Fan fault diagnosis method based on image information fusion technology

Abstract

The invention discloses a fan fault diagnosis method based on an image information fusion technology, and the method comprises the steps: S1, obtaining vibration data when a fan works, and segmenting the vibration data to obtain a sample data set; s2, substituting the sample data set into a fan vibration definition algorithm to obtain an index data set corresponding to vibration data; s3, performing data conversion and segmentation on the index data set to obtain a corresponding grayscale image data set; s4, dividing the grayscale image data set into a training set, a verification set and a test set in proportion; s5, constructing a convolutional neural network model, and performing multiple times of training and verification on the convolutional neural network model through the training set and the verification set to obtain a plurality of training models; s6, evaluating a plurality of training models one by one through an accuracy and loss function to obtain an optimal training model; and S7, testing the optimal training model through the test set to obtain a fault diagnosis result of the fan. The finally trained model can perform fault diagnosis on the fan through the data of the test set, and a fan fault diagnosis result with relatively high accuracy is obtained.

Description

Technical Field

[0001] The invention belongs to the technical field of fan fault diagnosis, and in particular relates to a fan fault diagnosis method based on image information fusion technology. Background Art

[0002] Fans are machines that rely on input mechanical energy to increase gas pressure and discharge gas. They are widely used in industrial production environments such as factories, mines, cooling towers, etc. For example, ventilation, dust removal and cooling of construction sites, ventilation and draft of boilers and industrial furnaces, drying and selection of grains, etc. Therefore, fault diagnosis of fans is particularly important in industrial production. The health of fans is a key factor in ensuring the normal operation of industrial production. Once a fan fails, it will affect the processing quality. In serious cases, it will cause the entire industrial equipment to shut down and cause economic losses.

[0003] Early fault diagnosis methods based on analysis of fan vibration signals, such as the Chinese invention patent with announcement number CN113269221B, which discloses a rolling bearing fault diagnosis method based on a multi-scale convolutional neural network, which extracts envelope order spectrum as a feature for fault identification through variational mode decomposition and energy operator. For example, the Chinese invention patent with announcement number CN114077847B discloses a variable working condition rolling bearing fault diagnosis method, system, equipment and medium, which proposes to optimize the parameter combination of VMD through the SHO optimization algorithm, extract the permutation entropy, singular value feature, and sample entropy feature of the signal after noise reduction for feature fusion as the final feature parameter. In addition to the schemes of the above two invention patents, there are also methods such as extracting signal features of the original vibration signal in the time domain, frequency domain and time-frequency domain, such as based on Fourier transform, empirical mode decomposition, integrated empirical mode decomposition and other methods to realize fault analysis and classification of the collected original vibration signal of the fan, so that the vibration signal features are easier to be identified and fault diagnosis is realized.

[0004] However, with the rise of deep learning, the mode of training convolutional neural network models will gradually replace the traditional fan fault diagnosis method. For example, the Chinese invention patent application with publication number CN114897029A discloses a method for integrated stove fan fault diagnosis based on a deep residual contraction network, which uses multiple vibration sensors to collect vibration signal data and uses the trained deep residual contraction network to detect the integrated stove fan fault category. However, since convolutional neural networks require a large amount of data support, the dimensional relationship between the data will cause the training model to overfit, that is, even if the convolutional neural network performs very well on the training set data, it will cause the accuracy of the test data to decrease, and ultimately fail to bring into play the superiority of convolutional neural networks for fan fault diagnosis. Summary of the invention

[0005] The purpose of the present invention is to provide a fan fault diagnosis method based on image information fusion technology, which converts fan data into image data to train a convolutional neural network model, and in the process of data conversion, the dimensional relationship between the data used in the training model can be eliminated. Finally, the fan fault diagnosis result of the trained model will be more accurate.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is as follows:

[0007] A fan fault diagnosis method based on image information fusion technology includes the following steps:

[0008] Step S1. Obtain vibration data of the fan when it is working, and segment the vibration data to obtain a sample data set;

[0009] Step S2. Substitute the segmented data of the sample data set into the fan vibration definition algorithm to obtain an index data set corresponding to the vibration data;

[0010] Step S3. Convert and segment the index data set to obtain corresponding grayscale image data;

[0011] Step S4. Divide the grayscale image dataset into a training set, a validation set, and a test set in proportion;

[0012] Step S5. Construct a convolutional neural network model, and train and verify it multiple times through training sets and verification sets to obtain multiple training models;

[0013] Step S6. Evaluate multiple training models one by one through accuracy and loss function to obtain the best training model;

[0014] Step S7: Test the best training model using the test set to obtain the fault diagnosis result of the wind turbine.

[0015] The present invention discloses a fan fault diagnosis method based on image information fusion technology. The method first segments the vibration data of the fan when it is working, then eliminates the dimensional relationship between the data by defining an algorithm, and then converts the data into image data to train a convolutional neural network model. In the process of data conversion, the dimensional relationship between the data used in the training model can be eliminated. Finally, the trained model can diagnose the fan fault through the data of the test set, and obtain a fan fault diagnosis result with higher accuracy.

[0016] Preferably, the step S1 comprises:

[0017] Step S1.1. When the fan is working, the vibration signal of the fan is collected by the sensor to obtain corresponding vibration data;

[0018] Step S1.2. Define a sliding window and its size through a sliding window algorithm, and segment the vibration data through a sliding window of this size to obtain a segmented sample data set.

[0019] Preferably, the fan vibration definition algorithm in step S2 includes a dimensionless index expression, a peak index expression, a waveform index expression, and a pulse index expression.

[0020] Preferably, the step S2 comprises:

[0021] Step S2.1. Let the sample data of the sample data set be x, obtain the corresponding probability density function p(x), and set the hyperparameters l and m to obtain the dimensionless index Expression of

[0022] Step S2.2. According to the dimensionless index The expression is to assign different values ​​to the hyperparameters l and m in turn, and get the peak index C f Expression, waveform index S f Expression of pulse index I f Expression of

[0023] Step S2.3. Substitute the segmented data in the sample data set into the peak index C in sequence f Expression, waveform index S f Expression of pulse index I f Calculate the index data x and substitute it into the dimensionless index The expression gets the corresponding multiple indicator data sets

[0024] Preferably, the dimensionless index The expression is as follows:

[0025]

[0026] When l→∞, m=2, the peak index C f The expression is as follows:

[0027]

[0028] Among them, X max Indicates the maximum value of sample data, X rms Represents the root mean square value of sample data;

[0029] When l=2, m=1, the waveform index S f The expression is as follows:

[0030]

[0031] Among them, X rms represents the root mean square value of the sample data, Represents the average amplitude of sample data;

[0032] When l→∞, m=1, the pulse index I f The expression is as follows:

[0033]

[0034] Preferably, the step S3 comprises: converting the plurality of indicator data sets into After data segmentation through the sliding window, the grayscale image data is converted using the grayscale image conversion formula to obtain the grayscale image data set as follows:

[0035] T x ={{T1,T2,···,T m},{T m ,T m+1 ,···,T 2m-1},···,{T 2m ,T 2m+1 ,···,T n}};

[0036] The grayscale image conversion formula is as follows:

[0037]

[0038] in, Indicates the minimum value in the indicator data. Indicates the maximum value in the indicator data.

[0039] Preferably, in step S4, the grayscale image dataset is divided into a training set, a validation set and a test set in a ratio of 0.7:0.2:0.1.

[0040] Preferably, step S5 includes: inputting the data of the training set to train the convolutional neural network model multiple times, and in each training iteration, verifying and evaluating the training results of the previous training through the verification set, and obtaining the best model based on the comparison of the verification and evaluation results.

[0041] Preferably, step S6 includes: performing model evaluation on the convolutional neural network model after each training through accuracy and loss function, and saving the best training model.

[0042] Preferably, step S7 comprises: after inputting the data of the test set into the optimal training model to obtain the fan diagnosis result, the fan diagnosis result and the diagnosis accuracy are analyzed by a normalized exponential function.

[0043] Beneficial effects:

[0044] The present invention discloses a fan fault diagnosis method based on image information fusion technology. The method first segments the vibration data of the fan when it is working, then eliminates the dimensional relationship between the data by defining an algorithm, and then converts the data into image data to train a convolutional neural network model. In the process of data conversion, the dimensional relationship between the data used in the training model can be eliminated. Finally, the trained model can diagnose the fan fault through the data of the test set, and obtain a fan fault diagnosis result with higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Shown is a general flow chart of a fan fault diagnosis method based on image information fusion technology according to an embodiment;

[0046] Figure 2 The figure shows a sub-flow chart of a fan fault diagnosis method based on image information fusion technology in an embodiment;

[0047] Figure 3 It is a schematic diagram showing a fan fault diagnosis method based on image information fusion technology in an embodiment for segmenting and dimensionless calculation of fan vibration data;

[0048] Figure 4 Shown is a schematic diagram of converting indicator data into grayscale image data according to an embodiment;

[0049] Figure 5 Shown is a schematic diagram of the trend of the model training accuracy and loss function of the training set and the validation set of the embodiment;

[0050] Figure 6 Shown is a schematic diagram of training a convolutional neural network model using grayscale image data according to an embodiment.

[0051] Reference numerals

[0052] 100, model training accuracy of the training set; 200, model training accuracy of the validation set; 300, loss function of the validation set; 400, loss function of the training set. DETAILED DESCRIPTION

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings and other implementation methods can be obtained based on these drawings without creative work.

[0054] The technical solution of the present invention is described in detail below with specific embodiments.

[0055] Example

[0056] like Figures 1 to 4 As shown, a fan fault diagnosis method based on image information fusion technology in this embodiment includes the following steps:

[0057] Step S1. Obtain vibration data of the fan when it is working, and segment the vibration data to obtain a sample data set;

[0058] Step S2. Substitute the segmented data of the sample data set into the fan vibration definition algorithm to obtain an index data set corresponding to the vibration data;

[0059] Step S3. Convert and segment the index data set to obtain a corresponding grayscale image data set;

[0060] Step S4. Divide the grayscale image dataset into a training set, a validation set, and a test set in proportion;

[0061] Step S5. Construct a convolutional neural network model, and train and verify it multiple times through training sets and verification sets to obtain multiple training models;

[0062] Step S6. Evaluate multiple training models one by one through accuracy and loss function to obtain the best training model;

[0063] Step S7: Test the best training model using the test set to obtain the fault diagnosis result of the wind turbine.

[0064] Preferably, step S1 comprises:

[0065] Step S1.1. When the fan is working, the vibration signal of the fan is collected by the sensor to obtain corresponding vibration data;

[0066] Step S1.2. Define a sliding window and its size through a sliding window algorithm, and segment the vibration data through a sliding window of this size to obtain a segmented sample data set.

[0067] Specifically, this embodiment adopts ten types of fan end vibration signals, and the signal data length is 119808. i ={x1,x2,···,x n} (i is the data representation category, n is the data length) to perform data segmentation. Data segmentation can effectively increase the number of samples and make full use of the fan fault data for analysis. According to the defined sliding window sliding value, the sliding window length is set to 50 and the overlap rate is set to 0.9. The sample data set obtained after segmentation is X i ={{x1,···,x 50}, {x 51 ,···,x 100},···{x n-49 ,···,x n}}(i is the data representation category, n is the data length) and the dimensionless index data is calculated once every 50 original data.

[0068] Specifically, the length of the sliding window of this embodiment is set to ω, and the sliding step s is set according to the sample data overlap rate. For example, if the repetition rate of the current sample and the next sample value is τ, the sliding step of the sliding window is set according to s=w*(1-τ).

[0069] Specifically, the vibration signal data of this embodiment is collected by an accelerometer sensor, such as a CWRU data set, which is collected by an acceleration sensor installed near a wind turbine bearing and records the normal operation status of the bearing and different types of bearing fault status.

[0070] Preferably, the fan vibration definition algorithm in step S2 includes a dimensionless index expression, a peak index expression, a waveform index expression, and a pulse index expression.

[0071] Preferably, step S2 comprises:

[0072] Step S2.1. Let the sample data of the sample data set be x, obtain the corresponding probability density function p(x), and set the hyperparameters l and m to obtain the dimensionless index Expression of

[0073] Step S2.2. According to the dimensionless index The expression is to assign different values ​​to the hyperparameters l and m in turn, and get the peak index C f Expression, waveform index S f Expression of pulse index I f Expression of

[0074] Step S2.3. Substitute the segmented data in the sample data set into the peak index C in sequence f Expression, waveform index S f Expression of pulse index I f Expression to obtain the corresponding multiple indicator data sets

[0075] Specifically, the peak index, waveform index and pulse index can meet the massive data required for deep model learning. Substituting the sample data set X into the three formulas i ={{x1,···,x 50}, {x 51 ,···,x 100},···{x n-49 ,···,x n} (i represents the category, n represents the data length), the indicator data set can be obtained (j is the indicator category, m is the data length), indicator data can effectively reduce the dimensional connection between data.

[0076] Preferably, dimensionless index The expression is as follows:

[0077]

[0078] When l→∞, m=2, the peak index C f The expression is as follows:

[0079]

[0080] Among them, X max Indicates the maximum value of sample data, X rms Represents the root mean square value of sample data;

[0081] The peak value index C of this embodiment f The expression can be substituted into the sample data to obtain the peak index of the sample data, thereby eliminating the dimensional connection of the sample data.

[0082] When l = 2, m = 1, the waveform index S f The expression is as follows:

[0083]

[0084] Among them, X rms represents the root mean square value of the sample data, Represents the average amplitude of sample data;

[0085] The waveform index S of this embodiment f The expression can be substituted into the sample data to obtain the waveform indicators of the sample data, thereby eliminating the dimensional connection of the sample data.

[0086] When l→∞, m=1, the pulse index I f The expression is as follows:

[0087]

[0088] The pulse index I of this embodiment f The expression can be substituted into the sample data to obtain the pulse index of the sample data, thereby eliminating the dimensional connection of the sample data.

[0089] Specifically, the dimensionless index is obtained by the ratio of two dimensional indicators. The peak index describes the signal fluctuation range, the waveform index describes the signal operation status, and the pulse index describes the pulse signal characteristics.

[0090] Specifically, the setting of l and m is based on which dimensioned index value is to be solved in the sample data. For example, when solving the average amplitude in the sample data, l = 1 (or m = 1); when solving the root mean square value in the sample data, l = 2 (or m = 2); when solving the maximum value in the sample data, l (or m) tends to infinity.

[0091] Specifically, the ratio of the peak value of the sample data divided by the root mean square value of the sample data is a dimensionless value, which represents the relationship between the amplitude in the sample data and the overall sample data, and can therefore be used to evaluate the peak characteristics of the signal. According to the dimensionless expression formula, when l approaches infinity, that is, the denominator represents the limiting behavior of the weighted average of x and its corresponding probability density function, the expression for solving the data peak value is obtained; the root mean square value is equal to the expected square root of the square of the variable x, and through the method of implementing the probability density function, only m=2 is required. Similarly, the peak value of the sample data is divided by the average amplitude of the sample data, and the formula for calculating the average amplitude using the probability density function is obtained by simply setting l or m to 1 in the dimensionless formula to obtain the formula obtained by multiplying x with its corresponding probability density function. Similarly, the dimensionless value obtained by dividing the root mean square value of the sample data by the average amplitude of the data is handled in the same way.

[0092] Furthermore, the peak value of the sample data is divided by the average amplitude of the sample data. This value represents the relationship between the amplitude in the sample data and the overall average level. Since the peak value can reflect the signal impact strength and the average amplitude reflects the average level of the signal, the higher the ratio, the more obvious the extreme value in the signal and the greater the signal pulse characteristics. This ratio can reflect the pulse characteristics and is therefore a pulse indicator; the root mean square value of the sample data is divided by the average amplitude of the sample data. Since the root mean square value can reflect the data vibration strength and the average amplitude reflects the average level of the data, the ratio can reflect the signal waveform characteristics including the degree of deviation, signal stability and other signal characteristics and is therefore a waveform indicator. If the ratio is more obvious (data deviation), then the amplitude frequency and other values ​​of this wave are also obvious, so there is no need to analyze the amplitude frequency of the wave separately. The waveform characteristics of the sample data can be analyzed through the ratio.

[0093] Preferably, step S3 includes: After data segmentation through the sliding window, the grayscale image data is converted using the grayscale image conversion formula to obtain the grayscale image data set as follows:

[0094] T x ={{T1,T2,···,T m},{T m ,T m+1 ,···,T 2m-1},···,{T 2m ,T 2m+1 ,···,Tn}};

[0095] The grayscale image conversion formula is as follows:

[0096]

[0097] in, Indicates the minimum value in the indicator data. Indicates the maximum value in the indicator data.

[0098] Specifically, this embodiment can enrich the feature data details and feature data contrast through the grayscale image data conversion process, and is simpler to implement and more conducive to model learning.

[0099] Preferably, in step S4, the grayscale image dataset is divided into a training set, a validation set and a test set in a ratio of 0.7:0.2:0.1.

[0100] Specifically, the indicator dataset (j is the indicator category, m is the data length) is divided into data through sliding windows (the window length is set to 1024, and the overlap rate is set to 0.9), and then converted into grayscale image data T j ={{T1,T2,···,T 1024},{T 1025 ,T 1026 ,···,T 2048},···,{T 2049 ,T 2050 ,···,T o}}(j is the indicator representing the category, o is the data length), grayscale image data can better play the classification function of the convolutional neural network model than the one-dimensional time series, and the grayscale image data set is divided into training set, validation set and test set in the ratio of 0.7:0.2:0.1.

[0101] Preferably, step S5 includes: inputting the data of the training set to train the convolutional neural network model multiple times, and in each training iteration, verifying and evaluating the training results of the previous training through the verification set, and obtaining the best model based on the comparison of the verification and evaluation results. Specifically, the model is optimized using an adaptive moment estimation (Adma) optimizer.

[0102] Preferably, step S6 includes: using accuracy and loss function to evaluate the convolutional neural network model after each training, and saving the best training model. Specifically, the cross entropy loss function is used to evaluate the prediction accuracy of the model, and the accuracy is the ratio of the total number of correct diagnoses of the model to the total number of samples in the input data set.

[0103] Specifically, the training set is used as input for convolutional neural network model training. The model weights can be obtained by using the training set for model training. Each training iteration is verified by the validation set. The model performance is evaluated using the accuracy and loss function. The validation set is used to verify the weights obtained by training the training data. The model training accuracy and loss function of the training set and validation set are compared. Figure 5 As shown, the best model weights and the best trained model are saved.

[0104] Preferably, step S7 includes: after inputting the test set data into the best training model to obtain the fan diagnosis result, parsing the fan diagnosis result and the diagnosis accuracy through a normalized exponential function. Specifically, the output data of the best training model is input into a soft-max function for classification.

[0105] Specifically, Figure 6 As shown, the network architecture of the convolutional neural network model of this embodiment sequentially sets convolution module 1, convolution module 2, convolution module 3, two convolution modules 4, two fully connected layers 1, one fully connected layer 2 (three fully connected layers are located at Figure 6 far right).

[0106] Furthermore, convolution module 1 and convolution module 2 both set convolution layers and maximum pooling layers in sequence, and the output channel sizes of the convolution layers are set to 64 and 218 in sequence. Convolution module 3 sets two convolution layers and maximum pooling layers in sequence, and the output channel size of the convolution layer is set to 256. The convolution kernel sizes are all set to 3*3, the padding is set to 1, and the step size is set to 1. The pooling kernel size of the maximum pooling layer is set to 2*2, without padding, and the step size is set to 2.

[0107] Furthermore, the convolution layer is used for data convolution operation, the maximum pooling layer is used to apply the maximum value operation on the grayscale image, reduce the spatial dimension of the feature map, and maintain the depth of the feature map. The convolution module 4 sequentially sets two convolution layers, a spatial-dropout layer, and a maximum pooling layer. The output channel size of the convolution layer is set to 512, the convolution kernel size is set to 3*3, the padding is set to 1, and the step size is set to 1. The pooling kernel size of the maximum pooling layer is set to 2*2, no padding, and the step size is set to 2. The spatial-dropout layer randomly sets some areas to zero.

[0108] Furthermore, during the training process of the training set data input into the convolutional neural network model, the Adma optimizer is used for model optimization, and the accuracy and cross entropy loss functions are used for model evaluation. The soft-max function is used as the classification function for diagnosis. The training set data is input into the convolutional neural network model. Each iteration of training updates the weights and uses the validation set for validation evaluation. The model training accuracy and loss function of the training set and validation set are shown in Figure 5As shown in the figure, the horizontal axis represents the number of iterations and the vertical axis represents the parameter value. This data shows that the scheme has excellent fan fault diagnosis capabilities. Finally, the test set data is input into the best training model, and the output data of the best training model is input into the soft-max function for classification to obtain the fan fault diagnosis results.

[0109] Furthermore, this embodiment introduces a Spatial-Dropout deactivation layer into the original convolutional neural network model to prevent the model from overfitting, instead of simply improving the accuracy of training data while ignoring the accuracy of the test set. After implementation, it was found that adding a deactivation layer after the convolution layer in the convolutional neural network model can effectively improve the diagnostic accuracy of the test set.

[0110] In this embodiment, a fan fault diagnosis method based on image information fusion technology first calculates dimensionless indicators of the fan raw data to eliminate the dimensional relationship between the raw data. The obtained multi-index data is then converted into a grayscale image, and then a sliding window method is used to divide the data. This can not only meet the data volume required for deep learning of the convolutional neural network model, but also give full play to the superiority of the convolutional neural network model in processing image data. Judging from the accuracy of the training and verification data and the loss function curve, the training model performance is significantly improved.

[0111] A fan fault diagnosis method based on image information fusion technology in this embodiment first uses the collected fan vibration signal to perform data segmentation, and obtains the segmented sample data set by sliding the value according to the defined sliding window; then introduces the dimensionless formula and selects the peak index, waveform index and pulse index, and substitutes them into the sample data set to obtain the index data set. Multi-index data can improve the performance of the deep learning model, and the dimensionless index effectively reduces the dimensional relationship between the data; then the index data set is normalized and converted into two-dimensional grayscale image data, and the superiority of convolutional neural network in processing image data classification is used to improve the diagnosis accuracy; then the grayscale image data set is divided into a training set, a validation set and a test set in a ratio of 0.7:0.2:0.1; finally, a convolutional neural network (VGG-SD) model is constructed, and a spatial dropout layer (Spatial-Dropout) is introduced in the convolutional neural network model to prevent the training model from overfitting and improve the diagnostic accuracy of the test set. It is found through experiments that the diagnostic accuracy of the model for the test set is significantly improved, and a high-accuracy fault diagnosis result of the fan can be obtained by inputting the data of the test set.

[0112] The above is a detailed description of an embodiment of a fan fault diagnosis method based on image information fusion technology provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the core idea of ​​the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, the present invention can also be improved and modified, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A fan fault diagnosis method based on image information fusion technology, characterized in that: The following steps are involved: Step S1. Obtain vibration data of the fan when it is working, and segment the vibration data to obtain a sample data set; Step S2. Substitute the segmented data of the sample data set into the fan vibration definition algorithm to obtain an index data set corresponding to the vibration data; Step S3. Convert and segment the index data set to obtain corresponding grayscale image data; Step S4. Divide the grayscale image data into a training set, a validation set, and a test set in proportion; Step S5. Construct a convolutional neural network model, and train and verify it multiple times through training sets and verification sets to obtain multiple training models; Step S6. Evaluate multiple training models one by one through accuracy and loss function to obtain the best training model; Step S7: Test the best training model using the test set to obtain the fault diagnosis result of the wind turbine.

2. The fan fault diagnosis method according to claim 1, characterized in that: The step S1 comprises: Step S1.

1. When the fan is working, the vibration signal of the fan is collected by the sensor to obtain corresponding vibration data; Step S1.

2. Define a sliding window and its size through a sliding window algorithm, and segment the vibration data through a sliding window of this size to obtain a segmented sample data set.

3. The fan fault diagnosis method according to claim 2, characterized in that: The fan vibration definition algorithm in step S2 includes a dimensionless index expression, a peak index expression, a waveform index expression, and a pulse index expression.

4. The fan fault diagnosis method according to claim 3, characterized in that: The step S2 comprises: Step S2.

1. Let the sample data of the sample data set be x, obtain the corresponding probability density function p(x), and set the hyperparameters l and m to obtain the dimensionless index Expression of Step S2.

2. According to the dimensionless index The expression is to assign different values ​​to the hyperparameters l and m in turn, and get the peak index C f Expression, waveform index S f Expression of pulse index I f Expression of Step S2.

3. Substitute the segmented data in the sample data set into the peak index C in sequence f Expression, waveform index S f Expression of pulse index I f Calculate the sample data x and substitute it into the dimensionless index The expression gets the corresponding multiple indicator data sets 5. The fan fault diagnosis method according to claim 4, characterized in that: The dimensionless index The expression is as follows: By dimensionless index The expression of l→∞, m=2, and the peak index C is obtained. f The expression is as follows: Among them, X max Indicates the maximum value of sample data, X rms Represents the root mean square value of sample data; By dimensionless index The expression of l = 2, m = 1, then the waveform index S is obtained f The expression is as follows: Among them, X rms represents the root mean square value of the sample data, Represents the average amplitude of sample data; By dimensionless index The expression of l→∞, m=1, then the pulse index I is obtained f The expression is as follows:

6. The fan fault diagnosis method according to claim 4, characterized in that: The step S3 comprises: The grayscale image data is converted using the grayscale image conversion formula, and the data is segmented using the sliding window to obtain the grayscale image data set as follows: T x ={{T1,T2,···,T m },{T m ,T m+1 ,···,T 2m-1 },···,{T 2m ,T 2m+1 ,···,T n }}; The grayscale image conversion formula is as follows: in, Indicates the minimum value in the indicator data. Indicates the maximum value in the indicator data.

7. The fan fault diagnosis method according to claim 1, characterized in that: In step S4, the grayscale image dataset is divided into a training set, a validation set and a test set in a ratio of 0.7:0.2:0.

1.

8. The fan fault diagnosis method according to claim 1 or 7, characterized in that: The step S5 includes: inputting the data of the training set to train the convolutional neural network model multiple times, and in each training iteration, verifying and evaluating the training results of the previous training through the verification set, and obtaining the best model based on the comparison of the verification and evaluation results.

9. The fan fault diagnosis method according to claim 8, characterized in that: The step S6 includes: using accuracy and loss function to evaluate the convolutional neural network model after each training, and saving the best training model.

10. The fan fault diagnosis method according to claim 9, characterized in that: The step S7 includes: after inputting the data of the test set into the optimal training model to obtain the fan diagnosis result, the fan diagnosis result and the diagnosis accuracy are analyzed by a normalized exponential function.

Images