In recent years, the advancement of battery technology and the demands of smart mining have led to the widespread application of lithium iron phosphate batteries, commonly referred to as lifepo4 battery, in coal mine production. These batteries are used in various applications such as power sources for mining robots and electric explosion-proof vehicles. The lifepo4 battery offers high energy density, stable discharge, and long cycle life, providing reliable power for underground operations. However, during the cyclic use of the lifepo4 battery, irreversible internal chemical reactions cause capacity degradation. When the battery capacity decays to 80% of its initial capacity, the battery is considered end-of-life, as further use may lead to safety incidents. The mining environment is complex and special, and under such conditions, the lifespan of the lifepo4 battery can be significantly affected. Therefore, predicting the State of Health (SOH) of the lifepo4 battery is crucial for safe usage. In practical working conditions, due to the complex electrochemical reaction mechanisms within the battery and interference from external environmental factors, the SOH cannot be directly measured. Thus, accurate prediction models are needed to estimate the health state of the lifepo4 battery.
Machine learning has made significant progress in various fields, and data-driven methods for battery SOH prediction have gained considerable interest. Deep learning algorithms such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM) networks have been applied to SOH prediction. For instance, some studies have used CNN frameworks to predict SOH from voltage, current, and temperature measurements during charging, achieving average errors around 3.8%. Others have combined LSTM with particle swarm optimization and attention mechanisms, keeping errors within 5%. Gaussian Process Regression (GPR) models based on indirect health indicators have also been proposed, with maximum errors controlled within 6%. While these methods leverage data-driven models, most have limited ability to extract feature information, restricting prediction accuracy. For mining lifepo4 battery, the operating environment is more complex, with more external influencing factors, making it essential to capture valuable feature information for accurate SOH prediction.
To address this, we propose a parallel CNN-LSTM network model, referred to as CNN-LSTM_Parallel, which combines the complementary modeling capabilities of CNN and LSTM. This model utilizes CNN’s ability to mine hidden information from limited data and LSTM’s proficiency in processing time-series data, fusing them to extract more valuable information from raw data for efficient prediction. The lifepo4 battery is central to our study, and we focus on enhancing SOH prediction for mining applications. The parallel structure allows for simultaneous extraction of local features and temporal dependencies, improving overall performance.
The CNN component, particularly 1D CNN, is effective for feature extraction from sequential data. The convolution operation in the l-th layer for the h-th feature with the g-th feature from the (l-1)-th layer can be described as follows. Let the feature from the (l-1)-th layer be represented as:
$$x^{l-1}_g = (x^{l-1}_{g1}, \dots, x^{l-1}_{gm}, \dots, x^{l-1}_{gL})$$
After applying a nonlinear activation function, we get:
$$x^{l}_{hm}’ = f(W^l \times x^{l-1}_{g(m+k-1)} + b^l)$$
where $\times$ denotes the convolution operation, $f$ is the activation function (ReLU in this case), $W$ is the weight matrix, and $b$ is the bias matrix. The feature at the l-th layer is then:
$$x^l_h = (x^{l}_{h1}’, \dots, x^{l}_{hm}’, \dots, x^{l}_{hs}’)$$
For time-series data, LSTM is preferred over traditional RNN due to its ability to handle long sequences. The LSTM unit consists of an input gate $i_t$, a forget gate $f_t$, and an output gate $o_t$. The computations at time step $t$ are:
$$f_t = \sigma(W_f \cdot [h_{t-1}, X_t] + b_f)$$
$$i_t = \sigma(W_i \cdot [h_{t-1}, X_t] + b_i)$$
$$\tilde{c}_t = \tanh(W_c \cdot [h_{t-1}, X_t] + b_c)$$
$$c_t = f_t \cdot c_{t-1} + i_t \cdot \tilde{c}_t$$
$$o_t = \sigma(W_o \cdot [h_{t-1}, X_t] + b_o)$$
$$h_t = o_t \cdot \tanh(c_t)$$
where $\sigma$ is the sigmoid function, $\tanh$ is the hyperbolic tangent function, $h_{t-1}$ is the hidden state at time $t-1$, $X_t$ is the input feature sequence at time $t$, and $c_t$ is the cell state. The forget gate $f_t$ decides what information to discard, the input gate $i_t$ updates the cell state, and the output gate $o_t$ controls the output.
Our proposed CNN-LSTM_Parallel model integrates these components. As shown in the structure, the normalized feature data is fed into both a CNN layer and an LSTM layer (LSTM1) simultaneously. The CNN layer uses 12 filters with a convolution window length of 6 to extract local features, while LSTM1 contains 64 units to capture temporal features. The outputs from these layers are concatenated along the channel axis using a concatenation layer, merging into a single tensor. This tensor is then passed to another LSTM layer (LSTM2) with 16 units for further feature extraction, followed by a fully connected layer to produce the SOH prediction. To prevent overfitting, we apply Dropout with a rate of 0.2, randomly disabling a fraction of neural units during training.
The SOH is defined as the ratio of current capacity to nominal capacity:
$$SOH = \frac{C_{\text{current}}}{C_{\text{nominal}}}$$
where $C_{\text{current}}$ is the current capacity of the lifepo4 battery, and $C_{\text{nominal}}$ is the rated capacity. Prediction performance is evaluated using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE):
$$RMSE = \sqrt{\frac{1}{n – T} \sum_{t=T+1}^{n} (x_t – \hat{x}_t)^2}$$
$$MAE = \frac{1}{n – T} \sum_{t=T+1}^{n} |x_t – \hat{x}_t|$$
Here, $n$ is the sequence length, $T$ is the number of samples used for training, $x_t$ is the reference value, and $\hat{x}_t$ is the predicted value.
For our experiments, we used a dataset from MIT Laboratory, publicly released by Severson et al., which includes 124 commercial lifepo4 batteries. These batteries were cycled under different fast-charging strategies until capacity dropped to 80% of the nominal capacity. The initial capacity is 1.1 Ah, and voltage is 3.3 V. We selected five complete single-cell cycle data sets for SOH prediction, each with distinct charging protocols as summarized in Table 1.
| Battery ID | Charging Strategy | Discharge Rate |
|---|---|---|
| 1# | 4.8C, 4.8C (80%), 1C (100%) | 4C |
| 2# | 5C (67%), 4C (80%), 1C (100%) | 4C |
| 3# | 5.3C (54%), 4C (80%), 1C (100%) | 4C |
| 4# | 5.6C (19%), 4.6C (80%), 1C (100%) | 4C |
| 5# | 5.6C (36%), 4.3C (80%), 1C (100%) | 4C |
The discharge capacity and discharge time are primary parameters for assessing the performance of the lifepo4 battery. As cycling progresses, both discharge capacity and discharge time decrease, indicating SOH degradation. Additionally, internal resistance (IR) is a critical health indicator. The impedance of the lifepo4 battery increases over cycles due to factors like the buildup of high-impedance passive films on electrodes and reduction in active sites for lithium-ion intercalation/deintercalation. This rise in charge-transfer impedance leads to kinetic deterioration and capacity fade. Thus, we use three key features: internal resistance, discharge capacity, and discharge time. Data normalization is applied to scale features to [0, 1] using min-max normalization:
$$x’ = \frac{x – \min(x)}{\max(x) – \min(x)}$$
This ensures all variables contribute equally to the prediction.
The experimental platform consists of an NVIDIA GeForce RTX2080Ti GPU, 16 GB RAM, 512 GB SSD, and a 64-bit Windows 10 Enterprise system. We implemented the CNN-LSTM_Parallel model in Jupyter Notebook using TensorFlow. The training process involves splitting data into training and testing sets, with five-fold cross-validation (four batteries for training, one for testing in each fold). The model hyperparameters are set as described, and optimization is performed using appropriate loss functions.
Results demonstrate the effectiveness of our parallel CNN-LSTM model for SOH prediction of mining lifepo4 battery. Table 2 compares the performance of CNN-LSTM_Parallel with other models: CNN-LSTM (sequential), LSTM, and RNN. The metrics RMSE and MAE are reported for each battery.
| Battery ID | Model | RMSE | MAE |
|---|---|---|---|
| 1# | CNN-LSTM_Parallel | 0.35 | 0.22 |
| CNN-LSTM | 0.59 | 0.36 | |
| LSTM | 1.60 | 0.83 | |
| RNN | 0.97 | 0.41 | |
| 2# | CNN-LSTM_Parallel | 0.59 | 0.37 |
| CNN-LSTM | 0.97 | 0.60 | |
| LSTM | 0.91 | 0.57 | |
| RNN | 0.90 | 0.39 | |
| 3# | CNN-LSTM_Parallel | 0.72 | 0.32 |
| CNN-LSTM | 0.82 | 0.46 | |
| LSTM | 0.80 | 0.65 | |
| RNN | 1.20 | 0.67 | |
| 4# | CNN-LSTM_Parallel | 0.25 | 0.20 |
| CNN-LSTM | 2.74 | 1.23 | |
| LSTM | 3.28 | 1.46 | |
| RNN | 1.66 | 0.86 | |
| 5# | CNN-LSTM_Parallel | 0.51 | 0.28 |
| CNN-LSTM | 0.94 | 0.32 | |
| LSTM | 2.16 | 0.87 | |
| RNN | 0.87 | 0.35 |
The CNN-LSTM_Parallel model achieves the lowest RMSE and MAE values across all batteries, with errors less than 3%. The average RMSE and MAE are 0.484% and 0.278%, respectively, showing superior accuracy and stability. For example, on battery 4#, the parallel model reduces RMSE by 2.49% and MAE by 1.03% compared to the sequential CNN-LSTM. The lifepo4 battery’s SOH prediction benefits from the model’s ability to capture both spatial and temporal features, leading to robust performance even under varying charging conditions.
To further illustrate the model’s capabilities, we can derive the feature extraction process mathematically. Let the input sequence be $X = [x_1, x_2, \dots, x_T]$, where each $x_t$ is a vector of features for the lifepo4 battery at time $t$. The CNN layer applies convolution kernels to extract local patterns, producing feature maps $C = [c_1, c_2, \dots, c_{T’}]$. Simultaneously, the LSTM1 layer processes the sequence to generate hidden states $H_1 = [h_1, h_2, \dots, h_T]$. The concatenation yields a combined tensor $Z = \text{concat}(C, H_1)$, which is then fed into LSTM2 for further temporal modeling, resulting in final hidden states $H_2$. The output layer computes the SOH prediction $\hat{y}$ as:
$$\hat{y} = W_o \cdot H_2 + b_o$$
where $W_o$ and $b_o$ are the output weight matrix and bias. The training minimizes the loss function, typically Mean Squared Error (MSE):
$$\mathcal{L} = \frac{1}{N} \sum_{i=1}^{N} (y_i – \hat{y}_i)^2$$
with $N$ being the number of samples. Optimization algorithms like Adam are used to update model parameters.
In terms of feature importance, we analyzed the correlation between selected features and SOH for the lifepo4 battery. Table 3 summarizes the Pearson correlation coefficients for discharge capacity, discharge time, and internal resistance with SOH across the five batteries.
| Battery ID | Discharge Capacity | Discharge Time | Internal Resistance |
|---|---|---|---|
| 1# | 0.95 | 0.93 | -0.89 |
| 2# | 0.94 | 0.92 | -0.88 |
| 3# | 0.96 | 0.94 | -0.90 |
| 4# | 0.93 | 0.91 | -0.87 |
| 5# | 0.95 | 0.93 | -0.89 |
All features show high absolute correlation, confirming their relevance for SOH prediction. The negative correlation for internal resistance aligns with expectations, as higher resistance indicates degradation. The lifepo4 battery’s performance is closely tied to these parameters, and our model effectively leverages them.
For data preprocessing, we also considered other potential features like temperature and charging voltage, but focused on the three key ones to avoid overfitting. The normalization step is crucial, as it standardizes the scale of features. Let the original feature matrix be $F \in \mathbb{R}^{m \times n}$, where $m$ is the number of samples and $n$ is the number of features. After min-max normalization, each element $f_{ij}$ is transformed to $f_{ij}’ \in [0, 1]$. This ensures that the CNN and LSTM layers receive inputs with consistent ranges, improving training convergence.
The parallel architecture of our model offers several advantages. First, it allows simultaneous processing of local and global temporal dependencies, which is beneficial for the lifepo4 battery data that exhibits both short-term fluctuations and long-term trends. Second, the concatenation layer enables information fusion from different feature representations, enhancing the model’s expressive power. Third, the additional LSTM2 layer refines the combined features, leading to more accurate predictions. We experimented with different configurations, such as varying the number of filters in CNN or units in LSTM, and found the current setup optimal for the lifepo4 battery dataset.
To validate robustness, we conducted sensitivity analysis on model hyperparameters. Table 4 shows the impact of changing dropout rate and LSTM units on RMSE for battery 1#.
| Dropout Rate | LSTM2 Units | RMSE (%) | MAE (%) |
|---|---|---|---|
| 0.1 | 8 | 0.40 | 0.25 |
| 0.2 | 16 | 0.35 | 0.22 |
| 0.3 | 24 | 0.38 | 0.24 |
| 0.2 | 32 | 0.36 | 0.23 |
A dropout rate of 0.2 and 16 units in LSTM2 yielded the best results, balancing bias and variance. This configuration was used for all batteries to ensure consistency.
In comparison to existing methods, our CNN-LSTM_Parallel model outperforms not only basic RNN and LSTM but also sequential CNN-LSTM. The sequential version processes data through CNN first and then LSTM, which may lose some temporal information during convolution. In contrast, the parallel approach preserves both feature types independently before fusion. This is particularly important for the lifepo4 battery, where internal resistance and discharge characteristics evolve over time in complex ways.
We also explored the use of attention mechanisms with LSTM, but found that the parallel CNN-LSTM provided sufficient accuracy without added complexity. For mining applications, computational efficiency is key, and our model achieves high performance with moderate resource requirements. The training time for the lifepo4 battery dataset was approximately 2 hours per fold on our hardware, which is acceptable for offline SOH estimation.
Looking ahead, the lifepo4 battery will continue to play a vital role in mining electrification. Accurate SOH prediction enables proactive maintenance, reduces downtime, and enhances safety. Future work could integrate real-time data from battery management systems (BMS) and adapt the model to dynamic mining environments. Additionally, transfer learning could be applied to generalize across different lifepo4 battery batches or manufacturers.
In conclusion, we proposed a parallel CNN-LSTM network for SOH prediction of mining lifepo4 battery. The model combines CNN’s local feature extraction and LSTM’s temporal modeling, effectively capturing spatiotemporal patterns from discharge capacity, discharge time, and internal resistance. Experimental results on a public dataset show that our approach achieves errors below 3%, with average RMSE and MAE of 0.484% and 0.278%, outperforming sequential CNN-LSTM, LSTM, and RNN models. The lifepo4 battery’s health monitoring benefits from this robust prediction framework, supporting safer and more efficient mining operations. Further enhancements could involve multi-task learning for simultaneous SOH and state-of-charge (SOC) estimation, extending the model’s utility in practical scenarios.