New energy equipment fault diagnosis system based on deep learning

Abstract

The invention provides a new energy equipment fault diagnosis system based on deep learning. According to the system, waveform data of an electronic power device in the new energy equipment is collected and analyzed, so that real-time detection and classification of equipment faults are realized. Specifically, the system firstly collects voltage and current waveforms of a power device through a sensor, and transmits the voltage and current waveforms as input signals to a convolutional neural network. And the convolutional neural network automatically extracts features in the signals through multi-layer convolution and pooling operation, outputs a classification result through a full connection layer, and identifies possible fault types of the equipment. The system can judge the credibility of the classification result through a preset threshold value, if the result is credible, a fault alarm or suggestion is output, and otherwise, manual inspection is further carried out. The method has the characteristics of high precision, high real-time performance and high automation degree, and can be widely applied to fault diagnosis and predictive maintenance of new energy equipment such as wind power, photovoltaic power generation, electric automobiles and the like.

Description

Technical Field

[0001] The present invention relates to fault diagnosis of electronic equipment, and in particular to a new energy equipment fault diagnosis system based on deep learning. Background Art

[0002] With the widespread application of new energy equipment, such as wind power generation, solar power generation, and electric vehicles, the monitoring and maintenance of their operating status is particularly important. Electronic power devices in new energy equipment, such as inverters and frequency converters, are core components for realizing power conversion and control. However, these power devices are easily affected by environmental factors, aging, and load changes during long-term operation, leading to failures. Once a failure occurs, it will not only reduce the efficiency of the system, but may even cause equipment downtime, resulting in serious economic losses.

[0003] Traditional fault diagnosis methods rely on experience and regular maintenance, which is not only time-consuming and labor-intensive, but also difficult to detect potential faults in a timely manner, such as: CN114035086A. In addition, as the complexity of equipment increases, the diversity of signal characteristics and noise interference make fault diagnosis more difficult. Summary of the invention

[0004] The present invention aims to at least solve the technical problems existing in the prior art, and in particular innovatively proposes a new energy equipment fault diagnosis system based on deep learning.

[0005] According to the needs of detecting new energy equipment, select a data acquisition card DAQ card with sufficient sampling rate and accuracy. Set the sampling rate, input range and channel configuration of the DAQ card, connect the current sensor to the output end of the power device of the wind power generation equipment, run the acquisition program, and start continuous acquisition of current waveform data. Ensure that the sampling clock of the DAQ card is strictly synchronized with the system timing t, and perform smoothing and filtering operations on the collected current waveform data:

[0006] The current conversion formula is:

[0007] Sampling point calculation: Where m is the sampling point number of the wind power equipment.

[0008] Among them, I in For current input, DAQ gain The gain setting for the data acquisition card (DAQ). The gain is the multiple by which the data acquisition card amplifies the input signal. By adjusting the gain, ensure that the input signal is within the measurement range of the DAQ card and has sufficient resolution. rangeThe input range of the data acquisition card indicates the maximum and minimum signal values ​​that the DAQ card can measure. The input range is usually expressed in voltage units, such as ±10V.

[0009] q s The sampling frequency represents the number of samples collected by the data acquisition card per second. The sampling frequency is an important parameter for data acquisition, which determines the highest frequency of the signal that can be captured. According to the Nyquist theorem, the sampling frequency should be at least twice the highest frequency of the measured signal to avoid aliasing.

[0010] Acquiring current data is more suitable for analyzing the circuit status, load changes, and fault diagnosis of new energy equipment. It provides detailed information on the operation of the generator and helps to deeply understand the behavior of new energy equipment.

[0011] 1. Data Collection and Preparation

[0012] 1.1 Data Collection

[0013] Waveform data acquisition: Collect the output voltage and current waveform data of the power device under normal working conditions and different fault conditions. Use a high-precision oscilloscope or data acquisition card to ensure that the sampling rate is high enough to capture subtle changes in the waveform.

[0014] The original waveform data x(t) of the power device is obtained, and then the data is processed by time-frequency analysis methods to identify potential electrical faults, such as transistor short circuit or component aging. This analysis helps predictive maintenance and prevent more serious failures.

[0015] 1.2 Data Preprocessing

[0016] Signal denoising: Apply filters or wavelet transforms to remove noise from waveforms to ensure data quality for input into deep learning models.

[0017] Wavelet transform can process the local characteristics of the signal very well, while Fourier transform has greater advantages in processing the global frequency characteristics. Hybrid wavelet-Fourier transform attempts to combine the advantages of these two transforms, which can not only analyze the local characteristics of the signal, but also process the global frequency information of the signal.

[0018]

[0019] Among them, F(a,b,ω) is the coefficient after hybrid wavelet-Fourier transform, which represents the time-frequency component of the signal at a given scale a, translation b and frequency ω;

[0020] x(t) is the waveform data of the power device, and the time variable is t;

[0021] a is the scale parameter, which controls the expansion and compression of the wavelet function, b is the translation parameter, which controls the position of the wavelet function on the time axis, and ω is the frequency parameter of the Fourier transform, which represents the frequency component of the signal;

[0022] ψ(t) is the wavelet basis function, which is used to extract the local features of the signal. -jωt It is the kernel function of Fourier transform, which is used to extract the frequency characteristics of the signal;

[0023] 2. Model selection and design

[0024] 2.1 Model selection

[0025] Use Convolutional Neural Network (CNN): CNN is good at processing two-dimensional data and extracting local features for analyzing the time-amplitude two-dimensional representation of the waveform.

[0026] 2.2 Model design

[0027] Assume that the length of the waveform data x(t) of the power device is N, sample the continuous signal x(t) and obtain the discrete signal vector Facilitates subsequent digital processing;

[0028]

[0029] First, the signal is feature extracted through several convolutional layers. Let the convolution kernel of the first layer be W (l) , and its convolution operation is

[0030] h (l) =f(W (l) *h( l-1 )+b (l) )

[0031] in,

[0032] h (l-1) is the output feature of the l-1th layer;

[0033] W (l) is the convolution kernel and weight matrix of the lth layer;

[0034] b (l) is the bias vector of the lth layer;

[0035] f(*) is the activation function;

[0036] After several layers of convolution, we get a high-level feature representation h (L) ;

[0037] h (L) =f(W (L) *h (L-1) +b (L) )

[0038] The high-level features h (L) Mapped to a space representing the wavelet basis function, the generated wavelet basis function is:

[0039]

[0040] Among them, h (L) and h (L-1) They are the Lth convolutional layer and the L-1th convolutional layer, W out and b out is the weight matrix and bias vector of the fully connected layer;

[0041] Fully connected layer: The extracted features are mapped to the fault type through the fully connected layer, and the final fault diagnosis result is output.

[0042]

[0043] in, is the classification probability vector, belonging to the fully connected layer of the convolutional neural network;

[0044] The deep learning model generates a probability distribution by analyzing the input waveform x(t) of the power device. Each element p in this probability distribution i Represents a certain fault type C i The probability that all p i The sum is 1, that is:

[0045]

[0046] Where n is the number of all fault types;

[0047] W P is the weight matrix of the fault classifier, b P is the bias vector of the fault classifier, and the softmax function converts the network output into a probability distribution;

[0048] Use gradient descent to update model parameters;

[0049] Repeat the above steps until the model converges.

[0050] 3. Fault diagnosis module

[0051] 3.1 Fault Detection and Diagnosis

[0052] The diagnostic results generate a diagnostic report based on the signal analysis results, which includes:

[0053] C = argmaxp i

[0054] Among them, C is the final fault type, argmaxp i Indicates that p i The fault type corresponding to the largest i.

[0055] 3.2 Credibility Judgment

[0056] In order to ensure the credibility of the fault classification results, a threshold p is set max ,If the maximum probability exceeds this threshold, the fault diagnosis result is considered credible and is used directly;

[0057] If p max >τ, then the fault type C is considered to be credible;

[0058] If p max <τ indicates that the credibility of the fault is insufficient and further analysis is required, and only a warning is sent to prompt manual inspection;

[0059] According to the characteristics of new energy equipment, a specific threshold τ is set for a specific fault type to measure the credibility of fault identification.

[0060] 3.3 Fault location: By analyzing the time domain and frequency domain characteristics of the fault waveform, the specific location of the fault in the equipment can be located according to the fault type. For example, the short circuit point or overload source can be located by combining spectrum analysis and time domain characteristics.

[0061] 3.4 Fault severity assessment: Based on waveform characteristics such as amplitude, frequency content, and harmonic content, the severity of the fault is assessed and repair suggestions are given. For example, a slight voltage fluctuation may indicate an impending fault, while severe harmonic distortion may mean that the equipment is about to fail.

[0062] Historical data comparison: Compare the currently identified waveform with historical data to determine the duration and evolution trend of the fault, so as to provide a more accurate maintenance time window and strategy.

[0063] 4. Functional Experiment

[0064] The inverter in an electric vehicle is responsible for converting the DC power provided by the battery into three-phase AC power to drive the motor. It is one of the key components in the electric vehicle power system. During the long-term operation of the inverter, a variety of faults may occur, such as open circuit faults, short circuit faults, and overload faults of IGBT (insulated gate bipolar transistor). In order to ensure the safe operation of electric vehicles, it is crucial to detect and classify these faults in a timely manner.

[0065] This experiment uses a hybrid wavelet-Fourier transform combined with deep learning to perform real-time diagnosis and classification of electric vehicle inverter faults. By collecting the waveform data of the inverter output current, possible fault types are detected and classified, and the accuracy and reliability of the system in actual use scenarios are verified.

[0066] 4.1 Experimental data collection

[0067] The experiment was conducted in a simulated electric vehicle drive system, which includes a three-phase inverter. The experimental data collected is the output current signal of the inverter, with a sampling frequency of 10kHz, and the signal is collected for 1 second each time, resulting in 10,000 sample points.

[0068] The collected data include:

[0069] Current waveform in normal working state;

[0070] IGBT open circuit fault waveform;

[0071] IGBT short circuit fault waveform;

[0072] Load overload fault waveform;

[0073] The experiment collected waveform data under each state, 100 sets of data each time, for a total of 400 sets of experimental data.

[0074] 4.2 Experimental signal waveform

[0075] Taking the IGBT short circuit fault as an example, the following is the collected inverter output current signal:

[0076]

[0077] Under the short-circuit fault condition, a sudden increase in the current waveform can be clearly observed, especially the high amplitude and rapidly changing frequency characteristics of the waveform.

[0078] 4.3 Classification of Deep Learning Models

[0079] The extracted time-frequency features are input into the pre-trained convolutional neural network (CNN) model for fault classification. The model structure includes 3 convolutional layers and 2 fully connected layers, and finally the probability distribution of the fault is output through the softmax layer.

[0080] Model classification results:

[0081] Normal state: 99.5%;

[0082] IGBT open circuit failure: 97.8%;

[0083] IGBT short circuit failure: 98.9%;

[0084] Load overload failure: 96.7%;

[0085] 4.4 Experimental Results and Analysis

[0086] Through the analysis of experimental data and the training of deep learning models, the final fault classification accuracy is 98.2%, which shows the effectiveness of this method in electric vehicle inverter fault diagnosis.

[0087] Specific experimental results:

[0088]

[0089] 4.5 Fault Location and Severity Assessment

[0090] In this experiment, the hybrid wavelet-Fourier transform not only helps to classify the fault type, but also shows the specific time when the fault occurs through the time-frequency diagram. For the short-circuit fault, its high-frequency component reaches a peak between 0.45 seconds and 0.47 seconds, confirming the specific time point of the fault. At the same time, the severity of the fault is evaluated by combining the change in waveform amplitude and frequency component.

[0091] Failure time: 0.45 seconds to 0.47 seconds

[0092] Fault severity: The high-amplitude waveform and significant high-frequency component indicate the severity of the IGBT short-circuit fault, which requires immediate shutdown and inspection.

[0093] 4.6 Experimental Summary

[0094] Through the implementation of this experiment, the hybrid wavelet-Fourier transform combined with deep learning technology can effectively detect various fault types in electric vehicle inverters and accurately locate the fault location. This method provides higher time-frequency analysis capabilities than traditional Fourier transform and wavelet transform, especially in capturing transient characteristics of fault signals. This provides strong support for the health management and predictive maintenance of electric vehicle power systems.

[0095] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0096] First, by combining convolutional neural networks (CNN) and hybrid wavelet-Fourier transform technology, the present invention greatly improves the real-time and accuracy of fault detection. Unlike traditional fault detection methods that rely on preset thresholds or fixed rules, the present invention can perform multi-scale, multi-frequency real-time analysis of waveform data of equipment power devices. This efficient time-frequency domain processing method enables the system to capture subtle abnormal changes in the waveform in a timely manner when a fault occurs, thereby achieving rapid response. Especially when sudden abnormalities occur during the operation of new energy equipment, the system can accurately identify the abnormal components in the signal, avoiding equipment damage or fault extension caused by delayed response or misjudgment of traditional methods.

[0097] Secondly, the present invention can efficiently classify a variety of complex fault modes through the automated learning capabilities of the deep learning model. The convolutional neural network automatically extracts and learns the waveform characteristics of different fault types by training a large amount of fault data. Therefore, the system can identify nonlinear and complex fault conditions, including short circuits, overloads, and device aging. In addition, with the continuous accumulation of data and optimization of the model, the system has adaptive characteristics. Even if new fault types appear, the system can achieve accurate classification and diagnosis of unknown faults through continuous learning, thereby greatly improving the comprehensiveness and accuracy of fault identification.

[0098] In terms of fault location, the present invention uses the time-frequency localization characteristics of the hybrid wavelet-Fourier transform to achieve accurate positioning of the fault location. By analyzing the output signal of the equipment at different frequencies and time periods, the system can identify the exact time of occurrence of the fault and the specific equipment components, thereby providing clear and accurate fault location information for equipment maintenance personnel. This not only significantly reduces the troubleshooting time, but also reduces the unnecessary maintenance costs caused by inaccurate positioning, and improves the efficiency of equipment maintenance.

[0099] Finally, through automated fault diagnosis, the present invention significantly reduces maintenance costs. The system can predict potential problems before a fault occurs and provide repair suggestions, thus avoiding unplanned downtime of the equipment. At the same time, accurate fault classification and location functions also greatly reduce the time and labor costs required for blind maintenance. This maintenance method based on intelligent analysis not only improves the operational efficiency of the equipment, but also significantly reduces operation and maintenance expenses.

[0100] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0102] Figure 1 It is a system schematic diagram of the present invention.

[0103] Figure 2 It is an analytical comparison chart of Fourier transform and hybrid wavelet-Fourier transform.

[0104] Figure 3 Fault diagnosis diagram. DETAILED DESCRIPTION

[0105] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0106] like Figure 1 As shown, the present invention discloses a new energy equipment fault diagnosis system based on deep learning, including the following contents:

[0107] According to the needs of detecting new energy equipment, select a data acquisition card DAQ card with sufficient sampling rate and accuracy. Set the sampling rate, input range and channel configuration of the DAQ card, connect the current sensor to the power device output of the wind power generation equipment, run the acquisition program, start continuous acquisition of current waveform data, ensure that the sampling clock of the DAQ card is strictly synchronized with the system timing t, and perform smoothing and filtering operations on the acquired current waveform data.

[0108] The current conversion formula is:

[0109] Sampling point calculation: Where m is the sampling point number of the wind power equipment.

[0110] Among them, I in For current input, DAQ gain The gain setting for the data acquisition card (DAQ). The gain is the multiple by which the data acquisition card amplifies the input signal. By adjusting the gain, ensure that the input signal is within the measurement range of the DAQ card and has sufficient resolution. range The input range of the data acquisition card indicates the maximum and minimum signal values ​​that the DAQ card can measure. The input range is usually expressed in voltage units, such as ±10V.

[0111] q s The sampling frequency represents the number of samples collected by the data acquisition card per second. The sampling frequency is an important parameter for data acquisition, which determines the highest frequency of the signal that can be captured. According to the Nyquist theorem, the sampling frequency should be at least twice the highest frequency of the measured signal to avoid aliasing.

[0112] Acquiring current data is more suitable for analyzing the circuit status, load changes, and fault diagnosis of new energy equipment. It provides detailed information on the operation of the generator and helps to deeply understand the behavior of new energy equipment.

[0113] 1. Data Collection and Preparation

[0114] 1.1 Data Collection

[0115] Waveform data acquisition: Collect the output voltage and current waveform data of the power device under normal working conditions and different fault conditions. Use a high-precision oscilloscope or data acquisition card to ensure that the sampling rate is high enough to capture subtle changes in the waveform.

[0116] The original waveform data x(t) of the power device is obtained, and then the data is processed by time-frequency analysis methods to identify potential electrical faults, such as transistor short circuit or component aging. This analysis helps predictive maintenance and prevent more serious failures.

[0117] 1.2 Data Preprocessing

[0118] Signal denoising: Apply filters or wavelet transforms to remove noise from waveforms to ensure data quality for input into deep learning models.

[0119] Wavelet transform can process the local characteristics of the signal very well, while Fourier transform has greater advantages in processing the global frequency characteristics. Hybrid wavelet-Fourier transform attempts to combine the advantages of these two transforms, which can not only analyze the local characteristics of the signal, but also process the global frequency information of the signal.

[0120]

[0121] Among them, F(a,b,ω) is the coefficient after hybrid wavelet-Fourier transform, which represents the time-frequency component of the signal at a given scale a, translation b and frequency ω;

[0122] x(t) is the waveform data of the power device, and the time variable is t;

[0123] a is the scale parameter, which controls the expansion and compression of the wavelet function, b is the translation parameter, which controls the position of the wavelet function on the time axis, and ω is the frequency parameter of the Fourier transform, which represents the frequency component of the signal;

[0124] ψ(t) is the wavelet basis function, which is used to extract the local features of the signal. -jωt It is the kernel function of Fourier transform, which is used to extract the frequency characteristics of the signal;

[0125] like Figure 2As shown in the figure, the second figure is the analysis diagram of the signal by hybrid wavelet-Fourier transform, and the third figure is the analysis diagram of the signal by Fourier transform. Although Fourier transform can capture the frequency components of the signal well, it cannot reflect the changes of these frequency components over time. The hybrid wavelet-Fourier transform method shows the time-frequency characteristics of the signal. In the figure, the color gradually changes from blue to red, red represents the part with stronger energy, and blue represents the part with weaker energy. The advantage of this method is that it can capture the local time information and global frequency characteristics of the signal at the same time, which is especially suitable for processing non-stationary signals (such as power signals in fault detection).

[0126] 2. Model selection and design

[0127] 2.1 Model selection

[0128] Use Convolutional Neural Network (CNN): CNN is good at processing two-dimensional data and extracting local features for analyzing the time-amplitude two-dimensional representation of the waveform.

[0129] 2.2 Model design

[0130] Assume that the length of the waveform data x(t) of the power device is N, sample the continuous signal x(t) and obtain the discrete signal vector Facilitates subsequent digital processing;

[0131]

[0132] First, the signal is feature extracted through several convolutional layers. Let the convolution kernel of the first layer be W (l) , and its convolution operation is

[0133] h (l) =f(W (l) *h (l-1) +b (l) )

[0134] in,

[0135] h (l-1) is the output feature of the l-1th layer;

[0136] W (l) is the convolution kernel of the lth layer, and the weight matrix

[0137] b (l) is the bias vector of the lth layer;

[0138] f(*) is the activation function

[0139] After several layers of convolution, we get a high-level feature representation h (L)

[0140] h (L) =f(W(L) *h (L-1) +b (L) )

[0141] The high-level features h (L) Mapped to a space representing the wavelet basis function, the generated wavelet basis function is:

[0142]

[0143] Among them, h (L) and h (L-1) They are the Lth convolutional layer and the L-1th convolutional layer, W out and b out is the weight matrix and bias vector of the fully connected layer;

[0144] Fully connected layer: The extracted features are mapped to the fault type through the fully connected layer, and the final fault diagnosis result is output.

[0145]

[0146] in, is the classification probability vector, belonging to the fully connected layer of the convolutional neural network;

[0147] The deep learning model generates a probability distribution by analyzing the input waveform x(t) of the power device. Each element p in this probability distribution i Represents a certain fault type C i The probability that all p i The sum is 1, that is:

[0148]

[0149] Where n is the number of all fault types;

[0150] W P is the weight matrix of the fault classifier, b P is the bias vector of the fault classifier, and the softmax function converts the network output into a probability distribution;

[0151] Use gradient descent to update model parameters;

[0152] Repeat the above steps until the model converges.

[0153] 3. Fault diagnosis module

[0154] 3.1 Fault Detection and Diagnosis

[0155] like Figure 3 As shown, the diagnostic results generate a diagnostic report based on the signal analysis results, which includes:

[0156] C = argmaxp i

[0157] Among them, C is the final fault type, argmaxp i Indicates that p i The fault type corresponding to the largest i.

[0158] 3.2 Credibility Judgment

[0159] In order to ensure the credibility of the fault classification results, a threshold p is set max ,If the maximum probability exceeds this threshold, the fault diagnosis result is considered credible and is used directly;

[0160] If p max >τ, then the fault type C is considered to be credible;

[0161] If p max <τ indicates that the credibility of the fault is insufficient and further analysis is required, and only a warning is sent to prompt manual inspection;

[0162] According to the characteristics of new energy equipment, a specific threshold τ is set for a specific fault type to measure the credibility of fault identification.

[0163] 3.3 Fault location: By analyzing the time domain and frequency domain characteristics of the fault waveform, the specific location of the fault in the equipment can be located according to the fault type. For example, the short circuit point or overload source can be located by combining spectrum analysis and time domain characteristics.

[0164] 3.4 Fault severity assessment: Based on waveform characteristics such as amplitude, frequency content, and harmonic content, the severity of the fault is assessed and repair suggestions are given. For example, a slight voltage fluctuation may indicate an impending fault, while severe harmonic distortion may mean that the equipment is about to fail.

[0165] Historical data comparison: Compare the currently identified waveform with historical data to determine the duration and evolution trend of the fault, so as to provide a more accurate maintenance time window and strategy.

[0166] 4. Functional Experiment

[0167] The inverter in an electric vehicle is responsible for converting the DC power provided by the battery into three-phase AC power to drive the motor. It is one of the key components in the electric vehicle power system. During the long-term operation of the inverter, a variety of faults may occur, such as open circuit faults, short circuit faults, and overload faults of IGBT (insulated gate bipolar transistor). In order to ensure the safe operation of electric vehicles, it is crucial to detect and classify these faults in a timely manner.

[0168] This experiment uses a hybrid wavelet-Fourier transform combined with deep learning to perform real-time diagnosis and classification of electric vehicle inverter faults. By collecting the waveform data of the inverter output current, possible fault types are detected and classified, and the accuracy and reliability of the system in actual use scenarios are verified.

[0169] 4.1 Experimental data collection

[0170] The experiment was conducted in a simulated electric vehicle drive system, which includes a three-phase inverter. The experimental data collected is the output current signal of the inverter, with a sampling frequency of 10kHz, and the signal is collected for 1 second each time, resulting in 10,000 sample points.

[0171] The collected data include:

[0172] Current waveform under normal working condition

[0173] IGBT open circuit fault waveform

[0174] IGBT short circuit fault waveform

[0175] Load overload fault waveform

[0176] The experiment collected waveform data under each state, 100 sets of data each time, for a total of 400 sets of experimental data.

[0177] 4.2 Experimental signal waveform

[0178] Taking the IGBT short circuit fault as an example, the following is the collected inverter output current signal:

[0179]

[0180] Under the short-circuit fault condition, a sudden increase in the current waveform can be clearly observed, especially the high amplitude and rapidly changing frequency characteristics of the waveform.

[0181] 4.3 Classification of Deep Learning Models

[0182] The extracted time-frequency features are input into the pre-trained convolutional neural network (CNN) model for fault classification. The model structure includes 3 convolutional layers and 2 fully connected layers, and finally the probability distribution of the fault is output through the softmax layer.

[0183] Model classification results:

[0184] Normal state: 99.5%;

[0185] IGBT open circuit failure: 97.8%;

[0186] IGBT short circuit failure: 98.9%;

[0187] Load overload failure: 96.7%;

[0188] 4.4 Experimental Results and Analysis

[0189] Through the analysis of experimental data and the training of deep learning models, the final fault classification accuracy is 98.2%, which shows the effectiveness of this method in electric vehicle inverter fault diagnosis.

[0190] Specific experimental results:

[0191]

[0192] 4.5 Fault Location and Severity Assessment

[0193] In this experiment, the hybrid wavelet-Fourier transform not only helps to classify the fault type, but also shows the specific time when the fault occurs through the time-frequency diagram. For the short-circuit fault, its high-frequency component reaches a peak between 0.45 seconds and 0.47 seconds, confirming the specific time point of the fault. At the same time, the severity of the fault is evaluated by combining the change in waveform amplitude and frequency component.

[0194] Failure time: 0.45 seconds to 0.47 seconds

[0195] Fault severity: The high-amplitude waveform and significant high-frequency component indicate the severity of the IGBT short-circuit fault, which requires immediate shutdown and inspection.

[0196] 4.6 Experimental Summary

[0197] Through the implementation of this experiment, the hybrid wavelet-Fourier transform combined with deep learning technology can effectively detect various fault types in electric vehicle inverters and accurately locate the fault location. This method provides higher time-frequency analysis capabilities than traditional Fourier transform and wavelet transform, especially in capturing transient characteristics of fault signals. This provides strong support for the health management and predictive maintenance of electric vehicle power systems.

[0198] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A new energy equipment fault diagnosis system based on deep learning, characterized in that: include: Signal processing module: collects waveform data of power devices, processes signal noise, and performs time domain and frequency analysis on input waveforms; Deep learning module: Apply convolutional neural network to the waveform data of power devices for training and fault analysis; Fault diagnosis module: Apply the results of deep learning to fault diagnosis and locate the fault type and severity through probability.

2. A new energy equipment fault diagnosis system based on deep learning according to claim 1, characterized in that: The signal processing module comprises: Obtain power device waveform data x(t); The current conversion formula is: Sampling point calculation: Among them, I in is the original current signal measured by the acquisition device, I out It is the current signal processed by the data acquisition card. gain The gain setting of the data acquisition card is range is the input range of the data acquisition card; m is the sampling point number of the wind power generation equipment, q s is the sampling frequency; Apply filters to remove noise from waveforms to ensure data quality for input into deep learning models; Use hybrid wavelet-Fourier transform to analyze the local characteristics of the signal and process the global frequency information of the signal; Among them, F(a,b,ω) is the coefficient after hybrid wavelet-Fourier transform, which represents the time-frequency component of the signal at a given scale a, translation b and frequency ω; x(t) is the waveform data of the power device, and the time variable is t; a is the scale parameter, which controls the expansion and compression of the wavelet function, b is the translation parameter, which controls the position of the wavelet function on the time axis, and ω is the frequency parameter of the Fourier transform, which represents the frequency component of the signal; ψ(t) is the wavelet basis function, which is used to extract the local features of the signal. -jωt It is the kernel function of Fourier transform, which is used to extract the frequency characteristics of the signal.

3. A new energy equipment fault diagnosis system based on deep learning according to claim 1, characterized in that: The deep learning module includes: Assume that the length of the waveform data x(t) of the power device is N, sample the continuous signal x(t) and obtain the discrete signal vector Facilitates subsequent digital processing; First, the signal is feature extracted through several convolutional layers. Let the convolution kernel of the lth layer be W (l) , and its convolution operation is: h (l) =f(W (l) *h( l-1) +b (l) ) in, h (l-1) is the output feature of the l-1th layer, W (l) is the convolution kernel of the lth layer, the weight matrix, b (l) is the bias vector of the lth layer, f(*) is the activation function; After several layers of convolution, we get a high-level feature representation h (L) : h (L) =f(W (L) *h (L-1) +b (L) ) The high-level features h (L) Mapped to a space representing the wavelet basis function, the generated wavelet basis function ψ(t) is: Among them, h (L) and h (L-1) They are the Lth convolutional layer and the L-1th convolutional layer, W out and b out is the weight matrix and bias vector of the fully connected layer; Fully connected layer: The extracted features are mapped to the fault type through the fully connected layer, and the final fault diagnosis result is output; in, is the classification probability vector, belonging to the fully connected layer of the convolutional neural network; The deep learning model generates a probability distribution by analyzing the input waveform x(t) of the power device. Each element p in this probability distribution i Represents a certain fault type C i The probability that all p i The sum is 1, that is: Where n is the number of all possible fault types; W P is the weight matrix of the fault classifier and b P is the bias vector of the fault classifier, and the softmax function converts the network output into a probability distribution; Use gradient descent to update model parameters; Repeat the above steps until the model converges.

4. A new energy equipment fault diagnosis system based on deep learning according to claim 1, characterized in that: The fault diagnosis module comprises: The diagnostic results generate a diagnostic report based on the signal analysis results, which includes: C=argmaxp i Among them, C is the final fault type, argmaxp i Indicates that p i The fault type corresponding to the largest i; In order to ensure the credibility of the fault classification results, a threshold p is set max ,If the maximum probability exceeds this threshold, the fault diagnosis result is considered credible and can be used directly; If p max >τ, the fault type C is considered to be credible; If p max <τ indicates that the credibility of the fault is insufficient and further analysis is required, and only a warning is sent to prompt manual inspection; According to the characteristics of new energy equipment, a specific threshold τ is set for a specific fault type to measure the credibility of fault identification.

Images