Multi-modal data-driven intelligent early warning method for charging operation of aviation equipment component logistics AGV

Abstract

The application discloses a multimodal data-driven aviation equipment component logistics AGV charging operation intelligent early warning method, comprising the following steps: acquiring multimodal charging data according to the real-time interface data of the intelligent guided vehicle AGV vehicle-mounted system and the charging pile; constructing and training a PINN model fused with battery physical constraints according to the multimodal charging data, wherein the PINN model takes charging time, battery temperature, AGV cumulative running time and battery SOC as input and outputs expected charging current; acquiring expected charging current and prediction uncertainty index according to the real-time prediction result of the current charging process of the trained PINN model; acquiring an adaptive early warning threshold based on the deviation of the expected charging current and the measured charging current and combining the prediction uncertainty index; and acquiring graded early warning information according to the comparison result of the deviation and the adaptive early warning threshold.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and aviation equipment support technology, and in particular relates to a multimodal data-driven intelligent early warning method for charging operations of AGVs in aviation equipment component logistics. Background Technology

[0002] As the digitalization of the aviation equipment manufacturing and support system continues to improve, intelligent guided vehicles (AGVs) are being used more and more widely in the logistics of aviation equipment components. During operation, AGVs need to go to charging piles to recharge and replenish their energy according to the actual situation. The safety and stability of their charging operations directly affect the cycle time of aviation production lines. This leads to the development of intelligent monitoring technology for the charging operations of AGVs in aviation equipment component logistics driven by multimodal data.

[0003] Currently, the AGV charging detection commonly used in industrial scenarios mainly relies on local data provided by the battery and simple threshold rules for early warning. These methods have obvious shortcomings, such as limited data dimensions, lack of physical constraints, and difficulty in achieving real-time predictive judgment of the charging process. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a multimodal data-driven intelligent early warning method for charging operations of AGVs used in the logistics of aviation equipment components. This method enables a novel intelligent monitoring system for charging operations that allows for real-time deviation analysis of the expected charging status from the actual measured values.

[0005] To achieve the above objectives, this invention provides a multimodal data-driven intelligent early warning method for charging operations of AGVs used in the logistics of aviation equipment components, comprising: Multimodal charging data is obtained based on the real-time interface data of the intelligent guided vehicle (AGV) onboard system and the charging pile. Based on the multimodal charging data, a PINN model that integrates battery physical constraints is constructed and trained. The PINN model takes charging time, battery temperature, AGV cumulative running time and battery SOC as inputs and outputs the desired charging current. Based on the real-time prediction results of the current charging process from the trained PINN model, the expected charging current and prediction uncertainty index are obtained. Based on the deviation between the expected charging current and the measured charging current, and combined with the prediction uncertainty index, an adaptive early warning threshold is obtained. Based on the comparison between the deviation and the adaptive warning threshold, graded warning information is obtained.

[0006] Optionally, based on real-time interface data between the intelligent guided vehicle (AGV) onboard system and the charging pile, multimodal charging data can be obtained, including: By interfacing with the AGV onboard system and charging pile data interface, the charging time, battery temperature, AGV cumulative running time, and battery SOC of the AGV are collected in real time to obtain the multimodal charging data.

[0007] Optionally, the battery physical constraints include: battery SOC kinetic differential constraints, constant voltage-constant current charging law constraints, constant voltage-constant current charging law constraints, charging safety current limit constraints, and current curve smoothness and SOC monotonicity constraints.

[0008] Optionally, a total loss function is constructed based on the battery physical constraints, and the PINN model is trained using the total loss function; The total loss function for: ; in, The weights of each physical constraint, Historical charging data collected when the AGV is in a healthy state. For SOC dynamic differential constraints, Constraints on the constant voltage-constant current charging law, Constrained by the temperature-internal resistance-voltage relationship, For charging safety current limit constraints, To constrain the smoothness of the current curve, As a constraint on the monotonicity of SOC, The weighting coefficients for the data consistency loss term. This represents the expected current predicted by the model.

[0009] Optionally, the PINN model includes: a multimodal input module, a feature encoding module, a charging phase sensing branch module, a physical constraint embedded correction module, and an output fusion module; The multimodal input module is used to standardize the input data; The feature encoding module is used to map standardized data into high-dimensional feature representations; The branching module for sensing the charging phase is used to obtain branching features based on the high-dimensional feature representation; The physical constraint embedded correction module is used to physically correct the branch features and obtain the corrected features; The output fusion module is used to fuse the branch features and correction features to obtain the expected charging current and prediction uncertainty index.

[0010] Optionally, the branch module for sensing the charging stage includes: a constant current unit and a constant voltage unit; The constant current unit is used to establish a mapping relationship between multimodal input characteristics and target charging current during the constant current charging stage, thereby characterizing the charging characteristics during the constant current stage. The constant voltage unit is used to establish a mapping relationship between multimodal input characteristics and charging current variation during the constant voltage charging stage, and to characterize the characteristics of charging current variation with battery state during the constant voltage stage.

[0011] Optionally, based on the deviation between the expected charging current and the measured charging current, and in conjunction with the prediction uncertainty index, the adaptive early warning threshold is obtained, including: Calculate the relative deviation based on the expected charging current and the measured charging current; Based on the relative deviation, and combined with the deviation statistics of the AGV's historical normal charging cycles, the basic dynamic threshold is calculated. Based on the prediction uncertainty index, the basic dynamic threshold is dynamically corrected to obtain the adaptive early warning threshold.

[0012] Optionally, before obtaining the graded early warning information based on the comparison result between the deviation and the adaptive early warning threshold, the method further includes: Based on the relative deviation, the deviation persistence index and the deviation change rate index are calculated within a preset sliding time window. The deviation persistence index and the deviation change rate index are used to characterize the degree of persistence and trend of the relative deviation in the time dimension. Based on the aforementioned deviation persistence index, deviation change rate index, and adaptive warning threshold, joint discrimination of different warning levels is achieved.

[0013] Optionally, obtaining graded early warning information based on the comparison result between the deviation and the adaptive early warning threshold includes: When the deviation exceeds the adaptive warning threshold and the risk situation is within a controllable range, it is triggered and classified as a prompt-level warning. When the deviation continues to exceed the adaptive threshold, or the deviation persistence index exceeds the preset range, it is triggered and classified as a warning level alert; When the deviation reaches the high-risk range, or when the deviation change rate indicator shows a rapid upward trend, it is classified as an emergency warning.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention improves the comprehensiveness and accuracy of detection by integrating multimodal data such as charging current, battery temperature, runtime, and SOC. By introducing PINN, the prediction results are made more physically reliable and less susceptible to the influence of limited data and changes in operating conditions. The PINN model predicts the normal charging current curve, enabling intelligent prediction of expected charging behavior and early identification of abnormal charging behavior. By preventing AGV downtime and task delays caused by charging failures, it ensures the stability and orderliness of the transportation process of aviation equipment components, improving operational efficiency and reliability. By introducing physical constraints related to the battery charging mechanism during model training to enhance the physical consistency and generalization ability of the prediction results, this invention maintains stable prediction performance under different battery types and environmental conditions, making it applicable to AGV systems of different models and operating states. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a multimodal data-driven intelligent early warning method for charging operations of AGVs used in the logistics of aviation equipment components, according to an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] This embodiment proposes a multimodal data-driven intelligent early warning method for charging operations of AGVs used in the logistics of aviation equipment components, such as... Figure 1 As shown, the specific steps include: Multimodal charging data is obtained based on the real-time interface data of the intelligent guided vehicle (AGV) onboard system and the charging pile. Based on the multimodal charging data, a PINN model that integrates battery physical constraints is constructed and trained. The PINN model takes charging time, battery temperature, AGV cumulative running time and battery SOC as inputs and outputs the desired charging current. Based on the real-time prediction results of the current charging process from the trained PINN model, the expected charging current and prediction uncertainty index are obtained. Based on the deviation between the expected charging current and the measured charging current, and combined with the prediction uncertainty index, an adaptive early warning threshold is obtained. Based on the comparison between the deviation and the adaptive warning threshold, graded warning information is obtained.

[0019] Specifically, this embodiment provides an intelligent solution to address the problems of insufficient monitoring dimensions, unstable predictions due to lack of physical constraints, and inability to provide proactive early warnings for abnormal charging behavior in AGVs during charging. This embodiment can establish a unified modeling framework for multimodal charging data, construct a PINN prediction model that satisfies the physical laws of batteries, and improve the comprehensiveness of charging status perception; it can achieve prediction of the desired charging curve and real-time deviation monitoring, and automatically alarm by setting dynamic thresholds based on the aviation graded early warning concept; it can improve the safety and continuous operation capability of aviation equipment component logistics systems and reduce potential battery safety risks.

[0020] Furthermore, based on the real-time interface data between the intelligent guided vehicle (AGV) onboard system and the charging pile, multimodal charging data is obtained, including: By interfacing with the AGV onboard system and charging pile data interface, the charging time, battery temperature, AGV cumulative running time, and battery SOC of the AGV are collected in real time to obtain the multimodal charging data.

[0021] Specifically, by interfacing with the AGV onboard system and charging pile data interface, the charging time of the AGV is collected in real time. Battery temperature AGV cumulative running time The battery SOC forms a unified multimodal data input structure, providing a complete dataset for PINN modeling.

[0022] Furthermore, the battery physical constraints include: battery SOC kinetic differential constraints, constant voltage-constant current charging law constraints, constant voltage-constant current charging law constraints, charging safety current limit constraints, and current curve smoothness and SOC monotonicity constraints.

[0023] Specifically, building a PINN model requires designing the model input, model output, physical constraints, loss function, and network architecture. In the technology proposed in this invention, the model input is designed as a four-dimensional input. This refers to four modal data points: charging time, battery temperature, AGV cumulative running time, and battery SOC; the model output is designed as follows: This refers to the expected charging current under normal charging behavior.

[0024] PINN introduces five types of physical constraints. Physical constraint 1 is the SOC dynamic differential constraint. PINN obtains the SOC derivative predicted by the network through automatic differentiation and constructs the SOC residual term, as shown in formula (1), to ensure that the predicted current is consistent with the dynamic relationship of SOC change.

[0025] ; Physical constraint 2 is a constant voltage-constant current charging law constraint. To ensure that the model learns the CC-CV curve characteristics, the physical residual is defined as Equation (2): ; Physical constraint 3 is a temperature-internal resistance-voltage relationship constraint. The residual is constructed based on the equivalent model of the influence of temperature and internal resistance on the current, as shown in formula (3): ; Physical constraint 4 is a charging safety current limit constraint. The soft constraint is defined based on the relationship between the maximum charging current of the lithium battery and the degree of aging, as shown in formula (4): ; Physical constraint 5 is the constraint on the smoothness of the current curve and the monotonicity of SOC, as shown in formulas (5) and (6): ; ; These physical constraints ensure that the network output conforms to the actual charging mechanism, avoiding the instability caused by purely data-driven approaches.

[0026] Based on the above physical constraints, the total loss function of PINN is constructed as Equation (7).

[0027] ; in, The weights of each physical constraint, Historical charging data collected when the AGV is in a healthy state.

[0028] Furthermore, the PINN model includes: a multimodal input module, a feature encoding module, a charging phase sensing branch module, a physical constraint embedded correction module, and an output fusion module; The multimodal input module is used to standardize the input data; The feature encoding module is used to map standardized data into high-dimensional feature representations; The branching module for sensing the charging phase is used to obtain branching features based on the high-dimensional feature representation; The physical constraint embedded correction module is used to physically correct the branch features and obtain the corrected features; The output fusion module is used to fuse the branch features and correction features to obtain the expected charging current and prediction uncertainty index.

[0029] Furthermore, the branch module for sensing the charging stage includes: a constant current unit and a constant voltage unit; The constant current unit is used to establish a mapping relationship between multimodal input characteristics and target charging current during the constant current charging stage, thereby characterizing the charging characteristics during the constant current stage. The constant voltage unit is used to establish a mapping relationship between multimodal input characteristics and charging current variation during the constant voltage charging stage, and to characterize the characteristics of charging current variation with battery state during the constant voltage stage.

[0030] Specifically, at the model architecture level, a structured PINN network architecture was designed to address the needs of charging stage differences, multi-physical constraint collaboration, and early warning. The architecture consists of five modules: a multimodal input layer, a feature encoding layer, a charging stage-aware branch network, a physical constraint embedded correction module, and an output fusion layer. The total number of hidden layers is 14, including 3 hidden layers in the feature encoding layer, 4 hidden layers in each sub-network of the charging stage-aware branch network, 2 hidden layers in the physical constraint embedded correction module, and 1 hidden layer in the output fusion layer. Each layer contains 64 neurons, and the Tanh function is used for all activation functions.

[0031] The multimodal input layer normalizes the four-modal input using a linear mapping. Map the input to the feature encoding layer. This is the weight matrix. This is the bias vector.

[0032] The feature encoding layer maps the input to a shared high-dimensional feature representation. Output features This provides a basic representation for the sensing branch network in the subsequent charging phase. Among these, It consists of a multi-layer fully connected network, with each layer using the Tanh function as the activation function. Batch normalization can be selectively applied, and the weight matrix and bias vector are learned during the training process.

[0033] The charging stage sensing branch network constructs a constant current stage subnetwork based on the physical differences between the constant current stage and the constant voltage stage during battery charging. With constant voltage stage subnetwork The branch networks have similar structures but independent parameters. Each sub-network consists of several fully connected layers. The activation function is also the Tanh function, which is used to learn the mapping relationship of the expected charging current under different charging stages, so as to avoid mutual interference of the physical characteristics of different charging stages.

[0034] The physical constraint embedded correction module is used to explicitly introduce battery physics laws during forward propagation, through differentiable function mapping. The output of the branch network is modified so that some physical constraints are no longer penalized solely by the loss function, but are directly reflected in the network structure, thereby improving the model's training stability and prediction reliability. The embedded physical constraint correction module is implemented by combining neural network layers with explicit differentiable physical mappings. The neural network layers are used to learn residual correction terms, and the explicit mappings ensure that the output satisfies physical boundaries and safety constraints.

[0035] The output fusion layer fuses the branch features with the physical corrections to obtain... At the same time, the predictive uncertainty index is obtained. This is used to characterize the stability and reliability of the model's prediction results under the current input conditions. It can be used to dynamically adjust subsequent warning thresholds, enabling the warning mechanism to have risk adaptive capabilities.

[0036] Subsequently, the PINN model was trained and deployed.

[0037] The PINN model was trained using historical charging data and physical constraints, and its parameters were optimized through cross-validation. After training, the model was packaged and deployed to an edge computing gateway within the factory's local area network to enable real-time inference of charging behavior. Through an API interface, the edge computing gateway maintains real-time communication with the AGV central scheduling and management system to support online monitoring and early warning functions.

[0038] Furthermore, based on the deviation between the expected charging current and the measured charging current, and in conjunction with the prediction uncertainty index, the adaptive warning threshold is obtained as follows: Calculate the relative deviation based on the expected charging current and the measured charging current; Based on the relative deviation, and combined with the deviation statistics of the AGV's historical normal charging cycles, the basic dynamic threshold is calculated. Based on the prediction uncertainty index, the basic dynamic threshold is dynamically corrected to obtain the adaptive early warning threshold.

[0039] Furthermore, before obtaining the graded early warning information, based on the comparison result between the deviation and the adaptive early warning threshold, the following steps are also included: Based on the relative deviation, the deviation persistence index and the deviation change rate index are calculated within a preset sliding time window. The deviation persistence index and the deviation change rate index are used to characterize the degree of persistence and trend of the relative deviation in the time dimension. Based on the aforementioned deviation persistence index, deviation change rate index, and adaptive warning threshold, joint discrimination of different warning levels is achieved.

[0040] Specifically, it enables intelligent early warning of charging anomalies based on the concept of aviation classification.

[0041] During actual charging, the measured current is read. Expected output of PINN Calculate the deviation between the two. .

[0042] ; To further characterize the evolution of charging anomalies, based on the deviation δ, a deviation persistence index and a deviation change rate index are introduced, defined as follows: ; ; Where T is the preset time window length.

[0043] To avoid false alarms or missed alarms caused by using a fixed threshold, this invention calculates the mean deviation based on the statistical characteristics of the AGV's historical normal charging cycles. with standard deviation Based on this, the basic dynamic early warning threshold was determined as follows: This basic threshold is used to reflect the acceptable range of deviation for the AGV under normal charging conditions.

[0044] In addition, the prediction uncertainty index output by PINN As a risk correction factor, it participates in early warning determination, dynamically adjusting the basic early warning threshold to obtain an adaptive threshold for uncertainty perception. ; in, This is the uncertainty adjustment coefficient.

[0045] Further, based on the comparison result between the deviation and the adaptive warning threshold, obtaining graded warning information includes: When the deviation exceeds the adaptive warning threshold and the risk situation is within a controllable range, it is triggered and classified as a prompt-level warning. When the deviation continues to exceed the adaptive threshold, or the deviation persistence index exceeds the preset range, it is triggered and classified as a warning level alert; When the deviation reaches the high-risk range, or when the deviation change rate indicator shows a rapid upward trend, it is classified as an emergency warning.

[0046] Specifically, referencing the "alert-warning-emergency" graded alarm logic in aviation systems, this invention classifies charging anomalies based on deviation amplitude, persistence, trend of change, and prediction uncertainty.

[0047] The warning levels include at least three levels, and the judgment criteria and response strategies are shown in Table 1: Table 1 The threshold ranges for each of the aforementioned warning levels can be configured or adaptively adjusted according to the AGV model, battery specifications, and operating environment; this embodiment does not impose any limitations on this. When a warning is triggered, the edge computing gateway maintains a long-term connection with the AGV central dispatch management system via the WebSocket protocol and generates structured warning information containing the AGV number, charging pile number, abnormal timestamp, warning level, deviation index, and suggested handling measures, which is then pushed to the dispatch management platform in JSON format.

[0048] The following is a detailed description of this embodiment: Step 1: Acquisition and preprocessing of multimodal charging data.

[0049] At the aerospace equipment component logistics site, AGVs are connected to the scheduling system via Wi-Fi or 5G industrial networks, while charging piles are connected to the factory's local area network via Ethernet. The data acquisition module subscribes to and reads real-time data streams, including real-time charging current, from the charging pile's server interface at a frequency of 1Hz through a predefined OPC UA protocol client. (Unit: A), Charging Voltage (Unit: V), Corresponding Time Series (Unit: seconds), the vehicle battery data is retrieved at the same frequency via the RESTful API interface provided by the AGV onboard controller. (0-1 normalized representation), battery temperature (Unit: °C) Cumulative AGV running time (Unit: h). During data processing, the high-precision clock of the charging pile is used as a reference. Linear interpolation is employed to align the timestamps of multi-source data to a unified time series. Subsequently, abnormal jump values ​​are further processed using... Data was discarded according to established principles. Short-term missing data (duration < 3s) was filled in using linear interpolation. The current and temperature signals were smoothed using mean filtering with a sliding window length of 5, resulting in a well-organized time-series training dataset. The recorded sample dataset is as follows: .

[0050] Dataset The sample size is generally no less than 5,000 continuous charging cycle records, covering different ambient temperatures (10℃~40℃) and different battery aging levels (cumulative running time 0~3000h), providing a sufficient data foundation for PINN model training.

[0051] Step 2: Construct a PINN charging behavior prediction model.

[0052] The input to the PINN model is defined as a four-dimensional vector. The output is the desired charging current. The network structure adopts a five-module design with a total of 14 hidden layers. The feature encoding layer consists of three fully connected hidden layers. Each sub-network of the charging stage sensing branch network contains four hidden layers. The physical constraint embedded correction module consists of two hidden layers. The output fusion layer includes one hidden layer and one output layer, with 64 neurons per layer. The activation function, except for the output layer, is the Tanh function. The constant current stage sub-network and the constant voltage stage sub-network have the same structure but do not share parameters to enhance the modeling capability for different charging stages. In addition to outputting the expected charging current, the output fusion layer also outputs the corresponding prediction uncertainty index. This metric is obtained through variance estimation of multiple random Dropout forward inferences and is used to reflect the stability of the model's prediction results under the current operating conditions.

[0053] The weights of each term in the loss function are determined by grid search as follows: In the data fitting term Historical charging data (approximately 300 complete charging cycles) of the same AGV in a healthy state.

[0054] Step 3: PINN model training, parameter optimization, and deployment.

[0055] Preprocessed dataset The dataset is randomly divided into training, validation, and test sets in a 7:2:1 ratio. Model training consists of two phases: the first phase involves global coarse-tuning using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 64. This is repeated 5000 times on the training set to minimize the loss function. The second stage involves local fine-tuning. The network structure is fixed, and the L-BFGS optimizer is switched to. Using the parameters obtained in the first stage as the initial point, unconstrained minimization iterations are performed until the convergence condition is met: the change in the loss function value over 100 consecutive iterations is less than [a certain value]. .

[0056] During training, model performance is evaluated on the validation set every 100 epochs, and the model parameters with the minimum validation loss are saved. Finally, the model's generalization ability is evaluated using the test set, requiring the mean absolute percentage error (MAPE) between the predicted and measured currents to be less than 5%. After training, the model parameters and forward inference program are packaged into Docker containers and deployed on the factory edge computing gateway (such as NVIDIA Jetson Xavier or equivalent computing power equipment) to achieve real-time prediction capability with a single inference latency of less than 50ms.

[0057] Step 4: Generate the expected current curve for normal charging based on the trained PINN model.

[0058] The trained PINN model is based on the real-time input of the AGV's current charging status. Output the expected charging current curve under normal conditions. .

[0059] Step 5: Real-time monitoring and anomaly identification of charging behavior.

[0060] The actual output current of the charging pile is calculated based on formula (8). With the expected charging current deviation .

[0061] Based on this, the deviation persistence index is further calculated within a sliding time window of length T=30s. deviation change rate index This is used to characterize the evolution trend of anomalies. Based on the historical deviation statistics of the AGV's most recent 100 normal charging operations, the mean is calculated. with standard deviation To obtain the basic threshold Further, an adaptive threshold is obtained. ,in (This can be adjusted based on the actual false alarm rate). The resulting detailed tiered early warning system is shown in Table 2. Table 2 Based on the deviation amplitude, duration, trend, and prediction uncertainty, charging anomalies are classified into three levels: alert, warning, and emergency, triggering corresponding handling strategies such as status alerts, parameter adjustments, or automatic charging interruption. When an alert is triggered, the edge computing gateway pushes structured JSON alert information to the AGV central scheduling system via the WebSocket protocol, including the AGV number, charging pile number, anomaly timestamp, alert level, and suggested handling measures.

[0062] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multimodal data-driven intelligent early warning method for charging operations of AGVs used in the logistics of aviation equipment components, characterized in that, include: Multimodal charging data is obtained based on the real-time interface data of the intelligent guided vehicle (AGV) onboard system and the charging pile. Based on the multimodal charging data, a PINN model that integrates battery physical constraints is constructed and trained. The PINN model takes charging time, battery temperature, AGV cumulative running time and battery SOC as inputs and outputs the desired charging current. A total loss function is constructed based on the battery physical constraints, and the PINN model is trained using the total loss function; The total loss function for: ； in, The weights of each physical constraint, Historical charging data collected when the AGV is in a healthy state. For SOC dynamic differential constraints, Constraints on the constant voltage-constant current charging law, Constrained by the temperature-internal resistance-voltage relationship, For charging safety current limit constraints, To constrain the smoothness of the current curve, As a constraint on the monotonicity of SOC, The weighting coefficients for the data consistency loss term. The expected current predicted by the model; ； ； ； ； ； ； in, The cumulative running time of the AGV is t, where t is the charging time of the AGV. Based on the real-time prediction results of the current charging process from the trained PINN model, the expected charging current and prediction uncertainty index are obtained. Based on the deviation between the expected charging current and the measured charging current, and combined with the prediction uncertainty index, an adaptive early warning threshold is obtained. Based on the comparison between the deviation and the adaptive warning threshold, graded warning information is obtained.

2. The intelligent early warning method for charging operations of AGVs in aviation equipment component logistics driven by multimodal data according to claim 1, characterized in that, Based on real-time interface data from the intelligent guided vehicle (AGV) onboard system and charging pile, multimodal charging data is obtained, including: By interfacing with the AGV onboard system and charging pile data interface, the charging time, battery temperature, AGV cumulative running time, and battery SOC of the AGV are collected in real time to obtain the multimodal charging data.

3. The intelligent early warning method for charging operations of AGVs in aviation equipment component logistics driven by multimodal data according to claim 1, characterized in that, The battery physical constraints include: battery SOC kinetic differential constraints, constant voltage-constant current charging law constraints, temperature-internal resistance-voltage relationship constraints, charging safety current limit constraints, and current curve smoothness and SOC monotonicity constraints.

4. The intelligent early warning method for charging operations of AGVs in aviation equipment component logistics driven by multimodal data according to claim 3, characterized in that, The PINN model includes: a multimodal input module, a feature encoding module, a charging phase sensing branch module, a physical constraint embedded correction module, and an output fusion module; The multimodal input module is used to standardize the input data; The feature encoding module is used to map standardized data into high-dimensional feature representations; The branching module for sensing the charging phase is used to obtain branching features based on the high-dimensional feature representation; The physical constraint embedded correction module is used to physically correct the branch features and obtain the corrected features; The output fusion module is used to fuse the branch features and correction features to obtain the expected charging current and prediction uncertainty index.

5. The intelligent early warning method for charging operations of AGVs in aviation equipment component logistics driven by multimodal data according to claim 4, characterized in that, The branch module for sensing the charging phase includes: a constant current unit and a constant voltage unit; The constant current unit is used to establish a mapping relationship between multimodal input characteristics and target charging current during the constant current charging stage, thereby characterizing the charging characteristics during the constant current stage. The constant voltage unit is used to establish a mapping relationship between multimodal input characteristics and charging current variation during the constant voltage charging stage, and to characterize the characteristics of charging current variation with battery state during the constant voltage stage.

6. The intelligent early warning method for charging operations of AGVs in aviation equipment component logistics driven by multimodal data according to claim 1, characterized in that, Based on the deviation between the expected charging current and the measured charging current, and in conjunction with the prediction uncertainty index, the adaptive early warning threshold is obtained as follows: Calculate the relative deviation based on the expected charging current and the measured charging current; Based on the relative deviation, and combined with the deviation statistics of the AGV's historical normal charging cycles, the basic dynamic threshold is calculated. Based on the prediction uncertainty index, the basic dynamic threshold is dynamically corrected to obtain the adaptive early warning threshold.

7. The intelligent early warning method for charging operations of AGVs in aviation equipment component logistics driven by multimodal data according to claim 6, characterized in that, Based on the comparison result between the deviation and the adaptive early warning threshold, the process before obtaining the graded early warning information further includes: Based on the relative deviation, the deviation persistence index and the deviation change rate index are calculated within a preset sliding time window. The deviation persistence index and the deviation change rate index are used to characterize the degree of persistence and trend of the relative deviation in the time dimension. Based on the aforementioned deviation persistence index, deviation change rate index, and adaptive warning threshold, joint discrimination of different warning levels is achieved.

8. The intelligent early warning method for charging operations of AGVs in aviation equipment component logistics driven by multimodal data according to claim 1, characterized in that, Based on the comparison result between the deviation and the adaptive warning threshold, the graded warning information is obtained as follows: When the deviation exceeds the adaptive warning threshold and the risk situation is within a controllable range, it is triggered and classified as a prompt-level warning. When the deviation continues to exceed the adaptive threshold, or the deviation persistence index exceeds the preset range, it is triggered and classified as a warning level alert; When the deviation reaches the high-risk range, or when the deviation change rate indicator shows a rapid upward trend, it is classified as an emergency warning.

Images