Pulse width modulation rectifier fault diagnosis method and system
By constructing simulation datasets for various fault types and adaptive neural network models, the problem of identifying multiple fault types in pulse width modulation rectifier fault diagnosis was solved, achieving high-precision and real-time fault diagnosis and improving the reliability and adaptability of diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-03-13
AI Technical Summary
Existing fault diagnosis methods for pulse width modulation rectifiers rely on single feature analysis, which makes it difficult to cover multiple fault types. They are also susceptible to noise and changes in operating conditions, resulting in low diagnostic accuracy and poor real-time performance. Furthermore, the scarcity of fault samples in actual engineering projects limits the generalization ability and reliability of data-driven methods.
A simulation dataset covering multiple fault types is constructed. Through multi-feature fusion and adaptive neural network models, high-precision, real-time fault diagnosis is achieved, including signal acquisition, feature extraction and fusion, intelligent classification and decision output. The adaptive neural network model is used for online real-time diagnosis.
It improves the reliability, adaptability and efficiency of fault diagnosis, can effectively identify a variety of fault types and adapt to different operating conditions, and achieves high-precision online real-time diagnosis.
Smart Images
Figure CN121659152A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power electronics technology, specifically to a method and system for diagnosing faults in pulse width modulation rectifiers. Background Technology
[0002] Pulse width modulation (PWM) rectifiers are widely used in power electronic systems, but their fault diagnosis faces many challenges. Existing methods mostly rely on single feature analysis or simple threshold judgment, which is difficult to cover multiple fault types and is susceptible to noise and changes in operating conditions, resulting in low diagnostic accuracy and poor real-time performance. In addition, fault samples are scarce in actual engineering, which limits the generalization ability and reliability of data-driven methods and cannot meet the needs of efficient diagnosis in complex industrial environments. Summary of the Invention
[0003] The purpose of this invention is to provide a fault diagnosis method for pulse width modulation rectifiers, which can achieve high-precision, real-time fault diagnosis through multi-feature fusion and adaptive neural network model, covering a variety of fault types and adapting to different operating conditions.
[0004] This invention provides a fault diagnosis method for a pulse width modulation rectifier, the method comprising the following steps:
[0005] Step 1: Construct a simulation dataset covering multiple fault types;
[0006] Step 2: Extract and fuse core features and auxiliary features from the acquired three-phase current signals to form an enhanced feature vector;
[0007] Step 3: Train an adaptive neural network classification model using the simulation dataset;
[0008] Step four: Deploy the trained adaptive neural network classification model to the hardware platform to perform online real-time diagnosis and decision-making for the pulse width modulation rectifier.
[0009] The present invention also provides a fault diagnosis system for a pulse width modulation rectifier, the system comprising:
[0010] The signal acquisition unit is used to capture three-phase current signals;
[0011] The feature fusion unit, connected to the signal acquisition unit, is used to extract and fuse features from the three-phase current signal and output an enhanced feature vector.
[0012] An intelligent classification unit is connected to the feature fusion unit. The intelligent classification unit is equipped with an adaptive neural network model for online classification of the enhanced feature vector.
[0013] The decision output unit is connected to the intelligent classification unit and is used to trigger fault code output, alarm signal or protection action based on the classification result.
[0014] This invention addresses the problem of sample scarcity by constructing a simulation dataset, enhances fault discrimination through multi-feature fusion, ensures classification accuracy through a neural network model, and enables online real-time diagnosis through hardware deployment, effectively improving the reliability, adaptability, and efficiency of fault diagnosis. Attached Figure Description
[0015] Figure 1 This is a system architecture diagram for fault diagnosis of the pulse width modulation rectifier described in this invention.
[0016] Figure 2 This is a flowchart of the technical solution described in this invention.
[0017] Figure 3 This is a flowchart of the multi-level confidence determination and fault handling process described in this invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0019] like Figure 1 As shown, the system comprises four parts: a signal acquisition unit, a feature fusion unit, an intelligent classification unit, and a decision output unit. The signal acquisition unit is equipped with a high-precision current sensor and an analog-to-digital converter sampling module to capture three-phase current signals; the feature fusion unit integrates a multi-scale wavelet transform module, a harmonic analysis module, and an amplitude tracking module to complete feature extraction and fusion operations; the intelligent classification unit uses an improved neural network model to perform online classification of the fused features; and the decision output unit triggers fault code output, alarm signals, or protection circuit actions based on the classification results.
[0020] To achieve the above objectives, the present invention adopts the following approach: Figure 2 The technical solution shown has the following specific steps:
[0021] Step 1: Generate a diverse dataset.
[0022] The core of this step is to build a high-fidelity, widely covered fault sample dataset to overcome the problem of scarce fault samples in actual engineering.
[0023] 1. High-precision simulation model construction: Using widely recognized power electronics simulation software, such as Matlab / Simulink combined with PLECS (piecewise linear electrical circuit simulation) or PSIM (a power electronics simulation software), a fully detailed simulation model of the three-phase pulse width modulation rectifier is established. This model must accurately include:
[0024] (1) Main power circuit: grid-side filter inductor, three-phase full-bridge IGBT (insulated gate bipolar transistor) module and its anti-parallel diodes, DC bus capacitor, load or DC grid model.
[0025] (2) Pulse width modulation module: Implements space vector pulse width modulation or carrier comparison modulation strategy.
[0026] (3) Dual closed-loop control algorithm: A current inner loop and a voltage outer loop structure based on PR (proportional resonance) or PI (proportional integral) decoupling control are adopted, and the DC voltage reference value is set to 750V. The control parameters are matched and adjusted according to the actual system characteristics to ensure that the dynamic and static response characteristics are consistent with the actual equipment.
[0027] 2. Fault Injection Mechanism: A flexible fault injection module is implemented in the simulation model, which can accurately simulate a variety of typical faults.
[0028] (1) Device-level faults: such as open circuit of a single or multiple IGBTs, which can be achieved by setting the on-resistance of the corresponding IGBT element to the maximum value or forcing the drive signal to zero.
[0029] (2) Parametric faults: simulate the change of parameters such as the capacitance value decay of the DC bus (linear or nonlinear decay of the nominal value) and the inductance value drift.
[0030] (3) Control-related faults: loss of analog pulse width modulation signal, signal interference or delay caused by abnormal communication of control board.
[0031] (4) Composite faults: Supports the simultaneous injection of two or more different types and severity of faults, such as IGBT open circuit and capacitor decay occurring at the same time.
[0032] 3. Fault mode coverage and operating condition settings:
[0033] (1) Diverse fault types: covering single device faults (such as single IGBT open circuit), multi-device faults (such as phase-to-phase short circuit, bridge arm shoot-through), key parameter faults (such as capacitor attenuation, inductor saturation), control signal faults (such as pulse width modulation missing, control delay) and various composite faults (such as IGBT open circuit, capacitor capacitance decay).
[0034] (2) Duration dynamics: including permanent faults (such as device breakdown), transient faults (such as transient pulse interference at the level of 0.1ms, instantaneous drive failure) and intermittent faults (such as periodic loss of pulse width modulation signal or communication error).
[0035] (3) Operating conditions: Injecting faults under different load levels (light load, rated load, overload) and different power grid conditions (normal voltage, voltage drop / rise, frequency fluctuation).
[0036] 4. Data Collection and Tag Generation:
[0037] (1) Collect key data such as three-phase AC input current, DC bus voltage, grid-side voltage and current.
[0038] (2) Based on the injected fault type, severity, duration and other feature dimensions, a clear and structured fault labeling system is established. The labels adopt one-hot encoding or integer classification encoding, where 0 represents normal and 1~N represent different fault categories and severity levels.
[0039] (3) Add Gaussian noise that conforms to the characteristics of the actual industrial environment to the original signal collected by simulation to enhance the robustness of the model.
[0040] 5. Dataset Generation and Format:
[0041] After the simulation runs, noisy signal data blocks are associated with and stored along with their corresponding fault labels. This results in a dataset file containing training, validation, and test sets for subsequent model training. The dataset generated in this step provides rich, accurate, and quantifiable fault samples, laying a solid foundation for data-driven diagnostic algorithms.
[0042] Step 2: Feature extraction and enhancement.
[0043] This step aims to extract and fuse the core and auxiliary features that are most fault-discriminating and robust from the raw current and voltage signals.
[0044] 1. Core Feature Extraction: The three-phase current signal is decomposed into wavelet packet coefficients using the Daubechies wavelet basis at multiple scales. The wavelet energy entropy of each sub-band is calculated. Finally, a wavelet energy entropy vector containing 16 components is extracted from each current channel as the core feature. This feature effectively characterizes the distribution and disorder of signal energy across frequency bands under different fault conditions. For example, changes in low-frequency energy entropy often correspond to slow dynamic faults such as device failure and switching anomalies.
[0045] 2. Auxiliary Feature Extraction and Optimization:
[0046] Total Harmonic Distortion (THD): The fundamental component amplitude V1 and the amplitudes of the h=2nd to 25th harmonic components Vh are accurately calculated using interpolation-based Fast Fourier Transform. An optimized THD is calculated, which sensitively reflects the degree of waveform distortion caused by switching transistor failures, abnormal pulse width modulation, etc.
[0047] Amplitude offset: Extracts the positive half-cycle peak sequence and negative half-cycle trough sequence within a single complete cycle or power frequency window of the AC current signal. This feature is sensitive to faults such as three-phase imbalance (e.g., single IGBT failure) and sampling bias.
[0048] 3. Feature fusion and adaptive enhancement:
[0049] The wavelet energy entropy vectors of the three-phase currents, the total harmonic distortion rate of the three phases, and the amplitude offset of the three phases are fused to form a preliminary enhanced feature vector. To further improve the representational ability of the features and the classification accuracy of the model, an adaptive feature weight mechanism is introduced. The resulting weighted enhanced feature space is used as the input to the neural network model to enhance the distinguishing ability of fault features.
[0050] Step 3: Construction and optimization of adaptive neural networks.
[0051] This step trains a high-performance, hardware-friendly adaptive neural network classification model based on the dataset generated in step one.
[0052] 1. Neural Network Architecture Design:
[0053] The number of nodes in the input layer is consistent with the dimension of the enhanced feature space generated in step two. The hidden layers employ a structure with skip connections, and the main hidden layer units all use the hyperbolic tangent function. Compared to the ReLU function, the hyperbolic tangent function exhibits richer nonlinear expressive power in the zero-mean region, which helps the model learn complex fault mapping relationships. The number of nodes in the output layer equals the total number of fault categories and normal states, supporting at least four fault modes. A normalized exponential function is used to convert the hidden layer output into a probability distribution of each fault type, with probability values between 0 and 1, and the sum of the probabilities of all categories being 1.
[0054] The core optimizer of the training algorithm adopts the Levenberg-Marquardt (LM) algorithm, which adaptively switches between the Gauss-Newton method and the gradient descent method. It is known for its strong second-order convergence characteristics and is especially suitable for small and medium-sized networks.
[0055] 2. Model training and storage:
[0056] Using the dataset generated in step one as input, divide it into training set, validation set and test set according to the proportions, with the training set accounting for 70%, the validation set accounting for 15% and the test set accounting for 15%.
[0057] The Levenberg-Marquardt algorithm is used for model training. Based on the basic LM update rule, a momentum factor is introduced, where represents the change in the gradient norm during the current iteration. This strategy accelerates convergence in flat regions and reduces the number of iterations required for convergence.
[0058] During training, accuracy and loss metrics are monitored using a validation set, and an early stopping strategy effectively prevents overfitting. Finally, the optimal neural network weight parameters, network layer configuration, and other key model coefficients are exported and stored for subsequent hardware deployment. The model trained in this way combines high accuracy and computational efficiency, laying the foundation for real-time diagnostics.
[0059] Step 4: Pulse Width Modulation Rectifier Fault Diagnosis.
[0060] This step involves deploying the trained neural network model onto an actual hardware platform to enable online real-time diagnosis and decision-making regarding the operating status of the pulse width modulation rectifier.
[0061] 1. System Deployment:
[0062] The model is deployed on a hardware experimental testing platform, relying on a CPU plus FPGA architecture to achieve layered processing, covering multi-scale wavelet transform modules, harmonic analysis modules, amplitude tracking modules, and improved neural network model modules. Model configuration is completed through the CPU, loading the trained neural network coefficients into the hardware.
[0063] 2. Real-time data processing workflow:
[0064] The system acquires three-phase AC current signals in real time via an analog-to-digital converter, with DC voltage and current serving as auxiliary inputs. A sliding time window mode is used to update the input data, with the window length set to cover 1-3 power frequency cycles; 50Hz corresponds to a 20ms-60ms window. Assuming a sampling frequency of 10kHz, the fixed step size is 0.1ms, meaning it slides forward by one sampling cycle, enabling high-frequency diagnostics. Each time the window is updated, data for the new window, including current and historical samples, is extracted from the buffer.
[0065] 3. Embedded Inference and Fault Diagnosis:
[0066] For each new window of data, the hardware platform efficiently executes the feature extraction process in step two, including three-phase wavelet energy entropy vector calculation, accurate calculation of total harmonic distortion rate, and weighted feature fusion to generate an enhanced feature vector. This feature vector is then input into the deployed neural network model for forward inference, obtaining a probability output vector representing normalized exponential functions of various fault and normal states. Here, Pnormal represents the probability of belonging to the i-th class, Pfault represents the probability of the pulse-width modulation rectifier being in a normal state, and Pnormal represents the probability of the pulse-width modulation rectifier exhibiting a certain type of fault.
[0067] Here, P is the probability output vector for all states, summing to 1. It includes normal states and various fault states. Pnormal represents the probability of the pulse width modulation rectifier operating normally, and Pfault represents the probability of the pulse width modulation rectifier experiencing a certain type of fault. For example, P = [0.01, 0.96, 0.01, 0.01, 0.01] indicates a possible fault type 1. Similarly, P = [0.96, 0.01, 0.01, 0.01, 0.01] indicates a normal state.
[0068] 4. Multi-level confidence level determination and fault handling:
[0069] Figure 3 This is a flowchart illustrating the multi-level confidence assessment and fault handling process described in this invention. Multi-level confidence assessment is performed based on the maximum probability value max(P) output by the neural network and its corresponding fault type:
[0070] Class I Fault (Emergency): When max(P)≥0.95, it is judged as a high-confidence emergency fault, such as open circuit of main power device, serious short circuit, failure of multiple IGBT transistors due to open circuit, etc. The protection is immediately triggered, blocking the drive pulse signal of all IGBTs and disconnecting the main contactor to minimize equipment damage and accident expansion.
[0071] Type II Fault (Early Warning): When 0.8 ≤ max(P) < 0.95, a fault is identified as potentially present and an early warning is required. A derating operation mode is then initiated, such as limiting output power. If subsequent verification confirms a fault, the fault is escalated to Type I. If verification fails, the early warning is reset.
[0072] Type III Fault (Low Confidence): When max(P) < 0.8, the confidence level is considered insufficient, and the fault cannot be immediately confirmed. The system records the abnormal event and its diagnostic probability, and continuously monitors it for one or two subsequent power frequency cycles. If the confidence level increases in the subsequent window and reaches the Type II or Type I standard, the corresponding operation is performed; if the low confidence level remains or disappears, only the event log is recorded.
[0073] Fault Handling: Once the diagnosis confirms that a fault of type I or type II has been verified, the system not only outputs the fault type, but also executes the corresponding refined protection or isolation strategy based on the severity level defined during training.
[0074] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A fault diagnosis method for a pulse width modulation rectifier, characterized in that, The method includes the following steps: Step 1: Construct a simulation dataset covering multiple fault types; Step 2: Extract and fuse core features and auxiliary features from the acquired three-phase current signals to form an enhanced feature vector; Step 3: Train an adaptive neural network classification model using the simulation dataset; Step four: Deploy the trained adaptive neural network classification model to the hardware platform to perform online real-time diagnosis and decision-making for the pulse width modulation rectifier.
2. The method according to claim 1, characterized in that, The step one involves constructing a simulation dataset covering multiple fault types, including: Establish a high-precision simulation model of a three-phase pulse width modulation rectifier; Device-level faults, parametric faults, control-related faults, and composite faults are injected into the high-precision simulation model. The high-precision simulation model was run under different load levels and different power grid conditions, and faults were injected. Collect key electrical signal data and generate corresponding fault tags; Noisy signal data blocks are associated with fault labels and stored to generate a dataset file containing training, validation, and test sets.
3. The method according to claim 2, characterized in that, The injected fault types include single-device faults, multi-device faults, critical parameter faults, control signal faults, and various composite faults; the duration of the faults includes permanent faults, transient faults, and intermittent faults.
4. The method according to claim 1, characterized in that, Step two involves extracting and fusing core features and auxiliary features from the original electrical signal, including: Multi-scale wavelet packet decomposition is performed on the three-phase current signal to calculate the wavelet energy entropy of each sub-band and form the wavelet energy entropy vector as the core feature. Calculate the total harmonic distortion rate of the three-phase current signal as the first auxiliary feature; Extract the positive half-cycle peak sequence and negative half-cycle trough sequence of the AC current signal within the power frequency window, and calculate the amplitude offset as a second auxiliary feature. The wavelet energy entropy vector, the total harmonic distortion rate, and the amplitude offset are fused together, and a feature weight adaptive mechanism is introduced to construct a weighted enhanced feature vector.
5. The method according to claim 1, characterized in that, Step three, training the adaptive neural network classification model, includes: Design a neural network structure in which the number of nodes in the input layer is consistent with the dimension of the enhanced feature vector, the hidden layer adopts a structure containing skip connections and uses the hyperbolic tangent function as the activation function, and the output layer uses the normalized exponential function. The Levenberg-Marquardt algorithm was used, and a momentum factor was introduced to train the neural network; During training, accuracy and loss metrics are monitored using a validation set, and an early stopping strategy is used to prevent overfitting. Store the optimal neural network weight parameters and network configuration obtained from training.
6. The method according to claim 1, characterized in that, Step four, which involves deploying the trained adaptive neural network classification model to the hardware platform, includes: Deploy neural network models in a hardware architecture consisting of a central processing unit and a field-programmable gate array; The model configuration is completed through the central processing unit, which loads the trained neural network coefficients into the hardware.
7. The method according to claim 6, characterized in that, Step four, which involves online real-time diagnosis and decision-making for the pulse width modulation rectifier, includes: The three-phase AC current signal is acquired in real time by an analog-to-digital converter, and the input data is updated using a sliding time window mode. For each new window of data, a feature extraction process is executed on the hardware platform to generate the enhanced feature vector; The enhanced feature vector is input into the deployed neural network model for forward inference to obtain the fault probability output vector.
8. The method according to claim 7, characterized in that, Step four, which involves online real-time diagnosis and decision-making for the pulse width modulation rectifier, also includes multi-level confidence assessment and fault handling: The fault level is determined based on the maximum probability value in the fault probability output vector. When the maximum probability value reaches the first threshold, it is determined to be an emergency fault, and the protection action is immediately triggered; When the maximum probability value reaches the second threshold but is lower than the first threshold, it is determined to be a warning fault, the reduced operation mode is activated and subsequent verification is carried out; When the maximum probability value is lower than the second threshold, it is judged as a low-confidence fault, the abnormal event is recorded and continuously monitored.
9. The method according to claim 8, characterized in that, The trigger protection action includes blocking the drive pulse signal of the insulated gate bipolar transistor and disconnecting the main contactor; the start of the derating operation mode includes limiting the output power of the pulse width modulation rectifier.
10. A fault diagnosis system for a pulse width modulation rectifier, characterized in that, The system includes: The signal acquisition unit is used to capture three-phase current signals; The feature fusion unit, connected to the signal acquisition unit, is used to extract and fuse features from the three-phase current signal and output an enhanced feature vector. An intelligent classification unit is connected to the feature fusion unit. The intelligent classification unit is equipped with an adaptive neural network model for online classification of the enhanced feature vector. The decision output unit is connected to the intelligent classification unit and is used to trigger fault code output, alarm signal or protection action based on the classification result.