Sequential network collaborative diagnosis method and system based on IGBT module switching speed and electromagnetic interference

Abstract

The invention discloses a sequential network collaborative diagnosis method and system based on IGBT module switching speed and electromagnetic interference, and relates to the field of power electronic technology state monitoring. At present, IGBT switching speed slowing signals and external electromagnetic interference signals in the MMC are easy to confuse and misjudge, and traditional intrusive measurement is low in reliability. The method comprises the steps that MMC radiation electromagnetic field time sequence signals are collected in a non-intrusive mode, after marking and preprocessing are conducted, a collaborative neural network containing a shared convolution feature extraction module and three sub-models is adopted, signal types are classified firstly, and then the switching speed state or the electromagnetic interference type is diagnosed in a targeted mode. According to the technical scheme, the characteristics representing device switching speed change and the characteristics of external interference can be effectively decoupled and extracted from complex time sequence signals containing noise and interference, and the accuracy and reliability of power sub-module health state evaluation are improved.

Description

Technical Field

[0001] This invention relates to the field of power electronics condition monitoring technology, and in particular to a time-series network collaborative diagnosis method and system based on IGBT module switching speed and electromagnetic interference. Background Technology

[0002] Modular multilevel converters (MMCs) are key core equipment in modern high-voltage direct current (HVDC) transmission, flexible AC transmission, and large-scale renewable energy grid integration. These converter systems are large in scale, consisting of hundreds or even thousands of cascaded submodules, each with power switching devices at its core. These numerous power switching devices form the basis for the stable operation of the MMC, and their health directly affects the safety and reliability of the entire power system. However, the devices in the power submodules must continuously withstand complex and harsh operating conditions during long-term operation, facing both gradual aging of internal components and electromagnetic interference from the external operating environment. On the one hand, under repeated electrical and thermal stress shocks, the devices experience cumulative physical damage, reflecting changes in their gradual switching characteristics and no longer meeting the requirements for loop switching (turn-on and turn-off) speeds. On the other hand, the converter valves operate in a complex, strong electromagnetic environment. When subjected to electromagnetic pulse signals caused by external events such as lightning or HEMPs, transient disturbances occur to the system, requiring identification of the affected power submodules for maintenance.

[0003] Traditional diagnostic methods for identifying switching speed and external electromagnetic interference (EMI) in power devices rely on invasive measurements of electrical signals from internal converter modules. This requires installing numerous internal sensors within the valve tower to collect voltage and current signals, drastically increasing system cost, size, and complexity. Furthermore, the introduced measurement circuitry and sensors themselves can become new potential fault points, ultimately reducing overall system reliability. For electromagnetic field probes, the switching process and radiated electromagnetic signals from external EMI exhibit similar characteristics, showing rapid, instantaneous changes. Single-model methods are prone to failure, potentially misinterpreting transient external disturbances as electromagnetic signals from internal device or circuit changes, leading to erroneous alarms and unnecessary downtime. Power devices radiate electromagnetic field signals during high-speed switching and when subjected to external EMI. Algorithms can infer operating status information from these signals. Therefore, a non-invasive, collaborative diagnostic method is needed to analyze these radiated signals, accurately decoupling and distinguishing between internal switching speed changes and external EMI events. Summary of the Invention

[0004] The technical problem to be solved and the technical task proposed by this invention is to improve and refine existing technical solutions, and to provide a time-series network collaborative diagnostic method and system based on IGBT module switching speed and electromagnetic interference, with the aim of accurately distinguishing between the gradual changes in the internal switching speed characteristics of MMC power devices and the signals of transient events of external electromagnetic interference through non-invasive measurement of electromagnetic signals. To this end, this invention adopts the following technical solution.

[0005] A time-series network-based diagnostic method for IGBT module switching speed and electromagnetic interference includes the following steps:

[0006] 1) Using a non-invasive method, collect the radiated electromagnetic field timing signal data generated during the operation of the modular multilevel converter; and label the timing signal data; the labeling includes signal type label, device switching speed status label, and electromagnetic interference type label; 2) Preprocess the labeled time series signal data, including cutting the long time series signal into time series segment samples of fixed length and performing amplitude normalization processing on the samples. 3) Construct a collaborative diagnostic neural network model, which includes a shared convolutional feature extraction module and three sub-models connected thereto: a signal type recognition sub-model, a device switching speed diagnosis sub-model, and an electromagnetic interference type recognition sub-model; 4) The collaborative diagnostic neural network model is trained step-by-step using the preprocessed time-series signal data; 5) Input the radiated electromagnetic field signal to be diagnosed into the trained collaborative diagnostic neural network model. First, the signal type identification sub-model determines the signal type. Then, based on the determination result, the device switching speed diagnostic sub-model or the electromagnetic interference type identification sub-model is conditionally called to output the diagnostic result of the switching speed status or electromagnetic interference type.

[0007] This technical solution effectively solves the challenge of accurately distinguishing between internal state changes and external transient interference in power devices under complex electromagnetic environments through non-invasive signal acquisition, structured data annotation, standardized preprocessing, collaborative model architecture, targeted training strategies, and diagnostic processes. Specifically, this solution only requires acquiring electromagnetic field signals radiated from the outside of the device, eliminating the need to install sensors or measurement circuits inside complex power electronic equipment. This achieves integrated, non-invasive online monitoring and diagnosis of the switching speed status inside the power module and the type of external electromagnetic interference. Multi-dimensional annotation of the raw time-series signal data provides three key information labels for subsequent model training: signal source category (internal switching / external interference), internal device health status (switching speed level), and external interference type. This constructs a structured diagnostic knowledge base, laying a solid foundation for accurate model learning. By performing fixed-length slicing and normalization preprocessing on long time-series signals, continuous monitoring data is transformed into standardized samples suitable for batch processing by neural networks, improving computational efficiency. Furthermore, it eliminates differences in signal amplitude dimensions, enhancing the convergence and stability of model training. The constructed neural network model adopts a collaborative architecture of "shared feature extraction + dedicated sub-models." The shared convolutional feature extraction module avoids redundant feature calculations, improving efficiency. The three dedicated sub-models have clearly defined roles, responsible for signal classification, switch speed diagnosis, and interference type identification, respectively. This task decoupling design makes the model structure clearer and more targeted. Step-by-step training of the collaborative diagnostic model allows each sub-model to be fully optimized on its most relevant subset of data, avoiding mutual interference between different types of signal features in a single model. This helps to improve the professionalism and accuracy of switch speed diagnosis and interference type identification, respectively. In actual diagnosis, the model automatically determines the type of input signal through an intelligent decision-making process of "classification first, routing later," and conditionally calls the corresponding dedicated diagnostic sub-model, reducing misjudgments and ensuring that the output diagnostic results (switch speed status or interference type) highly match the signal properties, improving the reliability and practicality of the diagnostic results.

[0008] As a preferred technical means: In step 1), the timing signal is obtained through electromagnetic transient simulation or experimental platform, including: establishing a benchmark model to obtain the signal at the nominal switching speed; simulating different levels of slowing down switching speed by adjusting the IGBT gate resistance or equivalent circuit parameters; and applying different types of external electromagnetic pulse interference to construct a signal sample set containing concurrent operating conditions.

[0009] This technical solution, by combining electromagnetic transient simulation and experimental platforms, can systematically construct a complete signal sample set covering nominal state, progressive aging state, and external interference state. A benchmark model is established to obtain the nominal speed signal, ensuring the accuracy of the health benchmark. By adjusting key parameters such as gate resistance to simulate the slowing state, the gradual change process of switching characteristics caused by aging or parameter drift is reproduced. Applying different types of external electromagnetic pulses covers the main interference sources that may be encountered in the actual environment. This construction method ensures that the sample library can highly realistically reflect the various single and concurrent operating conditions that the equipment may face in actual operation, providing a solid data foundation for training a robust diagnostic model. Compared to simply relying on fault data that is accidentally captured in the field, which is uncontrollable and difficult to label, this technical solution actively generates the required operating conditions on a simulation and controlled experimental platform. By adjusting specific and quantifiable physical or equivalent parameters such as IGBT gate resistance, different levels of switching speed degradation states can be accurately and repeatedly generated, establishing a deterministic correspondence between circuit parameter changes and signal characteristics. Simultaneously, this method can safely and efficiently apply high-intensity electromagnetic interference that is difficult to actively introduce into real systems, greatly overcoming the engineering bottlenecks of difficult fault sample acquisition, long cycles, and high labeling uncertainty, and significantly improving model development efficiency and data quality. This technical solution constructs a signal sample set containing concurrent operating conditions, i.e., simulating a composite scenario of slowed switching speed and external electromagnetic interference intrusion. By actively creating and labeling these complex and easily confused samples, the neural network model is forced to delve deeper into and distinguish the subtle differences between these two types of signals during the learning process, thereby significantly improving the model's generalization ability and diagnostic accuracy in real complex environments and avoiding misjudgments due to similar features. This technical solution directly associates the switching speed state with specific, operable circuit parameter adjustments and corresponds the interference type with specific standard interference sources. This makes the switching speed state labels and electromagnetic interference type labels for the samples have clear and explicit physical meaning and engineering correspondence, rather than vague state descriptions. Models trained with this type of data can more directly reflect the actual physical state or interference events, enhancing the interpretability and engineering guidance value of diagnostic results.

[0010] As a preferred technical means: the device switching speed status label includes nominal speed, moderate slowdown and severe slowdown; the electromagnetic interference type label includes high-altitude electromagnetic pulse interference, lightning interference and electrostatic discharge interference.

[0011] This technical solution uses a three-level labeling system for device switching speed status: nominal speed, moderate slowdown, and severe slowdown, rather than a simple dichotomy of normal and abnormal. This grading directly corresponds to the gradual process of device performance degradation, enabling diagnostic results to not only indicate the presence of a fault but also quantitatively assess the severity of aging. Moderate slowdown serves as an early warning, while severe slowdown points to a critical state requiring urgent maintenance, thus providing a refined decision-making basis for predictive maintenance and achieving a leap from fault alarm to health status grading assessment. Electromagnetic interference type labels include high-altitude electromagnetic pulse interference, lightning interference, and electrostatic discharge interference, avoiding vague external interference labels. This allows the diagnostic model to learn and distinguish the unique signal characteristics of different interference sources, identifying specific interference types. This has crucial engineering guidance value for tracing interference sources, assessing system electromagnetic vulnerability, and implementing targeted protection and hardening measures. The three-level discretization of switching speed transforms the continuous performance degradation problem into a clear classification problem, reducing the difficulty of model learning.

[0012] As a preferred technical means: in the preprocessing of step 2), a sliding time window is used to cut the continuous radiation electromagnetic field time domain waveform into time-series segments of equal length, and the amplitude of each segment is normalized.

[0013] This technical solution uses a sliding time window to segment continuous, variable-length real-time radiated electromagnetic field waveforms into a large number of equal-length, structurally regular time-series samples. This transforms streaming monitoring data into a standard input format suitable for batch training and parallel inference of deep learning models, improving data throughput and model computation efficiency, and providing a foundation for real-time online diagnosis. The sliding window traverses the entire signal in an overlapping or continuous manner, ensuring that no transient switching events or pulse-type electromagnetic interference occurring at any moment are missed and are always fully captured by one or more windows. This avoids truncation or loss of event features, enabling the model to learn the complete waveform shape and temporal evolution of each event, preserving key information for subsequent accurate classification and diagnosis. Amplitude normalization is performed on each time-series segment, effectively eliminating absolute amplitude differences caused by factors such as sensor sensitivity differences, installation location changes, or signal propagation attenuation. After being segmented and normalized, the time-series segments form a clean and consistent model input, providing optimal working conditions for the subsequent convolutional feature extraction module. This enables the convolutional kernel to more effectively extract comparable local features among different samples, laying a reliable low-level data foundation for accurate judgment of high-level sub-models.

[0014] As a preferred technical means: In step 3), the core architecture of the convolutional feature extraction module, the signal type identification sub-model, the device switching speed diagnosis sub-model, and the electromagnetic interference type identification sub-model are all one-dimensional convolutional neural networks, and the signal type identification sub-model, the device switching speed diagnosis sub-model, and the electromagnetic interference type identification sub-model share the same depth temporal features extracted by the convolutional feature extraction module during diagnosis.

[0015] The three diagnostic sub-models (signal type identification, switching speed diagnosis, and interference type identification) share the same deep temporal features extracted by the same convolutional feature extraction module. This ensures that for any input signal, the complex and time-consuming convolutional feature extraction process only needs to be executed once. The resulting features can be used by all subsequent dedicated sub-models simultaneously or on demand, avoiding the need to build and run a complete feature extraction network separately for each sub-task. This significantly reduces the number of model parameters, computational burden, and memory usage, making it possible to achieve low-latency real-time online diagnosis. Since all subsequent diagnoses are based on the same set of feature representations, signal type judgment, switching speed evaluation, and interference type identification are all completed in a unified feature space. This avoids additional biases or noise introduced by inconsistencies in the feature extraction process, enabling synergy between the first-level "classification" (signal type identification) and the second-level "diagnosis" (speed or interference identification). The convolutional feature extraction module ensures the homogeneity of the decision-making basis for different sub-tasks, fundamentally improving the inherent consistency and overall accuracy of the entire system's diagnostic logic.

[0016] As a preferred technical means: In step 4), the collaborative diagnostic neural network model is trained, including: The signal type recognition sub-model is trained using all time-series samples and their signal type labels; The device switching speed diagnostic sub-model is trained using only samples labeled as device switching signals and their switching speed state labels; The electromagnetic interference type identification sub-model is trained using only samples labeled as electromagnetic interference signals and their interference type labels.

[0017] This technical solution divides the training task into two parts: training a general classifier (signal type recognition sub-model) and two specialized diagnostic models (switching speed and interference type sub-models). The general model accesses all data and learns to macroscopically distinguish the fundamental difference between the source of the signal (internal switching vs. external interference); while each specialized model performs deep learning only within its own specific data domain (pure switching signal or pure interference signal), focusing on finely identifying state or type changes within its domain. This avoids the difficulty of using a single model to simultaneously learn cross-domain tasks with potentially similar features but vastly different physical meanings, making the learning objective of each sub-model more singular and focused, thereby achieving higher potential accuracy in its respective task. Switching speed variations and transient waveforms of electromagnetic interference may share similar "rapidly changing" characteristics in the time domain. If a model with all labels is used for end-to-end training, it is highly susceptible to learning confusing and discriminative features, leading to negative transfer—that is, learning one task while interfering with another. This technical solution, through data isolation, ensures that the switching speed diagnostic sub-model has never seen external interference samples, and the electromagnetic interference identification sub-model has never seen internal switching samples. This allows each specialized model to uninterruptedly mine and enhance the unique subtle feature patterns of its diagnostic target in the feature space, greatly improving the specificity and reliability of their respective diagnoses. From an optimization perspective, each sub-model only needs to process its relevant subset, reducing noise and gradient interference from irrelevant data, effectively improving convergence speed, and resulting in a more stable training process. At the same time, due to task simplification, the reliance on complex multi-task balancing strategies (such as loss weight adjustment) is reduced, making the training process simpler and more controllable.

[0018] As a preferred technique: In step 4), the model training process divides the sample data into a training set and a validation set, uses the Adam optimizer to update parameters, introduces a Dropout layer in the hidden layer of the network, and dynamically adjusts the learning rate based on the validation set loss using a learning rate scheduler to prevent overfitting.

[0019] This technical solution employs the Adam optimizer for parameter updates, leveraging its adaptive learning rate and momentum acceleration features to automatically adjust the learning step size for each parameter. It performs exceptionally well when handling complex neural networks with sparse gradients or parameters of varying scales, providing technical support for efficient and stable optimization in complex electromagnetic signal feature spaces. This helps accelerate training convergence and improve final performance. A Dropout layer is introduced into the network's hidden layers, randomly "dropping" a portion of neurons during training. This forces the network to avoid over-reliance on a few specific neurons or feature combinations, thereby learning more robust and generalizable feature representations. This enhances the model's diagnostic stability and reliability in the face of noise and sample variations in real-world scenarios. Based on validation set loss, the learning rate is dynamically adjusted through a learning rate scheduler, enabling control over the training process. This automatically adapts to the needs of different training stages and is a key technique for ensuring full model convergence and preventing oscillations or overfitting.

[0020] As a preferred technical means: after step 5), a post-processing step is also included: setting a confidence threshold for the probability distribution output by the device switching speed diagnostic sub-model or the electromagnetic interference type identification sub-model, and only when the highest category probability exceeds the threshold is the corresponding diagnostic result determined and output. The diagnostic result is used to evaluate the device health status or system electromagnetic compatibility.

[0021] Another technical solution of the present invention is to provide a timing network collaborative diagnostic system based on IGBT module switching speed and electromagnetic interference, the timing network collaborative diagnostic system comprising: The signal acquisition and labeling module is used to acquire the radiated electromagnetic field timing signal generated by the modular multilevel converter in a non-intrusive manner, and to label the timing signal. The labeling includes signal type label, device switching speed status label and electromagnetic interference type label. The signal preprocessing module is used to preprocess the labeled time series signal, including cutting the long time series signal into time series segment samples of fixed length and performing amplitude normalization processing on the samples. The collaborative diagnostic neural network model includes a shared convolutional feature extraction module and three sub-models connected to it: a signal type recognition sub-model, a device switching speed diagnostic sub-model, and an electromagnetic interference type recognition sub-model. The model training module is used to train the collaborative diagnostic neural network model step by step using the preprocessed data; The diagnostic execution module is used to input the radiated electromagnetic field signal to be diagnosed into the trained collaborative diagnostic neural network model. First, the signal type identification sub-model determines the signal type, and then, based on the determination result, conditionally calls the device switching speed diagnostic sub-model or the electromagnetic interference type identification sub-model to output the diagnostic result of the switching speed status or electromagnetic interference type.

[0022] This technical solution employs a non-invasive acquisition method, eliminating the need for numerous invasive sensors and measurement circuits inside the MMC valve tower. This avoids damaging the original converter structure, reducing system hardware complexity and the risk of additional fault points, while also not affecting the normal operation of the MMC. Multi-dimensional annotations (signal type labels, device switching speed status labels, and electromagnetic interference type labels) directly correspond to the core objectives of subsequent model training, providing accurate and targeted data support for step-by-step training and laying the foundation for diagnostic accuracy. By segmenting long time-series signals into fixed-length fragments, the signals meet the input format requirements of the neural network model, ensuring consistency of signal samples during training and diagnosis. Amplitude normalization eliminates interference caused by differences in signal amplitude under different operating conditions, allowing the model to focus more on the core features of the time-series signal (rather than amplitude magnitude), effectively improving the model's adaptability and robustness to signals in different scenarios. The shared convolutional feature extraction module avoids redundant computations from multiple sub-models repeatedly extracting features, improving feature extraction efficiency and resource utilization. The three sub-models have clear divisions of labor and strong specificity, focusing on single tasks: signal classification, switching speed diagnosis, and interference type identification, respectively. Compared to a "single model handling multiple tasks" design, this reduces feature interference between different diagnostic targets, enabling more accurate capture of the core features of the corresponding task and improving diagnostic accuracy. The model training module adopts a "step-by-step training" strategy, matching dedicated training data to the task objectives of different sub-models (training the signal classification sub-model with full data, the speed diagnosis sub-model with a subset of switching signals, and the interference identification sub-model with a subset of interference signals). This allows each sub-model to complete specialized optimization based on the adapted data, avoiding the problem of "excessive generalization but insufficient specialization" caused by mixed data training, and ensuring optimal diagnostic performance for each sub-model on its corresponding task. The conditional call process of "classification first, diagnosis later" can first identify the core attributes of the signal to be diagnosed (switching signal / interference signal) and then call the dedicated sub-model in a targeted manner. This process avoids misjudgment caused by the similarity of the characteristics of two types of signals, and significantly improves the accuracy of the diagnostic results. At the same time, the conditional call only starts the sub-model of the corresponding task, without having to run all sub-models, which reduces the amount of computation in the diagnostic process, improves the efficiency of real-time diagnosis, and adapts to the needs of online monitoring.

[0023] As a preferred technical means: the signal type identification sub-model, the device switching speed diagnosis sub-model, and the electromagnetic interference type identification sub-model share the same depth temporal features extracted by the convolutional feature extraction module during diagnosis; The convolutional feature extraction module is used to extract deep temporal features from the input radiated electromagnetic field time-series signal segment; The signal type identification sub-model has its input connected to the output of the convolutional feature extraction module, and is used to output the signal type judgment result based on the depth temporal features. A device switching speed diagnostic sub-model, whose input is connected to the output of the convolutional feature extraction module, is used to output device switching speed state diagnostic results based on the depth time-series features. An electromagnetic interference type identification sub-model, whose input is connected to the output of the convolutional feature extraction module, is used to output electromagnetic interference type identification results based on the depth temporal features. The output of the signal type identification sub-model is connected to a selector. The selector selects to route the deep timing features to the device switching speed diagnostic sub-model or the electromagnetic interference type identification sub-model based on the received signal type judgment result, and outputs the diagnostic result of the corresponding sub-model.

[0024] The three sub-models share the same deep temporal features output by the convolutional feature extraction module, eliminating the need for separate feature extraction chains for each sub-model. This avoids repetitive convolution, activation, and pooling operations on the signal, significantly reducing computational load and resource consumption during the diagnostic process and improving overall diagnostic efficiency, especially suitable for online real-time monitoring scenarios. Signal type identification, switching speed diagnosis, and interference type identification are all based on the same set of deep temporal features, ensuring that the feature inputs for the three tasks are from the same source and standardized. This avoids feature bias or information fragmentation caused by independent feature extraction by different sub-models, making the logical correlation between signal classification and specialized diagnosis stronger, effectively reducing the risk of misjudgment caused by inconsistent features, and improving the accuracy and reliability of diagnostic results. By using a selector to route features based on the signal type judgment result, only the corresponding specialized sub-model (switching speed diagnosis / electromagnetic interference identification) participates in subsequent calculations, without requiring all sub-models to run in full. This reduces redundant time consumption from invalid calculations and allows the diagnostic process to focus on the target task, resulting in clearer logic and faster response, adapting to the actual needs of MMC systems for rapid fault / interference determination. The design of a shared convolutional feature extraction module reduces the repetitive structure of the model, making the overall network architecture more compact and less redundant. At the same time, the three types of tasks are based on the same feature extraction logic, and subsequent optimizations to the model (such as adjusting the convolutional kernel size and pooling method) can be applied to all sub-models simultaneously without the need to optimize independent feature extraction links separately, thus reducing the design, debugging and maintenance costs of the model.

[0025] Beneficial effects: 1. This technical solution achieves type decoupling of electromagnetic signals through non-intrusive online monitoring. Combined with a hierarchical collaborative architecture of "classification first, diagnosis later", it can separate internal switching speed changes and external transient interference transient characteristics from a single external electromagnetic field signal source.

[0026] 2) This technical solution improves the accuracy of diagnosis and the intelligence level of operation and maintenance decision-making. By collaboratively diagnosing and identifying switching speed status and interference type and outputting conditional results, it provides operation and maintenance personnel with clear and accurate decision-making basis, improving the accuracy and intelligence level of diagnosis, and has significant engineering application value and economic benefits. Attached Figure Description

[0027] Figure 1 This is a flowchart of the model training and collaborative diagnostic model construction process of the present invention.

[0028] Figure 2 This is a flowchart illustrating the signal diagnosis and classification application of the present invention.

[0029] Figure 3 This is a diagram of the collaborative diagnostic neural network model architecture of the present invention.

[0030] Figure 4 This is a schematic diagram of the one-dimensional convolutional neural network (1D-CNN) structure of the present invention. Detailed Implementation

[0031] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0032] Example 1: A collaborative diagnostic method for IGBT module switching speed and electromagnetic interference using a timing neural network, such as... Figure 1 As shown, it includes the following steps: S1: Data Acquisition and Labeling. Acquire radiated electromagnetic field timing signal data of the modular multilevel converter under different operating conditions using non-invasive sensors (such as near-field probes). Perform multi-dimensional labeling on the timing signal data. The multi-dimensional labeling includes signal type labels to distinguish signal sources, switching (on / off) speed status labels to characterize the relative level of device switching speed, and electromagnetic interference type labels to identify specific interference types.

[0033] Acquiring timing signal data is accomplished through electromagnetic transient simulation platforms or experiments. Specifically, this involves: first, establishing a benchmark model reflecting actual stable operating conditions to obtain "nominal speed" state data; then, simulating different deviations in switching speed relative to the benchmark model by adjusting the gate resistance Rg or equivalent circuit parameters of the IGBT devices to obtain waveforms for "moderately slowed" and "severely slowed" states; applying different types of electromagnetic interference, such as lightning electromagnetic pulse (LEMP), high-altitude electromagnetic pulse (HEMP), or switching transients, to construct a composite concurrent operating condition; and acquiring the corresponding radiated electromagnetic field time-domain waveforms using near-field probes. Finally, multi-dimensional labeling of the waveforms includes switching speed state tags and electromagnetic interference types.

[0034] The switching speed status specifically includes "nominal speed", "moderately slower" and "severely slower"; among which, "nominal speed" corresponds to the switching speed reference of MMC under actual stable operating conditions, and "moderately slower" and "severely slower" correspond to different delay levels of switching speed relative to the reference; the electromagnetic interference types specifically include "HEMP interference", "lightning interference" and "electrostatic discharge interference".

[0035] S2: Data Preprocessing. The labeled time-series signal data is preprocessed. A sliding time window is used to cut the long time-domain waveform into fixed-length time-domain segments (e.g., a window covering several IGBT switching cycles). Subsequently, the amplitude of each segment is normalized to eliminate the influence of dimensions.

[0036] S3: Model Construction. Construct a collaborative diagnostic neural network model with a hierarchical diagnostic architecture. This model includes a convolutional feature extraction module to receive time-series sample data and extract its deep temporal features; and three sub-models: a signal type identification sub-model DL1, used to determine the signal type based on the temporal features; a device switching speed status diagnostic sub-model DL2, used to diagnose device switching time variations based on the temporal features; and an electromagnetic interference type identification sub-model DL3, used to identify the electromagnetic interference type based on the temporal signal features.

[0037] The convolutional feature extraction module, the signal type recognition sub-model DL1, the device switching speed state diagnosis sub-model DL2, and the electromagnetic interference type recognition sub-model DL3 all adopt a one-dimensional convolutional neural network as the core architecture. Furthermore, DL2 and DL3 share the same depth temporal features extracted by the convolutional feature extraction module during diagnosis. The features used for training and recognition of the three models come from the same one-dimensional convolutional layer, activation layer, and pooling layer.

[0038] S4: Model Training. Using preprocessed time-series signal data, the collaborative diagnostic neural network model is trained step-by-step and independently under supervision. Specifically, during model training, the three sub-models are trained independently. The signal type recognition sub-model DL1 is trained under supervision using all training samples and their corresponding signal type labels; the device switching speed state diagnosis sub-model DL2 is trained under supervision using only a subset of samples labeled "device switching signals" and their corresponding switching speed state labels; and the electromagnetic interference type recognition sub-model DL3 is trained under supervision using only a subset of samples labeled "electromagnetic interference signals" and their corresponding electromagnetic interference type labels. The model training process divides the time-series training data into training and validation sets; it uses the Adam optimizer to iteratively update the model parameters; it sets Dropout layers between the hidden layers of the model; and it optimizes the training process and prevents overfitting by using a learning rate scheduler based on the loss rate changes of the validation set, and dynamically adjusts the learning rate based on the validation set loss.

[0039] S5: Diagnostic Application. Acquire the long-term radiated electromagnetic signal of the MMC to be diagnosed, and perform the same segmentation and normalization preprocessing as in step S2 to obtain one or more samples to be diagnosed. Input the samples to be diagnosed into the trained collaborative diagnostic model. DL1 first determines the signal type: if it is determined to be a "device switching signal", then DL2 is called to output the switching speed status diagnostic result; if it is determined to be an "electromagnetic interference signal", then DL3 is called to output the electromagnetic interference type diagnostic result.

[0040] Furthermore, after the collaborative diagnostic calculation, the output results are post-processed, a preset confidence threshold is set for the probability distribution value output by the sub-model, and the corresponding category is determined only when the maximum probability value exceeds the confidence threshold. The diagnostic results are used to evaluate the switching speed status of the device and provide a basis for equipment health management and potential aging trend analysis. The identification results are also used to evaluate the electromagnetic compatibility environment inside the MMC valve tower.

[0041] This invention uses a non-invasive radiated electromagnetic field signal as the sole input. By designing a hierarchical collaborative architecture of "classification first, diagnosis later" and combining it with an independent training strategy for specific targets, it achieves decoupling and collaborative diagnosis of the gradual characteristics of internal switching speed and the transient characteristics of external electromagnetic interference. This solves the problem of misjudgment caused by the approximate characteristics of transient signals of two types of signals in a single multi-task model, and provides a new and highly reliable intelligent diagnostic method for the safe and economical operation of MMC.

[0042] Example 2 A timing network collaborative diagnostic system based on IGBT module switching speed and electromagnetic interference is provided, characterized by comprising: I. Signal Acquisition and Labeling Module This is used to acquire the radiated electromagnetic field timing signals generated during the operation of a modular multilevel converter in a non-invasive manner, and to label the timing signals, including signal type labels, device switching speed status labels, and electromagnetic interference type labels.

[0043] II. Signal Preprocessing Module Used for preprocessing labeled time series signals, including cutting long time series signals into time series segment samples of fixed length and performing amplitude normalization on the samples; The collaborative diagnostic neural network model includes a shared convolutional feature extraction module and three sub-models connected to it: a signal type recognition sub-model, a device switching speed diagnostic sub-model, and an electromagnetic interference type recognition sub-model.

[0044] The signal type identification sub-model, the device switching speed diagnosis sub-model, and the electromagnetic interference type identification sub-model share the same depth-based temporal features extracted by the convolutional feature extraction module during diagnosis.

[0045] The convolutional feature extraction module is used to extract deep temporal features from the input radiated electromagnetic field time-series signal segments.

[0046] The signal type recognition sub-model has its input connected to the output of the convolutional feature extraction module, and is used to output the signal type judgment result based on the depth temporal features.

[0047] The device switching speed diagnostic sub-model has its input connected to the output of the convolutional feature extraction module, and is used to output the device switching speed state diagnostic results based on the depth time series features.

[0048] The electromagnetic interference type identification sub-model has its input connected to the output of the convolutional feature extraction module, and is used to output the electromagnetic interference type identification result based on the depth temporal features.

[0049] The output of the signal type identification sub-model is connected to a selector. Based on the received signal type judgment result, the selector selects to route the deep timing features to the device switching speed diagnostic sub-model or the electromagnetic interference type identification sub-model, and outputs the diagnostic results of the corresponding sub-model.

[0050] III. Model Training Module This is used to train a collaborative diagnostic neural network model step-by-step using preprocessed data.

[0051] IV. Diagnostic Execution Module The collaborative diagnostic neural network model is used to input the radiated electromagnetic field signal to be diagnosed into the trained model. First, the signal type identification sub-model determines the signal type. Then, based on the determination result, the device switching speed diagnostic sub-model or the electromagnetic interference type identification sub-model is conditionally called to output the diagnostic result of the switching speed status or electromagnetic interference type.

[0052] In specific implementation, the working principle, control process and technical effects of the timing network collaborative diagnosis system based on IGBT module switching speed and electromagnetic interference provided in this embodiment of the invention are the same as those of the timing network collaborative diagnosis method based on IGBT module switching speed and electromagnetic interference in the above embodiments, and will not be repeated here.

[0053] The above are specific embodiments of the present invention, which demonstrate the substantial features and progress of the present invention. Based on the actual needs of use, equivalent modifications in shape, structure, etc., can be made to it according to the teachings of the present invention, and all such modifications are within the scope of protection of this solution.

Claims

1. A time-series network collaborative diagnostic method based on IGBT module switching speed and electromagnetic interference, characterized in that, Includes the following steps: 1) Using a non-invasive method, collect the radiated electromagnetic field timing signal data generated during the operation of the modular multilevel converter; and label the timing signal data; the labeling includes signal type label, device switching speed status label, and electromagnetic interference type label; 2) Preprocess the labeled time series signal data, including cutting the long time series signal into time series segment samples of fixed length and performing amplitude normalization processing on the samples. 3) Construct a collaborative diagnostic neural network model, which includes a shared convolutional feature extraction module and three sub-models connected thereto: a signal type recognition sub-model, a device switching speed diagnosis sub-model, and an electromagnetic interference type recognition sub-model; 4) The collaborative diagnostic neural network model is trained step-by-step using the preprocessed time-series signal data; 5) Input the radiated electromagnetic field signal to be diagnosed into the trained collaborative diagnostic neural network model. First, the signal type identification sub-model determines the signal type. Then, based on the determination result, the device switching speed diagnostic sub-model or the electromagnetic interference type identification sub-model is conditionally called to output the diagnostic result of the switching speed status or electromagnetic interference type.

2. The timing network collaborative diagnosis method based on IGBT module switching speed and electromagnetic interference according to claim 1, characterized in that: In step 1), the timing signal is obtained through electromagnetic transient simulation or experimental platform, including: establishing a benchmark model to obtain the signal at the nominal switching speed; simulating different levels of slowing switching speed by adjusting the IGBT gate resistance or equivalent circuit parameters; and applying different types of external electromagnetic pulse interference to construct a signal sample set containing concurrent operating conditions.

3. The timing network collaborative diagnosis method based on IGBT module switching speed and electromagnetic interference according to claim 1, characterized in that: The device switching speed status label includes nominal speed, moderate slowdown, and severe slowdown; the electromagnetic interference type label includes high-altitude electromagnetic pulse interference, lightning interference, and electrostatic discharge interference.

4. The timing network collaborative diagnosis method based on IGBT module switching speed and electromagnetic interference according to claim 1, characterized in that: In the preprocessing step 2), a sliding time window is used to cut the continuous radiated electromagnetic field time-domain waveform into time-series segments of equal length, and the amplitude of each segment is normalized.

5. The timing network collaborative diagnosis method based on IGBT module switching speed and electromagnetic interference according to claim 1, characterized in that: In step 3), the core architecture of the convolutional feature extraction module, the signal type identification sub-model, the device switching speed diagnosis sub-model, and the electromagnetic interference type identification sub-model are all one-dimensional convolutional neural networks, and the signal type identification sub-model, the device switching speed diagnosis sub-model, and the electromagnetic interference type identification sub-model share the same depth temporal features extracted by the convolutional feature extraction module during diagnosis.

6. The timing network collaborative diagnosis method based on IGBT module switching speed and electromagnetic interference according to claim 1, characterized in that: In step 4), the collaborative diagnostic neural network model is trained, including: The signal type recognition sub-model is trained using all time-series samples and their signal type labels; The device switching speed diagnostic sub-model is trained using only samples labeled as device switching signals and their switching speed state labels; The electromagnetic interference type identification sub-model is trained using only samples labeled as electromagnetic interference signals and their interference type labels.

7. The timing network collaborative diagnosis method based on IGBT module switching speed and electromagnetic interference according to claim 6, characterized in that: In step 4), the model training process divides the sample data into training and validation sets, uses the Adam optimizer to update parameters, introduces a Dropout layer in the hidden layers of the network, and dynamically adjusts the learning rate based on the validation set loss using a learning rate scheduler to prevent overfitting.

8. The timing network collaborative diagnosis method based on IGBT module switching speed and electromagnetic interference according to claim 1, characterized in that: Step 5) is followed by a post-processing step: setting a confidence threshold for the probability distribution output by the device switching speed diagnostic sub-model or the electromagnetic interference type identification sub-model. Only when the highest category probability exceeds the threshold is the corresponding diagnostic result determined and output. The diagnostic result is used to evaluate the device health status or system electromagnetic compatibility.

9. A timing network collaborative diagnostic system based on IGBT module switching speed and electromagnetic interference, characterized in that, include: The signal acquisition and labeling module is used to acquire the radiated electromagnetic field timing signal generated by the modular multilevel converter in a non-intrusive manner, and to label the timing signal. The labeling includes signal type label, device switching speed status label and electromagnetic interference type label. The signal preprocessing module is used to preprocess the labeled time series signal, including cutting the long time series signal into time series segment samples of fixed length and performing amplitude normalization processing on the samples. The collaborative diagnostic neural network model includes a shared convolutional feature extraction module and three sub-models connected to it: a signal type recognition sub-model, a device switching speed diagnostic sub-model, and an electromagnetic interference type recognition sub-model. The model training module is used to train the collaborative diagnostic neural network model step by step using the preprocessed data; The diagnostic execution module is used to input the radiated electromagnetic field signal to be diagnosed into the trained collaborative diagnostic neural network model. First, the signal type identification sub-model determines the signal type, and then, based on the determination result, conditionally calls the device switching speed diagnostic sub-model or the electromagnetic interference type identification sub-model to output the diagnostic result of the switching speed status or electromagnetic interference type.

10. The timing network collaborative diagnostic system based on IGBT module switching speed and electromagnetic interference according to claim 9, characterized in that: The signal type identification sub-model, the device switching speed diagnosis sub-model, and the electromagnetic interference type identification sub-model share the same depth temporal features extracted by the convolutional feature extraction module during diagnosis. The convolutional feature extraction module is used to extract deep temporal features from the input radiated electromagnetic field time-series signal segment; The signal type identification sub-model has its input connected to the output of the convolutional feature extraction module, and is used to output the signal type judgment result based on the depth temporal features. A device switching speed diagnostic sub-model, whose input is connected to the output of the convolutional feature extraction module, is used to output device switching speed state diagnostic results based on the depth time-series features. An electromagnetic interference type identification sub-model, whose input is connected to the output of the convolutional feature extraction module, is used to output electromagnetic interference type identification results based on the depth temporal features. The output of the signal type identification sub-model is connected to a selector. The selector selects to route the deep timing features to the device switching speed diagnostic sub-model or the electromagnetic interference type identification sub-model based on the received signal type judgment result, and outputs the diagnostic result of the corresponding sub-model.

Images