Open-circuit fault pre-judgment method and system of energy storage converter

Abstract

The embodiment of the invention provides an open-circuit fault pre-judgment method and system for an energy storage converter, which abandons direct analysis on original monitoring data and establishes a dynamic health reference changing in real time along with working conditions for the energy storage converter by using an electric heating physical model. And comparing an actual measurement value with the reference to generate a group of normalized state feature vectors of which the influence of working conditions is eliminated, so that effective decoupling of fault information and working condition noise is realized. Then, in order to accurately quantify the health level reflected by the multi-dimensional features, an unsupervised learning model based on an auto-encoder is introduced. According to the model, by learning massive health samples, normalized feature vectors can be intelligently fused into a single health factor which is highly sensitive to weak anomalies. And finally, the trend of the health factor is analyzed, so that the pre-judgment of the open-circuit fault can be realized.

Description

Technical Field

[0001] This invention relates to the field of energy storage converter fault detection technology, specifically to a method and system for predicting open-circuit faults in energy storage converters. Background Technology

[0002] As a key technology supporting new power systems and promoting the consumption of renewable energy, the safe and stable operation of energy storage systems is crucial. The energy storage converter, as the core interface connecting the energy storage system to the grid, directly determines the availability and economic benefits of the entire system. In the energy storage converter, power semiconductor devices such as insulated-gate bipolar transistors (IGBTs) are the execution units for energy conversion and also among the components with the highest failure rate. Open-circuit faults caused by bonding wire fatigue and solder layer degradation are the main cause of unplanned converter outages. Traditional operation and maintenance models are mostly reactive repairs after a fault occurs, which not only incurs high downtime costs but may also lead to secondary damage. Therefore, developing open-circuit fault prediction technology for energy storage converters, realizing the transformation from fault diagnosis to health management, and providing early warnings and scheduling maintenance plans, is of great significance for ensuring the safety of energy storage power stations and improving operation and maintenance efficiency.

[0003] However, achieving accurate open-circuit fault prediction faces significant technical challenges. Existing fault diagnosis methods, such as those based on output voltage harmonic analysis or Park vector mode tracing, are primarily used to detect faults that have already occurred and have relatively obvious characteristics. They are not sensitive to the weak characteristic signals in the early stages of device degradation and cannot meet the needs of prediction. Techniques attempting fault prediction generally face a core challenge: how to effectively decouple the degradation characteristics that truly reflect the device's health status from the parameter fluctuations caused by complex and ever-changing operating conditions. In practical energy storage applications, converters need to frequently respond to grid dispatch commands, and their operating parameters, such as load current and DC-side voltage, change drastically over a wide range. These changes in operating conditions cause critical monitoring parameters of IGBTs, such as on-state voltage drop and junction temperature, to fluctuate much more than the normal fluctuations caused by early faults. For example, an increase in load current directly leads to a significant rise in on-state voltage drop and junction temperature, with changes sufficient to completely overwhelm the millivolt-level voltage rise caused by bonding wire microcracks or the small temperature rise caused by solder layer voids. If fixed thresholds are set or trend analysis is performed directly on these raw monitoring data, it is very easy to generate a large number of false alarms and missed alarms, making the prediction system lack practical engineering value.

[0004] Therefore, an optimized scheme for predicting open-circuit faults in energy storage converters is desired. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art, and provides a method and system for predicting open-circuit faults in energy storage converters.

[0006] In a first aspect, embodiments of the present invention provide a method for predicting open-circuit faults in an energy storage converter, comprising: Obtain collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage; The collector current, collector-emitter voltage, junction temperature, case temperature and gate drive voltage are normalized based on the electrothermal model to obtain the state feature vector of the energy storage converter. Health factors are obtained by performing explicit health factor modeling based on autoencoder on the state feature vector of the energy storage converter. Health factor trend analysis was performed on the time series of health factors to determine the state of degradation.

[0007] Secondly, embodiments of the present invention provide an open-circuit fault prediction system for an energy storage converter, comprising: The data acquisition module is used to acquire collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage. The feature normalization module is used to perform feature normalization on collector current, collector-emitter voltage, junction temperature, case temperature and gate drive voltage based on the electrothermal model to obtain the state feature vector of the energy storage converter. The explicit health factor modeling module is used to perform explicit health factor modeling based on an autoencoder on the state feature vector of the energy storage converter to obtain the health factor. The Health Factor Trend Analysis module is used to perform health factor trend analysis on time series data to determine the state of degradation.

[0008] Compared with existing technologies, this invention provides a method and system for predicting open-circuit faults in energy storage converters. It abandons direct analysis of raw monitoring data and instead utilizes an electrothermal physical model to establish a dynamic health benchmark for the energy storage converter that changes in real time with operating conditions. By comparing actual measured values ​​with this benchmark, a set of normalized state feature vectors, free from the influence of operating conditions, is generated, thereby effectively decoupling fault information from operating noise. Subsequently, to accurately quantify the health level reflected by these multi-dimensional features, an unsupervised learning model based on an autoencoder is introduced. This model, by learning from massive amounts of health samples, can intelligently fuse the normalized feature vectors into a single health factor that is highly sensitive to subtle anomalies. Finally, by analyzing the trend of this health factor, accurate and robust prediction of open-circuit faults can be achieved, effectively overcoming the false alarms and missed alarms caused by drastic changes in operating conditions in traditional methods. Attached Figure Description

[0009] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0010] Figure 1 A flowchart of an open-circuit fault prediction method for an energy storage converter according to an embodiment of the present invention; Figure 2 This is a data flow diagram illustrating the open-circuit fault prediction method for an energy storage converter according to an embodiment of the present invention. Figure 3 The flowchart illustrates the open-circuit fault prediction method for energy storage converters according to an embodiment of the present invention, which performs feature normalization based on an electrothermal model on collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage to obtain the state feature vector of the energy storage converter. Figure 4 A flowchart is provided for the method of predicting open-circuit faults in energy storage converters according to an embodiment of the present invention, which involves explicit modeling of health factors based on an autoencoder on the state feature vector of the energy storage converter to obtain the health factors. Figure 5 This is a block diagram of an open-circuit fault prediction system for an energy storage converter according to an embodiment of the present invention. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0012] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0013] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0014] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0015] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0016] To address the technical problem that early degradation characteristics in the prediction of open-circuit faults in energy storage converters are easily overwhelmed by parameter fluctuations caused by drastic changes in operating conditions, this invention proposes a method for predicting open-circuit faults in energy storage converters. This method first acquires multi-dimensional operating data such as collector current, collector-emitter voltage, and junction temperature in real time. However, instead of directly using this raw data for judgment, it innovatively introduces a dynamic benchmark based on an electrothermal physics model. Specifically, the system looks up a theoretical health parameter value from a pre-set health model (such as a Vce lookup table) based on the current current and temperature, and compares the measured value with this theoretical value to calculate the normalized voltage characteristic and normalized thermal resistance characteristic. This normalization process effectively isolates the influence of operating condition changes, allowing the generated energy storage converter state feature vector to more purely reflect the device's own health status. Subsequently, this feature vector is input into an autoencoder neural network pre-trained with a large amount of health data. The core task of this network is to learn and reproduce the characteristic patterns of a healthy state. When the input vector deviates from this health pattern, a significant error occurs between the reconstructed output and the original input. This invention quantifies this reconstruction error into a single health factor, which acts as a highly sensitive anomaly indicator. A sustained increase in the value of this factor clearly reveals the degradation trajectory of the device. Finally, by analyzing the trend of the time series of this health factor, the degradation state can be accurately identified before a failure occurs, achieving reliable fault prediction.

[0017] Figure 1 This is a flowchart of an open-circuit fault prediction method for an energy storage converter according to an embodiment of the present invention. Figure 2 This is a data flow diagram illustrating the open-circuit fault prediction method for an energy storage converter according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the open-circuit fault prediction method for an energy storage converter according to an embodiment of the present invention includes the following steps: S100, acquiring collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage; S200, performing feature normalization based on an electrothermal model on the collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage to obtain an energy storage converter state feature vector; S300, performing explicit health factor modeling based on an autoencoder on the energy storage converter state feature vector to obtain a health factor; S400, performing health factor trend analysis on the time series of the health factor to obtain a degradation state.

[0018] Specifically, in step S100, the collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage are acquired. It should be understood that since an open-circuit fault in a power device within an energy storage converter is a gradual physical degradation process, this process inevitably causes slight but continuous changes in the device's electrical characteristics (such as on-state voltage drop) and thermal characteristics (such as thermal resistance). Therefore, in the technical solution of this invention, by acquiring the collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage, the electrothermal state of the power device under current operating conditions can be comprehensively and in real-time characterized. This provides the necessary, high-fidelity raw data input for subsequent feature normalization processing based on the electrothermal model, ensuring the completeness and accuracy of the data foundation for fault prediction analysis.

[0019] More specifically, in a concrete example of the invention, the data acquisition process is implemented through an integrated data acquisition system deployed within the energy storage converter. First, for high-frequency varying electrical quantities, the system measures the collector current using a high-precision Hall current sensor installed in the power circuit, and measures the collector-emitter voltage using a differential voltage probe connected in parallel across the collector and emitter of the power device. Simultaneously, it directly acquires the gate drive voltage signal from the gate driver output. These analog signals are sampled via a high-speed analog-to-digital converter. Second, for slowly varying thermal quantities, the system directly measures the case temperature using an NTC thermistor attached to the power device module substrate. The junction temperature is indirectly estimated using an online thermal model based on real-time calculated power losses and the measured case temperature. Finally, the digital signal processor within the controller is responsible for timestamping and synchronizing the data from all acquisition channels, fusing electrical and thermal data from different sampling rates into a structured data frame with a unified time reference for subsequent steps.

[0020] Specifically, in step S200, the collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage are normalized based on an electrothermal model to obtain the energy storage converter state feature vector. It should be understood that the absolute values ​​of directly measured parameters such as on-state voltage drop and thermal resistance fluctuate significantly with drastic changes in operating conditions such as load current and ambient temperature. This fluctuation can completely mask weak characteristic signals caused by early device degradation, leading to the failure of fault prediction. Therefore, in the technical solution of this invention, the collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage are further normalized based on an electrothermal model to obtain the energy storage converter state feature vector. This eliminates the influence of operating conditions on monitoring data and extracts standardized features that are only related to the intrinsic health state of the device. In this way, a state feature vector that is sensitive to device degradation but insensitive to changes in operating conditions can be generated, providing stable and reliable data input for subsequent construction of an accurate health factor model, significantly improving the accuracy and robustness of prediction.

[0021] Figure 3 This is a flowchart illustrating the open-circuit fault prediction method for energy storage converters according to an embodiment of the present invention, which performs feature normalization based on an electrothermal model on collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage to obtain the state feature vector of the energy storage converter. Figure 3 As shown, step S200 includes: S210, when the gate drive voltage is high, extracting the steady-state value of the collector-emitter voltage as the measured on-state voltage drop; S220, looking up the theoretical healthy on-state voltage drop in the healthy Vce lookup table using bilinear interpolation based on the collector current and case temperature; S230, calculating the normalized voltage characteristic based on the measured on-state voltage drop and the theoretical healthy on-state voltage drop; S240, calculating the instantaneous power loss based on the measured on-state voltage drop and the collector current; S250, calculating the measured resistance based on the instantaneous power loss, junction temperature, and case temperature; S260, dividing the measured resistance by the reference healthy resistance to obtain the normalized thermal resistance characteristic.

[0022] Accordingly, in steps S210 and S220, when the gate drive voltage is high, the steady-state value of the collector-emitter voltage is extracted as the measured on-state voltage drop. Furthermore, in the health Vce lookup table, the theoretical healthy on-state voltage drop is found based on the collector current and case temperature using bilinear interpolation. It should be understood that since the on-state voltage drop of a power device is not a fixed value, but a dynamic parameter closely related to its real-time collector current and its own temperature (such as case temperature), to accurately assess its health status, the measured voltage must be compared with a health benchmark under the same operating conditions, rather than a static threshold. Therefore, in the technical solution of this invention, when the gate drive voltage is high, the steady-state value of the collector-emitter voltage is extracted as the measured on-state voltage drop. Furthermore, in the health Vce lookup table, the theoretical healthy on-state voltage drop is found based on the collector current and case temperature using bilinear interpolation. This allows for the simultaneous acquisition of both the measured value reflecting the current true performance of the device and the theoretical benchmark value characterizing its ideal health status within each effective conduction cycle. This provides a pair of accurate data that are completely corresponding in time and operating conditions and can be directly compared for subsequent normalization calculations, thereby ensuring that the normalization results can truly reflect the voltage drop deviation caused by device degradation, rather than the normal fluctuations caused by changes in operating conditions.

[0023] More specifically, in a concrete example of the present invention, this process is executed by a digital signal processor embedded in the converter controller. First, the processor continuously monitors the gate drive signal. When it detects a transition from low to high, the processor starts a microsecond-level internal timer to avoid the turn-on transient process of the voltage waveform. After the timer expires, the processor immediately reads the sampled value of the collector-emitter voltage and uses it as the measured on-state voltage drop at that moment. Simultaneously, the processor obtains the current collector current and case temperature values ​​from a synchronized data acquisition system. Subsequently, the processor accesses a two-dimensional health Vce lookup table stored in its internal firmware. This table stores the corresponding healthy on-state voltage drop values ​​with discrete current and case temperature values ​​as coordinate axes. Based on the real-time collector current and case temperature values, the processor locates the four nearest data points surrounding the operating point in the lookup table and performs bilinear interpolation to accurately calculate the theoretical healthy on-state voltage drop under the current operating condition. Finally, the processor outputs the measured on-state voltage drop and the calculated theoretical healthy on-state voltage drop as a data pair.

[0024] Accordingly, in step S230, a normalized voltage characteristic is calculated based on the measured on-state voltage drop and the theoretical healthy on-state voltage drop. It should be understood that since the measured on-state voltage drop and the theoretical healthy on-state voltage drop are still absolute quantities related to the operating conditions, it is difficult to directly construct a unified, cross-operating-condition health assessment index. Therefore, in the technical solution of this invention, a normalized voltage characteristic is further calculated based on the measured on-state voltage drop and the theoretical healthy on-state voltage drop, thereby converting the two absolute voltage values ​​into a dimensionless ratio that quantifies the degree of deviation of the measured performance from the ideal healthy state. In this way, a standardized characteristic value can be generated that constantly approaches 1 under the ideal healthy state and monotonically increases with the degree of device degradation. This characteristic value eliminates the dependence on operating conditions and can be directly used for subsequent health state modeling and trend analysis.

[0025] Specifically, in one specific example of the present invention, the calculation of the normalized voltage characteristic is performed internally by a simple arithmetic operation within the digital signal processor. After obtaining the measured on-state voltage drop and the theoretical healthy on-state voltage drop from the previous stage, the processor immediately invokes its arithmetic logic unit to perform a division operation. Specifically, the processor uses the measured on-state voltage drop as the dividend and the theoretical healthy on-state voltage drop as the divisor to calculate the quotient. The result of this operation, a dimensionless floating-point value, is defined as the normalized voltage characteristic and stored in a designated memory address as a dimension component of the energy storage converter's state characteristic vector for use in subsequent processing steps.

[0026] Accordingly, in step S240, the instantaneous power loss is calculated based on the measured on-state voltage drop and collector current. It should be understood that since the thermal resistance of a power device is a key physical quantity characterizing the health of its heat dissipation path, and this thermal resistance is defined as the ratio of the device's temperature rise to the power loss that causes that temperature rise, the instantaneous power loss must be accurately quantified before calculating the measured thermal resistance. Therefore, in the technical solution of this invention, the instantaneous power loss is further calculated based on the measured on-state voltage drop and collector current to obtain the rate of heat generated inside the device due to conduction at a specific conduction moment. This provides an accurate power loss benchmark that is strictly synchronized in time with electrical and thermal measurements for subsequent calculations of the measured thermal resistance, ensuring the correct physical meaning and numerical accuracy of the thermal resistance calculation.

[0027] Specifically, in this embodiment of the invention, the calculation of the instantaneous power loss is performed by the digital signal processor in its internal arithmetic logic unit. The processor retrieves the measured on-state voltage drop and the corresponding collector current value acquired within the same data frame from its internal registers or cache. Then, the instantaneous power loss is calculated using the following formula:

[0028] in, To measure the on-state voltage drop, For collector current, The watt value representing the instantaneous conduction loss power is stored in a new memory unit and marked as instantaneous loss power, so that it can be directly called in the step of calculating the measured resistance.

[0029] Accordingly, in step S250, the measured resistance is calculated based on the instantaneous power loss, junction temperature, and case temperature. It should be understood that physical degradation of power devices (e.g., solder layer fatigue causing voids) directly leads to a deterioration of their heat dissipation path, manifested as an increase in thermal resistance, and this change is independent of the electrical load borne by the device. Therefore, in the technical solution of this invention, the measured resistance is further calculated based on the instantaneous power loss, junction temperature, and case temperature to quantify the heat transfer capability of the device from the chip junction to the case in real time using fundamental thermal laws. This provides a key performance indicator that directly reflects the physical integrity of the heat dissipation path, and the trend of this indicator can sensitively indicate potential failure modes related to thermal management.

[0030] Specifically, in this embodiment of the invention, the calculation of the measured resistance is performed in the digital signal processor immediately after the calculation of the instantaneous power loss. The processor first reads the calculated instantaneous power loss from memory, along with the junction temperature and case temperature acquired in the same data frame. Then, the measured resistance is calculated using the following formula:

[0031] in, For the junction temperature, Shell temperature, For instantaneous power loss, This is the measured resistance value at the current moment, expressed in kWh. This calculation result is stored in a new memory variable for use in subsequent normalized thermal resistance characteristic calculations.

[0032] Accordingly, in step S260, the measured resistance is divided by the reference healthy resistance to obtain the normalized thermal resistance characteristic. That is, the measured thermal resistance is divided by a preset reference healthy resistance value representing the new state of the device to obtain the second-dimensional characteristic, namely the normalized thermal resistance characteristic. Finally, the processing unit combines the calculated normalized voltage characteristic and normalized thermal resistance characteristic into a two-dimensional energy storage converter state characteristic vector and outputs it to the next processing module.

[0033] Specifically, in step S300, an explicit health factor model based on an autoencoder is performed on the state feature vector of the energy storage converter to obtain a health factor. It should be understood that since the state feature vector of the energy storage converter is a vector containing multiple dimensions of information (such as normalized voltage and normalized thermal resistance), directly performing trend analysis and threshold setting on the time series of the multidimensional vector is complex and unintuitive, and it is difficult to capture the nonlinear coupling relationship between the features of each dimension. Therefore, in the technical solution of this invention, the state feature vector of the energy storage converter is further subjected to an explicit health factor model based on an autoencoder to obtain a health factor. This utilizes the nonlinear mapping capability of neural networks to intelligently reduce the dimensionality of the multidimensional and complex feature vector and fuse it into a single, intuitive scalar index that can quantify the health level. In this way, a health factor highly sensitive to early, weak, and hidden abnormal patterns in the multidimensional feature space can be generated, greatly simplifying the subsequent trend analysis and fault judgment process and improving the sensitivity of prediction.

[0034] Figure 4 This document describes a flowchart illustrating the open-circuit fault prediction method for energy storage converters according to an embodiment of the present invention, which involves explicit modeling of health factors based on an autoencoder on the state feature vector of the energy storage converter to obtain health factors. For example... Figure 4 As shown, step S300 includes: S310, inputting the energy storage converter state feature vector into the autoencoder to obtain the reconstructed feature vector; S320, calculating the reconstruction error between the reconstructed feature vector and the energy storage converter state feature vector to obtain the health factor.

[0035] Accordingly, in step S310, the energy storage converter state feature vector is input into the autoencoder to obtain a reconstructed feature vector. It should be understood that since the core mechanism of the autoencoder model is to reconstruct data by learning the inherent patterns of health data, its ability to reconstruct abnormal patterns not seen in the training set is inevitably poor. Therefore, in the technical solution of this invention, the energy storage converter state feature vector is further input into the autoencoder to obtain a reconstructed feature vector. This utilizes the model as a verification standard, forcibly mapping any input feature vector to its understood health feature space. In this way, a reconstructed vector with the same dimension as the original input vector but containing only health pattern information can be generated. This vector provides a direct comparison benchmark for subsequently quantifying the deviation between the current state and the health state.

[0036] Specifically, in one particular example of the invention, the reconstruction process is executed on a digital signal processor within the converter controller. The processor first reads the energy storage converter state feature vector generated by the feature normalization module from memory. Then, the processor uses this vector as input to initiate the forward propagation computation of a pre-loaded autoencoder neural network model. This computation includes multiplying the input vector with the weight matrix of the encoder part, adding a bias term, and then processing it through an activation function to obtain a low-dimensional latent feature representation. Next, this latent feature representation is input to the decoder part, and through multiplication with the decoder weight matrix, addition of the bias term, and activation function processing, it is restored layer by layer to its original dimension. The final output of this forward propagation computation is a reconstructed feature vector with the same dimension as the input vector. The processor stores this vector in a designated register for use in the next step of calculating the reconstruction error.

[0037] Accordingly, in step S320, the reconstruction error between the reconstructed feature vector and the energy storage converter state feature vector is calculated to obtain the health factor. It should be understood that since the reconstructed feature vector is the best reproduction of the original input features by the autoencoder based on the health data distribution it has learned, while the original state feature vector reflects the actual operating state of the device, the degree of difference between these two vectors directly quantifies the distance of the current state from the established health pattern. Therefore, in the technical solution of this invention, the reconstruction error between the reconstructed feature vector and the energy storage converter state feature vector is further calculated to obtain the health factor, thereby condensing multi-dimensional feature deviation information into a single, dimensionless scalar value. This generates a health factor whose value is close to zero when the device is healthy and exhibits a monotonically increasing trend as device performance degrades. This factor, as a highly integrated health status indicator, provides a direct and quantitatively clear basis for subsequent trend analysis and threshold judgment for fault warning.

[0038] However, the original mechanism uses the mean square error of the autoencoder repetition error as the health factor, which inherently degenerates complex, physically coupled, multidimensional fault characteristics into a single, non-directional Euclidean distance metric. In the actual operation of energy storage converters, features such as normalized on-state voltage drop and normalized thermal resistance are not independent; their fluctuations and correlations under healthy conditions follow specific statistical laws. The mean square error method ignores this inherent covariance structure and cannot distinguish between a large error in a statistically common direction and a small error in a rare direction, the latter often being a symptom of early faults. This isotropic measurement makes the health factor sensitive to normal operating noise but insensitive to critical abnormal patterns that indicate physical mechanism degradation, resulting in insufficient reliability and accuracy in prediction. To address these shortcomings, this invention proposes a probabilistic health factor construction method based on Mahalanobis distance and chi-square distribution. This method no longer measures the magnitude of the error in isolation but assesses the probability of anomalies in the current error state within the health statistical model, thereby achieving accurate capture of early fault characteristics.

[0039] Specifically, in a preferred embodiment of the present invention, calculating the reconstruction error between the reconstructed feature vector and the energy storage converter state feature vector to obtain the health factor includes: calculating the positional difference between the reconstructed feature vector and the energy storage converter state feature vector to obtain an error vector; performing an inversion operation on the offline stored covariance matrix to obtain the inverse of the covariance matrix; and determining the health factor based on the error vector and the inverse of the covariance matrix.

[0040] More specifically, the error vector is obtained by calculating the positional difference between the reconstructed feature vector and the energy storage converter state feature vector, and the inverse of the offline-stored covariance matrix is ​​obtained. It should be understood that simple reconstruction error measures (such as mean square error) compress the deviations of multidimensional features into a single scalar without directionality, ignoring the inherent statistical correlation between the errors of each feature under healthy conditions. For example, the errors of normalized on-state voltage drop and normalized thermal resistance follow a specific covariance structure during healthy operation. A large error in a statistically common direction may simply be operating noise, while a small error in a rare direction is often a sign of early failure. Therefore, in the technical solution of this invention, the error vector is further obtained by calculating the positional difference between the reconstructed feature vector and the energy storage converter state feature vector, and the inverse of the offline-stored covariance matrix is ​​obtained. This simultaneously obtains the error vector representing the deviation from the current state, and a mathematical operator that can directionally weight the error vector and eliminate the correlation between features. This lays the foundation for subsequent anomaly assessment using Mahalanobis distance, ensuring that the calculation of health factors can distinguish between normal fluctuations and abnormal patterns indicating physical mechanism degradation, thereby achieving accurate capture of early fault characteristics.

[0041] In a specific example of the present invention, the calculation process is executed by a digital signal processor within the controller. First, the processor retrieves the original energy storage converter state feature vector and its corresponding reconstructed feature vector from memory, and performs element-wise subtraction of the vectors to generate an error vector representing the current state deviation. At the same time, the processor reads a pre-stored health state covariance matrix representing the health baseline distribution from its non-volatile memory. This matrix, obtained during the offline phase by collecting a large amount of health operation data and calculating the covariance of its error vector, comprehensively describes the distribution pattern of health errors and the correlation between different characteristic errors. Subsequently, the processor calls its internal linear algebra library to perform an inversion operation on the covariance matrix. Finally, the processor passes both the calculated real-time error vector and the inverse of the covariance matrix to the next computational unit for the final synthesis of health factors, laying the foundation for assessing the degree of anomalies from a statistical perspective.

[0042] More specifically, the health factor is determined based on the error vector and the inverse of the covariance matrix. It should be understood that, since the error vector alone is still a multidimensional quantity, its absolute magnitude (such as Euclidean distance or mean squared error) is an isotropic measure. It cannot distinguish between a large error in a statistically common direction and a small error in a rare direction, the latter often being a key early sign of physical mechanism degradation. This measure ignores the inherent covariance structure between the errors of various features in a healthy state. Therefore, in the technical solution of this invention, the health factor is further determined based on the error vector and the inverse of the covariance matrix, thereby introducing the advanced measure of Mahalanobis distance. By transforming the error space through the inverse of the covariance matrix, the decoupling of feature correlations and normalization of different dimensional scales are automatically achieved when measuring the degree of anomaly.

[0043] In a specific example of the present invention, the final determination of the health factor is completed by a digital signal processor through a two-step composite operation. First, the processor uses the real-time error vector obtained in the previous step. and its transpose, and the inverse of the health baseline covariance matrix. Perform a quadratic operation, which calculates the product of the transpose of the error vector, the inverse of the covariance matrix, and the error vector, to obtain the squared Mahalanobis distance of the current state. Expressed using the following formula:

[0044] in, For the error vector, It is the inverse of the covariance matrix. The squared Mahalanobis distance of the current state is used to generate a scalar anomaly score that takes into account all health statistics. Compared to the MSE, this score is extremely sensitive to weak but directional error signals that indicate early faults. Even if the absolute value of the error vector is small, if its direction is contrary to the main change direction of the health data, its squared Mahalanobis distance will increase significantly, thereby effectively amplifying fault symptoms and suppressing normal operating noise.

[0045] Then, although the Mahalanobis distance squared It is a superior outlier score, but its numerical value lacks intuitive probabilistic meaning, making it inconvenient to set a unified and interpretable warning threshold. According to multivariate statistical theory, the squared Mahalanobis distance of a random variable following a k-dimensional Gaussian distribution follows a chi-square distribution with k degrees of freedom. Distribution. Therefore, the processor further uses this value as input, and... By substituting the cumulative distribution function of the chi-square distribution with k degrees of freedom as the feature dimension, the final probabilistic health factor is obtained by converting it into a standardized index with a clear probabilistic meaning. Expressed using the following formula:

[0046] in, The square of the Mahalanobis distance to the current state. Let be the cumulative distribution function of the chi-square distribution. Let be the probability density function of a chi-square distribution with k degrees of freedom. For health factors. This allows for the abstraction of statistical distance. Mapped to a probability value between 0 and 1 This gives it clear interpretability, namely, "the current state is more abnormal than what percentage of healthy states." The resulting probabilistic health factor... It is a highly standardized and interpretable metric. When the device is fully healthy, The value will fluctuate around a low level; and when early failures occur, It will stably and significantly approach 1. This makes the setting of the warning threshold (e.g., 0.99 or 0.999) a reliable basis, greatly improving the robustness and versatility of the warning system.

[0047] Specifically, in step S400, a health factor trend analysis is performed on the time series of health factors to obtain the degradation state. It should be understood that since the instantaneously calculated health factor sequence inevitably contains random fluctuations and glitches caused by measurement noise, electromagnetic interference, or non-continuous operating condition disturbances, directly comparing these raw values ​​with thresholds can easily lead to false alarms, thereby reducing the reliability of the early warning system. Therefore, in the technical solution of this invention, a health factor trend analysis is further performed on the time series of health factors to obtain the degradation state, thereby filtering out high-frequency noise interference, extracting a smooth trend that truly reflects the gradual degradation of the physical performance of power devices, and making stable and reliable fault judgments based on this trend. This ensures that the final output degradation state is based on a judgment of the long-term, stable changes in the equipment's health status, rather than an overreaction to instantaneous, occasional fluctuations, thus significantly improving the accuracy and robustness of fault prediction.

[0048] More specifically, in this embodiment of the invention, performing health factor trend analysis on the time series of health factors to obtain the degradation state includes: smoothing the time series of health factors to obtain a smoothed health factor; and determining the degradation state based on a comparison between the smoothed health factor and a fault threshold. Specifically, this trend analysis and state determination process is periodically executed by the processing unit within the controller. First, the processing unit maintains a first-in-first-out queue to store the health factor values ​​calculated over the most recent N sampling periods. In each new sampling period, the latest health factor value is pushed into the queue, while the oldest value is removed. Subsequently, the processing unit performs a moving average filtering algorithm on the current N health factor values ​​in the queue, i.e., calculates the arithmetic mean of these values, and uses this mean as the smoothed health factor at the current moment. Next, the processing unit compares the smoothed health factor with a fault threshold (e.g., 0.999) preset in non-volatile memory. If the smoothed health factor continuously exceeds the fault threshold, the processing unit sets the system's degradation state flag to fault and generates a warning signal; otherwise, the state flag remains healthy. Ultimately, the identified degradation status is reported to the top-level management system of the energy storage converter to trigger the corresponding maintenance or protection procedures.

[0049] In summary, the open-circuit fault prediction method for energy storage converters according to embodiments of the present invention is explained. It abandons direct analysis of raw monitoring data and instead utilizes an electrothermal physical model to establish a dynamic health benchmark for the energy storage converter that changes in real time with operating conditions. By comparing actual measured values ​​with this benchmark, a set of normalized state feature vectors, free from the influence of operating conditions, is generated, thereby effectively decoupling fault information from operating condition noise. Subsequently, to accurately quantify the health level reflected by this multi-dimensional feature, an unsupervised learning model based on an autoencoder is introduced. This model, by learning from massive amounts of health samples, can intelligently fuse the normalized feature vectors into a single health factor that is highly sensitive to subtle anomalies. Finally, by analyzing the trend of this health factor, accurate and robust prediction of open-circuit faults can be achieved, effectively overcoming the false alarms and missed alarms caused by drastic changes in operating conditions in traditional methods.

[0050] Furthermore, an open-circuit fault prediction system for energy storage converters is also provided.

[0051] Figure 5 This is a block diagram of an open-circuit fault prediction system for an energy storage converter according to an embodiment of the present invention. Figure 5As shown, the open-circuit fault prediction system 100 for an energy storage converter according to an embodiment of the present invention includes: a data acquisition module 110 for acquiring collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage; a feature normalization module 120 for performing feature normalization on the collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage based on an electrothermal model to obtain a state feature vector of the energy storage converter; a health factor explicit modeling module 130 for performing explicit health factor modeling on the state feature vector of the energy storage converter based on an autoencoder to obtain health factors; and a health factor trend analysis module 140 for performing health factor trend analysis on the time series of health factors to obtain a degradation state.

[0052] Furthermore, the feature normalization module 120 is specifically used for: extracting the steady-state value of the collector-emitter voltage as the measured on-state voltage drop when the gate drive voltage is high; searching for the theoretical healthy on-state voltage drop based on the collector current and case temperature in the healthy Vce lookup table using bilinear interpolation; and calculating the normalized voltage feature based on the measured on-state voltage drop and the theoretical healthy on-state voltage drop.

[0053] As described above, the open-circuit fault prediction system 100 for the energy storage converter according to embodiments of the present invention can be implemented in various wireless terminals, such as servers with open-circuit fault prediction algorithms for energy storage converters. In one possible implementation, the open-circuit fault prediction system 100 for the energy storage converter according to embodiments of the present invention can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the open-circuit fault prediction system 100 for the energy storage converter can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the open-circuit fault prediction system 100 for the energy storage converter can also be one of many hardware modules of the wireless terminal.

[0054] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting open-circuit faults in an energy storage converter, characterized in that, include: Obtain collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage; The collector current, collector-emitter voltage, junction temperature, case temperature and gate drive voltage are normalized based on the electrothermal model to obtain the state feature vector of the energy storage converter. Health factors are obtained by performing explicit health factor modeling based on autoencoder on the state feature vector of the energy storage converter. Health factor trend analysis was performed on the time series of health factors to determine the state of degradation.

2. The method for predicting open-circuit faults in an energy storage converter according to claim 1, characterized in that, The collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage are normalized based on an electrothermal model to obtain the state feature vector of the energy storage converter, including: When the gate drive voltage is high, the steady-state value of the collector-emitter voltage is extracted as the measured on-state voltage drop. In the healthy Vce lookup table, the theoretical healthy on-state voltage drop is found based on the collector current and case temperature using bilinear interpolation. The normalized voltage characteristics are calculated based on the measured on-state voltage drop and the theoretical healthy on-state voltage drop.

3. The method for predicting open-circuit faults in an energy storage converter according to claim 2, characterized in that, The collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage are normalized based on an electrothermal model to obtain the state feature vector of the energy storage converter, including: Calculate the instantaneous power loss based on the measured on-state voltage drop and collector current; Calculate the measured resistance based on instantaneous power loss, junction temperature, and case temperature; Divide the measured resistance by the reference healthy resistance to obtain the normalized thermal resistance characteristic.

4. The open-circuit fault prediction method for energy storage converters according to claim 3, characterized in that, in, Based on the measured on-state voltage drop and collector current, the instantaneous power loss is calculated, including: calculating the instantaneous power loss using the following formula, where the formula is: in, To measure the on-state voltage drop, For collector current, It is instantaneous power loss; The calculation of the measured resistance based on instantaneous power loss, junction temperature, and case temperature includes: calculating the measured resistance using the following formula, where the formula is: in, For the junction temperature, Shell temperature, For instantaneous power loss, It is the actual measured resistance.

5. The method for predicting open-circuit faults in an energy storage converter according to any one of claims 1 to 4, characterized in that, Health factors are obtained by performing explicit health factor modeling based on an autoencoder on the state feature vector of the energy storage converter, including: The state feature vector of the energy storage converter is input into the autoencoder to obtain the reconstructed feature vector; The health factor is obtained by calculating the reconstruction error between the reconstructed feature vector and the energy storage converter state feature vector.

6. The method for predicting open-circuit faults in an energy storage converter according to claim 5, characterized in that, Calculating the reconstruction error between the reconstructed feature vector and the energy storage converter state feature vector to obtain the health factor includes: The position-based difference between the reconstructed feature vector and the energy storage converter state feature vector is calculated to obtain the error vector; The inverse of the covariance matrix is ​​obtained by performing an inversion operation on the offline stored covariance matrix; The health factor is determined based on the error vector and the inverse of the covariance matrix.

7. The method for predicting open-circuit faults in an energy storage converter according to claim 6, characterized in that, The health factor is determined based on the error vector and the inverse of the covariance matrix, including: calculating the health factor using the following formula: in, For the error vector, It is the inverse of the covariance matrix. The square of the Mahalanobis distance to the current state. Let be the probability density function of a chi-square distribution with k degrees of freedom. Let be the cumulative distribution function of the chi-square distribution. For health factors.

8. The method for predicting open-circuit faults in an energy storage converter according to any one of claims 1 to 4, characterized in that, Health factor trend analysis was performed on time series data of health factors to determine the state of decline, including: Smoothing filters are applied to the time series of health factors to obtain smoothed health factors. The degradation state is determined by comparing the smooth health factor with the fault threshold.

9. An open-circuit fault prediction system for an energy storage converter, characterized in that, include: The data acquisition module is used to acquire collector current, collector-emitter voltage, junction temperature, case temperature, and gate drive voltage. The feature normalization module is used to perform feature normalization on collector current, collector-emitter voltage, junction temperature, case temperature and gate drive voltage based on the electrothermal model to obtain the state feature vector of the energy storage converter. The explicit health factor modeling module is used to perform explicit health factor modeling based on an autoencoder on the state feature vector of the energy storage converter to obtain the health factor. The Health Factor Trend Analysis module is used to perform health factor trend analysis on time series data to determine the state of degradation.

10. The open-circuit fault prediction system for energy storage converters according to claim 9, characterized in that, The feature normalization module is also specifically used for: When the gate drive voltage is high, the steady-state value of the collector-emitter voltage is extracted as the measured on-state voltage drop. In the healthy Vce lookup table, the theoretical healthy on-state voltage drop is found based on the collector current and case temperature using bilinear interpolation. The normalized voltage characteristics are calculated based on the measured on-state voltage drop and the theoretical healthy on-state voltage drop.

Images