Fan gearbox wear state identification method and system based on neural network

Abstract

The invention discloses a method and a system for identifying the wear state of a fan gear box based on a neural network, and relates to the technical field of mechanical equipment state monitoring, and the method comprises the steps: combining sound wave features and text features related to the wear of the fan gear box, achieving the high-dimensional semantic alignment of two modal features through a CLIP-style comparison and alignment module, and carrying out the recognition of the wear state of the fan gear box; complementary information of multi-source data is fully utilized, and the accuracy and robustness of wear state recognition are remarkably improved. The characteristics of the sound wave signals are extracted by adopting high-level triangular descriptors, key sampling point information of a time domain and a frequency domain is fused, subtle changes of the sound wave signals under the abrasion working condition can be effectively captured, and noise interference under the complex working condition is resisted. A wear state recognition model is constructed based on a ViLT model, cross-modal information association is mined through a multi-layer self-attention mechanism and a cross attention mechanism, and precise modeling of the wear state is achieved.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment condition monitoring technology, and more specifically to a method and system for identifying the wear condition of a wind turbine gearbox based on a neural network. Background Technology

[0002] With the rapid development of the new energy industry, wind power, as an important component of clean energy, has seen its installed capacity and operational scale continuously expand. As the core transmission component of a wind turbine generator set, the gearbox plays a crucial role in converting the wind energy captured by the rotor into electrical energy. Its operating status directly affects the wind turbine's power generation efficiency, operational safety, and service life. During long-term high-load, variable-condition operation, gearboxes are highly susceptible to gear and bearing wear. Failure to promptly and accurately identify the wear condition and implement maintenance measures may lead to escalating faults, causing downtime and significant economic losses.

[0003] Currently, the identification of wear conditions in wind turbine gearboxes mainly relies on manual inspection and traditional sensor monitoring methods. Manual inspection depends on the experience and judgment of technicians, which suffers from low efficiency, strong subjectivity, and high missed detection rate, and is difficult to adapt to the routine monitoring needs of large-scale wind farms. Traditional sensor monitoring methods are mostly based on the analysis of single physical quantity data such as vibration and temperature, failing to make full use of the complementary information from multiple data sources, and lacking accuracy and robustness in identifying complex wear conditions.

[0004] With the development of acoustic monitoring and deep learning technologies, equipment status recognition methods based on multimodal data have gradually become a research hotspot. Multimodal methods, by fusing various data types such as sound waves and text, can provide more comprehensive equipment status information, laying the foundation for accurate recognition. However, existing multimodal wind turbine gearbox wear status recognition technologies suffer from the following problems: First, the extraction of key acoustic features related to gearbox wear is not accurate enough, and there is a lack of effective methods for describing acoustic signals under wear conditions; second, the multimodal data fusion mechanism is imperfect, failing to achieve deep semantic alignment between acoustic features and text features; and third, insufficient cross-modal information correlation mining during model training results in the recognition model's generalization ability and recognition accuracy failing to meet practical application requirements.

[0005] Therefore, how to propose a method and system for identifying the wear state of wind turbine gearboxes based on neural networks, and overcome the shortcomings of existing technologies, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for identifying the wear state of wind turbine gearboxes based on neural networks, to solve the problems of low accuracy, insufficient robustness, and poor multimodal data fusion in existing technologies, and to achieve accurate and efficient identification of the wear state of wind turbine gearboxes. To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying the wear condition of a wind turbine gearbox based on a neural network includes: Acquire acoustic signals during the operation of the wind turbine gearbox, and preprocess the acoustic signals to obtain time-domain and frequency-domain signals; The characteristic time-domain points in the time-domain signal are detected by a sliding window sampling method, and the characteristic frequency-domain points in the frequency-domain signal are detected by a power spectrum peak detection method. The characteristic time-domain points and characteristic frequency-domain points are used as sampling points. Based on the sampling points, the acoustic and textual features of the wear state of the wind turbine gearbox were extracted; After performing dimensional matching between text features and sound wave features, the input is fed into the trained CLIP-style contrast alignment module. In the high-dimensional semantic space, the text features and sound wave features are paired according to the maximum cosine similarity. The paired text features and sound wave features are then concatenated to obtain the fused features. After the fused features are converted into a tensor format adapted to the ViLT model, they are input into the pre-trained wear state recognition model to identify the wear state of the wind turbine gearbox.

[0007] Optionally, the extraction of acoustic and textual features of the wear state of the wind turbine gearbox based on sampling points includes: Local triangular features are extracted from each sampling point, and Fisher vector method is used to encode the local triangular features to obtain HTSD features. Principal component analysis is used to reduce the dimensionality of the HTSD features to obtain the acoustic features of the wear state of the wind turbine gearbox. The operating condition information corresponding to the acoustic signal is input into the trained VisualGPT model to obtain initial text features. The initial text features are then cleaned, and a word segmenter is used to convert the cleaned initial text features into a token sequence. The token sequence is then encoded using a text encoder to output the text features of the wear status of the wind turbine gearbox.

[0008] Optionally, the local triangle features include local triangle features of feature time-domain points and local triangle features of feature frequency-domain points, both of which include the area of ​​the triangle, the distance from the center of the triangle to the point, and the side length and included angle of the triangle.

[0009] Optionally, the training process of the CLIP-style contrast alignment module includes: Obtain the sample dataset, which includes sample text features and sample acoustic features; The sample text features and sample sound wave features are dimension-matched and mapped to a high-dimensional semantic space of the same dimension. Then, L2 normalization is performed. Within each batch, construct a sound wave-text similarity matrix to obtain N sound wave-text pairs, and calculate the bidirectional InfoNCE loss based on the N sound wave-text pairs; If the bidirectional InfoNCE loss is higher than a set threshold, the learnable parameters of the CLIP-style contrast alignment module are adjusted and iteratively trained until the bidirectional InfoNCE loss is lower than the set threshold, thus obtaining a trained CLIP-style contrast alignment module.

[0010] Optionally, the wear condition identification model is obtained by pre-training the ViLT model.

[0011] Optionally, training the ViLT model includes: Obtain a sample dataset, which includes fusion features corresponding to acoustic signals of different wear states of wind turbine gearboxes. After converting the fusion features into a tensor format adapted to the ViLT model, training samples are obtained. A ViLT model is constructed and trained using training samples to obtain a wear state recognition model.

[0012] Optionally, the construction of the ViLT model, which involves training the ViLT model using training samples to obtain a wear state recognition model, includes: Acquire training samples; input the training samples into the ViLT model, perform forward propagation on the training samples, and execute a multi-layer self-attention mechanism to capture the connection between each feature segment in the sound wave features and understand the contextual connection between each text word in the text features. After aligning the text words in the text features with the feature fragments in the sound wave features using the cross-attention mechanism, the loss value of the ViLT model is calculated. If the training stop condition is not met, the process proceeds to the step of inputting the training samples into the ViLT model after backpropagation based on the loss value. If the training stop condition is met, the trained ViLT model is used as the wear state recognition model. The training stop condition includes the number of iterations reaching a preset number and / or the loss value of the ViLT model being lower than a preset loss threshold.

[0013] Optionally, the connection between the various feature segments in the captured acoustic wave features includes: A Pooler layer is used to process the acoustic wave features to generate a fixed-dimensional feature vector. Then, a fully connected layer is used to perform a linear transformation on the feature vector to generate the relationship between the various feature segments in the acoustic wave features.

[0014] Optionally, the loss value can be calculated using the binary cross-entropy loss function, as shown in the formula: ; in, represents the binary cross-entropy loss value, used to measure the error between the prediction result of the wear condition recognition model and the true label. N represents the total number of training samples, corresponding to the number of acoustic-text fusion feature samples for different wear conditions of the wind turbine gearbox. This represents the true label of the i-th sample, with a value of 0 or 1. 1 indicates that the sample corresponds to a target wear state, and 0 indicates that the target wear state does not exist. This represents the probability that the model predicts the i-th sample has a target wear state, with a value range of [0,1], and is calculated by the output layer of the ViLT model.

[0015] Optionally, a neural network-based wind turbine gearbox wear condition identification system includes: Acquisition module: Used to acquire acoustic signals during the operation of the wind turbine gearbox; Preprocessing module: used to preprocess the acoustic signal to obtain time-domain and frequency-domain signals; Detection module: used to detect characteristic time-domain points in the time-domain signal using a sliding window sampling method, and to detect characteristic frequency-domain points in the frequency-domain signal using a power spectrum peak detection method, and to use the characteristic time-domain points and characteristic frequency-domain points as sampling points; Extraction module: used to extract acoustic and text features of the wear state of the wind turbine gearbox based on sampling points; The fusion module is used to input the text features and sound wave features into the trained CLIP-style contrast alignment module after dimensional matching. In the high-dimensional semantic space, the text features and sound wave features are paired according to the maximum cosine similarity. The paired text features and sound wave features are then concatenated to obtain the fused features. Recognition module: It is used to convert the fused features into a tensor format adapted to the ViLT model and input it into the pre-trained wear state recognition model to identify the wear state of the wind turbine gearbox.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for identifying the wear state of a wind turbine gearbox based on a neural network, which has the following beneficial effects: This invention proposes a method for identifying the wear state of a wind turbine gearbox based on a neural network, comprising: acquiring acoustic signals during the operation of the wind turbine gearbox; preprocessing the acoustic signals to obtain time-domain and frequency-domain signals; detecting feature time-domain points in the time-domain signals using a sliding window sampling method and feature frequency-domain points in the frequency-domain signals using a power spectrum peak detection method, and using the feature time-domain points and feature frequency-domain points as sampling points; extracting acoustic and text features of the wear state of the wind turbine gearbox based on the sampling points; inputting the text features and acoustic features into a trained CLIP-style comparison and alignment module after dimensional matching, pairing the text features and acoustic features in a high-dimensional semantic space by maximizing cosine similarity, and concatenating the paired text features and acoustic features to obtain fused features; converting the fused features into a tensor format adapted to the ViLT model and inputting it into a pre-trained wear state identification model to identify the wear state of the wind turbine gearbox. This invention combines acoustic and textual features related to wind turbine gearbox wear. A CLIP-style comparison alignment module achieves high-dimensional semantic alignment of the two modal features, fully utilizing complementary information from multi-source data to significantly improve the accuracy and robustness of wear state identification. A high-level triangular shape descriptor (HTSD) is used to extract acoustic signal features, fusing key sampling point information from the time and frequency domains. This effectively captures subtle changes in acoustic signals under wear conditions, combats noise interference under complex operating conditions, and maintains strong discriminative ability for wear state features. A wear state identification model is constructed based on the ViLT model, using multi-layer self-attention and cross-attention mechanisms to mine cross-modal information associations, achieving accurate modeling of wear states. Reasonable loss functions and stopping conditions are used during model training to ensure the model's generalization ability and recognition efficiency. This invention achieves automated and accurate identification of wind turbine gearbox wear states, eliminating reliance on manual experience, lowering the monitoring threshold, and helping maintenance personnel to promptly detect wear faults and take maintenance measures, reducing downtime losses and improving the operation and maintenance efficiency and economic benefits of wind farms. Attached Figure Description

[0017] 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 only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a schematic flowchart of a method for identifying the wear status of a wind turbine gearbox based on a neural network, provided by the present invention.

[0019] Figure 2The present invention provides a structural framework diagram of a wind turbine gearbox wear condition identification system based on neural networks. Detailed Implementation

[0020] 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, and 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.

[0021] This invention discloses a method for identifying the wear state of a wind turbine gearbox based on a neural network, such as... Figure 1 As shown, it includes: Acquire acoustic signals during the operation of the wind turbine gearbox, and preprocess the acoustic signals to obtain time-domain and frequency-domain signals; The characteristic time-domain points in the time-domain signal are detected by a sliding window sampling method, and the characteristic frequency-domain points in the frequency-domain signal are detected by a power spectrum peak detection method. The characteristic time-domain points and characteristic frequency-domain points are used as sampling points. Based on the sampling points, the acoustic and textual features of the wear state of the wind turbine gearbox were extracted; After performing dimensional matching between text features and sound wave features, the input is fed into the trained CLIP-style contrast alignment module. In the high-dimensional semantic space, the text features and sound wave features are paired according to the maximum cosine similarity. The paired text features and sound wave features are then concatenated to obtain the fused features. After the fused features are converted into a tensor format adapted to the ViLT model, they are input into the pre-trained wear state recognition model to identify the wear state of the wind turbine gearbox.

[0022] Furthermore, the extraction of acoustic and textual features of the wear state of the wind turbine gearbox based on sampling points includes: Local triangular features are extracted from each sampling point, and Fisher vector method is used to encode the local triangular features to obtain HTSD features. Principal component analysis is used to reduce the dimensionality of the HTSD features to obtain the acoustic features of the wear state of the wind turbine gearbox. The operating condition information corresponding to the acoustic signal is input into the trained VisualGPT model to obtain initial text features. The initial text features are cleaned to remove special symbols and format marks. The cleaned initial text features are converted into a token sequence using a word segmenter. The token sequence is encoded using a text encoder to output the text features of the wear status of the wind turbine gearbox.

[0023] Furthermore, the local triangle features include local triangle features at characteristic time-domain points and local triangle features at characteristic frequency-domain points, both of which include the area of ​​the triangle, the distance from the center of the triangle to the point, and the side length and included angle of the triangle.

[0024] Furthermore, the training process of the CLIP-style contrast alignment module includes: Obtain the sample dataset, which includes sample text features and sample acoustic features; The sample text features and sample sound wave features are dimension-matched and mapped to a high-dimensional semantic space of the same dimension. Then, L2 normalization is performed. Within each batch, construct a sound wave-text similarity matrix to obtain N sound wave-text pairs, and calculate the bidirectional InfoNCE loss based on the N sound wave-text pairs; If the bidirectional InfoNCE loss is higher than a set threshold, the learnable parameters of the CLIP-style contrast alignment module are adjusted and iteratively trained until the bidirectional InfoNCE loss is lower than the set threshold, thus obtaining a trained CLIP-style contrast alignment module.

[0025] Furthermore, the wear condition recognition model is obtained by pre-training the ViLT model.

[0026] Furthermore, training the ViLT model includes: A sample dataset is obtained, which includes fused features corresponding to acoustic signals of different wear states of wind turbine gearboxes. The fused features are then converted into a tensor format adapted to the ViLT model to obtain training samples. The tensor format has a dimension of (batch_size, feature_dim), where batch_size is the number of samples trained each time, and feature_dim is the dimension of the fused features.

[0027] A ViLT model is constructed and trained using training samples to obtain a wear state recognition model.

[0028] Specifically, the ViLT model includes an input layer, an embedding layer, a multi-layer Transformer encoder, and an output layer. The input layer receives fused features and transforms them into tensors that the model can process. The embedding layer includes position embedding and modal embedding. Position embedding adds positional encoding to the concatenated feature sequence, and modal embedding adds modal labels to the acoustic features and text features, respectively. The multi-layer Transformer encoder consists of N stacked Transformer blocks, each layer containing a multi-head self-attention mechanism and a feedforward network, used to capture the correlation between frequency bands in the acoustic features and the contextual relationship between words in the text features. The output layer includes a classification head and a loss calculation module. The classification head maps high-dimensional features to the wear state category space, and the loss calculation module calculates the error between the model's prediction and the true label.

[0029] Furthermore, the construction of the ViLT model, which involves training the ViLT model using training samples to obtain a wear state recognition model, includes: Acquire training samples; input the training samples into the ViLT model, perform forward propagation on the training samples, and execute a multi-layer self-attention mechanism to capture the connection between each feature segment in the sound wave features and understand the contextual connection between each text word in the text features. After aligning the text words in the text features with the feature fragments in the sound wave features using the cross-attention mechanism, the loss value of the ViLT model is calculated. If the training stop condition is not met, the process proceeds to the step of inputting the training samples into the ViLT model after backpropagation based on the loss value. If the training stop condition is met, the trained ViLT model is used as the wear state recognition model. The training stop condition includes the number of iterations reaching a preset number and / or the loss value of the ViLT model being lower than a preset loss threshold.

[0030] Furthermore, the connection between the various feature segments in the captured acoustic wave features includes: A Pooler layer is used to process the acoustic wave features to generate a fixed-dimensional feature vector. Then, a fully connected layer is used to perform a linear transformation on the feature vector to generate the relationship between the various feature segments in the acoustic wave features.

[0031] Furthermore, the loss value is calculated using the binary cross-entropy loss function, as shown in the formula: ; in, represents the binary cross-entropy loss value, used to measure the error between the prediction result of the wear condition recognition model and the true label. N represents the total number of training samples, corresponding to the number of acoustic-text fusion feature samples for different wear conditions of the wind turbine gearbox. This represents the true label of the i-th sample, with a value of 0 or 1. 1 indicates that the sample corresponds to a target wear state, and 0 indicates that the target wear state does not exist. This represents the probability that the model predicts the i-th sample has a target wear state, with a value range of [0,1], and is calculated by the output layer of the ViLT model.

[0032] In a specific implementation, an incremental optimization of a neural network-based method for identifying the wear state of a wind turbine gearbox is performed, and the specific steps are as follows: Step 1: Data Acquisition and Incremental Task Division: Collect operating acoustic signals and corresponding working condition information of wind turbine gearboxes under different wear conditions, including known wear types and potential new wear types, to construct a dataset; divide the dataset into multiple stage category sets according to wear type to form an incremental task sequence; The incremental task partitioning methods include: dividing the dataset equally into multiple incremental tasks according to wear type, or first taking half of the total number of wear types as the initial task, and then equally dividing the remaining wear types into multiple subsequent incremental tasks.

[0033] Step 2: Generate synthetic samples: Using a hybrid playback algorithm, the acoustic signal samples from the current incremental task are interpolated with randomly selected samples uniformly sampled from the old wear type data buffer to generate synthetic samples; the formula for generating synthetic samples is: ; in, These are the interpolation coefficients. For the synthetic sample, This indicates the sample selected from the old wear type buffer M through uniform sampling. This represents the acoustic signal sample for the current incremental task.

[0034] Step 3: Generation of acoustic wave features: The synthesized sample and the original acoustic signal sample are preprocessed to obtain time-domain and frequency-domain signals, respectively. A sliding window sampling method is used to detect characteristic time-domain points in the time-domain signal, and a power spectrum peak detection method is used to detect characteristic frequency-domain points in the frequency-domain signal. These characteristic time-domain points and characteristic frequency-domain points are used as sampling points. Local triangular features of each sampling point are extracted, and the Fisher vector method is used to encode these local triangular features to obtain high-level triangle shape descriptor (HTSD) features. The local triangle features include local triangle features at feature time-domain points and local triangle features at feature frequency-domain points, both of which include the area of ​​the triangle, the distance from the center of the triangle to the point, and the side length and included angle of the triangle. Principal component analysis (PCA) is used to reduce the dimensionality of the HTSD features to obtain the acoustic features of gearbox wear.

[0035] Step 4: Text Feature Generation and Feature Fusion Construction: The operating condition information corresponding to the acoustic signal is input into the trained VisualGPT model to obtain initial text features. The initial text features are cleaned to remove special symbols and format markers. A word segmenter is used to convert the cleaned initial text features into a token sequence. A text encoder is used to encode the token sequence to output the text features of the wear state of the wind turbine gearbox. After dimensional matching between the text features and the acoustic features, the data is input into a trained CLIP-style comparison and alignment module. In the high-dimensional semantic space, the text features and acoustic features are paired according to the principle of maximizing cosine similarity. The paired text features and acoustic features are then concatenated to obtain fused features.

[0036] Step 5: Training the unsupervised incremental model based on SimSiam: Construct a SimSiam model, which includes two identical neural network branches that share parameters. Each branch includes a feature extraction module, a projection head, and a prediction head. The fused features are subjected to two different data augmentation processes, including random rotation, adding noise, scaling or flipping, to obtain two randomly augmented views; Two random augmented views are input into the SimSiam model, and a stopping gradient mechanism is introduced, that is, during the training process, only the gradient is calculated for one branch, and the parameters of the other branch are not updated. The loss function is constructed based on the negative cosine similarity of the two network branches, and the SimSiam model is trained iteratively in multiple stages. The loss function formula is as follows: ; in, For cosine similarity, and These are the features obtained by mapping the two branches through the prediction head. and These are the branch vectors obtained from the feature extraction module for the two branches, respectively.

[0037] Step Six: Model Performance Evaluation A portion of data is randomly selected from the training set as the memory dataset. The memory dataset and the test set are input into the trained SimSiam model to extract features. The distance between the features of the memory dataset and each feature of the test set is calculated, including Euclidean distance, Manhattan distance, Chebyshev distance or Mahalanobis distance. The K closest samples are found and corresponding feature labels are generated. Based on the feature labels, the average accuracy, performance degradation rate and average forgetting rate of the model are calculated to complete the model performance evaluation.

[0038] Step 7: Wear condition identification: Combine the trained and performance-evaluated SimSiam model with the ViLT model to form a wear condition identification model that integrates unsupervised incremental learning. After feature extraction, text feature generation, and feature fusion of the acoustic signal of the wind turbine gearbox to be identified, it is input into the wear state identification model to identify the wear state of the wind turbine gearbox, including known wear types and newly added wear types.

[0039] Furthermore, the optimization training process for the wear condition recognition model includes: After completing the model training for the current incremental task, determine whether there is a next incremental task. If there is, store the wear type data of the current task into the cache according to a preset ratio, update the memory dataset, and then jump to step two to repeat the process of generating synthetic samples, extracting features, training the model, and evaluating performance. If there is no next incremental task, output the final wear status recognition model.

[0040] Specifically, the feature extraction module of the SimSiam model uses a ResNet18 residual network to obtain feature vectors with large receptive fields and small dimensions; the projection head uses a multilayer perceptron (MLP) structure to map features to a lower-dimensional space; and the prediction head uses an MLP structure to assist the network in learning more meaningful feature representations.

[0041] In a specific implementation, a wind turbine gearbox wear condition identification system based on a neural network, such as... Figure 2 As shown, it includes: Acquisition module: Used to acquire acoustic signals during the operation of the wind turbine gearbox; Preprocessing module: used to preprocess the acoustic signal to obtain time-domain and frequency-domain signals; Detection module: used to detect characteristic time-domain points in the time-domain signal using a sliding window sampling method, and to detect characteristic frequency-domain points in the frequency-domain signal using a power spectrum peak detection method, and to use the characteristic time-domain points and characteristic frequency-domain points as sampling points; Extraction module: used to extract acoustic and text features of the wear state of the wind turbine gearbox based on sampling points; The fusion module is used to input the text features and sound wave features into the trained CLIP-style contrast alignment module after dimensional matching. In the high-dimensional semantic space, the text features and sound wave features are paired according to the maximum cosine similarity. The paired text features and sound wave features are then concatenated to obtain the fused features. Recognition module: It is used to convert the fused features into a tensor format adapted to the ViLT model and input it into the pre-trained wear state recognition model to identify the wear state of the wind turbine gearbox.

[0042] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0043] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A neural network-based fan gearbox wear state identification method, characterized in that, The method comprises the following steps: acquiring an acoustic signal in the operation process of a fan gearbox, and preprocessing the acoustic signal to obtain a time domain signal and a frequency domain signal; detecting feature time domain points in the time domain signal by using a sliding window sampling method, and detecting feature frequency domain points in the frequency domain signal by using a power spectrum peak detection method, and taking the feature time domain points and the feature frequency domain points as sampling points; extracting acoustic features and text features of the wear state of the fan gearbox according to the sampling points; inputting the text features and the acoustic features into a trained CLIP-style contrast alignment module after dimension matching, pairing the text features and the acoustic features by maximizing the cosine similarity in a high-dimensional semantic space, splicing the paired text features and acoustic features to obtain fusion features; converting the fusion features into a tensor format suitable for a ViLT model, and inputting the fusion features into a pre-trained wear state recognition model to identify the wear state of the fan gearbox.

2. The neural network-based fan gearbox wear state recognition method according to claim 1, characterized in that, The method of extracting acoustic features and text features of the wear state of the fan gearbox according to the sampling points comprises the following steps: extracting local triangle features of each sampling point, encoding the local triangle features by using a Fisher vector method to obtain HTSD features, and performing dimension reduction processing on the HTSD features by using a principal component analysis method to obtain acoustic features of the wear state of the fan gearbox; inputting working condition information corresponding to the acoustic signal into a trained VisualGPT model to obtain initial text features, cleaning the initial text features, converting the cleaned initial text features into a token sequence by using a word segmenter, encoding the token sequence by using a text encoder, and outputting text features of the wear state of the fan gearbox.

3. The neural network-based fan gearbox wear state recognition method of claim 2, wherein, The local triangle features include local triangle features of the feature time domain points and local triangle features of the feature frequency domain points, and each of the local triangle features includes an area of a triangle, a distance from a center of the triangle to a point, and a length of a side of the triangle and an included angle.

4. The neural network-based fan gearbox wear state recognition method according to claim 1, characterized in that, The training process of the CLIP-style contrast alignment module comprises the following steps: obtaining a sample data set, wherein samples in the sample data set include sample text features and sample acoustic features; performing dimension matching on the sample text features and the sample acoustic features, and performing L2 normalization processing after mapping the sample text features and the sample acoustic features to a same-dimensional high-dimensional semantic space; constructing an acoustic-text similarity matrix in each batch to obtain N acoustic-text pairs, and calculating a bidirectional InfoNCE loss based on the N acoustic-text pairs; if the bidirectional InfoNCE loss is higher than a set threshold, adjusting learnable parameters of the CLIP-style contrast alignment module and iteratively training until the bidirectional InfoNCE loss is lower than the set threshold, and obtaining a trained CLIP-style contrast alignment module.

5. The neural network-based fan gearbox wear state recognition method according to claim 1, characterized in that, The wear state recognition model is obtained by pre-training a ViLT model.

6. The neural network-based fan gearbox wear state recognition method according to claim 5, characterized in that, The training of the ViLT model comprises the following steps: obtaining a sample data set, wherein the sample data includes fusion features corresponding to acoustic signals of different wear states of a fan gearbox, and obtaining training samples after converting the fusion features into a tensor format suitable for the ViLT model. The ViLT model is constructed, the ViLT model is trained through training samples, and a wear state recognition model is obtained.

7. The neural network-based fan gearbox wear state recognition method according to claim 6, characterized in that, The ViLT model is constructed, the ViLT model is trained through training samples, and a wear state recognition model is obtained. The training samples are obtained; the training samples are input into the ViLT model, after forward propagation of the training samples, a multi-layer self-attention mechanism is executed to capture the connection of each feature segment in the sound wave feature and understand the context connection of each text word in the text feature; After aligning the text words in the text feature with the feature segments in the sound wave feature using the cross-attention mechanism, the loss value of the ViLT model is calculated; If not, the ViLT model is backward propagated based on the loss value, and then the step of inputting the training samples into the ViLT model is jumped to; if yes, the trained ViLT model is used as the wear state recognition model; wherein the stop training condition includes that the number of iterative training reaches a preset number and / or the loss value of the ViLT model is lower than a preset loss threshold.

8. The neural network-based fan gearbox wear state recognition method according to claim 7, characterized in that, The connection of each feature segment in the sound wave feature is captured, including: The Pooler layer is used to process the sound wave feature to generate a fixed-dimensional feature vector, and then the feature vector is linearly transformed by the full connection layer to generate the connection of each feature segment in the sound wave feature.

9. The neural network-based fan gearbox wear state recognition method of claim 7, wherein, The loss value is calculated by using the binary cross-entropy loss function, and the formula is: ； wherein, represents a binary cross-entropy loss value, used to measure the error between the prediction result of the wear state recognition model and the true label, N represents the total number of training samples, and corresponds to the number of sound-wave-text fusion feature samples of different wear states of the fan gearbox, represents the true label of the i-th sample, and takes a value of 0 or 1, wherein 1 indicates that the sample corresponds to the target wear state, and 0 indicates that the target wear state does not exist, represents the probability that the model predicts the i-th sample to exist in the target wear state, and takes a value in the range of [0, 1], which is calculated by the output layer of the ViLT model. 10.A neural network-based fan gearbox wear state recognition system, characterized in that, It includes: The acquisition module is used to acquire the sound wave signal in the running process of the fan gearbox; The preprocessing module is used to preprocess the sound wave signal to obtain time domain signal and frequency domain signal; The detection module is used to detect the characteristic time domain points in the time domain signal by using the sliding window sampling method, and detect the characteristic frequency domain points in the frequency domain signal by using the power spectrum peak detection method, and the characteristic time domain points and characteristic frequency domain points are used as sampling points; The extraction module is used to extract the sound wave feature and text feature of the wear state of the fan gearbox according to the sampling points; The fusion module is used to input the text feature and the sound wave feature into the CLIP-style contrast alignment module after dimension matching, and pair the text feature and the sound wave feature in the high-dimensional semantic space according to the maximum cosine similarity, splice the paired text feature and sound wave feature to obtain the fusion feature; The recognition module is used to convert the fusion feature into a tensor format suitable for the ViLT model, and then input it into the pre-trained wear state recognition model to recognize the wear state of the fan gearbox.

Images