Inverter fault feature intelligent diagnosis and prediction method and system thereof

Abstract

The invention discloses an inverter fault feature intelligent diagnosis and prediction method and system, and the method comprises the following steps: collecting inverter operation data from a plurality of monitoring points, carrying out the preprocessing of a data vector, extracting features, combining the extracted features into a feature vector, constructing and training an intelligent diagnosis model, and recognizing an inverter fault mode. And correcting a diagnosis result by combining with an expert system rule, outputting a fault diagnosis result vector and a credibility value, predicting a future fault development trend, calculating a prediction fault probability and setting an early warning threshold. Therefore, various technical advantages can be fused, multi-monitoring-point data are collected, preprocessed and subjected to feature extraction, an intelligent diagnosis model is input, a result is corrected in combination with expert system rules, diagnosis accuracy and reliability are improved, the fault trend can be predicted, early warning can be set, operation and maintenance personnel can deal with and adapt to various fault types in time, and the fault diagnosis accuracy and reliability are improved. The overall intelligent level is enhanced, and the fault diagnosis prediction process is simply and efficiently realized.

Description

Technical Field

[0001] The present invention relates to the field of industrial informatization, and in particular to an intelligent diagnosis and prediction method for inverter fault characteristics and a system thereof. Background Art

[0002] With the widespread application of renewable energy around the world, photovoltaic power generation, as an important way to utilize clean energy, has continued to expand in scale. As a key device in the photovoltaic power generation system, the cascade H-bridge multilevel inverter is responsible for converting direct current into alternating current. Its operating status directly affects the efficiency and stability of the entire photovoltaic power generation system. However, during the long-term operation of the cascade H-bridge multilevel inverter, various faults are prone to occur due to the influence of environmental factors (such as temperature, humidity, dust, etc.), electrical stress (such as overvoltage, overcurrent, voltage fluctuation, etc.) and the aging of its own components. These faults will not only cause a decrease in the power generation of the photovoltaic power generation system, but may even cause the entire system to shut down, causing huge economic losses to photovoltaic power generation companies.

[0003] Traditional fault diagnosis methods for cascaded H-bridge multilevel inverters mainly include methods based on analytical models, methods based on signal processing, and methods based on expert systems. The analytical model-based method requires the establishment of an accurate mathematical model of the cascaded H-bridge multilevel inverter, but due to its complex internal structure and the changing actual operating environment, it is difficult to establish an accurate model, resulting in low diagnostic accuracy. The signal processing-based method mainly extracts fault features by analyzing the electrical signals (such as current, voltage, etc.) of the cascaded H-bridge multilevel inverter during operation, but the diagnostic ability for complex faults is limited and the requirements for signal processing algorithms are high. The expert system-based method relies on a large amount of expert knowledge and empirical rules, has low diagnostic efficiency, and is difficult to adapt to new fault types. Therefore, it is of great practical significance to study a cascaded H-bridge multilevel inverter fault feature intelligent diagnosis and prediction method and system that integrates multiple advanced technologies. Summary of the invention

[0004] The present invention aims to solve the technical problems in the above-mentioned technology at least to some extent.

[0005] To this end, the first aspect of the present invention discloses an inverter fault feature intelligent diagnosis and prediction method, comprising the following steps:

[0006] S1. Collect inverter operation data from multiple monitoring points. Each monitoring point corresponds to one or more sensors. The data vector collected at a specific time is expressed as X(t)=[x1(t),x2(t),…,x n (t)], where x i (t) represents the data collected by the i-th sensor at time t;

[0007] S2. For the data vector X(t) = [x1(t), x2(t), ..., x n (t)] to perform noise reduction and normalization preprocessing to obtain the preprocessed data vector X'(t) = [x'1(t), x'2(t), ..., x' n (t)];

[0008] S3. Extract the preprocessed data vector X'(t) = [x'1(t), x'2(t), ..., x' n (t)], the features include but are not limited to time domain features, frequency domain features and time-frequency domain features, wherein,

[0009] The time domain features include but are not limited to the mean μ i ,variance and kurtosis K i ;

[0010] The frequency domain feature obtains the spectrum amplitude distribution by performing fast Fourier transform on the time domain data, and analyzes whether there are abnormal harmonics;

[0011] The time-frequency domain features are decomposed by multi-layer wavelet packets to obtain wavelet packet coefficient matrices of different frequency bands and the energy changes of each corresponding sub-frequency band;

[0012] S4. Combining the extracted time domain features, frequency domain features and time-frequency domain features into a feature vector F(t)=[F1(t), F2(t),…, F m (t)] and used as the input of the intelligent diagnosis model, where F(t) = [F1(t), F2(t), …, F m (t)] includes all of the time domain features, the frequency domain features and the time-frequency domain features;

[0013] S5. Build and train an intelligent diagnosis model M that integrates convolutional neural networks and long short-term memory networks CNN-LSTM , used to analyze the eigenvector F(t)=[F1(t),F2(t),…,F m (t)] and identify the inverter failure mode, where

[0014] The fused convolutional neural network is used to gradually extract deep spatial features, and outputs a feature map Z(t) after activation function and pooling layer operations;

[0015] The long short-term memory network is used to receive the feature map Z(t) output by the fused convolutional neural network and expand it in time steps to learn the evolution law of fault features over time;

[0016] S6. Combined with the expert system rule R = {r1, r2, ..., r r}Correct the diagnosis result and output the fault diagnosis result vector Y DI (t)=[y1(t),y2(t),…,y q (t)] and the credibility value p DI (t), where q is the number of fault categories, y i (t)∈{0,1}, when y i (t)=0 means there is no fault. i (t) = 1 indicates the existence of the i-th type of fault;

[0017] S7. Use the autoregressive integrated moving average model to predict future fault development trends based on historical diagnostic data sequences and calculate the predicted fault probability And set warning thresholds When the failure probability When a fault occurs, the early warning mechanism is triggered to notify the operation and maintenance personnel to take measures.

[0018] The inverter fault feature intelligent diagnosis and prediction method disclosed in the present invention can effectively integrate a variety of advanced technologies, collect inverter operation data from multiple monitoring points, perform noise reduction, normalization preprocessing, and time domain, frequency domain, and time-frequency domain feature extraction, and then combine the features into vector inputs into an intelligent diagnosis model that integrates a convolutional neural network and a long short-term memory network. The diagnosis results are corrected in combination with expert system rules, thereby improving the accuracy and reliability of the diagnosis. At the same time, an autoregressive integral sliding average model is used to predict future fault development trends, set warning thresholds, and achieve early warning of faults, which helps operation and maintenance personnel to take timely measures to reduce problems such as power generation reduction and shutdown of photovoltaic power generation systems caused by inverter failures, reduce the economic losses of photovoltaic power generation companies, adapt to different fault types, and improve the overall intelligence level of fault diagnosis and prediction.

[0019] In addition, the inverter fault feature intelligent diagnosis and prediction method disclosed in the present invention may also have the following additional technical features:

[0020] In one embodiment of the present invention, in step S3, the mean μ i The calculation function is The variance The calculation function is The kurtosis K i The calculation function is

[0021] In one embodiment of the present invention, in step S3, the frequency domain feature is obtained by performing a frequency domain analysis on the data vector X'(t)=[x'1(t), x'2(t), ..., x' n (t)] performs fast Fourier transform to obtain the spectrum amplitude distribution X i(f), where f is the frequency variable and the spectrum amplitude is at the frequency point f j The value at |X i (f j )|, by analyzing the spectrum amplitude distribution, determine whether there are abnormal harmonics.

[0022] In one embodiment of the present invention, in step S3, the time-frequency domain features are obtained by performing a computation on the data vector X'(t)=[x'1(t), x'2(t), ..., x' n (t)] performs L-layer wavelet packet decomposition and decomposes the signal into 2 L subspaces of different frequency bands, and obtain the wavelet packet coefficient matrix And the energy of each sub-band To capture the time-frequency local anomalies caused by faults.

[0023] In one embodiment of the present invention, in step S4, the fused convolutional neural network includes but is not limited to the first convolutional layer having k1 convolution kernels with a size of h1×ω1, and the convolution operation calculation function is It is then processed by an activation function and reduced in dimension by a pooling layer to output the feature map.

[0024] In one embodiment of the present invention, in step S4, the input gate i of the long short-term memory network t The update calculation function is i t =σ[W xi Z(t)+W hi h t-1 +b i ], forget gate f t The update calculation function is f t =σ[W xf Z(t)+W hf h t-1 +b f ], output gate o t The update calculation function is o t =σ[W xo Z(t)+W ho h t-1 +b o ], memory unit c t The update calculation function is c t =f t c t-1 +i t tanh[W xc Z(t)+W hc h t-1 +b c ], where h t is the hidden state at time t, ht The calculation function is h t =o t tanh(c t ).

[0025] In one embodiment of the present invention, in step S5, the expert system rule R={r1, r2, ..., r r}Based on the feature vector and the intermediate output setting of the model, the rule form is "IF condition THEN conclusion".

[0026] In one embodiment of the present invention, in step S6, the autoregressive integrated moving average model is expressed as ARIMA (p, d, q), and the model parameters are estimated through historical data to predict the fault development trend in the next h time steps to obtain the prediction result sequence Calculate predicted failure probability

[0027] The second aspect of the present invention discloses an inverter fault feature intelligent diagnosis and prediction system, which is used to implement the inverter fault feature intelligent diagnosis and prediction method, and comprises:

[0028] A data acquisition module, used to collect inverter operation data from multiple monitoring points;

[0029] Data preprocessing module, used to perform noise reduction and normalization on the original signal;

[0030] A feature extraction module, used to extract time domain features, frequency domain features and time-frequency domain features;

[0031] Intelligent diagnostic module to analyze feature vectors and identify failure modes;

[0032] Expert system module, used to modify the diagnosis results according to rules;

[0033] The fault prediction module is used to predict future fault development trends and set early warnings.

[0034] Additional contents and advantages of the present invention will be given in the following description or may be understood through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The technical solutions and beneficial effects of the present invention will become apparent and easily understood from the following contents in conjunction with the accompanying drawings, wherein:

[0036] Figure 1 The process of the inverter fault characteristic intelligent diagnosis and prediction method of the present invention is as follows Figure 1 ;

[0037] Figure 2 The process of the inverter fault characteristic intelligent diagnosis and prediction method of the present invention is as follows Figure 2;

[0038] Figure 3 This is a system block diagram of the inverter fault feature intelligent diagnosis and prediction system of the present invention. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0040] The inverter fault feature intelligent diagnosis and prediction method and inverter fault feature intelligent diagnosis and prediction system disclosed in the present invention will be described below with reference to the accompanying drawings.

[0041] like Figure 1 and Figure 2 As shown, a method for intelligent diagnosis and prediction of inverter fault characteristics includes the following steps:

[0042] S1. Collect inverter operation data from multiple monitoring points. Each monitoring point corresponds to one or more sensors. The data vector collected at a specific time is expressed as X(t)=[x1(t),x2(t),…,x n (t)], where x i (t) represents the data collected by the i-th sensor at time t;

[0043] S2. For the data vector X(t) = [x1(T), x2(t),…, x n (t)] to perform noise reduction and normalization preprocessing to obtain the preprocessed data vector X'(t) = [x'1(t), x'2(t), ..., x' n (t)];

[0044] S3. Extract the preprocessed data vector X'(t) = [x'1(t), x'2(t), ..., x' n (t)], the features include but are not limited to time domain features, frequency domain features and time-frequency domain features, wherein,

[0045] Time domain features include but are not limited to mean μ i ,variance and kurtosis K i ;

[0046] The frequency domain features obtain the spectrum amplitude distribution by performing fast Fourier transform on the time domain data and analyze whether there are abnormal harmonics;

[0047] The time-frequency domain features are decomposed by multi-layer wavelet packets to obtain the wavelet packet coefficient matrices of different frequency bands and the corresponding energy changes of each sub-band;

[0048] Mean μi The calculation function is variance The calculation function is Kurtosis K i The calculation function is

[0049] The frequency domain features are obtained by transforming the data vector X'(t) = [x'1(t), x'2(t), ..., x' n (t)] performs fast Fourier transform to obtain the spectrum amplitude distribution X i (f), where f is the frequency variable and the spectrum amplitude is at the frequency point f j The value at |X i (f j )|, by analyzing the spectrum amplitude distribution, determine whether there are abnormal harmonics.

[0050] The time-frequency domain features are obtained by transforming the data vector X'(t) = [x'1(t), x'2(t), ..., x' n (t)] performs L-layer wavelet packet decomposition and decomposes the signal into 2 L subspaces of different frequency bands, and obtain the wavelet packet coefficient matrix And the energy of each sub-band To capture the time-frequency local anomalies caused by faults.

[0051] S4. Combine the extracted time domain features, frequency domain features and time-frequency domain features into a feature vector F(t) = [F1(t), F2(t), ..., F m (t)] and used as the input of the intelligent diagnosis model, where F(t) = [F1(t), F2(t), …, F m (t)] contains all time domain features, frequency domain features, and time-frequency domain features;

[0052] The fused convolutional neural network includes but is not limited to the first convolutional layer with k1 convolution kernels of size h1×ω1, and the convolution operation calculation function is It is then processed through an activation function and reduced in dimension through a pooling layer to output a feature map.

[0053] Input gate i of the LSTM network t The update calculation function is i t =σ[W xi Z(t)+W hi h t-1 +b i ], forget gate f t The update calculation function is f t =σ[W xf Z(t)+W hf ht-1 +b f ], output gate o t The update calculation function is o t =σ[W xo Z(t)+W ho h t-1 +b o ], memory unit c t The update calculation function is c t =f t c t-1 +i t tanh[W xc Z(t)+W hc h t-1 +b c ], where h t is the hidden state at time t, h t The calculation function is h t =o t tanh(c t ).

[0054] S5. Build and train an intelligent diagnosis model M that integrates convolutional neural networks and long short-term memory networks CNN-LSTM , used to analyze the eigenvector F(t)=[F1(t),F2(t),…,F m (t)] and identify the inverter failure mode, where

[0055] The fused convolutional neural network is used to gradually extract deep spatial features, and the output feature map Z(t) is output after the activation function and pooling layer operations;

[0056] The long short-term memory network is used to receive the feature map Z(t) output by the fused convolutional neural network and expand it by time steps to learn the evolution law of fault features over time;

[0057] Expert system rule R = {r1, r2, ..., r r}Based on the feature vector and the intermediate output setting of the model, the rule form is "IF condition THEN conclusion".

[0058] S6. Combined with the expert system rule R = {r1, r2, ..., r r}Correct the diagnosis result and output the fault diagnosis result vector Y DI (t)=[y1(t),y2(t),…,y q (t)] and the credibility value p DI (t), where q is the number of fault categories, y i (t)∈{0,1}, when y i (t)=0 means there is no fault. i(t) = 1 indicates the existence of the i-th type of fault;

[0059] The autoregressive integrated moving average model is expressed as ARIMA (p, d, q). The model parameters are estimated through historical data to predict the development trend of faults in the next h time steps and obtain the prediction result sequence. Calculate predicted failure probability

[0060] S7. Use the autoregressive integrated moving average model to predict future fault development trends based on historical diagnostic data sequences and calculate the predicted fault probability And set warning thresholds When the probability of failure When a warning occurs, the early warning mechanism is triggered to notify the operation and maintenance personnel to take measures;

[0061] Specifically, in actual application scenarios, monitoring points are deployed at multiple key locations of the inverter, such as sensors are set at the input side, output side, power module, control circuit and other locations of the inverter. Each monitoring point is equipped with one or more sensors of different types, such as current sensors, voltage sensors, temperature sensors, etc. Taking a typical three-phase inverter as an example, a current sensor and a voltage sensor are installed on the line on the input side, and three-phase current sensors and voltage sensors are also arranged on the output side. Multiple temperature sensors are installed on the power module to monitor the temperature of the power device. The data vector collected at a specific time t is X(t)=[x1(t),x2(t),…,x n (t)], where n is the total number of sensors. Assume that the current data collected on the input side is x1(t) = 5.2A, and the voltage data collected on the input side is x2(t) = 220.5V, and so on, to complete the collection of the entire inverter operation data.

[0062] The collected data vectors are processed by wavelet denoising algorithm. Taking the current data on the input side as an example, the data is first decomposed by wavelet, and a suitable wavelet basis (such as db4 wavelet basis) and decomposition layer number (such as 5 layers) are selected to decompose the original signal into sub-band signals of different frequencies. Then, the high-frequency noise components in each sub-band signal are suppressed by setting a threshold rule (such as a soft threshold rule). After reconstruction, the current data on the input side after denoising is obtained.

[0063] The minimum-maximum normalization method is used to normalize the noise-reduced data. For the current data on the input side, the minimum and maximum values ​​in the entire data set are calculated to obtain the preprocessed X'(t) = [x'1(t), x'2(t), ..., x' n (t)];

[0064] For the extraction of time domain feature pairs, regarding the mean calculation, for each sensor data sequence in the preprocessed data vector, its mean is calculated. Taking the output side B phase current data sequence as an example, assuming that the sequence has T = 1000 data points, its mean μ i The calculation is

[0065] Regarding variance calculation, the variance is also calculated using the output side B phase current data sequence. First, the square of the difference between each data point and the mean is calculated, and then the average is calculated. The calculation is

[0066] Regarding the kurtosis calculation, the kurtosis of the output side B phase current data sequence is calculated, the kurtosis K i The calculation is

[0067] The preprocessed data vector is subjected to a fast Fourier transform (FFT) to obtain the spectrum amplitude distribution. Taking the voltage data on the input side as an example, the spectrum amplitude distribution X is obtained after performing an FFT transformation on it. i (f), where the value range of the frequency variable f is determined according to the actual sampling frequency. Assuming the sampling frequency is f s =10kHz, then the frequency resolution is Where N is the number of FFT points, assuming it is 1024, that is, At the frequency point f j =50Hz, the spectrum amplitude of the voltage on the input side is calculated to be |X A相电压 (50)|=311.0V, analyze the spectrum amplitude distribution to determine whether there are abnormal harmonics. For example, if a large amplitude appears at a non-integer multiple of the fundamental frequency, it is determined that there are abnormal harmonics, indicating that there may be a fault;

[0068] The preprocessed data vector is decomposed by L=3 layers of wavelet packet to decompose the signal into 2 L =8 subspaces of different frequency bands. Taking the three-phase current data on the output side as an example, the wavelet packet coefficient matrix is ​​obtained after decomposition And the energy of each sub-band For the energy calculation of the output side A phase current in a certain sub-band (such as the third sub-band), it is assumed that the wavelet packet coefficient matrix corresponding to the sub-band is Its energy By analyzing the changes in the energy of each sub-frequency band, the local time-frequency anomalies caused by the fault can be captured. For example, if the energy of a sub-frequency band is stable at a low level during normal operation, but suddenly increases significantly at a certain moment (such as from 0.01 to 0.1), it may indicate that the fault characteristics corresponding to the frequency band appear (such as abnormal oscillation during the switching process of the power device, etc.);

[0069] For the input of the feature vector combination and model, the extracted time domain features (mean, variance, kurtosis, etc.), frequency domain features (spectrum amplitude distribution and abnormal harmonic judgment results) and time-frequency domain features (wavelet packet coefficient matrix and energy changes of each sub-band) are combined into a feature vector F(t) = [F1(t), F2(t), …, F m (t)], for example, the mean, variance, kurtosis of the current on the input side, its spectrum amplitude at each frequency point, and the energy of each sub-band are arranged in a certain order to form part of the feature vector. Ensure that the feature vector contains all the extracted feature information and use it as a fusion convolutional neural network and long short-term memory network M CNN-LSTM The input of the intelligent diagnosis model;

[0070] Regarding the fused convolutional neural network part, k1=32 convolution kernels are set, with a size of h1×ω1, h1=3, ω1=3. Taking the partial data of the voltage and current corresponding to the input side in the input feature vector as an example, the convolution operation calculation function is Among them, F i+m,j+n (t) is the data at the corresponding position in the input feature vector, is the convolution kernel weight of the i-th row and j-th column, is the bias term, It is the result of the first convolution layer. The local features of the input data, such as the local change features of the voltage and current waveforms, are extracted through the convolution operation. Then, it is processed by the activation function (such as the ReLU function) to enhance the nonlinear expression ability of the network. Finally, the dimension reduction operation is performed through the pooling layer (such as the maximum pooling layer, the pooling kernel size is 2×2), reducing the amount of data while retaining the main features, and outputting the feature map Z(t);

[0071] Subsequently, multiple convolutional layers and pooling layers can be set as needed to gradually extract deep spatial features. For example, the second convolutional layer is set with 64 convolution kernels of size 3×3 to further extract more abstract features to better identify the characteristic patterns under different fault modes of the inverter.

[0072] For the long short-term memory network part, the feature map Z(t) is received and expanded by time step. Taking a feature map data with a time series length of 10 as an example, through the input gate i t 、Forget gate t , output gate o t and memory unit c tBy updating these gates, the LSTM network can learn the evolution of fault characteristics over time, such as the gradual development trend and periodic changes of fault characteristics, so as to better predict and diagnose inverter faults. For example, in a fault scenario where the temperature gradually rises due to poor heat dissipation of the inverter power module, the LSTM network can learn the trend of temperature data rising over time, as well as the time correlation relationship with other related characteristics (such as current and voltage changes), and predict possible faults in advance.

[0073] Prepare a large number of inverter operation data samples with fault labels and divide them into training set, validation set and test set. For example, collect 1000 sets of inverter data under different working conditions (including normal operation and various simulated fault conditions), of which 700 sets are used as training sets, 200 sets as validation sets, and 100 sets as test sets. Use the training set to construct the M CNN-LSTM The intelligent diagnosis model is trained, and the weight and bias parameters of the model are adjusted through the back propagation algorithm to minimize the error function (such as the cross entropy loss function) between the prediction result and the actual fault label. During the training process, the model is evaluated using the validation set, and the performance indicators such as the accuracy and recall rate of the model on the validation set are monitored to prevent overfitting. When the performance of the model on the validation set no longer improves, the training is stopped to obtain a trained intelligent diagnosis model.

[0074] Based on the in-depth understanding of the inverter failure mechanism, a large amount of experimental data and actual operation and maintenance experience, the expert system rule R = {r1, r2, …, r r}, the rule format is "IF condition THEN conclusion", for example, a rule is "IF the average current on the output side is greater than 10A and the amplitude of the third harmonic in the spectrum is greater than 10% of the fundamental amplitude THEN the power module is faulty";

[0075] The preliminary diagnosis results output by the trained intelligent diagnosis model are compared and verified with the expert system rules. Assume that the intelligent diagnosis model preliminarily determines that there is a power module fault in the inverter at a certain moment (the element corresponding to the power module fault in the output fault diagnosis result vector is 1), but at this time, according to the expert system rules, although the average current on the output side is slightly greater than 10A, the amplitude of the third harmonic does not reach 10% of the fundamental amplitude. After comprehensive judgment combined with other relevant rules, it is believed that this may be due to a short-term current fluctuation caused by a sudden load change, rather than a power module failure. Therefore, the diagnosis result is corrected, and the judgment result of the power module fault is corrected to no fault (the element corresponding to the power module fault in the output fault diagnosis result vector is modified to 0), and according to factors such as the credibility setting of the rule, the corrected diagnosis result vector Y is given. DI (t)=[y1(t),y2(t),…,y q(t)] = [0, 0, 1, 0, ..., 0] means that only the third type of fault exists and other faults do not exist) and the credibility value p DI (t) = 0.8, which means that the confidence of the diagnosis result is 80%;

[0076] The autoregressive integrated moving average model is used to predict the fault development trend. The parameters p, d, and q of the autoregressive integrated moving average model are determined by analyzing the historical diagnostic data series. For example, after data stationarity test and autocorrelation and partial autocorrelation analysis, p = 2, d = 1, and q = 1 are determined. The model parameters are estimated using historical data, and the ARIMA (2, 1, 1) model is constructed to predict the fault development trend within the next h = 5 time steps.

[0077] Taking the prediction of the development trend of power module failure in the next five time steps as an example, according to the constructed ARIMA model, the prediction result sequence is obtained: Calculate predicted failure probability Assuming that at time t+3, the diagnostic result vector of the predicted power module fault is [0, 0, 0.6, …], the predicted fault probability is calculated to be Setting warning thresholds because Trigger the early warning mechanism. The early warning system notifies the operation and maintenance personnel through SMS, email or pop-up windows on the monitoring platform, informing them of possible power module failures and providing relevant prediction information (such as failure probability, expected time of occurrence, etc.) so that the operation and maintenance personnel can take timely measures, such as checking the heat dissipation of the power module, monitoring changes in related electrical parameters, and preparing to replace power devices that may be damaged, so as to prevent failures in advance or reduce losses caused by failures.

[0078] In addition, in order to better explain the present invention, in one embodiment of the present invention, the circuit composed of the cascaded H-bridge multi-level inverter is a key part for converting direct current into alternating current. During long-term operation, the circuit is affected by various factors such as environmental factors (such as temperature, humidity, dust, etc.), electrical stress (such as overvoltage, overcurrent, voltage fluctuation, etc.) and aging of its own components, and various faults may occur.

[0079] For the input side, the current and voltage data are collected on the DC power supply, denoted as I in and V in For the output side, the current and voltage data are also collected on the three-phase output line, and the current and voltage data of phase A, phase B, and phase C are collected respectively, which are recorded as I Aout 、V Aout ,I Bout 、V Bout ,I Cout and V Cout .

[0080] For each H-bridge module, voltage data is collected on the DC side of each H-bridge module, marked as V H1 、V H2 , ..., collect current data on the AC side, marked as I H1 ,I H2 , …;

[0081] As for temperature data, temperature sensors are installed at key locations of the inverter (such as near the power module) to collect temperature data, which are recorded as T1, T2, ...

[0082] In the expert system rules, the following rules can be preset: “IF I in >I in_max THEN input overcurrent fault", where I in_max is the maximum input current setting, “IF V in >V in_max THEN input overvoltage fault”, where V in_max The maximum input voltage is set as “IF I out >I out_max TH EN output overcurrent fault”, where I out_max The maximum output current is set as “IF V out >V out_max THEN output overvoltage fault", where V out_max is the maximum output voltage setting, and other rules are analogous to this.

[0083] For example, there is a preset rule: "IF the average current of phase A on the output side is greater than 8A and the absolute value of the difference between the DC side voltage of H-bridge module 1 and the output side phase A voltage is greater than 10V THEN H-bridge module 1 is faulty", assuming that after the collected data is processed, the average current of phase A on the output side is 9A, and the DC side voltage of H-bridge module 1 is V H1 =200V, output side A phase voltage is V Aout =190V, the absolute value of the difference between the two is |200-190|=10V, which meets the conditions in the expert system rules, so it is judged that the H-bridge module 1 is faulty.

[0084] Now collect a new set of data at time t:

[0085] The input side data is I in (t)=5A,V in (t) = 200 V;

[0086] The output side data is I Aout (t)=6A,V Aout (t)=180V,I Bout(t)=5.5A,V Bout (t)=178V,I Cout (t)=5.8A,V Cout (t) = 182 V;

[0087] H-bridge module 1 DC side voltage V H1 (t) = 195V, AC side current I H1 (t) = 3A, temperature data T1(t) = 40°C.

[0088] The autoregressive integrated moving average model (ARIMA) is used to predict the fault development trend. First, the historical diagnostic data sequence is analyzed, the model parameters p=1, d=0, q=1 are determined, and the ARIMA(1,0,1) model is constructed. Taking the prediction of the fault development trend of H-bridge module 1 in the next three time steps as an example, the prediction result sequence is obtained based on the collected new data and model. and Calculate predicted failure probability

[0089] Assuming that at time t+2, the diagnostic result vector of the predicted H-bridge module 1 fault is [0, 0, 0.4, …], the predicted fault probability is calculated to be According to the set warning threshold (assuming it is 0.3), due to The early warning mechanism is triggered to notify the operation and maintenance personnel to take measures, such as further checking the relevant circuits of the H-bridge module 1 and monitoring the changes in electrical parameters.

[0090] like Figure 3 As shown, an inverter fault feature intelligent diagnosis prediction system 100 is used to implement an inverter fault feature intelligent diagnosis prediction method, which includes:

[0091] The data acquisition module 101 is used to collect inverter operation data from multiple monitoring points;

[0092] The data preprocessing module 102 is used to perform noise reduction and normalization processing on the original signal;

[0093] The feature extraction module 103 is used to extract time domain features, frequency domain features and time-frequency domain features;

[0094] Intelligent diagnosis module 104, used to analyze feature vectors and identify fault modes;

[0095] Expert system module 105, used to modify the diagnosis result according to the rules;

[0096] The fault prediction module 106 is used to predict the future fault development trend and set early warning.

[0097] In summary, the inverter fault feature intelligent diagnosis and prediction method and inverter fault feature intelligent diagnosis and prediction system disclosed in the present invention can effectively integrate a variety of advanced technologies, collect inverter operation data from multiple monitoring points, perform noise reduction, normalization preprocessing, and time domain, frequency domain, and time-frequency domain feature extraction, and then combine the features into vectors to input the intelligent diagnosis model that integrates convolutional neural networks and long short-term memory networks. The diagnosis results are corrected in combination with expert system rules, thereby improving the accuracy and reliability of the diagnosis. At the same time, the autoregressive integral moving average model is used to predict future fault development trends, set warning thresholds, and achieve early warning of faults, which helps operation and maintenance personnel to take timely measures to reduce problems such as power generation reduction and shutdown of photovoltaic power generation systems caused by inverter failures, reduce the economic losses of photovoltaic power generation companies, adapt to different fault types, and improve the overall intelligence level of fault diagnosis and prediction.

[0098] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. An intelligent diagnosis and prediction method for inverter fault characteristics, characterized in that: The following steps are involved: S1. Collect inverter operation data from multiple monitoring points. Each monitoring point corresponds to one or more sensors. The data vector collected at a specific time is expressed as X(t)=[x1(t),x2(t),…,x n (t)], where x i (t) represents the data collected by the i-th sensor at time t; S2. For the data vector X(t) = [x1(t), x2(t), ..., x n (t)] to perform noise reduction and normalization preprocessing to obtain the preprocessed data vector X'(t) = [x'1(t), x'2(t), ..., x' n (t)]; S3. Extract the preprocessed data vector X'(t) = [x'1(t), x'2(t), ..., x' n (t)], the features include but are not limited to time domain features, frequency domain features and time-frequency domain features, wherein, The time domain features include but are not limited to the mean μ i ,variance and kurtosis K i ; The frequency domain feature obtains the spectrum amplitude distribution by performing fast Fourier transform on the time domain data, and analyzes whether there are abnormal harmonics; The time-frequency domain features are decomposed by multi-layer wavelet packets to obtain wavelet packet coefficient matrices of different frequency bands and the energy changes of each corresponding sub-frequency band; S4. Combining the extracted time domain features, frequency domain features and time-frequency domain features into a feature vector F(t)=[F1(t), F2(t),…, F m (t)] and used as the input of the intelligent diagnosis model, where F(t) = [F1(t), F2(t), …, F m (t)] includes all of the time domain features, the frequency domain features and the time-frequency domain features; S5. Build and train an intelligent diagnosis model M that integrates convolutional neural networks and long short-term memory networks CNN-LSTM , used to analyze the eigenvector F(t)=[F1(t),F2(t),…,F m (t)] and identify the inverter failure mode, where The fused convolutional neural network is used to gradually extract deep spatial features, and outputs a feature map Z(t) after activation function and pooling layer operations; The long short-term memory network is used to receive the feature map Z(t) output by the fused convolutional neural network and expand it in time steps to learn the evolution law of fault features over time; S6. Combined with the expert system rule R = {r1, r2, ..., r r }Correct the diagnosis result and output the fault diagnosis result vector Y DI (t)=[y1(t),y2(t),…,y q (t)] and the credibility value p DI (t), where q is the number of fault categories, y i (y)∈{0,1}, when y i (y)=0 means there is no such fault. i (t) = 1 indicates the existence of the i-th type of fault; S7. Use the autoregressive integrated moving average model to predict future fault development trends based on historical diagnostic data sequences and calculate the predicted fault probability l=1,2,…,h, and set the warning threshold When the failure probability When a fault occurs, the early warning mechanism is triggered to notify the operation and maintenance personnel to take measures.

2. The inverter fault characteristic intelligent diagnosis and prediction method according to claim 1, characterized in that: In step S3, the mean μ i The calculation function is The variance The calculation function is The kurtosis k i The calculation function is 3. The inverter fault characteristic intelligent diagnosis and prediction method according to claim 1, characterized in that: In step S3, the frequency domain features are obtained by performing a multi-step analysis on the data vector X'(t)=[x'1(t), x'2(t), ..., x' n (t)] performs fast Fourier transform to obtain the spectrum amplitude distribution X i (f), where f is the frequency variable and the spectrum amplitude is at the frequency point f j The value at |X i (f j )|, by analyzing the spectrum amplitude distribution, determine whether there are abnormal harmonics.

4. The inverter fault characteristic intelligent diagnosis and prediction method according to claim 1, characterized in that: In step S3, the time-frequency domain features are obtained by performing a multi-step analysis on the data vector X'(t)=[x'1(t), x'2(t), ..., x' n (t)] performs L-layer wavelet packet decomposition and decomposes the signal into 2 L subspaces of different frequency bands, and obtain the wavelet packet coefficient matrix And the energy of each sub-band To capture the time-frequency local anomalies caused by faults.

5. The inverter fault characteristic intelligent diagnosis and prediction method according to claim 1, characterized in that: In step S4, the fused convolutional neural network includes but is not limited to the first convolutional layer having k1 convolution kernels with a size of h1×ω1, and the convolution operation calculation function is: It is then processed by an activation function and reduced in dimension by a pooling layer to output the feature map.

6. The inverter fault characteristic intelligent diagnosis and prediction method according to claim 1, characterized in that: In step S4, the input gate i of the LSTM network t The update calculation function is i t =σ[W xi Z(t)+W hi h t-1 +b i ], forget gate f t The update calculation function is f t =σ[W xf Z(t)+W hf h t-1 +b f ], output gate o t The update calculation function is o t =σ[W xo Z(t)+W ho h t-1 +b o ], memory unit c t The update calculation function is c t =f t c t-1 +i t tanh[W xc Z(t)+Wx c h t-1 +b c ], where h t is the hidden state at time t, h t The calculation function is h t =o t tanh(c t ).

7. The inverter fault characteristic intelligent diagnosis and prediction method according to claim 1, characterized in that: In step S5, the expert system rule R={r1,r2,…,r r }Based on the feature vector and the intermediate output setting of the model, the rule form is "IF condition THEN conclusion".

8. The inverter fault characteristic intelligent diagnosis and prediction method according to claim 1, characterized in that: In step S6, the autoregressive integrated moving average model is expressed as ARIMA (p, d, q), and the model parameters are estimated through historical data to predict the fault development trend in the next h time steps to obtain the prediction result sequence Calculate predicted failure probability 9. An inverter fault feature intelligent diagnosis and prediction system, characterized in that: It is used to implement the inverter fault feature intelligent diagnosis and prediction method according to any one of claims 1 to 8, and comprises: A data acquisition module, used to collect inverter operation data from multiple monitoring points; Data preprocessing module, used to perform noise reduction and normalization on the original signal; A feature extraction module, used to extract time domain features, frequency domain features and time-frequency domain features; Intelligent diagnostic module to analyze feature vectors and identify failure modes; Expert system module, used to modify the diagnosis results according to rules; The fault prediction module is used to predict future fault development trends and set early warnings.

Images