A multi-channel fault diagnosis method based on DS-VMD deep convolutional neural network

Abstract

The application discloses a kind of multi-channel fault diagnosis methods based on DS-VMD deep convolutional neural network, it is related to mechanical fault diagnosis technical field.The application includes: S1, crane multi-channel acoustic emission AE monitoring signal is collected, and it is divided into signal sub-sequence sample set according to industrial frequency;S2, the signal sub-sequence sample set is decomposed to obtain several IMF components by DS-VMD;S3, the kurtosis of each IMF component is calculated, and the kurtosis weight of IMF component is set according to the kurtosis value, reconstructs time domain signal based on the kurtosis weight of IMF component, obtains multi-channel fault feature;S4, improved deep convolutional neural network is constructed;S5, the multi-channel fault feature extracted by DS-VMD is input into network to train and test, and the intelligent diagnosis of multi-channel fault is completed.The application can guarantee the diagnosis accuracy at the same time, improve the intelligent level of hoisting system, provide strong technical support for the health management and fault prediction of equipment.

Description

Technical Field

[0001] This application relates to the field of mechanical fault diagnosis technology, and in particular to a multi-channel fault diagnosis method based on DS-VMD deep convolutional neural network. Background Technology

[0002] As a core lifting equipment in modern engineering construction, cranes operate under complex conditions and bear high loads. Key load-bearing components such as the balance beam, A-frame, tie rods, and boom are prone to localized damage. This can lead to equipment downtime or, in severe cases, major safety accidents, while also increasing maintenance costs and shortening equipment lifespan. Therefore, developing high-precision, highly versatile intelligent fault diagnosis and structural health monitoring for cranes under complex operating conditions is of significant practical importance for ensuring the safe operation of lifting equipment, reducing maintenance costs, and extending service life.

[0003] Acoustic emission (AE) technology is a dynamic non-destructive testing method that uses transient elastic waves generated by the local energy release of materials to characterize the damage state of structures. Acoustic emission parameters such as amplitude, count, and energy can effectively reflect the stress evolution and damage degree of components. It has been applied in damage monitoring and fault diagnosis in fields such as rock, bearings, wood, and motor rotors. However, the application of acoustic emission technology in the fault diagnosis of critical crane structures is currently limited, and there is a lack of targeted and robust acoustic emission feature extraction methods, making it difficult to meet the engineering requirements for accurate fault diagnosis of lifting equipment.

[0004] Due to the coupling effect of multiple parameters such as lifting load, span, and slewing frequency, the acoustic emission signals of cranes are nonlinear, non-stationary, and have strong background noise characteristics, making fault features easily submerged by noise. Traditional signal processing methods such as wavelet transform, EEMD, LMD, and intrinsic time scale decomposition are mostly limited to single-channel data analysis and cannot fully utilize the fault information contained in the multiple sensor channels of the crane. Although variational mode decomposition (VMD) is superior to EMD and LMD in terms of mode mixing suppression, signal decomposition, and noise resistance, its decomposition effect is highly dependent on the number of modes K and the penalty factor α. Inappropriate parameter selection can easily lead to under-decomposition or over-decomposition, which seriously restricts the accuracy of subsequent acoustic emission feature extraction and diagnosis.

[0005] Meanwhile, existing deep learning fault identification models such as DBN, RNN, and GAN have inherent defects: average pooling assigns equal weights to features, easily blurring key fault features; max pooling only retains extreme value information, resulting in the loss of a large number of effective features, making it difficult to adapt to the dynamic feature mining and accurate fault mode identification of multi-channel acoustic emission signals of cranes. In summary, existing technologies suffer from insufficient utilization of multi-channel information, strong reliance on manual methods for VMD parameters, poor robustness of feature extraction, and severe feature loss in fault identification models. Therefore, there is an urgent need to develop an optimized decomposition and intelligent fault diagnosis method adapted to multi-channel acoustic emission signals of cranes. Summary of the Invention

[0006] In view of this, the technical objective of this application is to propose a multi-channel fault diagnosis method based on DS-VMD deep convolutional neural networks, aiming to solve the technical problems existing in current fault diagnosis methods for lifting systems, such as non-stationarity, difficulty in multi-channel signal processing, incomplete feature extraction, and insufficient automated identification capabilities. By introducing the DS-VMD method, fault information in multi-channel acoustic emission signals is effectively extracted, and signal features are automatically learned through deep convolutional neural networks to achieve accurate identification of fault modes in lifting systems. This method not only makes full use of the correlation between multi-channel signals but also overcomes the limitations of traditional methods in processing complex signals, improving the accuracy and robustness of fault diagnosis. Ultimately, this invention can improve the intelligence level of lifting systems while ensuring diagnostic accuracy, providing strong technical support for equipment health management and fault prediction.

[0007] To achieve the above objectives, this application adopts the following technical solution: A multi-channel fault diagnosis method based on DS-VMD deep convolutional neural network includes the following steps: S1. Collect multi-channel acoustic emission (AE) monitoring signals from the crane and divide them into signal sub-sequence sample sets according to industrial frequencies; S2. The differential search DS algorithm is used to adaptively optimize the number of modes K and the penalty factor α of variational mode decomposition (VMD), and a fitness function based on envelope entropy is constructed. The signal subsequence sample set is then decomposed by DS-VMD to obtain several IMF components. S3. Calculate the kurtosis of each IMF component, set the kurtosis weight of the IMF component according to the kurtosis value, reconstruct the time domain signal based on the kurtosis weight of the IMF component, and obtain the multi-channel fault characteristics. S4. Construct an improved deep convolutional neural network; S5. Input the multi-channel fault features extracted by DS-VMD into the network for training and testing, update the network weights using backpropagation, automatically identify different fault modes of the crane, and complete intelligent diagnosis of multi-channel faults.

[0008] Optionally, the step of performing DS-VMD decomposition on the multi-channel acoustic emission signal to obtain several IMF components includes: The differential search (DS) algorithm is used to simulate the seasonal migration behavior of the population. With the goal of minimizing the envelope entropy of the decomposed IMF components, the optimal combination of VMD parameters is adaptively searched to obtain several IMF components.

[0009] Optionally, the optimization objective of minimizing the envelope entropy of the decomposed IMF components includes: Select the envelope sequence after signal decomposition The entropy is used as the fitness value; For a given original acoustic emission signal s(j), its envelope entropy Calculated using the following formula: ; in, ;a ; a(j) is the envelope spectrum of s(j) obtained through the Hilbert transform. It is the normalized form of a(j); Where j is the total number of sampling points, and j is the index of the sampling point; The fitness function of the DS algorithm is: ; in, It is the envelope entropy of the k-th IMF.

[0010] Optionally, the improved deep convolutional neural network includes: Small-scale convolutional pooling layers are used to replace traditional average pooling and max pooling layers, combined with 3×3 small-scale convolutional kernels, depthwise separable convolution, and global average pooling.

[0011] Optionally, the small-scale convolutional pooling uses a small-scale convolutional layer with a stride of 2 and an activation function of ReLU for downsampling, and retains the feature information of all neurons in the region through adaptive weights of the convolutional kernel; Depth-separable convolution and channel-wise spatial convolution are used to replace tiling and fully connected layers with global average pooling.

[0012] As can be seen from the above technical solution, compared with the prior art, the multi-channel fault diagnosis method based on DS-VMD deep convolutional neural network of this application has the following beneficial effects: The signal decomposition effect is excellent. The differential search algorithm is used to adaptively optimize the key parameters of VMD, avoiding over-decomposition or under-decomposition caused by manual parameter selection. It effectively suppresses mode mixing and background noise, and is suitable for nonlinear and non-stationary acoustic emission signals of cranes.

[0013] Multi-channel information utilization fully breaks through the limitations of traditional single-channel analysis, integrates acoustic emission data from multiple sensors, comprehensively explores the damage characteristics of key components, and improves the ability to identify early and subtle faults.

[0014] High fault identification accuracy is achieved by using a deep convolutional neural network, which overcomes the feature ambiguity and information loss problems caused by traditional DBN, RNN, and GAN pooling methods. It autonomously learns deep fault features, resulting in higher diagnostic accuracy under complex working conditions.

[0015] The differential search algorithm has good optimization efficiency and generalization, fewer parameters, and strong global search capability, which is superior to traditional intelligent optimization algorithms. The whole method is highly automated, requires no manual intervention, and has strong engineering applicability.

[0016] Its outstanding value in safe operation and maintenance is that it can provide accurate early warning of critical faults in crane components, avoid sudden downtime and safety accidents, reduce blind maintenance, lower operation and maintenance costs, and extend equipment life. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 A flowchart of a multi-channel fault diagnosis method based on DS-VMD deep convolutional neural network provided in this application; Figure 2 A schematic diagram of the structure of the deep convolutional neural network algorithm proposed in this application; Figure 3 The algorithm provided in this application includes a small-scale pooling process; Figure 4 The overall structure of the fault diagnosis method proposed in this application; Figure 5 This diagram illustrates the setup for acquiring acoustic emission data from cranes in practical engineering applications; among which... Figure 5 (a) is an MQ1260-45 gantry crane; Figure 5 (b) is a rotating column; Figure 5 (c) is a type A frame; Figure 5 (d) represents the crane boom; Figure 5 (e) represents the gantry leg; Figure 5 The middle (f) represents the tower body; Figure 5 (g) represents the rotating platform beam; Figure 5 h represents the counterweight; Figure 5 (i) represents the SAMOS acoustic emission acquisition system; Figure 5 (j) represents 40 acoustic emission acquisition channels; Figure 6 The time-domain waveforms of the original acoustic emission signals of the crane under different operating conditions with a load of 0t are provided for this application. Figure 7 Time-domain waveforms of multi-channel acoustic emission signals under different operating conditions provided in this application; Figure 8 The analysis results of the multi-channel acoustic emission signals provided for this application are as follows: Figure 8 (a) is the Fourier spectrum. Figure 8 (b) is the envelope spectrum; Figure 9 The impact of different numbers of IMFs provided for this application on the fault classification accuracy of the proposed method; Figure 10 The confusion matrix obtained by processing the acoustic emission dataset of cranes under different load conditions provided in this application; Figure 11 Box plots of acoustic emission data from cranes obtained by three algorithms under different load conditions in 20 tests provided in this application; Figure 12 Two-dimensional visualization results of the learned fault features provided in this application, compared with those of three standard deep learning algorithms: Figure 12 In the middle (a), a Deep Belief Network (DBN) is shown. Figure 12 (b) is a convolutional neural network; Figure 12 (c) represents a recurrent neural network (RNN); Figure 12 In the middle (d), the designed DS-VMD deep convolutional neural network is shown. Figure 13 This application provides a comparative analysis of diagnostic results obtained through different methods. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0021] This application can be used in a wide variety of general-purpose or special-purpose computing environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor devices, distributed computing environments including any of the above devices, etc.

[0022] Example 1 Reference Figure 1 As shown, this application discloses a multi-channel fault diagnosis method based on DS-VMD deep convolutional neural network, including the following steps: S1. Collect multi-channel acoustic emission (AE) monitoring signals from the crane and divide them into signal sub-sequence sample sets according to industrial frequencies; S2. The differential search DS algorithm is used to adaptively optimize the number of modes K and the penalty factor α of variational mode decomposition (VMD), and a fitness function based on envelope entropy is constructed. The signal subsequence sample set is then decomposed by DS-VMD to obtain several IMF components. For example, the range of parameters (K, a) in VMD is initialized and set, where the number of modes k=1 and the balancing parameter a=400. The fitness function of DS is determined. Shannon information entropy is used as the standard for evaluating signal sparsity, reflecting the uncertainty of the signal. Therefore, the envelope sequence after signal decomposition is selected. The entropy is used as the fitness value. For a given original acoustic emission signal s(j), its envelope entropy is... Calculate using the following formula.

[0023] ; in, and a .

[0024] Furthermore, a(j) is the envelope spectrum of s(j) obtained through the Hilbert transform. This is the normalized form of a(j). A smaller envelope entropy indicates a sparser signal, thus containing more periodic components. Therefore, the maximum envelope entropy of the decomposed intrinsic mode functions (IMFs) is minimized to ensure good sparsity of all components; that is, the fitness function of the DS algorithm can be expressed as: ; in It is the envelope entropy of the k-th IMF.

[0025] S3. Calculate the kurtosis of each IMF component, set the kurtosis weight of the IMF component according to the kurtosis value, reconstruct the time domain signal based on the kurtosis weight of the IMF component, and obtain the multi-channel fault characteristics. The optimal parameter combination is obtained using the DS algorithm, and then the time-domain acoustic emission signal is processed by VMD.

[0026] Based on the characteristic frequencies of crane faults, the kurtosis of each IMF is calculated. Then, according to the magnitude of the corresponding kurtosis value, each IMF component is assigned a corresponding weight, and the time-domain acoustic emission signal is reconstructed, thereby extracting the crane fault characteristics more stably and accurately. Assuming the number of VMD modes determined by DS is K, and the signal of each layer is... Furthermore, the correlation kurtosis for each mode is... The corresponding reconstruction coefficient for each IMF is Therefore, the reconstructed signal x new It can be defined as: ; For x new Envelope spectrum analysis was performed to extract the corresponding fault characteristic frequencies.

[0027] S4. Construct an improved deep convolutional neural network; S5. Input the multi-channel fault features extracted by DS-VMD into the network for training and testing, update the network weights using backpropagation, automatically identify different fault modes of the crane, and complete intelligent diagnosis of multi-channel faults.

[0028] Optionally, the step of performing DS-VMD decomposition on the multi-channel acoustic emission signal to obtain several IMF components includes: The differential search (DS) algorithm is used to simulate the seasonal migration behavior of the population. With the goal of minimizing the envelope entropy of the decomposed IMF components, the optimal combination of VMD parameters is adaptively searched to obtain several IMF components.

[0029] Optionally, the optimization objective of minimizing the envelope entropy of the decomposed IMF components includes: Select the envelope sequence after signal decomposition The entropy is used as the fitness value; For a given original acoustic emission signal s(j), its envelope entropy Calculated using the following formula: ; in, ;a ; a(j) is the envelope spectrum of s(j) obtained through the Hilbert transform. It is the normalized form of a(j); Where j is the total number of sampling points, and j is the index of the sampling point; The fitness function of the DS algorithm is: ; in, It is the envelope entropy of the k-th IMF.

[0030] Optionally, the improved deep convolutional neural network includes: Small-scale convolutional pooling layers are used to replace traditional average pooling and max pooling layers, combined with 3×3 small-scale convolutional kernels, depthwise separable convolution, and global average pooling.

[0031] In this embodiment, the overall architecture of the proposed DS-VMD deep convolutional neural network is as follows: The small-scale convolutional kernels it contains downsample fault features using a skip-sampling method. A schematic diagram of the designed DS-VMD deep convolutional neural network algorithm is shown below. Figure 2 As shown. Specifically, the construction process of the DS-VMD deep convolutional neural network framework is as follows: A) Training phase of DS-VMD deep convolutional neural network: To improve generalization ability and prevent overfitting, the small-scale convolutional layers in the proposed method can be defined by the following equation: ; in, and These represent input and output, respectively. Furthermore, represents the convolution kernel, and b represents the scalar bias. It is a non-linear activation function, and * denotes the convolution symbol. Specifically, the designed small-scale convolutional layers can effectively reduce the number of feature parameters, thereby preventing overfitting of the proposed method.

[0032] Furthermore, the proposed DS-VMD deep convolutional neural network is trained during error backpropagation. The gradient of the entire network is calculated using the chain rule based on the gradients of the convolutional and pooling layers. The partial derivative of its loss function with respect to the l-th convolutional kernel and bias can be calculated using the following formula: ; Where L represents the loss function, Represents the convolution kernel. It is the p-th network activity value in the l-th layer that has not undergone a non-linear activation function. Furthermore, This represents the input of the (l-1)th layer. The loss function is relative to The partial derivatives, It is the p-th bias of the l-th layer.

[0033] Therefore, since pooling layers have no weights, only the derivative of the loss function with respect to the input neurons needs to be calculated. During small-scale pooling, backpropagation only passes gradients to the recorded neurons, while other neurons do not participate in gradient propagation. Its mathematical representation is as follows: ; in, Represents neurons, This indicates the location of the neuron's maximum activity during forward propagation.

[0034] B) Small-scale convolutional pooling: To address the shortcomings of commonly used pooling layers, the DS-VMD deep convolutional neural network proposed in this application includes a small-scale convolutional layer with a stride of 2 and an activation function of a rectified linear unit, used to replace the max-pooling layer for signal pooling. The input-output relationship of the signal through the convolutional layer can be calculated using the following formula: ; Where W and N represent the input and output of the signal, respectively, F and S are the scale of the convolution kernel and the stride, respectively, and P represents the amount of padding.

[0035] When the stride of the convolutional kernel is set to 2 to replace zero padding, the output size of the signal will be half the input size, thus achieving signal downsampling similar to that of a pooling layer. Furthermore, the proposed small-scale convolutional pooling method extracts effective feature information by adjusting the weights of the convolutional kernel itself, fully utilizing the activity of all neurons within the region, thereby compensating for the shortcomings of max pooling and average pooling. Figure 3 A schematic diagram of the proposed small-scale convolutional pooling is shown.

[0036] Specifically, the proposed small-scale convolutional pooling method includes an activation function, which increases the nonlinearity of the entire network and improves the feature learning and representation capabilities of traditional deep convolutional neural networks. However, the convolution operation provides the network with more parameters that need to be learned. Therefore, this patent proposes three strategies to overcome the problem of a sharp increase in parameters.

[0037] First, a small-scale 3×3 convolution kernel was chosen, which introduces relatively fewer parameters; Secondly, depthwise separable convolution kernels are used in the pooling convolutional layers, which can perform spatial convolution on each channel of the input individually; Finally, a global average pooling layer is used to replace the tiling layer and the first fully connected layer.

[0038] Example 2 To evaluate the feasibility of the proposed method, acoustic emission signals from cranes under different operating conditions were collected in the experiment. The results were then compared with existing advanced fault diagnosis methods, verifying the superiority of the proposed method.

[0039] The overall structure of the fault diagnosis method provided in this application is as follows: Figure 4 As shown; the overall fault diagnosis steps are described below: 1) Divide the original acoustic emission signal into subsequences using N times the industrial frequency, and form a subsequence sample dataset.

[0040] 2) The relevant parameters of VMD are adaptively optimized through the DS algorithm, and the above subsequence sample dataset is processed to generate a series of IMF components.

[0041] 3) Construct the DS-VMD deep convolutional neural network architecture and set the network hyperparameters (such as the number of network layers and convolution kernels per layer, non-linear activation functions and loss functions, small-scale convolution pooling methods, etc.), and then randomly initialize the network parameters.

[0042] 4) Set the number of training iterations and the number of iterations to compile the network model, and then perform the forward computation of the model.

[0043] 5) Calculate the cross-entropy loss function based on network results and data labels.

[0044] 6) Update the weights of the proposed DS-VMD deep convolutional neural network architecture using the backpropagation algorithm, output the fault diagnosis results, and evaluate the generalization performance of the proposed method.

[0045] Specifically, the feasibility assessment implementation methods include: 1) Data Acquisition and Processing To verify the effectiveness of the proposed method, acoustic emission datasets were acquired using an MQ1260-45 gantry crane under four load conditions (0 tons, 45 tons, 56.25 tons, and 60 tons). An overview of the experimental setup is as follows: Figure 5 As shown. Among them, Figure 5 This diagram illustrates the setup for acquiring acoustic emission data from cranes in practical engineering applications; among which... Figure 5 (a) is an MQ1260-45 gantry crane; Figure 5 (b) is a rotating column; Figure 5 (c) is a type A frame; Figure 5 (d) represents the crane boom; Figure 5 (e) represents the gantry leg; Figure 5 The middle (f) represents the tower body; Figure 5 (g) represents the rotating platform beam; Figure 5 h represents the counterweight; Figure 5 (i) represents the SAMOS acoustic emission acquisition system; Figure 5 (j) represents 40 acoustic emission acquisition channels.

[0046] Specifically, under each load condition, eight fault modes can be acquired in three operating states (i.e., hoisting, luffing, and rotation), defined as rotating column (RC), A-frame (AF), boom (CB), gantry leg (GL), tower (T), rotating platform beam (RP), counterweight (BW), and normal state (N). Furthermore, the acoustic emission data for each fault mode contains 160,000 data points. In particular, four datasets (sets A / B / C / D) are used to test the diagnostic performance of the proposed algorithm. Detailed descriptions of the four datasets are shown in Table 1, where each data sample contains 1024 data points. In addition, 20 sets of samples were collected under each load condition in the experiment, with 10 data samples randomly selected as the training dataset and the remainder as the test dataset. That is, the ratio of training samples to test samples is 1:1. The specific experimental parameters of the acoustic emission acquisition system are listed in Table 2.

[0047] Table 1

[0048] Table 2

[0049] 2) Fault diagnosis result analysis Based on the fault diagnosis flowchart of the proposed method, DS-VMD is first used to decompose the original acoustic emission signals of all samples to extract the fault characteristics of the crane. Specifically, for intuitive analysis, a sample under zero-load conditions is selected as an example. Figure 6 The time-domain acoustic emission signal waveforms of seven channels under different fault modes are shown under zero-ton conditions. Figure 6 It can be seen that it is difficult to directly identify the characteristics of acoustic emission signals by observing time-domain waveforms and Fourier spectra, because the acoustic emission signals of cranes are usually nonlinear and non-stationary, and their fault characteristics are easily submerged by strong background noise. Therefore, it is necessary to adopt an intelligent feature extraction method that does not require prior knowledge to analyze the acoustic emission data of cranes, reducing the dependence on human factors.

[0050] Figure 7 This paper presents the decomposition results of multi-channel acoustic emission data under different operating conditions using DS-VMD. From... Figure 7 It can be seen that the proposed DS-VMD method can effectively avoid endpoint effects and remove environmental noise, thereby improving the decomposition performance of multi-channel signals. To further detect fault modes under different operating conditions, Figure 8 Fourier spectra and envelope spectra of different fault modes are shown; among them, Figure 8 (a) is the Fourier spectrum. Figure 8 (b) is the envelope spectrum; from Figure 8It can be seen that the fault frequency and its harmonics are very obvious, indicating that the proposed method can effectively extract fault information. Secondly, the DS-VMD deep convolutional neural network designed in this application is trained using training data. Finally, an equal amount of test data as the training set is input into the trained DS-VMD deep convolutional neural network to achieve automatic fault mode identification. The main parameter settings of the proposed DS-VMD deep convolutional neural network algorithm are shown in Table 3.

[0051] Table 3

[0052] To investigate the fault identification performance of the proposed DS-VMD deep convolutional neural network algorithm as the modulus increases, different numbers of IMF components (i.e., 3, 5, 7, 9, 11, 13, and 15) were used as inputs to study the training process of the DS-VMD deep convolutional neural network and its diagnostic accuracy on the test dataset. The change in classification accuracy with the number of training iterations is shown below. Figure 9 As shown. From Figure 9 It can be seen that when the original sequences are used directly to train the proposed network, the convergence speed of the accuracy is slow and unstable, that is, the accuracy fluctuates greatly with the number of training iterations. Detailed comparison results of the eight sequences are listed in Table 4. As shown in Table 4, when the number of IMFs is 9, the proposed network has good generalization performance. Therefore, in this study, an IMF component of 9 is selected as the input to the designed DS-VMD deep convolutional neural network algorithm.

[0053] Table 4

[0054] at the same time, Figure 10 The confusion matrix of the proposed method under different load conditions in the first experiment is shown. The confusion matrix displays the identification results for all failure modes, with the x-axis and y-axis representing abbreviations for different failure modes. Furthermore, the diagonal elements of the matrix represent the number of samples correctly classified for a particular failure mode, while the remaining off-diagonal elements represent the number of misclassified samples, where one failure mode is predicted as another. Figure 10As shown, under zero load conditions, all failure modes were correctly identified except for one GL sample which was misclassified as RP, indicating that the proposed method achieved a classification accuracy of 99.90% (999 / 1000) in the first test. Furthermore, for the other three loads (45 tons, 56.25 tons, and 60 tons), the classification accuracies in the first test were 98.90% (989 / 1000), 99.10% (991 / 1000), and 99.80% (998 / 1000), respectively. Although the classification accuracy decreased slightly with increasing load compared to the zero load condition, the proposed method achieved an accuracy exceeding 98% under all load conditions.

[0055] To verify the superiority of the designed DS-VMD deep convolutional neural network algorithm, it was first compared with two traditional DCNN algorithms (using max pooling and average pooling, referred to as DCNN1 and DCNN2 respectively). Detailed comparison results of the three algorithms (DCNN1, DCNN2, and the method proposed in this application) are shown in Table 5, including the average test accuracy and standard deviation during the testing process. As shown in Table 5, the classification accuracy of the proposed method under different load conditions is significantly higher than that of the traditional DCNN algorithms.

[0056] Table 5

[0057] In addition, to present the comparison results more intuitively, Figure 11 The box plots for three algorithms (DCNN1, DCNN2, and the method proposed in this application) under different load conditions are further described. The upper and lower black lines connecting the boxes represent the upper and lower limits of classification accuracy, respectively, and the black dots represent the average accuracy after 20 trials. Furthermore, red plus signs indicate outliers, the red line within the box represents the median of classification accuracy, and the upper and lower blue lines within the box represent the upper and lower quartiles of classification accuracy, respectively. Therefore, from... Figure 11 As can be seen, the accuracy of DCNN1 and DCNN2 fluctuates significantly and exhibits outliers (see red plus signs), while the accuracy of the proposed method is more stable, with most of the accuracy remaining between 99.88% and 99.99%. Overall, the comparison results preliminarily verify the effectiveness of the proposed method for crane fault identification under different load conditions.

[0058] To quantitatively verify the effectiveness of the designed DS-VMD deep convolutional neural network algorithm in fault feature extraction, this application introduces the inter-class distance D and intra-class distance Db indices of the output features of the last layer of the network, which are calculated by the following formulas.

[0059] ; ; Where Ci is the covariance matrix of sample i, n is the number of sample classes, tr is the trace of the solution matrix, and P(w i M represents the proportion of category i samples to the total number of samples. i and M o These represent the mean values ​​of the sample matrix for class i and the overall sample matrix, respectively. Table 6 lists the detailed results of the inter-class distance and intra-class distance calculated using different methods. As can be seen from Table 6, compared with DCNN1 and DCNN2, the designed network extracts the largest inter-class features and the smallest intra-class features through small-scale convolutional pooling, thus demonstrating the optimal performance of the proposed algorithm and its best fault classification ability.

[0060] Table 6

[0061] 3) Comparative analysis with existing advanced methods To verify the stability of the designed fault diagnosis algorithm, this application compared it with three standard deep learning algorithms (i.e., Deep Belief Network (DBN), Recurrent Neural Network (RNN), and Convolutional Neural Network (CNN)). The data processing and parameter settings for all comparison methods were the same as those for the proposed algorithm, using acoustic emission data from a crane under no-load conditions as an example. Specifically, the t-distributed random neighborhood embedding (t-SNE) technique was employed to reduce the high-dimensional output features of the last hidden layer extracted by the above four methods to a two-dimensional vector distribution, aiming to visualize the output features of the trained network. Figure 12 The results of feature visualization of the test samples are displayed; among them, Figure 12 In the middle (a), a Deep Belief Network (DBN) is shown. Figure 12 (b) is a convolutional neural network; Figure 12 (c) represents a recurrent neural network (RNN); Figure 12 In the middle (d), the designed DS-VMD deep convolutional neural network is shown. like Figure 12 As shown, only the low-dimensional features extracted by the proposed algorithm can completely distinguish all faults.

[0062] To further demonstrate the effectiveness of the proposed method, this application also compares it with existing advanced fault diagnosis methods (such as EMD+DBN, WT+DCNN, LMD+CNN, VMD+DBN, VMD+CNN, VMD+DCNN). Figure 13 The classification accuracy is shown for five trials using the method described above. Figure 13As shown, the proposed method achieves a higher classification accuracy than all the comparative methods. Furthermore, Table 7 lists the detailed comparison results, including the average test accuracy and standard deviation across five trials. It is evident from Table 7 that the standard deviation of the proposed method is smaller than that of the other comparative methods, indicating that the proposed method exhibits better stability in diagnosing crane faults.

[0063] Table 7

[0064] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0065] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in connection with the embodiments disclosed in this application can be implemented in electronic hardware, computer software, or a combination of both.

[0066] To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality above. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0067] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-channel fault diagnosis method based on DS-VMD deep convolutional neural network, characterized in that, Includes the following steps: S1. Collect multi-channel acoustic emission (AE) monitoring signals from the crane and divide them into signal sub-sequence sample sets according to industrial frequencies; S2. The differential search DS algorithm is used to adaptively optimize the number of modes K and the penalty factor α of variational mode decomposition (VMD), and a fitness function based on envelope entropy is constructed. The signal subsequence sample set is then decomposed by DS-VMD to obtain several IMF components. S3. Calculate the kurtosis of each IMF component, set the kurtosis weight of the IMF component according to the kurtosis value, reconstruct the time domain signal based on the kurtosis weight of the IMF component, and obtain the multi-channel fault characteristics. S4. Construct an improved deep convolutional neural network; S5. Input the multi-channel fault features extracted by DS-VMD into the network for training and testing, update the network weights using backpropagation, automatically identify different fault modes of the crane, and generate intelligent multi-channel fault diagnosis.

2. The multi-channel fault diagnosis method based on DS-VMD deep convolutional neural network according to claim 1, characterized in that, The DS-VMD decomposition of the multi-channel acoustic emission signal to obtain several IMF components includes: The differential search (DS) algorithm is used to simulate the seasonal migration behavior of the population. With the goal of minimizing the envelope entropy of the decomposed IMF components, the optimal combination of VMD parameters is adaptively searched to obtain several IMF components.

3. The multi-channel fault diagnosis method based on DS-VMD deep convolutional neural network according to claim 2, characterized in that, The optimization objective of minimizing the envelope entropy of the decomposed IMF components includes: Select the envelope sequence after signal decomposition The entropy is used as the fitness value; For a given original acoustic emission signal s(j), its envelope entropy Calculated using the following formula: ； in, ;a ; a(j) is the envelope spectrum of s(j) obtained through the Hilbert transform. It is the normalized form of a(j); Where j is the total number of sampling points, and j is the index of the sampling point; The fitness function of the DS algorithm is: ； in, It is the envelope entropy of the k-th IMF.

4. The multi-channel fault diagnosis method based on DS-VMD deep convolutional neural network according to claim 1, characterized in that, The improved deep convolutional neural network includes: Small-scale convolutional pooling layers are used to replace traditional average pooling and max pooling layers, combined with 3×3 small-scale convolutional kernels, depthwise separable convolution, and global average pooling.

5. The multi-channel fault diagnosis method based on DS-VMD deep convolutional neural network according to claim 3, characterized in that, The small-scale convolutional pooling uses a small-scale convolutional layer with a stride of 2 and an activation function of ReLU for downsampling, and retains the feature information of all neurons in the region through adaptive weights of the convolutional kernel. Depth-separable convolution and channel-wise spatial convolution are used to replace tiling and fully connected layers with global average pooling.

Images