Adaptive snow goose algorithm driven residual network bearing fault diagnosis method and system

Abstract

The application discloses a bearing fault diagnosis method based on a residual network driven by an adaptive snow goose algorithm, and comprises the following steps: acquiring bearing fault data to be diagnosed; inputting the bearing fault data to be diagnosed into a trained fault diagnosis model to output a fault diagnosis result; wherein the training of the fault diagnosis model comprises the following steps: constructing a bearing fault sample data set; constructing a fault diagnosis model, which comprises a frequency domain attention module and a flexible residual network, and performing deep feature extraction on enhanced features through a double flexible residual block, the double flexible residual block is adjustable in the number of convolution layers and comprises two serial flexible residual paths, and a dynamic selection mechanism of a skip connection position and a cross-path cross-connection weight and global average pooling attention mechanism are introduced; introducing an adaptive snow goose algorithm to optimize network structure parameters of the fault diagnosis model, and based on the optimal network structure parameters, performing model training on the constructed sample data set to output a trained fault diagnosis model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent diagnosis of rotating equipment, and particularly relates to a bearing fault diagnosis method based on a multi-stage self-adaptive snow goose algorithm driven frequency domain enhanced residual network. BACKGROUND

[0002] As a core component in rotating machinery equipment, the operation reliability of bearings is crucial to the safe and stable operation of the entire mechanical system. In the industrial production process, bearings are often subjected to high-speed rotation and heavy load, which can easily cause faults such as inner ring cracks, outer ring wear, and rolling element peeling. These faults not only reduce the operating efficiency of the equipment, but also can cause equipment damage and safety accidents. Therefore, an efficient intelligent diagnosis model is urgently needed to provide technical support for the safe operation of the system.

[0003] Currently, deep learning-based fault diagnosis methods are widely used in the field of rotating machinery such as bearings. Convolutional neural networks (CNN) have achieved remarkable results in vibration signal analysis due to their excellent local feature extraction capabilities. Using CNN for bearing fault diagnosis can achieve good diagnostic accuracy, but the computational complexity of CNN is high, making it difficult to capture complex patterns in a larger range. Residual networks (ResNet) effectively solve the gradient vanishing problem of deep networks through a jump connection mechanism, and exhibit excellent performance in the field of fault diagnosis. However, existing researches mostly use fixed network structures, which are difficult to adaptively adjust according to different data characteristics and diagnosis tasks, limiting the generalization ability and diagnostic accuracy of the model.

[0004] In addition, traditional methods mainly focus on time domain features and do not fully utilize frequency domain information, which has limitations in dealing with complex spectral features. To overcome the limitations of fixed network structures, neural architecture search (NAS) technology has gradually become a research hotspot. Researchers have begun to apply the concept of automated network design to the field of fault diagnosis. However, traditional NAS methods have high computational costs, and existing methods mostly focus on simple search of structural parameters, lacking deep optimization of the internal mechanisms of the network, and unable to achieve a synergistic improvement of structural flexibility and feature expression ability.

[0005] With the rapid development of attention mechanisms, frequency domain attention technology has begun to attract widespread attention in the field of signal processing. Frequency domain analysis can reveal the spectral information contained in vibration signals and has unique sensitivity to different fault patterns. However, how to effectively integrate frequency domain feature extraction and attention mechanisms to tap the discriminative potential of frequency domain features still needs further research.

[0006] Snow goose algorithm (SGA) as a new swarm intelligence algorithm, mainly through the simulation of snow goose migration and foraging behavior to achieve global optimization. However, the traditional SGA is easy to fall into local optimum, slow convergence speed and other problems, it is difficult to meet the needs of complex network structure optimization. Therefore, how to enhance the performance of SGA in neural network optimization, still need in-depth study. SUMMARY

[0007] In order to solve the problems of intelligent selection of deep neural network topology configuration parameters and insufficient extraction of frequency domain key features of vibration signals in rotating machinery bearing fault diagnosis, the application provides a residual network bearing fault diagnosis method driven by an adaptive snow goose algorithm, which realizes high accuracy recognition and intelligent discrimination of bearing degradation types through deep coupling of adaptive snow goose algorithm (ISGA) and dual flexible residual network (DFResNet-FDAM) enhanced by frequency domain attention and collaborative use of time and frequency domain information, effectively solving the problems of lack of flexibility of network architecture and insufficient extraction of frequency domain information.

[0008] According to an aspect of the present application, a residual network bearing fault diagnosis method driven by an adaptive snow goose algorithm is provided, comprising:

[0009] Obtaining bearing fault data to be diagnosed;

[0010] Inputting the bearing fault data to be diagnosed into the trained fault diagnosis model to output the fault diagnosis result; wherein the training of the fault diagnosis model comprises:

[0011] Constructing a bearing fault sample data set;

[0012] Constructing a fault diagnosis model, including: a frequency domain attention module for performing frequency domain perception attention enhanced features on the input bearing fault sample; a flexible residual network for performing deep feature extraction on the enhanced features through a double flexible residual block, the double flexible residual block contains two series of flexible residual paths and the convolution layer number of each flexible residual block is adjustable, at the same time, the flexible residual network also introduces a dynamic selection mechanism of the jump connection position to determine the access position and output position of the flexible residual path, and introduces a cross-path cross-connection weight and a global average pooling attention mechanism to realize information intercommunication and feature enhancement between different residual paths;

[0013] Introducing an adaptive snow goose algorithm to optimize the network structure parameters of the fault diagnosis model, and based on the optimal network structure parameters, training the model on the constructed sample data set to output the trained fault diagnosis model.

[0014] As a further technical solution, the construction of the adaptive snow goose algorithm comprises:

[0015] Initialize the population of Snow Goose algorithm, and the encoding of each individual contains network structure parameters;

[0016] Establish a multi-stage evolution criterion function to dynamically identify the evolution stage of the algorithm according to the population dispersion index and fitness increment;

[0017] Introduce the opposite solution generation strategy to generate the opposite solution of the current candidate solution and guide it with the elite individuals, thereby widening the search range of the solution space;

[0018] Design an elite solution memory bank management mechanism to store high-quality solutions in historical iterations and guide population evolution through a weighted selection strategy;

[0019] Construct an inter-individual cooperative foraging mechanism to calculate the cooperation strength index between individuals and incorporate it into the position update equation to enhance the collective search ability of the population;

[0020] Adaptively switch the V-shaped formation flight strategy in the exploration stage and the linear flight strategy in the development stage according to the multi-stage evolution identification results.

[0021] As a further technical solution, an adaptive Snow Goose algorithm is introduced to optimize the network structure parameters, including:

[0022] Set the adaptive Snow Goose algorithm execution parameters, including population size, maximum evolution generation, elite memory bank capacity, and opposite learning probability;

[0023] For each individual in the current population, construct the network topology of the corresponding fault diagnosis model;

[0024] Iteratively train the constructed network topology on the training set samples;

[0025] Evaluate the classification performance of the network topology on the validation set samples, and use the identification accuracy as the fitness value of the individual;

[0026] Perform population evolution operations of the adaptive Snow Goose algorithm, and apply multi-stage adaptation, opposite solution generation, elite memory guidance, and cooperative foraging strategies;

[0027] When the maximum evolution generation or fitness convergence criteria are met, output the parameter configuration scheme of the global optimal individual.

[0028] As a further technical solution, based on the optimal network structure parameters, model training is performed on the constructed sample data set, including:

[0029] Construct the network topology of the fault diagnosis model according to the optimal network structure parameters obtained by the adaptive Snow Goose algorithm;

[0030] Set the hyperparameter configuration for network training;

[0031] Iterative training is performed on the optimal network structure parameters on the training sample set, and the performance of the trained network model is evaluated on the test sample set;

[0032] An output fault recognition result is output, including overall accuracy, macro-averaged precision, macro-averaged recall, and macro-averaged F1 value.

[0033] As a further technical solution, before iterative optimization, the following network structure parameter search space is defined:

[0034] The search interval of the number of convolution layers of the first flexible residual block and the second flexible residual block is defined.

[0035] The search interval of the number of skip connections of the first flexible residual block and the second flexible residual block is determined.

[0036] A candidate value set of the number of convolution kernels is configured.

[0037] As a further technical solution, the construction of the frequency domain attention module further includes:

[0038] Performing fast Fourier transform on the feature tensor of the input bearing fault sample;

[0039] Performing amplitude and phase decoupling separation processing on the frequency domain complex result after fast Fourier transform, respectively obtaining a frequency spectrum amplitude matrix and a phase angle matrix;

[0040] Designing a frequency channel attention weighting mechanism to realize feature selection in the channel dimension by learning the importance coefficients of different frequency components;

[0041] Constructing a frequency domain spatial attention focusing mechanism to focus on the saliency regions of different frequency positions in the frequency spectrum amplitude matrix;

[0042] Performing frequency domain feature parameterization enhancement and phase compensation operations to optimize the frequency spectrum representation through a trainable adaptive filter;

[0043] Using inverse fast Fourier transform to reconstruct the enhanced frequency domain representation back to the time domain space, and integrating the original time domain features through a gating fusion mechanism to obtain the final enhanced features.

[0044] As a further technical solution, a bearing fault sample dataset is constructed, including:

[0045] Collecting bearing vibration data under multiple rotational speed conditions;

[0046] Using a sliding window segmentation to serialize and split the original vibration signal, setting the window width and sliding step, and generating a fixed-length sample sequence;

[0047] Performing Z-score normalization processing on the segmented sample sequence;

[0048] The normalized samples are subjected to dataset segmentation, and the proportion of the training set and the test set is set.

[0049] According to an aspect of the present application, a residual network bearing fault diagnosis system driven by an adaptive snow goose algorithm is provided for implementing the method, comprising:

[0050] A data acquisition module is configured to acquire bearing fault data to be diagnosed.

[0051] A fault diagnosis module is configured to input the bearing fault data to be diagnosed into a trained fault diagnosis model and output a fault diagnosis result.

[0052] A bearing fault sample dataset is constructed.

[0053] The fault diagnosis model is constructed, comprising a frequency domain attention module configured to perform frequency domain perception attention enhanced features on the input bearing fault samples, and a flexible residual network configured to perform deep feature extraction on the enhanced features through a double flexible residual block, wherein the double flexible residual block comprises two series-connected flexible residual paths and the number of convolution layers of each flexible residual block is adjustable.

[0054] An adaptive snow goose algorithm is introduced to optimize the network structure parameters of the fault diagnosis model, and based on the optimal network structure parameters, the model is trained on the constructed sample dataset to output a trained fault diagnosis model.

[0055] According to an aspect of the present application, a residual network bearing fault diagnosis device driven by an adaptive snow goose algorithm is provided, comprising a memory and a processor, wherein the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the residual network bearing fault diagnosis method driven by the adaptive snow goose algorithm.

[0056] According to an aspect of the present application, a non-transitory computer readable storage medium is provided, which stores computer instructions for executing the residual network bearing fault diagnosis method driven by the adaptive snow goose algorithm.

[0057] Compared with the prior art, the present application has the following advantages:

[0058] 1. Through the integration of ISGA and DFResNet-FDAM, a diagnostic system integrating neural architecture intelligent search and frequency domain feature deep extraction is constructed, effectively solving the problems of poor flexibility of fixed network topology and insufficient depth of frequency domain information utilization, significantly improving the bearing fault recognition accuracy and network architecture search efficiency;

[0059] 2. The introduction of multi-stage adaptive mechanism, opposite solution generation, elite memory guidance and cooperative foraging strategy improves the global exploration ability and local fine search ability of traditional SGA algorithm; the design of double-path flexible residual structure and frequency domain perception attention module enables the network to adaptively adjust the topology configuration according to the fault vibration characteristics and deeply excavate the frequency spectrum discriminant information, enhancing the structural flexibility and feature expression ability of the deep learning architecture;

[0060] 3. Multi-level collaborative design of convolution unit, residual path, frequency domain transformation and attention mechanism is realized, and the model performance is effectively improved through intelligent topology search strategy, providing an effective technical means for intelligent diagnosis and preventive maintenance of rotating machinery bearing faults under complex operating conditions. BRIEF DESCRIPTION OF DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, a brief introduction of the drawings used in the embodiments or prior art description will be given below. Obviously, the drawings described below are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor based on these drawings.

[0062] Figure 1 is a flowchart of an adaptive snow goose algorithm driven residual network bearing fault diagnosis method in an embodiment of the present application;

[0063] Figure 2 is a training flowchart of a fault diagnosis model in an embodiment of the present application;

[0064] Figure 3 is a basic structure diagram of ResNet in an embodiment of the present application;

[0065] Figure 4 is a basic principle diagram of DF-ResNet in an embodiment of the present application;

[0066] Figure 5 is a basic principle diagram of SGA algorithm in an embodiment of the present application;

[0067] Figure 6 is a diagnostic flowchart of ISGA-DFResNet-FDAM in an embodiment of the present application. DETAILED DESCRIPTION

[0068] In view of the problems existing in the prior art, a new deep learning network adaptive construction method is urgently needed, which can effectively cope with the challenge of extracting significant features in the frequency domain of complex vibration signals, realize the intelligent adaptation of neural network architecture parameters and fault vibration modes, and improve the classification accuracy and model robustness level of bearing fault diagnosis. Therefore, the present application provides an adaptive snow goose algorithm driven residual network bearing fault diagnosis method, which realizes high-accuracy identification and intelligent discrimination of bearing degradation types through deep coupling of ISGA and DFResNet-FDAM and collaborative use of time-frequency domain information, effectively solving the problems of lack of flexibility of network architecture and not deep enough extraction of frequency domain information.

[0069] The present application mainly improves the intelligent level of rotating machinery bearing fault diagnosis in complex spectrum environment based on improved swarm intelligence algorithm and network topology search strategy, and provides an adaptive snow goose algorithm driven residual network bearing fault diagnosis method. The present application fully utilizes the advantages of ISGA neural topology search and DFResNet-FDAM frequency domain feature deep extraction, realizes high-accuracy identification of multiple fault modes through integration of the two, designs multi-level feature extraction strategy for complex vibration signals, and realizes intelligent collaboration of convolution unit, residual path, frequency domain transformation and attention mechanism. The diagnosis effect obtained by the present application is more reliable and more accurate, which effectively improves the bearing fault recognition accuracy and the equipment health monitoring level under complex operating conditions.

[0070] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. In addition, the technical features in each embodiment or single embodiment provided by the present application can be combined with each other to form new technical schemes, and such combination is not restricted by the order of steps and / or structure composition mode, but must be based on the realization by those skilled in the art. When the combination of technical schemes appears contradictory or unachievable, it should be considered that such combination of technical schemes does not exist and is not within the protection scope required by the present application.

[0071] The technical scheme of the present application provides an adaptive snow goose algorithm driven residual network bearing fault diagnosis method, aiming to improve the adaptive configuration ability of deep network topology structure to differentiated fault modes and the deep extraction level of frequency domain discriminative features, realize intelligent exploration of residual network architecture parameters, and guarantee high-accuracy fault diagnosis and equipment health state evaluation of bearings under complex operating environment.

[0072] Referring to Figure 1 The adaptive snow leopard algorithm driven residual network bearing fault diagnosis method provided by the embodiment of the application first acquires bearing fault data to be diagnosed; then inputs the bearing fault data to be diagnosed into a trained fault diagnosis model to output a fault diagnosis result.

[0073] The training process of the fault diagnosis model is as shown in Figure 2 The training process of the fault diagnosis model is as shown in

[0074] Step S1, constructing a bearing fault sample data set;

[0075] Step S2, constructing a fault diagnosis model, including: a frequency domain attention module, used for frequency domain perception attention enhanced features of input bearing fault samples; a flexible residual network, used for deep feature extraction of enhanced features through double flexible residual blocks, the double flexible residual blocks contain two series of flexible residual paths and the convolution layer number of each flexible residual block is adjustable, at the same time, the flexible residual network also introduces a dynamic selection mechanism of a skip connection position to determine the access position and output position of the flexible residual path, and introduces a cross-path cross-connection weight and a global average pooling attention mechanism to realize information intercommunication and feature enhancement between different residual paths;

[0076] Step S3, introducing an adaptive snow leopard algorithm to optimize the network structure parameters of the fault diagnosis model, and based on the optimal network structure parameters, training the model on the constructed sample data set to output the trained fault diagnosis model.

[0077] Specifically, the bearing fault sample data set constructed in step S1 includes the following bearing vibration signal acquisition and data preprocessing:

[0078] Step 1.1: Collecting bearing vibration data under multiple rotating speed conditions, covering multiple operating states such as healthy state, inner ring damage, outer ring damage and rolling element wear;

[0079] Step 1.2: Using sliding window segmentation technology to sequence and cut the original vibration signal, setting the window width and sliding step distance to generate sample sequences with fixed length;

[0080] Step 1.3: Performing Z-score normalization processing on the cut sample sequence to eliminate the influence of numerical dimension and amplitude scale;

[0081] Step 1.4: Data set segmentation, setting the proportion of training set and test set to ensure balanced distribution and sufficient number of each type of fault sample.

[0082] The topological architecture design of the flexible residual network in step S2 includes the following steps:

[0083] Step 2.1: constructing a flexible residual base module with variable convolution layers, realizing the elastic configuration of the network longitudinal depth by dynamically adjusting the number of convolution layers;

[0084] Step 2.2: designing a dynamic selection mechanism for the position of the jump connection, determining the access position and output position of the residual path according to the total number of jump connection parameters;

[0085] Step 2.3: establishing a series topology structure of double flexible residual blocks, enhancing the nonlinear fitting ability of the network through the front and rear cascading of the two flexible residual channels;

[0086] Step 2.4: introducing cross-path cross-connection weights and global average pooling attention mechanisms to realize information exchange and feature enhancement between different residual paths.

[0087] The basic structure of ResNet is described in Figure 3 , and the residual network effectively solves the gradient vanishing problem in the training of deep neural networks by introducing a jump connection mechanism. Define x and y as the input and output of the residual block, W i is the weight matrix of the i-th layer, and F(-) is the residual function. The core idea of the standard residual block is to learn the residual mapping instead of the direct mapping, and its mathematical expression can be represented as follows:

[0088] .

[0089] Define ReLu(-) as the ReLU activation function, Bn(-) as the batch normalization operation, and * as the convolution operation. For the standard residual block, the residual function usually contains two convolution layers, as shown in the following formula:

[0090] .

[0091] The basic principle of DF-ResNet in the embodiment of the application is shown in Figure 4 . The standard ResNet adopts fixed convolution layer number and jump connection mode, which limits the adaptability of the network to different data characteristics, so the embodiment of the application proposes a flexible residual block, which realizes the adaptive adjustment of the structure by introducing variable parameters.

[0092] Define Dp(-) as the Dropout function, and N c is the number of convolution layers of the flexible residual block. At this time, the output h i of the i-th layer convolution can be represented as follows:

[0093] .

[0094] Define N sis the number of flexible residual blocks for the skip connection, t is the t-th skip connection, c is the number of skip connections, and C is the position of the skip connection t The number of flexible residual blocks can be dynamically determined by the following formula:

[0095] .

[0096] Fb1 and Fb2 are defined as the first and second flexible residual blocks. To further enhance the expression ability of the network, an embodiment of the present application designs a double flexible residual block, which includes two flexible residual paths in series. This process can be represented by the following formula:

[0097] .

[0098] W c is the cross-connection weight, G(-) is the global average pooling, A is the attention weight, and is an element-level multiplication, so the cross-connection and attention fusion mechanism can be represented as follows:

[0099] .

[0100] Further, the construction of the frequency domain attention module in step S2 includes the following steps:

[0101] Step S21: Perform a fast Fourier transform operation on the input feature tensor to achieve a mathematical transformation from the time domain space to the frequency domain space;

[0102] Step S22: Perform amplitude and phase decoupling separation processing on the frequency domain complex result to obtain a frequency spectrum amplitude matrix and a phase angle matrix, respectively;

[0103] Step S23: Design a frequency channel attention weighting mechanism to achieve feature selection in the channel dimension by learning the importance coefficients of different frequency components;

[0104] Step S24: Construct a frequency domain spatial attention focusing mechanism to focus on the salient regions of different frequency positions in the frequency spectrum matrix;

[0105] Step S25: Perform frequency domain feature parameterization enhancement and phase compensation operations to optimize the frequency spectrum representation through a trainable adaptive filter;

[0106] Step S26: Use the inverse fast Fourier transform to reconstruct the enhanced frequency domain representation back to the time domain space, and integrate the original time domain features through a gating fusion mechanism to form a feature that includes the original time domain features and the enhanced time domain features, and input the two residual channels of the flexible residual network.

[0107] The traditional CNN network mainly extracts features in the time domain, and the use of frequency domain information is not sufficient, therefore, an embodiment of the present application proposes a frequency domain attention module (FDAM), which mines frequency domain discriminant features through Fourier transform (FFT).

[0108] Define FFT(-) as a Fourier function, B, C and L as batch size, channel number and signal length. B×C×L , first perform FFT transformation and decompose into amplitude and phase, this process can be represented by the following formula:

[0109] .

[0110] Define w1 and w2 as channel attention weights. ch The process of learning the importance weight of different frequency components can be represented by the following formula:

[0111] .

[0112] Define W cn as a spatial attention convolution weight. s The process of paying attention to the importance of different frequency positions can be represented by the following formula:

[0113] .

[0114] Define F a as a learnable adaptive filtering parameter, W p as a phase correction weight. The frequency domain features after attention weighting need to be enhanced in amplitude and corrected in phase, and this process can be represented by the following formula:

[0115] .

[0116] Define IFFT(-) as an inverse Fourier operation, W g as a gating weight. Finally, the enhanced features are output through inverse FFT transformation and gating mechanism, and this process can be represented by the following formula:

[0117] .

[0118] In step S3, the adaptive snow goose algorithm is designed, including the following steps:

[0119] Step 3.1: initialize the snow goose algorithm population, the code of each individual contains network topology parameters such as the number of convolution layers, the number of convolution kernels and the number of jump connections;

[0120] Step 3.2: establish a multi-stage evolution discriminant criterion function, dynamically identify the evolution stage of the algorithm according to the population dispersion index and fitness increment;

[0121] Step 3.3: Introduce the opposite solution generation strategy, generate the opposite solution of the current candidate solution and combine the elite individual to guide, widen the search range of the solution space;

[0122] Step 3.4: Design an elite solution memory bank management mechanism to store high-quality solutions in historical iterations and guide population evolution through a weighted selection strategy;

[0123] Step 3.5: Build an inter-individual cooperative foraging mechanism to calculate the cooperation strength index between individuals and integrate it into the position update equation to enhance the collective search ability of the population;

[0124] Step 3.6: According to the multi-stage evolution discrimination result, adaptively switch the V-shaped formation flight strategy of the exploration stage and the linear flight strategy of the development stage.

[0125] Optionally, before iterative search using the adaptive snow goose algorithm, the following steps of network topology parameter search space definition are further included:

[0126] Define the search interval of the number of convolution layers of the first flexible and second flexible residual blocks;

[0127] Determine the search interval of the number of skip connections of the first flexible and second flexible residual blocks;

[0128] Configure a candidate value set of the number of convolution kernels.

[0129] The basic principle of the SGA algorithm in the embodiment of the application is as shown in the figure. Figure 5 Snow goose algorithm (SGA) is a new type of swarm intelligence optimization algorithm proposed by Tian et al. in 2024, which realizes global optimization by simulating the migration and foraging behavior of snow geese.

[0130] Define X i,j As the position of the i-th snow goose in the j-th dimension, l j And u j As the upper and lower bounds of the problem, rand is a random number between 0 and 1. In the snow goose population, each snow goose represents a candidate solution, and the population initialization process can be represented as:

[0131] .

[0132] Define N as the snow goose population size, and D as the dimension. Wherein the snow goose population matrix X and the fitness matrix F can be represented by the following formula:

[0133] ,

[0134] .

[0135] Define V tLet p be the velocity vector, p be the random perturbation term, and p be the inertia weight coefficient. T is the maximum number of iterations, and t is the current number of iterations.

[0136] During long-distance migrations, snow geese form unique V-shaped formations to reduce flight drag and energy consumption. This formation effectively utilizes the updrafts generated by the leading individual, thereby improving overall flight efficiency. Based on this phenomenon, the speed update mechanism of the snow goose algorithm can be expressed by the following formula:

[0137] ,

[0138] .

[0139] definition Let i be the position of the i-th snow goose in generation t. This is the current globally optimal position. For other randomly selected individual locations, B and D are the attraction and repulsion factors, respectively.

[0140] In actual flock migrations, snow geese develop different behavioral patterns based on their flight experience and physical condition. Dominant individuals typically occupy the front position of the formation, while other individuals adopt a follower strategy. The SGA algorithm defines the top 1 / 5 of the population as leaders, and their position update formula can be expressed as:

[0141] .

[0142] definition The location of the nearest outstanding individuals. Let E be the position of adjacent individuals and E be the coordination factor. For other following individuals in the formation, their position update strategy can be expressed as:

[0143] .

[0144] R is defined as a random number between 0 and 1, ⊕ is a vector addition operation, and B(D) is a random vector generated by d-dimensional Brownian motion, used to simulate small disturbances during flight.

[0145] After the snow geese complete their formation adjustments, the flock enters a relatively stable straight-line flight phase. During this phase, individuals primarily maintain consistency in flight direction through local information exchange. The SGA algorithm employs a probabilistic selection mechanism to simulate this behavioral characteristic, which can be represented by the following equation:

[0146] .

[0147] The standard SGA algorithm suffers from problems such as being prone to getting trapped in local optima and having slow convergence speed. To address these issues, this invention proposes the Adaptive Snow Goose Algorithm (ISGA), which integrates multiple enhancement strategies to improve optimization performance.

[0148] (1) Multi-stage adaptive mechanism

[0149] Let σ(t) denote the population diversity, θ e To explore the threshold, Δf(t) is the fitness improvement, and is the stagnation threshold. The SGA algorithm uses a fixed search strategy, which is difficult to adapt to the needs of different optimization stages. The present invention designs a multi-stage adaptive mechanism that can dynamically adjust the search behavior according to the population state, and the stage discrimination function can be defined as follows:

[0150] .

[0151] Let be the optimal fitness value of the tth generation, be a positive number to prevent zero values. At this time, the population diversity and fitness improvement can be represented by the following formula:

[0152] .

[0153] (2) Opposite learning strategy

[0154] Let P b be the basic opposite probability, and α be the decay parameter. Opposite learning enhances the search ability of the algorithm by considering both the solution and its opposite solution. This process can be represented by the following formula:

[0155] .

[0156] Let β be the elite guiding coefficient. At this time, the opposite solution generation formula can be represented by the following formula:

[0157] .

[0158] (3) Elite memory management

[0159] Let M be the elite library capacity, e i be the ith elite solution. To improve the optimization performance of the SGA algorithm, an elite memory library is designed to store excellent solutions and guide the search process. This process can be represented by the following formula:

[0160] .

[0161] Let η be the learning rate, e s be the elite solution selected according to the fitness and diversity weighting. At this time, the position update of the elite guide can be represented by the following formula:

[0162] .

[0163] (4) Cooperative foraging mechanism

[0164] Definition C i,j is the cooperation strength between individual i and individual j, σ c is the spatial distance parameter, f a is the average fitness. To strengthen the coordination between populations, a sub-population cooperation mechanism is designed to enhance information exchange, which can be represented by the following formula:

[0165] .

[0166] Definition w v is the inertia weight, c1 and c2 are acceleration coefficients, and r1 is a random number. At this time, the velocity update formula of the cooperation guide can be expressed as follows:

[0167] .

[0168] To verify the optimization performance of the ISGA algorithm, the Shekel benchmark test function is selected for experiment, the test iteration number is set to 1000, and the ISGA is compared with the standard SGA, the DBO and the GWO. The mathematical expression of the Shekel test function is shown in the following formula:

[0169] .

[0170] Further, after designing the adaptive snow goose algorithm, the execution steps of the network topology search driven by the adaptive snow goose algorithm ISGA are as follows:

[0171] Step S31: Set the ISGA algorithm execution parameters, including the population size, the maximum evolution generation, the elite memory bank capacity, and the opposite learning probability;

[0172] Step S32: For each individual in the current population, construct the corresponding DFResNet-FDAM network topology;

[0173] Step S33: Perform fast iterative training on the constructed network topology on the samples of the training set;

[0174] Step S34: Evaluate the classification performance of the network topology on the validation set samples to identify the accuracy rate as the fitness value of the individual;

[0175] Step S35: Perform the population evolution operation of the ISGA algorithm, and apply the multi-stage adaptation, opposite solution generation, elite memory guidance, and cooperation foraging strategies;

[0176] Step S36: Determine whether the maximum evolution generation or the fitness convergence criterion is met. If not, return to step 6.2 to continue the search iteration;

[0177] Step S37: Output the parameter configuration scheme of the global optimal individual.

[0178] Further, after obtaining the optimal network structure parameters, the following steps of optimal network topology training and fault identification are performed:

[0179] Step S38: According to the optimal topology parameter scheme obtained by ISGA search, the final DFResNet-FDAM diagnostic network is constructed;

[0180] Step S39: Set the hyperparameter configuration of network training, including initial learning rate, small batch sample number, maximum iteration number, etc.

[0181] Step S310: Fully iterate and train the optimal network on the complete training sample set;

[0182] Step S311: Evaluate the performance of the trained network model on the independent test sample set;

[0183] Step S312: Output the fault identification result, including overall accuracy, macro-precision, macro-recall and macro-F1 value.

[0184] The ISGA-DFResNet-FDAM diagnostic flowchart in the embodiment of the application is shown in Figure 6

[0185] The ISGA-DFResNet-FDAM integrated framework combines the adaptive snow goose algorithm, double flexible residual network and frequency domain attention module organically, realizes end-to-end optimization of bearing fault diagnosis. The number of convolutional layers, the number of convolutional kernels and the number of jump connections and other structure parameters directly affect the feature extraction ability and diagnostic performance of the network, therefore, the ISGA algorithm is used to optimize these key parameters.

[0186] The specific implementation steps are as follows:

[0187] Step 1: Collect bearing fault data, standardize the original signal and perform sliding window sampling, divide the training set and test set according to the 7:3 ratio, and ensure the balanced distribution of each fault class sample.

[0188] Step 2: Set the population size, the maximum iteration number is 50, the initial elite memory bank capacity is 20, the opposition learning probability is 0.3, and the cooperative foraging weight is 0.1.

[0189] ​Step3: Optimize the cycle execution, construct the DF-ResNet+FDAM network according to the current population individual code, use the training data for fast training and evaluate the fitness on the validation set; secondly, use the multi-stage adaptive mechanism to update the population position, and improve the SGA algorithm by applying the opposite learning, elite memory and cooperative foraging strategy; finally, use the improved SGA algorithm to optimize the network hyperparameters, so as to obtain the current optimal solution.

[0190] Step4: Termination condition judgment, check whether the maximum iteration number or fitness convergence is reached, if not, return to Step3 to continue optimization; if yes, terminate the search and output the optimal network structure parameters.

[0191] Step5: Use the optimal structure parameters obtained by ISGA search to construct the final DF-ResNet-FDAM network, and fully train on the complete training set.

[0192] Step6: Evaluate the fault diagnosis performance of the optimal model on the test set, and output the final result.

[0193] Optionally, based on the trained fault diagnosis model, the following steps of multi-data source verification and robustness evaluation are also performed:

[0194] Step S313: Perform the first group of verification experiments on the bearing fault data source of Jiangnan University, covering three rotating speed conditions of 600rpm, 800rpm and 1000rpm;

[0195] Step S314: Perform cross-dataset generalization capability evaluation on the CWRU bearing fault data source of Case Western Reserve University, including multiple working conditions of different damage sizes;

[0196] Step S315: Design module ablation comparison experiment, respectively remove the FDAM module, flexible structure and ISGA improvement strategy, verify the performance contribution of each component unit;

[0197] Step S316: Perform model robustness evaluation, and examine the disturbance resistance of the model by changing the data structure;

[0198] Step S317: Set the benchmark comparison experiment, select ResNet, WDCNN, Transformer and other current mainstream deep learning architectures for performance comparison.

[0199] The embodiment of the application adopts the bearing fault dataset of Jiangnan University for verification analysis. The dataset is manufactured by artificial cutting method, and the bearing type includes four types of outer ring fault, inner ring fault, rolling element fault and normal.

[0200] Define z i is the original data, μi and delta i is the mean and standard deviation of the original data.

[0201] Secondly, the data is processed, the sliding window is used for sample construction, the training set and the test set are divided according to the 7:3 ratio, and the Z-score standardization method is used for data preprocessing, and the process can be represented by the following formula:

[0202] .

[0203] Finally, the CWRU bearing fault data set is used for cross-data set verification experiment.

[0204] The ISGA-DFResNet-FDAM integrated architecture realizes intelligent diagnosis of bearing faults and automatic optimization of network topology, considers global search performance and deep feature learning ability, and provides an effective technical scheme for rotating machinery fault diagnosis in a complex spectrum environment.

[0205] The adaptive snow goose algorithm driven residual network bearing fault diagnosis method can be completed in Python 3.13.

[0206] The implementation basis of each embodiment of the present application is realized by programmed processing of a device with processor function. Therefore, in engineering practice, the technical scheme and function of each embodiment of the present application are packaged into various modules. Based on this actual situation, on the basis of the above embodiments, the embodiments of the present application provide an adaptive snow goose algorithm driven residual network bearing fault diagnosis system, which is used to execute the adaptive snow goose algorithm driven residual network bearing fault diagnosis method in the above method embodiment.

[0207] The system comprises: a data acquisition module, configured to acquire bearing fault data to be diagnosed; and a fault diagnosis module, configured to input the bearing fault data to be diagnosed into a trained fault diagnosis model, and output a fault diagnosis result; wherein the training of the fault diagnosis model comprises: constructing a bearing fault sample data set; and constructing the fault diagnosis model, which comprises: a frequency domain attention module, configured to perform frequency domain perception attention feature enhancement on an input bearing fault sample; a flexible residual network, configured to perform deep feature extraction on the enhanced feature through a double flexible residual block, the double flexible residual block comprising two series-connected flexible residual paths and the number of convolution layers of each flexible residual block being adjustable, meanwhile, the flexible residual network further introduces a dynamic selection mechanism of a skip connection position to determine an access position and an output position of the flexible residual path, and introduces a cross-path cross-connection weight and a global average pooling attention mechanism to realize information intercommunication and feature enhancement between different residual paths; an adaptive snow goose algorithm is introduced to optimize network structure parameters of the fault diagnosis model, and based on the optimal network structure parameters, the model is trained on the constructed sample data set to output the trained fault diagnosis model.

[0208] The adaptive snow goose algorithm driven residual network bearing fault diagnosis system provided by the embodiment of the present application is used to solve the problems of difficulty in intelligent selection of deep neural network topology configuration parameters and insufficiency of vibration signal frequency domain key feature extraction in bearing fault diagnosis of rotating machinery, adopts the foregoing modules, and through deep coupling of the adaptive snow goose algorithm and the double flexible residual network with frequency domain attention enhancement and collaborative use of time-frequency domain information, realizes high-accuracy recognition and intelligent discrimination of bearing degradation types, and effectively solves the problems of lack of flexibility of network architecture and insufficiency of frequency domain information extraction.

[0209] It should be noted that the system embodiments provided by the present application are used to implement the methods in the method embodiments and the methods in other method embodiments provided by the present application, and the difference is only that the corresponding functional modules are set, the principle is basically the same as that of the above-mentioned system embodiments provided by the present application, as long as the technical personnel in the art improve the modules in the above-mentioned system embodiments on the basis of the above-mentioned system embodiments, refer to the specific technical solutions in other method embodiments, obtain the corresponding technical means through the combination of technical features, and the technical solutions composed of these technical means, on the premise of ensuring the practicability of the technical solutions, obtain the corresponding system class embodiments, and are used to implement the methods in other method class embodiments.

[0210] Based on the same inventive concept as any of the preceding embodiments, the embodiments of the application also provide a bearing fault diagnosis device driven by a residual network of an adaptive snow goose algorithm, comprising a memory and a processor, the memory storing program instructions executed by the processor, and the processor calling the program instructions to execute the bearing fault diagnosis method driven by the residual network of the adaptive snow goose algorithm.

[0211] Based on the same inventive concept as any of the preceding embodiments, the embodiments of the application also provide a non-transitory computer-readable storage medium storing computer instructions, which cause the computer to execute the bearing fault diagnosis method driven by the residual network of the adaptive snow goose algorithm.

[0212] In summary of the embodiments, the application is suitable for solving the problems of rigid topology structure of deep neural network, shallow mining of vibration signal spectrum features, and difficulty in intelligent matching of network configuration parameters. Firstly, a flexible residual network modeling system is constructed, a scalable convolution layer configuration mechanism is designed to realize elastic adjustment of network depth, a dynamic jump connection selection strategy is established to adapt to multi-path transmission of gradient information, a double-path residual block cascade structure is used to strengthen the non-linear transformation ability of features, and cross-path cross-connection and attention mechanism are combined to realize effective integration of multi-level features. Secondly, a frequency domain perception attention module is proposed, bidirectional mapping of time-frequency space is completed through fast Fourier transform operation, amplitude-phase decoupling analysis strategy is designed to deeply analyze the multi-dimensional discriminant attributes of the spectrum, and a double-layer attention architecture of frequency channel weighting and frequency domain space focusing is constructed. Thirdly, a multi-stage evolutionary snow goose algorithm is designed, the optimization stage of the algorithm is dynamically identified through the joint criterion of population diversity and fitness improvement amount, the search space coverage breadth is widened by introducing opposite solution generation and elite guidance strategy, a memory storage and weighted utilization mechanism of historical high-quality solutions is established, and the group intelligence of the algorithm is enhanced by fusing individual cooperation strength calculation and collective foraging behavior simulation. Finally, an integrated diagnosis architecture of ISGA-DFResNet-FDAM is constructed to realize intelligent exploration of network topology structure parameters and joint training of deep model weights, through the hierarchical cooperation of upper ISGA topology structure search and lower residual network depth learning, accurate classification of multiple types of damage such as bearing inner ring defect, outer ring crack and rolling element spalling is realized.

[0213] The application can effectively improve the classification accuracy of bearing fault diagnosis under strong noise background, improve the adaptive configuration ability and multi-working condition generalization recognition level of deep learning architecture, and ensure the predictive maintenance and safe and reliable operation of rotating equipment. The key points are:

[0214] 1. Multi-strategy collaborative improvement of ISGA algorithm, including multi-stage evolution discrimination mechanism to dynamically switch search strategies according to population state, opposite solution generation strategy to broaden the search range of solution space, elite memory management mechanism to save and use historical high-quality solutions, and cooperative foraging mechanism to enhance the information interaction ability between individuals.

[0215] 2. Adaptive topology design of double flexible residual network, including scalable convolutional layer number to realize dynamic configuration of network depth, dynamic jump connection position selection to adapt feature transmission path, double path flexible residual block concatenation to enhance nonlinear fitting capability, and cross-path cross-connection and attention fusion to strengthen multi-scale features, so that the network can adaptively adjust the topology parameters according to the fault vibration characteristics.

[0216] 3. Frequency domain perception attention enhancement module based on fast Fourier transform, which provides accurate frequency spectrum discrimination basis for deep feature extraction.

[0217] The terms "comprising" and "having" and any variations thereof in the specification and claims and above-described drawings are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or apparatus that includes a series of steps or units, not necessarily limited to which steps or units are clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or apparatus.

[0218] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the technical solutions of the present application.

Claims

1. A residual network bearing fault diagnosis method driven by an adaptive snow goose algorithm, characterized in that, include: Obtain the bearing fault data to be diagnosed; The bearing fault data to be diagnosed is input into the trained fault diagnosis model, and the fault diagnosis result is output; wherein, the training of the fault diagnosis model includes: Construct a bearing failure sample dataset; A fault diagnosis model is constructed, including: a frequency domain attention module, used to enhance the frequency domain perception attention features of the input bearing fault samples; and a flexible residual network, used to extract deep features from the enhanced features through dual flexible residual blocks. The dual flexible residual blocks contain two serial flexible residual paths and the number of convolutional layers in each flexible residual block is adjustable. At the same time, the flexible residual network also introduces a dynamic selection mechanism for skip connection positions to determine the access and output positions of the flexible residual paths, and introduces cross-path cross-connection weights and a global average pooling attention mechanism to realize information exchange and feature enhancement between differentiated residual paths. An adaptive snow goose algorithm is introduced to optimize the network structure parameters of the fault diagnosis model. Based on the optimal network structure parameters, the model is trained on the constructed sample dataset, and the trained fault diagnosis model is output. The construction of the adaptive snow goose algorithm includes: Initialize the Snow Goose algorithm population, with each individual's encoding containing network structure parameters; Establish a multi-stage evolutionary discrimination criterion function, and dynamically identify the evolutionary stage of the algorithm based on the population dispersion index and fitness increment; A strategy for generating complementary solutions is introduced to generate complementary solutions for the current candidate solutions and combine this with guidance from elite individuals, thereby broadening the search range of the solution space. Design an elite solution memory bank management mechanism to store high-quality solutions from historical iterations and guide population evolution through a weighted selection strategy; Construct a cooperative foraging mechanism among individuals, calculate the cooperative strength index among individuals and incorporate it into the position update equation to enhance the collective search ability of the population; The flight strategy is adaptively switched between the V-formation flight strategy in the exploration phase and the linear flight strategy in the development phase based on the results of multi-stage evolution.

2. The adaptive snow goose algorithm-driven residual network bearing fault diagnosis method according to claim 1, characterized in that, An adaptive snow goose algorithm is introduced to optimize the network's structural parameters, including: Set the execution parameters of the adaptive snow goose algorithm, including the number of individuals in the population, the maximum number of generations, the capacity of the elite memory bank, and the probability of adversarial learning; For each individual in the current population, construct the network topology of the corresponding fault diagnosis model; The constructed network topology is iteratively trained on the training set samples; The classification performance of the network topology is evaluated on the validation set samples, and the recognition accuracy is used as the fitness value of the individual. The population evolution operation of the adaptive snow goose algorithm is performed, and multi-stage adaptation, opposition solution generation, elite memory guidance and cooperative foraging strategies are applied. When the maximum number of generations or fitness convergence criteria are met, the parameter configuration scheme of the globally optimal individual is output.

3. The adaptive snow goose algorithm-driven residual network bearing fault diagnosis method according to claim 2, characterized in that, Based on the optimal network structure parameters, the model is trained on the constructed sample dataset, including: The network topology of the fault diagnosis model is constructed based on the optimal network structure parameters obtained by the adaptive snow goose algorithm. Configure the hyperparameters for network training; The optimal network structure parameters are iteratively trained on the training sample set, and the performance of the trained network model is evaluated on the test sample set. Output fault identification results, including overall accuracy, macro average precision, macro average recall, and macro average F1 score.

4. The adaptive snow goose algorithm-driven residual network bearing fault diagnosis method according to claim 2, characterized in that, Before performing iterative optimization, the network structure parameter search space is bounded as follows: Define the search range for the number of convolutional layers for the first and second flexible residual blocks; Determine the search interval for the number of jump connections between the first flexible residual block and the second flexible residual block; A set of candidate values ​​for configuring the number of convolutional kernels.

5. The adaptive snow goose algorithm-driven residual network bearing fault diagnosis method according to claim 1, characterized in that, The construction of the frequency domain attention module also includes: Perform a Fast Fourier Transform on the feature tensor of the input bearing fault samples; The amplitude and phase of the frequency domain complex result after the fast Fourier transform are decoupled and separated to obtain the spectral amplitude matrix and phase angle matrix respectively. Design a frequency channel attention weighting mechanism to achieve feature selection in the channel dimension by learning the importance coefficients of different frequency components; Construct a frequency domain spatial attention focusing mechanism to focus on salient regions at different frequency positions in the spectral amplitude matrix; Perform frequency domain feature parameterization enhancement and phase compensation operations to optimize the spectral representation through a trainable adaptive filter; The enhanced frequency domain representation is reconstructed back to the time domain using the inverse fast Fourier transform, and the original time domain features are integrated through a gating fusion mechanism to obtain the final enhanced features.

6. The adaptive snow goose algorithm-driven residual network bearing fault diagnosis method according to claim 1, characterized in that, Construct a bearing failure sample dataset, including: Collect bearing vibration data under various speed conditions; The original vibration signal is serialized and segmented using a sliding window method. The window width and sliding step are set to generate a sample sequence of fixed length. Perform Z-score normalization on the segmented sample sequences; The normalized samples are split into datasets, and the ratio of training set to test set is set.

7. An adaptive snow goose algorithm-driven residual network bearing fault diagnosis system, used to implement the adaptive snow goose algorithm-driven residual network bearing fault diagnosis method according to any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to acquire bearing fault data to be diagnosed. The fault diagnosis module is used to input the bearing fault data to be diagnosed into the trained fault diagnosis model and output the fault diagnosis result; wherein, the training of the fault diagnosis model includes: Construct a bearing failure sample dataset; A fault diagnosis model is constructed, including: a frequency domain attention module, used to enhance the frequency domain perception attention features of the input bearing fault samples; and a flexible residual network, used to extract deep features from the enhanced features through dual flexible residual blocks. The dual flexible residual blocks contain two serial flexible residual paths and the number of convolutional layers in each flexible residual block is adjustable. At the same time, the flexible residual network also introduces a dynamic selection mechanism for skip connection positions to determine the access and output positions of the flexible residual paths, and introduces cross-path cross-connection weights and a global average pooling attention mechanism to realize information exchange and feature enhancement between differentiated residual paths. An adaptive snow goose algorithm is introduced to optimize the network structure parameters of the fault diagnosis model. Based on the optimal network structure parameters, the model is trained on the constructed sample dataset, and the trained fault diagnosis model is output.

8. A residual network bearing fault diagnosis device driven by an adaptive snow goose algorithm, characterized in that, The method includes a memory and a processor, wherein the memory stores program instructions that are executed by the processor, and the processor invokes the program instructions to execute the adaptive snow goose algorithm-driven residual network bearing fault diagnosis method according to any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the adaptive snow goose algorithm-driven residual network bearing fault diagnosis method as described in any one of claims 1 to 6.

Images