Diagnostic method for fault of rotary gear of reciprocating manipulator
Through the combination of the diffusion model and the wavelet packet noise reduction layer, the data scarcity of gear fault diagnosis in robot slewing motion is solved, high-precision gear fault diagnosis is achieved, and the robustness and diagnostic efficiency of the model are improved.
Patent Information
- Application Number
- CN202411299127.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-18
- Publication Date
- 2024-12-03
AI Technical Summary
The prior art is difficult to effectively diagnose gear failures during the robot rotation movement, especially under variable load conditions, data is scarce and labels are small, resulting in difficulty in diagnosis.
The diffusion model is used to expand the labeled samples under no-load conditions, and combine the one-dimensional convolution network and the wavelet packet noise reduction layer to extract common features under different operating conditions through pseudo-label calibration and maximum mean square deviation adaptation mechanism. , realize the classification of gear failures.
Through sample expansion and wavelet packet noise reduction processing, the training sample size is increased, the model's robustness and diagnostic accuracy are improved, and high-precision diagnosis of gear failures under different operating conditions is achieved.
Smart Images

Figure CN119064000A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of fault diagnosis during the rotary motion of a manipulator, and in particular to a method for diagnosing a rotary gear fault of a reciprocating manipulator. Background Art
[0002] Manipulator slewing devices are widely used in industrial automation and robotics to achieve rotational movement and multi-degree-of-freedom operation of the manipulator. The device is driven by gear transmission, servo motor or stepper motor, enabling the manipulator to move in different directions. Typical applications include welding, spraying, handling, assembly and other processes that require flexible and multi-angle operation. The slewing device not only improves production efficiency, but also significantly improves operational accuracy and consistency, meeting the needs of modern industry for efficient and precision manufacturing. In addition, manipulator slewing devices are gradually being used in the fields of medical care, logistics, agriculture, etc., helping to achieve automation and intelligent transformation.
[0003] Gears are one of the important transmission components in industrial robots and automation equipment. Their main function is to transmit power and motion through the meshing of gears to achieve precise positioning and movement of the manipulator. The manipulator's rotary device frequently and quickly switches between loaded and unloaded conditions, so the device's moment of inertia and friction will change significantly, which makes the gears very prone to failure. Current methods for detecting gear faults are mostly based on the same working condition, that is, the same distribution of data for analysis. This makes it impossible to achieve effective diagnosis for data under different distributions when facing variable load conditions. At the same time, the generation and discovery of faults during the manipulator's rotary motion are somewhat random, so it is impossible to obtain sufficient label data. The amount of data that can be obtained is relatively scarce and the labels are few, which also brings difficulties to diagnosis. Summary of the invention
[0004] The object of the present invention is to provide a method for diagnosing a fault of a rotary gear of a reciprocating manipulator, so as to overcome the deficiencies mentioned in the background technology.
[0005] The technical solution to achieve the purpose of the present invention is:
[0006] A method for diagnosing a reciprocating motion manipulator rotary gear fault comprises:
[0007] Step 1, collecting vibration signals of the reciprocating manipulator under different gear faults in no-load and loaded working conditions;
[0008] Step 2: Expand the labeled samples under no-load conditions through the diffusion model, and combine the expanded samples with the unlabeled samples under no-load conditions and the unlabeled samples under loaded conditions to form a data set;
[0009] Step 3: Build a one-dimensional convolutional network model, and add a wavelet packet denoising layer to the convolutional network layer to perform denoising on the signal in the input network;
[0010] Step 4: Input the no-load and loaded working condition data into the convolutional network for feature extraction at the same time, use the labeled no-load samples to guide the model training, and make predictions for the unlabeled samples. Pseudo-labels are given to the unlabeled no-load samples and loaded samples whose prediction probabilities are higher than the set threshold. Then, the obtained pseudo-label samples are added to the model training, and the pseudo-label calibration of all unlabeled samples is realized in this way.
[0011] In step 5, a new joint distribution adaptive mechanism with the maximum mean square error as the difference metric is used to achieve distribution alignment of data under two working conditions, helping the model to extract common features of the same type of gear faults under different working conditions, and finally classify the gear faults.
[0012] Compared with the prior art, the present invention has the following significant advantages:
[0013] (1) The present invention generates labeled samples in no-load conditions through a diffusion model, thereby compensating for the scarcity of samples and enabling the network to obtain sufficient samples for training.
[0014] (2) The present invention adds a wavelet packet denoising layer to the network. There is no need to perform denoising on the signal in advance. The signal samples can be denoised after passing through the wavelet denoising layer, which reduces manual time and improves the robustness of the network.
[0015] (3) The present invention adds pseudo labels during training to help the model learn the information in unlabeled samples. At the same time, a pseudo label threshold is set. Pseudo label calibration can only be performed when the model's predicted probability for unlabeled samples exceeds the threshold.
[0016] (4) The present invention uses a new joint distribution adaptive mechanism with the maximum mean square error as the difference metric to align the distribution of data under two working conditions, helping the model to learn the common features in the data, thereby realizing the diagnosis of gear faults under different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 It is a flow chart of a specific implementation method for fault diagnosis of a reciprocating motion manipulator rotary gear in the present invention;
[0018] Figure 2 It is a schematic diagram of the model structure of the reciprocating motion manipulator in the present invention;
[0019] Figure 3 This is a comparison diagram of the signal before and after the wavelet packet denoising layer is input in the present invention;
[0020] Figure 4 A comparison chart of data division between the conventional method in the present invention and the method proposed herein;
[0021] Figure 5 Visualization diagram of diagnostic features of the model of the present invention and the conventional model;
[0022] Figure 6 It is a confusion matrix diagram of the diagnosis results of the model of the present invention and the conventional model.
[0023] Figure 7 This is a diagram showing the accuracy of the diagnosis results of the model of the present invention and the conventional model. DETAILED DESCRIPTION
[0024] The specific embodiments of the present invention are described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here. Figure 1 , Figure 2 , Figure 3 as well as Figure 4 , Figure 5 , the specific steps are:
[0025] 1) Collect vibration signals at the gearbox of the manipulator's rotating device, and collect vibration signals of five types of gears (normal, missing teeth, broken teeth, and cracked) under operation. One rotation of the manipulator is a group of samples. Four hundred groups of signal samples of each of the five gear states under no-load conditions are collected (three hundred groups are labeled samples, and one hundred groups are unlabeled samples), and four hundred groups of signal samples of each of the five gear states under loaded conditions are collected (all four hundred groups are unlabeled samples). Therefore, there are a total of two thousand groups of samples under no-load conditions, including one thousand five hundred groups of labeled samples and five hundred groups of unlabeled samples, and a total of two thousand groups of samples under loaded conditions (unlabeled samples). A total of four thousand groups of samples under no-load and loaded conditions constitute the original data set.
[0026] 2) The data samples collected under no-load conditions are expanded through the diffusion model. The diffusion model learns the information attenuation caused by noise, so as to generate similar signals after learning. The diffusion model structure can be divided into forward diffusion process and reverse recovery engineering. It defines a Markov chain of diffusion steps, which slowly inputs random noise into the original signal until the signal completely becomes a noise signal, and then learns the reverse diffusion process to construct the required signal samples from the noise. The specific training process is as follows:
[0027] Forward process:
[0028] Given training samples x0~p(x0), the final noise signal can be obtained through the following Markov process:
[0029]
[0030] In the formula, x h represents the training sample at the hth diffusion step; x h-1 represents the training sample at the h-1th diffusion step; p(x h |x h-1 ) represents x h Relative to x h-1 A is the total number of diffusion steps; β h represents the variance at the h-th diffusion step; I is the identity matrix, which has the same dimension as the input signal; Represents the training sample x h In the normal distribution The measured values in the normal distribution have a mean of The covariance is β h ·Ι,· means multiplication of two numbers.
[0031] Backward process:
[0032] Given a noise sample after a forward process It represents the sample x h It obeys the normal distribution, its mean vector is 0, and its covariance matrix is the identity matrix I.
[0033] Therefore, the steps of noise removal can be expressed as:
[0034]
[0035] In the formula, p(x h-1 |x h ) represents the training sample x at the hth diffusion step h Relative to the change of the original training sample x0, Represents the training sample x h-1 In the normal distribution The measured value in the normal distribution has a mean of μ(x h ,h), the covariance is ∑(x h ,h), where μ(x h ,h) represents the training sample x h The mean value, Σ(x h ,h) represents the training sample x h The sum of .
[0036] Objective function:
[0037] Ideally, the neural network will be trained according to the maximum likelihood objective so that the probability of the model P being assigned to each original training sample x0 is as large as possible. The loss function in the model can be expressed as:
[0038]
[0039] In the formula, is the loss function of the model, and E represents minimizing the noise ε of the h-th diffusion step. h and the model estimated noise ε θ The mean square error between Represents a neural network, the input of the model is in brackets, ε is the initial output noise, that is, the noise input at the first diffusion step, where represents the (1-β h ), β h is the variance at the h-th diffusion step.
[0040] The sample expansion of labeled samples under no-load condition is performed through diffusion model, so the samples under each type of gear operation are expanded by 300 sets of labeled samples, and they are combined with the original data to form an expanded data set. Therefore, the composition of the expanded data set is 3,500 sets of samples of five types of gear faults under no-load condition (2,000 sets of labeled samples and 1,500 sets of unlabeled samples).
[0041] Finally, the samples of the expanded data set are divided into training set, test set, and validation set in the form of 7:2:1.
[0042] 3) Build a one-dimensional convolutional neural network model, which consists of a wavelet packet denoising layer, a maximum pooling layer, four convolution modules, a global average pooling layer, and a fully connected layer. Set the network learning rate Ir to 0.002. The structure of the one-dimensional convolutional network model is as follows.
[0043]
[0044]
[0045] Wavelet packet denoising layer: Since the robot's rotary device has a large impact, most of the characteristic information in the vibration signal is concentrated in the high-frequency part of the signal. Therefore, decomposing the high-frequency components of the signal through wavelet packets can obtain more detailed information, and the filtering of useful information can be better reduced during the denoising process. Therefore, the first layer of the network uses the wavelet packet denoising layer to denoise the input sample signal. First, decompose the signal in the input model, set the number of decomposition layers to three, and select the wavelet basis as 'db4'. The wavelet packet decomposition process can be expressed as:
[0046]
[0047] In the formula, x(t) represents the original signal with a time length of t, represents the wavelet coefficient of the kth wavelet packet node in the jth layer, ψj,k Indicates the wavelet coefficient in the kth wavelet packet node of the jth layer The corresponding wavelet basis function.
[0048] Since the hard threshold directly sets the coefficients below the threshold to zero, and the coefficients above the threshold remain unchanged, the soft threshold is smoother than the hard threshold processing. It adds a shrinkage amount λ. When the wavelet coefficient is greater than the given threshold, the wavelet coefficient shrinks by the amount λ in the direction of decreasing amplitude, and becomes 0 when the wavelet coefficient is less than the given threshold. In this way, since the coefficients are reduced instead of simply set to zero, the structure of the signal can be better preserved, thereby reducing the error in the wavelet packet reconstruction process. The expression of the soft threshold can be expressed as:
[0049]
[0050] In the formula, is the wavelet coefficient representing the kth wavelet packet node in the jth layer, is the wavelet packet coefficient after threshold processing, λ is the threshold, sgn(·) is the sign function, if the input value in the function is positive, the function value is 1, if the input value is 0, the function value is 0, if the input value is negative, the function value is -1.
[0051] After performing wavelet packet decomposition and soft threshold processing on the input signal, the wavelet packet coefficients are reconstructed to achieve signal noise reduction. The reconstruction process can be expressed as:
[0052]
[0053] In the formula, is the signal with a reconstructed time length of t, i.e. the denoised signal, J is the total number of decomposition layers, represents the wavelet coefficient of the kth wavelet packet node in the jth layer, ψ j,k Indicates the wavelet coefficient in the kth wavelet packet node of the jth layer The corresponding wavelet basis function.
[0054] Therefore, the input sample of the one-dimensional convolutional network can be expressed as follows after the wavelet packet denoising layer:
[0055]
[0056] In the formula, x(t) is the sample signal with a time length of t in the network input. is the denoised signal, and self.wavepacket represents the wavelet packet denoising layer.
[0057] Max Pooling Layer: The pooling window size of the max pooling layer in this model is 12, the stride is 12, the padding size is 1, and the number of channels is 1.
[0058] Convolution module: Four convolution modules are set in this model. The first three modules of these four convolution modules are composed of convolution layers, normalization layers and maximum pooling layers, and the fourth convolution module is composed of convolution layers and maximum pooling layers. In the first convolution module, the convolution kernel size of the convolution layer is 64, the step size is 1, the padding size is 1, the number of channels is 16, the activation function is ReLU, the number of channels of the normalization layer is 16, the pooling window size of the maximum pooling layer is 2, the step size is 2, the padding size is 1, and the number of channels is 16; in the second convolution module, the convolution kernel size of the convolution layer is 5, the step size is 1, the padding size is 1, the number of channels is 64, the activation function is ReLU, the number of channels of the normalization layer is 64, the pooling window size of the maximum pooling layer is 2, the step size is 2 , padding size is 1, and the number of channels is 64; the convolution kernel size of the convolution layer in the third convolution module is 5, the stride is 1, the padding size is 1, the number of channels is 128, the activation function is ReLU, the number of channels of the normalization layer is 128, the pooling window size of the maximum pooling layer is 2, the stride is 2, the padding size is 1, and the number of channels is 128; the convolution kernel size of the convolution layer in the fourth convolution module is 5, the stride is 1, the padding size is 1, the number of channels is 256, the activation function is ReLU, and the number of channels of the normalization layer is 256.
[0059] Global Average Pooling Layer: The number of channels in this layer is 256, and the size in each channel is 1.
[0060] Fully connected layer: The input feature number of this layer is 256, the output feature number is 5, and the activation function is ReLU.
[0061] 4) Introduce pseudo labels during the training process to realize the use of unlabeled data. First, the model is trained on labeled empty data. After a certain number of rounds of training, the unlabeled empty and unlabeled loaded samples are predicted. If the model's prediction probability is greater than a given threshold, the set label is used as the unlabeled sample pseudo label, and in subsequent training, the sample is used as a labeled sample to train the model together. This process is repeated multiple times until all labels are assigned. The training process of pseudo labels can be expressed as:
[0062]
[0063] Where θ represents the threshold of a given pseudo label, x τ represents the τth unlabeled sample, y τ represents the pseudo label of the τth unlabeled sample, represents the predicted label of the τth unlabeled sample, P(y τ |x τ ) represents the model for sample x τ The predicted probability value. None means that the unlabeled sample is not assigned a pseudo label.
[0064] 5) Since the robot alternates between no-load and loaded conditions during its reciprocating motion, the model needs to diagnose the data under both conditions. However, the distribution of the two types of data is different, and simple classification will result in poor diagnostic results.
[0065] (1) Maximum mean difference (MMD) is used to calculate the distribution difference between loaded data and unloaded data. It uses the Gaussian kernel function to map the samples into the reproduction kernel Hilbert space, and then determines whether the feature distribution matches by calculating the maximum average distance between the samples in the two domains (unloaded data and loaded data). MMD can be expressed as:
[0066]
[0067] Among them, MMD H (X S ,X T ) indicates no-load data X S and loaded data X T The maximum mean difference in Hilbert space, represents the i-th unloaded data sample and the i-th loaded data sample n S Indicates the number of samples of no-load data, n T represents the number of samples with loaded data, H represents the Hilbert space (RKHS), and φ(·) represents the mapping function in RKHS.
[0068] The maximum mean square difference (MMSD) can comprehensively reflect the mean and variance information of data samples in the reproducing kernel Hilbert space to solve the problem that the relationship between MMD and the kernel function is confusing, resulting in poor difference representation. MMSD can be expressed as:
[0069]
[0070] in Indicates no-load data X S and loaded data X T The maximum variance difference in the Hilbert space, H1H2 represents the tensor product of the Hilbert spaces H1 and H2 to represent the original Hilbert space H, φ(·) and ξ(·) represent the mapping functions in two different RKHSs, represents the i-th no-load data sample, represents the fth loaded data sample, represents the wth no-load data sample, represents the rth loaded data sample, n S Indicates the number of samples of no-load data, n TIndicates the number of samples with loaded data.
[0071] However, it is not enough to reduce the distribution difference between the two domain data only by adjusting the marginal distribution, but also the conditional distribution of the data needs to be adjusted. Because marginal distribution adaptation improves the commonality of the two domain features by adjusting the distribution center of the two domains, conditional distribution adaptation pays more attention to the difference between the data itself to match the discriminant structure between the two domain data. Therefore, the present invention uses MSSD as the difference metric to calculate the marginal distribution difference and conditional distribution difference between the two domain data, and its loss function L MD It can be expressed as:
[0072]
[0073] Among them, MMSDH1H2(X S ,X T ) represents the overall distribution difference between unloaded data and loaded data based on the MMSD difference metric, that is, the marginal distribution difference between the two domain data, while represents the sum of the distribution differences of different faults between no-load data and loaded data based on the MMSD difference metric, that is, the conditional distribution difference between the two domains of data, C represents the total number of fault categories, X S Indicates no-load data, X T Indicates that there is data loaded. Indicates the no-load data of the Z-type fault, Indicates the loaded data of the Zth type fault.
[0074] (2) The model uses cross entropy loss L for classification loss function during training CE , which can be expressed as:
[0075]
[0076] Where M represents the number of labeled samples, N represents the number of unlabeled samples with pseudo labels, and q(y g ) represents the label prediction probability value of the g-th labeled sample, y τ Represents the label prediction probability value of the τth unlabeled sample.
[0077] Therefore, the overall objective function of the model will combine the two-domain distribution loss and the cross entropy loss, and finally diagnose the overall objective loss function of the model L total It can be expressed as:
[0078] L total =L CE +μL MD
[0079] Where μ is the adjustment parameter, μ~2η / (1+e -Kv)-η, η is the upper limit of the regularization coefficient, v is the number of current iterative training of the model, and K is the control change rate.
[0080] Through training with the overall loss function, the model convergence is achieved, and finally the diagnosis of the reciprocating manipulator rotating gear fault is completed.
[0081] Verification of model effect:
[0082] The diffusion model is used to expand the sample of the labeled empty data, and the expanded samples and the original data samples are combined into an expanded data set. The composition of the expanded data set and the original sample data set is as follows:
[0083]
[0084] At the same time, in order to verify the effect of the expanded sample on the model diagnosis effect, the effect of the original sample data set training model and the expanded data set training model were compared. The results of the comparative experiment are as follows:
[0085]
[0086]
[0087] The experimental results show that by expanding the labeled samples of no-load conditions through the diffusion model, the training samples of the model are increased, so that it can achieve better training results, and the pseudo-label calibration of unlabeled samples is faster. At the same time, the diagnostic accuracy when the unlabeled samples are labeled is also higher. Finally, the diagnostic accuracy of the model trained by the original data set is only 93%, while the diagnostic accuracy of the model trained by the expanded data set is as high as 96.6%.
[0088] The proposed method model and the model without wavelet packet denoising layer were experimented on the expanded data set and the original data set, and the comparative experimental results are as follows:
[0089] Sample Set The diagnostic accuracy of the model of the present invention Diagnostic accuracy of the model without wavelet denoising layer Original Dataset 93.4% 90.2% Expanding the dataset 96.5% 92.8%
[0090] At the same time, in order to compare the effectiveness of the method proposed in the invention, it was compared with the conventional experimental method, and multiple experiments were conducted, and the average of the diagnostic accuracy of multiple experiments was taken as the result for comparison. The results of the comparative experiment are as follows:
[0091]
[0092] It can be seen from the above experimental results that the method proposed in the present invention has a higher diagnostic accuracy improvement compared with conventional methods. In order to solve the problem of gear failure with small samples and few labels in reciprocating manipulators, the present invention expands the sample of labeled no-load working condition data through a diffusion model to solve the problem of insufficient samples, and then adds a wavelet packet noise reduction layer to the network to reduce the noise of the input signal, thereby reducing the impact of noise and improving the effect of the model on input signal feature extraction. At the same time, in order to make the model extract common features under different working conditions, that is, different data distributions, a joint distribution adaptive method with the maximum mean square error difference as the difference standard is set to realize the alignment of data in marginal distribution adaptation and conditional distribution adaptation, and finally realize the identification of gear faults under different working conditions. The invention can realize high-precision diagnosis and has a wide range of applications in the field of fault diagnosis of manipulators and other rotating machinery. It also provides a new method for fault diagnosis technology based on deep learning.
Claims
1. A method for diagnosing a reciprocating manipulator rotary gear fault, characterized in that: include: Step 1, collecting vibration signals of the reciprocating manipulator under different gear faults in no-load and loaded working conditions; Step 2: Expand the labeled samples under no-load conditions through the diffusion model, and combine the expanded samples with the unlabeled samples under no-load conditions and the unlabeled samples under loaded conditions to form a data set; Step 3: Build a one-dimensional convolutional network model, and add a wavelet packet denoising layer to the convolutional network layer to perform denoising on the signal in the input network; Step 4: Input the no-load and loaded working condition data into the convolutional network for feature extraction at the same time, use the labeled no-load samples to guide the model training, and make predictions for the unlabeled samples. Pseudo-labels are given to the unlabeled no-load samples and loaded samples whose prediction probabilities are higher than the set threshold. Then, the obtained pseudo-label samples are added to the model training, and the pseudo-label calibration of all unlabeled samples is realized in this way. In step 5, a new joint distribution adaptive mechanism with the maximum mean square error as the difference metric is used to achieve distribution alignment of data under two working conditions, helping the model to extract common features of the same type of gear faults under different working conditions, and finally classify the gear faults.
2. The method for diagnosing a reciprocating manipulator rotary gear fault according to claim 1, characterized in that: The expansion model training process in step 2 is divided into two parts. The first part is a fixed forward process, in which Gaussian noise is gradually added to the vibration signal of the robot when it is working in an unloaded condition until a pure noise signal is obtained. The second part is a learnable reverse denoising process, in which the data is gradually denoised from the pure noise data through neural network learning, and finally data approximating the original signal is obtained, thereby realizing the generation of labeled samples when the robot is working in an unloaded condition.
3. The method for diagnosing a reciprocating manipulator rotary gear fault according to claim 2, characterized in that: The forward process specifically includes: Given training samples x0~p(x0), the final noise signal can be obtained through the following Markov process: In the formula, p(x h |x h-1 ) represents the training sample x at h diffusion steps h Relative to the training sample x at the h-1th diffusion step h-1 A is the total number of diffusion steps; β h represents the variance at the h-th diffusion step; I is the identity matrix; Represents the training sample x h In the normal distribution The measured value in the normal distribution has a mean of The covariance is β h · Ι.
4. The method for diagnosing a reciprocating manipulator rotary gear fault according to claim 2, characterized in that: The backward denoising process includes: Given a noise sample after a forward process The steps of noise removal can be expressed as: In the formula, p(x h-1 |x h ) represents the training sample x at the hth diffusion step h Relative to the change of the original training sample x0, Represents the training sample x h-1 In the normal distribution The measured value in the normal distribution has a mean of μ(x h ,h), the covariance is ∑(x h ,h), where μ(x h ,h) represents the training sample x h The mean value, ∑(x h ,h) represents the training sample x h The sum of .
5. The method for diagnosing a reciprocating manipulator rotary gear fault according to claim 2, characterized in that: The neural network uses the maximum likelihood objective to train the neural network, and the loss function is expressed as: In the formula, is the loss function, E represents the minimum noise ε of the h-th diffusion step h and the model estimated noise ε θ The mean square error between Represents a neural network, the input of the model is in brackets, ε is the initial output noise, that is, the noise input at the first diffusion step, where represents the (1-β h ), β h represents the variance at the h-th diffusion step, and x0 is the original training sample.
6. The method for diagnosing a reciprocating manipulator rotary gear fault according to claim 1, characterized in that: The one-dimensional convolutional network model built in step 3 consists of a wavelet packet denoising layer, a maximum pooling layer, four convolution modules, a global average pooling layer, and a fully connected layer; The wavelet packet denoising layer first decomposes the signal in the input model, the number of decomposition layers is set to three, and the wavelet basis is selected as 'db4'; The pooling window size of the maximum pooling layer is 12, the stride is 12, the padding size is 1, and the number of channels is 1; The first three modules of the four convolution modules are composed of convolutional layers, normalization layers and maximum pooling layers, and the fourth convolutional module is composed of convolutional layers and maximum pooling layers; the convolution kernel size of the convolution layer in the first convolutional module is 64, the step size is 1, the padding size is 1, the number of channels is 16, the activation function is ReLU, the number of channels of the normalization layer is 16, the pooling window size of the maximum pooling layer is 2, the step size is 2, the padding size is 1, and the number of channels is 16; the convolution kernel size of the convolution layer in the second convolutional module is 5, the step size is 1, the padding size is 1, the number of channels is 64, the activation function is ReLU, and the normalization layer is The number of channels is 64, the pooling window size of the maximum pooling layer is 2, the stride is 2, the padding size is 1, and the number of channels is 64; the convolution kernel size of the convolution layer in the third convolution module is 5, the stride is 1, the padding size is 1, the number of channels is 128, the activation function is ReLU, the number of channels of the normalization layer is 128, the pooling window size of the maximum pooling layer is 2, the stride is 2, the padding size is 1, and the number of channels is 128; the convolution kernel size of the convolution layer in the fourth convolution module is 5, the stride is 1, the padding size is 1, the number of channels is 256, the activation function is ReLU, and the number of channels of the normalization layer is 256; The number of channels in the global average pooling layer is 256, and the size in each channel is 1; The input feature number of the fully connected layer is 256, the output feature number is 5, and the activation function is ReLU.
7. The method for diagnosing a reciprocating manipulator rotary gear fault according to claim 1 or 6, characterized in that: The wavelet packet decomposition process is: In the formula, x(t) represents the original signal with a time length of t, represents the wavelet coefficient of the kth wavelet packet node in the jth layer, ψ j,k Indicates the wavelet coefficient in the kth wavelet packet node of the jth layer The corresponding wavelet basis function; The soft threshold is used to set the threshold, and the expression is expressed as: In the formula, is the wavelet coefficient representing the kth wavelet packet node in the jth layer, is the wavelet packet coefficient after threshold processing, λ is the threshold, sgn(·) is the sign function, if the input value in the function is positive, the function value is 1, if the input value is 0, the function value is 0, if the input value is negative, the function value is -1; After the input signal is decomposed by wavelet packets and soft threshold processing, the wavelet packet coefficients are reconstructed to achieve signal noise reduction. The reconstruction process can be expressed as: In the formula, is the signal with a reconstructed time length of t, i.e. the denoised signal, J is the total number of decomposition layers, represents the wavelet coefficient of the kth wavelet packet node in the jth layer, ψ j,k Indicates the wavelet coefficient in the kth wavelet packet node of the jth layer The corresponding wavelet basis function.
8. The method for diagnosing a reciprocating manipulator rotary gear fault according to claim 1, characterized in that: The training process of pseudo labels can be expressed as: Where θ represents the threshold of a given pseudo label, x τ represents the τth unlabeled sample, y τ represents the pseudo label of the τth unlabeled sample, represents the predicted label of the τth unlabeled sample, P(y τ |x τ ) represents the model for sample x τ The predicted probability value. None means that the unlabeled sample is not assigned a pseudo label.
9. The method for diagnosing a reciprocating manipulator rotary gear fault according to claim 1, characterized in that: Step 5 specifically includes: The maximum mean square difference (MSSD) is used as the difference metric to calculate the marginal distribution difference and conditional distribution difference between the source domain and the target domain data. The loss function L MD It can be expressed as: in It represents the overall distribution difference between unloaded data and loaded data based on the MMSD difference metric, that is, the marginal distribution difference between the two domain data, while represents the sum of the distribution differences of different faults between no-load data and loaded data based on the MMSD difference metric, C represents the total number of fault categories, and X S Indicates no-load data, X T Indicates that there is data loaded. Indicates the no-load data of the Z-type fault, Indicates the load data of the Zth type fault; The model uses cross entropy loss L for classification loss function during training. CE : Where M represents the number of labeled samples, N represents the number of unlabeled samples with pseudo labels, and q(y g ) represents the label prediction probability value of the g-th labeled sample, y τ Represents the label prediction probability value of the τth unlabeled sample; The overall objective loss function L of the final diagnosis model total It can be expressed as: L total =L CE +μL MD Where μ is the adjustment parameter, μ~2η / (1+e -Kv )-η, η is the upper limit of the regularization coefficient, v is the number of current iterative training of the model, and K is the control change rate.
