AI-based modeling-based health status assessment method for integrated power supply boxes

Abstract

This invention discloses an AI-based modeling method for assessing the health status of integrated power supply boxes, comprising the following steps: S1, constructing a multivariate time series; S2, modeling the multivariate time series trajectory using a neural control differential equation model to obtain the latent space trajectory; S3, performing dynamic modal decomposition on the latent space trajectory to construct a modal health space; S4, generating an enhanced training sample set using a manifold interpolation hybrid method in the modal health space; S5, inputting the hybrid modal vectors from the enhanced training sample set into a health classification module; S6, inputting the latent space trajectory into a temporal semantic playback module to generate the final predicted state trajectory; S7, comparing the final predicted state trajectory with the actual historical state trajectory, and generating an anomaly marker and performing a state rollback operation when the semantic difference exceeds a set tolerance threshold. This invention integrates neural differential modeling, interpolation enhancement, and temporal semantic playback to achieve accurate assessment of the health status of integrated power supply boxes.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence modeling and condition monitoring technology, and in particular to a method for assessing the health status of an integrated power supply box based on AI modeling. Background Technology

[0002] Integrated power supply boxes, as core components of power systems, are widely used in industrial automation, power grid infrastructure, and new energy equipment to centrally manage power input and output, regulate power supply stability, and control loads. With increasing system integration and application complexity, the number of multivariate disturbances, electrical fluctuations, and environmental coupling factors during equipment operation has significantly increased, making health status assessment more difficult and unstable. Traditional power supply box fault detection and health assessment methods largely rely on static threshold settings, feature engineering extraction, and empirical rule judgments. These methods suffer from significantly reduced accuracy and timeliness when dealing with modern power systems characterized by high data dimensionality, strong dynamic changes, and complex operating modes.

[0003] In existing technologies, some research has begun to explore the introduction of artificial intelligence methods to model and assess the operational data of integrated power supply units. For example, some studies have used deep learning architectures such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs) to classify and model sensor data, thereby achieving anomaly detection or early warning. However, most of these methods remain at the "black-box modeling" stage, lacking the ability to explicitly model the state evolution mechanism and failing to effectively capture the continuous dynamic characteristics during the operation of the power supply unit. Furthermore, existing methods generally use discrete time series inputs, failing to utilize the structural information of variables in temporal continuity, resulting in insufficient understanding of state differential trends or trajectory changes by the model, thus affecting the accuracy of health status assessment.

[0004] On the other hand, regarding the issues of data augmentation and sample distribution, existing methods mainly employ popular image augmentation techniques (such as rotation and cropping) or simple numerical perturbations (such as Gaussian noise addition and translation) to expand the original sensor data. These methods cannot adapt to the semantic preservation constraints in multidimensional modal spaces, leading to distortion and severe spectral drift in the training samples after augmentation, thus weakening the model's generalization ability. Furthermore, there is currently no effective way to use modal information as the basic modeling unit, combined with interpolation strategies on the modal structure, to generate semantically continuous and frequency-consistent mixed samples, thereby improving the stability and robustness of the health status discrimination model.

[0005] Furthermore, although some methods have incorporated residual networks or regression modules to predict the remaining lifespan of equipment, most methods still treat time as a one-way input, ignoring the potential symmetry or reversibility of the equipment's operating trajectory over time. This structural deficiency prevents the model from self-verifying its predictions, and also hinders its ability to combine historical data to verify the reasonableness of predictions or identify anomalies. Consequently, it becomes difficult to reverse-engineer and correct complex evolution trajectories in multi-state transition scenarios.

[0006] Therefore, how to provide a health status assessment method for integrated power supply boxes based on AI modeling is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose an AI-based modeling method for assessing the health status of integrated power supply units. This invention integrates neural control differential equation modeling with modal space interpolation enhancement methods to construct the latent space trajectory and modal health space of the integrated power supply unit's operating status. Combined with temporal semantic playback and anomaly verification mechanisms, it achieves accurate assessment and dynamic correction of health status and remaining lifespan. This method can perform continuous and differentiable modeling of multivariate time series, extract orthogonal modal features with frequency structures, and generate highly consistent enhanced samples through nonlinear interpolation, effectively improving the stability and generalization ability of the assessment model. It possesses advantages such as strong modeling continuity, high recognition accuracy, high data utilization efficiency, and strong anomaly self-correction capability.

[0008] The AI-based modeling-based health status assessment method for integrated power supply boxes according to embodiments of the present invention includes the following steps:

[0009] S1. Collect the operating data of the integrated power supply box, and perform time synchronization, interpolation and normalization processing to construct a multivariate time series;

[0010] S2. The multivariate time series trajectory is modeled using a neural control differential equation model. The state latent space vector is initialized, and the evolution path of the state over time is modeled using a continuous differential control function to obtain the latent space trajectory.

[0011] S3. Perform dynamic mode decomposition on the latent space trajectory to extract several orthogonal mode vectors. Each mode vector constitutes a mode health space, and each mode vector corresponds to a set of frequency components and energy distribution.

[0012] S4. In the modal health space, a manifold interpolation mixing method is used to mix the modal vectors of two different samples proportionally to generate an enhanced training sample set.

[0013] S5. Input the mixed modality vector in the enhanced training sample set into the health classification module, and output the corresponding health status label and remaining life prediction value.

[0014] S6. Input the hidden space trajectory into the time semantic playback module, push back from the current time point, and generate the final predicted state trajectory within the historical time interval.

[0015] S7. Compare the final predicted state trajectory with the actual historical state trajectory to calculate the semantic difference value. When the semantic difference value exceeds the set tolerance threshold, generate an anomaly marker, replace the hidden space trajectory of the current time step with the corresponding actual historical state trajectory, and re-execute steps S3 to S5 to update the corresponding health status label and remaining life prediction value.

[0016] Optionally, the operating data includes voltage, current, temperature, humidity, power factor, switching status, and electromagnetic interference value.

[0017] Optionally, S2 specifically includes:

[0018] S21. The multivariate time series is used as the input path of the neural control differential equation model, wherein the input path is an ordered multidimensional data sequence arranged by time index;

[0019] S22. Perform cubic spline interpolation on the input path to construct a continuous path representation, so that it is differentiable in the time dimension, in order to meet the modeling requirements of continuous differential control functions in the neural control differential equation model.

[0020] S23. Input the multidimensional data corresponding to the first time slice in the input path into a fully connected neural network to generate an initial state hidden space vector, which serves as the initial state of the neural control differential equation model.

[0021] S24. Construct a continuous differential control function, which is implemented by a feedforward neural network. It receives the current state hidden space vector and the path value at the current time point as input and outputs the corresponding state derivative, which is used to define the differential relationship of the state evolution over time.

[0022] S25. Numerical integration of the continuous differential control function is performed using ordinary differential equations, and the state latent space vectors at each time point are generated sequentially from the initial state, following the time axis of the input path.

[0023] S26. The state latent space vectors at each time point are arranged in chronological order to form a latent space trajectory. The latent space trajectory is a continuously differentiable vector sequence used to represent the dynamic evolution of the integrated power supply box's operating state over the entire time range.

[0024] Optionally, S3 specifically includes:

[0025] S31. The acquired latent space trajectory is converted into a state matrix arranged sequentially by time steps, wherein each column of the state matrix represents a latent space vector of a state at a time point.

[0026] S32. The state matrix is ​​divided into several local time blocks by a sliding window to capture the dynamic change characteristics of the running state in different time periods.

[0027] S33. Perform dynamic mode decomposition on each local time block. The dynamic mode decomposition includes: constructing an input matrix and an output matrix. The input matrix is ​​composed of the state latent space vectors in consecutive time steps, excluding the last time point of each time block. The output matrix corresponds to the state latent space vectors after excluding the first time point. The output matrix is ​​approximated by the solution of the input matrix multiplied by the linear mapping matrix through least squares fitting. After obtaining the linear mapping matrix, the eigenvalues ​​and right eigenvectors are calculated. The right eigenvectors constitute the modal basis representing the state evolution mode.

[0028] S34. Perform eigenvalue decomposition on the modal basis to obtain the frequency information and corresponding modal amplitude of each mode, and construct a mode set composed of several modal vectors;

[0029] S35. The mode set is constrained by an orthogonalization method so that the mode vectors satisfy the linear independence condition, thereby obtaining several orthogonal mode vectors.

[0030] S36. A modal health space is formed by several orthogonal modal vectors, and each modal vector corresponds to a set of frequency components and energy distributions.

[0031] Optionally, S4 specifically includes:

[0032] S41. Select two different training samples in the modal health space, and the corresponding modal vectors are the first modal vector and the second modal vector, respectively.

[0033] S42. The manifold interpolation hybrid method includes: embedding the first mode vector and the second mode vector into a unified feature space, and constructing a hybrid mode vector according to a nonlinear interpolation function, wherein the nonlinear interpolation function adopts an adjustable curvature control method.

[0034] ;

[0035] in, Represents a mixed-mode vector. Represents the first mode vector. This represents the second mode vector. This represents the control proportionality coefficient, with a value range of (0,1). The interpolation trajectory forms a nonlinear surface path, which is used to generate a continuously distributed mixed mode vector.

[0036] S43. Perform principal frequency analysis on the mixed mode vector and calculate the proportion of frequency components within a predefined energy distribution range. The principal frequency analysis is based on fast Fourier transform and constrains the spectral characteristics of the mixed mode vector.

[0037] S44. When the overlap between the main frequency energy concentration region of the mixed mode vector and the original two mode vectors is higher than a preset overlap threshold, the frequency of the mixed mode vector is deemed to meet the frequency consistency condition.

[0038] S45. Generate an enhanced training sample set based on the mixed modal vectors that satisfy the frequency consistency condition.

[0039] Optionally, S5 specifically includes:

[0040] S51. Input the mixed modality vector in the enhanced training sample set into the health classification module. The health classification module includes an input layer, a two-layer fully connected neural network structure, a classification output layer, and a lifetime regression output layer.

[0041] S52, The input layer receives the mixed mode vector from the enhanced training sample set;

[0042] S53. The two-layer fully connected neural network structure includes a fully connected layer with a ReLU activation function and a fully connected layer with a Sigmoid activation function, respectively, for extracting nonlinear feature representations of mixed mode vectors;

[0043] S54. The classification output layer is a fully connected layer with a Softmax activation function, and the output dimension is 3, corresponding to the three categories defined in the health status label: 0 represents normal, 1 represents warning, and 2 represents serious.

[0044] S55. The lifetime regression output layer and the classification output layer are connected in parallel. The lifetime regression output layer is a linear regression layer containing one output neuron. Each component of the nonlinear feature representation is multiplied by a set weight parameter and then summed. A bias term is added to calculate the remaining lifetime prediction value.

[0045] Optionally, S6 specifically includes:

[0046] S61. Input the acquired latent space trajectory into the time semantic playback module. The latent space trajectory is converted into a state matrix arranged sequentially according to time steps. Each column of the state matrix represents a state latent space vector at a time point.

[0047] S62. Select the state latent space vector corresponding to the current time point as the playback starting point, and construct a historical playback time window of a set length forward. The historical playback time window contains several consecutive time steps of reverse prediction targets.

[0048] S63. The time semantic playback module includes a time-aware encoder and a state predictor. The time-aware encoder encodes the interval information of each time step within each of the historical playback time windows, converts it into a time feature vector, and concatenates it with each state latent space vector to form a prediction input sequence.

[0049] S64. The state predictor adopts a symmetrical state residual structure, which includes a forward prediction submodule and a reverse residual correction submodule. The forward prediction submodule is based on a recurrent neural network, receives the prediction input sequence, and generates the initial predicted state trajectory for each time step within the historical playback time window.

[0050] S65. The residual correction submodule constructs the state evolution residual trajectory for each time step according to the symmetry principle, and adds the state evolution residual trajectory to the corresponding initial predicted state trajectory to form the final predicted state trajectory:

[0051] ;

[0052] in, This represents the final predicted state trajectory. This represents the initial predicted state trajectory. This represents the latent space vector of the state at the current time point.

[0053] Optionally, S7 specifically includes:

[0054] S71. Align the final predicted state trajectory with the actual historical state trajectory, and extract the state latent space vector pairs at each time point according to the time step.

[0055] S72. Based on the state latent space vector pair, the semantic difference value of the entire trajectory is obtained by calculating the state latent space vector difference for each corresponding time step and accumulating the norm.

[0056] S73. When the semantic difference value exceeds the set tolerance threshold, it is determined that there is an abnormal deviation between the final predicted state trajectory and the actual historical trajectory, and an abnormal marker is generated.

[0057] S74. After triggering the anomaly marker, replace the hidden space trajectory of the current time step with the corresponding actual historical state trajectory, and re-execute steps S3 to S5 to update the corresponding health status label and remaining life prediction value.

[0058] The beneficial effects of this invention are:

[0059] First, this invention employs a neural control differential equation model to perform continuous trajectory modeling of the multivariate time series of power supply box operation. By introducing a continuous differential control function, it achieves precise modeling of state changes over time, overcoming the problem that traditional deep learning models lack the ability to model temporal continuity when processing discrete sequences. Compared to traditional RNN or LSTM models, this method not only maintains the continuous differentiability of the state evolution path but also provides numerically interpretable modeling of the dynamic state process through explicit differential equations, thereby improving the accuracy of capturing equipment operating trends.

[0060] Secondly, this invention introduces a dynamic mode decomposition method to divide the latent space state trajectory into temporal blocks and extract modes, obtaining mode vectors with frequency characteristics and structural orthogonality, and constructing a modal health space. This mode-based modeling approach enhances the model's ability to identify system behavior under different operating modes and adapts to state changes under complex operating conditions. Simultaneously, the orthogonality constraint of the mode vectors ensures the separability and complementarity of features in each dimension, which is beneficial for the efficient training of health state classification and lifetime regression models in downstream tasks.

[0061] Furthermore, to address the issues of insufficient and unevenly distributed samples, this invention proposes a sample augmentation method based on manifold interpolation hybridization. This method selects the modal vectors of two samples in the modal health space, performs embedding and nonlinear interpolation to generate a hybrid modal vector, and combines spectral overlap constraints to ensure the frequency consistency of the augmented samples. This method balances the semantic structure and physical consistency of the samples, improves the model's generalization ability and training stability, and overcomes the limitation of existing technologies that cannot guarantee the physical reliability of interpolated samples.

[0062] Furthermore, this invention introduces a temporal semantic playback module, used to predict historical state trajectories backward from the current time point. Combined with a symmetric state residual structure, it achieves structural modeling of historical trajectories by applying a symmetric residual correction based on the current state to the initial predicted trajectory. Compared to traditional one-way prediction methods, this module not only realizes bidirectional reasoning capability for state prediction but also provides a self-checking basis for the rationality of the trajectory.

[0063] Finally, by calculating the semantic difference between the predicted trajectory obtained from playback and the actual historical trajectory, this invention establishes an anomaly verification and rollback mechanism. When the semantic difference exceeds a set threshold, the current state is automatically replaced and the health status and lifespan prediction results are rolled back for reassessment. This significantly enhances the model's fault tolerance and self-correction capabilities, avoiding misjudgments or miscontrols caused by prediction deviations. Attached Figure Description

[0064] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0065] Figure 1 This is an overall flowchart of the AI-based modeling-based health status assessment method for integrated power supply boxes proposed in this invention.

[0066] Figure 2 This is a step-by-step flowchart of the dynamic modal decomposition process of the AI-based integrated power supply box health status assessment method proposed in this invention.

[0067] Figure 3 This is a structural block diagram of the temporal semantic playback module of the AI-based integrated power supply box health status assessment method proposed in this invention. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0069] refer to Figures 1-3 The AI-based modeling-based integrated power supply box health status assessment method includes the following steps:

[0070] S1. Collect the operating data of the integrated power supply box, and perform time synchronization, interpolation and normalization processing to construct a multivariate time series;

[0071] S2. The multivariate time series trajectory is modeled using a neural control differential equation model. The state latent space vector is initialized, and the evolution path of the state over time is modeled using a continuous differential control function to obtain the latent space trajectory.

[0072] S3. Perform dynamic mode decomposition on the latent space trajectory to extract several orthogonal mode vectors. Each mode vector constitutes a mode health space, and each mode vector corresponds to a set of frequency components and energy distribution.

[0073] S4. In the modal health space, a manifold interpolation mixing method is used to mix the modal vectors of two different samples proportionally to generate an enhanced training sample set.

[0074] S5. Input the mixed modality vector in the enhanced training sample set into the health classification module, and output the corresponding health status label and remaining life prediction value.

[0075] S6. Input the hidden space trajectory into the time semantic playback module, push back from the current time point, and generate the final predicted state trajectory within the historical time interval.

[0076] S7. Compare the final predicted state trajectory with the actual historical state trajectory to calculate the semantic difference value. When the semantic difference value exceeds the set tolerance threshold, generate an anomaly marker, replace the hidden space trajectory of the current time step with the corresponding actual historical state trajectory, and re-execute steps S3 to S5 to update the corresponding health status label and remaining life prediction value.

[0077] This invention achieves continuous modeling of the operating state of an integrated power supply box by introducing a neural control differential equation model, fully preserving the temporal evolution information of multivariate time-series data. Through differential modeling and numerical integration of the input sequence, the obtained state latent space trajectory possesses differentiability and structural continuity, enabling a more accurate characterization of the nonlinear dynamic changes during equipment operation. Furthermore, based on the latent space trajectory, a series of operations such as mode decomposition, mode enhancement, state classification, and lifetime prediction are performed, making the health assessment process not only logically closed-loop but also interpretable and error-correcting. By automatically triggering a backoff mechanism when the state deviates, the method exhibits good adaptability and operational safety. The overall structural design emphasizes the completeness of state evolution modeling and the stability of system evaluation, significantly improving the robustness and prediction accuracy of health state judgment, making it suitable for real-time operation monitoring and maintenance decision-making in complex electrical systems.

[0078] In this embodiment, the operating data includes voltage, current, temperature, humidity, power factor, switching status, and electromagnetic interference value.

[0079] In this embodiment, S2 specifically includes:

[0080] S21. The multivariate time series is used as the input path of the neural control differential equation model, wherein the input path is an ordered multidimensional data sequence arranged by time index;

[0081] S22. Perform cubic spline interpolation on the input path to construct a continuous path representation, so that it is differentiable in the time dimension, in order to meet the modeling requirements of continuous differential control functions in the neural control differential equation model.

[0082] S23. Input the multidimensional data corresponding to the first time slice in the input path into a fully connected neural network to generate an initial state hidden space vector, which serves as the initial state of the neural control differential equation model.

[0083] S24. Construct a continuous differential control function, which is implemented by a feedforward neural network. It receives the current state hidden space vector and the path value at the current time point as input and outputs the corresponding state derivative, which is used to define the differential relationship of the state evolution over time.

[0084] S25. Numerical integration of the continuous differential control function is performed using ordinary differential equations, and the state latent space vectors at each time point are generated sequentially from the initial state, following the time axis of the input path.

[0085] S26. The state latent space vectors at each time point are arranged in chronological order to form a latent space trajectory. The latent space trajectory is a continuously differentiable vector sequence used to represent the dynamic evolution of the integrated power supply box's operating state over the entire time range.

[0086] The process of constructing the neural control differential equation model enables continuous modeling of state evolution, no longer relying on traditional discrete sequence modeling methods. By introducing cubic spline interpolation into the input path, the continuous differentiability of the path in the time dimension is guaranteed, providing a mathematical foundation for differential function modeling. Simultaneously, a continuous differential control function is constructed using a feedforward neural network, which generates derivative information for the input state at each time step, and then accurately generates the latent space trajectory through numerical integration. This approach enhances the ability to model the micro-trends of state transitions, and can more sensitively reflect the dynamic evolution of the system. Furthermore, this structure also enables the model to possess explicit state interpretability, exhibiting higher stability and reliability in high-frequency disturbances or critical state identification, and is a key foundation for achieving end-to-end trajectory modeling and evolution prediction.

[0087] In this embodiment, S3 specifically includes:

[0088] S31. The acquired latent space trajectory is converted into a state matrix arranged sequentially by time steps, wherein each column of the state matrix represents a latent space vector of a state at a time point.

[0089] S32. The state matrix is ​​divided into several local time blocks by a sliding window to capture the dynamic change characteristics of the running state in different time periods.

[0090] S33. Perform dynamic mode decomposition on each local time block. The dynamic mode decomposition includes: constructing an input matrix and an output matrix. The input matrix is ​​composed of the state latent space vectors in consecutive time steps, excluding the last time point of each time block. The output matrix corresponds to the state latent space vectors after excluding the first time point. The output matrix is ​​approximated by the solution of the input matrix multiplied by the linear mapping matrix through least squares fitting. After obtaining the linear mapping matrix, the eigenvalues ​​and right eigenvectors are calculated. The right eigenvectors constitute the modal basis representing the state evolution mode.

[0091] S34. Perform eigenvalue decomposition on the modal basis to obtain the frequency information and corresponding modal amplitude of each mode, and construct a mode set composed of several modal vectors;

[0092] S35. The mode set is constrained by an orthogonalization method so that the mode vectors satisfy the linear independence condition, thereby obtaining several orthogonal mode vectors.

[0093] S36. A modal health space is formed by several orthogonal modal vectors, and each modal vector corresponds to a set of frequency components and energy distributions.

[0094] By performing dynamic modal decomposition on the latent space trajectory, the frequency distribution structure and modal amplitude features describing the system's operating state can be effectively extracted. A sliding window strategy is employed to divide the time-series data into local blocks, ensuring the dynamic adaptability and temporal localization capability of the decomposition process. By constructing input-output matrices and solving linear mapping operators, modal fitting for each local state change is achieved, enabling the extracted modal basis to effectively represent various state evolution patterns. Orthogonalization is then introduced to ensure that modal features possess linear independence in the spatial dimension, thereby improving the discriminative performance of downstream classification and regression models. The establishment of the modal health space provides a structurally stable and frequency-controllable modeling foundation for subsequent sample augmentation and classification tasks, helping to enhance the system's ability to identify non-stationary state changes.

[0095] In this embodiment, S4 specifically includes:

[0096] S41. Select two different training samples in the modal health space, and the corresponding modal vectors are the first modal vector and the second modal vector, respectively.

[0097] S42. The manifold interpolation hybrid method includes: embedding the first mode vector and the second mode vector into a unified feature space, and constructing a hybrid mode vector according to a nonlinear interpolation function, wherein the nonlinear interpolation function adopts an adjustable curvature control method.

[0098] ;

[0099] in, Represents a mixed-mode vector. Represents the first mode vector. This represents the second mode vector. This represents the control proportionality coefficient, with a value range of (0,1). The interpolation trajectory forms a nonlinear surface path, which is used to generate a continuously distributed mixed mode vector.

[0100] S43. Perform principal frequency analysis on the mixed mode vector and calculate the proportion of frequency components within a predefined energy distribution range. The principal frequency analysis is based on fast Fourier transform and constrains the spectral characteristics of the mixed mode vector.

[0101] S44. When the overlap between the main frequency energy concentration region of the mixed mode vector and the original two mode vectors is higher than a preset overlap threshold, the frequency of the mixed mode vector is deemed to meet the frequency consistency condition.

[0102] S45. Generate an enhanced training sample set based on the mixed modal vectors that satisfy the frequency consistency condition.

[0103] This method constructs hybrid samples using manifold interpolation in the modality health space, overcoming the limitations of traditional sample augmentation methods in terms of semantic consistency and frequency preservation. By constructing nonlinear interpolation trajectories and applying spectral consistency constraints, it ensures that the generated hybrid modality vectors maintain similarity in frequency distribution to the original modalities, avoiding semantic drift and mode distortion. This augmentation method can generate samples with natural and continuously varying structures within the modality space, effectively filling sparsely distributed regions in the sample space and improving the model's ability to generalize and identify abnormal states. Simultaneously, a frequency overlap detection mechanism filters out non-compliant samples, enhancing data quality control and effectively improving the stability, convergence speed, and evaluation accuracy of subsequent classification model training.

[0104] In this embodiment, S5 specifically includes:

[0105] S51. Input the mixed modality vector in the enhanced training sample set into the health classification module. The health classification module includes an input layer, a two-layer fully connected neural network structure, a classification output layer, and a lifetime regression output layer.

[0106] S52, The input layer receives the mixed mode vector from the enhanced training sample set;

[0107] S53. The two-layer fully connected neural network structure includes a fully connected layer with a ReLU activation function and a fully connected layer with a Sigmoid activation function, respectively, for extracting nonlinear feature representations of mixed mode vectors;

[0108] S54. The classification output layer is a fully connected layer with a Softmax activation function, and the output dimension is 3, corresponding to the three categories defined in the health status label: 0 represents normal, 1 represents warning, and 2 represents serious.

[0109] S55. The lifetime regression output layer and the classification output layer are connected in parallel. The lifetime regression output layer is a linear regression layer containing one output neuron. Each component of the nonlinear feature representation is multiplied by a set weight parameter and then summed. A bias term is added to calculate the remaining lifetime prediction value.

[0110] The health status recognition module combines classification and regression with a dual-output structure, accurately determining the current health level and simultaneously providing a life expectancy estimate, thus enhancing the system's practicality and decision support capabilities. The nonlinear feature extraction network utilizes ReLU and Sigmoid activation functions to extract features at multiple scales, demonstrating excellent perception of subtle anomalies in modality vectors. The classification output structure employs the Softmax function to output category probabilities, adapting to multi-level status labels and clearly distinguishing between normal, warning, and severe states. The regression output uses a linear combination structure, improving the ability to finely fit life expectancy trends. The overall design is rational, the model output is stable, and it exhibits good engineering deployment performance and predictive response capabilities.

[0111] In this embodiment, S6 specifically includes:

[0112] S61. Input the acquired latent space trajectory into the time semantic playback module. The latent space trajectory is converted into a state matrix arranged sequentially according to time steps. Each column of the state matrix represents a state latent space vector at a time point.

[0113] S62. Select the state latent space vector corresponding to the current time point as the playback starting point, and construct a historical playback time window of a set length forward. The historical playback time window contains several consecutive time steps of reverse prediction targets.

[0114] S63. The time semantic playback module includes a time-aware encoder and a state predictor. The time-aware encoder encodes the interval information of each time step within each of the historical playback time windows, converts it into a time feature vector, and concatenates it with each state latent space vector to form a prediction input sequence.

[0115] S64. The state predictor adopts a symmetrical state residual structure, which includes a forward prediction submodule and a reverse residual correction submodule. The forward prediction submodule is based on a recurrent neural network, receives the prediction input sequence, and generates the initial predicted state trajectory for each time step within the historical playback time window.

[0116] S65. The residual correction submodule constructs the state evolution residual trajectory for each time step according to the symmetry principle, and adds the state evolution residual trajectory to the corresponding initial predicted state trajectory to form the final predicted state trajectory:

[0117] ;

[0118] in, This represents the final predicted state trajectory. This represents the initial predicted state trajectory. This represents the latent space vector of the state at the current time point.

[0119] The temporal semantic replay module achieves inverse modeling of historical trajectories by constructing a symmetric state residual structure. By embedding time step interval information into a time-aware encoder, the model can identify the dynamic evolution rhythm within the replay interval and generate an initial trajectory based on the predicted input sequence. Furthermore, a reverse residual correction submodule is constructed using the symmetry principle, enabling the model to have correction capabilities and significantly improving the accuracy and interpretability of the predicted trajectory. The introduction of this module allows the system to not only have forward prediction capabilities but also achieve inverse temporal inference, thus possessing self-verification and anomaly replay functions, which is a crucial mechanism for achieving reliable trajectory prediction.

[0120] In this embodiment, S7 specifically includes:

[0121] S71. Align the final predicted state trajectory with the actual historical state trajectory, and extract the state latent space vector pairs at each time point according to the time step.

[0122] S72. Based on the state latent space vector pair, the semantic difference value of the entire trajectory is obtained by calculating the state latent space vector difference for each corresponding time step and accumulating the norm.

[0123] S73. When the semantic difference value exceeds the set tolerance threshold, it is determined that there is an abnormal deviation between the final predicted state trajectory and the actual historical trajectory, and an abnormal marker is generated.

[0124] S74. After triggering the anomaly marker, replace the hidden space trajectory of the current time step with the corresponding actual historical state trajectory, and re-execute steps S3 to S5 to update the corresponding health status label and remaining life prediction value.

[0125] The predicted state trajectory is compared one-to-one with the actual historical trajectory, and anomaly deviation identification is achieved based on a semantic difference measurement mechanism, establishing an effective self-checking and feedback closed-loop mechanism. The semantic difference between trajectories is measured by norm accumulation, and a tolerance threshold is set as the discrimination boundary. When the prediction error exceeds the threshold, state replacement and result rollback operations are triggered, thereby preventing abnormal trajectories from propagating to the final judgment result. This strategy improves the robustness and reliability of the entire health assessment process, enabling the system to have self-repair and anomaly suppression capabilities, making it suitable for scenarios with high reliability requirements.

[0126] Example 1:

[0127] To verify the feasibility of this invention in practice, it was applied to an integrated power supply box system for experimental verification. The system continuously collected key operating parameters during operation, including voltage, current, temperature, humidity, power factor, switching status, and electromagnetic interference values, forming a representative set of multivariate time series data.

[0128] First, the collected raw operational data is synchronized in time, and alignment is completed based on a unified timestamp; linear interpolation is used to fill in missing data points; then normalization is performed to map all features to intervals. The final result is a multivariate time series input path with a length of 6000 steps and 7-dimensional features in each step.

[0129] Subsequently, continuous modeling was performed using a neural control differential equation model: first, cubic spline interpolation was applied to the input path to make it differentiable in the time dimension; then, the multidimensional data corresponding to the first time slice was input into a three-layer fully connected network, outputting a 128-dimensional initial state latent space vector; a continuous differential control function implemented by a two-layer feedforward network was constructed, with the current latent space vector and time path value as inputs, and the state derivative as output; the fourth-order Runge-Kutta method was used to integrate the function on the time axis to obtain the state derivative. to The latent space vectors are arranged in order to form the latent space trajectory.

[0130] Perform dynamic mode decomposition on the latent space trajectory: After transforming the trajectory into a state matrix, use the window length... Step, sliding step The sliding window strategy forms local time blocks; an input matrix is ​​constructed for each block. With output matrix Solving the linear mapping matrix using the least squares method ,satisfy For the matrix Eigenvalues ​​and right eigenvectors are obtained by performing eigenvalue decomposition, and modal basis is extracted. Modal health space is obtained by frequency-amplitude analysis and orthogonalization using the Gram-Schmidt method.

[0131] Two sample mode vectors are randomly selected within this space. and Set the control ratio coefficient The mixed mode vector is generated using the following formula:

[0132] ;

[0133] right Performing a Fast Fourier Transform, if its dominant frequency energy is... Overlap rate of main frequency range If the frequency consistency test is passed, it will be included in the enhanced training sample set.

[0134] The enhanced training sample set is fed into the health classification module: the input layer has a dimension of 128, and is activated by a single ReLU layer. Fully connected layer and one Sigmoid activated layer Fully connected layers extract nonlinear features; the softmax output layer provides health status labels. Parallel linear regression layers calculate remaining lifetime After training on 10,000 samples for 50 rounds, the network achieved a label classification accuracy of 94.6% and a mean absolute error of 3.8 hours for remaining lifetime.

[0135] During the online assessment phase, at the current point in time The corresponding latent space vector As the starting point for playback, a historical playback window of 40 steps is constructed forward. The time-aware encoder encodes the interval between adjacent time steps into 16-dimensional temporal features, which are concatenated with the latent space vector to form the prediction input sequence; the forward prediction submodule (bidirectional LSTM) outputs the initial predicted trajectory. The residual correction submodule is based on:

[0136] ;

[0137] Constructing symmetric residuals yields the final predicted state trajectory. .

[0138] Next With respect to actual historical trajectory Perform semantic difference comparison: calculate the norm of the difference vector at each step and sum them up to obtain the overall difference. The experiment sets a tolerance threshold. ;when When the system determines that the prediction is abnormal, it automatically... replace The results were then reassessed by going back to the modal decomposition step, and the outcome was revised from severe / 2 hours to warning / 6 hours, consistent with subsequent on-site detection.

[0139] This embodiment demonstrates that the method can effectively solve the problems of insufficient dynamic capture, sparse and distorted samples, and accumulation of prediction errors in traditional evaluation schemes at three levels: continuous sequence modeling, modal space enhancement, and time playback correction, thereby achieving high-precision identification of the health status of integrated power supply boxes and reliable prediction of remaining lifetime.

[0140] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A health status assessment method for integrated power supply boxes based on AI modeling, characterized in that, Includes the following steps: S1. Collect the operating data of the integrated power supply box, and perform time synchronization, interpolation and normalization processing to construct a multivariate time series. The operating data includes voltage, current, temperature, humidity, power factor, switch status and electromagnetic interference value. S2. The neural control differential equation model is used to model the trajectory of multivariate time series, the state latent space vector is initialized, and the evolution path of the state with time is modeled through continuous differential control function to obtain the latent space trajectory. First, cubic spline interpolation is applied to the input path to make it differentiable in the time dimension. Then, the multidimensional data corresponding to the first time slice is input into a three-layer fully connected network, which outputs a 128-dimensional initial state latent space vector. A continuous differential control function implemented by a two-layer feedforward network is constructed, whose input is the current latent space vector and the time path value, and whose output is the state derivative. The fourth-order Runge-Kutta method is used to integrate this function on the time axis to obtain the initial state from... to The latent space vectors are arranged in order to form the latent space trajectory; S3. Perform dynamic mode decomposition on the latent space trajectory to extract several orthogonal mode vectors. Each mode vector constitutes a mode health space, and each mode vector corresponds to a set of frequency components and energy distribution. S4. In the modal health space, a manifold interpolation mixing method is used to mix the modal vectors of two different samples proportionally to generate an enhanced training sample set. S5. Input the mixed modality vector in the enhanced training sample set into the health classification module, and output the corresponding health status label and remaining life prediction value. S6. Input the hidden space trajectory into the time semantic playback module, push back from the current time point, and generate the final predicted state trajectory within the historical time interval. S7. Compare the final predicted state trajectory with the actual historical state trajectory to calculate the semantic difference value. When the semantic difference value exceeds the set tolerance threshold, generate an anomaly marker, replace the hidden space trajectory of the current time step with the corresponding actual historical state trajectory, and re-execute steps S3 to S5 to update the corresponding health status label and remaining life prediction value.

2. The method for assessing the health status of an integrated power supply box based on AI modeling as described in claim 1, characterized in that, S2 specifically includes: S21. The multivariate time series is used as the input path of the neural control differential equation model, wherein the input path is an ordered multidimensional data sequence arranged by time index; S22. Perform cubic spline interpolation on the input path to construct a continuous path representation, so that it is differentiable in the time dimension, in order to meet the modeling requirements of continuous differential control functions in the neural control differential equation model. S23. Input the multidimensional data corresponding to the first time slice in the input path into a fully connected neural network to generate an initial state hidden space vector, which serves as the initial state of the neural control differential equation model. S24. Construct a continuous differential control function, which is implemented by a feedforward neural network. It receives the current state hidden space vector and the path value at the current time point as input and outputs the corresponding state derivative, which is used to define the differential relationship of the state evolution over time. S25. Numerical integration of the continuous differential control function is performed using ordinary differential equations, and the state latent space vectors at each time point are generated sequentially from the initial state, following the time axis of the input path. S26. The state latent space vectors at each time point are arranged in chronological order to form a latent space trajectory. The latent space trajectory is a continuously differentiable vector sequence used to represent the dynamic evolution of the integrated power supply box's operating state over the entire time range.

3. The method for assessing the health status of an integrated power supply box based on AI modeling according to claim 1, characterized in that, S3 specifically includes: S31. The acquired latent space trajectory is converted into a state matrix arranged sequentially by time steps, wherein each column of the state matrix represents a latent space vector of a state at a time point. S32. The state matrix is ​​divided into several local time blocks by a sliding window to capture the dynamic change characteristics of the running state in different time periods. S33. Perform dynamic mode decomposition on each local time block. The dynamic mode decomposition includes: constructing an input matrix and an output matrix. The input matrix is ​​composed of the state latent space vectors in consecutive time steps, excluding the last time point of each time block. The output matrix corresponds to the state latent space vectors after excluding the first time point. The output matrix is ​​approximated by the solution of the input matrix multiplied by the linear mapping matrix through least squares fitting. After obtaining the linear mapping matrix, the eigenvalues ​​and right eigenvectors are calculated. The right eigenvectors constitute the modal basis representing the state evolution mode. S34. Perform eigenvalue decomposition on the modal basis to obtain the frequency information and corresponding modal amplitude of each mode, and construct a mode set composed of several modal vectors; S35. The mode set is constrained by an orthogonalization method so that the mode vectors satisfy the linear independence condition, thereby obtaining several orthogonal mode vectors. S36. A modal health space is formed by several orthogonal modal vectors, and each modal vector corresponds to a set of frequency components and energy distributions.

4. The method for assessing the health status of an integrated power supply box based on AI modeling according to claim 1, characterized in that, S4 specifically includes: S41. Select two different training samples in the modal health space, and the corresponding modal vectors are the first modal vector and the second modal vector, respectively. S42. The manifold interpolation hybrid method includes: embedding the first mode vector and the second mode vector into a unified feature space, and constructing a hybrid mode vector according to a nonlinear interpolation function, wherein the nonlinear interpolation function adopts an adjustable curvature control method. ; in, Represents a mixed-mode vector. Represents the first mode vector. This represents the second mode vector. This represents the control proportionality coefficient, with a value range of (0,1). The interpolation trajectory forms a nonlinear surface path, which is used to generate a continuously distributed mixed mode vector. S43. Perform principal frequency analysis on the mixed mode vector and calculate the proportion of frequency components within a predefined energy distribution range. The principal frequency analysis is based on fast Fourier transform and constrains the spectral characteristics of the mixed mode vector. S44. When the overlap between the main frequency energy concentration region of the mixed mode vector and the original two mode vectors is higher than a preset overlap threshold, the frequency of the mixed mode vector is deemed to meet the frequency consistency condition. S45. Generate an enhanced training sample set based on the mixed modal vectors that satisfy the frequency consistency condition.

5. The method for assessing the health status of an integrated power supply box based on AI modeling according to claim 1, characterized in that, S5 specifically includes: S51. Input the mixed modality vector in the enhanced training sample set into the health classification module. The health classification module includes an input layer, a two-layer fully connected neural network structure, a classification output layer, and a lifetime regression output layer. S52, The input layer receives the mixed mode vector from the enhanced training sample set; S53. The two-layer fully connected neural network structure includes a fully connected layer with a ReLU activation function and a fully connected layer with a Sigmoid activation function, respectively, for extracting nonlinear feature representations of mixed mode vectors; S54. The classification output layer is a fully connected layer with a Softmax activation function, and the output dimension is 3, corresponding to the three categories defined in the health status label: 0 represents normal, 1 represents warning, and 2 represents serious. S55. The lifetime regression output layer and the classification output layer are connected in parallel. The lifetime regression output layer is a linear regression layer containing one output neuron. Each component of the nonlinear feature representation is multiplied by a set weight parameter and then summed. A bias term is added to calculate the remaining lifetime prediction value.

6. The method for assessing the health status of an integrated power supply box based on AI modeling according to claim 1, characterized in that, S6 specifically includes: S61. Input the acquired latent space trajectory into the time semantic playback module. The latent space trajectory is converted into a state matrix arranged sequentially according to time steps. Each column of the state matrix represents a state latent space vector at a time point. S62. Select the state latent space vector corresponding to the current time point as the playback starting point, and construct a historical playback time window of a set length forward. The historical playback time window contains several consecutive time steps of reverse prediction targets. S63. The time semantic playback module includes a time-aware encoder and a state predictor. The time-aware encoder encodes the interval information of each time step within each of the historical playback time windows, converts it into a time feature vector, and concatenates it with each state latent space vector to form a prediction input sequence. S64. The state predictor adopts a symmetrical state residual structure, which includes a forward prediction submodule and a reverse residual correction submodule. The forward prediction submodule is based on a recurrent neural network, receives the prediction input sequence, and generates the initial predicted state trajectory for each time step within the historical playback time window. S65. The residual correction submodule constructs the state evolution residual trajectory for each time step according to the symmetry principle, and adds the state evolution residual trajectory to the corresponding initial predicted state trajectory to form the final predicted state trajectory: ; in, This represents the final predicted state trajectory. This represents the initial predicted state trajectory. This represents the latent space vector of the state at the current time point.

7. The method for assessing the health status of an integrated power supply box based on AI modeling according to claim 1, characterized in that, Specifically, S7 includes: S71. Align the final predicted state trajectory with the actual historical state trajectory, and extract the state latent space vector pairs at each time point according to the time step. S72. Based on the state latent space vector pair, the semantic difference value of the entire trajectory is obtained by calculating the state latent space vector difference for each corresponding time step and accumulating the norm. S73. When the semantic difference value exceeds the set tolerance threshold, it is determined that there is an abnormal deviation between the final predicted state trajectory and the actual historical trajectory, and an abnormal marker is generated. S74. After triggering the anomaly marker, replace the hidden space trajectory of the current time step with the corresponding actual historical state trajectory, and re-execute steps S3 to S5 to update the corresponding health status label and remaining life prediction value.

Images