Mechanical fault feature extraction method and model construction method, system and device

Abstract

The present application relates to a mechanical fault feature extraction method and a model construction method, system and device, belonging to the field of mechanical equipment fault diagnosis. The construction method comprises: collecting a fault data set; building a fault feature extraction initial model comprising a multi-order positioning attention module and a multi-order positioning dense network; training and verifying the fault feature extraction initial model based on the fault data set; and obtaining a mechanical fault feature extraction model based on the results of training and verification. The extraction model performs multi-order re-calibration on the data set of the non-stationary variable load signal, and then extracts the deep features of the non-stationary variable load signal. Through the two stages of cascaded global recognition and local positioning, the pulse signal segment is gradually and deeply positioned, and the problem of pulse feature extraction in the non-stationary variable load signal is solved.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment fault diagnosis, and specifically relates to mechanical fault feature extraction methods and model construction methods, systems, and equipment. Background Technology

[0002] In mechanical equipment fault diagnosis, fault feature extraction is a crucial step. However, current traditional fault diagnosis methods for variable load conditions involve complex and cumbersome feature extraction processes, making them unsuitable for the fault diagnosis needs of the big data era. Dense networks, with their dense connection mechanism, excel in feature extraction, achieving adaptive feature mining of massive amounts of data. However, the features extracted from different channels contain a huge amount of information, and even within the same channel, there are numerous signal segments. Excessive redundancy can affect the accuracy of fault diagnosis. Among these numerous signal segments, periodic pulse signal segments are key features reflecting fault behaviors such as impacts and gaps, containing crucial fault information. Therefore, accurately extracting pulse features will further improve the accuracy of fault diagnosis. However, under variable load conditions, signals exhibit non-stationary characteristics, the intervals between pulse signal segments are irregular, and they may be submerged by noise signals, making pulse signal extraction even more difficult.

[0003] For example, CN116894215A discloses a gearbox fault diagnosis method based on semi-supervised dynamic graph attention. This method includes: obtaining fixed-length diagnostic samples from the original vibration signal of the gearbox, then using FFT as input KNN graph nodes for a graph attention network; calculating the fuzzy distance between graph nodes using a pooling strategy; introducing a dynamic attention mechanism to address the problem of similar weight allocation for different types of nodes in a static graph attention network; constructing a dynamic multi-head graph attention gearbox fault diagnosis model and using the Softmax function as a classifier; achieving semi-supervised learning through a label propagation algorithm when there are few labeled samples; training the model using the Adam optimizer via backpropagation; and saving the trained fault diagnosis model for online diagnosis. While this method overcomes the limited performance of static attention and enhances the attention mechanism's focus and differentiation of different types of faults, it suffers from a complex diagnostic process, high computational load, and the inability to solve the problem of difficult pulse feature extraction in non-stationary variable load signals. For example, the gearbox imbalance fault diagnosis method disclosed in CN116975563A includes: signal acquisition, obtaining vibration signals of the gearbox under various operating conditions and different health conditions, constructing a training set and a test set based on the vibration signals, wherein the normal state samples are more numerous than the fault state samples; environmental simulation construction, establishing a Markov decision process for imbalance classification, designing a reward function, and establishing a data environment simulation required for fault diagnosis under imbalance; establishing a multi-operating-condition imbalance deep reinforcement learning network, based on the training set, the agent continuously interacts with the environment, training the agent to autonomously learn the optimal diagnosis strategy, the agent including at least two multi-scale feature deep attention networks with the same structure; fault identification, inputting the samples in the test set one by one into the trained agent, identifying the gearbox fault type according to the diagnosis strategy, and analyzing the diagnosis results. While this method can adaptively weight feature information from different channels using a channel attention mechanism to address the redundancy of feature information under different channels and recalibrate multi-scale feature information to highlight more important effective features in each state, thereby enhancing the generalization of fault diagnosis, it also suffers from complex processes and cannot solve the problem of extracting pulse features from non-stationary load-varying signals. It is evident that existing gearbox fault diagnosis methods suffer from complex diagnostic processes and difficulty in extracting pulse features from non-stationary load-varying signals. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method, system, and equipment for extracting mechanical fault features and constructing models.

[0005] The first objective of this invention is to provide a method for constructing a mechanical fault feature extraction model, comprising:

[0006] Collect fault datasets;

[0007] An initial model for fault feature extraction, comprising a multi-level localization attention module and a multi-level localization dense network, is constructed.

[0008] Based on the fault dataset, the initial model for fault feature extraction is trained and validated.

[0009] Based on the training and validation results, a mechanical fault feature extraction model is obtained.

[0010] In a specific embodiment of the present invention, the multi-level localization attention module is a two-level localization attention module, which includes a first-level attention module and a second-level attention module.

[0011] The multi-level localization dense network includes convolutional pooling, multiple merged improved dense layers, global feature pooling, average pooling, fully connected layers, and a softmax classifier.

[0012] The improved dense layer consists of a dense layer and a multi-level localization attention module embedded in the dense layer.

[0013] In a specific embodiment of the present invention, the number of the merged improved dense layers is 6.

[0014] In a specific embodiment of the present invention, the step of training and validating the initial model for fault feature extraction based on the fault dataset includes:

[0015] The training set from the fault dataset is used as input to extract fault features and train the initial model.

[0016] The validation set from the fault dataset is input into the trained fault feature extraction initial model for validation.

[0017] In a specific embodiment of the present invention, the step of inputting the training set from the fault dataset into the fault feature extraction initial model for training includes:

[0018] The convolutional model is trained with reduced feature dimensions based on the training set in the fault dataset, and the training set with reduced feature dimensions is output.

[0019] Training for multi-level recalibration is performed on the training set with reduced feature dimensions, and the multi-level recalibrated training set is output.

[0020] The fault feature classification is trained based on the multi-level recalibrated training set, and the mechanical fault category feature set is output.

[0021] In a specific embodiment of the present invention, when the multi-order recalibration is a second-order recalibration, it includes:

[0022] The input features are passed through a first-order attention module to obtain a first-order weight map;

[0023] The first-order relabeled features are obtained by multiplying the first-order weight map element by element with the input features.

[0024] The features recalibrated in the first order are passed through the second order attention module to obtain the second order weight map;

[0025] The second-order weighted graph is multiplied element-wise with the features of the first-order recalibration to obtain the features of the second-order recalibration, thus completing the second-order recalibration.

[0026] In a specific embodiment of the present invention, the step of inputting the validation set from the fault dataset into the trained fault feature extraction initial model for validation includes:

[0027] The validation set from the fault dataset is input into the trained fault feature extraction initial model, and the values ​​of the number of iterations and learning rate are tested using the grid search method.

[0028] Optimize grid hyperparameters based on the fault diagnosis accuracy and computational resource utilization after testing;

[0029] Under different network structures, the validation set in the fault dataset is input into the initial model after training to extract fault features, and validation experiments of attention module ablation are carried out.

[0030] In a specific embodiment of the present invention, the different network structures include a baseline network, a network employing a first-order attention module, and a network employing a second-order attention module.

[0031] The second objective of this invention is to provide a method for extracting mechanical fault features, comprising:

[0032] Input the dataset to be extracted into the mechanical fault feature extraction model described above;

[0033] The mechanical fault feature extraction model described above extracts fault category features from the dataset to be extracted and outputs mechanical fault category features.

[0034] The third objective of this invention is to provide a system for constructing a mechanical fault feature extraction model, comprising:

[0035] Collection module: Used to collect fault datasets;

[0036] Building module: Used to build an initial model for fault feature extraction that includes a multi-level localization attention module and a multi-level localization dense network;

[0037] Training and optimization module: used to train and validate the initial model for fault feature extraction based on the fault dataset;

[0038] Acquisition Module: Used to obtain a mechanical fault feature extraction model based on the training and validation results.

[0039] A fourth objective of the present invention is to provide an electronic device, characterized in that it comprises: a processor coupled to a memory;

[0040] The memory is used to store computer programs;

[0041] The processor is configured to execute the computer program stored in the memory, so that the electronic device performs the construction method described above.

[0042] A fifth objective of the present invention is to provide a computer-readable storage medium, characterized in that the computer-readable storage medium stores a program or instructions that, when the program or instructions are run on a computer, cause the computer to perform the construction method as described above.

[0043] The beneficial effects of this invention are:

[0044] The mechanical fault feature extraction method, model construction method, system, and equipment of the present invention, by building an initial fault feature extraction model including a multi-level positioning attention module and a multi-level positioning dense network, performs multi-level recalibration on the dataset of non-stationary load signal, and then extracts the deep features of the non-stationary load signal (such as gradually locating the pulse signal segment through a cascaded global identification and local positioning two-stage process), thereby achieving "end-to-end" fault diagnosis.

[0045] Therefore, this invention can remove redundant information through hierarchical feature recalibration, adaptively locate pulse signal segments, and realize pulse feature extraction in non-stationary load signals, thus solving the problem of difficult pulse feature extraction in non-stationary load signals. Moreover, the extraction process of this invention is simple.

[0046] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A flowchart illustrating a method for constructing a mechanical fault feature extraction model according to an embodiment of the present invention is shown;

[0049] Figure 2A diagram of a two-order localization dense network structure according to an embodiment of the present invention is shown;

[0050] Figure 3 A structural diagram of a two-order localization attention module according to an embodiment of the present invention is shown;

[0051] Figure 4 The following diagram illustrates the classification accuracy under different hyperparameters according to an embodiment of the present invention;

[0052] Figure 5 A visualization of intermediate layer features according to an embodiment of the present invention is shown;

[0053] Figure 6 A graph showing the sample imbalance evaluation results according to an embodiment of the present invention is provided.

[0054] Figure 7 A framework diagram of a system for constructing a mechanical fault feature extraction model according to an embodiment of the present invention is shown;

[0055] Figure 8 A frame diagram of an electronic device according to an embodiment of the present invention is shown;

[0056] In the diagram: Collection module 10; Construction module 20; Training and optimization module 30; Acquisition module 40; Electronic device 300; Processor 301; Memory 302. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] like Figure 1 As shown, a method for constructing a mechanical fault feature extraction model according to an embodiment of the present invention includes:

[0059] S1. Collect fault dataset;

[0060] S2. Construct an initial model for fault feature extraction that includes a multi-level localization attention module and a multi-level localization dense network;

[0061] S3. Based on the fault dataset, train and validate the initial model for fault feature extraction;

[0062] S4. Based on the training and validation results, a mechanical fault feature extraction model is obtained.

[0063] In some embodiments of the present invention, in step S1, the fault dataset is exemplarily pulse data in a non-stationary load change signal. Fault labels are added to the pulse data in the original non-stationary load change signal and fault samples are generated according to a specified length, thus forming a fault dataset. At the same time, the fault dataset is further divided into a training set, a validation set, and a test set.

[0064] In some embodiments of the present invention, in step S2, the multi-stage localization attention module is exemplarily a two-stage localization attention module (TAM) to solve the problem of difficulty in extracting pulse features in non-stationary variable load signals. The TAM includes a first-stage attention module (hereinafter referred to as SE) and a second-stage attention module (hereinafter referred to as SAM), specifically as follows: Figure 3 As shown;

[0065] The TAM is used to perform two-level calibration on the input element Fij (the j-th signal segment of the i-th channel). When it obtains greater weight and higher importance, it will play a key role in the diagnostic results of fault category characteristics.

[0066] The TAM, through the close cascading between the SE and SAM levels, allows attention to be deepened step by step, focusing on the most critical information in the diagnostic task, and ultimately achieving precise positioning of the pulse signal segment, which is beneficial for subsequent accurate fault diagnosis.

[0067] In some embodiments of the present invention, in step S2, the multi-level localization dense network includes convolutional pooling, multiple merged improved dense layers, global feature pooling, average pooling, fully connected layers, and a softmax classifier.

[0068] The improved dense layer consists of a dense layer and a multi-level localization attention module embedded in the dense layer;

[0069] When the multi-level localization attention module is a two-level localization attention module, the multi-level localization dense network is correspondingly a two-level localization dense network;

[0070] The two-stage localization dense network is based on a densely connected network and incorporates the two-stage localization attention module to establish a two-stage localization dense network (TLDN). Through a progressively deepening feature recalibration process, it achieves pulse feature extraction of the variable-load gearbox. The specific network structure is as follows: Figure 2 As shown;

[0071] One-dimensional vibration signals are input into the network. First, convolutional pooling (consisting of 1×7 convolutional layers and max pooling) reduces the feature dimensionality. Next, an improved dense layer (I-Dense Layer) is used to extract pulse features from non-stationary variable load signals. The I-Dense Layer introduces a TAM module on top of the original Dense Layer. The SE module in the TAM adaptively enhances important channel features using global information to obtain discriminative features. The SAM module in the TAM considers local information to locate key signal segments in the channel features. The SE and SAM modules are cascaded, fully leveraging their respective advantages to progressively deepen attention, ultimately focusing on the most critical information in the task, achieving precise location of pulse signal segments, and solving the problem of difficult pulse feature extraction from non-stationary variable load signals.

[0072] Furthermore, in this embodiment of the invention, the stacked 6 I-Dense Layers achieve the merging (Concat) of features from each channel through a dense connection mechanism, ultimately forming a global feature pool with multi-level recalibration; finally, the classification of fault features of the variable load gearbox is achieved through average pooling, fully connected layers, and a softmax classifier.

[0073] In some embodiments of the present invention, step S3 includes:

[0074] S3-1. Input the training set from the fault dataset into the fault feature extraction initial model for training;

[0075] S3-2. Input the validation set from the fault dataset into the trained fault feature extraction initial model for validation.

[0076] In some embodiments of the present invention, step S3-1 includes:

[0077] S3-1-1. Train the convolutional network to reduce the feature dimension based on the training set in the fault dataset, and output the training set with reduced feature dimension. The reduction of feature dimension is achieved through convolutional pooling operation in the multi-level localization dense network.

[0078] S3-1-2. Train multi-level recalibration based on the training set with reduced feature dimensions, and output the multi-level recalibrated training set, wherein the multi-level recalibration is achieved through an improved dense layer in a multi-level localization dense network.

[0079] S3-1-3. Train the fault feature classification based on the multi-level recalibrated training set, and output the mechanical fault category feature set.

[0080] In some embodiments of the present invention, when the multi-order localization dense network is a two-order localization dense network, the multi-order recalibration is a two-order recalibration, including:

[0081] A1. Input features are passed through a first-order attention module to obtain a first-order weight map;

[0082] A2. The first-order recalibrated features are obtained by multiplying the first-order weight map element by element with the input features.

[0083] A3. The features recalibrated in the first order are passed through the second order attention module to obtain the second order weight map;

[0084] A4. Multiply the second-order weighted graph element by element with the features of the first-order recalibration to obtain the features of the second-order recalibration, thus completing the second-order recalibration.

[0085] The specific expressions for steps A1-A4 are shown in equations (1)-(2):

[0086]

[0087] In equations (1)-(2), F, F', and F” represent the input features, the first-order recalibrated features, and the second-order recalibrated features, respectively; M c (F), M s (F′) represent the first-order weighted graph and the second-order weighted graph, respectively; This indicates element-wise multiplication.

[0088] In some embodiments of the present invention, step S3-2 includes:

[0089] S3-2-1. Input the validation set from the fault dataset into the initial model for fault feature extraction after training, and test the values ​​of the number of iterations and the learning rate using the grid search method.

[0090] S3-2-2. Optimize grid hyperparameters based on the fault diagnosis accuracy and computational resource utilization rate after testing;

[0091] S3-2-3. Under different network structures, the validation set in the fault dataset is input into the initial model after training to extract fault features, and a validation experiment of attention module ablation is carried out.

[0092] In some embodiments of the present invention, in step S3-2-3, the different network structures include a baseline network (DenseNet), a network using a first-order attention module (DenseNet-SE), and a network using a second-order attention module (DenseNet-SAM). By establishing the above three network structures, the impact of different attention modules in TAM on diagnostic performance is verified.

[0093] Based on the completion of testing the trained initial model for fault feature extraction using the validation set, the network parameters of the initial model for fault feature extraction were updated, thus obtaining the mechanical fault feature extraction model.

[0094] A method for extracting mechanical fault features according to an embodiment of the present invention includes:

[0095] X1. Input the dataset to be extracted into the mechanical fault feature extraction model obtained in the above embodiment;

[0096] X2. The mechanical fault feature extraction model obtained in the above embodiments is used to extract fault category features from the dataset to be extracted, and outputs mechanical fault category features.

[0097] The following illustrates the process of constructing a mechanical fault feature extraction model based on a fault dataset of gearbox vibration signals, as well as the evaluation process after the model diagnoses the fault category. The gearbox vibration signals are obtained through fault diagnosis tests conducted on a gearbox fault simulation test bench.

[0098] The test bench has a rated power of 3kW and mainly includes a gearbox, acceleration sensor, motor, data acquisition unit, etc.

[0099] I. Data Collection

[0100] The sampling frequency of the gearbox vibration signal is 12kHz, the fault sample length is 1024, the sliding window movement step is 800, and the batch size is 64. Data is obtained for five states: spalling, pitting, broken teeth, missing teeth, and normal. Three working conditions are set for each state, with 200 samples selected for each working condition, for a total of 3000 samples.

[0101] Add fault labels to the sample data and generate fault samples according to the specified length, as shown in Table 1;

[0102] Subsequently, the training set, test set, and validation set were divided into sets of [0.7, 0.15, 0.15], as shown in Table 2.

[0103] Table 1

[0104]

[0105] Table 2

[0106]

[0107]

[0108] 1. Model Building

[0109] This invention implements an initial model for extracting mechanical fault features using Python. The programming environment is AMD Ryzen 7 5800H CPU, NVIDIA GeForce RTX 3060, Python 3.6, and TensorFlow 2.0.

[0110] The initial model for mechanical fault feature extraction was trained using the training set to obtain the trained mechanical fault feature extraction model.

[0111] 2. Model Validation

[0112] (1) Selection of hyperparameters

[0113] A grid search was performed on the validation set using the combination of two hyperparameters: the number of iterations and the learning rate. The search range for the number of iterations was [20, 50, 100, 200], and the search range for the learning rate was [0.1, 0.01, 0.001, 0.0001]. The classification accuracy obtained under each parameter combination is as follows: Figure 4 As shown.

[0114] The results show that, except in a few cases, these two hyperparameters have little impact on the fault diagnosis results. Under most parameter combinations, the final accuracy of this method reaches over 95%, maintaining a high level and demonstrating good fault diagnosis performance. This indicates that the method has good hyperparameter stability, achieving excellent fault diagnosis results without excessive hyperparameter tuning, and has good potential for practical application. Considering both fault diagnosis accuracy and computational resource utilization, the number of iterations was set to 50 and the learning rate to 0.001 in subsequent experiments.

[0115] (2) Attention Module Ablation Experiment

[0116] To verify the impact of different attention modules in a two-order localization attention module on diagnostic performance, an ablation experiment was designed under simulated ambient noise conditions with a signal-to-noise ratio of -1 dB. Three network structures were established: ① a baseline network DenseNet without any attention modules; ② a DenseNet-SE network using only SE; and ③ a DenseNet-SAM network using only SAM.

[0117] The ablation test results are shown in Table 3. Except for peeling faults, TLDN (an abbreviation for Two-Stage Localization Dense Networks) showed better recognition accuracy than other networks for the other four types of fault states, especially in the normal sample category, where the recognition accuracy reached 100%. This indicates that even under noise interference, this method can still distinguish normal samples from fault samples with high accuracy, demonstrating that TLDN has good practical application potential.

[0118] Table 3

[0119]

[0120] Further analysis of overall diagnostic accuracy revealed that, compared to the baseline network DenseNet, the overall diagnostic performance of networks such as DenseNet-SAM, DenseNet-SE, and TLDN was improved, with accuracy increases of 0.4, 2.2, and 3.3 percentage points, respectively. The TLDN network, by fusing the two attention mechanisms mentioned above, significantly outperformed networks using a single attention mechanism in fault diagnosis. This indicates that the fusion of the two attention mechanisms can fully leverage their respective advantages. SE removes redundant information through global relationships, thereby guiding SAM to focus on local relationships. These two stages progress step-by-step, ultimately achieving accurate pulse signal localization and further improving fault diagnosis performance.

[0121] 3. Model Evaluation

[0122] (1) Noise immunity assessment

[0123] For non-stationary variable load signals, impulse features are easily submerged by noise. To evaluate the ability of this method to extract impulse features under noisy conditions, three different intensities of Gaussian white noise were added to the original sample set. The proposed TLDN method was then compared with three widely used networks (CNN, ResNet34, and DenseNet) and four common attention mechanisms (Self-Attention, CBAM, ECA, and No-local), with DenseNet as the base network for the four attention mechanisms. The results are shown in Table 4.

[0124] Table 4

[0125]

[0126] Under all three noise conditions, TLDN outperformed other networks in accuracy. Even in a strong noise environment with a signal-to-noise ratio of -1dB, TLDN still achieved a diagnostic accuracy of 98.89%, an improvement of 23.11 percentage points compared to CNN.

[0127] As noise intensity increases, the recognition rates of all diagnostic methods except TLDN decrease. However, due to varying noise robustness, the degree of accuracy decline also differs among these methods. CNN is most significantly affected by random noise, while the other methods are less affected and show little change in accuracy. This indicates that cross-layer connected network structures can still effectively extract features under noise interference, demonstrating good noise resistance.

[0128] In summary, the TLDN model of this invention maintains a diagnostic accuracy of over 98% under various noise conditions, demonstrating its strong robustness, stability, and noise resistance. Utilizing the TLDN model enables accurate extraction of pulse features even under noise interference, thereby achieving more precise fault diagnosis of variable load gearboxes.

[0129] To more intuitively illustrate the diagnostic process, the t-distributed stochastic neighbor embedding (t-SNE) algorithm is used to reduce the dimensionality and cluster the output features of the intermediate layers of the network. The visualization effect is as follows: Figure 5 As shown, the initial raw data distribution was chaotic and disordered, making it difficult to distinguish fault categories. Through layer-by-layer feature mining, after the second dense layer, samples of the same type gradually began to cluster; after the sixth dense layer, samples of the same fault category clustered together; the final fully connected layer feature visualization shows that the boundaries between different fault types are clear, and faults of the same type are tightly clustered together, indicating that this method has good fault classification performance even under noise interference.

[0130] (2) Imbalanced sample assessment

[0131] The TLDN model of this invention is compared with various networks, and the results are as follows: Figure 6 As shown, in five identical repeated trials comparing eight methods, CNN and ResNet34 both achieved accuracy below 70%, while the remaining methods based on DenseNet achieved accuracy above 80%, further highlighting the superiority of the DenseNet network. The TLDN model in this embodiment achieved the best diagnostic results in all five repeated trials, demonstrating that this method also has good diagnostic performance under imbalanced sample conditions, further validating the performance of the proposed method.

[0132] In summary, the embodiments of this invention propose a gearbox fault feature extraction model under variable load, which utilizes a dense network to achieve adaptive deep feature extraction of non-stationary variable load signals. At the same time, a two-order localization attention module is designed to accurately locate pulse signal segments, solving the problem of difficult pulse feature extraction in non-stationary variable load signals, and has strong feature extraction capabilities.

[0133] like Figure 7 As shown, a system for constructing a mechanical fault feature extraction model includes:

[0134] Collection module 10: Used to collect fault datasets;

[0135] Module 20: Used to build an initial model for fault feature extraction that includes a multi-level localization attention module and a multi-level localization dense network;

[0136] Training and optimization module 30: Used to train and validate the initial model for fault feature extraction based on the fault dataset;

[0137] Module 40: Used to obtain a mechanical fault feature extraction model based on the training and validation results.

[0138] like Figure 8 As shown, in some embodiments of the present invention, an electronic device is provided, the electronic device 300 including: a processor 301 coupled to a memory 302;

[0139] The memory 302 is used to store computer programs;

[0140] The processor 301 is configured to execute the computer program stored in the memory 302, so that the electronic device performs the method described in the above embodiments.

[0141] In some embodiments of the present invention, a computer-readable storage medium is provided that stores a program or instructions that, when executed on a computer, cause the computer to perform the methods described in the above embodiments.

[0142] According to embodiments of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, electronic device, or apparatus.

[0143] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a mechanical failure feature extraction model, characterized in that, include: Collect fault datasets; An initial model for fault feature extraction, comprising a multi-level localization attention module and a multi-level localization dense network, is constructed. Based on the fault dataset, the initial model for fault feature extraction is trained and validated. Based on the training and validation results, a mechanical fault feature extraction model is obtained.

2. The method of claim 1, wherein, The multi-level localization attention module is a two-level localization attention module, which includes a first-level attention module and a second-level attention module. The multi-level localization dense network includes convolutional pooling, multiple merged improved dense layers, global feature pooling, average pooling, fully connected layers, and a softmax classifier. The improved dense layer consists of a dense layer and a multi-level localization attention module embedded in the dense layer.

3. The method of claim 2, wherein, The number of the merged improved dense layers is 6.

4. The method of claim 1, wherein, The training, validation, and evaluation of the initial fault feature extraction model based on the fault dataset includes: The training set from the fault dataset is used as input to extract fault features and train the initial model. The validation set from the fault dataset is input into the trained fault feature extraction initial model for validation.

5. The method of claim 4, wherein, The step of inputting the training set from the fault dataset into the fault feature extraction initial model for training includes: The convolutional model is trained with reduced feature dimensions based on the training set in the fault dataset, and the training set with reduced feature dimensions is output. Training for multi-level recalibration is performed on the training set with reduced feature dimensions, and the multi-level recalibrated training set is output. The fault feature classification is trained based on the multi-level recalibrated training set, and the mechanical fault category feature set is output.

6. The method of claim 5, wherein the method further comprises: When the multi-order recalibration is converted to second-order recalibration, it includes: The input features are passed through a first-order attention module to obtain a first-order weight map; The first-order relabeled features are obtained by multiplying the first-order weight map element by element with the input features. The features recalibrated in the first order are passed through the second order attention module to obtain the second order weight map; The second-order weighted graph is multiplied element-wise with the features of the first-order recalibration to obtain the features of the second-order recalibration, thus completing the second-order recalibration.

7. The method according to any one of claims 4-6, wherein, The step of inputting the validation set from the fault dataset into the trained fault feature extraction initial model for validation includes: The validation set from the fault dataset is input into the trained fault feature extraction initial model, and the values ​​of the number of iterations and learning rate are tested using the grid search method. Optimize grid hyperparameters based on the fault diagnosis accuracy and computational resource utilization after testing; Under different network structures, the validation set in the fault dataset is input into the initial model after training to extract fault features, and validation experiments of attention module ablation are carried out.

8. The method of claim 7, wherein the method further comprises: The different network structures include a baseline network, a network using a first-order attention module, and a network using a second-order attention module.

9. A mechanical fault feature extraction method characterized by, include: Input the dataset to be extracted into the mechanical fault feature extraction model according to any one of claims 1-8; The mechanical fault feature extraction model according to any one of claims 1-8 extracts fault category features from the dataset to be extracted and outputs mechanical fault category features. 10.A system for constructing a mechanical failure feature extraction model, comprising: include: Collection module: Used to collect fault datasets; Building module: Used to build an initial model for fault feature extraction that includes a multi-level localization attention module and a multi-level localization dense network; Training and optimization module: used to train and validate the initial model for fault feature extraction based on the fault dataset; Acquisition Module: Used to obtain a mechanical fault feature extraction model based on the training and validation results.

11. An electronic device, comprising: include: Processor, the processor being coupled to memory; The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 8.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program or instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 8.

Images