A Smart Detection and Lifespan Early Warning System for Aluminum Electrolytic Capacitors

Abstract

This invention discloses an intelligent detection and lifespan early warning system for aluminum electrolytic capacitors. The system identifies real-time operating conditions and matches sampling strategies to simultaneously collect time-series sequences of capacitance, equivalent series resistance, and leakage current values. It groups and analyzes nonlinear correlations for different operating conditions to construct a dynamic correlation feature set. A graph neural network is used to extract cross-operating-condition spatial topological features and output a fused degradation index that integrates the degree of electrolyte drying and anodic oxide film degradation. A temporal convolutional network is combined to predict the probability of reaching a failure threshold within a preset window. Multiple early warning thresholds are matched to output early warning commands. This system solves the problems of existing detection technologies relying on manual sampling or fixed threshold judgments, which struggle to capture progressive aging characteristics in real time, and traditional lifespan prediction models failing to integrate the dynamic evolution of electrical parameters under multiple operating conditions, leading to large prediction deviations. This system achieves accurate quantitative assessment of the degradation state of aluminum electrolytic capacitors and early prediction of failure risks.

Description

Technical Field

[0001] This invention relates to the field of capacitor testing and condition monitoring technology, and in particular to an intelligent testing and lifespan early warning system for aluminum electrolytic capacitors. Background Technology

[0002] In the manufacturing and maintenance of aluminum electrolytic capacitors, capacitor performance degradation and failure early warning are directly related to the reliability and safety of electronic systems. Existing testing technologies mainly rely on manual sampling or fixed threshold judgments, making it difficult to capture in real time the progressive aging characteristics of capacitors, such as electrolyte desiccation and anodic oxide film degradation, leading to the neglect or misjudgment of potential failure risks. Meanwhile, traditional life prediction models often use data from single stress-accelerated experiments, failing to effectively integrate the dynamic evolution of electrical parameters under multiple operating conditions. This results in significant discrepancies between predicted and actual service life, hindering accurate quantitative assessment of remaining life and proactive maintenance decisions.

[0003] Therefore, there is an urgent need to provide a technical solution to address the above problems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an intelligent detection and lifespan warning system for aluminum electrolytic capacitors. The technical solution of this system is as follows: The adaptive acquisition module is used to identify the real-time operating conditions of the aluminum electrolytic capacitor, match the corresponding sampling strategy from multiple preset standard operating conditions according to the real-time operating conditions, and synchronously acquire the capacitance value time sequence, equivalent series resistance value time sequence, and leakage current value time sequence of the aluminum electrolytic capacitor under the real-time operating conditions according to the matched sampling strategy. The correlation feature construction module is used to group the capacitance value time series, the equivalent series resistance value time series and the leakage current value time series according to the real-time operating conditions, perform nonlinear correlation analysis on each group of time series, and construct a dynamic correlation feature set of the coupling degradation relationship between capacitance value, equivalent series resistance value and leakage current value in each group of time series, wherein different real-time operating conditions correspond to different dynamic correlation feature sets; The degradation state analysis module is used to input the dynamic correlation feature set into a pre-constructed graph neural network. The graph neural network uses the capacitance value, equivalent series resistance value and leakage current value in the dynamic correlation feature set as nodes and the coupling degradation relationship as the edge between nodes. It extracts the spatial topological structure features across operating conditions through graph convolution operation and outputs the fused degradation index. The fused degradation index is used to simultaneously characterize the degree of electrolyte drying and the degree of anodic oxide film degradation of aluminum electrolytic capacitors. The danger prediction module is used to input the fusion degradation index and the dynamic correlation feature set into the time-series prediction model based on the temporal convolutional network, and output the danger prediction probability value of the aluminum electrolytic capacitor reaching the failure threshold within the preset prediction window. The early warning execution module is used to match the predicted danger probability value with a preset multi-level early warning threshold and output an early warning command corresponding to the matching result.

[0005] The technical solution of this invention identifies real-time operating conditions through an adaptive acquisition module and matches a sampling strategy to synchronously acquire time-series sequences of capacitance, equivalent series resistance, and leakage current values. An association feature construction module analyzes nonlinear correlations for different operating conditions to construct a dynamic association feature set. A degradation state analysis module uses a graph neural network to extract cross-operating-condition spatial topological features and outputs a fused degradation index that integrates the degree of electrolyte drying and the degree of anodic oxide film degradation. A hazard prediction module combines a temporal convolutional network to predict the probability of reaching a failure threshold within a preset window. An early warning execution module matches multi-level early warning thresholds and outputs early warning commands. This solution solves the problems of existing detection technologies relying on manual sampling or fixed threshold judgments, which make it difficult to capture progressive aging characteristics in real time, and traditional life prediction models failing to integrate the dynamic evolution of electrical parameters under multiple operating conditions, leading to large prediction deviations. It achieves accurate quantitative assessment of the degradation state of aluminum electrolytic capacitors and early prediction of failure risks.

[0006] Compared with existing technologies, this invention achieves dynamic matching between sampling strategies and real-time operating conditions through an adaptive acquisition module, solving the problem of data acquisition mismatch when switching between multiple operating conditions with a fixed sampling frequency. Through an association feature construction module, it uses variational mode decomposition and mutual information entropy to construct multimodal coupled degradation relationships, overcoming the limitation of traditional linear correlation analysis in capturing the nonlinear degradation coupling characteristics between parameters. In the degradation state analysis module, it extracts cross-operating-condition spatial topological features and generates degradation indices through graph neural networks, achieving deep fusion of multi-operating-condition degradation information and joint quantitative characterization of electrolyte desiccation and anodic oxide film degradation, overcoming the one-sidedness of single-operating-condition feature representation of degradation states. Finally, through the multi-scale feature interaction and hazard prediction probability output of the temporal convolutional network in the hazard prediction module, it takes into account both short-term fluctuation characteristics and long-term degradation trends, solving the problem of large prediction bias in traditional lifetime prediction models.

[0007] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0009] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of an embodiment of the intelligent detection and lifespan warning system for aluminum electrolytic capacitors according to the present invention. Detailed Implementation

[0010] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0011] Figure 1 This diagram illustrates a structural schematic of an embodiment of an intelligent detection and lifespan warning system for aluminum electrolytic capacitors provided by the present invention. Figure 1 As shown, the intelligent detection and lifespan warning system for aluminum electrolytic capacitors includes: The adaptive acquisition module 110 is used to identify the real-time operating conditions of the aluminum electrolytic capacitor, match the corresponding sampling strategy from a number of preset standard operating conditions according to the real-time operating conditions, and synchronously acquire the capacitance value time sequence, equivalent series resistance value time sequence, and leakage current value time sequence of the aluminum electrolytic capacitor under the real-time operating conditions according to the matched sampling strategy.

[0012] Aluminum electrolytic capacitors refer to capacitors that use aluminum metal as the anode, an oxide film covering the surface as the dielectric, and liquid or solid electrolyte as the cathode. For example, in a switching power supply circuit, an aluminum electrolytic capacitor with a rated voltage of 450V and a nominal capacitance of 470μF is used to smooth the rectified DC voltage. Real-time operating conditions refer to the combination of operating conditions an aluminum electrolytic capacitor is currently under during operation. For example, an aluminum electrolytic capacitor may be operating at a temperature of 85℃, a DC voltage of 380V, and an effective ripple current of 1.2A. Standard operating conditions refer to typical operating condition categories pre-set in the system to classify the operating conditions of aluminum electrolytic capacitors. For example, the system may pre-set temperature operating conditions including 25℃, 65℃, and 85℃; voltage operating conditions including 60%, 80%, and 100% of the rated voltage; and ripple current operating conditions including 50%, 80%, and 100% of the rated ripple current.

[0013] The sampling strategy refers to the pre-set combination of acquisition parameters for different standard operating conditions, used to specify the sampling frequency and duration when collecting data in a time sequence. For example, for the superimposed operating conditions of 85℃, 380V, and 1.2A, the system-matched sampling strategy is a sampling frequency of 10Hz and a sampling duration of 30s. The capacitance value time sequence refers to the sequence formed by arranging the measured capacitance values ​​of aluminum electrolytic capacitors in chronological order. For example, 300 capacitance value data points are collected at a sampling frequency of 10Hz within a 30s sampling duration, and arranged in chronological order to form the capacitance value time sequence. The equivalent series resistance value time sequence refers to the sequence formed by arranging the measured equivalent series resistance values ​​of aluminum electrolytic capacitors in chronological order. For example, 300 equivalent series resistance value data points collected synchronously with the capacitance values ​​are arranged in chronological order to form the equivalent series resistance value time sequence. A leakage current value time sequence refers to a sequence of leakage current measurement data of aluminum electrolytic capacitors arranged in chronological order; for example, 300 leakage current data points collected synchronously with the capacitance value are arranged in chronological order to form a leakage current value time sequence.

[0014] The correlation feature construction module 120 is used to group the capacitance value time series, the equivalent series resistance value time series, and the leakage current value time series according to the real-time operating conditions, perform nonlinear correlation analysis on each group of time series, and construct a dynamic correlation feature set of the coupling degradation relationship between capacitance value, equivalent series resistance value and leakage current value in each group of time series, wherein different real-time operating conditions correspond to different dynamic correlation feature sets.

[0015] Nonlinear correlation analysis refers to the process of calculating the degree of correlation between time series sequences with different parameters using nonlinear analysis methods. For example, after performing variational mode decomposition on the time series sequences of capacitance values ​​and equivalent series resistance values ​​collected under 85℃ conditions, the mutual information entropy of the two on the same mode components is calculated, resulting in a nonlinear correlation coefficient of 0.72. Dynamic correlation feature set refers to a feature set containing information on the coupling degradation relationship between capacitance values, equivalent series resistance values, and leakage current values. This feature set is dynamically updated over time. For example, under 85℃ conditions, a three-dimensional correlation feature set is constructed using nonlinear correlation analysis, containing a coupling coefficient of 0.72 between capacitance values ​​and equivalent series resistance values, a coupling coefficient of 0.58 between capacitance values ​​and leakage current values, and a coupling coefficient of 0.64 between equivalent series resistance values ​​and leakage current values.

[0016] The degradation state analysis module 130 is used to input the dynamic correlation feature set into a pre-constructed graph neural network. The graph neural network uses the capacitance value, equivalent series resistance value and leakage current value in the dynamic correlation feature set as nodes and the coupling degradation relationship as the edge between the nodes. It extracts the spatial topological structure features across operating conditions through graph convolution operation and outputs the fused degradation index. The fused degradation index is used to simultaneously characterize the degree of electrolyte drying and the degree of anodic oxide film degradation of the aluminum electrolytic capacitor.

[0017] Graph neural networks (GNNs) refer to neural network models capable of processing graph-structured data and facilitating information transfer between nodes. For example, a graph structure with three nodes (capacitance, equivalent series resistance, and leakage current) and edge weights (coupling coefficients) can be constructed, and spatial topological relationships between nodes can be extracted using graph convolutional layers. Spatial topological features refer to the feature representations reflecting the spatial relationships and dependencies between nodes in a graph structure, extracted through graph convolution operations in a graph neural network. For example, after two layers of graph convolution operations, spatial topological feature vectors reflecting the coupling relationships among the three nodes (capacitance, equivalent series resistance, and leakage current) can be extracted from their respective edge weights. The fusion degradation index is a quantitative indicator used to comprehensively characterize the degradation degree of aluminum electrolytic capacitors, output after fusing spatial topological features from multiple operating conditions using a graph neural network. For example, a system output fusion degradation index of 0.76 comprehensively reflects the combined degradation state of the capacitor under operating conditions of 85℃, 380V, 1.2A and 65℃, 300V, 0.8A.

[0018] The degree of electrolyte desiccation refers to the quantitative value of the gradual evaporation and reduction of the electrolyte inside the aluminum electrolytic capacitor due to long-term operation or high-temperature environment. For example, a component of 0.82 in the fusion degradation index indicates that approximately 82% of the electrolyte inside the capacitor has evaporated. The degree of anodic oxide film degradation refers to the quantitative value of the defects or thinning of the oxide film on the anode surface of the aluminum electrolytic capacitor due to voltage stress or temperature stress. For example, a component of 0.63 in the fusion degradation index indicates that the anodic oxide film has significantly deteriorated but has not yet completely failed.

[0019] The danger prediction module 140 is used to input the fusion degradation index and the dynamic correlation feature set into the time-series prediction model based on the temporal convolutional network, and output the danger prediction probability value of the aluminum electrolytic capacitor reaching the failure threshold within a preset prediction window.

[0020] The time-series prediction model refers to a mathematical model that predicts future states based on historical time-series data. For example, a time-series prediction model built based on a temporal convolutional network, after inputting a fused degradation index of 0.76 and a dynamic correlation feature set, outputs a hazard prediction probability value for the next 7 days. The preset prediction window refers to a pre-set length parameter in the time-series prediction model used to predict future time intervals; for example, a preset prediction window of 168 hours predicts whether a capacitor will reach its failure threshold within the next 7 days.

[0021] The failure threshold refers to a pre-set critical value of electrical parameters used to determine if an aluminum electrolytic capacitor has lost its normal function. For example, when the capacitance value drops to 80% of its initial value, the equivalent series resistance value rises to three times its initial value, or the leakage current value rises to five times its initial value, the capacitor is deemed to have reached the failure threshold. The hazard prediction probability value refers to the probability value output by the time-series prediction model, reflecting the likelihood that the aluminum electrolytic capacitor will reach the failure threshold within a preset prediction window. For example, a hazard prediction probability value of 0.87 output by the time-series prediction model indicates that the probability of the capacitor reaching the failure threshold within the next 7 days is 87%.

[0022] The early warning execution module 150 is used to match the preset multi-level early warning thresholds according to the danger prediction probability value, and output an early warning command corresponding to the matching result.

[0023] The multi-level warning thresholds refer to several pre-set critical values ​​used to divide the predicted probability values ​​of hazards into intervals, with different intervals corresponding to different warning levels. For example, the multi-level warning thresholds include a level one warning threshold of 0.5 and a level two warning threshold of 0.8. A level one warning is triggered when the predicted probability value of a hazard is greater than or equal to 0.5 and less than 0.8, and a level two warning is triggered when it is greater than or equal to 0.8. The matching result refers to the warning level determined by comparing the predicted probability value of a hazard with the multi-level warning thresholds. For example, after comparing the predicted probability value of a hazard of 0.87 with the multi-level warning thresholds, the matching result is a level two warning. The warning command refers to the control signal generated based on the matching result to trigger the corresponding warning action. For example, when the matching result is a level two warning, a high-level signal is output to illuminate a red warning light and an emergency maintenance request command is sent to the maintenance terminal.

[0024] The technical solution of this embodiment identifies real-time operating conditions and matches sampling strategies to synchronously collect time-series sequences of capacitance, equivalent series resistance, and leakage current values ​​through an adaptive acquisition module 110. An association feature construction module 120 analyzes nonlinear correlations for different operating conditions to construct a dynamic association feature set. A degradation state analysis module 130 uses a graph neural network to extract cross-operating condition spatial topological features and outputs a fused degradation index that integrates the degree of electrolyte drying and the degree of anodic oxide film degradation. A hazard prediction module 140 combines a temporal convolutional network to predict the probability value of reaching the failure threshold within a preset window. An early warning execution module 150 matches multi-level early warning thresholds and outputs early warning commands. This solution solves the problems of existing detection technologies relying on manual sampling or fixed threshold judgment, which makes it difficult to capture progressive aging characteristics in real time, and traditional life prediction models failing to integrate the dynamic evolution law of electrical parameters under multiple operating conditions, resulting in large prediction deviations. It realizes accurate quantitative assessment of the degradation state of aluminum electrolytic capacitors and early prediction of failure risks.

[0025] In one optional approach, the operating condition adaptive acquisition module 110 presets multiple standard operating conditions, including temperature operating conditions, voltage operating conditions, and ripple current operating conditions. The sampling strategy includes sampling frequency and sampling duration set for different standard operating conditions. When the operating condition adaptive acquisition module 110 matches the corresponding sampling strategy from multiple standard operating conditions according to the real-time operating condition, if the real-time operating condition belongs to the superposition state of multiple standard operating conditions, the sampling strategy with the highest sampling frequency and the longest sampling duration after superposition is selected for acquisition.

[0026] Among them, temperature condition refers to the preset temperature range category of the ambient temperature or casing temperature of the aluminum electrolytic capacitor; for example, preset temperature conditions include three categories: low temperature condition 25℃, normal temperature condition 65℃, and high temperature condition 85℃. Voltage condition refers to the preset voltage range category of the ratio of the DC voltage applied across the aluminum electrolytic capacitor to its rated voltage; for example, preset voltage conditions include three categories: low voltage condition 60% of the rated voltage, medium voltage condition 80% of the rated voltage, and high voltage condition 100% of the rated voltage. Ripple current condition refers to the preset current range category of the ratio of the effective value of the AC ripple current flowing through the aluminum electrolytic capacitor to its rated ripple current; for example, preset ripple current conditions include three categories: low ripple condition 50% of the rated ripple current, medium ripple condition 80% of the rated ripple current, and high ripple condition 100% of the rated ripple current.

[0027] Here, sampling frequency refers to the number of times electrical parameter data is collected per unit time; for example, a sampling frequency of 10Hz is set for an 85℃ high-temperature operating condition, meaning 10 data points are collected per second. Sampling duration refers to the length of time a single sampling process lasts; for example, a sampling duration of 30s is set for an 85℃ high-temperature operating condition, meaning each sampling lasts 30s. Superimposed state refers to the state when the real-time operating condition of an aluminum electrolytic capacitor simultaneously belongs to multiple standard operating condition categories; for example, when the real-time operating conditions are 85℃, 380V, and 1.2A, it simultaneously belongs to the superimposed state of three standard operating conditions: high-temperature condition, high-voltage condition, and high-ripple condition.

[0028] Among the above-mentioned optional methods, by further presetting temperature conditions, voltage conditions, and ripple current conditions and their corresponding sampling frequencies and sampling durations, when the real-time conditions are in a state of multiple conditions superimposed, the strategy of selecting the sampling frequency with the highest frequency and the longest sampling duration after superposition is adopted. This solves the problem that a single fixed sampling mode is difficult to adapt to complex changes in operating conditions, and achieves synchronous guarantee of data acquisition accuracy and integrity in the scenario of multiple conditions superimposed.

[0029] In an optional manner, the nonlinear correlation analysis in the correlation feature construction module 120 employs variational mode decomposition to perform mode decomposition on the time series sequences of the capacitance values, the time series sequences of the equivalent series resistance values, and the time series sequences of the leakage current values, respectively, extracting the instantaneous frequency and instantaneous amplitude of each mode component, and calculating the mutual information entropy between different parameters on the same mode component based on the instantaneous frequency and instantaneous amplitude as a quantitative representation of the coupling degradation relationship.

[0030] Mode decomposition refers to a signal processing method that decomposes a time series into multiple modal components with different frequency components. For example, variational mode decomposition can be used to decompose a capacitance value time series acquired under 85℃ conditions into four modal components, each corresponding to a different frequency band. Instantaneous frequency refers to the instantaneous oscillation frequency of a modal component at a certain moment; for example, the instantaneous frequency of the first modal component of the capacitance value time series at time 5s is 0.2Hz. Instantaneous amplitude refers to the instantaneous amplitude of a modal component at a certain moment; for example, the instantaneous amplitude of the first modal component of the capacitance value time series at time 5s is 2.3mV.

[0031] Mutual information entropy refers to a nonlinear metric that measures the degree of interdependence between two random variables. For example, under 85℃ operating conditions, the mutual information entropy of the capacitance value time series and the equivalent series resistance value time series on the first modal component is 0.52, indicating a strong nonlinear correlation between the two in this frequency band. Quantitative characterization refers to expressing qualitative or non-numerical relationships in numerical form; for example, the coupling degradation relationship between capacitance value and equivalent series resistance value can be quantified as a numerical representation of mutual information entropy 0.52.

[0032] In the above-mentioned optional methods, variational mode decomposition is further used to perform mode decomposition on the time series of capacitance value, equivalent series resistance value and leakage current value, extract the instantaneous frequency and instantaneous amplitude of each mode component and calculate the mutual information entropy on the same mode component as a quantitative characterization of the coupling degradation relationship. This solves the problem that traditional linear correlation analysis is difficult to capture the nonlinear coupling characteristics between parameters, and realizes the refined quantification of the degradation mechanism at the multi-mode component level.

[0033] In one alternative approach, the dynamic association feature set constructed by the association feature construction module 120 is stored using a four-dimensional tensor structure. The first dimension of the four-dimensional tensor structure is the real-time operating condition index, the second dimension is the time step, the third dimension is the parameter type, and the fourth dimension is the modal component index. The four-dimensional tensor structure is segmented using a sliding window in the time step dimension.

[0034] The real-time operating condition index refers to a unique number used to identify and distinguish different real-time operating conditions; for example, the real-time operating condition of 85℃, 380V, and 1.2A is numbered as operating condition 1, and the real-time operating condition of 65℃, 300V, and 0.8A is numbered as operating condition 2. The time step refers to the sequence number of each sampling moment in the time sequence; for example, 300 data points collected at a sampling frequency of 10Hz within a 30s sampling period correspond to time steps 1 to 300 in chronological order. The parameter type refers to the different categories of electrical parameters; for example, capacitance, equivalent series resistance, and leakage current belong to three different parameter types. The modal component index refers to a unique number used to identify and distinguish different modal components; for example, the four modal components obtained after variational mode decomposition correspond to modal component indices 1, 2, 3, and 4, respectively.

[0035] The sliding window refers to a method of extracting data segments by moving a fixed-length window across a time series. For example, the sliding window length is set to 50 time steps and the step size is 10 time steps, and multiple overlapping data segments are extracted sequentially from a time series of 300 time steps.

[0036] Among the above optional methods, a four-dimensional tensor structure is further adopted to store the dynamically associated feature set. The data organization dimension is constructed with real-time operating condition index, time step, parameter type and modal component index, and a sliding window is used for segmentation on the time step dimension. This solves the problem of chaotic data storage and low access efficiency of multi-source heterogeneous feature data, and realizes the structured management and segmentation processing of high-dimensional time series features.

[0037] In one alternative embodiment, the graph neural network in the degradation state parsing module 130 includes an input layer, multiple stacked graph convolutional layers, and an output layer. The input layer is used to construct graph structure data with capacitance value, equivalent series resistance value, and leakage current value as nodes and the coupling degradation relationship as the initial edge weight value. The graph convolutional layers are used to perform spatial dimension convolution operations on the graph structure data to extract spatial topological features. The output layer is used to map the spatial topological features to the fused degradation index. The graph convolutional layers use skip connections to pass graph structure information of different levels.

[0038] In this context, the input layer refers to the layer structure in a graph neural network that receives graph-structured data and performs initial feature transformation. For example, the input layer receives graph-structured data with capacitance values, equivalent series resistance values, and leakage current values ​​as nodes, and coupling coefficients as edge weights, mapping the node features from the original dimensions to a 64-dimensional hidden space. The graph convolutional layer refers to the layer structure in a graph neural network that performs spatial domain convolution operations on the graph-structured data to extract the topological relationships between nodes. For example, the first graph convolutional layer performs convolution operations on the 64-dimensional node features output by the input layer, outputting a 128-dimensional node feature representation. The output layer refers to the layer structure in a graph neural network that maps the features of the last graph convolutional layer to the final output result. For example, the output layer maps the 256-dimensional node features output by the last graph convolutional layer to a 1-dimensional fusion degradation index through a fully connected network.

[0039] Graph structure data refers to a data organization form composed of nodes and edges, used to represent entities and their relationships. For example, capacitance, equivalent series resistance, and leakage current can be used as three nodes, and the coupling coefficient between them can be used as edges between the nodes, forming a three-node, three-edge graph structure. Graph structure information refers to the comprehensive information contained in graph structure data, including node attributes, edge weights, and topological connections. For example, graph structure information includes the feature vectors of the three nodes, the weights of the three edges, and the connections between the nodes.

[0040] In the above optional approach, a graph neural network architecture containing an input layer, multiple stacked graph convolutional layers, and an output layer is further constructed. Graph structure data is constructed using capacitance value, equivalent series resistance value, and leakage current value as nodes and coupling degradation relationship as the initial value of edge weights. Skip connections are used between graph convolutional layers to pass graph structure information at different levels, which solves the problem of insufficient feature extraction depth of a single convolutional layer and realizes multi-scale fusion of spatial topological features and accurate mapping of degradation index.

[0041] In one alternative approach, each graph convolutional layer employs a graph attention mechanism to assign attention weights to edges between different nodes. These attention weights are dynamically updated based on the similarity of node features. The graph attention mechanism introduces cross-condition constraints to ensure that the dynamically associated feature set corresponding to the same real-time condition maintains consistency in condition identification during the graph convolution process.

[0042] In graph attention mechanisms, attention weights are coefficients assigned to edges between different nodes, reflecting their importance. For example, in the graph convolution process corresponding to the 85℃ operating condition, an attention weight of 0.6 is assigned to the edge between the capacitance value and the equivalent series resistance value, 0.3 to the edge between the capacitance value and the leakage current value, and 0.1 to the edge between the equivalent series resistance value and the leakage current value. Cross-operating condition constraints are restrictions imposed during graph neural network training to ensure consistency of operating condition identification across the dynamically associated feature sets corresponding to the same real-time operating condition during graph convolution. For example, during training, it is required that the operating condition identification information of data from operating condition 1 and operating condition 2 always participates in the convolution operation as part of the node features when passing through the graph convolution layer.

[0043] In the above-mentioned optional methods, a graph attention mechanism is further introduced into the graph convolutional layer. The attention weights are dynamically updated based on the similarity of node features, and cross-working condition constraints are introduced to ensure that the dynamic associated feature sets corresponding to the same real-time working condition maintain the consistency of working condition labels during the graph convolution process. This solves the problem that fixed edge weights are difficult to reflect the differences in node importance, and realizes the adaptive weighting of graph structure information under attention guidance and the unification of cross-working condition feature labels.

[0044] In one alternative embodiment, the temporal convolutional network in the hazard prediction module 140 includes multiple dilated convolutional layers and residual connection structures. The dilated convolutional layers expand the receptive field of the temporal prediction model through an exponentially increasing dilation rate. The residual connection structures superimpose the input and output of the dilated convolutional layers. The temporal convolutional network performs parallel convolution operations on the fusion degradation index and the dynamic correlation feature set in the temporal dimension.

[0045] Dilated convolutional layers refer to convolutional layer structures that expand the receptive field of the convolutional kernel by introducing holes into the input sequence during convolution operations. For example, the first dilated convolutional layer might have a dilation rate of 1, the second dilated convolutional layer a dilation rate of 2, and the third dilated convolutional layer a dilation rate of 4. Residual connection structures refer to connections where the input of a convolutional layer is directly superimposed on its output, used to alleviate the vanishing gradient problem in deep networks. For example, the input and output features of a dilated convolutional layer are added element-wise before being passed to the next layer.

[0046] In dilation, the dilation rate refers to the distance parameter between elements of the convolution kernel in dilated convolution, used to control the expansion speed of the receptive field. For example, with a dilation rate of 2, the convolution kernel samples the input sequence once every other element, doubling the receptive field. The receptive field refers to the size of the input region corresponding to the output feature of a certain layer in a neural network. For example, after three dilated convolutional layers with dilation rates of 1, 2, and 4, the receptive field corresponding to the output feature covers the range of 15 time steps in the input sequence.

[0047] Parallel convolution operation refers to the process of using multiple convolution kernels with different parameters to perform convolution operations on input features simultaneously in a temporal convolutional network; for example, using three dilated convolutional layers with different dilation rates to perform parallel convolution on the fusion degradation index to extract short-term, medium-term, and long-term time-dependent features respectively.

[0048] In the above-mentioned optional approach, multiple dilated convolutional layers and residual connection structures are further configured in the temporal convolutional network. The receptive field of the temporal prediction model is expanded by the exponentially increasing dilation rate, and the input and output of the dilated convolutional layer are superimposed. Parallel convolution operations are performed on the fused degradation index and the dynamically associated feature set in the temporal dimension. This solves the problem of insufficient long-range dependency capability of traditional convolutional networks in temporal modeling and realizes high-precision capture and parallel processing of degradation trends over a wide time span.

[0049] In one alternative approach, before inputting the fusion degradation index and the dynamic association feature set into the temporal convolutional network, the danger prediction module 140 first aligns the fusion degradation index and the dynamic association feature set according to the same timestamp through a feature alignment layer, and performs normalization processing on the aligned features. The feature alignment layer uses a linear interpolation method to fill in data points with mismatched timestamps.

[0050] The feature alignment layer refers to a layer structure used to unify the alignment relationship of multiple features in the time dimension before inputting them into the time series prediction model. For example, the feature alignment layer receives a fused degradation index sequence and a dynamically associated feature set, aligning them according to the same timestamp. Aligned features refer to feature data that maintain a consistent correspondence in the time dimension after processing by the feature alignment layer; for example, the data at time step 10 in the fused degradation index sequence is matched with the data at time step 10 in the dynamically associated feature set to form a pair of input features.

[0051] Among them, the linear interpolation method refers to the interpolation method that estimates the value of the intermediate position based on the linear relationship between known data points. For example, when the fusion degradation index sequence has data points at time step 5 but the dynamic association feature set has no data points at time step 5, the values ​​of time step 4 and time step 6 in the dynamic association feature set are used for linear interpolation to estimate the value of time step 5.

[0052] In the above-mentioned optional approach, a feature alignment layer is further set at the input end of the time series prediction model. First, the fusion degradation index and the dynamic association feature set are aligned according to the same timestamp, and the aligned features are normalized. Then, the linear interpolation method is used to fill the data points with mismatched timestamps. This solves the prediction bias problem caused by the inconsistency of time granularity of multi-source features and realizes accurate alignment and collaborative modeling of heterogeneous time series data under a unified time reference.

[0053] In an alternative approach, when the graph neural network in the degradation state parsing module 130 outputs the fused degradation index, the spatial topological features output by the graph convolutional layer are fused across operating conditions using the following formula: in, This represents the fusion feature vector corresponding to the fusion degradation index. This represents the total number of real-time operating conditions. This indicates the total number of layers in the graph convolutional layer. Indicates the first The weighting coefficient of each real-time operating condition. Indicates the first The real-time operating condition is at the... Inter-layer fusion weights of layered graph convolutional layers Represents a non-linear activation function. Indicates the first The real-time operating condition is at the... The learnable weight matrix of a layered graph convolutional layer. Indicates the first The dynamic correlation feature set corresponding to the real-time operating condition is processed by the first... Spatial topological features output after a layer graph convolutional layer. Indicates the first The real-time operating condition is at the... The adjacency matrix of a layer graph convolutional layer Indicates the first The high-level features output by the last graph convolutional layer for each real-time working condition. Indicates the first The high-level features output by the last graph convolutional layer for each real-time working condition. This represents the hyperbolic tangent activation function. This indicates an element-wise multiplication operation. The learnable weight matrix represents the interaction across different operating conditions. Indicates the first The real-time operating condition and the first Cross-condition interaction weighting coefficients between real-time operating conditions.

[0054] It should be noted that the above formula extracts spatial topological features from the dynamic correlation feature set corresponding to each real-time operating condition through multiple graph convolutional layers. Within each graph convolutional layer, operating condition weight coefficients and inter-layer fusion weights are introduced to weight and accumulate the features. Simultaneously, after the last graph convolutional layer, a cross-operating condition interaction term is introduced. A hyperbolic tangent activation function is used to perform element-wise multiplication of the high-level features of any two different operating conditions, thus combining vertical fusion within operating conditions with horizontal interaction between operating conditions to construct a complete cross-operating condition feature fusion mechanism. Through the collaborative learning of operating condition weight coefficients, inter-layer fusion weights, and cross-operating condition interaction weights, the above formula enables the graph neural network to simultaneously capture the evolutionary patterns of degradation features at different levels within the same operating condition, as well as the synergistic coupling relationship of degradation features between different operating conditions. Finally, it outputs a fusion degradation index that integrates degradation information from multiple operating conditions, achieving a comprehensive quantitative characterization of the electrolyte drying degree and anodic oxide film degradation degree of aluminum electrolytic capacitors.

[0055] In the above-mentioned optional methods, the spatial topological features output by the graph convolutional layer are further weighted and fused and cross-condition interactively calculated by using a cross-condition fusion formula. By introducing condition weight coefficients, inter-layer fusion weights and cross-condition interaction weights, the problem of the one-sidedness of single condition features in representing the degradation state is solved, and the deep fusion and comprehensive quantitative evaluation of multi-condition degradation information is realized.

[0056] In an alternative approach, when the temporal prediction model in the hazard prediction module 140 outputs a hazard prediction probability value, the fused degradation index and the dynamically associated feature set are subjected to multi-scale feature interaction during the temporal prediction process using the following formula: in, Indicates the preset prediction window The predicted probability value of the aluminum electrolytic capacitor reaching the failure threshold is described later. This represents the number of parallel convolutional branches in a temporal convolutional network. Indicates the first The fusion weights of parallel convolutional branches Indicates the first Temporal convolutional network operations with parallel convolutional branches This represents the fusion degradation index. This represents the total number of real-time operating conditions. Indicates the first The dynamic correlation feature set corresponding to each real-time operating condition at time [time] The difference vector, This indicates that the difference vector at time [time] Time derivative, This indicates a feature concatenation operation. Indicates the first Learnable convolutional kernels corresponding to each real-time operating condition Indicates the first The dynamic weight matrix corresponding to each real-time operating condition. This represents the convolution operation. This indicates an element-wise multiplication operation.

[0057] It should be noted that the above formula performs feature transformation on the fused degradation index and the dynamic correlation feature set separately. The fused degradation index enhances the sensitivity of the degradation trend change rate by multiplying by a coefficient corrected by the mean of the time derivatives of the difference vectors for each operating condition. The dynamic correlation feature set is modulated by element-wise multiplication of a learnable convolutional kernel and a dynamic weight matrix before convolution to extract multi-scale degradation patterns. The two processed feature sets are concatenated and input into multiple parallel temporal convolutional network branches. The outputs of each branch are weighted by the fused weights and mapped to probability values ​​using the Sigmoid function. The above formula captures the changing trend of the degradation rate by introducing the time derivative of the difference vector, achieves adaptive modulation of features using the dynamic weight matrix, and extracts degradation patterns at different time scales by using parallel temporal convolutional network branches. This enables the time series prediction model to take into account both short-term fluctuation characteristics and long-term degradation trends, outputting high-precision hazard prediction probability values ​​for matching multi-level warning thresholds.

[0058] In the above-mentioned optional methods, the fused degradation index and the dynamically associated feature set are further processed in parallel branching operations and feature concatenation in a temporal convolutional network through a multi-scale feature interaction formula. The time derivative of the difference vector, the dynamic weight matrix, and the learnable convolutional kernel are introduced. After the mean of the fused degradation index and the time derivative of the difference vector is multiplied and corrected, the result of the convolution with the dynamically associated feature set modulated by the dynamic weight matrix is ​​concatenated and input into the multi-parallel branch temporal convolutional network for operation. This solves the problem that single-scale temporal features cannot take into account both short-term fluctuations and long-term trends, and realizes multi-dimensional dynamic prediction of failure risk probability.

[0059] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0060] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and do not imply a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0061] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An aluminum electrolytic capacitor intelligent detection and life warning system, characterized in that, The system includes: The adaptive acquisition module is used to identify the real-time operating conditions of the aluminum electrolytic capacitor, match the corresponding sampling strategy from multiple preset standard operating conditions according to the real-time operating conditions, and synchronously acquire the capacitance value time sequence, equivalent series resistance value time sequence, and leakage current value time sequence of the aluminum electrolytic capacitor under the real-time operating conditions according to the matched sampling strategy. The correlation feature construction module is used to group the capacitance value time series, the equivalent series resistance value time series and the leakage current value time series according to the real-time operating conditions, perform nonlinear correlation analysis on each group of time series, and construct a dynamic correlation feature set of the coupling degradation relationship between capacitance value, equivalent series resistance value and leakage current value in each group of time series, wherein different real-time operating conditions correspond to different dynamic correlation feature sets; The degradation state analysis module is used to input the dynamic correlation feature set into a pre-constructed graph neural network. The graph neural network uses the capacitance value, equivalent series resistance value and leakage current value in the dynamic correlation feature set as nodes and the coupling degradation relationship as the edge between nodes. It extracts the spatial topological structure features across operating conditions through graph convolution operation and outputs the fused degradation index. The fused degradation index is used to simultaneously characterize the degree of electrolyte drying and the degree of anodic oxide film degradation of aluminum electrolytic capacitors. The danger prediction module is used to input the fusion degradation index and the dynamic correlation feature set into the time-series prediction model based on the temporal convolutional network, and output the danger prediction probability value of the aluminum electrolytic capacitor reaching the failure threshold within the preset prediction window. The early warning execution module is used to match the predicted danger probability value with a preset multi-level early warning threshold and output an early warning command corresponding to the matching result.

2. The aluminum electrolytic capacitor intelligent detection and life warning system according to claim 1, characterized in that, The adaptive acquisition module for operating conditions includes multiple preset standard operating conditions such as temperature, voltage, and ripple current. The sampling strategy includes sampling frequency and sampling duration set for different standard operating conditions. When the adaptive acquisition module for operating conditions matches the corresponding sampling strategy from multiple standard operating conditions based on the real-time operating conditions, if the real-time operating conditions are simultaneously superimposed on multiple standard operating conditions, the sampling strategy with the highest sampling frequency and longest sampling duration after superposition is selected for acquisition.

3. The aluminum electrolytic capacitor intelligent detection and life warning system according to claim 2, characterized in that, The nonlinear correlation analysis in the associated feature construction module uses variational mode decomposition to perform mode decomposition on the time series sequences of the capacitance value, the equivalent series resistance value, and the leakage current value, respectively, to extract the instantaneous frequency and instantaneous amplitude of each mode component, and calculates the mutual information entropy between different parameters on the same mode component based on the instantaneous frequency and instantaneous amplitude as a quantitative representation of the coupling degradation relationship.

4. The aluminum electrolytic capacitor intelligent detection and life warning system according to claim 3, characterized in that, The dynamic associated feature set constructed by the associated feature construction module is stored in a four-dimensional tensor structure. The first dimension of the four-dimensional tensor structure is the real-time working condition index, the second dimension is the time step, the third dimension is the parameter type, and the fourth dimension is the modal component index. The four-dimensional tensor structure is segmented in the time step dimension using a sliding window.

5. The aluminum electrolytic capacitor intelligent detection and life warning system according to claim 4, characterized in that, The graph neural network in the degradation state analysis module includes an input layer, multiple stacked graph convolutional layers, and an output layer. The input layer is used to construct graph structure data with capacitance value, equivalent series resistance value, and leakage current value as nodes and the coupling degradation relationship as the initial edge weight value. The graph convolutional layers are used to perform spatial dimension convolution operations on the graph structure data to extract spatial topological features. The output layer is used to map the spatial topological features to the fused degradation index. The graph convolutional layers use skip connections to pass graph structure information of different levels.

6. The intelligent detection and lifespan early warning system for aluminum electrolytic capacitors according to claim 5, characterized in that, Each graph convolutional layer employs a graph attention mechanism to assign attention weights to edges between different nodes. These attention weights are dynamically updated based on the similarity of node features. The graph attention mechanism introduces cross-condition constraints to ensure that the dynamically associated feature set corresponding to the same real-time condition maintains consistency in condition identification during the graph convolution process.

7. The intelligent detection and lifespan warning system for aluminum electrolytic capacitors according to claim 1, characterized in that, The temporal convolutional network in the hazard prediction module includes multiple dilated convolutional layers and residual connection structures. The dilated convolutional layers expand the receptive field of the temporal prediction model through an exponentially increasing dilation rate. The residual connection structures superimpose the input and output of the dilated convolutional layers. The temporal convolutional network performs parallel convolution operations on the fusion degradation index and the dynamic correlation feature set in the temporal dimension.

8. The intelligent detection and lifespan warning system for aluminum electrolytic capacitors according to claim 7, characterized in that, Before inputting the fusion degradation index and the dynamic association feature set into the temporal convolutional network, the danger prediction module first aligns the fusion degradation index and the dynamic association feature set according to the same timestamp through a feature alignment layer, and then normalizes the aligned features. The feature alignment layer uses a linear interpolation method to fill in data points with mismatched timestamps.

9. The intelligent detection and lifespan warning system for aluminum electrolytic capacitors according to claim 6, characterized in that, When the graph neural network in the degradation state analysis module outputs the fused degradation index, it performs cross-condition fusion of the spatial topological features output by the graph convolutional layer using the following formula: in, This represents the fusion feature vector corresponding to the fusion degradation index. This represents the total number of real-time operating conditions. This indicates the total number of layers in the graph convolutional layer. Indicates the first The weighting coefficient of each real-time operating condition. Indicates the first The real-time operating condition is at the... Inter-layer fusion weights of layered graph convolutional layers Represents a non-linear activation function. Indicates the first The real-time operating condition is at the... The learnable weight matrix of a layered graph convolutional layer. Indicates the first The dynamic correlation feature set corresponding to the real-time operating condition is processed by the first... Spatial topological features output after a layer graph convolutional layer. Indicates the first The real-time operating condition is at the... The adjacency matrix of a layer graph convolutional layer Indicates the first The high-level features output by the last graph convolutional layer for each real-time working condition. Indicates the first The high-level features output by the last graph convolutional layer for each real-time working condition. This represents the hyperbolic tangent activation function. This indicates an element-wise multiplication operation. The learnable weight matrix represents the interaction across different operating conditions. Indicates the first The real-time operating condition and the first Cross-condition interaction weighting coefficients between real-time operating conditions.

10. The intelligent detection and lifespan warning system for aluminum electrolytic capacitors according to claim 8, characterized in that, When the temporal prediction model in the hazard prediction module outputs a hazard prediction probability value, the fused degradation index and the dynamically associated feature set are subjected to multi-scale feature interaction during the temporal prediction process using the following formula: in, Indicates the preset prediction window The predicted probability value of the aluminum electrolytic capacitor reaching the failure threshold is described later. This represents the number of parallel convolutional branches in a temporal convolutional network. Indicates the first The fusion weights of parallel convolutional branches Indicates the first Temporal convolutional network operations with parallel convolutional branches This represents the fusion degradation index. This represents the total number of real-time operating conditions. Indicates the first The dynamic correlation feature set corresponding to each real-time operating condition at time [time] The difference vector, This indicates that the difference vector at time [time] Time derivative, This indicates a feature concatenation operation. Indicates the first Learnable convolutional kernels corresponding to each real-time operating condition Indicates the first The dynamic weight matrix corresponding to each real-time operating condition. This represents the convolution operation. This indicates an element-wise multiplication operation.

Images