A neural network-based capacitor implosion anomaly early warning system

Abstract

The application relates to the technical field of artificial intelligence safety monitoring, and specifically provides a capacitor implosion abnormality early warning system based on a neural network, which comprises the following: a waveform abnormality detection module adopts a long short-term memory network encoder-decoder framework to perform abnormality detection on a discharge waveform, and outputs a waveform abnormality index; a micro short circuit evolution trend tracking module extracts a change rate feature based on equivalent series resistance and capacitor time sequence values, and outputs a micro short circuit risk coefficient; a dynamic causal intervention and counterfactual reasoning engine constructs a structural causal model, generates a counterfactual waveform abnormality index through an intervention operation, and adjusts and updates model parameters in a self-adaptive manner according to a causal effect degree measurement driving threshold value; and an implosion risk comprehensive reasoning module performs implosion risk comprehensive reasoning through a multi-physical field coupling hard constraint graph neural network, and outputs an implosion risk grade and an early warning confidence degree. High-sensitivity early warning of capacitor implosion risk is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence safety monitoring technology, specifically to a capacitor implosion anomaly early warning system based on neural networks. Background Technology

[0002] Capacitors, as energy storage units, are widely used in pulsed power systems, power electronic equipment, and new energy power generation systems. Under harsh conditions such as long-term operation with high voltage and high current pulses, capacitors face problems such as aging, dielectric breakdown, and increased equivalent series resistance. In severe cases, this can lead to micro-short circuits or even implosion failure, causing equipment damage and even threatening personnel safety. Therefore, achieving early warning of capacitor implosion risk has significant engineering value.

[0003] Existing capacitor condition monitoring and early warning technologies suffer from the following shortcomings: Fault detection relies on threshold judgments of a single physical quantity, which is a reactive measure and cannot provide effective early warnings. Analysis methods based on time-series features remain at the level of feature extraction and hierarchical classification, lacking the ability to dynamically model and predict the degradation process. While deep learning-based lifetime prediction and fault diagnosis methods can capture some nonlinear relationships, they are mostly purely data-driven black-box models, lacking physical constraints and exhibiting insufficient generalization ability under operating conditions outside the training data coverage, potentially leading to predictions that deviate from physical laws. Existing early warning systems are mostly open-loop architectures, lacking feedback adjustment mechanisms based on early warning results, and cannot dynamically optimize their parameters according to actual operating data, resulting in decreased early warning accuracy after long-term operation. Furthermore, the gradual evolutionary relationship between micro-short circuits and implosion is not fully considered. Micro-short circuits are important precursor signals to implosion, but their characteristics are weak and easily masked by noise; existing methods have not yet established a dynamic quantitative correlation between micro-short circuit symptoms and implosion risk. The combination of methods such as waveform anomaly detection, time-series feature extraction and physically constrained neural networks lacks systematicity, and a complete closed-loop system from micro-short circuit precursor identification to implosion risk warning has not yet been formed.

[0004] Therefore, there is a lack of a capacitor implosion anomaly early warning system in the existing technology that can integrate multi-source sensor data, embed physical mechanism constraints, and have closed-loop adaptive capabilities. Summary of the Invention

[0005] This invention proposes a capacitor implosion anomaly early warning system based on neural networks, which effectively solves the problems of existing capacitor condition monitoring and early warning technologies, such as post-event detection as the main method, lack of physical constraints in pure data-driven models, inability to adaptively optimize open-loop architectures, and difficulty in effectively identifying micro short-circuit precursor signals.

[0006] The technical solution adopted by this invention is as follows: This solution provides a capacitor implosion anomaly early warning system based on neural networks, including a multi-source sensing acquisition layer, an edge computing and data preprocessing unit, a waveform anomaly detection module, a micro-short circuit evolution trend tracking module, a dynamic causal intervention and counterfactual reasoning engine, an implosion risk comprehensive reasoning module, and an early warning decision and feedback optimization module.

[0007] The multi-source sensing acquisition layer collects multi-dimensional operational data of the capacitor; The edge computing and data preprocessing unit performs time synchronization, outlier removal and signal denoising on multidimensional running data to obtain high-quality preprocessed time-series data. The waveform anomaly detection module adopts a long short-term memory network encoder-decoder framework, including a preset anomaly judgment threshold, to detect anomalies in the discharge waveform in the preprocessed high-quality time series data, and outputs the waveform anomaly index and anomaly waveform events. The micro-short circuit evolution trend tracking module calculates the equivalent series resistance time series value and capacitance time series value based on the voltage, current and temperature data in the preprocessed high-quality time series data. By analyzing the evolution trend of the equivalent series resistance time series value and capacitance time series value, it identifies early signs of micro-short circuits and outputs the micro-short circuit risk coefficient. The micro-short circuit evolution trend tracking module includes a micro-short circuit sign identification classifier, which determines whether there are micro-short circuit signs in the current period and outputs the micro-short circuit confidence level. The dynamic causal intervention and counterfactual reasoning engine uses the micro-short circuit risk coefficient and waveform anomaly index as inputs to construct a structural causal model. This model includes a structural causal graph relating the micro-short circuit risk coefficient and the waveform anomaly index. Through intervention operations, a counterfactual scenario is simulated, generating a counterfactual waveform anomaly index. The difference between the counterfactual and actual waveform anomaly indices is used as a causal effect measure. When the causal effect measure exceeds a preset threshold, the anomaly judgment threshold of the waveform anomaly detection module is adaptively lowered according to a preset mapping relationship. The engine also uses abnormal waveform events confirmed by the waveform anomaly detection module as intervention results. Counterfactual reasoning is used to backtrack whether abnormal waveform events could have been detected without threshold adjustment, obtaining backtracking results. Based on these backtracking results, the parameters of the structural causal model and the decision boundary of the micro-short circuit symptom recognition classifier are updated. The implosion risk comprehensive reasoning module receives the waveform anomaly index and the micro-short circuit risk coefficient. It performs comprehensive reasoning on implosion risk through a multi-physics field coupled hard-constraint graph neural network. The multi-physics field coupled hard-constraint graph neural network constructs a capacitor bank topology graph, with each capacitor in the capacitor bank as a graph node and the thermal and electrical coupling relationships between capacitors as graph edges. The thermal balance equation, the equivalent series resistance degradation dynamics equation, and the capacitance value decay equation are used as hard constraints (forced to be satisfied at the network structure level) in the form of a set of partial differential equations for the message passing layer of the graph neural network. At the same time, the residuals are used as soft constraints (loss function regularization terms). The module outputs the implosion risk level and warning confidence. The early warning decision and feedback optimization module generates early warning signals and maintenance suggestions based on the implosion risk level and early warning confidence level. It also obtains the actual occurrence results of implosion events in actual system operation and feeds the actual occurrence results back to the waveform anomaly detection module and the implosion risk comprehensive reasoning module, dynamically adjusting the anomaly judgment threshold of the waveform anomaly detection module and the network parameters of the implosion risk comprehensive reasoning module.

[0008] Furthermore, the waveform anomaly detection module includes an encoder unit, a decoder unit, a waveform reconstruction unit, and an anomaly scoring unit. The encoder unit uses historical discharge waveform sequences under normal operating conditions as training samples and encodes the input discharge waveform sequences into fixed-length context vectors through a multi-layer long short-term memory network. The decoder unit receives the context vectors and decodes and predicts the discharge waveform sequences step by step through the multi-layer long short-term memory network. The waveform reconstruction unit inputs the discharge waveform to be detected into the trained encoder-decoder framework and outputs the reconstructed discharge waveform after encoding-decoding mapping. The anomaly scoring unit calculates the cosine similarity between the discharge waveform to be detected and the reconstructed discharge waveform, and calculates the waveform anomaly index based on the deviation between the cosine similarity and the normal waveform similarity benchmark.

[0009] Furthermore, the micro-short circuit evolution trend tracking module also includes a micro-short circuit symptom identification unit and a risk coefficient calculation unit. The micro-short circuit evolution trend tracking module calculates the equivalent series resistance time series value and capacitance time series value based on the voltage, current and temperature data in the preprocessed high-quality time series data, and extracts the first-order rate of change feature and the second-order rate of change feature to form a degradation trend feature vector. The micro-short circuit symptom identification unit inputs the degradation trend feature vector into the micro-short circuit symptom identification classifier to determine whether there is a micro-short circuit symptom in the current period and outputs the micro-short circuit confidence level. The risk coefficient calculation unit calculates the micro-short circuit risk coefficient based on the micro-short circuit confidence level and the key indicators in the degradation trend feature vector.

[0010] Furthermore, in the dynamic causal intervention and counterfactual reasoning engine, a structural causal model is constructed using the micro-short circuit risk coefficient as the causal variable and the waveform anomaly index as the outcome variable. The causal structural equation in the structural causal model is parameter-estimated using paired samples from preprocessed high-quality time-series data. The intervention operation is used to cut off the causal influence path of the micro-short circuit risk coefficient on the waveform anomaly index. The causal effect is measured as the difference between the actual waveform anomaly index and the counterfactual waveform anomaly index obtained under the intervention operation, which is used to quantify the causal driving strength of the micro-short circuit trend on the waveform anomaly. When the causal effect measurement exceeds a preset threshold, the anomaly judgment threshold of the waveform anomaly detection module is adaptively reduced. After the waveform anomaly detection module confirms the abnormal waveform event, the dynamic causal intervention and counterfactual reasoning engine performs counterfactual reasoning, updating the parameters of the structural causal model and the decision boundary of the micro-short circuit symptom identification classifier based on the backtracking results.

[0011] Furthermore, the multiphysics-coupled hard-constraint graph neural network in the implosion risk comprehensive reasoning module includes a graph construction unit, a hard-constraint embedding unit, a soft-constraint loss unit, and a risk propagation deduction unit. The graph construction unit constructs a topology graph of the capacitor bank using each capacitor in the capacitor bank as a graph node and the spatial proximity and electrical connection relationships between capacitors as graph edges. The hard-constraint embedding unit embeds the numerical solutions of the thermal balance equation, the equivalent series resistance degradation dynamics equation, and the capacitance attenuation equation as fixed nonlinear transformations in the message passing layer of the graph neural network. The soft-constraint loss unit adds a multiphysics-coupled soft-constraint loss term to the loss function of the graph neural network. The risk propagation deduction unit simulates the propagation path and cumulative effect of implosion risk in the capacitor bank through multi-layer message passing of the graph neural network, and outputs the implosion risk probability value and risk propagation direction indication for each capacitor node.

[0012] Compared with the prior art, the beneficial effects of the present invention are: (1) By constructing a multi-source sensing acquisition layer, real-time acquisition and time synchronization of all elements of multi-dimensional operating data such as capacitor voltage, current, temperature and discharge waveform are realized, forming a high-quality monitoring data foundation covering multiple dimensions of electrical and thermal systems. Through edge computing and data preprocessing units to perform time synchronization, outlier removal and signal denoising, the integrity and reliability of the data are improved, providing high-quality input support for subsequent intelligent early warning, and solving the problem that the traditional single physical quantity threshold judgment method can only achieve post-event detection and cannot provide effective early warning.

[0013] (2) A waveform anomaly detection module based on a long short-term memory network encoder-decoder framework is introduced. The encoder performs feature compression and latent space mapping on the normal discharge waveform sequence, and the decoder reconstructs the predicted waveform. Based on the cosine similarity deviation between the actual waveform and the predicted waveform, abnormal discharge events are identified, achieving high-sensitivity early identification of abnormal discharge waveforms. A micro-short circuit evolution trend tracking module is proposed. By calculating the equivalent series resistance time series value and capacitance time series value and extracting their first-order and second-order rate of change features, quantitative identification and tracking of early signs of micro-short circuits are achieved. This transforms the important precursor signal of micro-short circuits into a quantifiable risk coefficient and establishes a dynamic quantitative correlation between micro-short circuit signs and implosion risk. This effectively solves the problems of weak micro-short circuit features, easy noise submersion, and lack of dynamic quantitative correlation in existing methods.

[0014] (3) Deep coupling and collaboration between the waveform anomaly detection module and the micro-short circuit evolution trend tracking module based on the structural causal model are achieved through dynamic causal intervention and counterfactual reasoning engine. The counterfactual waveform anomaly index is generated through intervention operation, and the difference between the actual value and the counterfactual value is used as the causal effect measure, so that the system can distinguish between "waveform anomalies driven by micro-short circuit trends" and "waveform anomalies driven by other factors", and realize the enhancement of accurate early warning based on causal reasoning. The effectiveness of threshold adjustment is backtracked through counterfactual reasoning, and the parameters of the structural causal model and the decision boundary of the classifier are dynamically updated. The implosion risk comprehensive reasoning is carried out through multi-physics field coupled hard constraint graph neural network. The thermal balance equation, the equivalent series resistance degradation dynamic equation and the capacitance value decay equation are used as hard constraints of the graph neural network message passing layer in the form of partial differential equation system (forced to be satisfied at the network structure level), and at the same time, the residual form is used as soft constraints (loss function regularization term), so that the early warning results are not only more accurate, but also have physical interpretability and cross-condition stability, effectively solving the problem that the pure data-driven model lacks physical mechanism constraints and has insufficient generalization ability outside the training data coverage.

[0015] (4) A complete closed-loop architecture of "assessment-reasoning-decision-feedback" was constructed: the early warning decision and feedback optimization module generates early warning signals and maintenance suggestions based on the implosion risk level and early warning confidence, and obtains the actual occurrence results of implosion events in actual operation of the system. The actual occurrence results are fed back to the waveform anomaly detection module and the implosion risk comprehensive reasoning module to dynamically adjust the anomaly judgment threshold and network parameters. Through the bidirectional coupling and collaboration between the waveform anomaly detection module and the micro-short circuit evolution trend tracking module, the reverse adjustment mechanism of the micro-short circuit risk coefficient on the waveform anomaly judgment threshold enables the system to actively improve the sensitivity to discharge waveform anomalies when it detects the degradation trend of the internal structure of the capacitor, realizing the "early warning enhancement" effect of the precursor signal on subsequent detection. The feedback update mechanism of the micro-short circuit symptom identification classifier after waveform anomaly confirmation enables the system to continuously optimize the accuracy of early micro-short circuit identification by using actual abnormal events. It effectively solves the problems that existing early warning systems are mostly open-loop architectures, lack feedback adjustment mechanisms based on early warning results, and have decreased early warning accuracy after long-term operation, realizing full-chain intelligent monitoring from micro-short circuit precursor identification to implosion risk early warning. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall structure of a capacitor implosion anomaly early warning system based on a neural network according to the present invention. The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] Example 1: Please see Figure 1 This embodiment presents a capacitor implosion anomaly early warning system based on neural networks, which is applied to the online monitoring and implosion risk early warning of capacitors in pulse power systems and power electronic equipment. The system includes a multi-source sensing acquisition layer, an edge computing and data preprocessing unit, a waveform anomaly detection module, a micro-short circuit evolution trend tracking module, a dynamic causal intervention and counterfactual reasoning engine, an implosion risk comprehensive reasoning module, and an early warning decision and feedback optimization module.

[0019] A multi-source sensor acquisition layer is fixedly installed on the surface of the capacitor under test, as well as its power supply and discharge circuits. It is equipped with various types of sensor components, including: a patch-type platinum resistance temperature sensor deployed in the middle and top of the capacitor's sidewall to collect time-series data of the capacitor's surface temperature; high-precision voltage and current sensors deployed in the capacitor's power supply circuit to collect time-series data of the capacitor's terminal voltage and circuit current; and high-frequency voltage and current probes deployed in the capacitor's discharge circuit to collect capacitor discharge waveform data. Through the coordinated operation of these multiple sensor components, complete multi-dimensional operational data covering both electrical and thermal parameters is generated, providing a high-precision input basis for subsequent implosion risk early warning.

[0020] The edge computing and data preprocessing unit adopts an integrated industrial-grade embedded computer, equipped with multi-channel analog input interfaces, high-frequency signal acquisition interfaces, and Ethernet communication interfaces. The built-in GNSS timing module performs unified timestamp marking on multi-dimensional operational data, and executes outlier detection and adaptive filtering for noise reduction. When data packet loss occurs, linear interpolation completion and integrity verification are performed. When breakpoint data occurs, local encrypted caching is implemented. Finally, high-quality time-series data is obtained after time alignment, noise reduction, smoothing, and complete continuous preprocessing.

[0021] The waveform anomaly detection module includes an encoder unit, a decoder unit, a waveform reconstruction unit, and an anomaly scoring unit. It adopts a long short-term memory network encoder-decoder framework and uses a preset anomaly decision threshold to detect anomalies in the discharge waveform in the preprocessed high-quality time series data, and outputs the waveform anomaly index and anomaly waveform events. The micro-short circuit evolution trend tracking module includes a micro-short circuit symptom identification unit, a micro-short circuit symptom identification classifier, and a risk coefficient calculation unit. Based on voltage, current, and temperature data from preprocessed high-quality time series data, it calculates the equivalent series resistance time series value and capacitance time series value, extracts first-order and second-order rate of change features to form a degradation trend feature vector, and inputs the degradation trend feature vector into the micro-short circuit symptom identification classifier by the micro-short circuit symptom identification unit to determine whether there is a micro-short circuit symptom in the current period and outputs the micro-short circuit confidence level. The risk coefficient calculation unit calculates the micro-short circuit risk coefficient based on the micro-short circuit confidence level and key indicators in the degradation trend feature vector. The dynamic causal intervention and counterfactual reasoning engine uses the micro-short circuit risk coefficient and waveform anomaly index as inputs to construct a structural causal model. The structural causal model includes a structural causal graph between the micro-short circuit risk coefficient and the waveform anomaly index. By simulating counterfactual scenarios through intervention operations, a counterfactual waveform anomaly index is generated. The difference between the counterfactual waveform anomaly index and the actual waveform anomaly index is used as a causal effect measure. When the causal effect measure exceeds a preset threshold, the anomaly judgment threshold of the waveform anomaly detection module is adaptively reduced according to a preset mapping relationship. The dynamic causal intervention and counterfactual reasoning engine also uses the abnormal waveform events confirmed by the waveform anomaly detection module as intervention results. Through counterfactual reasoning, it backtracks whether the abnormal waveform events could be detected if the threshold adjustment had not been performed, obtains backtracking results, and updates the parameters of the structural causal model and the decision boundary of the micro-short circuit symptom recognition classifier based on the backtracking results. The implosion risk comprehensive reasoning module receives the waveform anomaly index and the micro-short circuit risk coefficient, and performs comprehensive reasoning of implosion risk through a multi-physics field coupled hard-constraint graph neural network. The multi-physics field coupled hard-constraint graph neural network includes a graph construction unit, a hard-constraint embedding unit, a soft-constraint loss unit, and a risk propagation inference unit. The graph construction unit constructs a topology graph of the capacitor bank with each capacitor in the capacitor bank as a graph node and the spatial proximity and electrical connection relationships between capacitors as graph edges. The hard-constraint embedding unit embeds the numerical solutions of the partial differential equation system consisting of the thermal balance equation, the equivalent series resistance degradation dynamics equation, and the capacitance value decay equation as fixed nonlinear transformations in the message passing layer of the graph neural network. The soft-constraint loss unit adds a multi-physics field coupled soft-constraint loss term to the loss function of the graph neural network. The risk propagation inference unit simulates the propagation path and cumulative effect of implosion risk in the capacitor bank through multi-layer message passing of the graph neural network, outputs the implosion risk probability value and risk propagation direction indication for each capacitor node, and outputs the implosion risk level and warning confidence based on the implosion risk probability value. The early warning decision and feedback optimization module generates early warning signals and maintenance suggestions based on the implosion risk level and early warning confidence level. It also obtains the actual occurrence results of implosion events in actual system operation and feeds the actual occurrence results back to the waveform anomaly detection module and the implosion risk comprehensive reasoning module. The module dynamically adjusts the anomaly judgment threshold of the waveform anomaly detection module and the network parameters of the implosion risk comprehensive reasoning module to achieve closed-loop feedback optimization.

[0022] Example 2: This embodiment is based on Embodiment 1. The waveform anomaly detection module uses a long short-term memory network encoder-decoder framework to detect anomalies in the discharge waveform and outputs a waveform anomaly index. The specific workflow includes: Step A1: Construct the training dataset. Collect historical discharge waveform data of the capacitor under normal operating conditions. Each discharge waveform contains 120 sampling points (corresponding to one complete discharge cycle), for a total of 10,000 normal discharge waveform samples. Perform zero-mean standardization on each sample to eliminate dimensional differences, forming a standardized training sample set.

[0023] Step A2: Construct a Long Short-Term Memory (LSTM) network encoder-decoder framework. The encoder consists of two stacked LSM networks, each with a 64-dimensional hidden state and a 1-dimensional input (corresponding to the waveform amplitude of a single sampling point). The decoder also consists of two stacked LSM networks, each with a 64-dimensional hidden state and a 1-dimensional output.

[0024] Step A3: Train the encoder-decoder framework. Using the training sample set constructed in Step A1, the discharge waveform sequence in the training sample set is used as the encoder input, and the same discharge waveform sequence is used as the decoder target output. The mean squared error is used as the loss function, and the Adam optimizer (learning rate 0.001, β1=0.9, β2=0.999) is used for training. The batch size is 64, and the training epochs are 200, until the loss function converges.

[0025] Step A4: Calculate the baseline value for normal waveform similarity. Input all normal discharge waveforms from the training sample set into the trained encoder-decoder framework. After encoding and decoding, each discharge waveform is reconstructed to obtain the corresponding predicted discharge waveform. Calculate the cosine similarity between each discharge waveform and its corresponding predicted discharge waveform to obtain the normal waveform similarity set. Calculate the mean of the normal waveform similarity set. with standard deviation .

[0026] Step A5: Online Anomaly Detection. The discharge waveform to be detected is normalized to zero mean and then input into the trained encoder-decoder framework to obtain the reconstructed discharge waveform. The cosine similarity between the detected discharge waveform and the reconstructed discharge waveform is calculated. Then, the waveform anomaly index is calculated: ; When the waveform abnormality index Exceeding the preset anomaly judgment threshold When the initial value is set to 3.0, it is determined that the current discharge waveform is abnormal and an abnormal waveform event is generated.

[0027] Example 3: This embodiment is based on Embodiment 2. The core computational process of the Long Short-Term Memory (LSTM) network encoder-decoder framework is as follows: Encoder for input discharge waveform sequence Encode, where The length of the input discharge waveform sequence (corresponding to 120 sampling points for a complete discharge cycle). For the first The discharge waveform amplitude at each sampling point (already normalized to zero-crossing mean). The encoder's cyclic calculation process is as follows: ; in, For the encoder in the first The hidden state at any given moment. This is the transformation function of the recurrent unit of the encoder's long short-term memory network, which includes the calculation of the forget gate, input gate, output gate, and memory unit update.

[0028] After the encoder completes encoding the entire input discharge waveform sequence, it outputs a context vector. That is, the hidden state of the encoder at the last moment. The context vector compresses all the time-dependent information of the input discharge waveform sequence.

[0029] The decoder receives the context vector. , as the initial hidden state of the decoder Gradually generate a reconstructed discharge waveform sequence The decoder's output at each time step depends not only on the prediction from the previous time step, but also on the current hidden state and context vector. ; in, For the decoder in the The predicted value output at time step. For the decoder in the The hidden state at any given moment. This is the transformation function for the recurrent unit of the Long Short-Term Memory network in the decoder.

[0030] The anomaly scoring unit calculates the discharge waveform to be detected. With reconstructed discharge waveform Cosine similarity between them: ; in, Cosine similarity, with values ​​ranging from 1 to 10. The closer the value is to 1, the more similar the two waveforms are. The anomaly scoring unit calculates the waveform anomaly index based on the deviation between the cosine similarity and the normal waveform similarity benchmark.

[0031] Example 4: This embodiment is based on Embodiment 1. The micro-short circuit evolution trend tracking module calculates the equivalent series resistance time series value and capacitance time series value based on the voltage, current, and temperature data in the preprocessed high-quality time series data. By analyzing the evolution trend of the equivalent series resistance time series value and capacitance time series value, it identifies early signs of micro-short circuits and outputs a micro-short circuit risk coefficient. The specific workflow includes: Step B1: Parameter Calculation. For each charge / discharge cycle, based on the capacitor terminal voltage... With loop current The transient response during the initial charging phase is measured by the voltage jump amplitude. With the corresponding current value Calculate the equivalent series resistance: ; in, For the first The equivalent series resistance value per charge-discharge cycle. This represents the voltage jump amplitude at the start of charging. This corresponds to the loop current value.

[0032] Based on the voltage change rate of the capacitor during the constant current discharge phase With discharge current Calculate the capacitance value: ; in, For the first The capacitance value per charge-discharge cycle. This refers to the discharge current during the constant current discharge phase. This represents the rate of change of the capacitor terminal voltage during the constant current discharge phase.

[0033] Step B2: Trend Extraction. For a length of... Calculate the first-order rate of change characteristic of the equivalent series resistance using a sliding window of one charge-discharge cycle: ; in, This represents the duration of a single charge / discharge cycle. The second-order rate of change characteristic of the equivalent series resistance is: ; Similarly, calculate the first-order rate of change characteristic of the capacitance value. With second-order rate of change characteristics .

[0034] Step B3: Micro-short circuit symptom identification. This involves identifying four rate-of-change characteristics. The degradation trend feature vector is used as input to the micro-short circuit symptom identification classifier to determine whether micro-short circuit symptoms exist in the current time period, and the micro-short circuit confidence score is output. .

[0035] Step B4: Risk Coefficient Calculation. The risk coefficient calculation unit calculates the micro-short circuit risk coefficient based on key indicators in the micro-short circuit confidence and degradation trend feature vector: ; in, It is a minimum-maximum normalization function; This is the statistical mean of the first-order rate of change of the equivalent series resistance under normal aging conditions. The absolute value of the first-order rate of change of capacitance is the statistical mean under normal aging conditions; the weighting coefficient is taken as... , , When the micro-short circuit risk coefficient Exceeding the preset threshold When the initial value is set to 0.6, it is determined that there are early signs of a micro short circuit.

[0036] Micro-short circuit risk coefficient After the calculation is completed, it is transmitted to the implosion risk comprehensive reasoning module, and compared with the waveform anomaly index output by the waveform anomaly detection module. Together, they constitute the dual-channel input features of the implosion risk comprehensive reasoning. The micro-short circuit risk coefficient reflects the gradual degradation trend of the internal structure of the capacitor (precursor signal), while the waveform anomaly index reflects the instantaneous abnormal state of the discharge behavior (confirmation signal). The two complement each other in the time dimension and jointly drive the subsequent multi-physics field coupled hard-constraint graph neural network reasoning.

[0037] Example 5: This embodiment is based on Embodiment 1. The training process of the multiphysics coupled hard constraint graph neural network is as follows: Step C1: Construct the training dataset. Collect operational data of the capacitor bank under normal operating conditions, aging degradation, and historical implosion events, totaling 5000 samples. For each sample, in addition to labeling the implosion risk (0 indicating no implosion, 1 indicating implosion), the corresponding physical state variables are also labeled, including the equivalent series resistance value. capacitance value and temperature value .

[0038] Step C2: Define hard constraints for the multiphysics coupled partial differential equation system. The hard constraint embedding unit embeds the multiphysics coupled hard constraints in the form of a system of partial differential equations within the message passing layer of the graph neural network. The specific equations are as follows: Heat balance equation: ; in, For heat capacity, For the temperature field, For time, This is the effective value of the current. For equivalent series resistance, For ambient temperature, For thermal resistance, Thermal conductivity, This is the gradient operator.

[0039] Equivalent series resistance degradation kinetic equation: ; in, The degradation rate coefficient, For activation energy, Boltzmann's constant, This is the initial equivalent series resistance. This is the current stress coefficient.

[0040] Capacitance decay equation: ; in, The decay rate coefficient, The activation energy for capacitor decay. This is the critical capacitance value for failure.

[0041] The numerical solution of the partial differential equation system is obtained through the finite difference method. As a fixed nonlinear transformation of the intermediate layer of the graph neural network, the output of the intermediate layer of the graph neural network must structurally satisfy the multiphysics coupling equation.

[0042] Step C3: Define the loss function. The total loss function is a weighted sum of the task loss term and the multiphysics coupling soft constraint loss term.

[0043] Task loss items Using the binary cross-entropy loss function: ; in, This refers to the batch sample size. For the first The true implosion risk label of each sample This represents the probability of implosion predicted by the network.

[0044] Multiphysics Coupled Soft Constraint Loss Term Based on the residuals of the capacitor thermal balance equation, the equivalent series resistance degradation kinetic equation, and the capacitance decay equation: Residual loss of heat balance equation Defined as: ; in, For the network to the first The predicted temperature value for each sample. For the network to the first The predicted power loss value for each sample. For the first The ambient temperature of each sample.

[0045] Equivalent series resistance degradation kinetic equation residual loss for: ; in, The equivalent series resistance degradation kinetic function is obtained by fitting data from accelerated aging experiments on capacitors.

[0046] Capacitance decay equation residual loss for: ; in, The capacitance decay kinetic function is obtained by fitting data from accelerated aging experiments on capacitors.

[0047] Multiphysics Coupled Soft Constraint Loss The weighted sum of the three losses: ; Wherein, the weight coefficient is taken , , .

[0048] The total loss function is: ; Wherein, the weight coefficient is taken .

[0049] Step C4: Model Training. The Adam optimizer (learning rate 0.001, β1=0.9, β2=0.999) is used for training, with a batch size of 32 and 300 training epochs, until the total loss function converges.

[0050] Step C5: Online Inference and Risk Level Mapping. The real-time collected and preprocessed waveform anomaly index, micro-short circuit risk coefficient, temperature, mean voltage, and RMS current of each capacitor are used to construct a node feature vector. This vector is then input into a trained multiphysics coupled hard-constraint graph neural network to obtain the implosion risk probability value for each capacitor node. With warning confidence Risk levels are mapped according to preset thresholds: when When it is judged as high risk, When it is determined to be medium risk, It was determined to be low risk at that time.

[0051] Example 6: This embodiment is based on Embodiment 1, and the dynamic causal intervention and counterfactual reasoning engine realizes deep coupling and collaboration between the waveform anomaly detection module and the micro-short circuit evolution trend tracking module based on the structural causal model.

[0052] In the dynamic causal intervention and counterfactual reasoning engine, the structural causal graph is constructed as follows: With micro short circuit risk coefficient Causal variables, waveform anomaly index For the outcome variable, a structural causal model containing direct causal edges and potential confounding factors is constructed. The causal structural equation in the structural causal model adopts a nonlinear function form, and the parameters are estimated using paired samples (micro-short-circuit risk coefficient, waveform anomaly index) from preprocessed high-quality time series data.

[0053] Intervention Operation This indicates that the micro-short circuit risk coefficient is fixed at a benchmark value. This cuts off the causal path of the micro-short circuit risk coefficient on the waveform anomaly index, leaving only the influence of other factors on the waveform anomaly index.

[0054] Counterfactual waveform anomaly index The calculation formula is: ; in, Let be the expectation operator, representing the conditional expectation value of the waveform anomaly index under intervention.

[0055] Causal effect measurement The calculation formula is: ; when This indicates that the micro-short-circuit trend has a causal driving effect on waveform anomalies, among which... A threshold is preset for causal effects. At this point, the anomaly detection threshold of the waveform anomaly detection module adaptively decreases according to the following mapping relationship: When causal effect is measured At that time, the waveform anomaly judgment threshold remains at its initial value. ; When causal effect is measured At that time, the waveform anomaly judgment threshold decreases linearly; When causal effect is measured When the waveform anomaly judgment threshold is reduced to the preset minimum value, .

[0056] In the dynamic causal intervention and counterfactual reasoning engine, the method for updating the structural causal model parameters and the decision boundary of the micro-short circuit symptom identification classifier based on counterfactual reasoning is as follows: Once the waveform anomaly detection module confirms an abnormal waveform event, the dynamic causal intervention and counterfactual reasoning engine records the moment the event occurs. and the waveform anomaly index at that moment and micro-short circuit risk coefficient .

[0057] The dynamic causal intervention and counterfactual reasoning engine performs counterfactual reasoning: calculating whether the anomalous waveform would be detected under the counterfactual condition of "if threshold adjustment were not performed". Specifically, it will... Calculate the counterfactual waveform anomaly index using a fixed baseline value. ,like (Initial threshold) The counterfactual reasoning result shows that the abnormal waveform will not be detected under the condition that the threshold is not adjusted.

[0058] If counterfactual reasoning results indicate that the anomaly would not be detected under unadjusted threshold conditions, then the event is marked as a "threshold-adjusted valid sample" to strengthen the causal effect weight of the micro-short-circuit risk coefficient on the waveform anomaly index in the structural causal model.

[0059] If counterfactual reasoning results indicate that the anomaly will still be detected under unadjusted threshold conditions, then the event is marked as an "independent anomaly sample" to weaken the causal effect weight of the micro-short-circuit risk coefficient on the waveform anomaly index in the structural causal model.

[0060] Based on the updated structural causal model parameters, the optimal decision boundary of the micro-short circuit symptom identification classifier is recalculated.

[0061] Example 7: This embodiment is based on Embodiment 1. The early warning decision and feedback optimization module includes an early warning decision unit, a maintenance suggestion generation unit, and a feedback optimization unit.

[0062] The workflow of the early warning decision-making unit is as follows: When the implosion risk level is high and the early warning confidence level is high... When a Level 1 warning signal (red warning) is generated, it is recommended to immediately shut down the machine for maintenance; when the implosion risk level is medium risk and the warning confidence level is [not specified], [the warning will be issued]. When the risk level is low, a level 2 warning signal (yellow warning) is generated, and it is recommended to arrange maintenance within 24 hours; when the risk level of implosion is low, only monitoring data is recorded and no warning signal is generated.

[0063] The maintenance recommendation generation unit generates maintenance recommendations, including suggested maintenance time windows and suggested maintenance measures, based on the implosion risk level and warning confidence level, combined with the current operating parameters of the capacitor.

[0064] The workflow of the feedback optimization unit is as follows: Step D1: Data Acquisition. During actual operation, continuously record all early warning signals generated by the early warning decision unit and their corresponding waveform anomaly indices. Micro-short circuit risk coefficient Probability of implosion and early warning confidence Simultaneously, record whether an implosion event actually occurred.

[0065] Step D2: Performance Evaluation. After every 30 days of operation, the collected actual implosion event information is compared with historical early warning records to calculate the early warning accuracy. With the false negative rate TP represents the number of correct warnings (implosion actually occurred after warning), FP represents the number of false alarms (implosion did not actually occur after warning), and FN represents the number of missed warnings (implosion actually occurred without warning).

[0066] Step D3: Threshold Adjustment. When the early warning accuracy falls below 85% or the false negative rate exceeds 10%, a threshold adjustment is triggered. This involves setting up a set of waveform anomaly indices from historical early warning records. With the set of micro-short circuit risk coefficients The samples are grouped according to whether an actual implosion event occurred, resulting in positive and negative sample sets. The optimal waveform anomaly index decision threshold is calculated using the method of maximizing the Youden exponent. and the optimal micro-short circuit risk coefficient decision threshold .Will and The data is sent back to the waveform anomaly detection module and the implosion risk comprehensive reasoning module to replace the original judgment threshold.

[0067] Step D4: Network Parameter Fine-tuning. When the false negative rate remains above 15%, the multi-physics coupled hard-constraint graph neural network parameters of the implosion risk comprehensive reasoning module are fine-tuned. Using newly added actual implosion event data and corresponding operational data as training samples (data collected within the last 90 days), and while maintaining the original network parameters, incremental learning is used for a small number of iterative training rounds (50 training rounds, learning rate reduced to 0.0001) to update the network parameters.

[0068] Through the aforementioned closed-loop feedback mechanism, the system achieves dynamic adaptive optimization of the decision threshold and network parameters, ensuring the accuracy and reliability of early warning decisions during long-term operation.

[0069] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A capacitor implosion anomaly early warning system based on neural networks, characterized in that, It includes a multi-source sensing acquisition layer, an edge computing and data preprocessing unit, a waveform anomaly detection module, a micro-short circuit evolution trend tracking module, a dynamic causal intervention and counterfactual reasoning engine, an implosion risk comprehensive reasoning module, and an early warning decision and feedback optimization module; The multi-source sensing acquisition layer collects multi-dimensional operational data of the capacitor; The edge computing and data preprocessing unit performs time synchronization, outlier removal and signal denoising on multidimensional running data to obtain high-quality preprocessed time-series data. The waveform anomaly detection module performs anomaly detection on the discharge waveform in the preprocessed high-quality time-series data and outputs the waveform anomaly index and abnormal waveform events. The micro-short circuit evolution trend tracking module outputs the micro-short circuit risk coefficient based on preprocessed high-quality time series data; The dynamic causal intervention and counterfactual reasoning engine takes the micro-short circuit risk coefficient and waveform anomaly index as inputs to construct a structural causal model and generate a counterfactual waveform anomaly index, which drives the adaptive adjustment of the anomaly judgment threshold in the waveform anomaly detection module and updates the structural causal model parameters through counterfactual reasoning. The implosion risk comprehensive reasoning module performs implosion risk comprehensive reasoning through a multi-physics field coupled hard-constraint graph neural network, and outputs the implosion risk level and warning confidence. The early warning decision and feedback optimization module generates early warning signals and maintenance suggestions based on the implosion risk level and early warning confidence level, and obtains the actual occurrence results of the implosion event, and provides feedback to adjust the anomaly judgment threshold and network parameters.

2. The capacitor implosion anomaly early warning system based on neural networks according to claim 1, characterized in that: The waveform anomaly detection module includes an encoder unit, a decoder unit, and a waveform reconstruction unit. The encoder unit encodes the input discharge waveform sequence into a context vector. The decoder unit receives the context vector and decodes the predicted discharge waveform sequence. The waveform reconstruction unit reconstructs the output discharge waveform.

3. The capacitor implosion anomaly early warning system based on neural networks according to claim 2, characterized in that: The waveform anomaly detection module also includes an anomaly scoring unit; the anomaly scoring unit calculates the cosine similarity between the discharge waveform to be detected and the reconstructed discharge waveform, and calculates the waveform anomaly index based on the cosine similarity.

4. The capacitor implosion anomaly early warning system based on neural networks according to claim 3, characterized in that: The micro-short circuit evolution trend tracking module includes a micro-short circuit symptom identification classifier, a micro-short circuit symptom identification unit, and a risk coefficient calculation unit; Generate degradation trend feature vectors based on preprocessed high-quality time-series data; Input the degradation trend feature vector into the micro-short circuit symptom identification classifier and output the micro-short circuit confidence score; The risk coefficient calculation unit calculates the micro-short circuit risk coefficient based on the micro-short circuit confidence level.

5. A capacitor implosion anomaly early warning system based on a neural network according to claim 4, characterized in that: The dynamic causal intervention and counterfactual reasoning engine takes the abnormal waveform events confirmed by the waveform anomaly detection module as the intervention result, and obtains the backtracking result through counterfactual reasoning backtracking. The parameters of the structural causal model and the decision boundary of the micro-short circuit symptom identification classifier are updated based on the backtracking results.

6. The capacitor implosion anomaly early warning system based on neural networks according to claim 1, characterized in that: The multiphysics coupled hard-constraint graph neural network includes graph building units; the graph building units construct a topology graph of the capacitor bank by using each capacitor in the capacitor bank as a graph node and the spatial proximity and electrical connection relationships between capacitors as graph edges.

7. A capacitor implosion anomaly early warning system based on a neural network according to claim 6, characterized in that: The multiphysics coupled hard-constrained graph neural network also includes a risk propagation simulation unit; the risk propagation simulation unit simulates the propagation path and cumulative effect of implosion risk in the capacitor bank, and outputs the implosion risk probability value and risk propagation direction indication for each capacitor node.

8. A capacitor implosion anomaly early warning system based on a neural network according to claim 7, characterized in that: The multi-physics coupled hard-constraint graph neural network also includes hard-constraint embedding units and soft-constraint loss units. The hard-constraint embedding units embed the numerical solutions of the thermal balance equation, the equivalent series resistance degradation dynamics equation, and the capacitance decay equation as fixed nonlinear transformations in the message passing layer of the graph neural network. The soft-constraint loss units add a multi-physics coupled soft-constraint loss term to the loss function of the graph neural network.

9. A capacitor implosion anomaly early warning system based on a neural network according to claim 1, characterized in that: The early warning decision-making and feedback optimization module includes an early warning decision-making unit and a maintenance suggestion generation unit; The early warning decision-making unit generates an early warning signal based on the implosion risk level and the early warning confidence level; The maintenance suggestion generation unit generates maintenance suggestions.

10. A capacitor implosion anomaly early warning system based on a neural network according to claim 9, characterized in that: The early warning decision-making and feedback optimization module also includes a feedback optimization unit; The feedback optimization unit collects actual implosion event information, calculates the early warning accuracy and false negative rate, and triggers the adjustment of the decision threshold and fine-tuning of network parameters.

Images