Rotary machinery fault detection method based on improved grey wolf optimization and improved multi-agent LSTM algorithm

Abstract

The invention belongs to the field of rotating machine fault detection, and relates to a rotating machine fault detection method based on improved grey wolf optimization and an improved multi-agent LSTM algorithm, which comprises the following steps: acquiring to-be-detected rotating machine data; preprocessing the rotating machinery data; performing feature extraction on the preprocessed data; inputting the extracted features into an improved multi-agent LSTM network to obtain a deep feature map; selecting the deep feature map by adopting an improved grey wolf optimization algorithm to obtain an optimal feature map; inputting the optimal feature map into a classification model to obtain a fault detection result; according to the method, the wolf pack optimization algorithm is improved through the improved multi-agent LSTM network and the particle swarm optimization algorithm, so that the convergence efficiency of the algorithm is improved, and the optimal feature map is obtained more quickly.

Description

Technical Field

[0001] The present invention belongs to the field of rotating machinery fault detection, and specifically relates to a rotating machinery fault detection method based on improved grey wolf optimization and improved multi-agent LSTM algorithm. Background Art

[0002] In modern industrial operations, the efficiency and reliability of machinery are paramount, directly impacting overall productivity and safety. Key components in these systems include rolling element bearings, essential machine elements that provide essential support for the mechanical structures of various machines, including rotating shafts and associated components. These bearings are integral to the smooth operation of machinery, facilitating the rotation of components while minimizing friction and wear. Therefore, the proper functioning of rolling element bearings is crucial to maintaining the operational integrity of a wide range of industrial equipment, from production lines to heavy machinery in industries such as automotive, aerospace, and energy production. However, rolling element bearings are constantly used in a variety of harsh environmental conditions, making them susceptible to wear and eventual failure. Factors such as high rotational speeds, heavy loads, exposure to contaminants, insufficient lubrication, and extreme temperatures accelerate the deterioration process. Over time, this degradation leads to a significant decrease in bearing performance, ultimately leading to failure. The consequences of rolling element bearing failure can be catastrophic, resulting in severe machine failure, disrupting operations and causing significant financial losses. The financial impact of such failures extends beyond the direct cost of repair or replacement; it also encompasses the broader impact of unplanned downtime, production delays, and potential contractual penalties. Furthermore, rolling element bearing failures pose significant safety risks. In many industrial environments, especially those involving heavy machinery, bearing failures can lead to mechanical damage, potentially causing dangerous incidents such as fire, explosion, or the release of hazardous materials. These events threaten not only the safety of the equipment but also the lives of people working near the machinery. Therefore, early detection and diagnosis of bearing failures are crucial not only to prevent costly failures but also to ensure the safety of workers and the operating environment. Consequently, machine fault diagnosis has become an important area of research, with a focus on developing advanced technologies to monitor and diagnose bearing conditions in real time. Accurate and timely fault diagnosis allows for the implementation of predictive maintenance strategies that reduce the risk of unexpected failures, a pressing issue. Summary of the Invention

[0003] To solve the problems existing in the above-mentioned prior art, the present invention proposes a rotating machinery fault detection method based on an improved gray wolf optimization and an improved multi-agent LSTM algorithm. The method comprises: obtaining rotating machinery data to be detected; preprocessing the rotating machinery data, inputting the preprocessed data into a rotating machinery fault detection model, and obtaining a detection result;

[0004] Training a rotating machinery fault detection model involves:

[0005] S101, obtaining a rotating machinery dataset, and preprocessing the data in the rotating machinery dataset to obtain a training set;

[0006] S102, extracting features from the data in the training set;

[0007] S103, input the extracted features into the improved multi-agent LSTM network to obtain a deep feature map;

[0008] S104, using the improved gray wolf optimization algorithm to select the deep feature map to obtain the optimal feature map;

[0009] S105, inputting the optimal feature map into the classification model to obtain a fault detection result;

[0010] S106: Construct a loss function of the model based on the fault detection results, adjust the model parameters, and complete the model training when the loss function converges.

[0011] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned rotating machinery fault detection methods based on the improved grey wolf optimization and improved multi-agent LSTM algorithm.

[0012] To achieve the above-mentioned objectives, the present invention also provides a rotating machinery fault detection device based on improved gray wolf optimization and improved multi-agent LSTM algorithm, comprising a processor and a memory; the memory is used to store computer programs; the processor is connected to the memory, and is used to execute the computer programs stored in the memory, so that the rotating machinery fault detection device based on improved gray wolf optimization and improved multi-agent LSTM algorithm can execute any of the above-mentioned rotating machinery fault detection methods based on improved gray wolf optimization and improved multi-agent LSTM algorithm.

[0013] Beneficial effects of the present invention:

[0014] 1. This invention leverages an improved multi-agent LSTM network, particularly the hyperbolic tangent function's processing of cell states, which effectively reduces prediction oscillations and improves the model's stability and accuracy. The design of multi-layer LSTM stacks, fully connected layers, and dropout layers further enhances the model's ability to capture and generalize complex fault characteristics. The multi-agent system's parallel processing and information exchange mechanisms improve system processing efficiency and fault detection reliability.

[0015] 2. This paper improves the wolf pack optimization algorithm by using the particle swarm optimization algorithm, thereby increasing the algorithm's convergence efficiency and enabling faster acquisition of the optimal feature map. This paper also uses the IBGWOV algorithm combined with binary optimization to efficiently select features from the LSTM network output. This algorithm cleverly combines the advantages of the gray wolf optimization algorithm (GWO) and the particle swarm optimization algorithm (PSO), while using the V-shaped transfer function (V2-TF) to convert the continuous solution into a binary feature selection scheme. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is the overall flow chart of the present invention;

[0017] Figure 2 This is a flow chart of the improved gray wolf optimization algorithm of the present invention for processing deep feature maps;

[0018] Figure 3 This is the improved LSTM structure diagram of the present invention;

[0019] Figure 4 Flowchart of how the improved multi-agent LSTM of the present invention processes input features. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] This paper addresses these gaps by proposing a novel hybrid framework that integrates an improved multi-agent LSTM network with a RF classifier, enhanced by an improved GWO algorithm. This approach directly addresses the lack of post-extraction optimization in current methods, ensuring that the most relevant features selected after LSTM processing are used in the final classification stage. This approach combines the sequential data processing strength of the improved multi-agent LSTM with the powerful classification capabilities of the RF, and optimizes feature selection using an improved GWO algorithm. This integrated approach aims to improve the overall diagnostic accuracy and robustness of the system, especially in complex and dynamic industrial environments.

[0022] A rotating machinery fault detection method based on improved grey wolf optimization and improved multi-agent LSTM algorithm, such as Figures 1 to 4 As shown, the method includes: obtaining rotating machinery data to be detected; preprocessing the rotating machinery data, inputting the preprocessed data into a rotating machinery fault detection model to obtain a detection result; training the rotating machinery fault detection model includes:

[0023] S101, obtaining a rotating machinery dataset, and preprocessing the data in the rotating machinery dataset to obtain a training set;

[0024] S102, extracting features from the data in the training set;

[0025] S103, input the extracted features into the improved multi-agent LSTM network to obtain a deep feature map;

[0026] S104, using the improved gray wolf optimization algorithm to select the deep feature map to obtain the optimal feature map;

[0027] S105, inputting the optimal feature map into the classification model to obtain a fault detection result;

[0028] S106: Construct a loss function of the model based on the fault detection results, adjust the model parameters, and complete the model training when the loss function converges.

[0029] In this embodiment, data preprocessing includes: using high-precision vibration sensors to collect vibration signals of bearings under different operating conditions, and using filtering, denoising and other technologies to preprocess the original signals, effectively removing noise and interference components in the signals, improving the quality and stability of the signals, and providing a reliable data basis for subsequent feature extraction and analysis.

[0030] In this embodiment, feature extraction from the data involves manually extracting time and frequency domain features from the preprocessed signal. In the time domain, features such as skewness, kurtosis, peak value, impulse factor, shape factor, and crest factor are calculated. These features reflect the distribution and fluctuation of the signal. In the frequency domain, features such as ball spin frequency (BSF), outer ball pass frequency (BPFO), and inner ball pass frequency (BPFI) are extracted to characterize the bearing's operating status at different frequency components. These features serve as key indicators for rotating machinery fault diagnosis and provide an important basis for subsequent fault determination.

[0031] In this example, the extracted features are fed into an improved multi-agent LSTM network. The multi-agent system allocates multi-source data based on feature characteristics, enabling parallel processing to improve efficiency. The LSTM network optimizes cell state outputs using a hyperbolic tangent function to reduce oscillations. A combination of multi-layer stacking, fully connected layers, and dropout layers enhances feature capture and generalization capabilities. After partitioning the dataset and training with the Adam optimizer, each agent outputs feature information, which is aggregated to form a deep feature map, preparing for subsequent classification tasks.

[0032] The improved multi-agent LSTM network is used to process the input features, including:

[0033] 1. Build a multi-agent system (MAS): Assign the extracted multi-dimensional feature data to different agents for processing according to rules that consider feature correlation. The specific method is as follows:

[0034] In order to combine the feature data with high correlation and assign them to the same agent, we first calculate the Pearson correlation coefficient between different features. Let the two feature vectors X = {x1, x2, ..., x n} and Y={y1,y2,…,y n Where n is the number of samples. The calculation formula of Pearson correlation coefficient r is:

[0035]

[0036] in, is the mean of feature X, is the mean of feature Y. Based on this, for all m features F1, F2, ..., F m , calculate the Pearson correlation coefficient between each pair of features and construct an m×m correlation matrix R, where R ij Represents feature F i and F j Then, set a correlation threshold θ (e.g. θ = 0.8), traverse the correlation matrix R, and for R ij , if |R ij |≥θ, then the feature F i and F j Finally, the grouped features are assigned to the same agent. If there are k groups, each group is assigned to an agent, and each agent is responsible for processing a group of highly correlated features.

[0037] Each intelligent agent has independent computing resources and processing capabilities, and can process the assigned feature data in parallel, improving the processing efficiency and response speed of the system.

[0038] 2. Improve LSTM network structure

[0039] Hyperbolic tangent function processing cell state: Based on the traditional LSTM network, in order to solve the problems of prediction oscillation and gradient disappearance, the hyperbolic tangent function is added to its basic cell structure, such as Figure 3 The hyperbolic tangent function is added at the top of the figure. The hyperbolic tangent function maps the input value to the interval [-1, 1], making the output more stable, reducing the prediction oscillation, and improving the efficiency and accuracy of the model.

[0040] Multi-layer LSTM stacking: Two LSTM layers are used, each with 128 hidden units. The first LSTM layer receives input data and mines the data's basic features and long-term dependency patterns. The second LSTM layer, based on the output of the first layer, further refines complex and advanced time-dependent features, improving the model's ability to capture the complex characteristics of rotating machinery fault signals.

[0041] Fully connected layer and dropout layer: A fully connected layer is added between the two LSTM layers to integrate and transform the features output by the LSTM layers, promoting the fusion of feature information from different dimensions. A dropout layer is added after the fully connected layer, with the dropout rate set to 0.5. The dropout layer randomly discards some neuronal connections, reducing the co-adaptability between neurons, effectively preventing model overfitting and enhancing the model's generalization ability.

[0042] 3. Network Training and Feature Extraction: The preprocessed feature data is divided into training, validation, and test sets. The training set is used for model parameter learning, the validation set is used to adjust model hyperparameters such as the learning rate and batch size during training, and the test set is used to evaluate the performance of the trained model. Labeled rotating machinery fault data is used to train the improved multi-agent LSTM network. The training process utilizes the Adam optimizer with an adaptive learning rate, dynamically adjusting the learning rate based on the model's training progress. Initially, a higher learning rate allows the model to converge quickly to the optimal solution. As training progresses, the learning rate is gradually reduced, allowing the model to fine-tune parameters and improve training accuracy and stability. The batch size is set to 32, and the number of epochs is set to 50. Each epoch of training traverses all data in the training set and updates parameters. Each trained LSTM network outputs its own feature information, which reflects the feature extraction results of each agent on the data it is responsible for.

[0043] 4. Aggregate multi-agent information to form a deep feature map

[0044] Information Transmission and Synchronization: MAS establishes a communication mechanism between agents. After completing its own LSTM network feature extraction, each agent encapsulates the output feature information according to a preset protocol and format and transmits it via a network communication link to a central coordinator or a designated aggregation node. To ensure the accuracy and timeliness of information transmission, a reliable data transmission protocol, such as TCP, is adopted, with timeout retransmission and error checking mechanisms. Furthermore, to ensure the temporal synchronization of the feature information of each agent, a timestamp mechanism is introduced, allowing the aggregation node to accurately determine the time point corresponding to each feature information.

[0045] Feature dimension matching and alignment: Since the source and type of feature data processed by different agents may be different, the feature information they output may differ in dimension and format. Therefore, before aggregation, the feature information of each agent needs to be dimensionally matched and aligned. First, the dimensional structure of each agent's feature information is analyzed to determine a unified feature dimension standard. Then, for feature information with inconsistent dimensions, interpolation, dimensionality reduction, or dimensionality increase are used to process it so that its dimension matches the unified standard. For example, for the output of certain agents with higher feature dimensions, methods such as principal component analysis (PCA) can be used for dimensionality reduction; for feature information with lower dimensions, interpolation algorithms can be used to increase the dimension.

[0046] Feature fusion and weighted combination: After matching and aligning feature dimensions, the aggregation node fuses the feature information of each agent. Using a weighted combination approach, the feature information output by each agent is assigned a weight based on its importance and reliability in fault detection. Weights can be learned and optimized using training data or set based on expert experience. The specific fusion formula is:

[0047]

[0048] Among them, F total represents the final deep feature map, n represents the number of agents, w i represents the weight of the feature information of the i-th agent, F i Represents the feature information output by the i-th agent. Through this weighted combination method, the advantages of each agent can be fully utilized, and the expressive power of the deep feature map and the accuracy of fault diagnosis can be improved.

[0049] In this embodiment, the improved Grey Wolf Optimization algorithm is used to select deep feature maps. Rotating machinery fault feature selection can be viewed as a binary problem. The IBGWOV algorithm, combined with binary optimization concepts, efficiently selects features from the LSTM network output. This algorithm cleverly combines the advantages of the Grey Wolf Optimization (GWO) algorithm and the Particle Swarm Optimization (PSO) algorithm, while using the V-shaped transfer function (V2-TF) to transform the continuous solution into a binary feature selection scheme.

[0050] The improved gray wolf optimization algorithm is used to process the feature map, including:

[0051] 1. Initialization phase: First, set a series of key parameters of the IBGWOV algorithm. These include the maximum number of iterations T, which determines the number of search rounds of the algorithm; and the population size P, which is the number of individuals participating in the search. For the GWO algorithm, the parameter a is set to control the search range. Its initial value is set to 2 and decreases linearly with the increase in the number of iterations. The formula is: Where t is the current iteration number. At the same time, determine the two random coefficients A and C, and is a random vector in the range [0, 1]. For the PSO algorithm, set the acceleration factors c1 and c2, typically set to 2, and the inertia weight iw, which can vary linearly or nonlinearly during the search. Additionally, determine the relevant parameters of V2-TF, which is formulated as V2(x) = |tanh(x)| and is used to map continuous solutions to the interval [0, 1].

[0052] 2. GWO-PSO Collaborative Search Optimization Phase: In this phase, GWO and PSO are not simply executed sequentially, but rather collaborate and optimize together. Each "wolf" in the initial population is considered a potential feature selection solution, represented by a binary code. Each "wolf" can also be considered a particle in the PSO. Based on the social hierarchy of GWO, the optimal solution (α wolf), the second-best solution (β wolf), and the third-best solution (δ wolf) are determined, with the remaining solutions being ordinary solutions (ω wolf).

[0053] In each iteration, each wolf updates its position based on the distance between itself and its prey (optimal solution). The distance and updated position are calculated using the following formula.

[0054]

[0055] Here Represents the location of the prey, is the current position of the wolf. After updating the position, use V2-TF to transform the new position, that is, calculate And convert it into a binary feature selection scheme based on a set threshold (such as 0.5). If the value is greater than the threshold, the feature is selected (coded as 1), otherwise it is not selected (coded as 0). In this process, the GWO algorithm, with its unique social hierarchy and hunting mechanism, can quickly conduct a preliminary search in the solution space, providing the PSO algorithm with a certain quality initial search direction and range, guiding the PSO algorithm to more effectively find potential high-quality feature selection solutions.

[0056] Based on the updated position of GWO, the PSO algorithm comes into play. Each particle (i.e., the feature selection scheme represented by each wolf) updates the formula according to its own speed:

[0057] vl new =iw×vl old +c1×rand1×(pbest-x old )+c2×rand2×(gbest-x old )

[0058] Update speed and position. Among them, vl old is the old velocity of the particle, x old is the particle’s old position, pbest and gbest are the particle’s personal best historical position and global best position respectively, rand1 and rand2 are random numbers in the range [0,1]. new ) and converts it into a binary feature selection representation. The PSO algorithm leverages its own speed update mechanism, combines individual self-awareness with group information sharing, and, based on GWO exploration, fine-tunes each feature selection scheme locally to discover optimal feature combinations. This allows the overall search process to escape local optima while also conducting in-depth searches in promising areas.

[0059] In each iteration, GWO and PSO achieve information interaction and co-evolution through specific mechanisms:

[0060] PSO feeds back the best solution to GWO: If the particle of PSO algorithm finds a better position (corresponding to a better feature selection solution), for example, particle pi from X pi Update to And when the fitness is higher, After receiving it, GWO will compare it with the current ω wolf position. If it is better, it will replace the position of a certain ω wolf and guide GWO to search from a better position in the next iteration.

[0061] GWO passes level information to PSO: the position and fitness information of α wolf, β wolf, and ω wolf generated by GWO in each iteration will be passed to PSO. In the PSO speed update formula vl new =iw×vl old +c1×rand1×(pbest-x old )+c2×rand2×(gbest-x old ), the determination of gbest will refer to the α wolf position. If it is better, the particle will tend to move in that direction. And the particle's pbest may also be adjusted according to the better solution of GWO. If X α 、X β 、X γ If there is a better position than the current pbest and certain conditions are met, it will be updated to that position, allowing the particle to better utilize the global search results of GWO to optimize the search path.

[0062] 3. Comprehensive sorting and selection stage: This stage is the key link of the IBGWOV algorithm, which aims to select the optimal feature subset from all solutions obtained in the GWO and PSO stages and the initial population.

[0063] Constructing a comprehensive solution set: Combine the P solutions obtained in the GWO phase, the P solutions obtained in the PSO phase, and the P solutions of the initial population to form a comprehensive solution set containing 3P solutions. Each solution is a binary-coded feature selection scheme.

[0064] Define a fitness function: To evaluate the quality of each solution, define a fitness function. This function comprehensively considers the impact of feature subsets on classification accuracy and the number of features. The formula is as follows:

[0065]

[0066] Among them, Accuracy is the accuracy obtained by using the feature subset to classify on the training set, which can be evaluated by the random forest classifier; FeatureCount is the number of features selected in the feature subset; TotalFeatures is the total number of features; ω1 and ω2 are weight coefficients, and ω1+ω2=1, which can be adjusted according to actual needs. For example, when more emphasis is placed on classification accuracy, the value of ω1 can be increased.

[0067] Calculate the fitness value: For each solution in the comprehensive solution set, use the fitness function above to calculate its fitness value. By running the random forest classifier on the training set, the classification accuracy corresponding to each solution is obtained, and the fitness value is calculated based on the number of features.

[0068] Sorting and Selection: Sort the 3P solutions in the comprehensive solution set in descending order based on the calculated fitness values. A higher fitness value indicates a better feature subset for that solution. The solution with the highest fitness value is selected as the optimal feature subset for this iteration. If the fitness value of the optimal solution meets the preset termination condition (e.g., reaching a certain fitness threshold), the iteration stops and the feature subset corresponding to that solution is used as the final feature selection result; otherwise, the next iteration continues.

[0069] In this embodiment, the classification model uses a random forest classifier. The random forest classifier receives the features selected by the IBGWOV algorithm and classifies the input features by constructing multiple decision trees. Each decision tree is trained based on different self-sampling samples, and a subset of features is randomly selected at the node to minimize impurity (usually measured by Gini impurity, the formula is where pk is the proportion of samples in class k. Ultimately, the classification result is determined by majority voting, enabling accurate judgment of the rotating machinery fault type, including healthy, outer race fault, ball fault, inner race fault, combined fault, and general degradation states.

[0070] In one embodiment of the present invention, the present invention also includes a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, it implements any of the above-mentioned rotating machinery fault detection methods based on the improved gray wolf optimization and improved multi-agent LSTM algorithm.

[0071] Those skilled in the art will appreciate that all or part of the steps in the above-described method embodiments can be implemented using hardware associated with a computer program. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps in the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0072] A rotating machinery fault detection device based on improved gray wolf optimization and improved multi-agent LSTM algorithm comprises a processor and a memory; the memory is used to store computer programs; the processor is connected to the memory and is used to execute the computer programs stored in the memory, so that the rotating machinery fault detection device based on improved gray wolf optimization and improved multi-agent LSTM algorithm performs any of the above-mentioned rotating machinery fault detection methods based on improved gray wolf optimization and improved multi-agent LSTM algorithm.

[0073] Specifically, the memory includes various media that can store program codes, such as ROM, RAM, magnetic disk, USB flash drive, memory card or optical disk.

[0074] Preferably, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0075] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation plans of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A rotating machinery fault detection method based on improved grey wolf optimization and improved multi-agent LSTM algorithm, characterized in that: include: Acquiring data of rotating machinery to be detected; Preprocessing the rotating machinery data, inputting the preprocessed data into the rotating machinery fault detection model to obtain the detection results; Training a rotating machinery fault detection model involves: S101, obtaining a rotating machinery dataset, and preprocessing the data in the rotating machinery dataset to obtain a training set; S102, extracting features from the data in the training set; S103, input the extracted features into the improved multi-agent LSTM network to obtain a deep feature map; S104, using the improved gray wolf optimization algorithm to select the deep feature map to obtain the optimal feature map; S105, inputting the optimal feature map into the classification model to obtain a fault detection result; S106: Construct a loss function of the model based on the fault detection results, adjust the model parameters, and complete the model training when the loss function converges.

2. A rotating machinery fault detection method based on improved grey wolf optimization and improved multi-agent LSTM algorithm according to claim 1, characterized in that: Data preprocessing includes: using high-precision vibration sensors to collect vibration signals of bearings under different operating conditions; using filtering and denoising technology to process the vibration signals to obtain signals without noise and interference.

3. A rotating machinery fault detection method based on improved grey wolf optimization and improved multi-agent LSTM algorithm according to claim 1, characterized in that: Feature extraction of data includes: extracting time domain features and frequency domain features of the vibration signal; the time domain features include the signal's skewness, kurtosis, peak value, pulse factor, shape factor and crest factor; the frequency domain features include the ball spin frequency, outer circle ball passing frequency and inner circle ball passing frequency.

4. A rotating machinery fault detection method based on improved grey wolf optimization and improved multi-agent LSTM algorithm according to claim 1, characterized in that: Improve the multi-agent LSTM network to process the extracted features including: Step 1: Extract multi-dimensional features in time domain, frequency domain and time-frequency domain from the preprocessed multi-source data; Step 2: Calculate the Pearson correlation coefficient between different features, construct a correlation matrix, set a correlation threshold, group features with correlations higher than the threshold, and assign the grouped features to different agents; Step 3: Divide the preprocessed feature data into training set, validation set and test set; Step 4: Use the labeled rotating machinery fault data to train the improved multi-agent LSTM network; Step 5: Each trained agent LSTM network outputs its own feature information; Step 6: Summarize the feature information output by all agents to form a deep feature map.

5. A rotating machinery fault detection method based on improved grey wolf optimization and improved multi-agent LSTM algorithm according to claim 1, characterized in that: The improved gray wolf optimization algorithm is used to select deep feature maps, including: Step 1: Initialize parameters; the parameters include the maximum number of iterations T, the population size P, the search range parameter a, and two random numbers A and C; set the acceleration factors c1, c2, and the inertia weight w; Step 2: Treat each wolf in the initial population as a potential feature selection scheme, and use the wolves in the initial population as particles in PSO. Step 3: Perform a local search on the population to determine whether the searched particle is the optimal solution. If so, pass the new position to the wolf pack, otherwise execute step 5. Step 4: Determine whether the fitness of the new position is better than that of the wolf ω. If so, replace the wolf ω with the new position. Step 5: Iteratively update the wolf pack; Step 6: Determine the social hierarchy structure of the new wolf pack and pass the new social hierarchy structure information to the particle swarm; Step 7: Update the particle velocity and position in the particle swarm algorithm; Step 8: Determine whether the position of wolf α is better based on the updated particle velocity and position information. If so, the particle moves toward wolf α; otherwise, update normally. Step 9: Determine whether there is a position with a better level. If so, update pbest; Step 10: Determine whether the current number of iterations reaches the maximum number of iterations. If not, add 1 to the current number of iterations and return to step 3. If yes, execute step 11. Step 11: Merge all solution sets and calculate fitness; Step 12: Sort the solution set and determine whether the sorted solution set meets the termination condition. If so, output the optimal solution. If not, return to step 3.

6. A rotating machinery fault detection method based on improved grey wolf optimization and improved multi-agent LSTM algorithm according to claim 5, characterized in that: The expression of the fitness function is: Among them, ω1 and ω2 are weight coefficients, Accuracy is the accuracy obtained by classifying the feature subset on the training set, FeatureCount is the number of features selected in the feature subset, and TotalFeatures is the total number of features.

7. A rotating machinery fault detection method based on improved grey wolf optimization and improved multi-agent LSTM algorithm according to claim 1, characterized in that: The classification model uses random forest classifier.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is executed by a processor to implement the rotating machinery fault detection method based on the improved grey wolf optimization and improved multi-agent LSTM algorithm according to any one of claims 1 to 7.

9. A rotating machinery fault detection device based on improved grey wolf optimization and improved multi-agent LSTM algorithm, characterized in that: including processor and memory; The memory is used to store computer programs; the processor is connected to the memory and is used to execute the computer programs stored in the memory, so that the rotating machinery fault detection device based on the improved gray wolf optimization and improved multi-agent LSTM algorithm can execute any one of claims 1 to 7 of the rotating machinery fault detection method based on the improved gray wolf optimization and improved multi-agent LSTM algorithm.

Images