Inverter performance degradation online monitoring method

Abstract

The invention relates to the technical field of power electronics, discloses an on-line monitoring method for performance degradation of an inverter, and is used for solving the problems of high working condition noise, large environment drift and difficulty in early degradation monitoring in the prior art. According to the method, in the PWM dead zone time of the inverter, detection pulses are injected into a power device, electromagnetic and acoustic responses of the power device are synchronously collected, and a two-dimensional acoustic-electric resonance fingerprint is constructed. A variational auto-encoder is used to learn health fingerprint data, and a health baseline model representing a normal state is established. During on-line monitoring, measurement is dynamically calibrated through environment anchor points obtained in real time, drift influences such as temperature are eliminated, the difference between a real fingerprint and a model reconstruction fingerprint is calculated, and a physical degradation index is obtained. Through active silent detection and dynamic environment compensation, the device has strong anti-interference performance and high stability, and can realize high-sensitivity early warning for physical degradation of the device.

Description

Technical Field

[0001] This invention relates to the field of power electronics technology, specifically to a method for online monitoring of inverter performance degradation. Background Technology

[0002] Power semiconductor devices, especially insulated-gate bipolar transistors (IGBTs), are core components in modern power electronic converters for energy conversion and control. Operating under high voltage, high current, high frequency, and varying temperature stresses for extended periods, their packaging structures, such as bonding wires and solder layers, inevitably undergo fatigue aging, leading to performance degradation and ultimately system failure. Therefore, real-time online monitoring of the health status of power devices, enabling early fault warnings and predictive maintenance, is crucial for improving the reliability and safety of power electronic systems.

[0003] Currently, the industry has proposed various online monitoring technologies for power devices. These technologies typically assess the health status indirectly by monitoring changes in certain electrical, thermal, or acoustic parameters during device operation. However, these existing technologies generally face significant challenges in practical applications. A core difficulty lies in the extremely strong electromagnetic interference and severe operating condition fluctuations generated by power converters during normal operation. This strong background noise environment often drowns out weak characteristic signals caused by early physical degradation, resulting in an extremely low signal-to-noise ratio.

[0004] Furthermore, many traditional parameters used as health indicators, such as on-state voltage drop, switching time, or junction temperature, are not only related to the aging degree of the device, but also tightly coupled with its real-time load current, bus voltage, and operating temperature, among other external operating conditions. This complex coupling makes it extremely difficult to accurately separate subtle changes caused purely by aging from mixed signals, easily leading to misjudgments or missed detections.

[0005] Therefore, existing methods are often not sensitive enough to early, subtle physical structural damage (such as tiny voids in the solder layer or initial peeling of bond wires). Typically, changes in these parameters are only reliably detected when performance degradation has accumulated to a significant degree, by which time the optimal time for early maintenance has been missed. In summary, there is an urgent need for a novel online monitoring solution that can operate stably in high-noise environments, effectively decouple operating conditions from aging effects, and possess high sensitivity to the initial stages of physical degradation. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an online monitoring method for inverter performance degradation, solving the technical problem that existing technologies struggle to achieve high-sensitivity and high-stability online monitoring of early physical performance degradation of power devices under the strong interference and variable operating conditions of actual power converter operation.

[0007] To achieve the above objectives, the first aspect of this invention provides an online monitoring method for inverter performance degradation. The core idea of ​​this method is to actively excite and capture multimodal response signals that directly reflect the physical structural state of the inverter's power devices within a specific non-conducting gap during operation, and combine this with dynamic environmental calibration technology to achieve accurate and interference-resistant assessment of early performance degradation.

[0008] Specifically, the present invention provides an online monitoring method for inverter performance degradation, comprising the following steps:

[0009] First, during the dead time of the inverter's pulse width modulation signal, a preset probe pulse is injected into the target power device. This dead time is the instant when both the upper and lower bridge arms of the power module are reliably turned off; probes performed during this period can fundamentally isolate interference from high-power operation of the main circuit. Simultaneously, the electromagnetic and acoustic signals generated by the target power device in response to the probe pulse are collected to construct an acoustic-electric resonant fingerprint.

[0010] In an optional implementation, the step of constructing the acoustic-electric resonance fingerprint specifically includes: treating the synchronously acquired, time-series electromagnetic signals and acoustic signals as two independent dimensions, precisely aligning and combining them on the time axis to form a two-dimensional fingerprint matrix that can comprehensively characterize the transient response characteristics of the device.

[0011] Subsequently, the acoustic-electric resonance fingerprint is input into a pre-constructed health baseline model. This health baseline model is used to characterize the fingerprint pattern of the target power device in a healthy state.

[0012] In a preferred embodiment, the health baseline model is a variational autoencoder model. This model performs self-supervised learning on multiple acoustic-electric resonance fingerprints collected under the initial health state of the inverter. It automatically learns the high-dimensional distribution characteristics of the health fingerprint data without fault labels and compresses and maps them into a low-dimensional latent space, thereby constructing a latent space health manifold that can numerically represent the health state. The training process of this model aims to minimize a specific loss function L. VAE :

[0013] L VAE =L recon +β·D KL (q φ (z|F)||p(z));

[0014] Among them, L recon The reconstruction loss is used to ensure that the model can accurately reconstruct the original health fingerprint from the latent space; D KΠKL divergence is used as a regularization term to ensure a smooth and regular distribution in the latent space; F is the acoustic-electric resonance fingerprint used for training; q φ (z|F) is the posterior distribution of the encoder output, p(z) is the preset prior distribution, and β is the balance factor.

[0015] Finally, based on the output of the health baseline model, a performance degradation index for the target power device is calculated and determined. This index is used to quantify the degree to which the current device state deviates from its healthy state.

[0016] To further improve the accuracy and environmental adaptability of monitoring, in a preferred embodiment, a step of dynamically calibrating the evaluation benchmark is included before inputting the acoustic-electric resonance fingerprint into the model. This step includes: acquiring real-time environmental parameters synchronized with the acquisition of the acoustic-electric resonance fingerprint to form an environmental anchor point; and using the environmental anchor point to dynamically calibrate the evaluation benchmark of the health baseline model to compensate for measurement drift caused by environmental changes (especially temperature). The real-time environmental parameters may include the real-time temperature of the target power device and / or background noise characteristics acquired during normal switching of the inverter.

[0017] Based on this, the calculation process of the performance degradation index is further optimized. Specifically, the process includes: inputting the currently acquired acoustic-electric resonance fingerprint into the encoder of the healthy baseline model to obtain an original latent vector; simultaneously, calculating an environmental offset vector through a pre-trained mapping network based on synchronously acquired environmental anchor points; using the environmental offset vector to compensate and calibrate the original latent vector to obtain a calibrated latent vector; subsequently, inputting the calibrated latent vector into the decoder of the model to reconstruct a calibrated reconstructed fingerprint.

[0018] Finally, the performance degradation index is determined by calculating the error between the currently acquired acoustic-electric resonance fingerprint and the calibrated reconstructed fingerprint, and can be quantified using the following formula:

[0019]

[0020] Among them, F new For the currently acquired acoustic-electric resonance fingerprint, F′ cal_new For the reconstructed fingerprint after calibration, ||·|| F The PDI value represents the Frobenius norm. This PDI value can objectively reflect the degree of degradation at the physical level of the device.

[0021] Furthermore, to enhance the saliency of fingerprint features, in a further embodiment, the spectral characteristics of the injected preset probe pulse can be adaptively optimized. Specifically, its spectral characteristics can be adaptively adjusted and focused on the resonant frequency point most sensitive to the target power device after analyzing the initial response of the target power device, thereby obtaining a response signal with a higher signal-to-noise ratio with lower excitation energy.

[0022] This invention provides a method for online monitoring of inverter performance degradation. It has the following beneficial effects:

[0023] 1. This invention cleverly utilizes the brief "silent" window between power device switching actions by actively probing during the dead time of the inverter's PWM control signal. During this moment, the strong electromagnetic switching noise of the system's main circuit naturally disappears, resulting in a naturally high signal-to-noise ratio for the acquired response signal. This design fundamentally avoids the challenge of signal extraction under complex operating conditions, obtaining pure feature signals without the need for complex filtering algorithms, thus laying a solid foundation for subsequent high-precision degradation assessment.

[0024] 2. This invention deeply integrates information from two different physical modes: electromagnetic response and acoustic response. The electromagnetic response accurately captures changes in electrical path integrity caused by bonding wire degradation, while the acoustic response directly reveals subtle changes in the internal mechanical structure of the device caused by solder layer voids, delamination, etc. The two are coupled and complementary, forming a comprehensive health snapshot that reflects both the electrical and mechanical characteristics of the device. Compared to single-parameter monitoring, this provides stronger characterization capabilities and higher sensitivity for diverse progressive failure modes.

[0025] 3. This invention constructs a health baseline model using a self-supervised variational autoencoder, with a significant advantage being that it requires no pre-labeled fault samples. In real-world engineering, acquiring and labeling fault data covering all failure modes is costly and virtually impossible. This invention establishes a complete "normal" behavior model simply by learning the data distribution of power devices in their initial healthy state, greatly reducing reliance on prior knowledge and fault data, thus making this method highly practical and universally applicable in engineering.

[0026] 4. By introducing a dynamic environmental anchor calibration mechanism, this invention can proactively identify and quantify compensation for systematic measurement drift caused by factors such as ambient temperature and background noise intensity. This mechanism accurately separates environmental influences from the original measurement characteristics, ensuring that the final output physical degradation index purely reflects the physical performance degradation of the device itself. This effectively avoids false alarms caused by fluctuations in environment and operating conditions, significantly improving the stability and reliability of monitoring results during long-term, continuous operation.

[0027] 5. This invention precisely focuses the energy of the probe pulse onto the most sensitive inherent resonant frequency band of the device under test. This is akin to a precise acoustic "percussion" of a physical system, capable of highly efficient excitation and amplification of its structural response. Therefore, when the device experiences minor changes in its resonant characteristics due to early structural damage such as microcracks caused by aging, these changes are significantly amplified, thereby achieving extremely high detection sensitivity for early-stage faults and providing the possibility for truly predictive maintenance. Attached Figure Description

[0028] Figure 1 This is a functional block diagram of the online monitoring system for inverter performance degradation of the present invention;

[0029] Figure 2 This is a flowchart of the online monitoring method for inverter performance degradation according to the present invention;

[0030] Figure 3 This is a schematic diagram illustrating the performance degradation assessment principle based on a healthy baseline manifold of the present invention. Detailed Implementation

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please refer to the attached figures*-*. This embodiment of the invention provides an online monitoring method for inverter performance degradation, implemented through a monitoring device deployed in the inverter system under test. This monitoring device can be physically integrated with the inverter or exist as an external module. It mainly consists of several functional units working collaboratively to execute the monitoring process of this invention.

[0033] The monitoring system consists of a multimodal sensing unit, a pulse injection unit, a synchronous data acquisition unit, and a central processing unit. These units are interconnected to form a complete information acquisition and processing chain.

[0034] The multimodal sensing unit is designed to non-invasively capture the physical response signals generated by power devices when stimulated. Specifically, the unit includes at least one ultra-wideband electromagnetic probe and a microelectromechanical system (MEMS) ultrasonic sensor. These two types of sensors are closely and co-located in the near-field of the target power device's package to ensure simultaneous, high-fidelity capture of transient signals from both electromagnetic and acoustic modes originating from the same physical event.

[0035] The pulse injection unit functions to generate a standardized excitation source for exciting the target power device. This unit is connected to the inverter's gate drive circuit and is controlled by the central processing unit. During operation, it precisely injects a pre-set probe pulse with a steep edge and extremely low energy into the gate of the power device within a specified dead-time window. The pulse parameters are carefully designed so that its energy is sufficient to excite a measurable transient response within the device's internal physical structure, but far from enough to enable effective conduction, thus ensuring that the probe activity does not interfere with the normal operation of the inverter's main circuitry.

[0036] The synchronous data acquisition unit serves as a bridge connecting physical sensing and digital processing. This unit features multiple parallel channels and high temporal resolution and synchronization accuracy. It is responsible for performing strictly synchronized analog-to-digital conversion and data buffering of electromagnetic and acoustic signals from the multimodal sensing unit. Its high synchronization accuracy ensures the consistency of the two modal signals in terms of timestamps, providing a reliable data foundation for the subsequent construction of a structured acoustic-electric resonance fingerprint.

[0037] The central processing unit (CPU) is the control and computational core of the entire monitoring system, typically composed of a digital signal processor (DSP), a field-programmable gate array (FPGA), or a combination thereof. This unit performs multiple responsibilities: First, it analyzes the inverter's PWM control logic, accurately identifies the start and end times of the dead time, and issues commands to the pulse injection unit; second, it controls the startup and data transmission of the synchronous data acquisition unit; finally, and most importantly, the CPU executes pre-set algorithm models, completing the entire process from raw signal processing, acoustic-electric resonance fingerprint construction, health baseline model loading, dynamic environmental calibration to the final performance degradation index calculation, and storing the evaluation results or reporting them to the upper-level monitoring system via a communication interface.

[0038] In summary, the workflow of the method of the present invention at the system level is as follows: the central processing unit initiates and schedules the active and non-disruptive transient excitation of the target power device through the pulse injection unit, the multi-modal sensing unit captures its physical intrinsic response, the synchronous data acquisition unit completes the digitization, and finally the central processing unit performs a series of precise algorithm calculations to obtain a quantitative indicator that can accurately reflect the health status of the device.

[0039] In one specific embodiment of the present invention, the construction process of the acoustic-electric resonance fingerprint is meticulously designed to obtain a standardized data sample that can best reflect the physical intrinsic characteristics of the target power device and possesses high signal-to-noise ratio and high repeatability. This process forms the basis for all subsequent analyses and evaluations, and its specific implementation involves comprehensive consideration of the detection timing, excitation source, and signal synthesis method.

[0040] A key feature of this method is that the signal excitation and acquisition do not occur during the normal switching cycle of the inverter carrying the main power flow, but rather during the dead time of the pulse width modulation (PWM) signal. The fundamental reason for choosing this specific time window is that the power circuit is in a high-impedance state during the dead time, with no main current flowing through it, thus greatly suppressing the strong electromagnetic background noise generated by load fluctuations, bus voltage ripple, and main circuit switching operations. This "active silent detection" strategy creates an extremely clean electromagnetic environment for capturing the weak physical response signals generated by low-energy excitation sources, fundamentally solving the technical problem in traditional online monitoring methods where useful signals are submerged in strong background noise.

[0041] Within a defined dead-time window, the pulse injection unit injects a standardized, nanosecond-level probe pulse, denoted as p(t), into the gate of the target power device. This probe pulse is designed with steep rise and fall edges, functioning similarly to a precise physical "knock," capable of exciting complex, high-frequency transient oscillations within the semiconductor lattice, metallization layer, bonding wires, and packaging materials inside the device. The energy of this pulse is strictly controlled to ensure that it serves only as an excitation source for a single physical probe and does not have any substantial impact on the device's macroscopic electrical state (such as on or off).

[0042] At the instant the probe pulse p(t) is injected, the multimodal sensing unit deployed in the near field of the device is synchronously triggered to capture its physical response. The near-field electromagnetic radiation signal captured by the ultra-wideband electromagnetic probe, generated by the rapid redistribution of charge carriers within the device and the transient high-frequency current loop, is denoted as time series x. e (t); simultaneously, the acoustic vibration signal captured by the MEMS ultrasonic sensor, generated by the piezoelectric effect and the propagation of stress waves from the microstructure, is denoted as the time series x. a (t). These two time-series signals together constitute a complete multi-dimensional, multi-physics description of the same physical excitation event.

[0043] To integrate these two one-dimensional time-series signals into a structured data sample that is easy for subsequent model processing, this invention proposes constructing a two-dimensional acoustic-electric resonance fingerprint. Specifically, the synchronously acquired electromagnetic response signal x e (t) and acoustic response signal x a (t) are two independent channels, precisely aligned in the time dimension, and stacked into a two-dimensional matrix F. This matrix is ​​the acoustic-electric resonance fingerprint, which can be mathematically expressed as:

[0044]

[0045] in, The sampling time points represent discrete points; N is the total number of sample points collected each time. The fingerprint matrix F constructed in this way not only preserves the complete time series information within each signal, but more importantly, it maintains a strict phase-time correspondence between the two channels, making it possible for deep learning models to mine complex coupled features across modalities.

[0046] In one specific embodiment of the present invention, the process of establishing a health baseline model for the acoustic-electric resonance fingerprint is a key step in achieving high-precision degradation assessment. The core of this process lies in the ability to construct a mathematical model that accurately characterizes the normal behavior pattern of a target power device, without relying on any pre-labeled fault samples, solely by learning the fingerprint data of the target power device in its initial healthy state. To this end, the present invention employs a generative deep learning model, namely a variational autoencoder, to complete this self-supervised learning task.

[0047] The variational autoencoder was chosen because its unique structure allows it to learn the intrinsic probability distribution of the data, rather than simply a deterministic mapping. The model consists of a two-part neural network—an encoder and a decoder—that work together to compress high-dimensional input data into a low-dimensional, structured latent space, from which the original data can be reconstructed. This characteristic makes it well-suited for modeling "normal" data patterns and highly sensitive to anomalous data deviating from these patterns.

[0048] The encoder q θ (z|F), in this embodiment, can be implemented by a convolutional neural network (CNN) because it can efficiently extract the spatial and temporal coupling features contained in the two-dimensional acoustic-electric resonance fingerprint matrix F. The encoder's function is to map the input fingerprint matrix F to a Gaussian distributed parameter in a low-dimensional latent space, i.e., the mean vector μ. z Sum of logarithmic variance vector The decoder p θ (F′|z) employs a transposed convolutional network symmetrical to the encoder structure. Its function is to sample a point z from the latent space and upsample it to reconstruct a reconstructed fingerprint matrix F′ with the same dimensions as the original fingerprint. The latent variable z is obtained by sampling from the distribution defined by the encoder through a reparameterization technique, i.e., z = μ. z +∈·σ z , where ∈ is from the standard normal distribution Random noise extracted from the sample.

[0049] The training process of the health baseline model is as follows: First, in the initial stage of inverter system operation, a large number of acoustic-electric resonance fingerprint samples representing the initial health state of the target power device are collected under various typical ambient temperatures. Using this clean health dataset, the variational autoencoder model is trained end-to-end, with the optimization objective being to minimize a comprehensive loss function L. VAE The loss function consists of two parts:

[0050] L VAE (F,F′;φ,θ)=L recon +β·D KL (q φ (z|F)||p(z))

[0051] The first part is the reconstruction loss L. recon It measures the difference between the original input fingerprint F and the reconstructed fingerprint F′ obtained through the encoding-decoding process. This loss term drives the model to learn the ability to compress and decompress healthy fingerprints as losslessly as possible. In this embodiment, mean squared error can be used to calculate:

[0052]

[0053] Where F ij and F′ ij These are the elements in the original fingerprint matrix and the reconstructed fingerprint matrix, respectively, where N is the time series length of the fingerprint.

[0054] The second part is the Kullback-Leibler (KL) divergence term D. KL (q φ (z|F)||p(z)). This term acts as a regularizer, measuring the posterior probability distribution q generated by the encoder. φ The distance between (z|F) and a pre-defined standard Gaussian prior p(z). This regularization term forces all healthy fingerprints, after being encoded into the latent space, to be distributed as compactly as possible and clustered near the origin, thus forming a regular and continuous distribution region. β is a hyperparameter used to balance reconstruction quality with the degree of latent space regularization.

[0055] Once the model training converges, the high-probability-density regions in the latent space occupied by the acoustic-electric resonance fingerprints of all healthy states, as learned by the model, collectively constitute a mathematically complete healthy baseline manifold. This manifold is the final embodiment of the health baseline model constructed in this invention, which provides a precise benchmark for subsequent judgment of whether new and unknown fingerprints deviate from the "healthy" category.

[0056] In a preferred embodiment of the present invention, a dynamic environmental anchor calibration mechanism is introduced to further improve the stability and reliability of monitoring results during long-term operation. This mechanism aims to proactively identify and compensate for systematic drift caused by environmental factors, particularly temperature changes, to the measurement system (including the sensor itself and the signal transmission link). This ensures that the final evaluation results more purely reflect the physical performance degradation of the target power device itself, rather than environmental disturbances.

[0057] The core of this mechanism lies in constructing an environmental anchor vector that can represent the current instantaneous environmental state. In this embodiment, the environmental anchor vector V env It consists of two key real-time parameters. First, the real-time surface temperature T of the target power device module, directly measured by a temperature sensor. Second, during the inverter's normal PWM switching operation (non-dead time), a background noise fingerprint is collected, and an energy feature B that stably reflects the electromagnetic environment intensity under this operating condition is extracted from it; for example, the root mean square value of the background noise signal. These two parameters together constitute the environmental anchor-point vector.

[0058] V env =[T,B]

[0059] Among them, temperature T directly reflects the influence of thermal effect on the physical properties of device materials and sensor sensitivity, while background noise energy B indirectly reflects the intensity of the current load condition. The combination of the two can comprehensively characterize the environmental conditions affecting the measurement.

[0060] To decouple environmental impacts, this invention designs a mapping network specifically for learning and compensating for environmental effects, denoted as g. ψo This mapping network is typically a lightweight feedforward neural network, whose input is the real-time acquired environmental anchor vector V. env The output is an environmental offset vector Δz with the same latent spatial dimension as the health baseline model. envo .

[0061] Δz env =g ψ (V env )

[0062] The mapping network g ψ The training of the network g is performed simultaneously with the training of the health baseline model. When collecting healthy fingerprint samples, the environmental anchor vector corresponding to each fingerprint is recorded simultaneously. The network g is trained by analyzing the systematic positional drift of healthy acoustic-electric resonance fingerprints in the latent space under different environmental anchor points. ψ This enables it to accurately predict the state of any environment V. envThe resulting potential vector offset Δz envo .

[0063] In the actual online monitoring phase, this dynamic environmental anchor calibration mechanism operates as follows: when a new acoustic-electric resonance fingerprint F... new After being collected, the data is first mapped to the latent space by the encoder of the health baseline model, resulting in an uncalibrated raw latent vector z. raw At the same time, the system acquires and F new Acquire synchronous environmental anchor vector V env_new And input it into the pre-trained mapping network g ψ Calculate the offset vector Δz under the current environment. envo .

[0064] Finally, by subtracting the corresponding environment offset vector from the original latent vector, an environment-calibrated latent vector z can be obtained. cal The mathematical expression for this calibration process is:

[0065]

[0066] The physical meaning of this operation is that, regardless of the current actual measurement environment, this mechanism can, through mathematical transformation, "pull back" the measurement results to a unified reference coordinate system consistent with the initial training environment. In this way, fingerprint changes caused by environmental factors such as temperature fluctuations are effectively filtered out, ensuring that subsequent health assessments are based entirely on a stable and comparable benchmark, thereby greatly improving the accuracy of monitoring.

[0067] In a specific implementation of this invention, once the inverter system enters online operation, the central processing unit periodically executes a complete health assessment process to continuously monitor the performance degradation of the target power devices. This process integrates the aforementioned core technology modules such as fingerprint construction, model evaluation, and dynamic calibration, ultimately outputting an indicator that can intuitively quantify the health status of the devices.

[0068] A complete online evaluation cycle is implemented through the following steps: First, the system enters standby mode and continuously monitors the inverter's PWM control signal. Once the arrival of the next dead time is detected, an active probe is immediately triggered. As mentioned earlier, a standardized probe pulse is injected into the target power device at this time, and its generated electromagnetic and acoustic response signals are simultaneously acquired to construct the acoustic-electric resonance fingerprint of the current moment, denoted as F. new At the same time, the system records real-time environmental parameters that are completely synchronized with this data collection, forming the current environmental anchor vector V. env_new .

[0069] Next, the newly acquired fingerprint matrix F new The input is fed into the encoder portion of the trained health baseline model, which is embedded in the central processing unit. The encoder performs forward propagation computation on this input, mapping it to the model's low-dimensional latent space to obtain an original latent vector V. env_new .

[0070] Subsequently, the environmental calibration mechanism is initiated. The synchronously acquired environmental anchor vector V... emv_new The input is fed into a pre-trained environment influence mapping network gψgψ, which outputs a specific offset vector Δz corresponding to the current environment. emv_new Subsequently, the system performs vector subtraction to remove the environmental influences from the original latent vectors, obtaining the calibrated latent vectors.

[0071] z cal_new =z raw_new -Δz emv_new

[0072] Obtain the environmentally calibrated latent vector z caLnew Then, it is input into the decoder part of the health baseline model. Based on this vector, the decoder reconstructs, to the best of its ability, a value it considers to be closest to z in the baseline health state. cal_new The fingerprint matrix representing the state. This reconstructed fingerprint, denoted as F′, is the reconstructed fingerprint after environmental calibration. cal new .

[0073] Finally, this invention defines a physical degradation index to quantify the health status of a target power device. The core idea of ​​this index is that if a device is healthy, then its true fingerprint F... new After environmental calibration, the fingerprint should be reconstructed almost perfectly by the healthy baseline model. Conversely, if the device experiences physical performance degradation (such as bond wire fatigue, solder layer void deterioration, etc.), its true fingerprint structure will undergo fundamental changes. These changes are beyond the understanding and reconstruction capabilities of the healthy model, resulting in a suboptimal reconstructed fingerprint F′. cal new With the original fingerprint F new Significant differences were observed between them.

[0074] Therefore, the physical degradation index is precisely defined as the currently acquired acoustic-electric resonance fingerprint F. new The corresponding reconstructed fingerprint F′ after calibration cal new The square of the Euclidean distance between them is also the square of the Frobenius norm of the difference matrix between them. The formula for its calculation is as follows:

[0075]

[0076] The PDI value is a scalar that directly reflects the "distance" the current device state deviates from its inherent healthy manifold. A stable sequence of low PDI values ​​indicates that the device is in good operating condition. However, when the PDI value exhibits an irreversible and continuous upward trend, it clearly indicates that physical performance degradation is occurring within the device, thus providing a timely and reliable basis for decision-making regarding predictive maintenance.

[0077] In a more preferred embodiment of the invention, the design of the injected probe pulse itself is further optimized to achieve more sensitive capture of the degradation characteristics of the target power device. The core of this embodiment lies in upgrading the probe pulse from a fixed, wide-spectrum excitation source to an energy-focused excitation source that adaptively matches the inherent resonant characteristics of the device under test.

[0078] The technical motivation behind this optimization lies in the fact that every physical system, including power devices themselves, possesses inherent resonant frequencies determined by its physical structure and material properties. The system will produce the strongest response when the frequency of the external excitation coincides with these resonant frequencies. Therefore, by actively identifying and utilizing these resonant frequencies, a response signal with a significantly higher signal-to-noise ratio can be obtained with lower excitation energy, thereby improving monitoring sensitivity.

[0079] The implementation of this adaptive resonant detection includes an initial "resonance characteristic identification" phase. This phase is performed during the initial baseline modeling of the device. At this time, a probe signal with a very wide spectrum, such as a linearly modulated (Chirp) pulse or an approximate Dirac impulse pulse, is injected into the target power device. The response signal generated by this broadband excitation (e.g., the electromagnetic response signal x) is then processed. e The data (t) is collected and its spectrum is analyzed. The Fast Fourier Transform (FFT) is typically used to calculate its spectral density function X. e (f):

[0080]

[0081] in, This represents the Fourier transform operator. It is derived from the spectral density function |X|. e By searching for peak points in (f)|, one or more main resonant frequency points of the device in its current healthy state can be determined, denoted as {f}. r1 ,f r2 ,…,f rk}

[0082] After identifying these key resonant frequencies, the system will synthesize an optimized probe pulse accordingly. The energy of the optimized pulse is no longer uniformly distributed across a wide frequency band, but is precisely concentrated near one or more identified resonant frequency points. For example, this optimized probe pulse can be combined into a multi-tone signal, in the form of:

[0083]

[0084] Among them, A i and φ i These correspond to the i-th resonant frequency f. ri The amplitude and phase of the pulse, w(t) is a time window function used to ensure that the duration of the pulse is finite.

[0085] In all subsequent online monitoring cycles, the system will use this optimized probe pulse, "tailor-made" for this specific device. opt (t) is used instead of the general broadband pulse. Because the excitation energy is precisely applied to the most sensitive frequency band of the device, the resulting acoustic-electric resonant fingerprint will have significantly improved amplitude and signal-to-noise ratio.

[0086] The direct benefit of this preferred embodiment is that when a device experiences minor physical degradation (e.g., early cracking of the bonding wire), changes in its physical structure first cause a slight drift in its resonant frequency. Since the excitation used in this method is focused on these resonant points, it is extremely sensitive to frequency drift or changes in response amplitude. These changes are significantly amplified and reflected in the acquired acoustic-electric resonance fingerprint. Ultimately, this translates into earlier and larger jumps in the Physical Degradation Index (PDI) of the inverter performance online monitoring method, thereby achieving advanced detection and early warning of incipient faults.

[0087] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for on-line monitoring of performance degradation of an inverter, characterized in that, The method comprises the following steps: a. injecting a preset detection pulse into the target power device during the dead time of the pulse width modulation signal of the inverter, and synchronously collecting the electromagnetic signal and the acoustic signal generated in response to the target power device to construct an acoustic-electric resonance fingerprint; b. inputting the acoustic-electric resonance fingerprint into a pre-constructed health baseline model; c. calculating and determining a performance degradation index of the target power device based on the output of the health baseline model.

2. The method of claim 1, wherein, The step of constructing the acoustic-electric resonance fingerprint specifically comprises aligning and combining the synchronously collected electromagnetic signal and acoustic signal as two dimensions on the time axis to form a two-dimensional fingerprint matrix.

3. The method of claim 1, wherein, The health baseline model is a variational autoencoder model, which constructs a latent space health manifold representing the healthy state through self-supervised learning on multiple acoustic-electric resonance fingerprints collected in the healthy state.

4. The method of claim 1, wherein, Before step b is performed, the following steps are further included: obtaining real-time environmental parameters synchronized with the acoustic-electric resonance fingerprint collection to constitute an environmental anchor point, and dynamically calibrating the evaluation benchmark of the health baseline model using the environmental anchor point.

5. The method of claim 4, wherein, The real-time environmental parameters at least include the real-time temperature of the target power device and the background noise characteristics collected during normal switching of the inverter.

6. The method of claim 4, wherein the step of determining the performance degradation of the inverter is performed by using a neural network. The step of calculating the performance degradation index specifically comprises inputting the acoustic-electric resonance fingerprint into the encoder of the health baseline model to obtain an original latent vector, calculating an environmental offset vector according to the environmental anchor point, and calibrating the original latent vector to obtain a calibrated latent vector, inputting the calibrated latent vector into the decoder of the model to obtain a calibrated reconstructed fingerprint, and finally determining the performance degradation index by calculating the error between the acoustic-electric resonance fingerprint and the calibrated reconstructed fingerprint.

7. The online monitoring method for inverter performance degradation according to claim 6, characterized in that, The performance degradation index (PDI) is determined by the following formula: where F new is the current collected acoustic-electric resonance fingerprint, F′ cal_new is the calibrated reconstructed fingerprint, and ∥·∥ F is the Frobenius norm.

8. The method of claim 1, wherein, The spectral characteristics of the preset detection pulse are adaptively adjusted and focused on the resonance frequency point of the target power device based on the analysis of the initial response of the target power device.

9. The method of claim 6, wherein the step of determining the performance degradation of the inverter is performed by using a neural network. The environmental offset vector is calculated by a pre-trained mapping network, and the input of the mapping network is the environmental anchor point, and the output is the environmental offset vector.

10. The method of claim 3, wherein, The variational autoencoder model is trained by minimizing the following loss function L VAE Training is performed: L VAE = L recon + β · D KL (q φ (z | F) || p(z)); where L recon is the reconstruction loss, D KL is the KL divergence, β is a balancing factor, q φ (z|F) is the posterior distribution of the encoder output, p(z) is the prior distribution, and F is the acoustic-electric resonance fingerprint used for training.

Images