents its working state. Based on the usage time and aging degree of the power battery, all discharge data segments in the entire lifecycle data are divided.

Figure 1 Distribution curves of some features after preprocessing

2. Quantification of Cell Abnormalities Based on 1D CNN-LSTM: By quantifying cell abnormalities for power battery fault diagnosis, a real-time voltage estimation model based on 1D CNN-LSTM is established to simulate the voltage response of a healthy battery under current operating conditions, obtaining voltage reference values. By comparing these reference values with the actual sampled values of each cell, a scoring method is established to quantify the performance deviation of each cell relative to the healthy battery.

3. Model Verification and Case Analysis: This method was applied to cases of thermal runaway incidents in power batteries. The evaluation results show that this method has significant application effects for both types of thermal runaway incidents caused by cell abnormalities or overall deterioration.

1. This method can yield clearer quantification results for cell abnormalities. Compared to methods that rely on comparisons between individual batteries, this method uses the early healthy state as a reference to quantify the performance deviations of each battery relative to the same reference, thus obtaining more accurate assessments of cell abnormalities.

Figure 3 Example of real-time voltage estimation results based on 1D CNN-LSTM

2. This method can identify significant abnormalities in faulty cells compared to other cells up to 7 days in advance, and it can even detect potential risks in discharge segments occurring a year before an accident. For cases of overall deterioration without obvious cell inconsistencies, it can track the overall performance degradation process within 7 days prior to an incident.

4. Innovations and Significance

The paper proposes a power battery fault analysis method based on the quantification of cell abnormalities using 1D CNN-LSTM, establishing a fusion deep learning model that combines one-dimensional convolutional neural networks and long short-term memory networks (1D CNN-LSTM). This method can identify early potential risks and track the deterioration process, holding significant theoretical significance and engineering application value.