Temperature sensor fault diagnosis and fault data reconstruction method, system, equipment and medium

Abstract

The invention belongs to the technical field of wind turbine generator fault diagnosis and data reconstruction, and discloses a temperature sensor fault diagnosis and fault data reconstruction method, system and device and a medium. Performing optimization on hyper-parameters of the temperature sensor fault diagnosis model based on a firefly improved sparrow search algorithm, performing fault diagnosis on the wind turbine generator stator winding temperature sensor based on the optimized temperature sensor fault diagnosis model, constructing a fault data reconstruction model when a fault exists, and performing fault diagnosis on the wind turbine generator stator winding temperature sensor based on the fault data reconstruction model. And optimizing hyper-parameters of the temperature sensor fault diagnosis model based on a seagull optimization algorithm, and performing data reconstruction on sensor fault data of the wind turbine generator stator winding temperature sensor based on the optimized fault data reconstruction model. According to the technical scheme, an efficient, intelligent and accurate solution is provided for fault diagnosis and data reconstruction of the stator winding temperature sensor of the wind turbine generator.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind turbine fault diagnosis and data reconstruction, and in particular relates to a temperature sensor fault diagnosis and fault data reconstruction method, system, equipment and medium. Background Art

[0002] Wind turbines are one of the core components for achieving wind energy conversion. With the continuous increase in their capacity and complexity, the failure rate of components has increased, the reliability has decreased, and the maintenance cost has also increased significantly. For this reason, large-scale wind turbines are generally equipped with remote central monitoring and data acquisition systems (SCADA), which obtain the necessary real-time environmental parameters and operating status parameters through various sensors to complete wind turbine startup, pitch change, monitoring data over-limit alarm and brake protection functions. However, according to statistics, more than 40% of wind turbine failures are caused by sensor failures or related system failures. Sensor failures will directly cause the control system to receive erroneous information and issue erroneous control instructions, thereby endangering equipment safety and personnel safety. At the same time, wind turbines are difficult to detect during daily operation and their maintenance costs are expensive. Therefore, it is necessary to take measures to diagnose sensor failures and reconstruct failure data to ensure the normal operation of the system.

[0003] Many scholars have conducted research in the field of wind turbine sensor fault diagnosis and studied wind turbine sensor faults based on neural networks. However, the existing solutions are applicable to nonlinear abnormality diagnosis systems and are prone to fall into local optimal solutions, resulting in a decrease in prediction accuracy. Therefore, it is necessary to take optimization measures to improve the accuracy of the model. Summary of the invention

[0004] The purpose of the present invention is to provide a temperature sensor fault diagnosis and fault data reconstruction method, system, device and medium to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above object, the present invention provides a temperature sensor fault diagnosis and fault data reconstruction method, comprising:

[0006] S1: Acquire a historical data set of a wind turbine generator set, wherein the historical data set includes historical operating parameters of the wind turbine generator set, equipment status parameters of the wind turbine generator set, and historical parameters of the external environment;

[0007] S2: construct a temperature sensor fault diagnosis model based on the DBN deep belief network, train the temperature sensor fault diagnosis model based on the historical data set, optimize the hyperparameters of the temperature sensor fault diagnosis model based on the firefly improved sparrow search algorithm during the model training process, and obtain an optimized temperature sensor fault diagnosis model; perform fault diagnosis on the stator winding temperature sensor of the wind turbine based on the optimized temperature sensor fault diagnosis model, and output the sensor fault diagnosis result;

[0008] S3: When the sensor fault diagnosis result is greater than a preset threshold, a fault exists, and fault data of a period corresponding to the fault state is selected from the historical data set;

[0009] S4: constructing a fault data reconstruction model based on an LSTM long short-term memory network, training the fault data reconstruction model based on the historical data set and the fault data, optimizing the hyperparameters of the temperature sensor fault diagnosis model based on the Seagull optimization algorithm during the model training process, and obtaining an optimized fault data reconstruction model;

[0010] S5: Reconstruct the sensor fault data of the wind turbine stator winding temperature sensor based on the optimized fault data reconstruction model.

[0011] Optionally, it also includes performing data cleaning, correlation analysis and standardization on the data in the historical data set in sequence to obtain a preprocessed historical data set, and training the temperature sensor fault diagnosis model and the fault data reconstruction model based on the preprocessed historical data set.

[0012] Optionally, the step S2 specifically includes:

[0013] Dividing the historical data set into a training set, a test set, and a validation set according to a preset ratio;

[0014] Acquire several hyperparameter combinations of the temperature sensor fault diagnosis model as diagnosis model hyperparameter combinations;

[0015] Based on the training set, the temperature sensor fault diagnosis model corresponding to each of the diagnosis model hyperparameter combinations is trained, and the performance of the temperature sensor fault diagnosis model corresponding to each of the diagnosis model hyperparameter combinations is verified and evaluated through the validation set, and the evaluation result is used as the evaluation score corresponding to each of the diagnosis model hyperparameter combinations;

[0016] Constructing a sparrow population for sparrows with each of the hyperparameter combinations of the diagnostic model, updating and iterating the sparrow population based on the evaluation scores of each of the sparrows in combination with the firefly improved sparrow search algorithm, and outputting the optimal hyperparameter combination of the diagnostic model;

[0017] An optimized temperature sensor fault diagnosis model is constructed based on the optimal diagnosis model hyperparameter combination.

[0018] Optionally, updating and iterating the sparrow population based on the evaluation score of each sparrow in combination with the firefly improved sparrow search algorithm specifically includes:

[0019] S11: Randomly initialize the position of the sparrow population;

[0020] S12: updating the predator position, the joiner position and the guard position, calculating the evaluation score corresponding to each of the sparrows, and updating the position of each sparrow in the sparrow population based on the calculated evaluation score;

[0021] S13: updating the position of the sparrows through the firefly disturbance, calculating the evaluation score corresponding to each of the sparrows after the firefly disturbance, and updating the position of each sparrow in the sparrow population based on the calculated evaluation score;

[0022] Repeat steps S11 to S13 to perform iterative update until the iterative update reaches a preset number of iterations, and output the optimal diagnostic model hyperparameter combination.

[0023] Optionally, the step S4 specifically includes:

[0024] Dividing the historical data set and the fault data into a training set, a test set and a validation set according to a preset ratio;

[0025] Acquire several hyperparameter combinations of the fault data reconstruction model as reconstruction model hyperparameter combinations;

[0026] Based on the training set, the fault data reconstruction model corresponding to each of the reconstruction model hyperparameter combinations is trained, and the performance of the fault data reconstruction model corresponding to each of the reconstruction model hyperparameter combinations is verified and evaluated through the verification set, and the evaluation result is used as the evaluation score corresponding to each of the reconstruction model hyperparameter combinations;

[0027] Constructing a seagull population for seagulls using each of the reconstruction model hyperparameter combinations, updating and iterating the seagull population based on the evaluation scores of each of the seagulls in combination with a seagull optimization algorithm, and outputting an optimal reconstruction model hyperparameter combination;

[0028] An optimized fault data reconstruction model is constructed based on the optimal reconstruction model hyperparameter combination.

[0029] A temperature sensor fault diagnosis and fault data reconstruction system, comprising:

[0030] A data acquisition module is used to obtain a historical data set of the wind turbine generator set, wherein the historical data set includes historical operating parameters of the wind turbine generator set, equipment status parameters of the wind turbine generator set, and historical parameters of the external environment;

[0031] A fault diagnosis module is used to construct a temperature sensor fault diagnosis model based on a DBN deep belief network, train the temperature sensor fault diagnosis model based on the historical data set, optimize the hyperparameters of the temperature sensor fault diagnosis model based on the firefly-improved sparrow search algorithm during the model training process, and obtain an optimized temperature sensor fault diagnosis model; perform fault diagnosis on the stator winding temperature sensor of the wind turbine based on the optimized temperature sensor fault diagnosis model, and output the sensor fault diagnosis result; when the sensor fault diagnosis result is that a fault exists, select fault data of the time period corresponding to the fault state in the historical data set;

[0032] A data reconstruction module is used to build a fault data reconstruction model based on an LSTM long short-term memory network, train the fault data reconstruction model based on the historical data set and the fault data, optimize the hyperparameters of the temperature sensor fault diagnosis model based on the Seagull optimization algorithm during the model training process to obtain an optimized fault data reconstruction model; and reconstruct the sensor fault data of the wind turbine stator winding temperature sensor based on the optimized fault data reconstruction model.

[0033] An electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute a temperature sensor fault diagnosis and fault data reconstruction method.

[0034] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for diagnosing faults of a temperature sensor and reconstructing fault data is implemented.

[0035] The technical effects of the present invention are:

[0036] The present invention aims at the fault diagnosis problem of the stator winding temperature sensor of the wind turbine generator set, and establishes a deep belief network prediction model based on the firefly disturbance improved sparrow search optimization algorithm to determine whether the sensor is faulty, thereby realizing efficient and accurate fault diagnosis of the stator winding temperature sensor of the wind turbine generator set. According to the correlation and timing between the temperature sensor parameters, a multi-input single-output long-short-term memory fault data reconstruction model is established, and the model parameters are optimized by the seagull optimization algorithm, and data is reconstructed based on the optimized model. The present invention not only provides a fault diagnosis method, but also covers the reconstruction of fault data, forming a complete set of wind turbine generator set stator winding temperature sensor fault diagnosis and data recovery methods. The present invention combines advanced deep learning technology and optimization algorithms to provide an efficient, intelligent and accurate solution for the fault diagnosis and data reconstruction of the stator winding temperature sensor of the wind turbine generator set. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0038] Figure 1 is a structural diagram of a DBN model in an embodiment of the present invention;

[0039] Figure 2 A schematic diagram of the correlation between input parameters of the FISSA-DBN model in an embodiment of the present invention;

[0040] Figure 3 It is a box plot of all input parameters of the FISSA-DBN model in the embodiment of the present invention;

[0041] Figure 4 is a curve diagram of the alarm threshold and the early warning threshold in the embodiment of the present invention;

[0042] Figure 5 is a standard RNN model structure diagram in an embodiment of the present invention;

[0043] Figure 6 This is a multi-input single-output diagram of the RNN model in an embodiment of the present invention;

[0044] Figure 7 This is a standard LSTM cell structure diagram in an embodiment of the present invention;

[0045] Figure 8 A schematic diagram of the migration and attack modes of seagulls in an embodiment of the present invention;

[0046] Fig. 9 The stator winding 2 temperature sensor fault state monitoring δ in the embodiment of the present invention E Graphs;

[0047] Fig.10 The stator winding 6 temperature sensor fault state monitoring δ in the embodiment of the present invention E Graphs;

[0048] Fig.11 This is a schematic diagram of the result after data reconstruction based on the trained SOA-LSTM model in an embodiment of the present invention;

[0049] Fig.12 is a difference diagram between an actual fault value and a reconstructed value in an embodiment of the present invention;

[0050] Fig.13 Flowchart of an embodiment of the present invention. DETAILED DESCRIPTION

[0051] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as limiting the present invention, but should be understood as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0052] Embodiment 1

[0053] like Figure 1 - Fig.13 As shown, in this embodiment, a method for temperature sensor fault diagnosis and fault data reconstruction is provided, including: obtaining a historical data set of a wind turbine, wherein the historical data set includes historical operating parameters of the wind turbine, equipment status parameters of the wind turbine, and historical parameters of the external environment; constructing a temperature sensor fault diagnosis model based on a DBN deep belief network, training the temperature sensor fault diagnosis model based on the historical data set, optimizing the hyperparameters of the temperature sensor fault diagnosis model based on a firefly-improved sparrow search algorithm during the model training process, and obtaining an optimized temperature sensor fault diagnosis model; and diagnosing the stator winding of the wind turbine based on the optimized temperature sensor fault diagnosis model. The temperature sensor performs fault diagnosis and outputs the sensor fault diagnosis result; when the sensor fault diagnosis result is greater than a preset threshold, a fault exists, and fault data of a time period corresponding to the fault state is selected from the historical data set; a fault data reconstruction model is constructed based on an LSTM long short-term memory network, and the fault data reconstruction model is trained based on the historical data set and the fault data. During the model training process, the hyperparameters of the temperature sensor fault diagnosis model are optimized based on the Seagull optimization algorithm to obtain an optimized fault data reconstruction model; based on the optimized fault data reconstruction model, the sensor fault data of the stator winding temperature sensor of the wind turbine generator set is reconstructed.

[0054] This embodiment aims at the fault diagnosis problem of the stator winding temperature sensor of the wind turbine, and establishes a deep confidence network prediction model based on the firefly disturbance improved sparrow search optimization algorithm. The input parameters of the model are determined by the expert system and Pearson correlation analysis method, and the exponential weighted moving average threshold method is used to determine whether the sensor is faulty. According to the correlation and timing between the temperature sensor parameters, a multi-input single-output long-term and short-term memory fault data reconstruction model is established, and the model parameters are optimized by the seagull optimization algorithm. Finally, the actual data of a wind farm is analyzed to verify the feasibility of the model. The unoptimized DBN model and the DBN (WOA-DBN) model optimized by the whale optimization algorithm (WOA) are compared.

[0055] Establishment of fault diagnosis model for stator winding temperature sensor of wind turbine:

[0056] Deep Belief Network: DBN is an unsupervised artificial neural network, which is formed by stacking multiple restricted Boltzmann machines (RBM). Figure 1 shown.

[0057] Assuming that the DBN model contains l hidden layers, its joint probability is shown in formula (1):

[0058] p(υ,h 1 ,h 2 ,...,h l )=p(υ|h 1 )p(h 1 |h 2 )...p(h l-1 |h l ) (1)

[0059] Where: p(h j-1 |h j ),j=2,3,…,l represents the conditional probability distribution between the (j-1)th layer and the jth layer of the hidden layer. The energy between the visible layer neurons and the hidden neurons is defined as shown in formula (2):

[0060]

[0061] Where: θ = {ω ij ,b i ,c j} represents the network parameters; ω ij refers to the weight between the neurons in the visible layer and the neurons in the hidden layer; b i is the bias of the neurons in the explicit layer; c j is the bias of the hidden layer neurons. m is the number of neurons in the visible layer, and n is the number of neurons in the hidden layer.

[0062] The input X of the DBN network model is expressed as:

[0063] X=υ=[υ1,υ2,υ3,...,υ m ] (3)

[0064] Where: X is the relevant parameter set in SCADA, such as wind speed, generator voltage, generator current, etc., which can reflect the status or fault characteristics of the wind turbine.

[0065] Based on the above energy function, formula (1) can be expressed as:

[0066]

[0067] Where: Z is the allocation function, used to adjust e -E(υ,h)There is no connection between the explicit layer and the hidden layer. The activation functions of the explicit layer neurons and the hidden layer neurons of the network are expressed as:

[0068]

[0069] Where: σ(x) is the Sigmoid function, expressed as:

[0070]

[0071] Train RBM through contrastive divergence learning algorithm to find the optimal weight ω ij , the new weights and biases are continuously updated, and the update expression is:

[0072]

[0073] c i ←c i-1 +λ(h i -h i+1 )(9)

[0074] b j ←b j-1 +λ(υ i -υ i+1 )(10)

[0075] Where: λ is the learning rate and s represents the number of RBMs.

[0076] Firefly Improved Sparrow Search Algorithm: During the establishment of the deep confidence network model, the network parameters are randomly initialized, which will lead to the generation of local optimal solutions. The Firefly Improved Sparrow Search Algorithm Optimized Deep Confidence Network (FISSA-DBN) is used to optimize the deep confidence network. By seeking the optimal value of the DBN network learning rate and the number of hidden neurons in each layer of the RBM, a higher diagnostic accuracy can be achieved in the global sample space.

[0077] FISSA uses firefly perturbation to update the positions of all sparrows and the best sparrow to improve search performance, compare the sparrows after disturbance with those before disturbance, and update the sparrow's position if it is better.

[0078] The relative fluorescence brightness of fireflies is:

[0079]

[0080] In formula (11), I0 is the maximum fluorescence brightness of the firefly, which is related to the objective function value. The better the objective function value, the higher the brightness. γ is the light intensity absorption coefficient. The fluorescence will gradually weaken with the increase of distance and the absorption of the propagation medium. i,j is the spatial distance between fireflies i and j.

[0081] The attractiveness of fireflies is:

[0082]

[0083] In formula (12), P0 is the maximum attraction.

[0084] The position update formula of firefly i being attracted to firefly j is shown in (13):

[0085] x i =x i +P*(x j -x i )+q*(rand-1 / 2) (13)

[0086] In formula (13), q∈[0,1] is the step size factor; rand is a random number that follows a uniform distribution on [0,1]. The FISSA algorithm is used to optimize the DBN model. The DBN model sets two hidden layers. According to the best fitness position x obtained by formula (13), the number of neurons and the learning rate of the two hidden layers of the DBN model are optimized, as shown in formula (14):

[0087]

[0088] In formula (14), x(k), k = 1, 2, 3 are the optimal fitness values ​​obtained after optimization, α represents the number of neurons in the first hidden layer of RBM after optimization, β represents the number of neurons in the second hidden layer of RBM after optimization, and λ represents the learning rate obtained after optimization.

[0089] Research on FISSA improved deep belief network algorithm: Since DBN networks are prone to fall into local optimal solutions, the firefly improved sparrow search algorithm is used to optimize the deep belief network and establish a FISSA-DBN model. The specific steps of the FISSA optimization process of DBN are as follows.

[0090] 1) Set the parameters of FISSA. Based on multiple experiments, the number of iterations is selected as 10, the population size is 40, the proportion of predators is 20%, the safety threshold T is 0.8, the proportion of joiners is 30%, and the number of sparrows aware of the danger is 5.

[0091] 2) Optimize parameter settings. Randomly initialize the position of the sparrow population and set the parameter ranges of α, β and λ, α∈[1,100], β∈[1,100], λ∈[0.01,1].

[0092] 3) Calculate the fitness value. Calculate the fitness value of the sparrow according to the objective function, update the predator position, joiner position and guard position, calculate the fitness and update the sparrow position, and save the best position individual. The fitness expression is:

[0093] Best_pos(t)=x i =Function(α,β,λ) (15)

[0094] 4) Firefly disturbance: Use firefly disturbance to update the position of the sparrow, calculate the fitness value and update the position of the sparrow, and save the best position individual.

[0095] 5) Optimization ends. Determine whether the stopping condition is met based on the defined maximum number of iterations. If it is met, exit and output the optimal parameter vector; otherwise, return to step 2 and repeat the above steps until the optimal parameters are obtained.

[0096] Since the parameters α, β and λ in the DBN fault diagnosis model are randomly given, the setting of the initial parameters has a great influence on the performance of the DBN model. Therefore, the optimal parameters are selected by FISSA, and the FISSA optimization DBN model method is used to study the fault diagnosis of the wind turbine winding temperature sensor. The FISSA-DBN-based method has the advantages of high robustness, fast learning speed and high diagnostic accuracy. According to the statistics of existing literature, this method has not been applied in the fault diagnosis of wind turbine sensors. Therefore, it is necessary to use this method to study the fault diagnosis and prediction of wind turbine sensors.

[0097] The stator temperature sensor fault diagnosis process based on FISSA-DBN is as follows:

[0098] 1) Data selection: According to the operation records of wind turbines, the original operation data measured by the stator winding temperature sensor of the wind turbines are extracted from the SCADA data.

[0099] 2) Data preprocessing: Based on expert domain knowledge, the overall prediction model condition parameters are selected from the original SCADA data, the data information is further processed according to the probability of wind farm accidents, and the erroneous and unnecessary data in the original data are deleted.

[0100] 3) Determination of prediction model parameters: According to the stator winding temperature sensor of the research object, the preprocessed data group obtained in step 2 is used to analyze the correlation between each parameter and the stator winding temperature output parameter using the Pearson correlation coefficient. Parameters with a strong correlation with the stator winding temperature and the stator winding temperature parameters are selected as the input parameters of the model. The model output parameter is the stator winding temperature measurement value.

[0101] 4) Data processing: Standardize the selected data, normalize the values ​​to the interval [0,1], and use the data samples to train or test the prediction model. Select 20,000 sets of historical normal data in each state as samples, 80% of the random data as training data, and 20% as test data.

[0102] 5) Model topology: Set relevant parameters such as the population size of sparrows, number of iterations, producer ratio, and DBN structure.

[0103] 6) DBN parameter optimization: FISSA is used to optimize the learning rate of the RBM layer and the number of neurons in the hidden layer to obtain the optimal initial vector of the DBN model.

[0104] 7) Model training: According to the research object of wind turbine operation status monitoring, the model is trained and tested to obtain the upper and lower thresholds of the exponentially weighted moving average threshold (EWMA) control chart. The performance of the prediction model is evaluated by the root mean square error (RMSE), mean absolute percentage error (MAPE) and mean absolute error (MAE).

[0105]

[0106] Among them, x(t) and They represent the actual value and predicted value of the sample respectively, and N represents the total number of samples.

[0107] 8) Application of prediction model: The actual wind farm SCADA stator winding temperature measurement values ​​are input into the prediction model for prediction, and the sensor is judged whether it is faulty based on the prediction results and EWMA.

[0108] Determination of prediction model parameters:

[0109] Sensor determination: A total of 6 wind turbine stator winding temperature sensors are installed, with 2 for each phase of the three-phase winding, marked as stator temperature sensors 1-6. The data comes from the SCADA system data of the same wind farm provided by a wind power company in East China. The capacity of each unit is 2MW. Data with a sampling interval of 10 minutes for a total of 5 years from January 1, 2013 to December 30, 2017 are selected for analysis. Combined with the fault records, the stator winding temperature sensor No. 2 of wind turbine No. 19 in 2013 and the stator winding temperature sensor No. 6 of wind turbine No. 7 in 2014 were selected as the research objects.

[0110] Determination of input parameters: The accuracy of the prediction model is directly related to the input parameters of the model. Different parameter selection will lead to deviations in the prediction results. However, there are many parameters in the SCADA data, and not every parameter is related to the stator winding temperature. Therefore, in order to improve the accuracy of fault diagnosis and shorten the fault diagnosis time, the input parameters of the prediction model are selected by expert system judgment and Pearson correlation coefficient method. After expert system judgment, 44 different types of state parameters are selected from 150 parameters in the SCADA data, and then the Pearson correlation coefficient method is used to calculate and further determine the input parameters from the 44 parameters. Since there are correlation characteristics between the various components when the wind turbine is running, the state parameters will affect each other. However, it is necessary to select appropriate input parameters, and it is not advisable to select too many to avoid information redundancy of the model and affect the accuracy of the prediction. After repeated experiments, the Pearson correlation evaluation standard value is selected as 0.75, where the inverter grid-side current is related to the grid-side active power, and the inverter grid-side active power is selected. The generator current is related to the generator torque, and the generator current is selected. The correlations among the 6 generator stator temperature measurements are very high, so the generator stator temperature 4 with the highest relative value is selected as the input parameter. There are 2 parameters of main bearing temperature, and the main bearing temperature 1 is selected as the input parameter. In summary, it can be obtained that the input parameters of the prediction model are hub speed, cabin weather station wind speed, inverter grid-side active power, generator side power, generator operating frequency, generator current, generator stator temperature 4, main bearing temperature 1 and cabin temperature, a total of 9 parameters. The parameters of the stator winding temperature sensor fault diagnosis model are shown in Table 1, and the correlations between the parameters are shown in Table 1. Figure 2 As shown in Figure 2, the box plots of all input parameters of the model are shown in Figure 2. Figure 3 shown.

[0111] Table 1 Stator winding temperature sensor fault diagnosis model parameters

[0112]

[0113] Fault threshold determination: The DBN network encodes and decodes the input data and then outputs it. The input variables and outputs have the same physical meaning. The DBN network parameters are tuned through FISSA to make the model better fit the input data. Calculate the reconstruction error δ between the predicted value and the actual value of the prediction model E , and use EWMA as the basis for judging whether the state of the stator winding temperature sensor of the wind turbine is abnormal. When the wind turbine is operating normally, the SCADA monitoring variable data satisfies a stable internal relationship. When a fault occurs, the reconstruction error trend will change significantly, and its reconstruction error expression is:

[0114]

[0115] In formula (19), x(i) is the actual value, is the predicted value.

[0116] By analyzing δ E The change trend or mutation degree of the temperature sensor of the stator winding of the wind turbine can be monitored. The threshold line is set by using the exponential weighted moving average method to detect small drift of the mean value, which is suitable for continuous monitoring of δ E 's changing trend.

[0117] Statistics Z of EWMA Control Chart t for

[0118] Z t =λδ Et +(1-λ)Z t-1 (20)

[0119] In formula (20), λ is the weight factor, which represents the historical δ E The impact on the current EWMA, λ∈(0,1], the value is 0.2; t is the time; δ Et are independent random variables.

[0120] The mean of the EWMA statistic is

[0121]

[0122] In formula (22), is the δ of the corresponding equipment of the same type of wind turbine E The mean and variance of the EWMA statistic are

[0123]

[0124] In formula (23), n is the sampling length; is δ E The square of the standard deviation.

[0125] The upper and lower limit functions of the EWMA control chart for time-based wind turbine operating status monitoring are defined as

[0126]

[0127] In formulas (24) and (25), γ is the alarm coefficient. After repeated verification, the value is 4, which means that when the predicted reconstruction error is 4 times the mean, a fault alarm will be issued. κ is usually set to 2 or 3. When it is 2, it represents the limit of the fault warning threshold, and when it is 3, it represents the limit of the fault alarm threshold.

[0128] From formula (24) and (25), we can see that U c (t) and U L (t) is a set of data that changes with time. As t increases, these two values ​​tend to be constants, so the threshold line takes the last constant value. When the wind turbine is operating normally or shut down, δ E The value of is maintained below the EWMA threshold line and can operate stably for a long time, but when δ E When the change trend exceeds the threshold line and remains on the threshold line, it means that there is a large deviation between the actual value and the reconstructed value, the original equilibrium state has changed, and the operating state is abnormal. Figure 4 It is a curve chart of alarm threshold and warning threshold.

[0129] Establishment of wind turbine sensor fault data reconstruction model: Sensor fault data will affect the normal operation of the system, so it is necessary to reconstruct the fault output data. Since wind farm data has strong timeliness and relevance, the data collected by the sensor is time series data, which will not mutate, has a strong correlation in time, and the correlation between parameters is strong, so it is very suitable to use LSTM network multi-input single output to realize sensor fault data reconstruction. However, the LSTM cell nucleus has the characteristics of forgetting and retaining the collected data information, which will affect the accuracy of data reconstruction. Therefore, the SOA algorithm is used to select the optimal parameters of the LSTM parameters, which not only has fast convergence speed and simple model structure, but also can improve the reconstruction accuracy of fault data.

[0130] Long Short-Term Memory Networks: Figure 5 This is a standard RNN model structure diagram. Figure 6 This is a diagram of the RNN model with multiple inputs and a single output. This form is an extension of the basic model. A Softmax function is added to the output layer at the last time point. Taking 4 input parameters as an example, the output expression is shown in (27). LSTM is an improved version of RNN. It adds input gate, forget gate, and output gate on the basis of RNN. Since it has three gates, its structure is slightly more complicated. The standard LSTM cell structure is as follows: Figure 7 shown.

[0131]

[0132] Y=Softmax(Vh4+c) (27)

[0133] Figure 8 As can be seen from the figure, the process of LSTM is as follows: each moment has an output and a cell state. The output is h above, and the state is represented by C. Then, for the next moment, the output of the previous moment and the input of this moment are taken as a whole as the input of the cell, which is also the input to control the three gates. First, it must pass through a forget gate, using a sigmoid function with a value between 0 and 1. If the output of the forget gate is close to 0, it means that the output of the previous moment is forgotten as much as possible, and if it is close to 1, it means that it is memorized. Then, the remaining part of the previous state (C t-1 and f t The product of the previous state) continues to move forward, and it needs to add the input information of this state. The input information is realized according to the situation of the input gate. The input gate represents the amount of input that passes through. The signal before entering the input gate is first operated by tanh to compress the value range to the range of [-1,1]. The input signal processed by the input gate is added to the residual of the previous state to obtain the cell state at this moment. The cell state completes the update from t-1 to t. The relationship between the output and state C is the function of the output gate. The output gate is multiplied by the cell state after tanh, and the result is the output at time t. The expressions between the various gates are shown in equations (28)-(33).

[0134] f t =σ(W f x t +U f h t-1 +b f ) (28)

[0135] i t =σ(W i x t +U i h t-1 +b i ) (29)

[0136] o t =σ(W o x t +U o h t-1 +b o ) (30)

[0137] g t =tanh(W c x t +U c ht-1 +b c ) (31)

[0138] C t =f t C t-1 +i t g t (32)

[0139] h t =o t *tanh(C t ) (33)

[0140] In formulas (28)-(33), f t is the activation value of the forget gate, i t is the activation value of the input gate, o t is the activation value of the output gate, h t and h t-1 is the output of the memory unit at time step t and time step t-1. t and C t-1 are the states of the memory cells at time step t and time step t-1 respectively. The W value represents the weight matrix of the input data, the U value represents the weight matrix output by the LSTM cell at the previous moment, and b is the state offset.

[0141] Seagull Optimization Algorithm: LSTM is a neural network that has the ability to memorize long-term and short-term information. The traditional training method uses gradient descent. However, the gradient descent method is prone to fall into local optimality, and the swarm intelligence optimization algorithm has a strong global search capability. This embodiment uses the Seagull Optimization Algorithm (SOA) to directly optimize the number of hidden units, maximum training cycle, initial learning rate and L2 parameters of LSTM to improve the solution accuracy of LSTM. The migration and attack methods of seagulls are as follows: Figure 8 shown.

[0142] Seagull migration behavior: During migration, the algorithm simulates how a flock of seagulls moves from one location to another. In this process, the main thing is to avoid collisions with other seagulls. The algorithm uses an additional variable A to calculate the new position of the seagull.

[0143] C s (t) = A*P s (t) (34)

[0144] In formula (34), C s (t) represents the new position that does not conflict with other seagulls, P s (t) represents the current position of the seagull, t represents the current iteration, and A represents the movement behavior of the seagull in the given search space.

[0145] A=f c -(t*(f c / Max iteration )) (35)

[0146] In formula (35), f c The frequency of variable A can be controlled, and its value decreases linearly from 2 to 0. After avoiding the position overlap with other seagulls, the seagull will move in the direction of the best position.

[0147] M s (t) = B*(P bs (t)-P s (t)) (36)

[0148] In formula (36), M s (t) represents the direction of the optimal position, and B is a random number responsible for balancing global and local searches.

[0149] B=2*A 2 *r d (37)

[0150] In formula (37), r d is a random number in the range [0,1]. When the seagull moves to a position where it does not collide with other seagulls, it moves in the direction of the best position to reach the new position.

[0151] D s (t)=|C s (t)+M s (t) (38)

[0152] In formula (38), D s (t) is the latest position of the seagull.

[0153] Seagull attack behavior: Seagulls perform spiral movements in the air. Fig. 9 In the above equation, the motion behavior in the x, y, and z planes is described as shown in equation (39).

[0154]

[0155] In formula (39), r is the radius of each spiral, θ is a random angle value in the range [0,2π], u and v are related constants of the spiral shape, and e is the base of the natural logarithm. The attack position of the seagull is shown in formula (40).

[0156] P s (t) = D s (t)*x*y*z+P bs (t) (40)

[0157] In formula (40), P s (t) is the attack position of the seagull

[0158] SOA-LSTM model construction:

[0159] Principle of LSTM network data reconstruction: Sensor fault data reconstruction is to replace the original fault data with the optimal estimated value of the sensor based on the results of fault diagnosis, so as to ensure that the system continues to operate under normal circumstances. Therefore, the sensor fault data reconstruction process is essentially a process of finding the optimal replacement value for the fault value. The SCADA data of the wind farm monitoring system is characterized by strong time series, and the previous data and the next data are temporally correlated, and the data will not suddenly jump. The LSTM multi-input single-output network can fuse the data of multiple parameters before a certain time, the current state and the current input information, etc., to predict the next data. Since the LSTM network has strong correlation time and space sequence information in processing data, its predicted value is basically a reflection of the true value. The LSTM network can not only realize fault diagnosis, but also its predicted value is the reconstruction value of the fault data.

[0160] When a sensor is detected to have a fault in a certain period of time, the LSTM network with multiple inputs and a single output is used to reconstruct the sensor fault data. The basic principle is to obtain the associated parameters of the sensor through various methods according to a certain sampling time interval of the sensor. For example, if 3 associated parameters are selected, the normal m data before the parameters are taken, and the first parameter is marked as x1', x'2,..., x' m , the second parameter is labeled x1',x'2',...,x' m , the third parameter is labeled x1”',x’2”,...,x” m ', forming the first 3×m cell, together as the first input value of the LSTM network. Then the data is shifted back one position, and the first parameter is marked as x'2, x'3, ..., x' m+1 , the second parameter is labeled x'2',x'3',...,x" m+1 , the third parameter is labeled x'2",x'3",...,x" m ' +1 , forming the second 3×m cell, together as the second input value of the LSTM network, and so on, forming a time series set of multiple columns. Build an LSTM multi-input single-output prediction model, and use the constructed time series set for model training. Similarly, build a multi-parameter time series set of fault data, and use the trained LSTM model for prediction. After multiple predictions by the LSTM network, you can get the prediction data set y1, y2, …, y m, its fault data is replaced by the prediction data set in the corresponding period to achieve reconstruction of the fault data.

[0161] SOA optimization process: SOA is a new meta-heuristic algorithm proposed in recent years. It mainly imitates the seagulls' behaviors in finding and attacking prey and migrating in nature. The SOA swarm intelligent optimization algorithm has excellent characteristics such as high solution accuracy, stable and reliable algorithm, strong global search ability and fast convergence speed. It can improve the local optimum caused by gradient descent in the process of continuous iterative training of LSTM. The steps of the SOA algorithm are as follows:

[0162] 1) Parameter initialization: For the SOA parameter initialization settings, the population size is set to 50, the spatial dimension is 4, the maximum number of iterations is 200, the search lower boundary is set to [2,2,10E-5,10E-6], and the upper boundary is set to [200,100,1,1]. A, B and Max in the initialization parameter expressions (34) to (38) are reation ;

[0163] 2) Parameter assignment: f c Set to 2, and u and v to 1. d ∈[0,1],θ∈[0,2π];

[0164] 3) Initial value calculation: calculate the fitness value of each seagull, that is, the objective function value;

[0165] 4) Calculate D s : According to formulas (34) to (38), calculate the optimal position of each seagull;

[0166] 5) Calculate P s (x): Calculate the optimal attack position for each seagull according to formulas (39) to (40);

[0167] 6) Update position: Update the best position information and fitness value of the seagull, t = t + 1;

[0168] 7) Judgment: If t>Max iteration , execute step 8, otherwise jump to step 3;

[0169] 8) Output: Output the best seagull position and fitness value.

[0170] Run the SOA algorithm to get the best fitness value, and assign the value in the best fitness value Best_pos to the number of hidden units (numHiddenUnits), maximum training cycle (maxEpochs), initial learning rate (InitialLearnRate) and L2 parameter (L2Regularization) of LSTM. The expressions are as follows.

[0171] numHiddenUnits=Best_pos(1) (41)

[0172] maxEpochs = Best_pos(2) (42)

[0173] InitialLearnRate=Best_pos(3) (43)

[0174] L2Regularization=Best_pos(4) (44)

[0175] The specific steps of SOA optimization of LSTM network are as follows:

[0176] 1) Data selection: Select the normal historical data of the relevant parameters of the anemometer and stator winding No. 2 temperature sensor analyzed previously.

[0177] 2) Data processing: Establish a data sequence, determine the training input sample and output sample vector, as well as the test input sample and output sample vector. The m values ​​in the sample data are finally determined by error through multiple experiments. The selected data is normalized and the data range is [0, 1].

[0178] 3) Model topology: Set the population size, number of iterations, spatial dimension, boundary range of the seagulls, and initialize A, B, and Max reation Assign values ​​to other parameters and set the input and output dimensions of LSTM parameters.

[0179] 4) Parameter optimization: The SOA algorithm is used to find the optimal value, and the optimal value is assigned to the number of hidden units, maximum training cycle, initial learning rate and L2 parameter of the LSTM network to obtain the optimal initial vector of the LSTM.

[0180] 5) Model training: Train the model based on the previously selected training data sequence, test it with test data, and verify the reconstruction accuracy of the model.

[0181] 6) Fault data reconstruction: Based on the previous analysis, the faulty anemometer and stator winding 2 temperature data are selected, and the trained model is used to reconstruct the data.

[0182] SOA-LSTM model training and fault data reconstruction process:

[0183] Model training and testing: After the SOA-LSTM model is established, the model is trained and tested. The specific process is as follows:

[0184] 1) 20,000 historical normal data of relevant parameters of the research sensor are selected, of which 80% are randomly selected as training data, and each relevant parameter time series is recorded as (a1, a2, ..., a 16000 ), the time series of the developed sensor is recorded as (b1, b2, …, b 16000 ), and the remaining 20% ​​is used as test data. The time series of the developed sensor is recorded as (y1, y2, ..., y 4000 ).

[0185] 2) Construction of training data time series. Assuming that the number of sensor-related parameters is n, the m data before the relevant parameters are used to predict the m+1th data of the sensor under study. The training data input model time series sample consists of 16000-m cells, denoted as (X1, X2, …, X 16000-m ), each cell is an n×m matrix. The expression of X1 is shown in (45), the expression of X2 is shown in (46), and so on. 16000-m The expression of is shown in (47). The output model time series sample vector is recorded as (Y1, Y2, ..., Y 16000-m ), where Y1 is b m+1 , Y2 is b m+2 , and Y 16000-m for b 16000 .

[0186]

[0187] 3) Construction of test data time series. Similar to the training data time series, the test data input model time series sample consists of 4000-m cells, denoted as (Z1, Z2, …, Z 4000-m ). Among them, Z1, Z2…Z 4000-m The matrix expression of is similar to equation (45). The output model time series sample vector is recorded as (C1, C2, ..., C 4000-m ), where C1 is y m+1 , C2 is y m+2 , and similarly we get C 4000-m for y 4000 .

[0188] 4) Time series normalization: The training and test data are normalized according to formula (48).

[0189]

[0190] 5) SOA-LSTM model settings: According to the previous introduction, set the SOA-LSTM model parameters, and the error is expressed as RMSE.

[0191] 6) Model training. Select different m values. Since the selected SCADA data is 10 minutes of data, the range of m is [4-17], which means that the data from 0.4 hours to 3 hours before the relevant parameters are used to predict the m+1 number of sensors under study. Train the SOA-LSTM model. After the training is completed, save the model.

[0192] 7) Model testing: Based on the trained model, test the model performance using test data, calculate the RMSE value of the predicted value and the actual value of the test data obtained for each different m, and select the m and corresponding model that minimize the RMSE.

[0193] Fault data reconstruction process: According to the test of the SOA-LSTM model, select the appropriate m value, take the anemometer fault data and stator winding 2 temperature sensor fault data analyzed above, and use the SOA-LSTM model to reconstruct the fault data. The specific process is as follows.

[0194] 1) Select the fault data of the sensor under study and the normal data of related parameters, and mark the normal data of related parameters as (x1, x2, ..., x k ).

[0195] 2) Construction of fault data time series: The input model time series sample consists of km cells, denoted as (X1, X2, …, X k-m ). Among them, X1, X2…X k-m The matrix expression of is similar to equations (45)-(47).

[0196] 3) The input data time series is normalized according to formula (48).

[0197] 4) Using the trained SOA-LSTM model, the input fault data (X1, X2, …, X k-m ) is used for prediction, and the model prediction result is recorded as

[0198] 5) The prediction results are taken as fault reconstruction data and saved as input data of the system.

[0199] The specific algorithm example of this embodiment is:

[0200] Stator winding temperature sensor fault based on FISSA-DBN: Based on the SCADA data of the wind turbine stator temperature sensor state recognition model, the program runs in the Matlab R2019a environment. The sample data selects 20,000 historical normal data in 2017, and 80% of the samples are randomly selected as model training data, and the remaining samples are used as test data. In order to facilitate accurate analysis, the research data selects the stator winding 2 temperature measurement value in 2013 when the fault occurred before, that is, data points 39400-40100, and all data are normalized. The number of input layer nodes of the DBN model is 9, the number of output layer nodes is 1, the number of iterations is set to 10, the batch size is set to 2, the momentum is set to 0, the FISSA algorithm population size is set to 40, and the number of iterations is set to 10. The model is trained and the FISSA algorithm is used to find the optimal fitness value. The values ​​of α, β and λ in formula (14) are 4, 82 and 0.0215 respectively.

[0201] The SCADA data of wind turbine No. 19 in 2013 was verified, and the selected data was 2013-10-02T14:55~2013-10-07T11:35, that is, data points 39400~40100. After prediction by FISSA-DBN model, Fig. 9 For stator winding 2 temperature sensor fault status monitoring δ E Graph. Fig. 9 The first occurrence of δ at 39409 E Above the upper limit of the alarm threshold, it indicates that the temperature of stator winding 2 is abnormal. The subsequent curve fluctuates around the alarm threshold until point 39704. E It starts to rise straightly, indicating that the temperature fault of stator winding 2 is obvious. At 40002, i.e. 2013-10-06T19:15, δ E By checking the records, the wind turbine was shut down at that time, and it can be seen that the initial alarm signal appeared more than 4 days earlier than the actual shutdown time.

[0202] The SCADA data of wind turbine No. 7 in 2014 was verified, and the selected data was 2014-05-07T12:25~2014-09-03T17:55, that is, data points 18001~34000. After prediction by FISSA-DBN model, Fig.10 The stator winding 6 temperature sensor fault status monitoring δ E Graph. Fig.10 At 19819, δ E Above the upper limit of the alarm threshold, it indicates that the temperature fault of stator winding 6 is obvious. At 30498 points, that is, 2014-08-10T10:15, δ EBy checking the records, the actual downtime of the wind turbine was 2014-06-02T17:55, that is, data point 20609. It can be seen that the alarm signal appeared more than 5 days earlier than the actual downtime.

[0203] Reconstruction of stator winding temperature sensor fault data based on SOA-LSTM: By setting different m values, the RMSE value of each test data is obtained. According to the RMSE value, the m corresponding to the minimum value is taken to form a time series. For the reconstruction of stator winding temperature sensor fault data, the input parameters are selected from the 9 parameters in Table 1. Assuming that the stator winding 2 temperature sensor fails, the input parameters are selected from 10,000 groups of processed SCADA data, of which 8,000 groups are selected for training data and 2,000 groups are selected for test data. The training and test input and output parameter data matrices are set according to equations (45)-(47). The SOA-LSTM model is trained and tested, and the corresponding RMSE values ​​when different m are taken are shown in Table 2. It can be seen from Table 2 that when m is 12, the corresponding RMSE value is the smallest, so the reconstruction accuracy obtained by using the first 12 data to predict the 13th data is higher.

[0204] Table 2 RMSE values ​​corresponding to different m

[0205] m-value RMSE m-value RMSE 4 1.0787 5 1.5627 6 1.1617 7 1.1149 8 1.8869 9 1.6802 10 1.8124 11 1.2154 12 1.0656 13 1.9259 14 1.1308 15 1.9294 16 1.5219 17 1.1245

[0206] The SOA-LSTM model is used to reconstruct the fault data of the stator winding 2 temperature sensor of the No. 19 wind turbine in 2013. The data segment is 38000-42000. The value of m is 12, that is, the 38000-38011 of the 9 input parameters are used to predict the 38012th number of the stator winding 2 temperature sensor, the 38001-38012 of the input parameters are used to predict the 38013th number of the stator winding 2 temperature sensor, and so on, until the 41988-41999 of the input parameters predict the 42000th number, constructing a 3988×1 cell input array, each cell consists of a 9×12 matrix. After the trained SOA-LSTM model data is reconstructed, the results are as follows Fig.11 As shown in the figure, the difference between the reconstructed value and the fault value is Fig.12 shown.

[0207] A temperature sensor fault diagnosis and fault data reconstruction system comprises: a data acquisition module, which is used to obtain a historical data set of a wind turbine, wherein the historical data set comprises historical operating parameters of the wind turbine, equipment status parameters of the wind turbine and historical parameters of the external environment; a fault diagnosis module, which is used to construct a temperature sensor fault diagnosis model according to a DBN deep belief network, train the temperature sensor fault diagnosis model based on the historical data set, optimize the hyperparameters of the temperature sensor fault diagnosis model based on the firefly improved sparrow search algorithm during the model training process, and obtain an optimized temperature sensor fault diagnosis model; and optimize the wind turbine based on the optimized temperature sensor fault diagnosis model. The stator winding temperature sensor is used for fault diagnosis and the sensor fault diagnosis result is output; when the sensor fault diagnosis result is that a fault exists, the fault data of the time period corresponding to the fault state is selected from the historical data set; a data reconstruction module is used to construct a fault data reconstruction model according to the LSTM long short-term memory network, and the fault data reconstruction model is trained based on the historical data set and the fault data. During the model training process, the hyperparameters of the temperature sensor fault diagnosis model are optimized based on the Seagull optimization algorithm to obtain the optimized fault data reconstruction model; based on the optimized fault data reconstruction model, the sensor fault data of the stator winding temperature sensor of the wind turbine generator set is reconstructed.

[0208] An electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute a temperature sensor fault diagnosis and fault data reconstruction method.

[0209] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for diagnosing faults of a temperature sensor and reconstructing fault data is implemented.

[0210] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for temperature sensor fault diagnosis and fault data reconstruction, characterized in that: include: S1: Acquire a historical data set of a wind turbine generator set, wherein the historical data set includes historical operating parameters of the wind turbine generator set, equipment status parameters of the wind turbine generator set, and historical parameters of the external environment; S2: construct a temperature sensor fault diagnosis model based on the DBN deep belief network, train the temperature sensor fault diagnosis model based on the historical data set, optimize the hyperparameters of the temperature sensor fault diagnosis model based on the firefly improved sparrow search algorithm during the model training process, and obtain an optimized temperature sensor fault diagnosis model; perform fault diagnosis on the stator winding temperature sensor of the wind turbine based on the optimized temperature sensor fault diagnosis model, and output the sensor fault diagnosis result; S3: When the sensor fault diagnosis result is greater than a preset threshold, a fault exists, and fault data of a period corresponding to the fault state is selected from the historical data set; S4: constructing a fault data reconstruction model based on an LSTM long short-term memory network, training the fault data reconstruction model based on the historical data set and the fault data, optimizing the hyperparameters of the temperature sensor fault diagnosis model based on the Seagull optimization algorithm during the model training process, and obtaining an optimized fault data reconstruction model; S5: Reconstruct the sensor fault data of the wind turbine stator winding temperature sensor based on the optimized fault data reconstruction model.

2. A temperature sensor fault diagnosis and fault data reconstruction method according to claim 1, characterized in that: It also includes performing data cleaning, correlation analysis and standardization on the data in the historical data set in sequence to obtain a preprocessed historical data set, and training the temperature sensor fault diagnosis model and the fault data reconstruction model based on the preprocessed historical data set.

3. A temperature sensor fault diagnosis and fault data reconstruction method according to claim 1, characterized in that: The step S2 specifically includes: Dividing the historical data set into a training set, a test set, and a validation set according to a preset ratio; Acquire several hyperparameter combinations of the temperature sensor fault diagnosis model as diagnosis model hyperparameter combinations; Based on the training set, the temperature sensor fault diagnosis model corresponding to each of the diagnosis model hyperparameter combinations is trained, and the performance of the temperature sensor fault diagnosis model corresponding to each of the diagnosis model hyperparameter combinations is verified and evaluated through the validation set, and the evaluation result is used as the evaluation score corresponding to each of the diagnosis model hyperparameter combinations; Constructing a sparrow population for sparrows with each of the hyperparameter combinations of the diagnostic model, updating and iterating the sparrow population based on the evaluation scores of each of the sparrows in combination with the firefly improved sparrow search algorithm, and outputting the optimal hyperparameter combination of the diagnostic model; An optimized temperature sensor fault diagnosis model is constructed based on the optimal diagnosis model hyperparameter combination.

4. A temperature sensor fault diagnosis and fault data reconstruction method according to claim 3, characterized in that: Based on the evaluation scores of the sparrows, the sparrow population is updated and iterated in combination with the firefly improved sparrow search algorithm, specifically including: S11: Randomly initialize the position of the sparrow population; S12: updating the predator position, the joiner position and the guard position, calculating the evaluation score corresponding to each of the sparrows, and updating the position of each sparrow in the sparrow population based on the calculated evaluation score; S13: updating the position of the sparrows through the firefly disturbance, calculating the evaluation score corresponding to each of the sparrows after the firefly disturbance, and updating the position of each sparrow in the sparrow population based on the calculated evaluation score; Repeat steps S11 to S13 to perform iterative updates until the iterative update reaches a preset number of iterations, and output the optimal diagnostic model hyperparameter combination.

5. A temperature sensor fault diagnosis and fault data reconstruction method according to claim 1, characterized in that: The step S4 specifically includes: Dividing the historical data set and the fault data into a training set, a test set and a validation set according to a preset ratio; Acquire several hyperparameter combinations of the fault data reconstruction model as reconstruction model hyperparameter combinations; Based on the training set, the fault data reconstruction model corresponding to each of the reconstruction model hyperparameter combinations is trained, and the performance of the fault data reconstruction model corresponding to each of the reconstruction model hyperparameter combinations is verified and evaluated through the verification set, and the evaluation result is used as the evaluation score corresponding to each of the reconstruction model hyperparameter combinations; Constructing a seagull population for seagulls using each of the reconstruction model hyperparameter combinations, updating and iterating the seagull population based on the evaluation scores of each of the seagulls in combination with a seagull optimization algorithm, and outputting an optimal reconstruction model hyperparameter combination; An optimized fault data reconstruction model is constructed based on the optimal reconstruction model hyperparameter combination.

6. A temperature sensor fault diagnosis and fault data reconstruction system, characterized in that: include: A data acquisition module is used to obtain a historical data set of the wind turbine generator set, wherein the historical data set includes historical operating parameters of the wind turbine generator set, equipment status parameters of the wind turbine generator set, and historical parameters of the external environment; A fault diagnosis module is used to construct a temperature sensor fault diagnosis model based on a DBN deep belief network, train the temperature sensor fault diagnosis model based on the historical data set, optimize the hyperparameters of the temperature sensor fault diagnosis model based on the firefly-improved sparrow search algorithm during the model training process, and obtain an optimized temperature sensor fault diagnosis model; perform fault diagnosis on the stator winding temperature sensor of the wind turbine based on the optimized temperature sensor fault diagnosis model, and output the sensor fault diagnosis result; when the sensor fault diagnosis result is that a fault exists, select fault data of the time period corresponding to the fault state in the historical data set; A data reconstruction module is used to build a fault data reconstruction model based on an LSTM long short-term memory network, train the fault data reconstruction model based on the historical data set and the fault data, optimize the hyperparameters of the temperature sensor fault diagnosis model based on the Seagull optimization algorithm during the model training process to obtain an optimized fault data reconstruction model; and reconstruct the sensor fault data of the wind turbine stator winding temperature sensor based on the optimized fault data reconstruction model.

7. An electronic device, characterized in that: It comprises a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute a temperature sensor fault diagnosis and fault data reconstruction method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements a temperature sensor fault diagnosis and fault data reconstruction method as described in any one of claims 1 to 5.

Images