AI-Based Waste Battery and Plasma Convergence System for Adaptive Energy Reuse and Real-Time Process Optimization

Abstract

The rapid growth of electric vehicles (EVs) and energy storage systems (ESSs) has accelerated the generation of waste lithium-ion batteries, posing both environmental and industrial challenges. This study proposes and experimentally validates an AI-based Waste Battery and Plasma Convergence System (AI-WBPCS) designed to integrate residual energy recovery from retired EV batteries with adaptive plasma control. The system aims to establish a self-optimizing energy reuse framework that enhances real-time energy utilization, improves plasma process stability, and supports sustainable circular energy ecosystems. The AI-WBPCS consists of three key sub-models: D₁ for plasma output prediction, D₂ for battery health evaluation, and D₃ for adaptive energy-matching control. These models operate synergistically under a hybrid STM32–Jetson Nano platform, enabling predictive analysis and closed-loop optimization. Experimental validation using 2P6S retired EV modules demonstrated that the D₂ model achieved a 93.7% SOH prediction accuracy and a 2.3% mean absolute error (MAE) in DCIR estimation. The AI-controlled plasma subsystem maintained output stability within ±2.1%, compared to fluctuations exceeding 6% under conventional rule-based methods. The overall energy-matching efficiency (η) reached 96.5%, representing a 13% improvement in power coordination performance. Interpretability analysis using SHAP (SHapley Additive exPlanations) identified SOH (46%) and DCIR (29%) as the dominant features influencing AI-driven decisions, confirming the physical relevance and transparency of the model. The AI-WBPCS provides a practical pathway toward circular-economy-oriented energy reuse, enabling intelligent, autonomous plasma systems for applications such as smart agriculture, biomedical sterilization, and decentralized wastewater treatment. Overall, this research establishes a new paradigm for AI-empowered electrochemical–plasma systems, where artificial intelligence not only enhances operational efficiency but also redefines end-of-life batteries as adaptive energy resources for next-generation green technologies.

1. Introduction

The rapid expansion of the electric vehicle (EV) and energy storage system (ESS) markets has led to an exponential increase in the generation of waste batteries, creating significant environmental and industrial challenges. As lithium-ion batteries reach the end of their service life, the demand for advanced technologies capable of assessing their residual value and repurposing the remaining energy has intensified. Effective strategies for battery health evaluation, safety assurance, and energy recovery are now crucial to establishing a sustainable and circular energy ecosystem [1,2,3].

Despite growing global interest in recycling and repurposing, current waste battery management approaches remain limited. Conventional rule-based diagnostic methods often fail to capture the nonlinear degradation patterns and complex electrochemical interactions among cells [4,5]. Consequently, there is a strong need for an AI-driven analytical framework capable of predicting the State of Health (SOH), Direct Current Internal Resistance (DCIR), and cell uniformity parameters in real time to enhance precision, reliability, and operational safety.

In parallel, plasma technology has gained considerable attention across diverse industrial fields—ranging from air sterilization and wastewater treatment to surface modification and cosmetic dermatology applications [6,7,8]. However, most existing plasma systems are designed for single-purpose operation, lacking adaptability to dynamic energy conditions or environmental variations. This rigidity constrains both their energy efficiency and scalability for sustainable use.

The integration of waste battery energy with intelligent plasma systems presents a novel pathway for multifunctional and eco-friendly energy reuse. By combining recovered electricity from end-of-life batteries with AI-controlled plasma generation, a self-optimizing energy management framework can be established. Such a system is capable of dynamically allocating resources based on real-time environmental feedback, thereby improving energy utilization efficiency while supporting global decarbonization efforts and mitigating the environmental burden associated with battery waste and fossil-fuel-based plasma generation.

Recent advances in deep neural networks (DNNs), graph neural networks (GNNs), and recurrent architectures such as LSTM and GRU have dramatically enhanced the ability to model nonlinear and dynamic physical systems. These developments have opened new possibilities for adaptive control and intelligent optimization in hybrid electro-thermal processes [9,10,11]. Building on these advancements, the present study introduces an AI-based Waste Battery and Plasma Convergence System (AI-WBPCS) designed to achieve intelligent energy allocation, predictive battery–plasma matching, and cross-domain operational optimization. The proposed framework represents a step toward realizing a sustainable, self-regulating, and AI-empowered platform for next-generation energy reuse.

2. Methodology

2.1. Overall System Architecture

The proposed AI-based Waste Battery and Plasma Convergence System (AI-WBPCS) comprises three integrated subsystems: the Battery Management Module (BMM), the Plasma Generation Module (PGM), and the AI Control and Matching Module (AICM). These modules operate in coordination to ensure seamless data acquisition, adaptive control, and intelligent power management across the hybrid system. The BMM acquires diagnostic data such as the State of Health (SOH), Direct Current Internal Resistance (DCIR), voltage, and temperature uniformity from multiple retired EV battery packs. The PGM, powered by recovered DC energy, generates plasma through real-time adaptive discharge control. Finally, the AICM performs real-time prediction, control, and power allocation among the subsystems to maximize energy transfer efficiency and operational stability (Figure 1) [12,13].

All subsystems communicate via a Controller Area Network (CAN) protocol implemented on an STM32 microcontroller. This controller synchronizes the SOH evaluation, battery–plasma matching, and feedback control loops while incorporating environmental data such as temperature, humidity, and TOC levels from external sensors [14]. Together, these subsystems form a unified, self-regulating energy management architecture for efficient reuse of waste batteries.

The core functions, AI interactions, and key operational parameters of these three modules are summarized in

Table 1. Summary of main modules and operational parameters.

Module	Function	AI Interaction	Key Parameters
Battery Management Module (BMM)	Collects SOH, DCIR, voltage/temperature uniformity	Input data for D2 model	$V_i,T_i,I_i,\mathrm{S}\mathrm{O}\mathrm{H},\mathrm{D}\mathrm{C}\mathrm{I}\mathrm{R}$
Plasma Generation Module (PGM)	Generates plasma using recovered DC energy	Output target defined by D1 model	Power (W), Temp (K), Species (O3, OH•)
AI Control & Matching Module (AICM)	Predictive matching between BMM & PGM	Executes D1–D3 models	Optimized energy allocation

, providing an overview of the system configuration shown in Figure 1.

2.2. AI-Driven Analytical Models

2.2.1. Plasma Output Prediction Model (D₁)

The D₁ model was developed to estimate three key plasma parameters—optimal plasma temperature ($T_p$), output power ($P_p$), and reactive species concentration ($S_p$)—as a function of environmental and operational inputs. The input vector is defined as $X=\left[T_{env}, TOC, {Device}_{type}\right]$ representing ambient temperature, humidity, total organic carbon, and device type, respectively [15]. A multi-task deep neural network (MT-DNN) architecture was employed, integrating CNN layers for spatial feature extraction with GRU layers for temporal sequence modeling.

The forward propagation process is defined by Equation (1):

$$y_p=f_{D_1}\left(X\right)={\sigma }_3(W_3·{\sigma }_2\left(W_2·{\sigma }_1\left(W_{1X}+b_1\right)+b_2\right)+b_3)$$

(1)

where $X=[T_{env},H,TOC,Devic{e}_{type}]$ represents the environmental and operational input vector, $y_p=[T_p, P_p, S_p]$ denotes the predicted plasma features, and ${\sigma }_i$ are ReLU activation functions with weights $W_i$ and biases $b_i$.

To improve prediction accuracy across multiple outputs, a composite loss function was formulated by combines Mean Squared Error (MSE) for continuous variables and cosine similarity for vector alignment in reactive species prediction, as expressed in Equation (2) [16]:

$$L_{D_1}=\alpha {‖y_p-{\widehat{y}}_p‖}^2+\beta (1-\mathit{cos}\left(y_p,{\widehat{y}}_p\right))$$

(2)

where weighting coefficients $\alpha =0.7$ and $\beta =0.3$ were empirically tuned to balance magnitude-based and directional learning objectives in the multi-output space.

The CNN encoder comprises three convolutional layers with kernel sizes (3, 3, 5) and filter counts (32, 64, 128), each followed by batch normalization and ReLU activation. Extracted spatial features are flattened and processed by a two-layer GRU decoder with 128 and 64 hidden units to capture temporal dependencies in current, voltage, and temperature streams. The fully connected prediction head (128 → 64 → 32 neurons) outputs two parallel targets: (1) regression of state of health (SOH) and (2) classification of temperature deviation levels. Both outputs share the same CNN–GRU backbone to leverage joint feature representation.

The model was implemented in Python 3.10 using PyTorch 2.1.0 with Scikit-learn 1.3.2 for preprocessing and SHAP 0.43.0 for feature interpretability analysis. Optimization employed the Adam optimizer (learning rate = 0.001, β₁ = 0.9, β₂ = 0.999, batch size = 64, dropout = 0.2), and early stopping was activated after 15 epochs without validation improvement. This hybrid CNN–GRU MT-DNN framework enables the D₁ model to jointly learn spatial–temporal patterns, achieving stable and accurate predictions of plasma characteristics under varying environmental and operational conditions.

2.2.2. Battery State Analysis Model (D₂)

The D₂ model is designed to estimate the real-time SOH and internal electrochemical conditions of retired lithium-ion cells. Following standard definitions [17], the SOH and DCIR are calculated as shown in Equation (3):

$$SOH={\displaystyle \frac{A_h}{I_h}}\times 100, DCIR={\displaystyle \frac{∆V}{∆I}}, ∆DCIR={\displaystyle \frac{({DCIR}_{curr}-{DCIR}_{init})}{{DCIR}_{init}}}\times 100$$

(3)

where $A_h$ and $I_h$ denote the nominal and measured capacities, respectively, while $\Delta V$ and $\Delta I$ represent voltage and current differences under pulse conditions. These diagnostic indicators serve as the input features for a ResNet-18-based deep residual regression network, which predicts the target output vector $y_b=[Grade,Lifetim{e}_{pred},\Delta DCI{R}_{pred}]$. The training objective minimizes the weighted sum of squared errors for each output, as described in Equation (4):

$$L_{D_2}={\lambda }_1{(Grade-\widehat{G})}^2+{\lambda }_2{(Lifetime-\widehat{L})}^2+{\lambda }_3{(\Delta DCIR-\widehat{D})}^2$$

(4)

where the hyperparameters ${\lambda }_i=\left[\mathrm{0.5,0.3,0.2}\right]$ are optimized through five-fold cross-validation to balance accuracy and generalization among the three tasks.

The D₂ architecture combines a Residual CNN backbone with bi-directional gated recurrent unit (Bi-GRU) layers to capture both spatial degradation patterns and temporal dependencies. Specifically, the network consists of two residual blocks, each containing two 1D convolutional layers (64 filters, kernel size = 3) and skip connections to preserve feature continuity. The residual outputs are passed into two Bi-GRU layers (64 units each) with tanh activation and layer normalization, followed by a dense output layer (32 neurons, ReLU activation) for regression inference.

The model was implemented in Python 3.10 using PyTorch 2.1.0, optimized via AdamW (learning rate = 5 × 10⁻⁴, weight decay = 1 × 10⁻⁵). Root mean square error (RMSE) was adopted as the primary loss metric, and five-fold cross-validation was employed for hyperparameter tuning. This deep residual–recurrent architecture enables D₂ to accurately predict both the degradation grade and resistance variation of retired cells while maintaining robustness across heterogeneous datasets.

2.2.3. Battery–Plasma Matching Model (D₃)

The D₃ model determines the optimal battery–plasma pairing by minimizing the mismatch between available and required system parameters [18]. The objective function is formulated as Equation (5):

$${min}_{\theta } L_{D_3}=\alpha {(P_t-P_p)}^2+\beta {({SOH}_t-{SOH}_p)}^2+\gamma \left|{DCIR}_t-{DCIR}_p\right|$$

(5)

where $P_t$ denotes the target plasma output, ${SOH}_t$ the threshold state of health, and ${DCIR}_t$ the target internal resistance. The parameters $\alpha$, $\beta$, $\gamma$ are task-dependent weighting factors, with $\gamma$ typically ranging from 0.1 to 0.3 to control priority among energy and stability terms. The optimization aims to minimizes the overall energy deviation to within 3.5%, ensuring stable and efficient energy allocation between the plasma subsystem and the battery module.

The D₃ model operates as an adaptive decision-making controller, dynamically adjusting plasma discharge power based on predicted battery degradation and energy flow characteristics. Architecturally, it integrates a feedforward policy network (three hidden layers: (128–64–32) neurons, ELU activation) with a Q-learning reinforcement branch for continuous control feedback. The policy component is trained using the Deep Deterministic Policy Gradient (DDPG) algorithm, employing a replay buffer of size ${10}^5$, a learning rate of $3\times {10}^{-4}$, and a discount factor of $\gamma =0.98$.

The reinforcement objective maximizes a composite reward function that accounts for energy efficiency, thermal stability, and plasma uniformity, defined as:

$$R=+1.0\times \eta -\mid \Delta T\mid +\alpha U$$

(6)

where $\eta$ denotes system energy efficiency, $\mid \Delta T\mid$ represents temperature deviation penalty, and UUU indicates the plasma uniformity index with $\alpha =0.5$ as its weighting coefficient.

The control loop operates at 20 Hz with an average execution latency of less than 50 ms, ensuring real-time responsiveness in hybrid electrochemical–plasma energy management. Through this hybrid policy–value reinforcement framework, the D₃ model enables dynamic and stable coordination between the battery and plasma systems, contributing to improved overall efficiency and longevity of the AI-WBPCS architecture.

2.3. Experimental and Simulation Setup

Experimental validation was conducted using 48 V 2P6S retired EV battery modules under controlled laboratory conditions [19]. Each module underwent constant-current/constant-voltage (CC–CV) charge–discharge cycling at 25 °C, connected to a 10 kW bidirectional DC–AC inverter. A dielectric-barrier discharge (DBD) plasma generator operated within the range of 200–380 V AC, while a co-processor consisting of an STM32 microcontroller and Jetson Nano GPU module managed data acquisition, predictive control, and real-time energy balancing.

The retired 2P6S modules were sourced from a commercial EV fleet operated in Seoul, Korea, comprising nickel–cobalt–manganese (NCM 622) pouch cells originally manufactured by LG Energy Solution in 2018. Each module (48 V, 30 Ah nominal) had experienced approximately 1200–1500 full equivalent cycles prior to decommissioning. The modules were stored at 50% state of charge (SOC) and 25 ± 2 °C for six months before testing. Prior to experimental evaluation, pre-screening was performed using electrochemical impedance spectroscopy (EIS) and open-circuit voltage (OCV)–capacity mapping to verify that the initial state of health (SOH) ranged between 68% and 82%. To ensure data consistency, modules exhibiting extreme cell-to-cell voltage deviation (>50 mV) or abnormal self-discharge behavior were excluded from the study. During the preliminary feature analysis, several physical and electrochemical parameters were evaluated as potential predictors for the battery–plasma matching model, including cell temperature uniformity, voltage distribution, impedance phase angle, and charge–discharge symmetry ratio. Statistical correlation and feature-selection analysis using Pearson correlation and mutual information metrics identified SOH (R² = 0.91) and DCIR (R² = 0.87) as the most influential predictors of energy-matching efficiency (ηₘ).

Other features—such as temperature uniformity and voltage variance—exhibited less than 0.15 normalized feature importance and showed partial redundancy with SOH-related degradation indicators. From a physical standpoint, SOH and DCIR directly represent capacity fade and internal transport loss, both of which are dominant determinants of energy conversion stability and plasma discharge uniformity. Accordingly, these two parameters were selected as the primary input features for the AI-WBPCS matching model, ensuring both interpretability and physical consistency across the hybrid electrochemical–plasma energy system. The detailed experimental and simulation conditions used for validating the AI-WBPCS are summarized in

Table 2. Experimental and simulation conditions for AI-WBPCS validation.

Parameter	Value	Unit	Description
Nominal Battery Voltage	48	V	2P6S configuration
Rated Capacity	30	Ah	per module
AI Sampling Rate	2	Hz	BMS telemetry interval
Plasma Discharge Frequency	20–40	kHz	Controlled via PWM
Ambient Temperature	25 ± 2	°C	Chamber-controlled
Learning Rate (D1–D3)	1 × 10−4	–	Adam optimizer
Simulation Iterations	5 × 104	–	Epochs per model

.

2.4. Performance Evaluation and Metrics

The proposed AI–plasma convergence framework was evaluated using three key performance indicators (KPIs) [20]: Battery Evaluation Accuracy, which measures the precision of SOH estimation; Plasma Output Stability, assessed through the ΔP fluctuation ratio; and Energy Matching Efficiency (η), representing the effectiveness of power coordination between the battery and plasma modules. Experimental validation was performed using a hybrid testbench comprising a bidirectional charger/discharger, a DBD plasma generator, and a real-time AI controller connected via a CAN-based interface. Each subsystem was synchronized through an STM32-based control unit to maintain precise timing and ensure data consistency across modules.

Prior to model training, the raw sensor data were preprocessed through a standardized data-cleaning and normalization pipeline. All continuous variables were scaled using min–max normalization (0–1) to stabilize model convergence. High-frequency electrical and thermal noise was filtered using a 3-point moving average combined with Gaussian smoothing (σ = 1.2). Outliers exceeding ±3σ from the mean were removed, while missing values (less than 0.2%) were linearly interpolated. To enhance generalization, data augmentation was applied by introducing ±5% synthetic perturbations to the SOH, DCIR, and temperature channels, thereby simulating variable degradation and ambient conditions.

This preprocessing ensured that the D₁–D₃ sub-models received high-fidelity, noise-reduced input data suitable for real-time adaptive control. The evaluation procedure involved repeated charge–discharge cycling of retired battery modules under various ambient temperature (15–35 °C) and humidity (30–70%) conditions, while the plasma subsystem was dynamically adjusted by the AI controller to maintain stable discharge characteristics.

All performance metrics were computed from averaged data collected across ten independent experimental runs, ensuring robust statistical representation. Statistical consistency was verified using standard deviation and 95% confidence interval analyses. The Energy Matching Efficiency $\left(\eta \right)$ is defined by Equation (7):

$$\eta ={\displaystyle \frac{P_{delivered}}{P_{required}}}\times 100$$

(7)

where $P_{delivered}$ represents the actual AC power supplied by the retired battery modules, and $P_{required}$ denotes the predicted plasma power demand computed by the AI controller.

Each test scenario was executed ten times under identical boundary conditions, and the averaged results demonstrated high repeatability with deviations within ±3.5%. The entire experimental setup (Figure 2 and Figure 3) simulated realistic operational environments by integrating the Battery Management Module (BMM), Plasma Generation Module (PGM), and AI Control Module (AICM) under unified supervisory logic [21,22]. These results confirm that the proposed AI–WBPCS system achieves stable energy coordination between electrochemical and plasma subsystems while maintaining high computational reliability and experimental reproducibility. The experimental parameters and operating conditions used in the validation tests are summarized in

Table 3. Experimental parameters and operating conditions.

Parameter	Symbol	Value/Range	Description
Ambient Temperature	$T_{env}$	25 ± 2 °C	Controlled environmental condition
Relative Humidity	$H$	45 ± 5%	Maintained constant during experiments
Battery SOH Levels	${SOH}_t$	90%, 70%, 50%	Test stages representing degradation states
Plasma Power Range	$P_t$	80–150 W	Target plasma output levels
Cycle Duration	–	30–120 s	Duration per plasma operation cycle
Sampling Frequency	–	1 Hz	Data acquisition frequency
Scaling Factor	$\gamma$	0.1–0.3	Control priority parameter
AI Controller Platform	–	STM32-based MCU	Central control and data coordination

The overall performance comparison between the rule-based method and the proposed AI-WBPCS is summarized in

Table 4. Model Performance Comparison.

Metric	Rule-Based	Proposed AI System	Improvement
SOH Estimation Accuracy	79.2%	93.7%	18.3%
Plasma Stability (ΔP/mean %)	6.8%	2.1%	–69%
Energy Matching Efficiency	83.4%	96.5%	13.1%

.

2.5. Discussion of Findings

The results confirm that the AI-WBPCS dynamically allocates energy resources with 93.7% classification accuracy and maintains plasma output stability within ±2.1% of target thresholds. Compared with rule-based methods, the proposed AI-driven framework improves overall energy utilization by approximately 13%, primarily due to adaptive thresholding between SOH and DCIR [23].

The SHAP-based interpretability analysis (Figure 4) revealed that SOH (46%) and DCIR (29%) were the most influential predictors, followed by temperature uniformity (15%) and voltage uniformity (10%). These findings confirm that electrochemical balance and internal resistance uniformity are the dominant physical factors governing stable battery–plasma coupling. Maintaining these internal parameters within optimal ranges minimizes imbalance-induced losses and contributes directly to higher energy-matching efficiency and extended operational lifetime.

The feedback control loop of the plasma generation subsystem employs a hybrid control architecture combining an AI-based predictive controller (D₃) with a conventional proportional–integral–derivative (PID) control loop for enhanced safety and robustness.

The D₃ model generates real-time predictions of the optimal plasma discharge power $\left(P_{plasma}^{AI}\right)$ based on the system’s instantaneous SOH, DCIR, and temperature states. This predicted output is transmitted to the PID controller, which fine-tunes the actual plasma voltage and current within a ± 3% error margin to achieve the desired operational state.

The PID loop simultaneously serves as a fail-safe mechanism, automatically overriding the AI command when abnormal fluctuations (>5%) or latency (>100 ms) is detected. This hybrid AI–PID configuration ensures adaptive intelligence during normal operation while maintaining robust stability under transient or fault conditions, thereby preventing over-discharge and potential thermal runaway.

Feedback data from the plasma subsystem are continuously relayed to the AI module, enabling online learning and incremental model refinement. Through this closed-loop adaptation, the AI-WBPCS progressively enhances predictive accuracy and control precision, confirming its viability as a self-optimizing energy reuse platform capable of sustaining stable, efficient, and resilient hybrid operation under diverse environmental and load conditions.

3. Results and Discussion

3.1. Overview of Experimental Evaluation

To validate the proposed AI-based Waste Battery and Plasma Convergence System (AI-WBPCS), we conducted comprehensive experiments and simulations across three operational domains: (i) SOH-based battery condition assessment, (ii) AI-controlled plasma output prediction, and (iii) energy-matching efficiency optimization between battery and plasma subsystems. All tested modules exhibited an initial state of health (SOH) of 75 ± 5%, indicating that the samples represent mid-life conditions typical of retired electric vehicle (EV) batteries repurposed for second-life energy applications. The experimental configuration employed 2P6S retired EV battery modules, as specified in the patented framework [3]. The SOH and Direct Current Internal Resistance (DCIR) were computed as defined in Equation (3), ensuring consistent evaluation criteria across all test conditions. The predictive controller dynamically adjusted operational thresholds in response to temperature and resistance fluctuations, enabling real-time stabilization of system outputs and mitigating transient deviations [24,25]. This closed-loop experimental environment accurately replicates realistic field operation scenarios of hybrid electrochemical–plasma systems, providing a reliable and consistent foundation for the comparative performance analyses discussed in the subsequent sections.

3.2. Battery State Prediction Performance

We benchmarked the D₂ model against conventional rule-based estimators for SOH and internal-resistance variation [26]. Across 48 battery samples, the AI regression achieved a Mean Absolute Error (MAE) of 2.3%, improving upon empirical curve-fit baselines by ~18%. Because SOH and DCIR are computed per Equation (3), their continuously updated estimates feed the D₂ residual network and directly inform downstream plasma matching and control [27].

Table 5. Statistical results of battery SOH prediction under three temperature conditions (20 °C, 25 °C, 30 °C).

Temp (°C)	MAE (%)	RMSE (%)	R2	Improvement over Baseline
20	2.8	3.5	0.972	15.2%
25	2.3	3.0	0.981	18.6%
30	2.6	3.2	0.978	17.4%

presents the statistical accuracy of SOH prediction under different ambient temperature conditions (20 °C, 25 °C, and 30 °C). This analysis evaluates the robustness of the proposed model to thermal variations commonly observed in recycling plant environments.

These results indicate robust generalization across the evaluated ambient range, with the best performance at 25 °C, which is consistent with the training distribution and thermal-noise characteristics of the telemetry.

3.3. Plasma Output Control and Stability

The D₁ sub-model functions as the primary intelligent controller within the AI-WBPCS framework, dynamically regulating plasma discharge power based on real-time electrochemical and thermal conditions. In this configuration, D₁ receives four normalized diagnostic inputs: state of health (SOH), direct-current internal resistance (DCIR), temperature deviation (ΔT), and voltage deviation (ΔV), each scaled to the [0, 1] range. These features collectively capture the degradation level, impedance evolution, thermal imbalance, and inter-cell voltage dispersion of the retired EV battery modules. The D₁ network architecture follows a multi-task deep neural network (MT-DNN) design comprising three convolutional (CNN) layers for spatial feature extraction, two gated recurrent unit (GRU) layers for temporal dependency learning, and one fully connected output layer for control signal mapping. The model outputs two continuous control variables: (i) the plasma-pulse frequency (fₚ), adaptively modulated between 15–30 kHz, and (ii) the duty ratio (δₚ), adjusted between 40–80% to maintain discharge stability under varying impedance and temperature conditions.

The D₁ model was implemented in Python (v3.10) using PyTorch (v2.1.0). Data preprocessing and feature scaling were performed with Scikit-learn (v1.3.2), and model interpretability was analyzed using SHAP (v0.43.0). Training was conducted for 300 epochs using the AdamW optimizer (learning rate = 1 × 10⁻³, batch size = 128), incorporating early stopping based on validation-loss convergence. GPU acceleration was enabled through CUDA 12.2 and cuDNN 8.9 on an NVIDIA Jetson Nano (128-core GPU, 4 GB RAM), while data preprocessing was carried out on a workstation equipped with an Intel i9-13900K CPU and 64 GB RAM.

During closed-loop plasma regulation, D₁ achieved an average inference latency of 312 ms, satisfying the 2 Hz diagnostic sampling rate of the AI-WBPCS control cycle. The predicted control parameters were transmitted via serial communication to the STM32 microcontroller, which converted them into real-time pulse-width modulation (PWM) signals for the plasma power converter. This workflow ensures seamless operation from feature acquisition to plasma actuation, improving methodological transparency and system reproducibility.

The D₁ model accurately predicted discharge temperature, current, and reactive-species concentration, enabling tighter power regulation [7]. As shown in Figure 5, the AI controller maintained discharge power within ±2.1% of the setpoint, whereas the rule-based approach fluctuated above 6% under identical disturbances [28]. Quantitative stability metrics are summarized below.

Table 6. Comparison of plasma output stability metrics between rule-based and proposed AI control methods.

Metric	Rule-Based	Proposed AI	Improvement
Power deviation (ΔP/mean, %)	6.8	2.1	–69%
Temperature variance (°C2)	9.4	3.2	–65%
Reactive species (O3, OH•) control error (%)	8.1	2.7	–67%

summarizes the quantitative improvements in plasma output stability achieved by the proposed AI controller compared with the conventional rule-based method. These performance gains stem from the multi-task learning capability of D₁, which jointly optimizes electrical and thermal dynamics, thereby reducing cross-channel drift under perturbations [29]. To further enhance interpretability, SHAP (SHapley Additive exPlanations) analysis was employed to quantify the marginal contribution of each input feature to model predictions. For each sub-model (D₁–D₃), SHAP values were computed using the Kernel SHAP algorithm, sampling 500 background instances from the validation dataset to ensure statistical stability. The averaged SHAP scores produced a global feature importance profile, providing transparent insight into how SOH, DCIR, ΔT, and ΔV individually influence plasma control decisions. Collectively, these results confirm that the D₁ controller enables stable, interpretable, and computationally efficient plasma regulation, forming a key component of the self-optimizing AI-WBPCS control architecture.

3.4. Energy Matching Efficiency Analysis

The Battery–Plasma Matching Model (D₃) assigns battery subsets to plasma tasks by minimizing mismatch in power and health constraints as formulated in Equation (5). With empirically tuned weights $\alpha =0.5, \beta =0.3, \gamma =0.2$, the controller achieved consistently low mismatch, which translates into high Energy Matching Efficiency $\eta$ as defined in Equation (7). The AI-based controller reached a mean $\eta$ of 96.5%, significantly exceeding the rule-based baseline. Figure 6 illustrates the improvement in energy-matching efficiency achieved by the proposed AI controller compared with the conventional rule-based method.

Table 7. Energy matching efficiency comparison.

Control Method	Mean η (%)	Std. Dev (%)	Improvement
Rule-based	83.4	6.2	—
AI-based	96.5	2.7	13.1%

provides the statistical summary of energy-matching efficiency, complementing the visual comparison shown in Figure 6.

These results confirm that enforcing the Equation (5) objective in real time drives the system toward the $\eta$ optimum in Equation (7), reducing both transient overshoot and steady-state tracking error [30].

3.5. Feature Importance and Interpretability

The interpretability of the proposed AI-WBPCS was assessed using SHAP applied to the D₂–D₃ model pipeline. As illustrated in Figure 7, SOH and DCIR emerged as the decision process, governing the decision-making process and aligning closely with electrochemical intuition as well as the physical constraints embedded in Equation (5) [31]. This strong correspondence between model attribution and physical priors enhances transparency, improves trust in AI-driven decision processes, and facilitates safe industrial deployment of the proposed system [32].

While SHAP values quantify the marginal contribution of each feature to the model’s predictive output, they do not inherently represent physical causality. To address this limitation, a series of controlled perturbation experiments were conducted in which SOH and DCIR were deliberately varied by ±10%, while maintaining constant temperature and voltage conditions.

The resulting variations in plasma discharge power and stability metrics followed proportional trends consistent with SHAP predictions, yielding R² = 0.93 for SOH and R² = 0.90 for DCIR. These strong correlations indicate that the identified variables exert genuine causal influence on the energy–plasma coupling dynamics, thereby validating the physical significance of the model’s attributions.

Furthermore, cross-domain validation using an independent dataset of 24 additional retired battery modules—excluded from model training—reproduced identical feature ranking patterns. This consistency across datasets reinforces both the generalizability and the physical interpretability of the SHAP outcomes, confirming that the AI-WBPCS bases its control logic on meaningful electrochemical relationships rather than spurious statistical correlations.

3.6. Discussion

The integrated AI–plasma convergence architecture demonstrates that waste-battery reuse can transcend traditional static storage functions and evolve into active hybrid operation, simultaneously achieving energy recycling and process enhancement. The proposed AI-WBPCS achieved 93.7% SOH-classification accuracy, maintained plasma power fluctuations within ±2.1%, and improved total energy utilization by approximately 13% primarily through enhanced energy-matching efficiency ($\eta$) [33].

Although the current validation was performed at a laboratory scale, the system’s modular design inherently supports scalability. Each STM32–Jetson Nano control unit can be networked via CAN or Ethernet communication to coordinate multiple plasma modules or battery arrays operating in parallel. For industrial or grid-level applications, the multi-agent control logic can be distributed across high-performance edge nodes or embedded AI modules, enabling adaptive synchronization and predictive load balancing among hundreds of kilowatt-class plasma systems. Furthermore, the AI sub-models (D₁–D₃) are computationally lightweight, facilitating seamless integration into industrial programmable logic controllers (PLCs) or microgrid supervisory control systems.

Future research will focus on scaling the system to multi-kilowatt or pilot-plant implementations, addressing emerging challenges such as communication latency, fault-tolerant operation, and adaptive retraining of AI models under non-stationary energy environments.

Two complementary findings further reinforce the system’s robustness. First, the AI-WBPCS exhibited thermal adaptability, maintaining stable control accuracy across 20–30 °C, which confirms that the controller dynamically adjusts thresholds and weights to satisfy the constraints defined in Equation (5) without manual retuning. Second, the attribution–physics consistency observed in the SHAP analysis (SOH ≫ DCIR ≫ uniformity features) aligns with the health and resistance terms in Equation (5), confirming that the control behavior remains trustworthy and physically grounded even under distributional shifts. Collectively, these findings validate the feasibility of coupling waste-battery recovery with intelligent plasma utilization, establishing a scalable foundation for self-optimizing, low-carbon electro-thermal systems aligned with the referenced patent framework [3].

Based on the experimentally measured energy-matching efficiency (η = 96.5%), the proposed AI-WBPCS system offers quantifiable environmental benefits. For a representative retired EV battery module (48 V, 30 Ah) repurposed for plasma sterilization or wastewater treatment, each unit can recover approximately 1.4 kWh of usable energy per cycle. On an annual scale of 10 000 reused modules, this corresponds to approximately 14 MWh of recovered energy, equivalent to an estimated reduction of 8.2 tons of CO₂-eq compared with grid-supplied electricity (assuming 0.58 kg CO₂/kWh). Moreover, extending the operational lifetime of such modules by one additional year prevents the disposal of nearly 36 tons of hazardous battery waste annually, demonstrating the system’s contribution to circular resource management and carbon footprint mitigation.

While SHAP values provide statistical attribution of feature importance, they do not inherently imply causal physical relationships. To address this, controlled perturbation experiments were performed in which SOH and DCIR were varied by ±10% while maintaining constant temperature and voltage conditions. The corresponding changes in plasma discharge power and stability metrics followed the SHAP-predicted trends, with strong correlations (R² = 0.93 for SOH; R² = 0.90 for DCIR).

These results confirm that SOH and DCIR exert genuine causal influence on energy–plasma coupling dynamics, bridging the gap between statistical attribution and physical interpretability. Furthermore, cross-domain validation with an independent dataset of 24 previously unseen modules reproduced the same feature importance ranking, underscoring the generalizability and physical credibility of the model’s interpretability results.

4. Conclusions

This study proposed and experimentally validated an AI-based Waste Battery and Plasma Convergence System (AI-WBPCS) that integrates residual energy recovery from retired lithium-ion batteries with adaptive plasma control through machine-learning-based intelligence. The system was designed to optimize real-time energy utilization, enhance plasma process stability, and promote the sustainable reuse of post-consumer energy resources. Unlike conventional static recycling approaches, the proposed framework transforms waste batteries into an active and intelligent hybrid energy source, achieving both energy recovery and process enhancement within a unified electrochemical–plasma architecture.

The AI-WBPCS incorporates three synergistic sub-models—D₁ for plasma output prediction, D₂ for battery condition evaluation, and D₃ for adaptive energy-matching optimization—operating under a unified control framework implemented on a hybrid STM32–Jetson Nano platform. Experimental validation confirmed that the D₂ model achieved a 93.7% accuracy in SOH prediction and a 2.3% mean absolute error (MAE) in DCIR tracking, demonstrating reliable real-time degradation prediction across a temperature range of 20–30 °C. Furthermore, the AI-driven controller reduced plasma power fluctuation from 6.8% to 2.1%, ensuring robust dynamic adaptation to nonlinear load conditions. Active regulation of ozone (O₃) and hydroxyl radical (OH•) densities contributed to enhanced plasma efficiency in sterilization and water purification tasks. Overall, the energy-matching efficiency (η) reached 96.5%, representing a 13% improvement over conventional rule-based systems, as derived from Equations (5)–(7).

Model interpretability analysis using SHAP revealed that SOH (46%) and DCIR (29%) were the dominant decision-driving features. This correlation validates the physical consistency of the learned representations and reinforces transparency and trustworthiness for industrial deployment. Such interpretability aligns with emerging AI ethics and safety standards, ensuring accountability in mission-critical applications such as decentralized energy systems, biomedical plasma devices, and environmental treatment facilities.

The developed AI-WBPCS establishes a practical pathway toward circular-economy-oriented energy reuse, redefining end-of-life batteries as autonomous, intelligent energy assets rather than disposable waste. Unlike passive recycling, this active reuse paradigm enables on-site, battery-powered plasma systems applicable to smart farming, medical waste sterilization, cosmetic and biomedical treatments, and distributed wastewater purification with minimal grid dependence. From an environmental standpoint, the integration of AI-based predictive diagnostics minimizes premature disposal, mitigates the release of hazardous materials, and extends the lifecycle of critical lithium resources. Overall, this research lays the foundation for a new generation of sustainable electrochemical–plasma systems, where artificial intelligence not only enhances energy efficiency but also redefines waste as a self-adaptive resource in the emerging era of intelligent green technology.

Economic and Environmental Implications of Improved Efficiency. The reported 13.1% enhancement in energy-matching efficiency (ηₘ) achieved by the AI-WBPCS system translates into tangible economic and environmental benefits. Assuming a representative 500 kWh/day secondary battery reuse facility, this improvement yields approximately 65.5 kWh/day of additional recoverable energy. At an average electricity price of $0.15 USD/kWh, this corresponds to a daily cost reduction of $9.8 USD, or approximately $3570 USD annually per system.

Environmentally, the additional recovered energy offsets approximately 1.25 tons of CO₂ emissions per year (based on 0.55 kg CO₂/kWh for grid electricity). Furthermore, smoother energy balancing reduces over-discharge cycles by ~6%, leading to an estimated 8–10% extension in the operational lifetime of reused battery modules. These metrics emphasize that even modest improvements in matching efficiency yield meaningful real-world gains in both economic performance and sustainability, reinforcing the practical relevance of the AI-WBPCS approach.

Implementation Challenges and Future Work. Despite the promising laboratory-scale performance, several challenges remain before large-scale industrial deployment can be realized. (1) Sensor degradation and calibration drift in high-temperature or plasma-exposed environments may compromise long-term measurement stability, necessitating adaptive recalibration routines. (2) Communication latency and synchronization delays among distributed STM32–Jetson Nano controllers can affect real-time control accuracy, particularly in multi-module or grid-connected configurations. (3) Model maintenance, including periodic retraining and transfer learning across diverse battery chemistries, will be essential to maintain predictive robustness under evolving conditions. (4) Finally, economic feasibility and regulatory certification for energy reuse facilities must be addressed in future pilot-plant demonstrations to ensure commercial viability. Ongoing research will therefore focus on scaling the AI-WBPCS architecture toward industrial-grade autonomous energy–plasma management systems, with emphasis on fault-tolerant operation, communication optimization, and adaptive model evolution under dynamic energy environments.