Double-branch feature fusion harmonic reducer fault diagnosis method and device

Abstract

The invention relates to a harmonic reducer fault diagnosis method based on double-branch feature fusion. The harmonic reducer fault diagnosis method comprises the following steps: synchronously acquiring multi-sensor data of a harmonic reducer acquired by vibration, acoustic emission and current sensors; processing the multi-sensor data through a one-dimensional convolutional neural network to obtain data in a two-dimensional matrix form; the data in the two-dimensional matrix form is input into a pre-trained model, and the processing process of the model is as follows: feature extraction is performed on the data in the two-dimensional matrix form through an MCWGraphKAN branch feature extraction module and a GBiMama branch feature extraction module to obtain a first feature and a second feature, the first feature and the second feature are fused and then mapped to a sample space through a full-connection layer, and the data in the two-dimensional matrix form are obtained; and fault classification is realized. The method achieves the fault diagnosis of the harmonic reducer through a deep learning method, and is used for solving the problems that a single sensor cannot fully express the operation state of the harmonic reducer, and the fault feature extraction of a single-branch network is insufficient.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning and fault diagnosis, and particularly relates to a double-branch feature fusion harmonic reducer fault diagnosis method and device. BACKGROUND

[0002] The existing method has certain limitations in performing fault diagnosis on the harmonic reducer, mainly in the following two aspects:

[0003] 1. Single sensor (such as vibration, acoustic emission, current) is easily affected by the deployment location and external environment when collecting signals, and different sensors have different sensitivity and response time to faults. Using a single transmission signal as the input of the fault diagnosis model is difficult to fully reflect the running state of the equipment, which may lead to inaccurate and uncertain diagnosis.

[0004] 2. When facing complex working conditions or slight fault signals, the typical fault diagnosis method of deep learning usually focuses on a single feature space, such as convolutional neural network (CNN) extracting local spatial features, Long Short-Term Memory (LSTM) network focusing on time sequence dependence, and graphical convolutional networks (GCN) focusing on topological association. It is difficult to capture multi-dimensional key information in fault data at the same time, which leads to insufficient fault feature extraction and affects the accuracy and robustness of the final diagnosis. SUMMARY

[0005] The purpose of the present application is to at least solve one of the deficiencies of the prior art, and provide a double-branch feature fusion harmonic reducer fault diagnosis method.

[0006] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0007] Specifically, a double-branch feature fusion harmonic reducer fault diagnosis method is proposed, which includes the following:

[0008] Synchronously acquiring multi-sensor data of the harmonic reducer collected by vibration, acoustic emission and current sensors;

[0009] Processing the multi-sensor data through a one-dimensional convolutional neural network to obtain data in the form of a two-dimensional matrix;

[0010] Inputting the data in the form of a two-dimensional matrix into a pre-trained model, and the processing process of the model is as follows,

[0011] respectively through the MCW_GraphKAN branch feature extraction module and the GBiMamba branch feature extraction module to obtain first features and second features;

[0012] feature extraction is performed on the first features and the second features to obtain first features and second features;

[0013] The first features and the second features are fused, and then mapped to a sample space through a fully connected layer to realize classification of the fault;

[0014] The MCW_GraphKAN branch feature extraction module constructs a weighted graph by taking the reciprocal of the normalized Euclidean distance as the edge weight on the basis of the GCN, and uses parallel graph convolution layers of different scales to extract node features. In the message passing process, KAN is used to replace MLP, and the GBiMamba branch feature extraction module is a bidirectional state space modeling structure.

[0015] Further, specifically, the process of processing the multi-sensor data through the one-dimensional convolutional neural network to obtain the data in the form of a two-dimensional matrix includes:

[0016] A one-dimensional convolutional neural network 1D-CNN is used to extract features from the input sequence. If there are N sensors for synchronous data acquisition, each sensor records a time series signal with a length of L, then the multi-sensor data will be stacked into a two-dimensional matrix form, represented as:

[0017] X=[x1,x2,…,x N ] T ∈R N×L ;

[0018] wherein x i ∈R T represents the time series signal of the i-th sensor.

[0019] Further, specifically, the process of extracting first features from the data in the form of a two-dimensional matrix through the MCW_GraphKAN branch feature extraction module includes:

[0020] C out one-dimensional convolution kernels are defined, each kernel has a size of K, each convolution kernel performs weighted convolution on all input channels, and slides along the time dimension with a step size S to obtain an output feature map. Subsequently, a max-pooling operation is used to further downsample the features in the time dimension to obtain a high-dimensional feature representation h i ∈R d of each input sample, which constitutes a feature matrix:

[0021] H=[h1,h2,…,h N ] T ∈RN×d ;

[0022] where N represents the number of samples, and d represents the feature dimension of each sample;

[0023] The feature vector of each sample extracted by the CNN is represented as a node in the feature map, and the feature similarity between two samples is calculated by the Euclidean distance to construct an adjacency matrix A. To avoid the calculation redundancy and noise propagation caused by the full connection graph, the adjacency matrix A is further processed to be sparse, and only the connection between each node and its first k most similar neighbor nodes, i.e. the first k smallest distance samples, is reserved to construct a k-NN graph:

[0024]

[0025] Here, D ij represents the Euclidean distance between the i-th sample and the j-th sample, normlize(·) represents a normalization function, is the set of the k nearest neighbor nodes of node i, and argmin{·} represents the index corresponding to the smallest k values;

[0026] where the adjacency matrix A is defined as:

[0027]

[0028] The edge weight matrix W is defined as:

[0029]

[0030] After the feature map is constructed, the feature map is input into a parallel graph convolution layer with multiple different scales, and the features are aggregated from different numbers of neighbors, i.e. two layers of graph convolution modules are used. Each layer of graph convolution uses a scale size of order K=1, 2, 3, respectively, to extract neighborhood features under multiple scales, and uses KAN to replace MLP, which is defined as:

[0031]

[0032] In the formula, H is the output matrix after multi-scale graph convolution; X is the input graph node feature matrix; is a normalized graph Laplacian matrix; T k (·) is the k-th Chebyshev polynomial; KAN k (×) represents the k-th corresponding nonlinear mapping; p represents the number of scales; K1,K2,L,K p is the scale size; k1,k2,L,k p is the order of the Chebyshev polynomial; [·] is a feature splicing operator that splices the output features of different scales;

[0033] The information of each stage of the aggregation is passed to the next layer, and the input feature of the second layer graph convolution is taken as the first feature.

[0034] Further, the method further comprises,

[0035] In the double-layer Chebyshev network in the two-layer graph convolution module, a residual connection mechanism is introduced to improve the feature expression capability of the graph neural network in the deep structure and suppress the over-smoothing phenomenon.

[0036] Further, specifically, the process of extracting the second feature from the data in the form of a two-dimensional matrix through the GBiMamba branch feature extraction module includes:

[0037] The data in the form of a two-dimensional matrix is divided into small blocks in the time direction, each small block is mapped to a high-dimensional feature space through linear transformation to obtain block embedding Patch Embeddings, and at the same time, position encoding is added to each block embedding to consider the spatial position of the sequence. The sequenced representation processed by the embedding layer is taken as input and transmitted to Mamba for processing. An inverse Mamba is added in the forward Mamba, and an adaptive fusion strategy based on a gating mechanism is used for context feature fusion. Specifically, the output features x f ,x b are respectively encoded by a tanh activation function, then the contribution of the two directions of features is determined by a sigmoid-activated gating weight, and the features of the two directions are fused by weighting:

[0038] h f =tanh(W f ·x f )

[0039] h b =tanh(W b ·x b )

[0040] z=σ(W z ·[x f ,x b ])

[0041] h=z·h f +(1-z)·h b ;

[0042] In the formula, x f , x b are forward and reverse output features respectively; W f , W bare the weight matrices for forward and backward computation respectively; tanh(x) is the hyperbolic tangent function, which mainly serves to compress the numerical range and enhance the non-linear expression; h f and h b are the hidden states for forward and backward respectively; [·] is the feature dimension splicing operator; W z is the weight matrix of the gating mechanism; s(x) is the sigmoid activation function, which is used to compress the result to the interval [0,1]; z is the gating coefficient, which is used to control the contribution degree of forward and backward information; h represents the output feature fused through the gating mechanism.

[0043] Further, specifically, the parameter matrix is shared between the forward and backward Mamba to avoid increasing the memory occupation.

[0044] Further, specifically, the training process of the model comprises:

[0045] The health state of the harmonic reducer under a single working condition is simulated, the health state includes a normal state and a typical fault state, and multi-sensor data of the harmonic reducer collected by vibration, acoustic emission and current sensors are synchronously acquired to form a data set;

[0046] The data set is expanded and maximum-minimum normalized to obtain a processed data set, and the processed data set is divided to obtain a training set, a verification set and a test set;

[0047] The training set is input into the model, a cross-entropy loss function is used as an optimization target, and errors between model output results and true results are minimized;

[0048] After the training is completed, a weight file with detection capability for harmonic reducer fault data is obtained;

[0049] The weight file is used for the test set, and the diagnostic rate of the model is observed;

[0050] The recognition effects of multi-channel data and single-channel data as input are compared, and accuracy, precision, recall and F1-score are used to verify the model performance.

[0051] The application also provides a harmonic reducer fault diagnosis device with double-branch feature fusion, comprising the following:

[0052] A data acquisition module is configured to synchronously acquire multi-sensor data of a harmonic reducer collected by vibration, acoustic emission and current sensors;

[0053] A data preprocessing module is configured to process the multi-sensor data through a one-dimensional convolutional neural network to obtain data in a two-dimensional matrix form.

[0054] A fault diagnosis module is configured to input the data in the two-dimensional matrix form into a pre-trained model, and the processing procedure of the model is as follows,

[0055] The data in the two-dimensional matrix form is respectively subjected to feature extraction through an MCW_GraphKAN branch feature extraction module and a GBiMamba branch feature extraction module to obtain first features and second features.

[0056] The first features and the second features are fused, and then mapped to a sample space through a full connection layer to realize classification of faults.

[0057] The first features and the second features are fused, and then mapped to a sample space through a full connection layer to realize classification of faults.

[0058] The MCW_GraphKAN branch feature extraction module constructs a weighted graph by taking the reciprocal of the normalized Euclidean distance as the edge weight on the basis of the GCN, and uses parallel graph convolution layers of different scales to extract node features, and uses KAN to replace MLP in the message passing process, and the GBiMamba branch feature extraction module is a bidirectional state space modeling structure.

[0059] The method and device have the following beneficial effects:

[0060] The method and device can capture the running state features of the harmonic reducer from different physical angles and enhance the perception ability of the model to abnormal states. BRIEF DESCRIPTION OF DRAWINGS

[0061] The above and other features of the present disclosure will become more apparent from the following detailed description of embodiments taken in conjunction with the accompanying drawings, in which like reference characters indicate the same or similar elements throughout the drawings, and in which:

[0062] The above and other features of the present disclosure will become more apparent from the following detailed description of embodiments taken in conjunction with the accompanying drawings, in which like reference characters indicate the same or similar elements throughout the drawings, and in which:Figure 1 The principle diagram of a harmonic reducer system to which a harmonic reducer fault diagnosis method with double-branch feature fusion according to the application is applied is shown.

[0063] Figure 2 The overall flow chart of the realization of the harmonic reducer fault diagnosis method with double-branch feature fusion according to the application is shown.

[0064] Figure 3 The MCW_GraphKAN branch structure diagram proposed by the application is shown.

[0065] Figure 4 The GBiMamba branch overall structure diagram proposed by the application is shown.

[0066] Figure 5 The internal structure diagram of the GBiMamba with fusion gate mechanism proposed by the application is shown.

[0067] Figure 6 a), b), c), d), e), f) in the table are respectively different model confusion matrix comparisons.

[0068] Figure 7 a), b), c), d), e), f) in the table are respectively different model tSNE graph comparisons. DETAILED DESCRIPTION

[0069] The concept, specific structure and generated technical effects of the application will be described clearly and completely in combination with the embodiments and the drawings, so as to fully understand the purpose, scheme and effects of the application. It should be noted that the embodiments in the application and the features in the embodiments can be combined with each other without conflict. The same reference signs used in the drawings indicate the same or similar parts.

[0070] Using a single sensor may not guarantee the accuracy and completeness of the information obtained due to the influence of the deployment location and external environment interference, and cannot reflect the overall state of the device. Different types of sensors provide complementary information, vibration is an inevitable dynamic response during the operation of a mechanical system, and through a vibration sensor, dynamic parameters such as vibration amplitude, frequency and phase during device operation can be captured, focusing on abnormal vibration characteristics of mechanical structures. The acoustic emission sensor mainly captures the elastic wave signals released by the micro-damage of the material, and is highly sensitive to crack initiation and propagation, friction and wear, and can capture early features before the device has obvious vibration abnormalities. The current signal is an important physical quantity that reflects the driving state of the motor, load changes and fault characteristics, and has the advantages of non-intrusive, strong real-time and rich information. Multi-sensor fusion can reduce the probability of misjudgment and omission through mutual verification and supplementation of multiple sensor information. Moreover, different sensors have different sensitivity and response time to faults, and some faults may only show weak changes in a certain sensor signal at the initial stage. Through multi-sensor fusion, fault signs can be detected earlier to achieve early diagnosis and diagnosis. Therefore, the present application adopts three different types of sensors, i.e. vibration, acoustic emission and current, to collect data. In order to obtain effective data samples, the vibration sensor and the acoustic emission sensor should be as close to the fault source as possible, and according to the structure of the speed reducer, the outer surface of the rigid wheel is determined as the sensor mounting surface. The current sensor adopts a perforated installation method to collect the U-phase current on the motor driver. Then the host computer realizes the synchronous collection of multiple sensors.

[0071] The graph convolution network breaks through the inherent constraints of the convolutional neural network (CNN) in non-Euclidean space data modeling, and its core is the effective aggregation of scattered nodes in the graph sample. Through hierarchical progressive graph convolution operations, the correlation information between nodes with different local structures is deeply mined. However, the traditional GCN is constructed based on an unweighted graph, considering that the importance of neighbor nodes is the same, ignoring the inherent semantic and correlation differences between nodes. Secondly, GCN can only aggregate a certain number of neighbor node feature information, limiting its ability to capture global features. In addition, GCN relies on a multi-layer perceptron (MLP) to perform linear transformation and nonlinear activation on node features to aggregate neighbor node information and update node representation, but the linear weight structure of MLP has problems of insufficient feature extraction and limited diagnostic precision when facing highly nonlinear distribution of complex fault features. The KAN (Kolmogorov-Arnold Networks) can realize dynamic and learnable activation patterns on the edges and nodes of the graph by replacing the linear weights of MLP with a spline parameterized univariate function. Therefore, the present application designs a multi-scale weighted graph Kolmogorov-Arnold network (MCW_GraphKAN) based on GCN:

[0072] ①The reciprocal of the normalized Euclidean distance is used as the edge weight to construct a weighted graph;

[0073] ②Node feature extraction is performed using parallel graph convolution layers of different scales;

[0074] ③In the message passing process, KAN is used to replace MLP, which can enhance the fitting ability of the model to nonlinear fault features while retaining the ability to learn graph structure features.

[0075] Although MCW_GraphKAN can extract local features of multiple sensors and mine the graph structure relationship between different samples, it cannot extract long-term dependencies and global context features in time series. Although LSTM and GRU models can extract time series dependency features, their training process cannot be processed in parallel, which can easily cause gradient vanishing or explosion problems. Although the Transformer supports parallelization and can extract global features, its quadratic complexity results in high computational resource consumption. Mamba is a new architecture based on a selective state space model (SSM), which not only supports parallel computation but also significantly reduces computational resources. BiMamba is a bidirectional state space modeling structure, and the present application selects BiMamba as another branch and adds a gate-based adaptive fusion strategy. The BiMamba branch with a gating mechanism (GBiMamba) can efficiently extract forward and backward dependencies in time series and improve the ability to perceive global features of sequences.

[0076] Since MCW_GraphKAN and GBiMamba can extract different types of features, the present application proposes a method for parallel extraction of different types of features of multiple sensor signals using two branches, and fuses the extracted different types of features. This method effectively learns the complementary information of multiple sensors and overcomes the limitations of single models in feature expression ability and diagnostic performance. The model flowchart is shown in Figure 2 .

[0077] Based on the above, in Embodiment 1, the present application proposes a double-branch feature fusion harmonic reducer fault diagnosis method, which includes the following:

[0078] Synchronously acquiring multiple sensor data of a harmonic reducer collected by vibration, acoustic emission, and current sensors;

[0079] Processing the multiple sensor data through a one-dimensional convolutional neural network to obtain data in the form of a two-dimensional matrix;

[0080] Inputting the data in the form of a two-dimensional matrix into a pre-trained model, and the processing process of the model is as follows,

[0081] respectively, through the MCW_GraphKAN branch feature extraction module and the GBiMamba branch feature extraction module to obtain first features and second features;

[0082] respectively, through the MCW_GraphKAN branch feature extraction module and the GBiMamba branch feature extraction module to obtain first features and second features;

[0083] The first features and the second features are fused, and then mapped to a sample space through a fully connected layer to realize classification of the fault;

[0084] The MCW_GraphKAN branch feature extraction module constructs a weighted graph by taking the reciprocal of the normalized Euclidean distance as the edge weight on the basis of the GCN, uses parallel graph convolution layers of different scales to extract node features, and uses KAN to replace MLP in the message passing process, and the GBiMamba branch feature extraction module is a bidirectional state space modeling structure.

[0085] In this embodiment 1, the multi-sensor signals are fused before model input, which can capture the operating state features of the harmonic reducer from different physical angles and enhance the perception ability of the model to abnormal states. In addition, the MCW_GraphKAN and

[0086] The double-branch feature fusion method of the GBiMamba fully excavates the complementary information of different types of features, thereby enhancing the representation ability and classification performance of the model to complex fault modes. The method proposed in the present application can highlight the advantages of multi-sensor fusion in fault diagnosis by comparing with other methods using a single sensor. And by comparing with other mainstream methods, it is shown that considering global features and graph structure features at the same time plays a certain role in promoting the fault diagnosis of the harmonic reducer.

[0087] In addition, based on the actual operating conditions of the harmonic reducer, a harmonic reducer test platform is built to facilitate the collection of fault signal data sets required for model training. In a preferred embodiment, the harmonic reducer test platform comprises a servo motor, a motor driver, a coupling, a harmonic reducer, a swing arm load, a PLC, vibration, acoustic emission, current sensors, a synchronous acquisition board card, and an upper computer. The system principle is as shown in Figure 1 .

[0088] Specifically, in application, the present application includes the following main steps,

[0089] I. Constructing a fault data set

[0090] 1.1 According to the common failure modes of the harmonic reducer of the industrial robot, the application simulates five kinds of health states of the harmonic reducer under a single working condition, including one normal state and four typical fault states, wherein the failure modes involved are: slight wear of the inner wall of the flexspline, serious wear of the inner wall of the flexspline, slight wear of the outer ring of the wave generator, and serious wear of the outer ring of the wave generator.

[0091] 1.2 The harmonic reducer is installed on the harmonic reducer test platform, the input speed of the servo motor is adjusted to 600 rpm, and the load block is added to 4 kg. A three-axis acceleration sensor is used to collect vibration signals in x, y and z directions; a closed-loop Hall current sensor is used to collect the current of the servo motor; acoustic emission signals are collected and collected in real time with vibration and current signals, a total of 5 channel signals. In order to ensure the quality of the collected signals, the sampling frequency f s is set to 32KHz, and the collection time is 1min.

[0092] II. Dual-branch fusion feature extraction network

[0093] 2.1 MCW_GraphKAN branch feature extraction module

[0094] ①In order to effectively extract the time sequence local features in the original multi-sensor signals, this paper first uses one-dimensional convolutional neural network (1D-CNN) to extract features from the input sequence. Specifically, assuming that there are N sensors for synchronous data collection, each sensor records a time sequence signal with length L, then the original multi-sensor data can be stacked into a two-dimensional matrix form, denoted as:

[0095] X=[x1,x2,…,x N ] T ∈R N×L ;

[0096] Where x i ∈R T represents the time sequence signal of the i-th sensor

[0097] Define C out one-dimensional convolution kernel, each kernel size is K, each convolution kernel is weighted convolution on all input channels, and is slid along the time dimension with a step size S to obtain the output feature map. Subsequently, a max-pooling operation is used to further downsample the features in the time dimension to obtain a high-dimensional feature representation h i ∈R d of each input sample, which forms a feature matrix:

[0098] H=[h1,h2,…,h N ] T ∈R N×d ;

[0099] Wherein, N represents the sample quantity, and d represents the feature dimension of each sample

[0100] The input is subjected to feature extraction through multi-layer convolution and pooling, so that high-level semantic features can be extracted layer by layer, and the recognition ability of the model to complex time sequence changes is enhanced.

[0101] ②The feature vector of each sample extracted by the CNN is represented as a node in the graph, and then the feature similarity between two samples is calculated through the Euclidean distance to construct an adjacency matrix A. In order to avoid the calculation redundancy and noise propagation caused by the full connection graph, the adjacency matrix A is further sparsified, only the connection between each node and its first k most similar neighbor nodes (i.e. the first k smallest distance samples) is reserved, and a k-NN graph is constructed:

[0102]

[0103] Here, D ij represents the Euclidean distance between the ith sample and the jth sample, normlize(·) represents a normalization function, is the set of the nearest k neighbor nodes of node i, and argmin{·} represents the index corresponding to the smallest k values.

[0104] Further, the adjacency matrix A is defined as:

[0105]

[0106] In the traditional graph convolution operation, an unweighted graph is used, and the importance of neighbors is considered to be the same. This processing method ignores the heterogeneity of different neighbor nodes in the feature propagation process. In the present application, the edge weight between two connected nodes is set to the reciprocal of the normalized Euclidean distance. The edge weight matrix W is defined as:

[0107]

[0108] ③After the above graph is constructed, the graph is input into a plurality of parallel graph convolution layers with different scales, and the features are aggregated from different numbers of neighbors. Specifically, the present application uses two layers of graph convolution modules, each layer of graph convolution adopts a scale size of order K=1, 2, 3, extracts neighborhood features under multiple scales, and uses KAN to replace MLP, which can be defined as:

[0109]

[0110] In the formula, H is the output matrix after multi-scale graph convolution; X is the input graph node feature matrix; is a normalized graph Laplacian matrix; T k (·) is the kth Chebyshev polynomial; and KANk (x) represents the nonlinear mapping corresponding to the k-th order; p represents the number of scales; K1, K2, L, K p is the size of the scale; k1, k2, L, k p is the order of Chebyshev polynomial; [·] is a feature concatenation operator, which concatenates the output features of different scales;

[0111] The information aggregated by each order is passed to the next layer, and the input features of the second layer graph convolution are taken as the output features of the first branch.

[0112] ④In the double-layer Chebyshev network, a residual connection mechanism is introduced to improve the feature expression ability of the graph neural network in the deep structure and suppress the over-smoothing phenomenon. The structure is as shown in Figure 3 .

[0113] 2.2 Feature extraction module based on GBiMamba branch

[0114] ①The original Mamba mostly processes one-dimensional embedded samples, while ViM (Vision Mamba) uses a block method to process two-dimensional images. Therefore, the present application learns from the ViM block idea, divides the original multi-sensor signal stacked into a two-dimensional matrix form in the time direction into small blocks, maps each small block to a high-dimensional feature space through linear transformation to obtain patch embeddings. At the same time, position encoding is added to each patch embedding to consider the spatial position of the sequence. The sequenced representation processed by the embedding layer is taken as input and transmitted to the Mamba for processing, as shown in Figure 4 . An inverse Mamba enhancement model is added to the forward Mamba to enhance the understanding ability of the context. In order to avoid increasing the memory occupancy, the parameter matrix is shared between the forward and inverse Mambas. In order to effectively fuse the context features, the present application adopts an adaptive fusion strategy based on a gating mechanism, as shown in Figure 5 . Specifically, first, the internal encoding of each output feature x f , x b is performed using the tanh activation function, and then the contribution of the two directions is determined by the sigmoid-activated gating weight, and the features of the two directions are fused by weighting:

[0115] h f =tanh(W f ·x f )

[0116] h b =tanh(W b ·x b )

[0117] z=σ(W z·[x f ,x b ])

[0118] h = z·h f +(1-z)·h b ;

[0119] In the formula, x f x b These are the forward and backward output features, respectively; W f W b These are the weight matrices for forward and backward computation, respectively; tanh(×) is the hyperbolic tangent function, whose main function is to compress the numerical range and enhance the nonlinear expression; h f and h b These represent the forward and backward hidden states, respectively; [·] is the feature dimension concatenation operator; W z represents the weight matrix of the gating; s(×) is the sigmoid activation function, used to compress the result to the [0,1] interval; z is the gating coefficient, used to control the contribution of forward and backward information; h represents the output feature after fusion through the gating mechanism.

[0120] ② The different types of features extracted from the two branches are weighted and averaged and fused, and then mapped to the sample space through a fully connected layer to achieve fault classification.

[0121] III. Online Training

[0122] 3.1 The dataset is expanded using a sliding window method. This invention selects 1024 data points as the unit length to truncate the original multi-sensor signals, dividing them according to a 50% overlap rate, resulting in 350 data sets per class. The dataset is then divided into training, validation, and test sets at 60%, 10%, and 30% ratios, resulting in a total of 1750 training sets, 175 validation sets, and 525 test sets. To avoid model dependence on signal amplitude and to facilitate faster network training convergence, max-min normalization is applied to the dataset.

[0123]

[0124] 3.2 Input the training set into the model and use the cross-entropy loss function as the optimization objective to minimize the error between the model output and the true result.

[0125] 3.3 After training is completed, a weight file with the ability to detect fault data of harmonic reducers is obtained.

[0126] 3.4 Use the weight file on the test set to check the model's diagnostic rate.

[0127] 3.5In order to verify the effectiveness of the multi-channel fusion method proposed in the application in fault identification compared with single channel, the application sets up a comparative experiment, taking the identification effect of multi-channel data and single channel data as input as a comparison, using accuracy (Accuracy), precision (Presicion), recall (Recall), F1 score (F1-score) to verify the model performance, which can systematically evaluate the advantages of multi-channel fusion strategy in improving feature expression ability and fault classification accuracy. As shown in Table 1 below.

[0128]

[0129] Table 1

[0130] 3.6In order to reflect the superiority of the method proposed in the application in the harmonic reducer fault diagnosis task, the model is compared with

[0131] MCNN-MSA-BiGRU, CoordAtt-LSTM, TST, ResNet50, MRF-GCN model, and the performance of the model is verified using the above-mentioned indexes, and the diagnosis effect is shown in Table 2 below,

[0132]

[0133] Table 2

[0134] The confusion matrix of different models is shown in Table 3. Figure 6 The tSNE of different models is shown in Table 4. Figure 7

[0135] Although the description of the application has been quite detailed and particularly described for several embodiments, it is not intended to be limited to any of these details or embodiments or any special embodiment, but should be considered as providing a broad possibility interpretation for the prior art by referring to the appended claims, so as to effectively cover the intended range of the application. In addition, the application is described above in the embodiments that the inventors can foresee, the purpose is to provide a useful description, and non-essential modifications to the application that have not yet been foreseen can still represent equivalent modifications of the application.

[0136] The above is only a preferred embodiment of the application, and the application is not limited to the above-mentioned embodiments, as long as the same means achieve the technical effect of the application, it should belong to the protection scope of the application. The technical solutions and / or embodiments within the protection scope of the application can have various modifications and changes.​

Claims

1. A dual-branch feature fusion harmonic reducer fault diagnosis method, characterized in that, The method comprises the following steps: Synchronously acquiring multi-sensor data of a harmonic reducer collected by vibration, acoustic emission and current sensors; Processing the multi-sensor data through a one-dimensional convolutional neural network to obtain data in the form of a two-dimensional matrix; Inputting the data in the form of a two-dimensional matrix into a pre-trained model, and the processing process of the model is as follows, Respectively extracting first features and second features from the data in the form of a two-dimensional matrix through MCW_GraphKAN branch feature extraction modules and GBiMamba branch feature extraction modules; Fusing the first features and the second features, and then mapping them to a sample space through a fully connected layer to realize classification of faults; The MCW_GraphKAN branch feature extraction module constructs a weighted graph by taking the reciprocal of the normalized Euclidean distance as the edge weight on the basis of GCN, uses parallel graph convolution layers of different scales to extract node features, and uses KAN to replace MLP in the message passing process, and the GBiMamba branch feature extraction module is a bidirectional state space modeling structure.

2. The dual branch feature fused harmonic reducer fault diagnostic method of claim 1, wherein, Specifically, the process of processing the multi-sensor data through a one-dimensional convolutional neural network to obtain data in the form of a two-dimensional matrix comprises: Using a one-dimensional convolutional neural network 1D-CNN to extract features from an input sequence, if there are N synchronously data-acquired sensors, each sensor records a time series signal with a length of L, then the multi-sensor data will be stacked into a two-dimensional matrix form, which is represented as: X = [x1, x2,..., x N ] T ∈ R N×L ; where x i ∈R T denotes the timing signal of the i-th sensor.

3. The dual branch feature fused harmonic reducer fault diagnostic method of claim 2, wherein, Specifically, the process of extracting first features from the data in the form of a two-dimensional matrix through MCW_GraphKAN branch feature extraction modules comprises: Definition C out one-dimensional convolution kernel, each kernel has a size of K, each convolution kernel performs weighted convolution on all input channels and slides along the time dimension with a step size of S to obtain an output feature map, then a max-pooling operation is performed to further down-sample the features in the time dimension to obtain a high-dimensional feature representation h i ∈R d , to form a feature matrix: H = [h1, h2,..., h N ] T ∈ R N×d ; Wherein, N represents the number of samples, and d represents the feature dimension of each sample; The feature vector of each sample extracted by the CNN is represented as a node in a feature map, and then the feature similarity between two samples is calculated through the Euclidean distance to construct an adjacency matrix A. In order to avoid the calculation redundancy and noise propagation caused by the fully connected graph, the adjacency matrix A is further processed to be sparse, and only the connection between each node and its first k most similar neighbor nodes is reserved, that is, the first k smallest distance samples are selected to construct a k-neighbor graph k-NN Graph: Here, D ij represents the Euclidean distance between the ith sample and the jth sample, normlize(·) represents a normalization function, is the set of the k nearest neighbor nodes of node i, and arg min{·} represents the index corresponding to the minimum k values. Wherein, the adjacency matrix A is defined as: The edge weight matrix W is defined as: After the feature map is constructed, the feature map is input into a plurality of parallel graph convolution layers with different scales, and the features are aggregated from different numbers of neighbors, that is, two layers of graph convolution modules are used, each layer of graph convolution adopts a scale size with an order of K=1, 2, 3, extracts neighborhood features under multiple scales, and uses KAN to replace MLP, which is defined as: In the formula, H is the output matrix after multi-scale graph convolution; X is the input graph node feature matrix; is a normalized graph Laplacian matrix; T k (·) is the k-th order Chebyshev polynomial; KAN k (×) represents the k-th order corresponding nonlinear mapping; p represents the number of scales; K1, K2, L, K p is the scale size; k1, k2, L, k p is the order of the Chebyshev polynomial; [·] is a feature splicing operator, which splices the output features of different scales; The information aggregated by each order is transmitted to the next layer, and the input features of the second layer of graph convolution are taken as the first features.

4. The dual branch feature fused harmonic reducer fault diagnostic method of claim 3, wherein, The method further comprises: Introducing a residual connection mechanism in the double Chebyshev network in the two-layer graph convolution module to improve the feature expression capability of the graph neural network in the deep structure and suppress the over-smoothing phenomenon.

5. The dual branch feature fused harmonic reducer fault diagnostic method of claim 1, wherein, Specifically, the process of feature extraction of the data in the form of a two-dimensional matrix through the GBiMamba branch feature extraction module to obtain second features includes: The data in the form of a two-dimensional matrix is divided into small blocks in the time direction, each small block is mapped to a high-dimensional feature space through a linear transformation to obtain patch embeddings, and at the same time, position encoding is added to each block embedding to consider the spatial position of the sequence. The sequence representation processed by the embedding layer is input into the Mamba for processing. A reverse Mamba is added in the forward Mamba, and an adaptive fusion strategy based on a gating mechanism is used for context feature fusion. Specifically, first, the output features x f , x b are respectively internally encoded by using a tanh activation function, then the contribution of the two directions is determined by using a sigmoid-activated gating weight, and the features of the two directions are fused by weighting. h f = tanh(W f · x f ) h b = tanh(W b · x b ) z = σ(W z · [x f , x b ]) h = z - h f + (1 - z) - h b ; where x f and x b are forward and backward output features, respectively; W f and W b are weight matrices for forward and backward computation, respectively; tanh(x) is a hyperbolic tangent function to compress the value range and enhance the non-linear representation; h f and h b represent the hidden states for forward and backward, respectively; [·] is a feature dimension concatenation operator; W z is a weight matrix for gating; s(x) is a sigmoid activation function to compress the result to the interval [0, 1]; z is a gating coefficient to control the contribution degree of forward and backward information; h represents the output feature fused through the gating mechanism.

6. The dual branch feature fused harmonic reducer fault diagnostic method of claim 5, wherein, Specifically, the parameter matrix is shared between the forward and reverse Mamba to avoid increasing the memory occupancy.

7. The dual branch feature fused harmonic reducer fault diagnostic method of claim 1, wherein, Specifically, the training process of the model includes: Simulate the health state of the harmonic reducer under a single working condition, the health state including a normal state and a typical fault state, and synchronously acquire multi-sensor data of the harmonic reducer collected by vibration, acoustic emission and current sensors to form a data set; The data set is expanded and maximum-minimum normalized to obtain a processed data set, and the processed data set is divided to obtain a training set, a validation set and a test set; The training set is input into the model, and a cross-entropy loss function is used as an optimization objective to minimize the error between the model output result and the true result; After the training is completed, a weight file with detection capability for harmonic reducer fault data is obtained; The weight file is used for the test set to view the diagnostic rate of the model; The recognition effects of multi-channel data and single-channel data as input are compared, and accuracy, precision, recall and F1-score are used to verify the model performance.

8. A dual-branch feature fusion harmonic reducer fault diagnosis device, characterized in that, It includes the following: A data acquisition module is configured to synchronously acquire multi-sensor data of a harmonic reducer collected by vibration, acoustic emission and current sensors; A data preprocessing module is configured to process the multi-sensor data through a one-dimensional convolutional neural network to obtain data in the form of a two-dimensional matrix; A fault diagnosis module is configured to input the data in the form of a two-dimensional matrix into a pre-trained model, and the processing process of the model is as follows, The data in the form of a two-dimensional matrix is respectively extracted through an MCW_GraphKAN branch feature extraction module and a GBiMamba branch feature extraction module to obtain first features and second features; The first features and the second features are fused, and then mapped to a sample space through a fully connected layer to realize classification of faults; The MCW_GraphKAN branch feature extraction module constructs a weighted graph by taking the reciprocal of the normalized Euclidean distance as the edge weight on the basis of the GCN, uses parallel graph convolution layers of different scales to extract node features, and uses KAN to replace MLP in the message passing process, and the GBiMamba branch feature extraction module is a bidirectional state space modeling structure.

Images