Multi-entropy fusion engine fault diagnosis method based on MSCNN

Abstract

The invention aims to construct a multi-entropy fusion engine fault diagnosis method based on MSCNN, and the method comprises the steps: carrying out the decomposition and preprocessing of vibration signal data through employing CEEMDAN, obtaining an IMF component, calculating the approximate entropy, permutation entropy and fuzzy entropy, and constructing a multi-entropy feature fusion sample matrix as a data sample; dividing the data sample into a training set and a test set; building a fault diagnosis model, and initializing parameters of the fault diagnosis model; disorganizing the training set, inputting the disorganized training set into the fault diagnosis model, and spreading the disorganized training set forwards to construct a cross entropy loss function; performing back propagation on the fault diagnosis model until a preset requirement is met, and obtaining a trained fault diagnosis model; and inputting the test set into the trained fault diagnosis model to complete engine fault diagnosis and classification identification.

Description

Technical Field

[0001] The present invention relates to the field of engine fault diagnosis, and specifically to a multi-entropy fusion engine fault diagnosis method based on MSCNN. Background Technique

[0002] As an indispensable core component in mechanical equipment, during the operation process with high load and high intensity, the engine not only has to bear huge load pressures but also needs to cope with various unpredictable external interferences. At the same time, due to the engine often being in a complex, changeable and even harsh working environment, its various components are prone to different forms of degradation and damage. To ensure the stable operation of mechanical equipment, complete production tasks on time, improve the economic benefits of enterprises and ensure the safety of employees, it is particularly important to conduct regular inspections and maintenance on the engine. Common engine failure forms include fracture failure, fatigue failure and wear failure of components, etc. If these problems are not dealt with in time, they may have a serious impact on equipment and production safety.

[0003] Two important processes in traditional fault diagnosis methods are signal feature extraction and state classification. At present, since the acquisition of vibration signals is relatively easy and vibration signals can more directly reflect the operating state of the engine, they are widely used in the feature extraction of the engine. With the development of the feature extraction field, multi-information fusion methods have been widely studied, and more accurate and valuable feature information can be obtained than single information.

[0004] Although significant progress has been made in engine fault diagnosis technology, it still faces multiple challenges. First, the operating conditions of the engine are complex and the environment is changeable, and the signal acquisition process is easily affected by noise interference, resulting in a decrease in the accuracy of fault feature extraction. Second, the fault modes of the engine are diverse and the feature information is high-dimensional. Traditional methods are prone to ignoring key features or introducing redundant information when processing multi-dimensional data, thus affecting the diagnostic effect. In addition, the differences in different engine structures and operating conditions make it difficult for a general diagnostic model to be fully adapted, and the lack of pertinence has become a key factor restricting the popularization of diagnostic technology.

[0005] In view of this, an engine fault diagnosis method that ensures the effectiveness of the extracted degradation features of the engine operating state and improves the fault recognition rate is urgently needed to be studied. Summary of the Invention

[0006] The purpose of the present invention is to construct a multi-entropy fusion engine fault diagnosis method based on MSCNN to significantly improve the accuracy and robustness of diagnosis.

[0007] The technical solution of the present invention is as follows:

[0008] A multi-entropy fusion engine fault diagnosis method based on MSCNN, comprising:

[0009] Collect the vibration signal data of the engine to be diagnosed;

[0010] Use CEEMDAN to decompose and preprocess the vibration signal data, calculate its approximate entropy, permutation entropy and fuzzy entropy after obtaining the IMF components, and construct a multi-entropy feature fusion sample matrix as the data sample;

[0011] Divide the data sample according to different operating states of the engine, and divide the data sample into a training set and a test set according to a certain proportion;

[0012] Build a fault diagnosis model and initialize the parameters of the fault diagnosis model;

[0013] Shuffle the training set and input it into the fault diagnosis model, and perform forward propagation to construct the cross-entropy loss function Loss, that is;

[0014]

[0015] In the formula: N is the number of categories in the model input; y i is the true label of the division; is the classification prediction probability of the model;

[0016] Use the optimization algorithm to perform backpropagation on the fault diagnosis model until the preset requirements are met to obtain a trained fault diagnosis model;

[0017] Save the model parameters in the trained fault diagnosis model, input the test set into the trained fault diagnosis model, and complete the engine fault diagnosis and classification recognition.

[0018] The steps for obtaining the approximate entropy include:

[0019] Define the time series X(i), and construct a vector given the pattern dimension q, that is:

[0020] X(n) = [x(n), x(n + 1),..., x(n + q - 1)];

[0021] where, X(n) is a q-dimensional vector; n = 1, 2,..., i - q + 1;

[0022] Define d[X(n), X(m)] as the maximum range of the distance between vectors X(n) and X(m), that is:

[0023]

[0024] Given a similarity tolerance r (r > 0), obtain the ratio of the maximum range of the distance to i - q + 1, perform a logarithmic operation on it, and then find its average value for all n, that is:

[0025]

[0026] Wherein: is the ratio of the maximum distance range to i - q + 1;

[0027] Let q = q + 1, and we get and C q+1 (r), and the approximate entropy is: A ApEn (q, r, i) = φ q (r) - φ q+1 (r).

[0028] The steps for obtaining the permutation entropy include:

[0029] Perform phase space reconstruction on the original time series and convert it into a matrix Y with j rows and q columns as follows:

[0030]

[0031] Wherein: q is the embedding dimension; t is the delay factor;

[0032] Sort the elements in each row of the matrix Y in ascending order, and let the indices of the sorted elements be λ α , α = 1, 2,..., q. Through sorting, the indices in each row of the matrix represent the permutation pattern using the following formula:

[0033] S(w) = (λ 1 , λ 2 ,... λ q ), w = 1, 2,..., j, j ≤ q!;

[0034] Wherein: there are q! different permutation patterns for the q - dimensional phase space mapping;

[0035] The permutation entropy of the time series is obtained as:

[0036]

[0037] Wherein: P w is the occurrence probability that the number of occurrences of each permutation pattern is divided by the total number of occurrences of q! different permutation patterns.

[0038] The steps for obtaining the fuzzy entropy include:

[0039] Perform phase space reconstruction on the original time series to obtain the time series Y, that is:

[0040] Y = [x(i), x(i + 1),, x(i + q - 1),] - x 0 (i);

[0041]

[0042] where: q is the embedding dimension; x 0 (i) is the mean value;

[0043] Define the distance between the two decomposed time series Y(i) and Y(j) as That is:

[0044]

[0045] Introduce the fuzzy membership function and use the fuzzy function to calculate the similarity between two time series, that is:

[0046]

[0047] where: r is the similarity tolerance; i, j = 1, 2,..., K - m + 1, and i ≠ j; K is the length of the time series;

[0048] The fuzzy entropy of the original time series is:

[0049]

[0050] where: Ψ q (r) is the defined function.

[0051] The using of CEEMDAN to decompose and preprocess the vibration signal data includes: the step of adding Gaussian white noise to the vibration signal in a step-by-step addition manner, specifically including:

[0052] Add Gaussian white noise to the original signal X(n) and decompose it, that is:

[0053] E(x(n) + ε 0 ω i (n)) = C 1 (n) + r(n);

[0054] where: Ei() is the i-th IMF component of EMD; ε o ω i (n) is the Gaussian white noise; C 1 (n) is the first-order IMF component; r(n) is the residual signal;

[0055] Perform i times of EMD decomposition on the original signal to obtain the average value of N first-order IMFs That is:

[0056]

[0057] After removing the first-order modal component, add Gaussian white noise to the original signal again to obtain a new signal, and perform EMD decomposition on it again to obtain the second-order modal component, that is:

[0058]

[0059] Add Gaussian white noise to the original signal again to obtain a new signal, and perform EMD decomposition on it again to obtain K-order modal components. When the residual signal obtained is a monotonic function, a total of K IMF components and a residual signal are obtained, that is:

[0060]

[0061] The fault diagnosis model includes:

[0062] The fault diagnosis model is an MSCNN network, and its structure and parameters are as follows:

[0063]

[0064] The training process of the fault diagnosis model includes:

[0065] The data input passes through a convolutional kernel with a smaller size to extract local features, which are composed of three groups of convolutional layers and pooling layers in total;

[0066] Use three parallel larger convolutional layers and pooling layers to construct a feature fusion module to extract comprehensive feature information; and after extracting the comprehensive feature information, continue with a convolutional kernel with a smaller size for sampling to complete the learning process;

[0067] After completing the learning, summarize the local feature information and comprehensive feature information to the fully connected layer, and finally realize the diagnosis and classification of faults through the Softmax classifier.

[0068] Add a Dropout layer between the fully connected layer and the Softmax classifier to avoid overfitting problems caused by too many parameters.

[0069] The beneficial effects of the present invention at least include:

[0070] The method described in the present invention combines multi-entropy fusion and multi-scale convolutional neural network to solve the problems of insignificant feature extraction effect, high dimension of neural network input information, and large diagnostic error in the current research in the field of fault diagnosis and classification. Since each entropy has different feature representation effects, some operating states are relatively close in the calculation results of one entropy value, but have obvious differences in the other two cases, and the three entropy values complement each other and cross-verify. Therefore, a single entropy feature cannot effectively distinguish different operating states and faults, and the fusion feature matrix formed by fusing the three entropy features can effectively distinguish the operating states of the engine, providing a basis and data foundation for the next step of fault diagnosis. Description of the Drawings

[0071] Figure 1 It is the basic model structure of CNN;

[0072] Figure 2 It is the network structure of MSCNN;

[0073] Figure 3 It is the engine fault diagnosis flow chart. Specific implementation manners

[0074] The present application will be further described below with reference to the accompanying drawings.

[0075] Currently, engine fault diagnosis methods are mainly divided into model-based methods, signal-based methods, and data-driven methods. The model-based method diagnoses faults by establishing an engine physical model and using theoretical derivation and mathematical analysis. Its advantage is high accuracy, but it has strict requirements for modeling accuracy and is difficult to adapt to complex working conditions. The signal-based method analyzes the signal characteristics such as vibration, sound wave, temperature, and pressure collected during the operation of the engine to identify potential faults. This method has the characteristics of strong real-time performance and wide applicability, but its effect is limited in an environment with large noise interference. In recent years, with the development of big data technology and artificial intelligence, the data-driven method has rapidly emerged. By using machine learning and deep learning algorithms to mine and analyze massive operation data, it can realize fault prediction and health management of complex systems.

[0076] To ensure the effectiveness of the extracted degradation features of the engine operation state, it is necessary to extract the multi-domain features of the signal during the feature extraction process; for this purpose, an engine fault diagnosis method based on multi-entropy fusion and MSCNN is proposed; specifically, three groups of parallel convolutional layers and pooling layers are added in the middle of the traditional CNN to enhance the non-linear expression ability of the model. Compared with the traditional CNN, the MSCCN model can capture more fault information and then transmit it to the subsequent layers, enhancing the non-linear expression ability of the model while only adding a small number of model parameters to the network. With a shorter training time, it has a higher fault recognition rate compared with other fault diagnosis and recognition models.

[0077] Concept description:

[0078] Entropy is a thermophysics concept used to characterize the amount of information. In recent years, many experts and scholars have introduced entropy features into the field of feature extraction to realize the dataization of the operation state by calculating various entropy values of vibration signals. As a dimensionless index characterizing the complexity of time series, the larger the entropy value, the greater the disorder complexity of the signal. The degree of signal complexity is of great significance in the state evaluation and fault diagnosis of engines, can resist the interference of environmental factors, and has good applications in the field of feature extraction.

[0079] Specific embodiment I:

[0080] The present invention provides an embodiment: a multi-entropy fusion engine fault diagnosis method based on MSCNN, including: collecting vibration signal data of the engine to be diagnosed; using CEEMDAN to decompose and preprocess the vibration signal data to obtain the approximate entropy, permutation entropy, and fuzzy entropy of the IMF components, and constructing a multi-entropy feature fusion sample matrix as data samples; dividing the data samples according to different operating states of the engine, and dividing the data samples into a training set and a test set according to a ratio; building a fault diagnosis model and initializing the parameters of the fault diagnosis model; shuffling the training set and inputting it into the fault diagnosis model, and performing forward propagation to construct a cross-entropy loss function; using an optimization algorithm to perform backpropagation on the fault diagnosis model until a preset requirement is met to obtain a trained fault diagnosis model; saving the model parameters in the trained fault diagnosis model, inputting the test set into the trained fault diagnosis model, and completing engine fault diagnosis and classification recognition.

[0081] Specifically, the calculation method of approximate entropy (ApEn) is to calculate the pattern self-similarity of the time series and use the change of the approximate entropy value to achieve the purpose of identifying the change of the signal sequence. It has the advantages of less data calculation amount and can meet most time series, and the obtained result is effective. The calculation method is as follows:

[0082] (1) Define the time series X(i), and construct a vector given the pattern dimension q, that is:

[0083] X(n) = [x(n), x(n + 1),..., x(n + q - 1)]

[0084] Where: X(n) is a q-dimensional vector; n = 1, 2,..., i - q + 1.

[0085] (2) Define d[X(n), X(m)] as the maximum range of the distance between vectors X(n) and X(m), that is:

[0086]

[0087] (3) Given a similarity tolerance r (r > 0), calculate the ratio of the maximum range of the distance to i - q + 1, perform a logarithmic operation on it, and then find its average value for all n, that is:

[0088]

[0089] Where: is the ratio of the maximum range of the distance to i - q + 1.

[0090] (4) Let q = q + 1, and repeat the above steps to obtain and C q+1 (r), for the actual sequence, i cannot approach infinity, so its approximate entropy can be simplified as:

[0091] A ApEn (q, r, i) = φ q (r) - φ q+1 (r)

[0092] Permutation Entropy (PmEn) improves the calculation efficiency and anti-interference ability by introducing the idea of permutation, and has high sensitivity to mutations in the signal. Its calculation method is as follows:

[0093] (1) Perform phase space reconstruction on the original time series and transform it into a matrix Y with j rows and q columns as follows:

[0094]

[0095] Where: q is the embedding dimension; t is the delay factor (the value should be a positive integer. If t = 1, the sequence is defined the same as the approximate entropy sequence).

[0096] (2) Reorder the elements in each row of the matrix in ascending order. Let the indices of the elements after sorting be λα (α = 1, 2,..., q). Through sorting, the indices of each row in the matrix can be expressed in the following formula for the permutation method, that is:

[0097] S(w) = (λ 1 , λ 2 , λ q ), w = 1, 2,..., j, j ≤ q!

[0098] Where: there are q! different permutation methods for the q-dimensional phase space mapping.

[0099] (3) The permutation entropy of the time series is obtained as:

[0100]

[0101] Where: Pw is the occurrence probability that the number of occurrences of each permutation method is divided by the total number of occurrences of q! different permutation methods.

[0102] Fuzzy Entropy (FuEn) introduces a fuzzy membership function on the basis of sample entropy. After introducing this idea, the algorithm can more directly characterize the internal characteristics of signal information. Its calculation method is as follows:

[0103] (1) Perform phase space reconstruction on the original time series to obtain the time series Y, that is:

[0104] Y = [x(i), x(i + 1),..., x(i + q - 1)] - x 0 (i)

[0105]

[0106] Where: q is the embedding dimension; x 0 (i) is the mean value.

[0107] (2) Define the distance between the two decomposed time series Y(i) and Y(j) as That is:

[0108]

[0109] (3) Introduce the fuzzy membership function and calculate the similarity between the two time series using the fuzzy function, that is:

[0110]

[0111] Where: r is the similarity tolerance. i, j = 1, 2,..., K - m + 1, and i ≠ j; K is the length of the time series.

[0112] (4) The fuzzy entropy of the original time series is:

[0113]

[0114] Where: Ψ q (r) is the defined function.

[0115] Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is further optimized on the basis of the improved Ensemble Empirical Mode Decomposition. Gaussian white noise is added to the vibration signal in a step-by-step manner. This method can effectively remove the residual Gaussian white noise in the IMF components, reduce the reconstruction error, and has better feature extraction effect. The calculation method of CEEMDAN is as follows:

[0116] (1) Add Gaussian white noise to the original signal X(n) and decompose it, that is:

[0117] E(x(n) + εφ 0 ω i (n)) = C 1 (n) + r(n)

[0118] Where: Ei() is the i-th IMF component of EMD; εφ o ω i(n) is Gaussian white noise; C1(n) is the first-order IMF component; r(n) is the residual signal.

[0119] (2) Perform i times of EMD decomposition on the original signal, and calculate the average value of N first-order IMFs That is:

[0120]

[0121] (3) After removing the first-order modal component, add Gaussian white noise to the signal again to obtain a new signal, and perform EMD decomposition on it again to obtain the second-order modal component, that is:

[0122]

[0123] (4) Repeat steps (1)-(3) until the obtained residual signal is a monotonic function, and the algorithm ends. At this time, a total of K IMF components and residual signals are obtained, that is:

[0124]

[0125] CEEMDAN solves the problems of endpoint effect and modal aliasing in EMD and EEMD by adding white noise step by step, improves the signal scale feature extraction ability, and provides support for the next multi-entropy fusion matrix.

[0126] In this embodiment, a Convolutional Neural Network (CNN) is used as the model, and its basic structure is as Figure 1 shown, Figure 1 shows the main structural components of the CNN. Among them, the main function of the input layer is to receive data, which are often expressed as discrete matrix forms in the computer. In the CNN, in order to effectively extract the features of the data and perform dimensionality reduction, multiple convolutional layers and pooling layers are usually designed to be stacked alternately to gradually deepen the understanding and processing of the input data. The main role of the alternating stacking of convolutional layers and pooling layers is to compress the data by performing a pooling operation after convolution. The convolution ability is mainly determined by the number of convolutional kernels. Since the CNN is often composed of multiple convolutional layers, the convolutional kernel parameters between different convolutional layers are also different. Appropriate parameter selection can improve the data extraction ability of the CNN. After the data is processed by the convolutional layer and the pooling layer, it will further pass through the fully connected layer to extract and integrate key features. Finally, these features are classified through the output layer to obtain the final classification result.

[0127] (1) Convolutional layer

[0128] The convolutional layer is a core component of the CNN, and its role is to extract key features from the input data. The convolutional kernel exists in the form of a two-dimensional matrix, and the sizes and parameter settings of different convolutional kernels are also different. The essence of the convolutional operation is to multiply the elements of the convolutional kernel and the local area of the input data point by point and add up the products to form a new value. This operation will continuously slide on the input data matrix according to the preset stride and repeat the convolutional calculation.

[0129] The number of convolutional operations is affected by the sliding stride. Each time it slides, a convolutional operation will be performed in the corresponding area and the result will be input into the feature map. After the sliding operation is completed, all the operation results together constitute a feature map. Feature map Q out The size of can be calculated by the following formula:

[0130]

[0131] In the formula: Q in is the size of the input data; W is the size of the padding for the boundary of the input data; F is the size of the convolutional kernel; N is the sliding stride.

[0132] The core operation concept of the convolutional layer lies in local perception and weight sharing. While ensuring the training accuracy of the model, it significantly reduces the number of weight parameters, thereby reducing the training difficulty of the network.

[0133] The idea of local perception is inspired by the working principle of the human visual nerve, emphasizing only focusing on local information rather than global data. Compared with traditional neural networks, the convolutional layer no longer requires each input data point to establish a connection with a neuron. This local connection method greatly reduces the number of weights, thereby improving the training efficiency of the neural network. In addition, the neurons in the convolutional layer share the same weights, and these weights match the size of the convolutional kernel, further reducing the number of weight parameters.

[0134] (2) Activation function

[0135] In order to enhance the non-linear fitting ability of the neural network, when the input data completes the convolutional operation, an activation function is usually introduced. The introduced activation function must meet three requirements: First, the activation function must be non-linear, which helps to increase the depth of the neural network and enables it to fit more complex functional relationships; Second, the activation function must be differentiable, so that it can ensure that the network parameters can be optimized through the gradient descent algorithm during the training process; Finally, the activation function also needs to have the characteristic of simple calculation to reduce the computational burden.

[0136] (3) Pooling layer

[0137] The main function of the pooling layer is to downsample the input feature map and reduce the spatial dimension. The dimensionality reduction operation not only helps to reduce the amount of data, computational volume and complexity, but also improves the fault tolerance of the network. During the pooling process, the sliding window method is widely used, and different methods can be selected according to actual needs to extract data within the pooling area.

[0138] Among the pooling methods, the max-pooling method and the average-pooling method are commonly used. The max-pooling method selects the maximum value within the sliding window as the output, and arranges these output values in order in the new dimensionality reduction matrix, and finally completes the traversal of the entire input feature map to obtain the output result. The average-pooling method focuses on calculating the arithmetic mean of the data within the sliding window and uses it as the output value.

[0139] (4) Fully connected layer

[0140] After the feature maps obtained by the data through convolution and pooling processing are mapped to the sample space, the fully connected layer is responsible for converting these feature maps into one-dimensional vectors. In a CNN, the classifier consists of multiple fully connected layers, forming a multi-layer perceptron, which is responsible for accurate discrimination based on the extracted feature vectors. Each neuron in the fully connected layer is connected to the neurons in the previous layer according to its weight coefficient, and its calculation formula is as follows:

[0141] y = f(Wx + b)

[0142] In the formula: f(·) is the activation function; x is the input of the fully connected layer; W is the weight; b is the bias.

[0143] The MSCNN structure proposed by the present invention is as follows:

[0144] Although CNN has achieved remarkable achievements in many fields, it still has certain limitations. Among them, the size of the convolutional kernel has a non-negligible impact on the performance of CNN. The size of the convolutional kernel not only determines the granularity of the features that the network can capture, but also directly relates to the size of its receptive field. However, if only a single-size convolutional kernel is used, CNN may only be able to capture the feature information at a certain specific level, resulting in incomplete or one-sided information extraction. To solve this problem and extract the feature information in the data more fully and comprehensively, a multi-scale convolutional neural network is proposed. MSCNN uses convolutional kernels of different sizes, enabling the network to perform feature extraction simultaneously within different receptive field ranges, thereby achieving multi-level and multi-angle recognition and extraction of the input data, and has been successfully applied to the field of fault diagnosis.

[0145] The single convolutional kernel used in the CNN network cannot extract local and global features simultaneously. To improve the multi-scale extraction ability of the model, the MSCNN structure proposed in this embodiment is as Figure 2As shown, the specific model parameters are shown in Table 1. For multi-scale information extraction, to ensure the model's ability to capture fault features, two sizes of convolutional kernels are selected for convolutional processing, and the pooling method adopts the maximum pooling method. The specific process is as follows: The data input first passes through small-sized convolutional kernels to extract local features, which consists of three groups of convolutional layers and pooling layers in total; then it enters the feature fusion module, which is composed of three parallel structures of convolutional layers and pooling layers, and large convolutions are used in this part to increase the receptive field of the model; after comprehensively extracting feature information, a small-sized convolutional kernel is connected for sampling. After learning, the features learned at each scale are summarized to the fully connected layer, and finally, the fault diagnosis and classification are realized through the Softmax classifier. A Dropout layer is added between the fully connected layer and the Softmax classifier to avoid overfitting problems caused by too many parameters.

[0146] Table 1: MSCNN model parameters:

[0147]

[0148] The engine fault diagnosis method based on multi-entropy fusion and MSCNN described in this embodiment is mainly divided into three links when in use: vibration signal preprocessing, model training, and testing. The flow chart is as Figure 3 shown.

[0149] Specifically, it includes:

[0150] (1) Use CEEMDAN to decompose and preprocess the engine vibration signal data, calculate the approximate entropy, permutation entropy, and fuzzy entropy of the IMF components, and construct a multi-entropy feature fusion sample matrix.

[0151] (2) For different operating states, divide the data samples and divide the data samples into a training set and a test set according to a certain proportion.

[0152] (3) Build a fault diagnosis model and initialize the parameters of the model.

[0153] (4) Shuffle the labeled training set and input it into the MSCNN, and perform forward propagation to construct a cross-entropy loss function.

[0154] (5) Use the optimization algorithm to perform backpropagation on the model, set the parameters of the network. When the training requirements are met, the training of the MSCNN is completed; otherwise, return to step (4).

[0155] (6) Save the trained model parameters, input the test set into the trained MSCNN, and complete the engine fault diagnosis and classification recognition.

[0156] The above are only several specific implementation scenarios of the present invention. However, the present invention is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention. The above serial numbers of the present invention are only for description and do not represent the advantages or disadvantages of the implementation scenarios.

Claims

1. A multi-entropy fusion engine fault diagnosis method based on MSCNN, characterized in that: include: Collect vibration signal data of the engine to be diagnosed; The vibration signal data is decomposed and preprocessed by using CEEMDAN, and the approximate entropy, permutation entropy and fuzzy entropy of the IMF components are calculated to construct a multi-entropy feature fusion sample matrix as a data sample; According to different operating states of the engine, the data samples are divided to obtain a training set and a test set; Building a fault diagnosis model and initializing parameters of the fault diagnosis model; The training set is shuffled and input into the fault diagnosis model, and propagated forward to construct a cross entropy loss function; Back-propagating the fault diagnosis model until it meets preset requirements to obtain a trained fault diagnosis model; The model parameters in the trained fault diagnosis model are saved, and the test set is input into the trained fault diagnosis model to complete engine fault diagnosis and classification identification.

2. The multi-entropy fusion engine fault diagnosis method based on MSCNN according to claim 1 is characterized in that: The step of obtaining the approximate entropy comprises: Define the time series X(i), construct a vector given the pattern dimension q, that is: X(n)=[x(n),x(n+1),...,x(n+q-1)]; Where X(n) is a q-dimensional vector; n = 1, 2, ..., i-q + 1; Define d[X(n),X(m)] as the maximum value of the range of the distance between vectors X(n) and X(m), that is: Given a similarity tolerance r (r>0), obtain the ratio of the maximum value of the distance range to i-q+1, perform a logarithmic operation on it, and then calculate its average value for all n, that is: Where: It is the ratio of the maximum value of the distance range to i-q+1; Let q = q + 1, we get and C q+1 (r), the approximate entropy is: A ApEn (q,r,i)=φ q (r)-φ q+1 (r).

3. The multi-entropy fusion engine fault diagnosis method based on MSCNN according to claim 1 is characterized in that: The step of obtaining the permutation entropy comprises: The original time series is reconstructed in phase space and transformed into the following matrix Y with j rows and q columns, namely: Where: q is the embedding dimension; t is the delay factor; Rearrange the elements of each row in the matrix Y in ascending order, and let the index of each element after arrangement be λ α , α=1,2,…,q, through sorting, the index of each row in the matrix is ​​expressed by the following formula, namely: S(w)=(λ1,λ2,…λ q ),w=1,2,…j,j≤q!; Where: There are q! different arrangements of the q-dimensional phase space mapping; The permutation entropy of the time series is obtained as: Where: P w The probability of each arrangement being present divided by the total number of q! different arrangements being present is q! 4. The multi-entropy fusion engine fault diagnosis method based on MSCNN according to claim 1 is characterized in that: The step of obtaining the fuzzy entropy comprises: The original time series is reconstructed in phase space to obtain the time series Y, that is: Y=[x(i),x(i+1),...,x(i+q-1),]-x0(i); Where: q is the embedding dimension; x0(i) is the mean; The distance between the two decomposed time series Y(i) and Y(j) is defined as Right now: The fuzzy membership function is introduced and the similarity between two time series is calculated using the fuzzy function, namely: Where: r is the similarity tolerance; i, j = 1, 2, ..., K-m + 1, and i ≠ j; K is the length of the time series; The fuzzy entropy of the original time series is: Where: q (r) is the defined function.

5. The multi-entropy fusion engine fault diagnosis method based on MSCNN according to claim 1 is characterized in that: The step of decomposing and preprocessing the vibration signal data by using CEEMDAN includes: adding Gaussian white noise to the vibration signal in a step-by-step manner, specifically including: Add Gaussian white noise to the original signal X(n) and decompose it, that is: E(x(n)+ε0ω i (n))=C1(n)+r(n); Where: Ei() is the i-th IMF component of EMD; ε o ω i (n) is Gaussian white noise; C1(n) is the first-order IMF component; r(n) is the residual signal; Perform EMD decomposition on the original signal i times to obtain the average value of N first-order IMFs Right now: After removing the first-order modal component, Gaussian white noise is added to the original signal again to obtain a new signal, and the new signal is decomposed by EMD again to obtain the second-order modal component, namely: Gaussian white noise is added to the original signal again to obtain a new signal, and EMD decomposition is performed again to obtain K-order modal components. When the residual signal is a monotonic function, a total of K IMF components and residual signals are obtained, namely:

6. The multi-entropy fusion engine fault diagnosis method based on MSCNN according to claim 1 is characterized in that: The fault diagnosis model comprises: The fault diagnosis model is a MSCNN network, and its structure and parameters are as follows:

7. The multi-entropy fusion engine fault diagnosis method based on MSCNN according to claim 6 is characterized in that: The training process of the fault diagnosis model includes: The data input passes through a smaller convolution kernel to extract local features. It consists of three groups of convolution layers and pooling layers. A feature fusion module is constructed using three parallel large convolutional layers and pooling layers to extract comprehensive feature information. After extracting comprehensive feature information, a smaller convolution kernel is used for sampling to complete the learning process. After learning is completed, the local feature information and the overall feature information are aggregated to the fully connected layer, and finally the fault diagnosis and classification are realized through the Softmax classifier.

8. The multi-entropy fusion engine fault diagnosis method based on MSCNN according to claim 7 is characterized in that: include: A Dropout layer is added between the fully connected layer and the Softmax classifier to avoid overfitting problems caused by too many parameters.

Images