Intelligent industrial machine fault diagnosis method and device based on DenseNet neural network

Abstract

The application provides a kind of intelligent industrial machinery fault diagnosis method and device based on DenseNet neural network, and the method steps are as follows: collecting audio data related to the fault of the to-be-tested machine, analyzing the collected audio data, studying the law of fault waveform, summarizing the change phenomenon and fault reason of common fault type audio waveform, and pre-classifying the data set;Extract the Mel-frequency spectrum feature of the audio data;The new Mel-frequency spectrum feature extracted by the DenseNet neural network is subjected to secondary feature extraction, and the feature after dimension reduction obtained by dimension reduction processing of the new Mel-frequency spectrum feature;The feature after dimension reduction is used as the input sample of multi-classification SVM for intelligent fault diagnosis.The method provided by the application not only retains the time domain information, but also extracts the frequency domain information;At the same time, because industrial fault is often accompanied by the characteristics of small fault data, the SVM classifier with high small sample recognition degree is selected for fault classification, which effectively improves the fault recognition precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial fault diagnosis, and in particular to an intelligent industrial machine fault diagnosis method and device based on a DenseNet neural network. BACKGROUND

[0002] With the continuous updating of technology, the real-time of detection and the detection method put forward new requirements for the detection method. The method of fault detection is no longer constrained to contact identification methods such as sensors, and has gradually changed from contact to non-contact. Sound recognition of industrial faults has been recognized, accepted and developed by people. A large amount of information about the environment we live in is contained in sound. In actual industrial production, people can judge the running state of industrial equipment according to the sound of the equipment. For example, whether the motor is faulty can be judged according to the sound of the motor running, and whether the gear or track is faulty can be judged according to the sound when the gear or track rotates. In recent years, the deep learning method developed from artificial neural networks has made a lot of research results in the fields of sound recognition, visual target recognition and target detection. Sound recognition of faults has gradually become intelligent. The diversity of faults makes the fault characteristics, fault causes and the influence of fault feature points all have great differences. How to find the required recognition feature with the greatest influence has become a new research problem.

[0003] Data-driven modeling methods do not need to understand the specific mechanism of the industrial process. Compared with the mechanism modeling method, the data-driven modeling method has better universality. With the continuous development of deep learning technology, the data-driven modeling method has been greatly expanded. Compared with the traditional feature extraction method designed based on prior knowledge and analysis, deep learning as an end-to-end learning method can learn useful feature representations independently, without a large amount of prior knowledge and engineering skills. In addition, deep learning can transform or combine low-level features through deep network structures to obtain higher-level abstract features, thereby learning deep features with hierarchical structures, having better feature learning ability and the ability to obtain more essential data features. However, not all industrial data information is helpful for modeling and fault detection. Industrial data unrelated to modeling and fault detection will lead to low efficiency and poor generalization performance. SUMMARY

[0004] In order to solve the problem of low efficiency and poor generalization performance of the traditional fault detection and diagnosis method, the present application provides an intelligent industrial machine fault diagnosis method and device based on a DenseNet neural network.

[0005] According to one aspect of the present application, the present application provides an intelligent industrial machine fault diagnosis method based on a DenseNet neural network, comprising the following steps:

[0006] S1: obtaining audio data related to mechanical fault to be tested, building a database, and labeling and organizing the original database;

[0007] S2: preprocessing the collected audio data to obtain standardized fault audio data;

[0008] S3: extracting the mel-frequency spectrum features of the audio from the standardized fault audio data;

[0009] S4: building a DenseNet neural network;

[0010] S5: inputting the mel-frequency spectrum features into the DenseNet neural network for layer-by-layer unsupervised training, and after the training is completed, extracting new mel-frequency cepstral coefficients through the trained DenseNet neural network;

[0011] S6: performing feature dimension reduction on the new mel-frequency cepstral coefficients to obtain reduced features;

[0012] S7: inputting the reduced features into an SVM support vector machine for final fault diagnosis.

[0013] Preferably, S2 comprises:

[0014] S2.1: designing a high-pass filter:

[0015] H(z) = 1 - k * z^(-1)

[0016] wherein k is a filter coefficient, which is between 0.9-1.0 (preferably 0.95), and z represents the audio data to be processed;

[0017] S2.2: passing the audio data through the high-pass filter to obtain standardized fault audio data.

[0018] Preferably, the fully connected layer is removed from the original DenseNet121 model to obtain the DenseNet neural network.

[0019] Preferably, the DenseNet neural network comprises: 1 convolutional layer, 1 pooling layer, 4 dense blocks, and 3 transition layers; the convolutional layer, the pooling layer, and the dense blocks are connected in sequence, and one transition layer is connected between each two dense blocks.

[0020] Preferably, step S3 comprises:

[0021] S3.1: performing windowing and framing processing on the standardized fault audio data, and then performing FFT operation;

[0022] S3.2: performing logarithmic operation on the obtained feature vector, and then performing KL transformation;

[0023] S3.3: performing STFT operation on the obtained original audio sequence;

[0024] S3.4: constructing a mel filter;

[0025] S3.5: transposing the STFT transformation result, then performing matrix multiplication with the mel filter, and finally taking logarithm to obtain the mel spectrum feature of the audio.

[0026] Preferably, step S3 further comprises:

[0027] adjusting the extracted mel spectrum feature to scale the mel spectrum graph to a size of 224*224.

[0028] Preferably, in step S6, the T-SNE method is used to perform feature dimension reduction on the new mel cepstrum coefficient to obtain the reduced feature.

[0029] According to the second aspect of the present application, the present application further provides an intelligent industrial machine fault diagnosis device based on DenseNet neural network, comprising the following modules:

[0030] A data acquisition module is configured to acquire audio data related to the fault of the machine to be tested, build a database, and label and organize the original database.

[0031] A preprocessing module is configured to preprocess the collected audio data to obtain standardized fault audio data.

[0032] A first feature extraction module is configured to extract the mel spectrum feature of the audio from the standardized fault audio data.

[0033] A network building module is configured to build a DenseNet neural network.

[0034] A training and second feature extraction module is configured to input the mel spectrum feature into the DenseNet neural network for layer-by-layer unsupervised training, and after the training is completed, extract a new mel cepstrum coefficient through the trained DenseNet neural network.

[0035] A feature dimension reduction module is configured to perform feature dimension reduction on the new mel cepstrum coefficient to obtain a reduced feature.

[0036] A fault diagnosis module is configured to input the reduced feature into a SVM support vector machine for final fault diagnosis.

[0037] The technical solution provided by the present application has the following beneficial effects:

[0038] Firstly, the mel-frequency spectrum features of the audio data are preliminarily extracted; the mel-frequency spectrum features are secondarily extracted through a DenseNet neural network, the extracted new mel-frequency cepstrum coefficients are dimensionally reduced to serve as the dimensionally reduced features of the fault samples, the dimensionally reduced features are taken as the input samples of a multi-classification SVM to perform intelligent fault diagnosis, the problem of gradient disappearance is reduced, feature propagation is enhanced, feature reuse is encouraged, and the number of parameters is greatly reduced, and the SVM support vector machine is more suitable for a smaller training set and can reduce the overfitting phenomenon. BRIEF DESCRIPTION OF DRAWINGS

[0039] The specific effects of the present application will be further described below in combination with the drawings and embodiments, and the drawings are as follows:

[0040] Figure 1 is an execution flow diagram of the intelligent industrial mechanical fault diagnosis method based on the DenseNet neural network in the embodiment of the present application;

[0041] Figure 2 is a structural schematic diagram of the DenseNet neural network in the embodiment of the present application;

[0042] Figure 3 is a feature distribution diagram of the T-SNE dimensionally reduced features of the entire audio set in the embodiment of the present application;

[0043] Figure 4 is a feature distribution diagram of the T-SNE dimensionally reduced features of the fault training set in the embodiment of the present application;

[0044] Figure 5 is a feature distribution diagram of the T-SNE dimensionally reduced features of the fault test set in the embodiment of the present application;

[0045] Figure 6 is a structural diagram of the intelligent industrial mechanical fault diagnosis device based on the DenseNet neural network in the embodiment of the present application. DETAILED DESCRIPTION

[0046] In order to have a clearer understanding of the technical features, objects and effects of the present application, the specific embodiments of the present application will be described in detail with reference to the drawings.

[0047] Embodiment one:

[0048] Please refer to Figure 1 The embodiment of the present application provides an intelligent industrial mechanical fault diagnosis method based on a DenseNet neural network, including the following steps:

[0049] S1: acquiring audio data related to a mechanical fault to be measured, building a database, and labeling and arranging an original database;

[0050] S1 specifically includes:

[0051] S1.1: The acquired audio data is divided into a training set and a test set in a ratio of 7:3;

[0052] S1.2: The type of audio data, detection interval and other information are entered into an excel file;

[0053] S2: The collected audio data is preprocessed to obtain standardized fault audio data;

[0054] S2 specifically includes:

[0055] S2.1: A high-pass filter is designed:

[0056] H(z) = 1-k*z^(-1)

[0057] Wherein, k is the filter coefficient, which is between 0.9-1.0 (preferably 0.95 in this embodiment), and z represents the audio data to be processed;

[0058] S2.2: The audio data is passed through the high-pass filter to obtain standardized fault audio data.

[0059] S3: Extracting the Mel-frequency spectrum feature of the audio from the standardized fault audio data;

[0060] S3 specifically includes:

[0061] S3.1: The standardized fault audio data is windowed and framed, and then FFT operation is performed;

[0062] S3.2: The obtained feature vector is subjected to logarithmic operation, followed by KL transformation;

[0063] S3.3: The obtained original audio sequence is subjected to STFT operation;

[0064] S3.4: A Mel filter is constructed;

[0065] S3.5: The STFT transformation result is transposed, then multiplied by the Mel filter, and finally logarithm is taken to obtain the Mel-frequency spectrum feature of the audio.

[0066] S4: Building a DenseNet neural network;

[0067] Reference Figure 2 and Table 1, the DenseNet neural network suitable for the present application is constituted by applying and modifying the structure of DenseNet121.

[0068] The DenseNet neural network comprises 4 dense blocks (116 layers), 1 convolutional layer, 1 pooling layer, 3 transition layers, and a total of 120 weight layer groups, and removes a full connection layer;

[0069] Table 1. DenseNet121 neural network structure

[0070]

[0071]

[0072] From Figure 2 As can be seen from Table 1, the output of each layer is regarded as the input of the subsequent layer.

[0073] 1) After the data is input, the data (named x0) is first subjected to 1x1 convolution to adjust the channel number to 4*32, and then 3x3 convolution is used to obtain a 32-channel feature layer, at this time a feature layer x1 with a shape of (h, w, 32) is obtained.

[0074] 2) Stack x1 obtained after convolution and x0 of the initial input to obtain a new feature layer x2, the x2 feature layer will retain the features of the initial input x0 and also retain the x1 features after convolution processing.

[0075] 3) Repeat the convolution and stacking of 1) and 2), the features before convolution processing will be retained all the time, and the features after convolution processing will also be retained. When the network degree is continuously deepened, all the previous layers and the subsequent layers can be densely connected.

[0076] The network adds a combination function of 3 consecutive operations between layers, which obtains the feature mappings x0, x1,…, xi-1 of all layers before the i-th layer through batch processing normalization, ReLU, pooling and convolution. i-1 , as input, and is expressed as:

[0077] x i = H i ([x0,x1,…,x i-1 ]) (1)

[0078] In the formula, [x0,x1,…,x i-1 ] is the connection of the feature mappings of the 0th, 1st, 2nd,…, i-1th layers.

[0079] In order to facilitate down-sampling, the DenseNet neural network divides the network into multiple densely connected dense blocks, such as Figure 2The size of the feature map is changed by making a transition layer between the convolution and pooling layers. The advantages of this dense connection are: alleviating the problem of gradient disappearance, enhancing feature propagation, encouraging feature reuse, and greatly reducing the number of parameters, which can also reduce the phenomenon of overfitting on a smaller training set.

[0080] S5: input the mel spectrum feature into the DenseNet neural network for layer-by-layer unsupervised pre-training, and extract new mel cepstral coefficients through the trained DenseNet neural network after training is completed;

[0081] In this embodiment, the DenseNet121 is removed from the last 1000 classification fully connected layer, and a multi-classification SVM is added instead of the classification layer.

[0082] S6: performing feature dimension reduction on the new mel cepstral coefficients to obtain reduced features;

[0083] In S6, T-SNE is preferably used for dimension reduction to 2 dimensions.

[0084] The T-SNE dimension reduction method has good dimension reduction effect and outstanding visualization effect.

[0085] The core idea of T-SNE is to use a Gaussian distribution at high-dimensional features and a t-distribution at low-dimensional features, so that the distance can be described by a probability distribution. High-dimensional data is represented by x, x i represents the i-th sample, P ij represents the probability of the i-th sample being distributed around the sample j. Wherein formula (1): P ij is the conditional probability distribution at high-dimensional features, formula (2): Q ij is the conditional probability distribution obtained at low-dimensional features.

[0086]

[0087]

[0088] In the formula, σ is the Gaussian distribution variance corresponding to the data point x i , and k is the number of neighborhood points. The optimized gradient

[0089]

[0090] The algorithm flow of T-SNE is as follows:

[0091] (1) Set the input vector x and the output vector y;

[0092] (2) Set the parameters needed: iteration number N, learning rate L, and transformation momentum a(t);

[0093] (3) initialize y;

[0094] (4) iteratively compute the value of y as follows:

[0095]

[0096] By Figure 3 , 4 , the distribution of the feature points after dimension reduction in 5 can be observed to observe the difference between each feature.

[0097] S7: input the dimension-reduced features into the SVM support vector machine for final fault diagnosis.

[0098] In step S7, by adding the SVM support vector machine, the number of parameters is greatly reduced by using the small sample classification characteristics, and the overfitting phenomenon can also be reduced on a smaller training set.

[0099] The present application uses a nonlinear support vector machine, maps the training sample from the original space to a higher dimensional space through a nonlinear mapping function φ(x), so that the sample is linearly separable in this space. Let φ(x) represent the feature vector after mapping x, and the model corresponding to the division hyperplane in the feature space can be represented as:

[0100] f(x) = ωφ(x) + b

[0101] Where the variable ω reflects the complexity of the function, which is a linear combination of the nonlinear mapping function φ(x), and b is the bias. By converting the dual problem, the initial problem of solving ω and b is converted into a dual problem of solving α. Find all α (i.e. find all support vectors, dual variables), ω and b can be determined. Then the class of the data point can be determined by calculating the distance of the data point to the hyperplane.

[0102] The calculation process is as follows:

[0103] First, the minimum interval of the function is reduced to 1:

[0104]

[0105] s.t.1-y i (ωφ(x)+b)≤0,i=1,…,m

[0106] Its dual problem is α i and α j are its dual variables:

[0107]

[0108]

[0109] From the above formula, it can be seen that in the dual problem, there is φ(x i ) T φ(x j ), that is, the inner product after mapping to the feature space, so it is necessary to find a formula for calculating the inner product in the high-dimensional space, that is, the kernel function.

[0110] The selected kernel function of the application is a Gaussian kernel:

[0111]

[0112] K(xi,x j )≤φ(x i ), φ(x j )≥φ(x i ) T φ(x j )

[0113] The dual problem of the above formula is converted into:

[0114]

[0115]

[0116] The finally obtained classification function f(x) is:

[0117]

[0118]

[0119] Industrial production is often accompanied by various noises, so that the sample points cannot satisfy that the interval is greater than or equal to 1 or fall in the space of another class, so a slack variable ξ (ξ≥0) is introduced for each sample point here, so that it is greater than or equal to 1 after adding the slack variable ξ, and a penalty parameter C (C>0) is added, and the optimization target is changed to:

[0120]

[0121] s.t.1-y i (ω T x i +b)≤ξ,i=1,…,m,ξ≥0

[0122] The Lagrange function is constructed:

[0123]

[0124] Where μ i ≥0 is the Lagrange multiplier.

[0125] Finally, the Lagrange function is solved to obtain the values of alpha, omega and b in sequence.

[0126] Using SVM support vector machine for industrial fault diagnosis, it is found through experiments that the recognition accuracy of equipment fault state is as follows:

[0127] (1) The recognition rate of diagnosing whether the equipment is faulty is 100%, and the classification effect is as shown in Figure 3 ;

[0128] (2) The self-test recognition rate of the training set for diagnosing the specific fault of the equipment is 100%, and the classification effect is as shown in Figure 4 ;

[0129] (3) The test set test recognition rate of diagnosing the specific fault of the equipment is 87.5%, and the classification effect is as shown in Figure 5 ;

[0130] From the experimental results, it can be concluded that the present application has good industrial fault diagnosis capability and can achieve the expected goal.

[0131] Example 2

[0132] Referring to Figure 6 , the embodiment of the present application also provides an intelligent industrial machine fault diagnosis device based on DenseNet neural network, comprising the following modules:

[0133] The data acquisition module 01 is used for acquiring audio data related to the fault of the machine to be tested, building a database, and labeling and arranging the original database;

[0134] The preprocessing module 02 is used for preprocessing the collected audio data to obtain standardized fault audio data;

[0135] The first feature extraction module 03 is used for extracting the mel frequency spectrum feature of the audio from the standardized fault audio data;

[0136] The network building module 04 is used for building a DenseNet neural network;

[0137] The training and second feature extraction module 05 is used for inputting the mel frequency spectrum feature into the DenseNet neural network for layer-by-layer unsupervised pre-training, and after training, extracting new mel cepstrum coefficients through the trained DenseNet neural network;

[0138] The feature dimension reduction module 06 is used for performing feature dimension reduction on the new mel cepstrum coefficients to obtain reduced features;

[0139] The fault diagnosis module 07 is used for inputting the reduced features into the SVM support vector machine for final fault diagnosis.

[0140] It should be noted that, in the present document, the terms "comprising", "comprising" or any other variant thereof are intended to cover non-exclusive inclusions, so that a process, method, article or system that includes a list of elements not only includes those elements, but also includes other elements not explicitly listed, or inherent to such a process, method, article or system. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of other identical elements in the process, method, article or system that includes the element.

[0141] The above-mentioned embodiment numbers of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments. In the unit claims in which several devices are listed, several of these devices can be embodied by the same hardware item. The use of the words first, second, and third does not indicate any order, and these words can be interpreted as identifiers.

[0142] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for intelligent industrial machine fault diagnosis based on DenseNet neural network, characterized in that, The method comprises the following steps: S1: obtaining audio data related to mechanical fault to be tested, building a database, and labeling and organizing the original database; S2: preprocessing the collected audio data to obtain standardized fault audio data; S3: extracting the mel-frequency spectrum feature of the audio from the standardized fault audio data; comprising: S3.1: windowing and framing the standardized fault audio data, and then performing FFT operation; S3.2: performing logarithmic operation on the obtained feature vector, and then performing KL transformation; S3.3: performing STFT operation on the obtained original audio sequence; S3.4: constructing a mel filter; S3.5: transposing the STFT transformation result, then performing matrix multiplication with the mel filter, and finally taking the logarithm to obtain the mel-frequency spectrum feature of the audio; S4: building a DenseNet neural network; removing the fully connected layer from the original DenseNet121 model to obtain the DenseNet neural network; the DenseNet neural network comprises: 1 convolutional layer, 1 pooling layer, 4 dense blocks, and 3 transition layers; the convolutional layer, the pooling layer, and the dense blocks are connected in sequence, and a transition layer is connected between every two dense blocks; S5: inputting the mel-frequency spectrum feature into the DenseNet neural network for layer-by-layer unsupervised training, and after the training is completed, extracting new mel-frequency cepstral coefficients through the trained DenseNet neural network; S6: performing feature dimension reduction on the new mel-frequency cepstral coefficients to obtain reduced features; S7: inputting the reduced features into an SVM support vector machine for final fault diagnosis. 2.The intelligent industrial machine fault diagnosis method based on DenseNet neural network according to claim 1, wherein S2 Comprising: S2.1: designing a high-pass filter: H(z)=1-k*z^(-1) Wherein, k is the filter coefficient, and z represents the audio data to be processed; S2.2: passing the audio data through the high-pass filter to obtain standardized fault audio data. 3.The intelligent industrial machine fault diagnosis method based on DenseNet neural network according to claim 1, characterized in that, Step S3 further comprises: Adjusting the extracted mel-frequency spectrum feature to scale the mel-frequency spectrum graph to a size of 224*224. 4.The intelligent industrial machine fault diagnosis method based on DenseNet neural network according to claim 1, characterized in that, In step S6, the T-SNE method is used to perform feature dimension reduction on the new mel-frequency cepstral coefficients to obtain reduced features.

5. A DenseNet neural network-based intelligent industrial machine fault diagnosis device, characterized in that, Comprising the following modules: A data acquisition module for acquiring audio data related to mechanical fault to be tested, building a database, and labeling and organizing the original database; A preprocessing module for preprocessing the collected audio data to obtain standardized fault audio data; A first feature extraction module for extracting the mel-frequency spectrum feature of the audio from the standardized fault audio data; comprising: Windowing and framing the standardized fault audio data, and then performing FFT operation; Performing logarithmic operation on the obtained feature vector, and then performing KL transformation; Performing STFT operation on the obtained original audio sequence; Constructing a mel filter; Transposing the STFT transformation result, then performing matrix multiplication with the mel filter, and finally taking the logarithm to obtain the mel-frequency spectrum feature of the audio; The network building module is used for building a DenseNet neural network; a fully connected layer is removed from an original DenseNet121 model to obtain the DenseNet neural network; the DenseNet neural network comprises one convolutional layer, one pooling layer, four dense blocks and three transition layers; the convolutional layer, the pooling layer and the dense blocks are sequentially connected, and one transition layer is connected between each two dense blocks; The training and second feature extraction module is used for inputting the mel spectrum feature into the DenseNet neural network for layer-by-layer unsupervised training, and extracting new mel cepstral coefficients through the trained DenseNet neural network after the training is completed; The feature dimension reduction module is used for performing feature dimension reduction on the new mel cepstral coefficients to obtain reduced features; The fault diagnosis module is used for inputting the reduced features into an SVM support vector machine for final fault diagnosis.

Images