In the present industrial revolution era, the industrial mechanical system becomes incessantly highly intelligent and composite. So, it is necessary to develop data-driven and monitoring approaches for achieving quick, trustable, and high-quality analysis in an automated way. Fault diagnosis is an essential process to verify the safety and reliability operations of rotating machinery. The advent of deep learning (DL) methods employed to diagnose faults in rotating machinery by extracting a set of feature vectors from the vibration signals. This paper presents an Intelligent Industrial Fault Diagnosis using Sailfish Optimized Inception with Residual Network (IIFD-SOIR) Model. The proposed model operates on three major processes namely signal representation, feature extraction, and classification. The proposed model uses a Continuous Wavelet Transform (CWT) is for preprocessed representation of the original vibration signal. In addition, Inception with ResNet v2 based feature extraction model is applied to generate high-level features. Besides, the parameter tuning of Inception with the ResNet v2 model is carried out using a sailfish optimizer. Finally, a multilayer perceptron (MLP) is applied as a classification technique to diagnose the faults proficiently. Extensive experimentation takes place to ensure the outcome of the presented model on the gearbox dataset and a motor bearing dataset. The experimental outcome indicated that the IIFD-SOIR model has reached a higher average accuracy of 99.6% and 99.64% on the applied gearbox dataset and bearing dataset. The simulation outcome ensured that the proposed model has attained maximum performance over the compared methods.

In recent times, the operational status observance and fault analysis of rotating machinery is highly significant. Rotating machineries are becoming essential equipment in the industrial sector [

In recent times, an extensive application of the Internet of Things (IoT), advanced intelligent sensing devices, and data collection methodologies are applied vastly in rotating machinery automation. Most of the monitor data like vibrations, sound, temperature, power, and pressure of rotating machinery can be attained effortlessly, and the previous data saves the health details of rotating machinery from starting to the termination of the service. Hence, engineers compute fault diagnosis using statistical analysis of massive historical information. Currently, the data-driven fault diagnosis technique is well-known and used in several applications [

Contrasting from conventional fault diagnosis models relied on the signal processing method; intelligent diagnosing schemes are used for extracting applicable features from monitoring data in the industrial sector. The general intelligent diagnosis models have 3 phases namely, feature extraction, feature selection (FS), and fault classification. Initially, feature extraction transforms the actual data signals gathered by numerous sensors in both the time and frequency domain to reliable representative features for fault identification. Secondly, FS eliminates lower sensitivity and unwanted data from collected features. Thirdly, fault identification feeds the collected features to the fault classifier and compute pattern analysis and, lastly, results in classification results by frequent iterative training. By the utilization and validation, the predefined approaches have inferior feature extraction potential because of the shallow network architecture, and it is not easy to apply in the alternate application, especially for big data [

This paper develops an Intelligent Industrial Fault Diagnosis using Sailfish Optimized Inception with Residual Network (IIFD-SOIR) Model. The proposed model involves three processes such as signal representation, feature extraction, and classification. Initially, the Continuous Wavelet Transform (CWT) is applied to achieve a pre-processed representation of raw vibration signals. Afterward, Inception with ResNet v2 (IRV2) based feature extraction model is employed to create a set of high-level features. It is chosen over the other DL models because it possesses a shortcut connection at the left of each module. In addition, it has roughly the computational cost of Inception-v4. Additionally, the training of the IRV2 model is faster and got slightly better final accuracy than Inception-v4. Also, the way of fixing the hyperparameters of the IRV2 model necessitates knowledge and widespread trial and error. As there are no simpler and easy methods available for fixing the hyperparameters of the IRV2, the proposed model makes use of a sailfish optimizer (SFO) to tune them. Lastly, a multilayer perceptron (MLP) is applied as a classification tool to identify the faults capably. The utilization of SFO for the hyperparameter tuning of IRV2 in the fault diagnosis process shows the novelty of the work. Extensive experimentation takes place to ensure the effective outcome of the IIFD-SOIR method on the gearbox dataset and a motor bearing dataset.

The organization of the paper is given as follows. Section 2 briefs the related works, Section 3 proposes the IIFD-SOIR model, Section 4 simulates the presented model, and finally, Section 5 concludes the paper.

Awan et al. [

CNN is significantly capable of extracting effective features and is mainly employed in image analysis [

Rotating machinery is a function of different rotating speeds and loads. For performing fault identification in several functioning states, the vibration signal from the machine in the total speed and load range required for obtaining to train it [

By resampling frequency

Every data has a comparable length behind pre-processing at sampling frequencies that are similar multiples of the rotational frequencies. The wavelet transform decays a signal in the time-frequency field by utilizing relatives of wavelet functions. Scaling and translation of an essential wavelet function are defined by:

CWT take over and made the localization design of the short-time Fourier transform (STFT). A CWT is used for signal time-frequency diagnosis and processing. A CWT of a signal

Getting every wavelet coefficient in a matrix

Massive image recognition necessitates a more complicated CNN architecture and more calculations, which takes longer to train and calculate. Conversely, a massive image is reducing the result of tiny local features and decrease the sensitivity and accuracy of fault analysis. For accommodating this, CWTS cropping is performed using 3 rules:

A cropping effect should include at worst the CWT coefficients of 1 whole rotating duration.

The length of the one side of the square outcome should be superior to 2q.

When the pixel's coordinate points are greater than the coordinate points on the time axis, the pixel cannot be used as a result.

CNN is a variation of multilayer FC feedforward neural networks (FFNN) that can remove local features to classify data in an automated way. It is extremely utilized in varied computer vision functions. While several variations of the CNN method are made, a structure of the usual CNN is created with a convolution layer, pooling or sub-sampling layer, and FC layer as in a typical multiple NN.

The convolution layer is the important building block of CNN. It can be generally developed for a group of learnable kernels and one trainable bias for every feature map. In the convolutional layer, all filters are linked to the local patches in the feature map of the preceding layer [

Behind the convolution layer, it is necessary of adding a pooling layer among the CNN layers. It joins the outcome of the neighboring neurons at 1 layer to an individual neuron in the subsequent layer. Individual groups amongst several feature maps are optimal for obtaining further abstract feature illustrations. It can be used for shorting the calculations and manage overfitting by decreasing the dimensionality of the input for reducing the count of parameters. When an input map is available, after that resultant map with diminished size would be attained using a pooling function that is illustrated as:

Inception-ResNet-V2 (IRV2) was developed by Google Company in 2018 which is applied in place of existing approaches for the fault diagnosis of machinery. It is defined as the integration of GoogLeNet and ResNet. This method is composed of 10 portions, where each portion has its responsibility in role orientation as well as function. Here, Inception is a common network with parallel layer infrastructure used in GoogleNet. The filters have parallel connections with various sizes of 1x1, 3x3, and 5x5. The tiny size leads to a convolutional kernel and extracts the image features effectively and limits the model variables. When compared with other sizes, the large-scale convolution kernel would maximize the variables of the model matrix, hence various small-scaled convolution kernels are interchanged in a parallel fashion for eliminating the functional variables. Consequently, the method is applied extensively and more reliable when compared with the former network with Inception. Inception v1–v4 is the general approach of GoogleNet. Thus, the residual learning enabled ResNet is an extension of ILSVRC 2015 that applies 152 layers. ResNet's core assumption is to incorporate a direct link to this method, which is referred to as Highway Network informally. The traditional network structure is defined as a non-linear conversion of functional input, whereas Highway Network enables a limited ratio of a result in the existing network layer. As a result, the actual input data is forwarded directly to the upcoming layer. At the same time, ResNet secures the data by direct transmission of input to output. The entire network has to know the variations among input and output that signifies the learning objectives as well as complexities. ResNet-50, ResNet-101, and ResNet-152 are few modules in ResNet. In Residual-Inception system (

In the Inception with the ResNet v2 model, few main hyperparameters exist namely kernel size, filter count, hidden node count, and penalty coefficient, which majorly influence the overall results. Practically, it is time-consuming and hard to select the proper combination of parameters. To choose the optimal parameters of Inception with the ResNet v2 model, the SFO algorithm is employed. In general, SFO [

MLP is defined as a NN method with several hidden layers, and neurons among adjacent layers are linked together. The structural representation of this method is depicted in

The parameter selection is composed of newly presented MLP depends upon the experience and experiment. The hidden layer selection is computed by comparing the experiment by fixing 2, 4, 6, and 8 hidden layers and the attained results demonstrate the layer with enhanced time cost whereas the accuracy is not maximized. If the layer is fixed as 2, the classification accuracy is reduced. Hence, the accuracy and time cost can be balanced by 4 hidden layers applied in this approach. The count of neurons present in a hidden layer is fixed based on multiple trial performance, and the principle is balanced with time cost and accuracy. The activation functions, as well as loss functions, are ReLU and softmax cross-entropy along with logits is employed, correspondingly. The flow of extraction is composed of 3 phases namely, sample selection, model training, and classification generation.

The performance of the proposed model is simulated using the Python tool. To ensure the effective outcome of the presented model in the identification of various fault class labels, two datasets namely automotive gearbox and bearing fault from Case Western Reserve University Bearing Data Center [

Average accuracy analysis of the IIFD-SOIR model with the existing models is also made. The experimental values stated that the FFT-KNN model has exhibited worse performance with the least average accuracy of 86.353%. At the same time, the FFT-SVM model has offered slightly better performance with an average accuracy of 97.956%. Followed by, the FFT-DBN, CNN2, and CNN models have appeared moderate and closer average accuracy of 98.271%, 98.289%, and 98.304% respectively. Though the FFT-SAE model has obtained a high average accuracy of 99.231%, the proposed IIFD-SOIR model has demonstrated superior performance with an average accuracy of 99.6%.

In addition, on the applied bearing dataset 6, the IIFD-SOIR, CNN, CNN2, and FFT-KNN methods have accomplished high accuracy of 99.56%, 99.09%, 99.10%, and 98.53% while the FFT-SAE, FFT-DBN, and FFT-SVM frameworks have depicted reasonable performance with the accuracy of 97.32%, 97.78%, 92.52% correspondingly. In addition, on the applied bearing dataset 7, the IIFD-SOIR, FFT-SVM, CNN, and CNN2 methodologies have achieved superior accuracy of 100%, 100%, 100%, and 100% while the FFT-DBN, FFT-SAE, and FFT-KNN technologies have represented the least performance with the accuracy of 99.51%, 99.43%, 98.97% respectively. Followed by, on the applied bearing dataset 8, the IIFD-SOIR, CNN, FFT-KNN, and FFT-DBN schemes have obtained higher accuracy of 98.80%, 98.20%, 95.49%, and 95.22%, and the FFT-SVM, FFT-SAE, and CNN2 technologies have illustrated minimum performance with the accuracy of 95.31%, 95.15%, 94.18% correspondingly. Moreover, on the applied bearing dataset 9, the FFT-SVM, FFT-SAE, IIFD-SOIR, and CNN schemes have accomplished the best accuracy of 100%, 100%, 99.96%, and 99.94% whereas the FFT-DBN, CNN, and FFT-KNN techniques have exhibited considerable function with the accuracy of 99.76%, 99.94%, 99% respectively. Next, on the applied bearing dataset 10, the IIFD-SOIR, CNN, CNN2, and FFT-KNN frameworks have achieved the best accuracy of 99.97%, 99.94%, 99.94%, and 97.61% while the FFT-SVM, FFT-DBN, and FFT-SAE methods have depicted lower function with the accuracy of 87.80%, 96.19%, 95.49% respectively.

Average accuracy analysis of the IIFD-SOIR method with the previous approaches is also made. The experimental measures have revealed that the FFT-SVM technique has represented inferior function with the lower average accuracy of 97.106%. Simultaneously, the FFT-SVM technique has provided moderate function with an average accuracy of 97.827%. Besides, the FFT-DBN, FFT-SAE, and CNN2 methodologies have displayed considerable and closer average accuracy of 98.447%, 98.224%, and 98.084% correspondingly. Although the CNN approach has attained maximum average accuracy of 99.179%, the projected IIFD-SOIR framework has depicted a supreme function with an average accuracy of 99.647%.

Methods | Gearbox dataset | Bearing dataset | ||
---|---|---|---|---|

Training | Testing | Training | Testing | |

FFT-KNN | 90.84 | 86.35 | 98.28 | 97.83 |

FFT-SVM | 98.58 | 97.95 | 98.75 | 97.11 |

FFT-DBN | 100 | 98.27 | 99.46 | 98.45 |

FFT-SAE | 100 | 99.23 | 99.11 | 98.22 |

CNN | 99.33 | 98.30 | 99.61 | 99.18 |

CNN2 | 98.87 | 98.29 | 99.12 | 98.08 |

IIFD-SOIR | 100 | 99.60 | 99.83 | 99.65 |

On the applied bearing gearbox dataset the FFT-KNN method has accomplished poor function of an average training and testing accuracy of 98.28% and 97.83% correspondingly. Likewise, the FFT-SVM approach has attained a moderate function of average training and testing accuracy of 98.75% and 97.11% respectively. In line with this, the FFT-SAE model has obtained maximum performance of average training and testing accuracy of 99.11% and 98.22% respectively. Then, the CNN2 technique has accomplished moderate performance of average training and testing accuracy of 99.12% and 98.08% correspondingly. Meanwhile, the FFT-DBN approach has attained the least function of average training and testing accuracy of 99.46% and 98.45% respectively. At the same time, the CNN framework has accomplished the maximum function of average training and testing accuracy of 99.61% and 99.18% respectively. Additionally, the IIFD-SOIR model has achieved optimal performance of average training and testing accuracy of 99.83% and 99.65% respectively. From the above-mentioned experimental values, it is evident that the IIFD-SOIR model has resulted in effective performance over the compared methods due to the following reasons. The employed IRV2 model achieves a faster training rate with certainly better accuracy over the Inception v4 model. Besides, the parameter tuning of the DL model using the SFO algorithm also plays a vital role in the improved classification performance.

This paper has developed an IIFD-SOIR model to identify faults in rotating machinery. Initially, the data acquisition process takes place to collect the data. Then, the CWTS model is applied to preprocess and crop the vibration signals. Followed by, the SFO algorithm tuned Inception with ResNet v2 model is applied as a feature extractor. The parameter tuning of Inception with the ResNet v2 model takes place using the SFO algorithm. Finally, MLP is applied as a classification model to identify the different kinds of faults. Extensive experimentation takes place to ensure the outcome of the IIFD-SOIR model on the gearbox dataset and a motor bearing dataset. The experimental outcome indicated that the IIFD-SOIR model has reached a higher average accuracy of 99.6% and 99.64% on the applied gearbox dataset and bearing dataset. The IIFD-SOIR model can be employed as an appropriate tool for diagnosing faults in rotating machienry. In the future, the IIFD-SOIR model can be employed in real-time industries for diagnosis faults.