A fan fault recognition method and system integrating periodic characteristics and improved self-attention mechanism

Abstract

This invention discloses a wind turbine fault identification method and system incorporating periodic features and an improved self-attention mechanism. The method includes: preprocessing various data collected from the wind turbine under test, extracting the corresponding frequency domain information for each type of periodic data, and fusing it with aperiodic data; performing data offset prevention processing on the fused data, and then using it as input to an improved self-attention mechanism model; and outputting predictions of the occurrence and non-occurrence of different wind turbine faults from the improved self-attention mechanism model. The improved self-attention mechanism model adds a mean-dividing operation to the self-attention model. The method and system of this invention can detect multiple faults in a single run, greatly improving fault detection efficiency, and can globally fuse all feature information while mitigating data offset and improving prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine fault detection technology, and relates to a wind turbine fault identification method and system that incorporates periodic features and an improved self-attention mechanism. Background Technology

[0002] Ventilation fans, as a type of ventilation equipment, have a wide range of applications, such as jet fans in tunnels. Improving the safety and reliability of these fans, as well as reducing their maintenance costs, is of profound significance. Currently, an increasing number of technological solutions combine artificial intelligence with fan technology for status and fault monitoring. However, existing solutions often have some problems, such as: most existing methods can only detect one specific fault per run, making it difficult to detect multiple faults or any fault of the fan; they often do not use the frequency domain information of the input data, only obtaining local information and not global information; in addition, most existing methods do not consider the problem of data offset or address the impact of data offset on prediction accuracy. Summary of the Invention

[0003] The purpose of this invention is to provide a wind turbine fault identification method and system that incorporates periodic features and an improved self-attention mechanism. This method considers the frequency domain information of the input data, and through data preprocessing and the use of the improved self-attention mechanism, it enables the model to consider more global information during inference, effectively mitigating data bias. Furthermore, this method can detect multiple faults simultaneously with a single computation.

[0004] To address the aforementioned technical problems, this invention provides a wind turbine fault identification method incorporating periodic features and an improved self-attention mechanism, comprising:

[0005] The system continuously collects all available data from the wind turbine under test, which is divided into two main categories: periodic data and aperiodic data. For each type of periodic data, the system preprocesses to extract its corresponding frequency domain information and then fuses it with the aperiodic data.

[0006] The fused data is processed to prevent data shift, and then used as input to the improved self-attention mechanism model;

[0007] The improved self-attention mechanism model adds a division by mean operation to the self-attention model. Specifically, a set of parameter matrices to be trained is set, and the input data of the network in the self-attention model or the output data of the previous layer network are combined with the parameter matrices and then divided by the average value of the data. The output of the improved self-attention mechanism model is the state of different faults of the wind turbine, which are encoded in a "one-hot encoding method".

[0008] In the above technical solution, firstly, each type of data is uniformly grouped according to time sequence. For each type of periodic data, each group of data is subjected to odd extension or even extension and then discrete Fourier transform to calculate the amplitude, argument, and frequency of the transformed data. Thus, the amplitude data, argument data, and frequency data of each group of periodic data are obtained. The obtained periodic data is then fused with the grouped aperiodic data and merged according to the group number. Each group is merged to obtain a matrix, which consists of all the aperiodic data in the group and the amplitude data, argument data, and frequency data of each periodic data obtained above.

[0009] Furthermore, the sampling frequencies of various data should be kept consistent. If they are inconsistent, the frequency with the highest sampling frequency shall be used. For data missing at corresponding moments due to sampling frequencies lower than the highest sampling frequency, interpolation or other methods can be used to fill in the missing data.

[0010] Furthermore, the process of preventing data offset in the fused data specifically involves: obtaining a matrix data where each row corresponds to a different category of data, including various non-periodic data, amplitude data after preprocessing various periodic data, argument data after preprocessing various periodic data, and frequency data after preprocessing various periodic data; calculating the mean and standard deviation of each row, and then subtracting the corresponding mean from each element of each row and dividing by the corresponding standard deviation.

[0011] Furthermore, the improved self-attention mechanism model adds a mean-division operation to the self-attention model, specifically: dividing the input matrix or the output data of the previous layer network by... matrix, The size is N×k, and the parameter matrix to be trained is set. and The matrix size is k×q. The matrix size is q×1, where q needs to be greater than k; calculate The average of all elements in the matrix is Perform the mean division operation using the following formula:

[0012]

[0013] The resulting matrix is ​​then subjected to further self-attention calculations.

[0014] Furthermore, during the model training phase, the cross-entropy loss function is used to update each model parameter.

[0015] A wind turbine fault identification system incorporating periodic characteristics and an improved self-attention mechanism includes:

[0016] The data acquisition module is used to continuously collect various detectable data from the fan under test.

[0017] The data processing module is used to preprocess each type of periodic data to extract its corresponding frequency domain information and fuse it with the non-periodic data; the fused data is then processed to prevent data offset and used as input to the fault judgment module.

[0018] The fault judgment module is used to judge the fault status of the wind turbine based on the input data and an improved self-attention mechanism model. The improved self-attention mechanism model is a self-attention model that adds a mean-dividing operation.

[0019] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above.

[0020] The beneficial effects of this invention are:

[0021] The method of this invention can detect multiple faults in a single run, greatly improving fault detection efficiency. Furthermore, the method extracts the frequency domain information of the collected periodic parameter data and integrates it with other non-periodic parameter data. The overall data, after anti-data-shift processing, is input into an improved self-attention mechanism model for further processing. Due to the computational method of the improved self-attention mechanism model, it can naturally and globally integrate all feature information. Anti-data-shift processing calculates the mean and standard deviation, then performs normalization on the data, ensuring that each parameter is processed within the same range. This prevents inconsistent data ranges for different parameters caused by excessively large parameter ranges, thus mitigating data shift and improving prediction accuracy. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating the wind turbine fault identification method incorporating periodic features and an improved self-attention mechanism provided in this embodiment of the invention.

[0024] Figure 2 This is a schematic diagram of the structure of the improved self-attention mechanism model in an embodiment of the present invention; Detailed Implementation

[0025] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] During operation, the fan generates various status data, such as front bearing temperature, rear bearing temperature, end-side bearing temperature, horizontal vibration displacement, vertical vibration displacement, blade acceleration, total fan pressure, total fan pressure efficiency, static pressure, static pressure efficiency, airflow, outlet air velocity, ambient noise, and oil tank temperature. These parameters are collected by sensors, with the collection frequency determined by the sensors in the acquisition module. For example, if 1000 data points are collected per second, the collection frequency is 1000Hz. After the sensors collect the parameter data, the method and system of this invention are used to process it. Upon completion, the system outputs the faults present in the fan.

[0027] like Figure 1 The diagram shown is a flowchart of the model training phase of the method of the present invention, which consists of 6 steps.

[0028] Step A is the "Sensor Acquisition of Periodic and Aperiodic Data" step. In this step, the sensor collects data, which is divided into periodic data and aperiodic data. Whenever conditions permit, both types of data are collected. Periodic data includes things like the horizontal vibration displacement of the fan, the vertical vibration displacement of the fan, etc. Aperiodic data includes things like the temperature of the fan's front bearing, the temperature of the fan's front-side bearing, etc.

[0029] Step B is the "Fourier Transform of Periodic Data" step. In this step, the periodic data is subjected to a Fourier transform to obtain the frequency, amplitude, and phase angle corresponding to each type of periodic data.

[0030] Step C is "Preventing Data Offset Processing". In this step, some processing is performed on each data point to mitigate data offset.

[0031] Step D is "Improved Self-Attention Mechanism Model Processing". In this step, the data from step C or F is input into the "Improved Self-Attention Mechanism Model" for processing. In this step, this patent innovatively improves the self-attention mechanism model, specifically as follows: Figure 2 As shown, mean normalization is introduced into the self-attention model to mitigate data bias.

[0032] Step E is "calculating the loss function value". In this step, the data processed in step D is combined with the label values ​​in the training data to calculate the loss function value.

[0033] Step F is "updating model parameters". In this step, based on the loss function value calculated in step E, the derivative of each parameter is calculated, and then the updated model parameters are calculated. The execution continues from step D or terminates based on the set stopping conditions.

[0034] Once the model is trained, fault prediction can be performed.

[0035] Specifically, according to a specific embodiment of the present invention, the method may include the following steps:

[0036] Step 1: Collect all available parameters. Here, we assume the following parameters were collected. Define X1 as the front bearing temperature, X2 as the rear bearing temperature, X3 as the front-side bearing temperature, X4 as the horizontal vibration displacement, X5 as the vertical vibration displacement, X6 as the blade acceleration, X7 as the total pressure of the fan, X8 as the total pressure efficiency of the fan, X9 as the static pressure of the fan, and X... 10 Represents the static pressure efficiency of the fan, X 11 Represents fan air volume, X 12 Represents the fan outlet wind speed, X 13 Represents the sound around the wind turbine, X 14 The sample sequence collected for 14 parameters, including the fan oil tank temperature, is defined as Y, representing the fan state at the moment these 14 parameters were sampled. This state can be defined as various damaged or normal states, such as "bearing damaged," "motor damaged," etc., and is encoded using a one-hot encoding method. In this patent, four cases are used as examples: "bearing normal," "bearing faulty," "motor normal," and "motor faulty." According to the one-hot encoding method, whether the bearing is normal can be encoded in the first position, using 0 to represent normal bearing and 1 to represent a faulty bearing. Similarly, whether the motor is normal can be encoded in the second position, using 0 to represent normal motor and 1 to represent a faulty motor. Therefore, "bearing normal" and "motor normal" can be encoded as [0, 0], "bearing normal" and "motor faulty" as [0, 1], "bearing faulty" and "motor normal" as [1, 0], and "bearing faulty" and "motor faulty" as [1, 1].

[0037] in, This represents the data value of the m-th parameter collected at time t. (The last part, "and X", appears to be a typo and can be left as is.)m The state value of the wind turbine at each moment is Y = {y t The value range of yt is as described above, and it is encoded using one-hot encoding, i.e., one of the four types: [0, 0], [0, 1], [1, 0], and [1, 1]. Here, it is assumed that n data points are collected for each parameter.

[0038] Step 2: For X1, X2, ..., X 14 These 14 sample sequences were grouped according to sample order, with each group consisting of k samples, resulting in a total of [number missing]. The groups were assigned and numbered as Group 1, Group 2, ..., Group 3. For each group, k needs to be divisible by n. Y also needs to be grouped in this order.

[0039] After grouping, you can get

[0040]

[0041]

[0042] Where m takes values ​​from 1 to 14.

[0043] Step 3: For X4, X5, X6, and X 13 Each set of data is either odd-extended or even-extended before undergoing a discrete Fourier transform. The transform formula is as follows:

[0044]

[0045] Where X[l] represents the transformed sequence, l ranges from 1 to k, X[1] represents the first complex value of the transformed sequence, X[2] represents the second complex value of the transformed sequence, and so on. S represents the number of samples in each group, which is equal to k here. x[1] represents the first sample in the group, x[2] represents the second sample in the group, and so on. The frequency corresponding to X[l] f sample This corresponds to the sampling frequency. For example, for X4, f sample This refers to the sampling frequency of the horizontal vibration displacement. Here, the sampling frequency for each parameter should be consistent. If they are inconsistent, the frequency with the highest sampling frequency should be used. For data missing at corresponding moments due to sampling frequencies lower than the highest sampling frequency, interpolation or other methods can be used to fill in the missing data.

[0046] For example, for After performing a Fourier transform on this set of samples, the transformed sequence is as follows:

[0047]

[0048] After this step is completed, we can obtain X4, X5, X6, and X. 13 The sequences after performing the Discrete Fourier Transform are as follows:

[0049]

[0050] Meanwhile, the frequency corresponding to each sample after the discrete Fourier transform is

[0051]

[0052] The values ​​of r are 4, 5, 6, and 13.

[0053] Step 4: Based on the results calculated in Step 3

[0054]

[0055] The value of r is 4, 5, 6, or 13.

[0056] Each sample is a complex number; therefore, its amplitude and argument can be calculated. Let the calculated amplitude be...

[0057]

[0058] Argument is

[0059]

[0060] The value of r is 4, 5, 6, or 13.

[0061] Step 5: Based on the grouped data obtained in Step 2, using parameters 1, 2, 3, 7, 8, 9, 10, 11, and 12...

[0062]

[0063] Where m takes the values ​​1, 2, 3, 7, 8, 9, 10, 11, and 12.

[0064] The frequency data of parameters 4, 5, 6, and 13 obtained in step 3

[0065]

[0066] The value of r is 4, 5, 6, or 13.

[0067] The amplitude data of parameters 4, 5, 6, and 13 obtained in step 4.

[0068]

[0069] The value of r is 4, 5, 6, or 13.

[0070] The argument data of parameters 4, 5, 6, and 13 obtained in step 4.

[0071]

[0072] The value of r is 4, 5, 6, or 13.

[0073] Merge them according to their group numbers. Each group will be merged into a matrix. Let the merged result be denoted as...

[0074]

[0075] Where i is the group number, and in total Group.

[0076]

[0077] achievable The size of the matrix is ​​21×k.

[0078] Step 6: Set the batch_size to a certain value, here assumed to be 32, but this can be modified according to the actual situation. Calculation If total_step is a decimal, then round it up.

[0079] Step 7: Set step = 1.

[0080] Step 8: Remove There are a total of batch_size matrices.

[0081] Step 9: Calculate each matrix using the methods in Steps 10 and 11. In Steps 10 and 11, take... Taking matrices as an example, i is a matrix from step 8.

[0082] Step 10: Calculation The mean of each row, after calculation, is denoted as... calculate The standard deviation of each row, after calculation, is denoted as .

[0083] Step 11: Subtracting the mean of the row from each element of the matrix and then dividing by the variance of the row, the result is denoted as . The size of the matrix is ​​21×k.

[0084] Step 12: After steps 9 to 11 are completed, you can obtain... There are a total of batch_size matrices.

[0085] Step 13: Let n = 1

[0086] Step 14: Obtain the matrix generated in Step 12 There are a total of batch_size matrices, or, based on the result of step 20, the result of step 19 can be obtained. There are a total of batch_size matrices, which are renamed accordingly. If you haven't set it before and Then let's define a parameter matrix that needs to be trained. and The matrix size is k×q. The matrix size is q×1, where q must be greater than k. Otherwise, use the previously defined matrix. and

[0087] Step 15: Calculate the total batch_size matrices obtained in Step 12 using the following formula, where... Let's take an example. Calculation. The average of all elements in the matrix is

[0088]

[0089] Among them, due to Its size is q×1, and The calculated result is 21 × q, therefore, when calculating addition here, it is... Each row of the calculation results and Add the corresponding elements together. Therefore, after the calculation is complete, The matrix size is 21×q.

[0090] At this point, a total of [number] can be obtained. There are a total of batch_size matrices.

[0091] Step 16: Calculate the total batch_size matrices obtained in Step 15 using the following formula, where... Let's take an example.

[0092] set up calculate

[0093] A i =Softmax(Q i ×K i )

[0094]

[0095] Where Softmax is the Softmax function. Thus, we can obtain...

[0096] There are a total of batch_size matrices.

[0097] Step 17: Regarding step 14 The matrix and each batch_size matrix obtained in step 16 are calculated using the following formula, where... and Let's take an example.

[0098]

[0099] Here, LayerNorm refers to the Layer Normalization function.

[0100] At this point, a total of [number] can be obtained. There are a total of batch_size matrices.

[0101] Step 18: If you haven't set it before and Then set a parameter matrix that needs to be trained. and The matrix size is q×u. The matrix size is u×1, where u must be greater than q. Otherwise, use the previously defined matrix. and

[0102] The batch_size matrices obtained in step 17 are calculated using the following formula, where... Let's take an example.

[0103]

[0104] Among them, due to Its size is u×1, and The calculated result is 21 × u, therefore, when calculating addition here, it is... Each row of the calculation results and Add the corresponding elements together. Therefore, after the calculation is complete, The matrix size is 21×u.

[0105] At this point, a total of [number] can be obtained. There are a total of batch_size matrices.

[0106] Step 19: Calculate the batch_size matrices obtained in Steps 17 and 18 using the following formula, where... and Let's take an example.

[0107]

[0108] Here, LayerNorm refers to the Layer Normalization function.

[0109] At this point, a total of [number] can be obtained. There are a total of batch_size matrices.

[0110] Step 20: Define a variable `layer_num`, which can be any value, typically between 6 and 20. Set `n` to a value 1 greater than itself. If `n` is greater than `layer_num`, proceed to step 21; otherwise, repeat steps 14 through 19.

[0111] Step 21: After completing step 20, at the current step value, you will get Since the batch_size was set to 32 in step 6, a total of layer_num × 32 matrices can be obtained. Here, the value of n is in the range of 1 to layer_num.

[0112] Step 22: Define a parameter matrix to be trained. and The matrix size is u×1. The matrix size is 21×1. Next, define a parameter matrix that needs to be trained. and The matrix size is 21×2. The matrix size is 1×2.

[0113] Step 23: Obtain the results from Step 21 A total of batch_size matrices are generated, which was set to 32 in step 6 and is calculated using the following formula (here, we use...). Let's take an example.

[0114]

[0115]

[0116] Wherein, Sigmoid is the Sigmoid function. It is a 1×2 matrix.

[0117] According to the definition in step 1, The first element Let be the probability that the model predicts "bearing failure" for the wind turbine within the time period corresponding to the i-th data segment. If the probability is greater than 0.5, the model predicts that the bearing is faulty; otherwise, the model predicts that the bearing is normal. The second element Let be the probability that the model predicts "motor malfunction" for the wind turbine within the time period corresponding to the i-th data segment. If the probability is greater than 0.5, the model predicts that the motor is malfunctioning; otherwise, the model predicts that the motor is normal.

[0118] Thus, we can obtain There are a total of batch_size matrices of size 1×2, which were set to 32 in step 6.

[0119] Step 24: Obtain and The label Y corresponding to the time (step-1)*batch_size+1 Y (step-1)*batch_size+2 ...Y (step-1)*batch_size+batch_size .

[0120] Step 25: Based on the results generated in Steps 23 and 24 and Y (step-1)*batch_size+1 Y (step-1)*batch_size+2 ...Y (step-1)*batch_size+batch_size Calculate using the following formula. Here we take... and Y i Let's take an example.

[0121]

[0122] CrossEntropy is the cross-entropy loss function, which is used to calculate... and Y i The cross-entropy loss value.

[0123] At this point, the loss can be obtained. 1 loss 2 ...loss 32 There are a total of batch_size values, which were set to 32 in step 6.

[0124] Step 26: Apply the loss obtained in step 25 1 loss 2 ... loss 32 There are a total of 32 values, which are calculated using the following formula.

[0125]

[0126] In step 6, batch_size has been set to 32.

[0127] Step 27: Obtain the result calculated in Step 26, and apply it to the creation made in Step 14. and The creation in step 18 and The creation in step 22 and and Where n has a value from 1 to layer_num, and the gradient of the loss with respect to them is calculated according to the following formula.

[0128]

[0129] Step 28: Based on the gradient results calculated in Step 27, update the values ​​of each independent variable as follows:

[0130]

[0131]

[0132]

[0133]

[0134]

[0135]

[0136]

[0137]

[0138] Where i ranges from 1 to layer_num. β is the learning rate, which can be changed according to the actual situation, and is generally set to 0.002.

[0139] Step 29: Set step to a value one greater than itself. If step is greater than total_step, then execute step 30; otherwise, execute steps 8 to 28.

[0140] Step 30: Repeat steps 7 to 29 for a total of epochs. The epoch value can be set according to the actual situation, generally set to 5-10.

[0141] Step 31: At this point, the training phase is complete, and you can now obtain the trained data. Save it.

[0142] It is important to note during the prediction phase that the selection of the k value in step 2 should be consistent with that in the training phase.

[0143] It can be obtained in the prediction stage. Here, n represents the total number of samples collected. The result allows direct reading of the fan's status. For example, in this example, based on the definition in step 1, The first element Let be the probability that the model predicts "bearing failure" for the wind turbine within the time period corresponding to the i-th data segment. If the probability is greater than 0.5, the model predicts that the bearing is faulty; otherwise, the model predicts that the bearing is normal. The second element Let be the probability that the model predicts "motor malfunction" for the wind turbine within the time period corresponding to the i-th data segment. If the probability is greater than 0.5, the model predicts that the motor is malfunctioning; otherwise, the model predicts that the motor is normal.

[0144] The prediction phase has now ended.

[0145] This example also provides a wind turbine fault identification system incorporating periodic features and an improved self-attention mechanism to implement the aforementioned fault identification method, including:

[0146] The data acquisition module is used to continuously collect various detectable data from the fan under test.

[0147] The data processing module is used to preprocess each type of periodic data to extract its corresponding frequency domain information and fuse it with the non-periodic data; the fused data is then processed to prevent data offset and used as input to the fault judgment module.

[0148] The fault judgment module is used to judge the fault status of the wind turbine based on the input data and an improved self-attention mechanism model. The improved self-attention mechanism model is a self-attention model that adds a mean-dividing operation.

[0149] The above modules are used to implement steps 1-2, 3-12 and 7-23 in the above fault identification method, respectively. Therefore, the specific implementation methods can be referred to the description of each part of the corresponding embodiments, which will not be repeated here.

[0150] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A wind turbine fault identification method incorporating periodic features and an improved self-attention mechanism, characterized in that, include: The system continuously collects all available data from the wind turbine under test, which is divided into two main categories: periodic data and aperiodic data. For each type of periodic data, the system preprocesses to extract its corresponding frequency domain information and then fuses it with the aperiodic data. The fused data is processed to prevent data shift, and then used as input to the improved self-attention mechanism model; The process of preventing data offset in the fused data specifically involves: fused matrix data, where each row corresponds to a different category of data, including various non-periodic data, amplitude data after preprocessing various periodic data, argument data after preprocessing various periodic data, and frequency data after preprocessing various periodic data; calculating the mean and standard deviation of each row, and then subtracting the corresponding mean from each element of each row and dividing by the corresponding standard deviation; The improved self-attention mechanism model adds a mean-division operation to the self-attention model. Specifically, this involves dividing the input matrix or the output data of the previous layer network by a mean-division operation. matrix, The size is N×k, and the parameter matrix to be trained is set. and , The size of the matrix is , The size of the matrix is , Need to be greater than ;calculate The average of all elements in the matrix is Perform the mean division operation using the following formula: ， The resulting matrix is ​​then subjected to further self-attention calculations. The output of the improved self-attention mechanism model represents the occurrence and non-occurrence of different faults in the wind turbine, and is encoded using a "one-hot encoding method".

2. The wind turbine fault identification method incorporating periodic features and an improved self-attention mechanism according to claim 1, characterized in that, First, the collected data of each type are uniformly grouped according to time sequence. For each type of periodic data, each group of data is subjected to odd extension or even extension and then discrete Fourier transform to calculate the amplitude, phase angle and frequency of the transformed data. Thus, the amplitude data, phase angle data and frequency data of each group of periodic data are obtained. The obtained periodic data are then merged with the grouped aperiodic data and merged according to the group number. Each group is merged to obtain a matrix. The matrix consists of all the aperiodic data of the group and the amplitude data, phase angle data and frequency data of each periodic data obtained above.

3. The wind turbine fault identification method incorporating periodic features and an improved self-attention mechanism according to claim 2, characterized in that, The sampling frequency of all data should be consistent. If they are inconsistent, the highest sampling frequency shall be used. For the missing data at the corresponding time when the sampling frequency is lower than the highest sampling frequency, data filling shall be performed.

4. The wind turbine fault identification method incorporating periodic features and an improved self-attention mechanism according to claim 1, characterized in that, During the model training phase, the cross-entropy loss function is used to update each model parameter.

5. A wind turbine fault identification system incorporating periodic characteristics and an improved self-attention mechanism, characterized in that, To implement the method as described in any one of claims 1-4, comprising: The data acquisition module is used to continuously collect various detectable data from the fan under test. The data processing module is used to preprocess each type of periodic data to extract its corresponding frequency domain information and fuse it with the non-periodic data; the fused data is then processed to prevent data offset and used as input to the fault judgment module. The fault judgment module is used to judge the fault status of the wind turbine based on the input data and an improved self-attention mechanism model. The improved self-attention mechanism model is a self-attention model that adds a mean-dividing operation.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1-4.

Images