Harmonic reducer intelligent fault diagnosis method based on DRGAF-SwinTransform

Abstract

The invention belongs to the technical field of harmonic reducer fault diagnosis, and particularly relates to a harmonic reducer intelligent fault diagnosis method based on DRGAF-SwinTransform. Comprising the following steps: S100, collecting encoder signals of an input shaft and an output shaft of the harmonic reducer, and calculating to obtain a transmission error signal; s200, variational mode decomposition is carried out on the transmission error signal, and two mode components with the highest crest factor are selected from the mode components obtained through decomposition to be identified as fault mode components; s300, performing feature fusion on the two identified fault mode components by using a multi-mode Grubrum angle field, and converting the two identified fault mode components into an RGB image; and S400, constructing a DRGAF-SwinTransform diagnosis model, inputting the RGB image obtained in the step S300 into the model to train the model, obtaining probability values of different faults, and realizing intelligent fault diagnosis of the harmonic reducer based on the DRGAF-SwinTransform.

Description

Technical Field

[0001] The present invention belongs to the technical field of harmonic reducer fault diagnosis, and in particular relates to an intelligent fault diagnosis method for a harmonic reducer based on DRGAF-SwinTransformer. Background Art

[0002] Harmonic reducers are the core mechanisms of industrial robots, aerospace equipment, and medical machinery. They have complex operating conditions and require frequent loading, unloading, and switching actions. Therefore, harmonic reducers are prone to various faults, which in turn cause abnormal vibration and noise in the harmonic reducer. In severe cases, it may even cause equipment damage and result in serious economic losses. Therefore, it is necessary to conduct state detection and fault diagnosis research on harmonic reducers.

[0003] Transmission error is one of the important indicators to measure the performance of harmonic reducer. It represents the difference between the ideal motion and the actual motion of the gear system, that is, the angular difference between the ideal position and the actual position of the output gear. The input and output angles are collected by encoders installed on the input and output shafts, and then the transmission error is calculated using the transmission error calculation formula. At present, the lumped parameter model of the planetary gearbox has been successfully applied to obtain the fault TE response, revealing that different types of local faults of planetary gear teeth will produce different fault events. The harmonic reducer consists of a wave generator, a flexible bearing, a flexible wheel, and a rigid wheel. Its system structure is complex, and motion and force are transmitted through the deformation of the flexible wheel. In terms of transmission error, the error sources include processing errors, assembly errors, etc., which are similar to planetary gearboxes. When a harmonic reducer fails, the meshing of the faulty gear teeth will cause poor meshing. Compared with the healthy state, the transmission error amplitude will have a sudden change, and different fault events will occur. In addition, the technology of using residual transmission error for traditional gear fault diagnosis has been proposed, and it has been verified that transmission error can be used as an effective fault diagnosis information source. The encoder has high accuracy and is less interfered with, and can provide high-precision signals for fault diagnosis. Summary of the invention

[0004] The present invention aims to solve the problem that the current method for intelligent fault diagnosis of harmonic reducer transmission errors is relatively inadequate.

[0005] The present invention provides the following technical solution: a harmonic reducer intelligent fault diagnosis method based on DRGAF-SwinTransformer, comprising the following steps:

[0006] S100: Collect the input and output shaft encoder signals of the harmonic reducer and calculate the transmission error signal;

[0007] S200: performing variational modal decomposition on the transmission error signal, and selecting two modal components with the highest crest factors from the modal components obtained by the decomposition to identify them as fault modal components;

[0008] S300: Apply multimodal Gram angle field to perform feature fusion on the two identified fault modal components and convert them into RGB images;

[0009] S400: Construct a DRGAF-SwinTransformer diagnostic model, input the RGB image obtained in step S300 into the model for training, obtain probability values ​​of different faults, and implement intelligent fault diagnosis of harmonic reducers based on DRGAF-SwinTransformer.

[0010] Furthermore, in step S100, the transmission error signal TE(t) is calculated as follows:

[0011]

[0012] Among them, TE(t) is the transmission error signal; A in (t) is the input shaft encoder signal; A out (t) is the output shaft encoder signal; p is the reduction ratio of the harmonic reducer.

[0013] Further, in step S200, the variational modal decomposition decomposes the transmission error signal TE(t) into a series of intrinsic modal components by constructing and solving a constrained variational problem;

[0014] The steps of variational mode decomposition are:

[0015] S201: Construct a constrained variational problem, denoted as:

[0016]

[0017] Among them, u k is a series of decomposed eigenmode components, u k ={u1,u2,...,u k}; {ω k} is the center frequency of each eigenmode component, {ω k}={ω1,ω2,...,ω k}; δ(g) is the DiracDelta function; the symbol * represents the convolution operation; k is the number of extracted intrinsic mode components; To find partial derivatives; j is the imaginary unit; π is the circumference of a circle; e is a natural constant; t represents the time variable; s represents the bandwidth of each mode function;

[0018] S202: Introduce the quadratic penalty parameter α and the Lagrangian operator λ(t), transform the constructed constrained variational problem into an unconstrained variational problem, and obtain the augmented Lagrangian function, which is recorded as:

[0019]

[0020] Wherein, L is the augmented Lagrangian function; λ(t) is the Lagrangian multiplier; α is the penalty factor; f(t) is the transmission error signal TE(t) to be decomposed; In order to seek partial guidance;

[0021] S203: Iteratively update u using the multiplier alternating direction algorithm k ,ω k ,λ, and solve the saddle point of the unconstrained variational problem as the final solution.

[0022] Furthermore, in step S200, the calculation formula of the crest factor C is as follows:

[0023]

[0024] Where |x| peak is the peak value of the waveform; x rms is the effective value of the waveform; N c is the period of the waveform; x i is the instantaneous value of the waveform; max(|x|) is the maximum absolute value of the waveform.

[0025] Furthermore, in step S300, multi-modal Gram angle and field transformation is performed on the two fault modal components, including the following steps:

[0026] S301: Divide the two fault modal components into three equal parts, normalize them respectively, and convert them into polar coordinate points;

[0027] S302: The polar coordinates of the two fault modal components are respectively converted into Gram angle fields to obtain two matrices;

[0028] S303: the main diagonals of the matrices are set to 0 respectively, the matrix of the first fault modal component retains the upper triangular information, the matrix of the second fault modal component retains the lower triangular information, and the multi-modal Gram angle field is obtained by adding them together;

[0029] S304: Calculate the multi-modal Gram angle field by pairwise correspondence of the three data of each fault modal component, encode them as red, green and blue channels of the image, and merge them into an RGB image.

[0030] Furthermore, in step S301, the two fault modal components are divided into three equal parts, normalized respectively, and converted into polar coordinate points, and the expression is as follows:

[0031]

[0032] Among them, x is the first fault mode component; y is the second fault mode component, x ij ,yij is the sampling point on the original time series; is the sampling point after normalization; maxX i 、maxY i , minX i 、minY i are the maximum and minimum values ​​of the original time series respectively; t i is the timestamp; N is the constant factor for the regularized polar coordinate generation space; θ ij , is the phase angle; r ij is the polar coordinate radius;

[0033] The polar coordinates of the first fault modal component are <θ ij ,r ij >, the polar coordinates of the second fault mode component are

[0034] Furthermore, the specific process of step S302 is to convert the polar coordinates of the two fault modal components <θ ij ,r ij > and The inner product of the Gram angle field is performed respectively, where the polar coordinates of the first fault modal component <θ ij ,r ij The inner product formula of > is as follows:

[0035]

[0036] in, and are the normalized sampling points at the jth and mth timestamps respectively; θ ij and θ im They are and The corresponding phase angle;

[0037] Substitute the inner product result into the Gram matrix as an element of the Gram matrix to realize the Gram angle and field transformation. The Gram matrix formula is as follows:

[0038]

[0039] Where G is the Gram matrix;

[0040] The Gram angle and field formulas are as follows:

[0041] Polar coordinates of the second fault mode component The matrix after Gram angle and field transformation is as follows:

[0042]

[0043] Furthermore, the specific process of step S303 is to set the main diagonals of the two matrices to 0 respectively, the matrix of the first fault modal component retains the upper triangular information, and the matrix of the second fault modal component retains the lower triangular information, and the following formula is obtained:

[0044]

[0045] Adding the two matrices gives the multimodal Gram angle field:

[0046]

[0047] Furthermore, the specific process of step S304 is that the matrices of the three copies of data of the two fault modal components respectively calculate the multi-modal Gram angle fields to obtain DGAF1, DGAF2, and DGAF3, and the expressions are as follows:

[0048]

[0049] And encode them as the red, green and blue channels of the image, merge them into an RGB image, complete DRGAF encoding, and realize multimodal data fusion.

[0050] Further, the DRGAF-SwinTransformer diagnostic model in step S400 includes a PatchPartition block, a PatchMerging block, a SwinTransformer block, a fully connected layer, and an output layer;

[0051] The PatchPartition block divides the input RGB image into blocks to obtain local feature representation;

[0052] The PatchMerging block merges the local feature representations obtained by the PatchPartition block to obtain a higher-level global feature representation;

[0053] The self-attention mechanism of the SwinTransformer block enables the model to establish connections between features at different positions, and its multi-layer perceptron further mixes features. After multiple stages of processing, the feature map is subjected to global average pooling, and the feature values ​​of each channel are averaged to obtain a global feature vector and sent to the fully connected layer.

[0054] The global feature vector is mapped to the output space of the model through the fully connected layer;

[0055] The output layer uses the Softmax activation function to convert the output into a probability distribution, determine the fault type of the harmonic reducer, and realize intelligent fault diagnosis of the harmonic reducer.

[0056] Compared with the prior art, the advantages of the present invention are:

[0057] The present invention provides a harmonic reducer intelligent fault diagnosis method based on DRGAF-SwinTransformer, which uses variational mode decomposition (VMD) technology to effectively decompose the transmission error signal, extract the modal component most related to the fault, and combine the multimodal Gram angular field (DRGAF) image coding method to realize dual-modal signal fusion, and use the SwinTransformer model for intelligent fault diagnosis, which significantly improves the accuracy and robustness of harmonic reducer fault diagnosis. This method not only makes full use of the fault-sensitive information in the transmission error signal, but also realizes efficient identification of harmonic reducer faults under different working conditions through the powerful feature extraction capability of the deep learning model, which can be effectively applied to the fault diagnosis of harmonic reducers and improve the operating reliability of harmonic reducers. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a flow chart of the present invention;

[0059] Figure 2 It is a flow chart of data acquisition and raw signal processing;

[0060] Figure 3 It is a flow chart of signal DRGAF encoding processing;

[0061] Figure 4 It is the network structure of the DRGAF-SwinTransformer model;

[0062] Figure 5 This is the DRGAF encoding result diagram;

[0063] Figure 6 Confusion matrix for four different methods;

[0064] Figure 7 t-SNE scatter plots of four different methods;

[0065] Figure 8 It is the confusion matrix of the test set under different working conditions;

[0066] Fig. 9 It is the t-SNE scatter plot under different working conditions. DETAILED DESCRIPTION

[0067] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0068] like Figure 1 to Figure 3 As shown: A harmonic reducer intelligent fault diagnosis method based on DRGAF-SwinTransformer, comprising the following steps:

[0069] S100: Collect the input and output shaft encoder signals of the harmonic reducer and calculate the transmission error signal;

[0070] S200: performing variational modal decomposition on the transmission error signal, and selecting two modal components with the highest crest factors from the modal components obtained by the decomposition to identify them as fault modal components;

[0071] S300: Apply multimodal Gram angle field to perform feature fusion on the two identified fault modal components and convert them into RGB images;

[0072] S400: Construct a DRGAF-SwinTransformer diagnostic model, input the RGB image obtained in step S300 into the model for training, obtain probability values ​​of different faults, and implement intelligent fault diagnosis of harmonic reducers based on DRGAF-SwinTransformer.

[0073] The input shaft encoder and output shaft encoder on the harmonic reducer are both absolute encoders, and the encoder line count is 36,000.

[0074] In step S100, the transmission error signal TE(t) is calculated as follows:

[0075]

[0076] Among them, TE(t) is the transmission error signal; A in (t) is the input shaft encoder signal; A out (t) is the output shaft encoder signal; p is the reduction ratio of the harmonic reducer.

[0077] In step S200, the variational modal decomposition decomposes the transmission error signal TE(t) into a series of eigenmodal components by constructing and solving a constrained variational problem;

[0078] The steps of variational mode decomposition are:

[0079] S201: Construct a constrained variational problem, denoted as:

[0080]

[0081] Among them, u k is a series of decomposed eigenmode components, u k ={u1,u2,...,u k}; {ω k} is the center frequency of each eigenmode component, {ω k}={ω1,ω2,...,ω k}; δ(g) is the DiracDelta function; the symbol * represents the convolution operation; k is the number of extracted intrinsic mode components; To find partial derivatives; j is the imaginary unit; π is the circumference of a circle; e is a natural constant; t represents the time variable; s represents the bandwidth of each mode function;

[0082] S202: Introduce the quadratic penalty parameter α and the Lagrangian operator λ(t), transform the constructed constrained variational problem into an unconstrained variational problem, and obtain the augmented Lagrangian function, which is recorded as:

[0083]

[0084] Wherein, L is the augmented Lagrangian function; λ(t) is the Lagrangian multiplier; α is the penalty factor; f(t) is the transmission error signal TE(t) to be decomposed; In order to seek partial guidance;

[0085] S203: Iteratively update u using the multiplier alternating direction algorithm k ,ω k ,λ, and solve the saddle point of the unconstrained variational problem as the final solution.

[0086] In step S200, the calculation formula of the crest factor C is as follows:

[0087]

[0088] Where |x| peak is the peak value of the waveform; x rms is the effective value of the waveform; N c is the period of the waveform; x i is the instantaneous value of the waveform; max(|x|) is the maximum absolute value of the waveform.

[0089] When a local fault occurs in the harmonic reducer, the low-frequency component is affected within a change cycle, and the change in transmission error caused by the fault will be hidden in the transmission error of the entire cycle and is not obvious; the impact of the local fault on the high-frequency component will appear at each meshing of the faulty gear, which is specifically manifested as a periodic increase in the amplitude of the high-frequency component of the transmission error, the appearance of periodic peaks, and obvious fault characteristics. The crest factor is usually used to describe the transient characteristics of the signal. The larger its value, the more obvious the peak of the signal, and there may be pulses or mutations. When a fault occurs, the high-frequency component will produce periodic mutations at each meshing point of the faulty gear teeth, forming periodic peaks, so the crest factor can well identify the modal components most related to the fault.

[0090] Specifically, after obtaining the modal components through variational modal decomposition, the modal components most relevant to the fault can be effectively identified by comparing the crestfactor values ​​of each modal component.

[0091] In step S300, multi-modal Gram angle and field transformation is performed on two fault modal components, including the following steps:

[0092] S301: Divide the two fault modal components into three equal parts, normalize them respectively, and convert them into polar coordinate points;

[0093] The expression is as follows:

[0094]

[0095] Among them, x is the first fault mode component; y is the second fault mode component, x ij ,y ij is the sampling point on the original time series; is the sampling point after normalization; maxX i 、maxY i , minX i 、minY i are the maximum and minimum values ​​of the original time series respectively; t i is the timestamp; N is the constant factor for the regularized polar coordinate generation space; θ ij , is the phase angle; r ij is the polar coordinate radius;

[0096] The polar coordinates of the first fault modal component are <θ ij ,r ij >, the polar coordinates of the second fault mode component are

[0097] S302: The polar coordinates of the two fault modal components <θij ,r ij > and The inner product unique to the Gram angle field is performed respectively. This inner product is the inner product form with a penalty term defined by the Gram angle field. For data points that are close in time scale, the inner product can effectively suppress signal interference. Among them, the polar coordinates of the first fault modal component <θ ij ,r ij The inner product formula of > is as follows:

[0098]

[0099] in, and are the normalized sampling points at the jth and mth timestamps respectively; θ ij and θ im They are and The corresponding phase angle;

[0100] From the inner product formula and polar coordinate conversion formula, we can know that the inner product result is Only with Related, that is, only related to each time stamp The value is related, so the inner product result is Finally, X i Decide by X i From the extraction process, we know that X i There is an impulse amplitude that occurs at a certain frequency, and there is a gradual amplitude change near the impulse amplitude, so the inner product result is A similar amplitude change will also be formed;

[0101] This inner product can effectively suppress signal interference. Since the data points are close in time scale, when differential mode interference occurs, there is Δθ ij =-Δθ im , substituting into the inner product formula, we get:

[0102]

[0103] Among them, Δθ ij is θ ij The increment of Δθ im is θ im The increment of

[0104] Finally, the inner product result is substituted into the Gram matrix as an element of the Gram matrix to realize the Gram angle and field transformation. The Gram matrix formula is as follows:

[0105]

[0106] Where G is the Gram matrix;

[0107] The Gram angle and field formulas are as follows:

[0108] Depend on The calculation formula is to do the inner product process unique to the Gram angle field on the transformed vector. There will be a modal component X containing fault information i The corresponding amplitude changes, while the Gram angle and the various elements of the field are so The changes will eventually be reflected in the results of Gram's angle and field;

[0109] Polar coordinates of the second fault mode component The matrix after Gram angle and field transformation is as follows:

[0110]

[0111] S303: the main diagonals of the matrices are set to 0 respectively, the matrix of the first fault modal component retains the upper triangular information, the matrix of the second fault modal component retains the lower triangular information, and the multi-modal Gram angle field is obtained by adding them together;

[0112] The specific process is to set the main diagonals of the two matrices to 0 respectively, the matrix of the first fault mode component retains the upper triangular information, and the matrix of the second fault mode component retains the lower triangular information, and the following formula is obtained:

[0113]

[0114] Adding the two matrices gives the multimodal Gram angle field:

[0115]

[0116] Among them, the multi-modal Gram angle field inputs two modal data and uses the improved Gram angle field encoding method to achieve signal fusion. The data redundancy of the Gram angle field and the inability of the single-modal signal to fully describe the working state of the harmonic reducer, resulting in low fault diagnosis accuracy, are solved;

[0117] S304: Calculate the multi-modal Gram angle fields corresponding to the three data matrices of the two fault modal components respectively, encode them as red, green and blue channels of the image, and merge them into an RGB image;

[0118] The specific process is that the matrices of the three copies of the data of the two fault modal components respectively calculate the multi-modal Gram angle field to obtain DGAF1, DGAF2, and DGAF3. The expressions are as follows:

[0119]

[0120] And encode them as the red, green and blue channels of the image, merge them into an RGB image, complete DRGAF encoding, and realize multimodal data fusion.

[0121] The DRGAF method fuses the two modal data of the transmission error. The three DRGAF matrices contain the data of the two modalities respectively, realizing multimodal signal fusion. At the same time, they are encoded into three-channel information of RGB images, that is, the three-channel data of RGB contain information of the two modalities, further realizing information fusion and enhancing feature expression capabilities.

[0122] The DRGAF-SwinTransformer diagnostic model in step S400 includes a PatchPartition block, a PatchMerging block, a SwinTransformer block, a fully connected layer, and an output layer;

[0123] The PatchPartition block divides the input RGB image into blocks to obtain local feature representation;

[0124] The PatchMerging block merges the local feature representations obtained by the PatchPartition block to obtain a higher-level global feature representation;

[0125] The self-attention mechanism of the SwinTransformer block enables the model to establish connections between features at different positions, and its multi-layer perceptron further mixes features. After multiple stages of processing, the feature map is subjected to global average pooling, and the feature values ​​of each channel are averaged to obtain a global feature vector and sent to the fully connected layer.

[0126] The global feature vector is mapped to the output space of the model through the fully connected layer;

[0127] The output layer uses the Softmax activation function to convert the output into a probability distribution, determine the fault type of the harmonic reducer, and realize intelligent fault diagnosis of the harmonic reducer.

[0128] The specific steps are as follows:

[0129] S401: Decompose the input RGB image into multiple local blocks of the same size and overlap each other through Patch Partition, effectively capture the local information of the image, enable it to have the ability to interact with cross-block information, and avoid the loss of important information. The process expression is as follows:

[0130] PatchPartition:X={X1,X2,...,Xn}

[0131] Where X is the input RGB image, X irepresents the ith local block (Patch) after partitioning, and n is the total number of local blocks after partitioning; for example, for an image of size H×W, the Patch Partition process will partition the image into local blocks, each of which contains a P×P pixel area in the image;

[0132] S402: Merge every two adjacent rows and every two adjacent columns of the local feature map (i.e., the local feature representation obtained by Patch Partition) through Patch Merging to obtain a The new feature map of size, that is, the dimension of the feature will be doubled, and a higher-level global feature representation will be obtained; the process expression is as follows:

[0133] PatchMerging:Y={Merge(X i ,X i+1 ),Merge(X i+2 ,X i+3 ),...}

[0134] Among them, X is the set of local blocks obtained by Patch Partition, Y is the set of merged feature blocks, and Merge(·) is a merging operation, which is usually an element-by-element addition or averaging operation, used to merge adjacent local blocks into a new block. For example, if the size of the original feature map is 4P×4P, after a Patch Merging operation, the size of the new feature map will be 2P×2P, and each merged block contains the information of the original four blocks;

[0135] S403: The self-attention mechanism of the Swin Transformer block enables the model to establish connections between features at different positions. Its multi-layer perceptron further mixes features. The W-MSA and SW-MSA sub-blocks are applied alternately in two consecutive Swin Transformer blocks to capture local and global features and obtain a more comprehensive perspective. The calculation expression of two consecutive Swin Transformer blocks is as follows:

[0136]

[0137] in, Represent the output features of W-MSA and SW-MSA respectively; Z l , Z l+1 Represent the output features of the MLP modules in the first block and the second block respectively;

[0138] When calculating self-attention, Swin Transformer introduces a relative position deviation for each attention head, calculated as follows:

[0139]

[0140] Where Q, K, V are the learned weight matrices used to calculate the query, key, and value respectively; d represents the dimension of the query and key; B represents the bias matrix;

[0141] S404: After multiple stages are processed, the feature map is subjected to global average pooling, and the feature values ​​of each channel are averaged to obtain a global feature vector and sent to the fully connected layer. Let F be a feature map with a size of H×W×C, where H is the height of the feature map, W is the width, and C is the number of channels. The global average pooling operation GAP(F) averages the feature values ​​of each channel of F to obtain a global feature vector V. This process can be described by the following expression:

[0142]

[0143] Where, for each channel c (c ranges from 1 to c), Represents the value of the cth channel at position (i, j) in the feature map F, V c is the average value of the cth channel in the global feature vector V;

[0144] Therefore, the calculation of the global eigenvector V can be expressed as:

[0145] V=[V1,V2,...,V C ]

[0146] S405: The global feature vector of the fully connected layer is mapped to the output space of the model. The output layer applies the Softmax activation function to convert the output into a probability distribution, and the fault type of the harmonic reducer is determined to achieve intelligent fault diagnosis of the harmonic reducer. Let Z be the input feature vector of the fully connected layer, W be the weight matrix of the fully connected layer, b be the bias vector, and Y be the output of the fully connected layer. The calculation of the fully connected layer can be expressed as:

[0147] Y=WZ+b

[0148] Among them, the dimension of W is the number of output categories K multiplied by the length of the input feature vector N, that is, W∈R K×N ; The dimension of b is K, that is, b∈R K ; The dimension of Y is also K, that is, Y∈R K ;

[0149] Next, we apply the Softmax activation function to transform Y into a probability distribution The definition of the Softmax function is element-by-element calculation. For each element y in Y i , the Softmax function is calculated as follows:

[0150]

[0151] in, is the probability of the ith category in the probability distribution, Yes i The denominator is the sum of the exponents of all elements in Y;

[0152] Therefore, for the entire output vector Y, the Softmax function converts each element into a probability value to obtain the final probability distribution

[0153] so, Each element in Represents the probability that the input feature vector Z belongs to the i-th category.

[0154] Experimental case:

[0155] In order to verify the effectiveness of the method proposed in this paper, the HMCS-32-80-I model reducer was used as the experimental object to carry out method verification. The effectiveness of the proposed method was verified through two experimental cases: ① using ablation experiments to verify the effectiveness of the multimodal Gram angular field (DRGAF) method; ② using diagnostic accuracy as an indicator, the comparison methods are: Resnet 50 model, EfficientNet model, Visual Transformer model, vgg16 model.

[0156] Different working conditions were set during the data collection process: the load was set to 138N·m, and the speed was set to three different speeds of 600r / min, 400r / min and 200r / min; the speed was set to 600r / min, and the load was set to three different loads of 138N·m, 92N·m and 46N·m. The experimental data collected single fault data in 4 health states, including normal, steel wheel tooth breakage, flexible wheel tooth breakage, and flexible bearing wear. There are 4 single health states in total, and 100 samples are collected for each single health state, for a total of 400 samples. Subsequently, two composite fault data of steel wheel tooth breakage plus flexible bearing wear and flexible wheel tooth breakage plus flexible bearing wear were collected, for a total of 200 samples. The length of each sample is 2048.

[0157] All samples are decomposed to obtain the two modal components most relevant to the fault of each sample. They are respectively sampled by sliding window to obtain single modal sample data. Each single modal sample contains 1350 data points. Then, the two modal data at the same time are combined based on time to obtain sample data, that is, each sample data contains 2700 data points, and the number of samples for each health state is 100. Finally, DRGAF is used to encode them into RGB images. The encoded images are as follows: Figure 5 As shown in the figure, 80% of the RGB images were randomly selected as the training set, and the remaining 20% ​​were used as the test set. Five experiments were conducted under each working condition, and the classification accuracy was used as the evaluation criterion to quantitatively test the model performance.

[0158] Firstly, the ablation experiment is used to verify the superiority of the proposed VMD-multimodal Gram angular field (DRGAF) method. Based on the Swin Transformer model, three different image encoding methods and the VMD-multimodal Gram angular field (DRGAF) method are used for experiments, and the experimental results are analyzed and compared. GASF-0: input the original transmission error data, and the image encoding method uses the original Gram angular field; GASF-1: input a modal component most relevant to the fault, and the image encoding method uses the original Gram angular field; DRGAF-0: input a modal component most relevant to the fault and the original transmission error data, and the GAF matrix uses DRGAF; Method 4 is the proposed method.

[0159] In order to better demonstrate the experimental results of the proposed method for different fault types, the confusion matrix is ​​used to visualize the prediction results, with the horizontal axis representing the predicted label and the vertical axis representing the true label. Figure 6 (a) to (d) list the confusion matrices of the above four different methods. Comparing the four confusion matrices, it can be seen that with the addition of VMD and DRGAF, the accuracy of model prediction is improved, which shows the effectiveness of VMD and DRGAF. At the same time, compared with other fault types, the classification of flexible bearing wear (WGW) and rigid wheel tooth breakage + flexible bearing wear (FSBT&WGW) is more difficult and prone to misjudgment. The classification accuracy of the first three methods in these two categories has decreased to varying degrees, while the proposed method can still maintain a high level of diagnostic accuracy, proving that the proposed method has a strong classification ability.

[0160] In order to more intuitively illustrate the performance of the model, this paper uses t-Distributed Stochastic Neighbor Embedding (t-SNE) to project the features extracted from the last hidden layer of the model into two-dimensional space to visualize the data. Figure 7As shown, it can be seen that with the addition of VMD and DRGAF, the clustering effect of the features extracted by the model is better. The first three methods have unclear boundaries, while the proposed method has obvious classification boundaries and good clustering effect, which further proves the effectiveness of the proposed method.

[0161] Secondly, in order to verify the advantages of this method, the images generated by the VMD-DRGAF method were input into other image processing models and the results were compared. Four representative models were selected for comparative study. The comparative models include: Resnet 50 model, vgg16 model, EfficientNet model, Visual Transformer model. In order to avoid randomness of the results, the experiment randomly divided the training and test samples, and the average value of 5 random tests was calculated.

[0162] From the experimental results, all five models showed high accuracy, which once again verified the effectiveness of the proposed VMD-DRGAF feature extraction method. Comparing the prediction results of different models, the proposed diagnostic method showed the highest accuracy in all tested working conditions, with an average accuracy of 99.33%, which is higher than other methods and shows obvious advantages.

[0163] At the same time, for different working conditions, the accuracy of the proposed method is maintained above 98%. In addition, in order to better demonstrate the experimental results of the proposed method for different fault types, Figure 8 The confusion matrix of the test set under different working conditions is shown. It can be found that the recognition accuracy of each fault under different working conditions is maintained above 96%; Fig. 9 The t-SNE scatter plots under different working conditions are shown, and it can be observed that the clustering effect under each working condition is very good, the class boundaries are clear, and there are few misclassifications. From the perspective of accuracy and feature dimension, the proposed method can accurately classify various faults under different working conditions and has strong robustness and generalization ability.

[0164] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be 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 present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent fault diagnosis method for harmonic reducer based on DRGAF-SwinTransformer, characterized in that: The following steps are involved: S100: Collect the input and output shaft encoder signals of the harmonic reducer and calculate the transmission error signal; S200: performing variational modal decomposition on the transmission error signal, and selecting two modal components with the highest crest factors from the modal components obtained by the decomposition to identify them as fault modal components; S300: Apply multimodal Gram angle field to perform feature fusion on the two identified fault modal components and convert them into RGB images; S400: Construct a DRGAF-SwinTransformer diagnostic model, input the RGB image obtained in step S300 into the model for training, obtain probability values ​​of different faults, and implement intelligent fault diagnosis of harmonic reducers based on DRGAF-SwinTransformer.

2. According to claim 1, a harmonic reducer intelligent fault diagnosis method based on DRGAF-SwinTransformer is characterized by: In step S100, the transmission error signal TE(t) is calculated as follows: Among them, TE(t) is the transmission error signal; A in (t) is the input shaft encoder signal; A out (t) is the output shaft encoder signal; p is the reduction ratio of the harmonic reducer.

3. The intelligent fault diagnosis method for harmonic reducer based on DRGAF-SwinTransformer according to claim 2 is characterized by: In step S200, the variational modal decomposition decomposes the transmission error signal TE(t) into a series of eigenmodal components by constructing and solving a constrained variational problem; The steps of variational mode decomposition are: S201: Construct a constrained variational problem, denoted as: Among them, u k is a series of decomposed eigenmode components, u k ={u1,u2,...,u k }; {ω k } is the center frequency of each eigenmode component, {ω k }={ω1,ω2,...,ω k }; δ(g) is the DiracDelta function; the symbol * represents the convolution operation; k is the number of extracted intrinsic mode components; To find partial derivatives; j is the imaginary unit; π is the circumference of a circle; e is a natural constant; t represents the time variable; s represents the bandwidth of each mode function; S202: Introduce the quadratic penalty parameter α and the Lagrangian operator λ(t), transform the constructed constrained variational problem into an unconstrained variational problem, and obtain the augmented Lagrangian function, which is recorded as: Wherein, L is the augmented Lagrangian function; λ(t) is the Lagrangian multiplier; α is the penalty factor; f(t) is the transmission error signal TE(t) to be decomposed; In order to seek partial guidance; S203: Iteratively update u using the multiplier alternating direction algorithm k ,ω k ,λ, and solve the saddle point of the unconstrained variational problem as the final solution.

4. The intelligent fault diagnosis method for harmonic reducer based on DRGAF-SwinTransformer according to claim 3 is characterized by: In step S200, the calculation formula of the crest factor C is as follows: Where |x| peak is the peak value of the waveform; x rms is the effective value of the waveform; N c is the period of the waveform; x i is the instantaneous value of the waveform; max(|x|) is the maximum absolute value of the waveform.

5. The intelligent fault diagnosis method for harmonic reducer based on DRGAF-SwinTransformer according to claim 4 is characterized by: In step S300, multi-modal Gram angle and field transformation is performed on two fault modal components, including the following steps: S301: Divide the two fault modal components into three equal parts, normalize them respectively, and convert them into polar coordinate points; S302: The polar coordinates of the two fault modal components are respectively converted into Gram angle fields to obtain two matrices; S303: the main diagonals of the matrices are set to 0 respectively, the matrix of the first fault modal component retains the upper triangular information, the matrix of the second fault modal component retains the lower triangular information, and the multi-modal Gram angle field is obtained by adding them together; S304: Calculate the multi-modal Gram angle field by pairwise correspondence of the three data of each fault modal component, encode them as red, green and blue channels of the image, and merge them into an RGB image.

6. The intelligent fault diagnosis method for harmonic reducer based on DRGAF-SwinTransformer according to claim 5 is characterized by: In step S301, the two fault modal components are divided into three equal parts, normalized respectively, and converted into polar coordinate points. The expression is as follows: Among them, x is the first fault mode component; y is the second fault mode component, x ij ,y ij is the sampling point on the original time series; is the sampling point after normalization; max X i 、max Y i , min X i 、min Y i are the maximum and minimum values ​​of the original time series respectively; t i is the timestamp; N is the constant factor for generating the regularized polar coordinate space; is the phase angle; r ij is the polar coordinate radius; The polar coordinates of the first fault modal component are <θ ij ,r ij >, the polar coordinates of the second fault mode component are 7. The intelligent fault diagnosis method for harmonic reducer based on DRGAF-SwinTransformer according to claim 6 is characterized by: The specific process of step S302 is to convert the polar coordinates of the two fault modal components <θ ij ,r ij > and The inner product of the Gram angle field is performed respectively, where the polar coordinates of the first fault modal component <θ ij ,r ij The inner product formula of > is as follows: in, and are the normalized sampling points at the jth and mth timestamps respectively; θ ij and θ im They are and The corresponding phase angle; Substitute the inner product result into the Gram matrix as an element of the Gram matrix to realize the Gram angle and field transformation. The Gram matrix formula is as follows: Where G is the Gram matrix; The Gram angle and field formulas are as follows: Polar coordinates of the second fault mode component The matrix after Gram angle and field transformation is as follows:

8. The intelligent fault diagnosis method for harmonic reducer based on DRGAF-SwinTransformer according to claim 7 is characterized by: The specific process of step S303 is to set the main diagonals of the two matrices to 0 respectively, the matrix of the first fault modal component retains the upper triangular information, and the matrix of the second fault modal component retains the lower triangular information, and the following formula is obtained: Adding the two matrices gives the multimodal Gram angle field:

9. The intelligent fault diagnosis method for harmonic reducer based on DRGAF-SwinTransformer according to claim 8 is characterized by: The specific process of step S304 is to calculate the multi-modal Gram angle field corresponding to the matrices of the three copies of the two fault modal components respectively, and obtain DGAF1, DGAF2, and DGAF3, which are expressed as follows: And encode them as the red, green and blue channels of the image, merge them into an RGB image, complete DRGAF encoding, and realize multimodal data fusion.

10. The intelligent fault diagnosis method for harmonic reducer based on DRGAF-SwinTransformer according to claim 1 or 9, characterized in that: The DRGAF-SwinTransformer diagnostic model in step S400 includes a PatchPartition block, a PatchMerging block, a SwinTransformer block, a fully connected layer, and an output layer; The PatchPartition block divides the input RGB image into blocks to obtain local feature representation; The PatchMerging block merges the local feature representations obtained by the PatchPartition block to obtain a higher-level global feature representation; The self-attention mechanism of the SwinTransformer block enables the model to establish connections between features at different positions, and its multi-layer perceptron further mixes features; After multiple stages of processing, the feature map is subjected to global average pooling, the feature values ​​of each channel are averaged, a global feature vector is obtained and sent to the fully connected layer; The global feature vector is mapped to the output space of the model through the fully connected layer; The output layer uses the Softmax activation function to convert the output into a probability distribution, determine the fault type of the harmonic reducer, and realize intelligent fault diagnosis of the harmonic reducer.

Images