Fine-Grained Dismantling Decision-Making for Distribution Transformers Based on Knowledge Graph Subgraph Contrast and Multimodal Fusion Perception

Abstract

Distribution transformers serve as critical nodes in smart grids, and management of their recycling plays a vital role in the full life-cycle management for electrical equipment. However, the traditional manual dismantling methods often exhibit a low metal recovery efficiency and high levels of hazardous substance residue. To facilitate green, cost-effective, and fine-grained recycling of distribution transformers, this study proposes a fine-grained dismantling decision-making system based on a knowledge graph subgraph comparison and multimodal fusion perception. First, a standardized dismantling process is designed to achieve refined transformer decomposition. Second, a comprehensive set of multi-dimensional evaluation metrics is established to assess the effectiveness of various recycling strategies for different transformers. Finally, through the integration of multimodal perception with knowledge graph technology, the system achieves automated sequencing of the dismantling operations. The experimental results demonstrate that the proposed method attains 99% accuracy in identifying recyclable transformers and 97% accuracy in auction-based pricing. The residual oil rate in dismantled transformers is reduced to below 1%, while the metal recovery efficiency increases by 40%. Furthermore, the environmental sustainability and economic value are improved by 23% and 40%, respectively. This approach significantly enhances the recycling value and environmental safety of distribution transformers, providing effective technical support for smart grid development and environmental protection.

1. Introduction

With the continuous advancement of smart grid [1,2,3] construction and the large-scale renewal of power equipment, full life-cycle management for electrical [4] assets has become a critical component in ensuring the efficient and stable operation of the power grid. As key nodes connecting transmission networks with end users, transformers play a pivotal role in the recycling [5], dismantling, and reuse phases of electrical equipment life-cycle management [6]. Retired transformers often contain highly polluting insulating oil, which can generate large quantities of toxic substances, such as polychlorinated biphenyls (PCBs), after prolonged use. Improper disposal of these materials can lead to serious soil and water contamination, posing significant threats to environmental sustainability [7,8,9,10].

The traditional manual dismantling methods [11] exhibit several limitations, including inconsistent decision-making criteria, non-standardized procedures, low operational efficiency, safety risks, and a lack of transparency. Furthermore, these methods yield limited economic returns [12] and environmental benefits while potentially enabling the recirculation of defective products. These challenges collectively underscore the need for efficient, intelligent, and automated dismantling systems and decision-making frameworks [13].

Currently, dismantling decisions predominantly depend on the expertise of recycling personnel. This approach not only imposes stringent technical requirements on workers but also hinders the implementation of customized recycling strategies for complex transformers. Consequently, such methods fail to meet the granular, full-chain management demands of modern power equipment. To improve recycling efficiency while simultaneously optimizing environmental and economic outcomes, the development of intelligent dismantling decision-making systems has emerged as a critical facilitator for advancing smart grid infrastructure.

With the rapid development of resource recycling techniques, scholars and enterprises worldwide have conducted extensive research on automated transformer dismantling and decision-making. Significant progress has been made in automation technologies globally, with researchers introducing multimodal vision algorithms to enhance the recognition and dismantling efficiency of complex internal transformer components [13,14,15]. Advances in robotics have introduced innovative solutions for transformer disassembly. By integrating visual systems and intelligent planning algorithms, robotic systems can dynamically adjust dismantling plans in real time based on sensor data, improving both stability and efficiency. For instance, Japan and Germany have implemented automated dismantling equipment that combines robotic arms with machine vision to accurately identify transformer housing bolts, enabling automatic shell separation and precise component sorting. In scrap wire recovery, automated stripping machines such as pneumatic wire strippers have demonstrated a 25% efficiency improvement over manual operations. Furthermore, studies indicate that adjusting the extension length of push rods allows for effective multi-diameter wire cutting [16,17].

Regarding recycling decision-making [18], several studies have focused on automated strategies using life-cycle assessment (LCA) and multi-objective optimization [18,19,20,21]. Some researchers have combined big data technologies with automated dismantling equipment to facilitate dynamic dismantling of complex power assets. Others have developed hybrid decision models based on the Analytic Hierarchy Process (AHP), integrating deep learning algorithms to construct evaluation systems for dismantling sequences and optimize the dismantling paths in real time.

In recent years, knowledge graph technology has been increasingly applied in both electrical equipment recycling and broader process industries to model complex entity–relation structures, support semantic queries, and enable dynamic reasoning [22,23]. By constructing domain-specific ontologies for equipment components, materials, processes, and standards, knowledge graphs facilitate the integration of heterogeneous data sources and provide a unified framework for decision support. In process industries such as chemical manufacturing, petrochemicals, and pharmaceuticals, knowledge graphs underpin systems, enabling real-time monitoring, fault diagnosis, and process optimization through semantic inference and machine-learning–augmented embeddings [24].

Although prior research has made strides in improving dismantling efficiency and environmental sustainability, challenges remain in achieving robustness and generalization across diverse transformer models. This paper aims to develop a standardized, knowledge-driven decision-making [25,26] framework to enhance the accuracy and intelligence of transformer recycling.

The proposed system integrates multimodal perception, knowledge graphs, and optimization algorithms to achieve closed-loop intelligent dismantling:

• During the state recognition phase, sensor data (e.g., gas concentration, resistance) and imaging results are fused and processed using Transformer and EfficientNetv2 architectures for deep feature extraction [27];
• In the knowledge management layer, a Neo4j-based relational knowledge graph is constructed, and TransR with a GAT is applied for entity–relation embedding [25,26];
• For decision execution, NSGA-II multi-objective optimization and reinforcement learning simulate dismantling sequences and generate green-value-optimized plans [28,29].

The main contributions of this work are as follows:

• The development of a composite evaluation system integrating economic and environmental value, with weight optimization for the dismantling sequence decisions;
• Improved accuracy and generalization in equipment status identification and material recovery modeling through knowledge graph reasoning and embeddings;
• The implementation of a scalable, intelligent recycling platform for complex transformers leveraging multimodal sensing.

2. Recovery and Disassembly of Retired Distribution Transformers

Building upon previous research, this study categorizes transformer dismantling technology into three main stages: green recovery of transformer oil, separation of the transformer cover, and sorting of the internal components. Each dismantling step is briefly introduced with the corresponding granularity-specific methods and standardized procedures.

2.1. Recovery of Discarded Transformer Insulating Oil

The recovery of insulating oil from waste transformers can generally be classified into three categories: physical purification, chemical regeneration [30], and thermal regeneration. In this study, the standardized procedure adopts thermal regeneration to achieve complete oil recovery. The process begins with the extraction of the insulation oil from the discarded transformer and continues until the oil flow rate reaches zero. To remove residual oil adhered to internal structures, hot air circulation is applied to vaporize the remaining oil. A mixture of oil vapor and gas is extracted under controlled pressure and temperature conditions until full separation is achieved. Finally, low-frequency induction heating is used to dry and collect the remaining traces of oil, completing the recovery process.

2.2. Dismantling of Transformer Covers

The dismantling techniques for waste transformer covers are categorized based on their connection types: screw-fastened and welded. An RGB-D camera is used to capture both 2D images and 3D point cloud data for the transformer cover. A multimodal fusion algorithm is applied to enhance the robustness and improve the recognition accuracy under complex environmental conditions.

First, the system employs a pre-trained EfficientNetV2 model [27] to automatically identify the cover-to-shell connection type. The model input consists of fused multimodal feature maps, and the output classifies the connection as either screw-fastened or welded.

For screw-fastened covers, the system initially checks whether the screws are stripped or damaged. If such issues are detected, the operation switches to a cutting procedure. Screw localization is carried out using an improved YOLOv8 model enhanced with a Mask Attention mechanism, enabling precise localization of the screw positions. Additionally, a template-matching method based on the shape and distribution characteristics of the embedded nuts is used to initialize the query, further improving detection in complex backgrounds. Once the nut positions are obtained, coordinate data is sent to the robotic arm, which performs the screw removal task.

For welded covers, semantic segmentation techniques are used to generate welding cut lines for robotic processing. The STDC network is utilized to segment the cover and its components, accurately identifying the spatial distribution of each part. Cutting paths are then planned using graphical processing techniques, and SegFormer is employed to generate precise welding cut lines. These lines and planned paths are subsequently transmitted to the robotic arm to execute the cutting operation.

2.3. Dismantling of the Internal Components in a Discarded Transformer

After the transformer’s cover is removed, the primary internal components consist of windings, studs, silicon steel sheets, and clamping parts. For the standardized automated dismantling procedure, semantic segmentation of the internal components is a fundamental step. This study adopts a multimodal semantic segmentation scheme based on the ODIN model (Figure 1), which leverages multi-view RGB-D cameras to collect internal 3D point cloud data combined with depth information [31]. By employing multi-scale feature fusion, the system accurately detects and segments the internal transformer components to ensure high-precision localization and segmentation.

RGB-D cameras are installed on top of and on the four sides of the transformer to capture multi-view RGB images and corresponding depth maps. A neural network is used to extract 2D image features, which are then fused into 3D representations using a k-nearest neighbor algorithm and positional encoding. The precise locations of the internal components are used to generate optimized cutting paths. The system first identifies and removes the screws and copper plates beneath the cover, separating it from the internal structure. Then, it cuts the studs surrounding the windings to detach the clamping plates and windings. Finally, the silicon steel sheets are processed either by slicing or prying them apart to separate them from the windings.

3. A Technical Proposal for Precision Recycling of Distribution Transformers

To maximize the full life-cycle value of power equipment and balance the trade-offs among transformer recycling costs, reuse benefits, and environmental protection requirements, this paper establishes a differentiated classification system for various recyclable materials within transformers. A comprehensive green evaluation and value assessment framework is proposed for the recycling and disposal of electrical equipment. Based on this framework, a refined recycling and disposal workflow is developed to optimize both the environmental and economic benefits in the life-cycle management of power assets.

3.1. Value Evaluation and Metric Design for Recycling Procedures

To enable scientific quantification and dynamic optimization of the recycling value of distribution transformers, this study develops a multi-objective evaluation framework that considers three key dimensions: economic performance, environmental benefits, and technical generalizability.

In the economic dimension, the evaluation focuses on three main indicators: the overall auction value of the transformer, the revenue from the recovery of dismantled metal, and the cost of dismantling. The overall economic value from transformer auctions is determined based on real-world auction examples and expert evaluations. The value is calculated using a comprehensive index system incorporating insulation performance, sealing performance, electrical performance, service duration, visual inspection, failure rate, defect rate, and the original price of the transformer.

Among these indicators, service duration, failure rate, and defect rate exert a negative influence on the economic value; deterioration in these factors leads to a reduction in the auction value. In contrast, the insulation performance, sealing performance, electrical performance, and visual inspection results positively contribute to the assessed value of the transformer. To eliminate the dimensional inconsistency among the performance indicators and ensure their balanced contribution to the overall evaluation, each metric is normalized using a fuzzy membership function [32]. This prevents any single indicator from disproportionately influencing the evaluation results.

For positively correlated indicators, the normalization is defined as follows:

$$R\left(X_i\right)={\displaystyle \frac{X_i-X_{min}}{X_{max}-X_{min}}}$$

(1)

For negatively correlated indicators, the normalization is defined as follows:

$$R\left(X_i\right)={\displaystyle \frac{1-(X_i-X_{min})}{X_{max}-X_{min}}}$$

(2)

Here, $R\left(X_i\right)$ denotes the normalized membership value of each indicator, obtained through min–max normalization. The values of $X_{min}$ and $X_{max}$ correspond to the minimum and maximum values of the indicator, respectively. The result $R\left(X_i\right)\in [0,1]$, where a higher value indicates a better performance for the evaluated metric.

The three key performance indicators—insulation performance, sealing performance, and electrical performance—are obtained through self-designed experimental testing procedures. The experimental protocols were designed in accordance with practical operational scenarios and referred to relevant standards such as GB/T 4776 [33] and IEC 60076 [34], ensuring the consistency, repeatability, and reliability of the performance evaluation.

In the calculation of the evaluation indicators, the maximum and minimum values for each metric are determined based on empirical data and expert recommendations. By setting appropriate bounds, all indicators are normalized to ensure consistency and comparability across different evaluation scenarios.

The overall equipment condition score $\omega$ is then used to determine the baseline auction price through a stepwise grading mechanism. The formula is defined as follows:

$$\omega =\sum _{i=1}^{10}R\left(X_i\right)·{\beta }_i$$

(3)

The weight coefficient ${\beta }_i$ for each evaluation indicator is determined using the entropy weight method, dynamically optimized by integrating expert judgment and the entropy-based objective weighting scheme to derive the optimal solution.

The equipment condition score $\omega$ is used to classify transformers into discrete condition levels, each corresponding to a different starting price ratio, as shown in

Table 1. Classification of the transformer condition levels based on the equipment condition score $\omega$.

Condition Level	Score Interval ($\mathit{\omega }$)	Starting Price Ratio (% of Original Price)
Excellent	$[0.84,1.00]$	80%
Good	$[0.7043,0.84)$	60%
Fair	$[0.4822,0.7043)$	30%
Poor	<0.4822	Recycled via dismantling

.

The assessment of the transformer dismantling value is primarily divided into two components: the gross profit of the overall dismantling process and the net value of each individual dismantling operation.

The gross dismantling value is estimated based on the unit price of recovered metals $P_m$ and the metal content $Q_m$ of each transformer model, as recorded in the database. This provides a coarse-grained approximation of the economic value derived from material recovery.

$$V_{\mathrm{gross}}=\sum _{m=1}^k{Q}_m·P_m$$

(4)

The value of a single dismantling operation can be divided into four major components:

• The value of the dismantled products;
• The labor cost;
• The energy consumption cost;
• The equipment depreciation cost.

The value of the dismantled products is calculated based on the recoverable metal content. Two cases are taken into consideration:

• If the dismantled product contains only one type of metal, its recovery value is computed as

  $$V_p=Q_m·P_m$$

  where $Q_m$ is the quantity of metal m, and $P_m$ is its unit recovery price.
• If the product contains multiple types of recoverable metals, the recovery value is determined by its comprehensive market recovery price $V_m$, i.e.,

  $$V_p=\sum _{m=1}^k{V}_m$$

The labor cost $C_{\mathrm{labour}}$ associated with a single dismantling operation is computed based on the number of workers $N_j$, the unit hourly wage $S_j$, and the operation time $T_j$ for each worker category j. The unit hourly wage $S_j$ is determined by combining the base monthly salary $S_b$ and the dismantling performance bonus $S_p$. To normalize monthly wages, a standard conversion of 21 working days per month and 8 working hours per day is applied.

The labor cost is calculated as

$$C_{\mathrm{labour}}=\sum _{j=1}^n{N}_j·\left(S_b+S_p\right)·T_j$$

(5)

The energy consumption cost $C_{\mathrm{energy}}$ is computed using the total energy consumed E and the unit price of energy $P_e$:

$$C_{\mathrm{energy}}=E·P_e$$

(6)

The cost of equipment wear and depreciation is calculated based on the workload-based depreciation method. The net residual value of the equipment is estimated according to its anticipated salvage price in the secondary market, with a benchmark residual rate set at 5% of the original asset value. The total expected working hours are computed based on the rated service lifespan of the equipment and its average daily operational duration.

The depreciation cost per unit workload is calculated as

$${\mathrm{Depreciation}}_{\mathrm{unit}}={\displaystyle \frac{\mathrm{Original}\mathrm{Equipment}\mathrm{Value}\times (1-\mathrm{Residual}\mathrm{Value}\mathrm{Rate})}{\mathrm{Estimated}\mathrm{Total}\mathrm{Operating}\mathrm{Hours}}}$$

(7)

The equipment cost for a specific dismantling operation is calculated as the product of the unit operating time and the depreciation per unit workload:

$$C_{\mathrm{equipment}}=T_{\mathrm{unit}}\times {\mathrm{Depreciation}}_{\mathrm{unit}}$$

(8)

where $T_{\mathrm{unit}}$ denotes the actual operating time of the equipment for a single dismantling process.

For dismantling operations of varying levels of granularity, the total dismantling cost is obtained by aggregating the costs associated with each constituent dismantling procedure involved.

From an environmental perspective, it remains challenging to comprehensively quantify the full life-cycle recovery indicators of distribution transformers. To partially evaluate the environmental benefits of the recycling process, this study focuses on two primary aspects: the reduction in landfill pollution achieved by dismantling at different levels of granularity and the reduction in carbon emissions due to recycling of net metals, thereby avoiding new resource extraction and processing.

Among transformer waste, the two major hazardous pollutants are used transformer oil and various plastic components. This study establishes evaluation indicators based on the intrinsic properties of these pollutants and their corresponding recycling and disposal strategies.

Used transformer oil contains various toxic substances, such as polychlorinated biphenyls (PCBs). If disposed of through a landfill, these chemicals may contaminate groundwater. If incinerated while adhering to residual metals, it may release dioxins and other carcinogens, posing significant environmental risks.

Meanwhile, discarded transformer plastic insulation components often contain flame retardants and other chemical additives. These materials are difficult to degrade naturally and are prone to releasing microplastics, leading to ecological disruption and contributing to “white pollution”.

To evaluate the green benefits of managing these two major sources of pollution, an emission reduction effectiveness approach is employed. This involves quantifying the decrease in pollutant leakage, assigning hazard grades based on the pollution severity, and applying weighted adjustments according to the reliability of pollutant recovery channels. This ensures that hazardous waste is directed, as much as possible, to specialized treatment facilities.

The environmental benefit index $G_{\mathrm{eff}}$ is calculated as follows:

$$G_{\mathrm{eff}}=\sum _i(C_i·R_i·(1+D_{\mathrm{loss}})·m$$

(9)

$C_i$, $R_i$, $D_{\mathrm{loss}}$, and m denote the pollutant concentration, the environmental risk factor, the degradation level, and the gain from recycling channels, respectively.

Regarding the carbon emission benchmark for metal recycling, since the carbon intensity of the primary metal production processes (including mining, smelting, and refining) is significantly higher than that of secondary (recycled) metal processing, the net content of recycled metals can be used to evaluate the reduction in carbon emissions at a given decomposition granularity. Furthermore, due to the varying carbon emission reduction potentials of different metals, this method provides a relatively accurate assessment of the impact of the decomposition granularity on the total carbon emissions.

To account for the green value loss during recycling operations, it is quantitatively analyzed by converting the loss into an equivalent increase in carbon emissions.

The carbon emissions function is given by

$$f_{{\mathrm{CO}}_2}=\sum \left({\beta }_i·{\epsilon }_i·q_i·V_i·R_i\right)+F$$

(10)

where ${\beta }_i$, ${\epsilon }_i$, $q_i$, $V_i$, $R_i$, and F denote the carbon emission coefficient, the standard conversion factor, the total consumption per unit operation, the CO₂ equivalent conversion factor, the metal content, and the gas leakage volume, respectively.

The technical versatility index is constructed based on three dimensions: equipment compatibility, standardization of the process flow, and operational complexity. A hierarchical scoring method is employed for quantitative evaluation.

Equipment compatibility encompasses two aspects: coverage of the transformer types and tool adaptability. The transformer type coverage is calculated as the ratio of the number of transformer models that can be processed to the total number of mainstream market models, multiplied by 100%. Tool adaptability is assessed based on the proportion of specialized tools used.

Process flow standardization is evaluated by the degree of modularization of operational steps.

Operational complexity is assessed through two sub-indicators: the technical requirements for dismantling personnel and dismantling efficiency.

Each sub-indicator is rated using a five-point Likert scale. The final technical versatility score is calculated by applying weighted aggregation across all sub-indicators.

Furthermore, all three indicators can be uniformly evaluated from an economic perspective. Green value is quantified by introducing pollution control costs, while technical versatility can be converted into economic value through reductions in equipment investments and training costs.

3.2. The Transformer Recycling Decision System

To achieve high technical versatility, this paper proposes a standardized disassembly process based on the current technologies for end-of-life transformer dismantling and recycling. In this framework, the disassembly granularity of discarded transformers is transformed into a binary decision structure, indicating whether or not to execute each defined process module. The standard disassembly procedure is divided into six functional modules: transformer oil recovery; separation and treatment of the transformer cover plate; detection and disassembly of the internal components; group cutting of the transformer windings; fine disassembly of the winding coils; and the generation of disassembly reports.

To achieve synergistic optimization of both economic and environmental benefits in the transformer recycling process, this study proposes a data-driven dynamic disassembly decision-making workflow. The core of the framework comprises four functional modules:

• Multi-source data fusion;
• State-adaptive assessment;
• Dynamic adjustment of the disassembly strategies;
• Closed-loop model optimization.

This framework enables the system to respond intelligently to heterogeneous transformer conditions, external economic parameters, and environmental constraints, thereby supporting more efficient and sustainable disassembly decisions in real time.

3.2.1. Multi-Source Data Fusion

To support subsequent transformer value assessments and recycling decision-making, the system collects three categories of data through sensors, historical databases, and market monitoring platforms:

• Transformer intrinsic data: This includes both static parameters and real-time monitoring data, such as the transformer model, years of operation, the dielectric strength of the insulating oil, and performance test results.
• Environmental and economic data: This covers external dynamic parameters, including fluctuations in metal recovery prices, the average market prices of transformer components, hazardous waste disposal costs, and green benefit metrics.
• Process parameters: These record operational data during disassembly, such as equipment compatibility scores and process time efficiency.

After acquiring the intrinsic data, heterogeneous information is subjected to data cleaning, normalization, and feature selection. This process yields a feature vector that effectively characterizes the operational condition of the transformer.

3.2.2. Condition Assessment and Disassembly Path Planning

To prevent the circulation of substandard transformers back onto the market, whole-unit transformer recovery must be based on verified performance compliance. A pre-trained classification model is employed to assess a transformer’s condition and estimate its remaining service life. This evaluation determines whether a transformer is suitable for direct market reuse.

For transformers deemed eligible for whole-unit recovery, the equipment condition index is updated using the entropy weight method, generating a composite weight factor $\omega$ to support transformer auction value estimation.

For dismantled transformers, a knowledge graph is queried based on transformer model information to retrieve the optimal historical disassembly paths. Each path consists of a combination of standardized disassembly procedures.

The standardized disassembly procedures are as follows:

• Insulating oil recovery and treatment: A self-developed insulating oil recovery device is used to recover waste oil from the transformer, producing a de-oiled and dried transformer. The recovery time is selected as an optional process parameter based on the recovery time and the efficiency curve.
• Transformer housing disassembly: Transformers are classified into two types based on the cover plate fastening method. For screw-fastened cover plates, a screw position identification diagram is generated according to the method in Section 2.2, and a robotic arm equipped with a hex screwdriver removes the screws to obtain the cover and screws. For welded screw cover plates, the system generates a component recognition diagram and cutting lines, and a laser cutting device performs the cutting.
• Cover plate segmentation: The system uses object detection technology to accurately identify internal components; cuts the screws and copper plates beneath the cover; and separates the cover from the transformer interior.
• Stud separation: The windings surrounding the transformer studs are cut, and the studs are separated and classified along with the tap changer.
• Silicon steel sheet separation: A robotic arm first lifts the upper silicon steel sheets and then cuts the separated winding to access the lower silicon steel sheets.
• Winding insulation removal: A wire-stripping device is used to separate the metal conductor from the insulation.
• Fine processing of the transformer conductors: A wire-stripping device is used to separate copper wire from other metals, obtaining relatively pure copper.

The constructed disassembly path utility function balances the maximization of the net economic value and the environmental benefits of the disassembly process:

$$U={\lambda }_1·V_{\mathrm{enc}}+{\lambda }_2·V_{\mathrm{env}}$$

(11)

where ${\lambda }_1$ and ${\lambda }_2$ are weighting coefficients, determined through expert scoring and fitted based on the external carbon value and pollution treatment costs.

3.2.3. Dynamic Adjustment of the Disassembly Strategies

To achieve high efficiency and maximize the resource recovery in the transformer disassembly process, a real-time data-driven dynamic optimization framework is introduced. This framework enables online decision adjustments guided by multimodal sensing.

Multimodal sensing is primarily realized through the deployment of a sensor network at key nodes of the disassembly system. For the transformer oil recovery unit, an oil quality monitoring module is deployed, which utilizes near-infrared spectrometers and electrochemical sensors to monitor the real-time oil quality and prevent contamination leakage. For internal transformer disassembly, a component condition monitoring unit is deployed, incorporating industrial cameras and laser displacement sensors to track the status of the components during disassembly, including the stacking state of silicon steel sheets and the deformation level of coils. Additionally, power meters are installed to record and monitor the energy consumption of disassembly equipment in real time, detecting abnormal consumption.

Based on real-time internal data from the transformer, the disassembly process parameters are dynamically adjusted to achieve the optimal disassembly performance.

To avoid sequence conflicts and equipment congestion in the disassembly system, real-time planning of the disassembly path is implemented. Furthermore, if the system detects a timeout in any operation, it generates an exception report and triggers parallel disassembly tasks to ensure optimal disassembly efficiency.

3.2.4. Closed-Loop Model Optimization

After disassembly is completed, the system records the cost of each disassembly operation, the types of materials recovered, and the associated profits. It then calculates the deviation between the actual recovery value and the estimated recovery value. Disassembly operations with deviations exceeding the predefined threshold are flagged for root cause analysis. If the deviation is caused by material misjudgment, reverse tracing is performed for the corresponding transformer to support quality supervision.

Leveraging the generalization capability of large-scale models, disassembly reports are generated based on the collected disassembly data, accompanied by auxiliary suggestions for process improvements.

To ensure continuous improvement in the overall transformer recycling decision system and maintain compliance with evolving policies and technological standards, an incremental knowledge graph expansion mechanism is employed. Triplets are automatically extracted from disassembly reports and newly issued policies. In the event of a detected conflict between new and existing knowledge, a manual review process is triggered.

4. The Fine-Grained Transformer Recycling Decision Algorithm and Experiments

Based on the overall design of the fine-grained transformer recycling decision system, this paper proposes a decision-making algorithm framework that integrates performance evaluation, value optimization, disassembly optimization, and knowledge iteration.

4.1. The Whole-Transformer Recovery Evaluation Algorithm and Experiments

Based on an extensive analysis of the literature and the operational conditions of equipment in power material enterprises, this study proposes a reverse logistic value evaluation model for electrical equipment. For the evaluation of the whole-transformer recovery, the assessment framework consists of six primary indicators: insulation performance, sealing performance, electrical performance, service duration, visual inspection, failure rate, and defect rate.

Insulation performance: For transformers, insulation performance directly determines operational safety. Understanding a transformer’s insulation condition is crucial for estimating its remaining service life. Good insulation ensures that the transformer still meets the basic requirements for operation and prevents insulation failure, which could result in fire hazards. Insulation performance reflects the aging degree of insulation materials, and aging in transformers is irreversible. For example, a 100-point decrease in the degree of polymerization (DP) of insulation paper leads to a 50% reduction in mechanical strength and a 300% increase in repair costs, severely impacting auction value and recovery safety. To evaluate insulation performance, five experimental tests are conducted:

• An insulation resistance test of the core and clamping parts;
• An insulation paper test;
• An insulation resistance test of the windings and bushings;
• A dissolved gas analysis (DGA) of insulating oil;
• An analysis of the furfural content of insulating oil.

The insulation resistance test of the core and clamping parts assesses the insulation state between the core and the ground, identifying risks of metallic short circuits. This test follows the national standard GB/T 6451 [35]. For transformers rated 66 kV and above, the resistance must exceed 100 MΩ. If it is below 50 MΩ, whole-unit recovery is prohibited.

The insulation paper test primarily evaluates the DP value of the insulation paper, which directly reflects its mechanical strength and chemical aging. Insulation paper is categorized into three levels:

• Excellent: A DP > 400;
• Qualified: A 200 ≤ DP ≥ 400;
• Unqualified: A DP < 200.

The insulation resistance $R_{60}$ test for windings and bushings primarily reflects the overall insulation degradation of the winding, as well as moisture ingress into the bushings. According to the national standard GB 50150 [36], the measured resistance value should not be less than 70% of the value obtained during the previous test cycle.

The dissolved gas analysis (DGA) of the insulating oil is employed to detect latent electrical discharges and thermal faults. The analysis mainly focuses on the concentrations of total hydrocarbons, hydrogen, and acetylene. Acceptable thresholds for operational safety are defined as follows: the total hydrocarbon and hydrogen concentrations should be less than 150 ppm, while the acetylene concentration should not exceed 1 ppm.

The content of furfural compounds in insulating oil serves as a key indicator for predicting the remaining service life of a power transformer. When the furfural concentration is below 1.5 mg/L, the transformer is considered to have a relatively long residual lifespan and to thus be in excellent condition. In contrast, concentrations exceeding 4 mg/L are indicative of severe paper insulation aging and an insufficient remaining life, warranting transformer decommissioning.

The Insulation Health Index (IHI) is computed using a neural network model. The model takes as input a five-dimensional feature vector comprising normalized results from four diagnostic tests:

$$\mathbf{x}=\left[{\displaystyle \frac{DP}{700}},{\displaystyle \frac{R_{\mathrm{core}}}{1000}},{\displaystyle \frac{1}{{log}_{10}(\mathrm{Furan}+1)}},{\displaystyle \frac{R_{60}}{R_{\mathrm{standard}}}}\right]$$

(12)

A multilayer perceptron (MLP) is employed to compute dynamic weights for these four experimental indicators. By fusing the features, the model generates the IHI score and adaptively learns the relative importance of each indicator in evaluating the insulation performance. Specifically, the neural network comprises three fully connected layers with 128, 64, and 32 neurons, respectively, all activated using the ReLU function. The model is trained using the Adam optimizer with an initial learning rate of 0.001 for 100 epochs and a batch size of 32. To prevent overfitting, a Dropout layer with a drop rate of 0.5 is introduced. Additionally, an early stopping strategy is applied by monitoring the validation loss during training to enhance the model’s generalization capability and overall performance.

Sealing performance: Sealing performance characterizes the capability of power equipment to maintain operational stability under varying environmental conditions. Inadequate sealing may lead to frequent maintenance and component replacement; increase the risk of mechanical wear and corrosion; and potentially result in the leakage of insulating oil and harmful gases. Such leakages not only compromise equipment reliability but also pose environmental risks, contravening the principles of green and sustainable operation. Sealing integrity is evaluated through an oil column pressure test, where the oil level is raised to 0.6 m above the top of the conservator. If no leakage is observed after 12 h, the sealing performance is considered excellent.

Electrical performance: Electrical performance is a core indicator of a transformer’s condition and remaining asset value. High electrical performance suggests strong potential for reuse, helping to avoid premature decommissioning due to technological obsolescence. Moreover, transformers with superior electrical characteristics are more competitive in the secondary market and retain higher residual value. In this study, electrical performance is assessed using three diagnostic methods: the induced voltage withstand test, partial discharge (PD) measurements, and a winding direct current (DC) resistance test.

• The induced voltage test assesses the dielectric withstanding capacity of the transformer under elevated electric field stress and is evaluated by comparing the results with factory acceptance test data.
• Partial discharge measurements utilize phase-resolved partial discharge (PRPD) patterns to identify the type and severity of discharge activity. Devices exhibiting internal discharge defects are subject to further disassembly and inspection.
• Winding DC resistance measurements are corrected to a standard reference temperature of 20 °C to ensure accuracy and comparability.

Service duration: The service duration of a transformer is a critical parameter reflecting its degree of aging. Prolonged operation typically corresponds to increased mechanical degradation and a reduced functional performance, thereby lowering the residual value. Moreover, the service duration provides essential input for expert systems in estimating the remaining useful life of a transformer. To facilitate standardized assessments, the rated service life of the transformer is used as the upper bound for this parameter.

Visual inspection: Visual inspection provides a direct and intuitive assessment of the physical condition of a transformer, including visible signs of mechanical damage, surface corrosion, wear, and aging. Such defects also indirectly reflect the quality of the transformer’s historical maintenance and may influence both repair costs and operational safety. In this study, a multi-view 3D imaging system is employed to acquire comprehensive exterior images of the transformer. A modified YOLOv8 deep learning model is applied to perform defect detection, outputting the size and type of each detected defect. The output of the model is structured as

$$\mathrm{Output}=\left\{(x,y,w,h,p)\right\}$$

(13)

where x, y represent the center coordinates of the defect; w, h denote the width and height; and p indicates the confidence score associated with the defect classification. The detailed labels used for classification are shown in

Table 2. Defect category labels.

Label ID	Defect Type	Risk Level
0	Surface Cracks	High
1	Local Corrosion	Medium
2	Oil Leakage Traces	Medium
3	Mechanical Deformation	Very High

.

The final appearance inspection score is computed by integrating the defect type, geometric characteristics, and associated risk levels. This multi-factor evaluation enables a more accurate assessment of the visual condition of the transformer.

Once individual performance indicators have been scored through the corresponding diagnostic tests, a comprehensive transformer condition index is computed using a weighted aggregation model. The overall condition score $\omega$ is defined as follows:

$$\omega =\sum _{i=1}^{10}R\left(X_i\right)·{\beta }_i$$

(14)

The initial weights ${\beta }_i$ are assigned based on expert knowledge and recommendations. The initial weight values for each evaluation indicator are listed in

Table 3. Initial weight values of evaluation indicators.

Indicator Name	Initial Weight ${\mathit{\beta }}_{\mathit{i}}$
Insulation Performance	0.25
Sealing Performance	0.10
Electrical Performance	0.15
Visual Inspection	0.05
Service Duration (years)	0.20
Failure Rate (events/year)	0.15
Defect Rate (%)	0.10

. Subsequently, the weights are dynamically updated and optimized using the entropy weight method to enhance the objectivity and robustness of the evaluation model.

Experimental Results of Whole-Unit Recovered Transformers

To validate the practical applicability and effectiveness of the proposed evaluation model, this study selected 50 whole-unit recovered transformers as research subjects. A systematic testing and assessment campaign was conducted, wherein each transformer underwent a quantitative evaluation across seven key indicators: insulation performance, sealing performance, electrical performance, service duration, visual inspection, failure rate, and defect rate.

In the experiment, the evaluation scores for each indicator were first normalized and then fed into the comprehensive assessment model. Based on the initial expert-defined weights ${\beta }_i$ and the dynamic updating mechanism provided by the entropy weight method, the overall condition score $\omega$ for each transformer was calculated. A subset of the experimental results is presented in

Table 4. Assessment results for selected recovered transformers.

ID	Insul.	Electr.	Seal.	Appear.	Life (Years)	Fail (/Year)	Defect (%)	$\mathit{\omega }$	Condition
T1	0.88	0.75	0.92	0.60	15	0.12	3.5	0.812	Good
T2	0.45	0.50	0.70	0.40	28	0.30	7.8	0.523	Fair
T3	0.25	0.35	0.50	0.30	32	0.45	10.2	0.341	Poor

below:

The evaluation results indicate that transformers with a composite score above 0.75 are generally in good condition and exhibit high potential for reuse. Those with scores between 0.5 and 0.75 are considered to be in fair condition, with certain components potentially requiring repair or replacement. In contrast, transformers with scores below 0.5 are assessed to be in poor condition and are not recommended for direct reuse.

The findings are summarized as follows:

• Whole-unit recovery decision accuracy: The proposed model achieved a recovery decision accuracy of 98% in determining whether a transformer could be reused as a whole unit.
• Performance indicator evaluation: The average deviation between the automated performance assessment model and expert evaluations across key indicators was less than 5%, demonstrating high consistency;
• Auction value prediction: The predicted auction prices using the proposed method exhibited the lowest mean absolute error compared to the actual transaction prices, with an 8.7% reduction in error relative to that in expert-based valuation.

The assessment outcomes show strong alignment with the experts’ conclusions from on-site inspections, thereby validating the effectiveness and engineering applicability of the proposed evaluation framework. These findings demonstrate that the methodology enables accurate and reliable assessment of transformers’ condition; supports informed recovery path decisions; and facilitates rational and data-driven auction pricing. Consequently, it contributes to the dual optimization of both the economic and environmental value of recoverable transformers.

4.2. The Transformer Disassembly Decision Algorithm

For transformers identified as disassembly-type, this study proposes a decision-making algorithm based on knowledge graph reasoning. The algorithm takes the transformer model and a disassembly objective—either economic value maximization or pollution minimization—as the input. A set of disassembly process candidates is retrieved through subgraph matching, whereby the SPARQL query algorithm is used to extract subgraphs from the knowledge base associated with the given transformer model.

These disassembly knowledge subgraphs contain information such as the component structure, material composition, and historical disassembly cases. Based on these historical records, a set of candidate disassembly processes is generated:

$$\mathrm{P}=\left\{P_1,P_2,\dots ,P_n\right\}$$

(15)

In the value evaluation logic chain of the dismantling process, we define the state variables as follows: the dismantling stage t, the set of remaining components $C_t$, and the accumulated value $V_t$; the dismantling granularity $P_t$ is defined as the decision variable. This model allows for the flexible incorporation of various constraint conditions to impose operational and safety restrictions on the transformer recycling process, thereby ensuring compliance with practical and regulatory standards.

For each generated disassembly value evaluation logic chain, a multi-dimensional value evaluation index system is introduced and formulated as an objective function:

$$max\sum _{t=1}^T({\lambda }_1·V_{enc,t}+{\lambda }_2·V_{env,t})$$

(16)

Based on this framework, this study adopts the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) to optimize multi-objective functions. The objective functions jointly consider economic value and environmental value.

NSGA-II enhances the convergence speed and the global search capability by employing fast non-dominated sorting and crowding distance computation, effectively maintaining the population diversity and balancing conflicts among objectives. It outputs a set of Pareto-optimal solutions, from which decision-makers can select the most appropriate dismantling strategy based on their practical requirements.

The algorithm parameters are configured as follows: population size = 150, the maximum number of generations = 250, crossover probability = 0.85, and mutation probability = 0.05. The selection operator is tournament selection based on crowding comparison, the crossover operator is Simulated Binary Crossover (SBX), and the mutation operator is polynomial mutation [37].

The disassembly process is demonstrated and monitored in real time. For each disassembly operation, the system outputs detailed information including the single-metal disassembly product and its corresponding disassembly value; the manual labor cost associated with the individual disassembly step; the energy consumption and equipment wear, with the latter calculated based on asset depreciation; and the cumulative disassembly value and the total cost up to the current process stage.

This real-time monitoring and evaluation framework enables precise tracking of the resource utilization and economic performance throughout the disassembly life cycle.

Experimental Validation and Comparative Analysis

During the experimental process, waste transformer insulating oil was recovered using a self-developed comprehensive insulating oil recovery system. This system comprises four functional modules: waste oil extraction, residual oil vaporization, exhaust gas treatment, and online monitoring. A schematic diagram of the system’s structure and its physical implementation is provided in Figure 2 and Figure 3.

The waste oil extraction module performs high-flow-rate oil removal from the upper interface of the transformer, while simultaneously extracting residual oil from the bottom interface until the flow sensor indicates a residual oil flow rate of zero. The residual oil vaporization module utilizes a hot air blower to vaporize internal residual oil under constant-pressure and -temperature conditions for one minute, after which a vacuum pump is activated to extract the oil–gas mixture. This process is repeated until the oil–gas saturation no longer changes, at which point low-frequency induction heating under a negative pressure is applied.

The exhaust gas treatment module cools, filters, and adsorbs the oil–gas mixture until the exhaust emissions meet the regulatory standards. The online monitoring module, which includes pressure and temperature sensors and an oil–gas composition analyzer, employs a neural-network-based saturation analysis algorithm to monitor the residual oil recovery status. It regulates the power output accordingly to prevent damage to the internal components.

A total of 50 transformer samples across three model types were selected for the experiment. Based on prior experimental tests, relevant transformer data were obtained. The evaluation metrics included the technical feasibility of disassembly, the economic value of recovery decision-making, environmental (green) value, and component safety.

The average time required for complete recovery of the insulating oil from the 50 transformers was 132.1 min. With regard to the residual oil, only 24 units had measurable remaining insulating oil, resulting in an overall residual oil rate of less than 1%. The success rate for dismantling transformer covers and internal components reached 98%, and no damage to the core recovery components was observed in any of the successfully completed dismantling experiments.

From the perspective of economic value, the proposed dismantling decision-making method resulted in a metal recovery rate of 94.2%, representing an average 18% increase in economic value compared to that under the conventional dismantling methods.

In terms of environmental (green) value, the proposed dismantling decision-making approach achieved a 100% pollutant recovery rate, enabled complete recovery of waste oil, and yielded a 6.8% average improvement in the carbon recovery value.

Regarding component safety, a full life-cycle reverse logistic management strategy was implemented for all dismantled transformer parts, effectively preventing the issue of an uncertain component flow commonly associated with traditional dismantling practices.

4.3. Knowledge Graph Construction and Updating

To enable knowledge-driven and dynamically optimized disassembly decision-making for transformers, this section proposes a multi-source data-fusion-based methodology for constructing and incrementally updating a domain-specific knowledge graph.

The construction of the knowledge system begins with an analysis of the industry characteristics, categorizing the core elements involved in power equipment disassembly into three main types:

• Equipment-related entities (e.g., transformers and internal components such as cores and windings);
• Material entities (recording key parameters such as metal density and recycling value);
• Process entities (covering the tool selection, disassembly steps, and environmental compliance requirements).

The knowledge graph establishes the relationships between transformer models and the optimal disassembly procedures, annotates hazardous substance attributes of components, tracks variations in the recycling efficiency across different processes, and assigns temporal labels to each data node. This results in an inferable knowledge network capable of supporting decision automation.

The integration of multi-source data presents several challenges: First, heterogeneous data types must be handled, including structured databases, unstructured documents, and real-time sensor streams; second, information conflicts must be resolved—for example, discrepancies in the disassembly methods for the same model reported in different sources. To address these issues, a three-level conflict resolution strategy is implemented: newly acquired data takes precedence over outdated information. Source authority is ranked in the order of national standards > peer-reviewed literature > enterprise records. Regional policy differences are considered during interpretation.

To address the extraction of information from unstructured technical documents, this study proposes a method based on Large Language Models (LLMs) utilizing prompt engineering. By constructing domain-specific prompt templates, the model is guided to automatically identify and extract key transformer parameters—such as the model number, rated capacity, voltage level, cooling method, and winding materials—as well as technical process keywords, including vacuum oil filling, drying treatment, and insulation testing. The extracted results are then structured into standardized JSON format.

To improve the extraction accuracy and consistency, few-shot prompt learning is incorporated to fine-tune the LLM on transformer-related tasks. Additionally, a domain terminology dictionary is integrated to enhance the model’s knowledge and understanding. On top of the automatic extraction, a manual verification mechanism is introduced to review and refine the results, balancing intelligent processing efficiency with the rigor of human oversight.

The current prototype knowledge graph comprises approximately 3800 entities and 12,500 relationships, with an average node degree of 3.4. All information is stored in the form of RDF triples (subject–predicate–object), enabling efficient semantic retrieval and reasoning tasks. To enhance the interpretability and reasoning capacity of the knowledge graph, the system employs the TransR (Translation on Relation) embedding algorithm. This method maps entities and relations into distinct vector spaces and utilizes a relation-specific projection matrix to learn more accurate entity associations.

To ensure the timeliness of data and the reliability of decision-making within the knowledge graph, each node is assigned a timestamp attribute. During graph utilization, the system continuously calculates the “freshness” of the node information and dynamically adjusts its weight using an exponential decay function. Nodes that have not been updated for an extended period are assigned lower weights to minimize the influence of outdated information on the decision logic.

The exponential decay function is defined as follows:

$$w_{\mathrm{eff}}=w_0·e^{-\lambda (t_{\mathrm{now}}-t_{\mathrm{update}})}$$

(17)

where $w_{\mathrm{eff}}$ is the current effective weight of a node, $w_0$ is the initial weight, $t_{\mathrm{now}}$ denotes the current time, $t_{\mathrm{update}}$ represents the last update time of the node, and $\lambda$ is the decay factor.

This mechanism effectively reduces the impact of stale information on the reasoning process and enhances the adaptability and relevance of the knowledge graph in dynamic decision-making scenarios.

An event-driven incremental update mechanism enables the system to respond dynamically to external changes. Upon the detection of regulatory updates, the deployment of new disassembly tools, or fluctuations in metal prices exceeding 5%, the system automatically triggers targeted subgraph reconstructions.

Changes in environmental regulations immediately update the regulatedBy tags of affected process nodes, ensuring compliance alignment. New tool entries establish canBeDisassembledWith relations with the corresponding components. Meanwhile, price changes directly revise the hasEconomicValue field associated with the material nodes.

The update engine recalculates the dependency paths and performs partial graph merges and reasoning operations to synchronize the knowledge graph with the evolving operational context.

By digitizing domain expertise and visualizing the flow of data across entities and processes, the system overcomes the inefficiencies of traditional recycling workflows that heavily rely on manual expertise. This is further reinforced by a continuous learning mechanism, allowing the system to evolve and become more intelligent over time. This self-adaptive capability establishes a solid technical foundation for developing a green, efficient, and intelligent resource recovery ecosystem.

As shown in Figure 4, the transformer model node connects to downstream process and material valuation nodes through multiple semantic relations such as hasComponent, containsMaterial, and requiresProcess.

This subgraph consists of six functional node types: devices (yellow), components (light blue), materials (red), processes (light green), tools (pink), and policy/standard documents (gray). Directed edges represent semantic relationships: hasComponent links devices and components, containsMaterial links components to materials, requiresProcess connects components with processes, canBeDisassembledWith maps components to tools, and regulatedBy connects processes to regulations. Materials also point to economic evaluation nodes via hasEconomicValue.

This visualized subgraph enables an intuitive understanding of the core entities and their interactions within the knowledge management layer, supporting downstream tasks such as semantic querying, embedding training, and decision reasoning.

This structure can be extended to cover the complete ontology comprising 18 entity types and 32 relation types for full-scale knowledge representation.

5. Intelligent Decision Support System Design

5.1. The System Architecture

The intelligent decision support system for transformer recovery proposed in this study adopts a four-layer collaborative architecture that integrates data acquisition, knowledge management, intelligent decision-making, and human–machine interaction, thereby forming a dynamically optimized decision-making loop. The design of the system architecture strikes a balance between technical depth and user-friendliness, with each layer fulfilling its specific functions under a coordinated mechanism.

The data perception layer serves as the foundational component of the system, responsible for the real-time acquisition and standardized processing of multi-source heterogeneous data. On the equipment side, high-precision sensor arrays are deployed: a high-performance liquid chromatograph continuously monitors the furfural concentration in the insulating oil, a digital megohmmeter dynamically tracks changes in the insulation resistance, and a 51.2-megapixel industrial camera performs multi-angle scanning to construct 3D surface models of the transformer, capable of identifying micro-cracks as small as 0.1 mm. On the market and environmental side, the system builds a centralized data platform that connects in real time to APIs from metal future exchanges to monitor fluctuations in copper and aluminum prices, and it extracts updates from the Ministry of Ecology and Environment’s regulatory database to track changes in environmental standards. This results in a comprehensive, multi-dimensional data pool encompassing equipment condition, market trends, and policy constraints. Data processing is carried out using intelligent cleaning strategies. A sliding window algorithm is used to detect abnormal sensor values, and missing data is predicted and filled using an XGBoost model. Finally, feature engineering techniques are applied to generating standardized decision feature vectors.

The knowledge management layer is built on a Neo4j graph database, which supports the construction of a dynamically evolving domain knowledge graph. This graph includes 18 categories of entities and 32 types of relationships. Key technologies include an attention-mechanism-based entity alignment algorithm to resolve naming conflicts in multisource data, a spatio-temporal-aware knowledge update model to trigger automatic revision upon policy changes, and a hybrid embedding method combining TransR and a GAT to improve the process recommendation accuracy.

The decision analysis layer serves as the system’s intelligence core and adopts a multi-model collaborative decision-making mechanism. The equipment health assessment module integrates seven indicators, including the insulation health index (IHI) and the sealing health index (SHI), using the entropy weight method and outputs a visual radar chart of the health rating. The value prediction engine is built on a meta-learning framework that automatically switches between regression models when market fluctuations exceed a set threshold, ensuring a residual value prediction error remains within ±8%. The disassembly optimization system combines the NSGA-II algorithm with deep reinforcement learning to simulate tens of thousands of disassembly paths and identify the solution that produces the optimal economic and environmental results under hard environmental constraints.

The application service layer creates an immersive decision-making environment. It uses a three-dimensional visualization interface to present holographic images of equipment, supporting touch-based rotation and transparent inspection of defective areas. The intelligent reporting system innovatively applies knowledge-graph-based narrative generation techniques. A decision simulation platform offers sandbox functionality, where users can drag environmental, economic, and technical sliders to observe the real-time changes in the value curves of different disposal strategies. The system is fully integrated with ERP systems through standardized APIs, enabling disassembly instructions to be transmitted directly to robotic controllers on site.

5.2. The Human–Machine Interaction Interface

The application service layer designed in this study displays key decision-making parameters throughout system operation. The data perception interface (see Figure 5) first presents the detection data by individual indicators, offering an intuitive basis for subsequent decision support (see Figure 6). The system outputs real-time values for each key performance metric of the transformer (see Figure 7), along with the corresponding decision recommendations.

For the disassembly process, the system provides stepwise outputs of the operational parameters for each procedure while automatically generating and storing work orders to ensure traceability and post-process auditing.

6. Conclusions and Future Work

This paper proposes an automated transformer recycling decision-making system based on multimodal perception and a knowledge subgraph comparison. By constructing standardized dismantling workflows and a comprehensive power equipment knowledge graph, the system demonstrates a strong performance in optimizing the value recovery and ensuring regulatory compliance throughout the transformer recycling process. The experimental results show that the system achieves high recovery success rates and operational safety while significantly enhancing the transformer reuse efficiency and economic value. Additionally, it enables full-chain traceability and management of power equipment recovery, ensuring the security of reverse logistics in the power grid domain.

In terms of its practical applicability, the system adopts lightweight multimodal feature extraction models and efficient knowledge graph querying mechanisms, allowing for deployment on industrial-grade servers in cloud environments without the need for high-performance computing resources. The modular architecture also supports scalability to a wide variety of transformer models and materials. By updating the knowledge base with new entity types, material attributes, and standardized dismantling workflows, the system can be flexibly extended to support diverse transformer types at a low adaptation cost.

To address the integration challenges in real-world applications, such as data heterogeneity, sensor precision limitations, and compatibility with the existing disassembly workflows, small-scale pilot tests are currently underway. These trials aim to validate the system’s robustness in industrial environments and guide subsequent iterative development.

Future work will focus on improving the adaptability of the system to emerging recycling technologies further and refining the knowledge graph update mechanisms. This includes enhancing semantic reasoning capabilities and optimizing the decision strategies under evolving regulatory and economic contexts.