Harmonic reducer fault diagnosis method based on artificial intelligence algorithm
By converting the vibration signal of the harmonic reducer into a frequency domain Gram matrix and using the ResNet101 model for fault diagnosis, the problem of insufficient signal clarity and accuracy in the prior art is solved, and high-accuracy fault diagnosis is achieved.
Patent Information
- Application Number
- CN202410911784.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-09
- Publication Date
- 2024-11-26
AI Technical Summary
Existing harmonic reducer fault diagnosis methods have signal clarity and accuracy problems when processing complex signals, especially due to artifacts and insufficient resolution caused by noise and complex structures, which makes it difficult for traditional methods to accurately capture fault characteristics.
Using an artificial intelligence-based method, by converting one-dimensional vibration signals into frequency domain Gramma matrix and performing two-dimensional Fourier transformation, images are generated to extract feature information, and training and classification are used for ResNet101 model, and model parameters are dynamically optimized to Improve diagnostic accuracy.
The feature expression and diagnostic accuracy of harmonic reducer fault signals are improved, and the interpretability of fault status is enhanced. The accuracy of the model on different data sets reaches as high as 99%, which is better than the traditional model.
Smart Images

Figure CN119027710A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of harmonic reducer fault diagnosis, and relates to a harmonic reducer fault diagnosis method based on an artificial intelligence algorithm. Background Art
[0002] Intelligent devices such as industrial robots are an indispensable pillar of manufacturing progress, and their service quality directly affects the long-term safety and stability of manufacturing operations. In the literature [1], harmonic reducers, as one of the key components of industrial robots, play a vital role in the transmission process of robot joint arms. During their manufacturing and operation stages, harmonic reducers are prone to various types of damage and failures. Therefore, the research on fault detection and monitoring technology of harmonic reducers has important scientific and theoretical value and practical engineering application significance, because it can ensure the safe and stable operation of key equipment in the manufacturing process.
[0003] Faults in harmonic reducers exhibit various characteristics and complex structures, which poses a challenge to monitoring. Traditional signal detection and fault diagnosis methods mainly rely on analyzing the time domain or frequency domain information of vibration, speed, noise and other signal data at a specific time point. These methods often ignore the global information contained in the frequency domain characteristics of the signal. However, the signals captured from harmonic reducers exhibit complex transient behaviors that cannot be accurately represented by local time domain details alone. Therefore, a method that can summarize its operating status is needed. As a result, traditional fault detection and monitoring techniques are significantly limited.
[0004] Multidimensional information extraction methods can transform detailed features into two-dimensional or three-dimensional visual information. These methods not only focus on the details of local signals, but also reveal the global features embedded in the data. Therefore, more and more researchers have begun to use multidimensional information extraction methods to transform one-dimensional data into two-dimensional or three-dimensional information. In reference [2], Wigner-Ville time-frequency transform was used; in reference [3], the spectrum graph method was used; in reference [4], the autocorrelation method was used; in reference [5], the overlapping window method, as well as short-time Fourier transform and wavelet transform techniques were used. Given the complexity of the internal and external structures of the harmonic reducer and its operating environment, the diagnostic signal may contain a lot of noise and redundant information. Therefore, it is necessary to implement a suitable transformation algorithm to achieve excellent visibility and interpretability of multidimensional data.
[0005] AlexNet promoted the widespread application of deep learning in computer vision and other fields as the first large-scale deep neural network using the ReLU (rectified linear unit) activation function. Compared with traditional sigmoid or tanh activation functions, ReLU speeds up the training of deep neural networks. Sun Shuguang et al. combined time-frequency images with the AlexNet model in the literature [6] and applied it to the field of fault diagnosis. Before the emergence of the ResNet model, VGG and GoogLeNet were highly respected models. The advantage of the VGG model lies in its simple and unified network structure, which mainly consists of repeated 3x3 convolutional layers and 2x2 maximum pooling layers. In the literature [7], this simple and unified structure makes the VGG model easy to understand and implement. On the other hand, GoogLeNet introduced the Inception module, which is a parallel structure that can perform convolution operations of various sizes in parallel at the same hierarchical level, enabling the model to capture information at different spatial scales. This design enables GoogLeNet to use computing resources more efficiently while improving the performance of the model. In the literature [8], these two models are also frequently used in the field of fault diagnosis. The shortcomings of these algorithms in fault classification.
[0006] In two-dimensional or three-dimensional visualization methods, the Wigner-Ville time-frequency transform has cross-term interference, which easily leads to artifacts in the time-frequency image, affecting the clarity and accuracy of the signal. Although the spectrogram method can provide information about the distribution of the signal in the frequency domain, its time-frequency resolution is limited by the Heisenberg uncertainty principle and cannot obtain high resolution in both time and frequency. In addition, for complex signals, the spectrogram may not provide enough details. The autocorrelation method is suitable for the analysis of periodic signals, but it is less effective for non-periodic or complex signals. In addition, the autocorrelation method is sensitive to noise and is easily affected by noise, resulting in a decrease in the accuracy of feature extraction. The overlapping window method can improve the resolution of time-frequency analysis to a certain extent, but its computational complexity is high, especially for large-scale data, the amount of calculation is large, and it is difficult to process in real time. In addition, the choice of window has a great influence on the analysis results, and different windows may lead to different results. The time-frequency resolution of the short-time Fourier transform is fixed, and it is impossible to achieve the optimal resolution in both time and frequency at the same time. Although the wavelet transform has the ability of multi-resolution analysis, its transformation results depend on the selected mother wavelet. Different mother wavelets will lead to different analysis results and lack a unified standard. In terms of model algorithms, AlexNet's relatively shallow network structure may not be able to capture sufficiently deep feature information when dealing with complex fault modes. VGG may not perform well in classifying some complex fault modes, especially for nonlinear and non-stationary fault signals. In fault diagnosis, signals often contain noise, and GoogLeNet may be unstable when dealing with noisy data.
[0007] [1]Yiming He,Jihong Chen,Xing Zhou,Shifeng Huang,In-situ faultdiagnosis for the HRs of industrial robots via multi-scale mixedconvolutional neural networks,Journal of Manufacturing Systems,Volume 66,2023,Pages 233-247.
[0008] [2]D.Griffin and Jae Lim,"Signal estimation from modified short-timeFourier transform,"in IEEE Transactions on Acoustics,Speech,and SignalProcessing,vol.32,no.2,pp.236-243,April 1984.
[0009] [3]R.Balan,"On signal reconstruction from its spectrogram,"201044thAnnual Conference on Information Sciences and Systems(CISS),Princeton,NJ,USA,2010,pp.1-4,doi:10.1109 / CISS.2010.5464828.
[0010] [4]Ya-Qing Tu,Yan-Lin Shen,Phase correction autocorrelation-basedfrequency estimation method for sinusoidal signal,Signal Processing,Volume130,2017,Pages 183-189.
[0011] [5] Huan Zhao, Shaofang He, Analysis of speech signals' characteristics based on MF-DFA with moving overlapping windows, Physica A: Statistical Mechanics and its Applications, Volume 442, 2016, Pages 343-349, ISSN 0378-4371.
[0012] [6]Ismail Fawaz,H.,Lucas,B.,Forestier,G.et al.InceptionTime:FindingAlexNet for time series classification.Data Min Knowl Disc 34,1936-1962(2020).
[0013] [7] L.Wen,
[0014] [8] Z.Wang, W.Shangguan, C.Peng and B.Cai, "Afault diagnosis method based on data feature reconstruction and deep transfer learning," 2022 IEEE Asia-Pacific Conference on Image Processing, Electronics and Computers (IPEC), Dalian, China, 2022, pp.1-5. Summary of the invention
[0015] In order to solve the above technical problems, the purpose of the present invention is to provide a harmonic reducer fault diagnosis method based on artificial intelligence algorithm.
[0016] A harmonic reducer fault diagnosis method based on artificial intelligence algorithm of the present invention comprises:
[0017] Step 1: Collect the fault data and normal parts data of four types of harmonic reducers to form sample data;
[0018] Step 2: Preprocess the sample data;
[0019] Step 3: Convert the preprocessed one-dimensional data from rectangular coordinates to polar coordinates, construct the Gram matrix and convert it into the frequency domain Gram matrix, and then convert it into an image;
[0020] Step 4: Use random sampling method to divide the images into training set and validation set in a ratio of 7:3, and preprocess the training set and validation set respectively;
[0021] Step 5: Input the training set data into ResNet101 to extract image feature information, and use these features to train the classification task. At the same time, during the entire training process, dynamically monitor the accuracy indicators of the model on the training set and validation set to continuously optimize the model parameters and evaluate the model performance.
[0022] The present invention provides a harmonic reducer fault diagnosis method based on artificial intelligence algorithm, which has the following beneficial effects:
[0023] (1) The frequency domain Gram method is used to extract the harmonic reducer fault signal information, making the signal feature expression richer. The Gram angle field is used to convert the one-dimensional vibration signal into a grayscale image, retaining the correlation between the one-dimensional signal and the time series. The main diagonal elements of the Gram angle field correspond to the values of different timestamps, so they have a unique mapping relationship with the time series and fully reflect all the information of the time series. The off-diagonal elements represent the correlation of data at different time points and reflect the internal correlation of the time series, which allows the characteristics of the vibration signal to be fully preserved through the image.
[0024] (2) The image is converted from the spatial domain to the frequency domain through a two-dimensional Fourier transform, and the grayscale distribution function of the image is converted into the frequency distribution function of the image. In the image, the horizontal axis usually represents the spatial frequency, reflecting the periodicity and regularity of the signal, while the brightness of the image depicts the different periods and specific fluctuations of different fault signals. The vertical axis represents the frequency of the signal and shows the distribution of the signal at different frequencies. The brightness in the image can indicate the increase in energy of different faults at specific frequencies. This helps people intuitively understand the characteristics of different fault signals and enhances the interpretability of various fault states in the harmonic reducer. All of these help the model better capture the characteristics of the fault signal, thereby improving the accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 It is a flow chart of a harmonic reducer fault diagnosis method based on artificial intelligence algorithm of the present invention. DETAILED DESCRIPTION
[0026] like Figure 1 As shown, a harmonic reducer fault diagnosis method based on artificial intelligence algorithm of the present invention comprises:
[0027] Step 1: Collect the fault data and normal parts data of four types of harmonic reducers to form sample data.
[0028] In the specific implementation, the fault data and normal data of four harmonic reducers were collected through acceleration sensors and Siemens software. All vibration signals were collected on a self-made experimental platform designed by China Midea Group for harmonic reducer fault testing. In the actual production line, four common harmonic reducer faults were selected for research.
[0029] The faults include: fracture fault of the flexible bearing in the wave generator, misalignment fault of the input shaft, pitting corrosion fault at the tooth root of the flexible wheel, and pitting corrosion fault at the tooth tip of the flexible wheel.
[0030] Step 2: Preprocess the sample data.
[0031] In the specific implementation, in step 2, the sample data is normalized, de-biased and mean filtered; the sample data is normalized so that the normalized data is between [-1, 1], and the calculation formula is as follows:
[0032]
[0033] in, is the normalized value; x i is the i-th sample data; max(X) is the maximum value in the sample data; min(X) is the minimum value in the sample data.
[0034] Step 3: Convert the preprocessed one-dimensional data from rectangular coordinates to polar coordinates, construct the Gram matrix and convert it into the frequency domain Gram matrix, and then convert it into an image, specifically:
[0035] Step 3.1: Use the Gram angle field to map the one-dimensional sample signal from the normalized time series to the polar coordinate system, and transform x i Mapping to the arccosine function, mapping time to radius, the conversion function is:
[0036]
[0037] Φ=[φ1 φ2 … φ n ] (4)
[0038] Construct the following Gram matrix:
[0039]
[0040] Wherein, in formula (2), Φ represents the polar coordinate angle set, arccos is the inverse cosine function, which is used to convert the normalized time series value into an angle; is the normalized time series value, represents the normalized time series set; in formula (3), γ i Represents the normalized time index value, t i represents the original time index of the time series, N1 represents the total length of the time series; in formula (5), G D represents the Gram matrix, φ1 is the polar coordinate angle corresponding to the first point in the time series, φ n is the polar coordinate angle corresponding to the nth point in the time series.
[0041] Step 3.2: Scale each element of the obtained Gram matrix to between 0 and 255, corresponding to the size of each pixel in the image, to convert the two-dimensional Gram matrix into a Gram angle field using the following formula:
[0042] I(j,k)=int(127.5×(G(j,k)+1)) (6)
[0043] Where I(j,k) represents the pixel value at point (j,k) in the image, int(·) is the rounding function, and G(j,k) represents the element value corresponding to the j-th row and k-th column in the Gram matrix.
[0044] Step 3.3: Convert the Gram angle field into the frequency domain Gram matrix through two-dimensional Fourier transform:
[0045]
[0046] Since the pixels of an image are discrete and finite, and in the computer field, computers process digital signals, under these constraints, the two-dimensional Fourier transform of an image can only be performed in a discrete manner. The formula for the two-dimensional discrete Fourier transform is shown in equation (7); where F(μ, v) is the Fourier transform result in the frequency domain; M and N are the width and height of the image, respectively; μ and v are coordinates in the frequency domain, where μ in the frequency domain corresponds to the X-axis and v to the Y-axis.
[0047] Step 3.4: Convert the frequency domain Gram matrix into a two-dimensional image. Each element in the frequency domain Gram matrix corresponds to the color of a pixel in the image, and the size of the original image is equal to the number of rows and columns of the matrix.
[0048] Step 4: Use random sampling method to divide the images into training set and validation set in a ratio of 7:3, and preprocess the training set and validation set respectively, as follows:
[0049] Step 4.1: Use random sampling method to divide the images into training set and validation set in a ratio of 7:3.
[0050] Step 4.2: In the image preprocessing of the training set, the system first reads in all the pixel data of the image, and then performs a variety of data augmentation operations, including random scaling, cropping to a specified size, random horizontal flipping, and normalization to ensure that the model can learn various changes in the image. After these operations, the size of the training set images is uniformly adjusted to 224×224×3.
[0051] Step 4.3: For the image preprocessing of the validation set, after reading in all the pixel data, the system first performs scaling, and then performs center cropping and normalization to ensure that the processing results of the validation set images are consistent with the training set in size and format, that is, they are also 224×224×3. This consistency helps the fair evaluation and performance testing of the model on different datasets.
[0052] Step 5: Input the training set data into ResNet101 to extract image feature information, and use these features to train the classification task. At the same time, during the entire training process, dynamically monitor the accuracy indicators of the model on the training set and the validation set to continuously optimize the model parameters and evaluate the model performance. Specifically:
[0053] Step 5.1: Initialize the network weights, set the learning rate to 0.0001, the batch size to 16, and the training cycle to 300.
[0054] Step 5.2: Load the training set data in batches. For each batch of data in the training set, get a batch of images and labels from the data loader.
[0055] Step 5.3: Clear the historical gradient information of model parameters.
[0056] Step 5.4: Input the training set data of the current batch into the ResNet101 model and perform forward calculation to generate predicted values.
[0057] In specific implementation, the activation function used in the Resnet101 model is the ReLu function, and the formula is as follows:
[0058] ReLU(x)=max(0,x) (8)
[0059] Here, x represents the output of the previous layer in the ResNet101 network.
[0060] Step 5.5: Calculate the loss between the model prediction value and the actual label as the error measure of the current batch of training data;
[0061] In specific implementation, the loss function used is the cross entropy loss function:
[0062]
[0063] in, is the value of the cross entropy loss function, which is used to measure the difference between the model prediction and the true label. The smaller the loss value, the more accurate the model prediction. A is the number of samples, that is, the total number of samples in the batch. C is the number of categories, that is, the total number of categories in the classification task. i,c The true label of sample i. If the true category of sample i is c, then y i,c =1, otherwise y i,c =0, is the probability that sample i is predicted to be of category c.
[0064] Step 5.6: Calculate the predicted category of each training data sample in the current batch, calculate the number of accurate predictions for this batch and add them to the total number of accurate predictions.
[0065] Step 5.7: Based on the calculated loss value, perform the back propagation algorithm to calculate the gradient information of the model parameters.
[0066] Step 5.8: Based on the gradient information obtained by back-propagation calculation, update the parameters of the neural network model to reduce the loss value.
[0067] Step 5.9: Count the losses, add the losses of the current batch to the total losses, and add the losses of the current batch to the loss list.
[0068] Step 5.10: Repeat steps 5.2-5.9 until all training sets are loaded and training is completed. Save the accuracy and loss rate calculated during the training process for subsequent analysis and model evaluation.
[0069] Step 5.11: Load the validation set data. For each batch of data in the validation set, get a batch of images and labels from the data loader.
[0070] Step 5.12: Input the validation set data into the model and perform forward propagation to calculate the output.
[0071] Step 5.13: Calculate the loss between the predicted values and the true labels.
[0072] Step 5.14: Accuracy calculation, calculate the predicted category of each validation set data sample, calculate the number of accurate predictions for this batch and add it to the total accurate prediction number.
[0073] Step 5.15: Count the losses, add the losses of the current batch to the total losses, add the losses of the current batch to the loss list, repeat 5.11-5.15 until all validation sets are loaded and the model validation is completed, and save the accuracy and loss rate calculated during the validation process.
[0074] Step 5.16: If the number of training times reaches the training cycle, the loop ends, otherwise continue to loop and execute steps 5.2 to 5.16.
[0075] In the fixed working condition data set, the model accuracy of the method of the present invention is 99%, in the variable load data set, the model accuracy is 99.4%, and in the variable speed data set, the model accuracy is 94.7%. Specifically, as shown in Table 1, in the fixed working condition data set, ResNet101 is 0.3% higher than ResNet50, 3.5% higher than VGGNet-16, and 2.9% higher than the Swin-Tiny model. In the variable load data set, the accuracy is 0.7% higher than ResNet50, 5.5% higher than VGGNet-16, and 5.6% higher than the Swin-Tiny data set. In the variable speed data set, the accuracy is 0.9% higher than ResNet50, 0.1% higher than VGGNet-16, and 9.3% higher than the Swin-Tiny data set. This shows that the ResNet101 model has certain advantages in training FDG images.
[0076] Table 1
[0077] Model Fixed working condition Variable load Variable speed AlexNet 88.8% 87.2% 92.5% VGGNet-16 95.5% 93.9% 94.6% GoogleNet 85.3% 94.5% 86.7% Swin-Tiny 96.1% 93.8% 85.4% ResNet50 98.7% 98.7% 93.8% ResNet101 99.0% 99.4% 94.7%
[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the concept of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A harmonic reducer fault diagnosis method based on artificial intelligence algorithm, characterized in that: include: Step 1: Collect the fault data and normal data of four types of harmonic reducers to form sample data; Step 2: Preprocess the sample data; Step 3: Convert the preprocessed one-dimensional data from rectangular coordinates to polar coordinates, construct the Gram matrix and convert it into the frequency domain Gram matrix, and then convert it into an image; Step 4: Use random sampling method to divide the images into training set and validation set in a ratio of 7:3, and preprocess the training set and validation set respectively; Step 5: Input the training set data into ResNet101 to extract image feature information, and use these features to train the classification task. At the same time, during the entire training process, dynamically monitor the accuracy indicators of the model on the training set and validation set to continuously optimize the model parameters and evaluate the model performance.
2. The harmonic reducer fault diagnosis method based on artificial intelligence algorithm according to claim 1 is characterized in that: In step 1, the fault data and normal part data of four harmonic reducers are collected by using an acceleration sensor and Siemens software; The faults include: fracture fault of the flexible bearing in the wave generator, misalignment fault of the input shaft, pitting corrosion fault at the tooth root of the flexible wheel and pitting corrosion fault at the tooth tip of the flexible wheel.
3. The harmonic reducer fault diagnosis method based on artificial intelligence algorithm according to claim 1 is characterized in that: In step 2, the sample data is normalized, debiased, and mean filtered. The sample data is normalized so that the normalized data is between [-1, 1]. The calculation formula is as follows: in, is the normalized value; x i is the i-th sample data; max(X) is the maximum value in the sample data; min(X) is the minimum value in the sample data.
4. The harmonic reducer fault diagnosis method based on artificial intelligence algorithm according to claim 1 is characterized in that: The step 3 is specifically as follows: Step 3.1: Use the Gram angle field to map the one-dimensional sample signal from the normalized time series to the polar coordinate system, and transform x i Mapping to the inverse cosine function, mapping time to radius, the conversion function is: Φ=[φ1 φ2 … φ n ] (4) Construct the following Gram matrix: Wherein, in formula (2), Φ represents the polar coordinate angle set, arccos is the inverse cosine function, which is used to convert the normalized time series value into an angle; is the normalized time series value, represents the normalized time series set; in formula (3), γ i Represents the normalized time index value, t i represents the original time index of the time series, N1 represents the total length of the time series; in formula (5), G D represents the Gram matrix, φ1 is the polar coordinate angle corresponding to the first point in the time series, φ n is the polar coordinate angle corresponding to the nth point in the time series; Step 3.2: Scale each element of the obtained Gram matrix to between 0 and 255, corresponding to the size of each pixel in the image, to convert the two-dimensional Gram matrix into a Gram angle field using the following formula: I(j,k)=int(127.5×(G(j,k)+1)) (6) Where I(j,k) represents the pixel value at point (j,k) in the image, int(·) is the rounding function, and G(j,k) represents the element value corresponding to the j-th row and k-th column in the Gram matrix. Step 3.3: Convert the Gram angle field into a frequency domain Gram matrix through two-dimensional Fourier transform: Since the pixels of an image are discrete and limited, and in the computer field, computers process digital signals, under these constraints, the two-dimensional Fourier transform of the image can only be performed in a discrete manner. The formula for the two-dimensional discrete Fourier transform is as shown in Equation (7); where F(μ,v) is the Fourier transform result in the frequency domain; M and N are the width and height of the image respectively; μ and v are the coordinates in the frequency domain, where μ corresponds to the X-axis and v corresponds to the Y-axis in the frequency domain; Step 3.4: Convert the frequency domain Gram matrix into a two-dimensional image. Each element in the frequency domain Gram matrix corresponds to the color of a pixel in the image. The size of the original image is equal to the number of rows and columns of the matrix.
5. The harmonic reducer fault diagnosis method based on artificial intelligence algorithm according to claim 1 is characterized in that: The step 4 is specifically as follows: Step 4.1: Use random sampling to divide the images into training and validation sets in a ratio of 7:3; Step 4.2: During preprocessing of the training set images, the system first reads in all pixel data for the image. It then performs various data augmentation operations, including random scaling, cropping to a specified size, random horizontal flipping, and normalization, to ensure the model can learn various variations in the image. After these operations, the training set images are uniformly resized to 224×224×3. Step 4.3: For validation set image preprocessing, after reading in all pixel data, the system first performs scaling, then center cropping, and normalization to ensure that the validation set image processing results are consistent with the training set in size and format, that is, also 224×224×3. This consistency helps to fairly evaluate and perform performance testing of the model on different datasets.
6. The harmonic reducer fault diagnosis method based on artificial intelligence algorithm according to claim 1 is characterized in that: The step 5 is specifically as follows: Step 5.1: Initialize the network weights, set the learning rate to 0.0001, the batch size to 16, and the training epochs to 300; Step 5.2: Load the training set data in batches. For each batch of data in the training set, get a batch of images and labels from the data loader. Step 5.3: Clear the historical gradient information of the model parameters; Step 5.4: Input the training set data of the current batch into the ResNet101 model and perform forward calculation to generate the predicted value; Step 5.5: Calculate the loss between the model's predicted value and the actual label as the error measure for the current batch of training data. Step 5.6: Calculate the predicted category for each training data sample in the current batch, calculate the number of accurate predictions for this batch and add it to the total number of accurate predictions; Step 5.7: Based on the calculated loss value, perform the backpropagation algorithm to calculate the gradient information of the model parameters; Step 5.8: Update the parameters of the neural network model based on the gradient information obtained by backpropagation to reduce the loss value; Step 5.9: Calculate the loss, add the loss of the current batch to the total loss, and add the loss of the current batch to the loss list; Step 5.10: Repeat steps 5.2-5.9 until all training sets are loaded and training is complete. Save the accuracy and loss rates calculated during training for subsequent analysis and model evaluation. Step 5.11: Load the validation set data. For each batch of data in the validation set, get a batch of images and labels from the data loader. Step 5.12: Input the validation set data into the model and perform forward propagation to calculate the output; Step 5.13: Calculate the loss between the predicted value and the true label; Step 5.14: Accuracy calculation: calculate the predicted category of each validation set data sample, calculate the number of accurate predictions for this batch and add it to the total accurate prediction number; Step 5.15: Calculate the loss, add the loss of the current batch to the total loss, add the loss of the current batch to the loss list, and repeat 5.11-5.15 until the validation set is fully loaded and the model validation is completed; save the accuracy and loss rate calculated during the validation process; Step 5.16: If the number of training times reaches the training cycle, the loop ends; otherwise, continue to loop and execute steps 5.2 to 5.
16.
7. The harmonic reducer fault diagnosis method based on artificial intelligence algorithm according to claim 6 is characterized in that: The activation function used in the Resnet101 model in step 5.4 is the ReLu function, and the formula is as follows: ReLU(x)=max(0,x) (8) Here, x represents the output of the previous layer in the ResNet101 network.
8. The harmonic reducer fault diagnosis method based on artificial intelligence algorithm according to claim 6 is characterized in that: The loss function used in step 5.5 is the cross entropy loss function: in, is the value of the cross entropy loss function, which is used to measure the difference between the model prediction and the true label. The smaller the loss value, the more accurate the model prediction. A is the number of samples, that is, the total number of samples in the batch. C is the number of categories, that is, the total number of categories in the classification task. i,c The true label of sample i. If the true category of sample i is c, then y i,c =1, otherwise y i,c =0, is the probability that sample i is predicted to be of category c.
Citation Information
Cited By
Elevator running state evaluation method based on multi-domain comparative learning
CN121580095A
Explosion-proof intelligent electro-hydraulic actuator for coal mine and fault self-diagnosis method thereof
CN121594058A
