Multi-state same-scale quantitative evaluation method and system for harmonic reducer of industrial robot

Abstract

The invention provides a multi-state same-scale quantitative evaluation method and system for a harmonic reducer of an industrial robot, and belongs to the field of harmonic reducer fault quantitative evaluation. The method aims to solve the problems that quantitative analysis of the fault degree of a harmonic reducer is difficult and the fault degrees of different positions cannot be evaluated under a unified scale in an existing method. The method comprises the following steps: performing continuous wavelet transform on an acoustic emission signal to construct a two-dimensional time-frequency graph data set; then, a convolution attention module is provided to improve ResNet so as to fully mine deep features of the time-frequency graph; a multi-kernel function is introduced to improve SVDD, and an MKSVDD evaluation model is constructed based on the depth features of the harmonic reducer in the normal state; calculating the distance between the characteristics of different fault degrees and the normal state sphere center, constructing an evaluation index, and obtaining a quantitative evaluation curve through fitting; according to the structure of the harmonic reducer and an acoustic emission signal propagation mechanism, a relative compensation distance is provided to construct a multi-state evaluation index, and quantitative evaluation of different states under a unified scale is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of quantitative evaluation of harmonic reducer faults, and in particular to a method and system for quantitative evaluation of an industrial robot harmonic reducer in multiple states and at the same scale. Background Art

[0002] The “Robot+” Application Action Plan clearly proposes to accelerate the industrialization process of robots. As an important part of the automation field, the demand for industrial robots in practical applications is constantly increasing, and the application fields are becoming more and more extensive [1]. As the “joint” of industrial robots, the normal operation of harmonic reducers is crucial for the safe and reliable operation of industrial robots [2]. Harmonic reducers are generally under high load and high torque working conditions, and are prone to performance degradation and failure. In the worst case, it will cause production stagnation and endanger the lives of production line workers [3-4]. During the operation cycle of harmonic reducers, they usually go through the complete process of normal operation, different degrees of fault manifestation and complete failure. In this situation, accurate and quantitative evaluation of the operating status of harmonic reducers is a prerequisite for ensuring the normal operation of industrial robots. Therefore, based on fault diagnosis technology, it is of great significance to study the degradation law of harmonic reducers from intact to a series of fault levels, so as to achieve quantitative evaluation of the health status of harmonic reducers.

[0003] In recent years, with the continuous development of artificial intelligence, deep learning methods have been widely used in the field of fault diagnosis of rotating machinery such as harmonic reducers due to their powerful adaptive feature extraction capabilities. Reference [5] uses convolutional neural networks to extract deep features of multi-sensor fusion time-frequency graphs, thereby realizing fault diagnosis of harmonic reducers. Reference [6] uses generative adversarial networks combined with multi-scale convolutional neural networks to realize fault diagnosis under unbalanced data of harmonic reducers. Reference [7] uses convolutional neural networks to extract deep features of harmonic reducer voltage signals to realize fault diagnosis of harmonic reducers. Reference [8] extracts a fault diagnosis method that adapts to information fusion subdomains and effectively realizes fault classification under different working conditions in unsupervised scenarios. Reference [9] proposes to use one-dimensional convolutional neural networks to extract deep features of original signals for fault diagnosis, and verifies the effectiveness of the proposed method using real-time operation data of industrial robot harmonic reducers. Reference

[10] proposes a harmonic reducer fault diagnosis method based on personalized federated learning. While protecting data privacy, it uses convolutional neural networks to extract deep features of vibration signals, effectively realizing fault diagnosis of harmonic reducers.

[0004] Deep learning methods provide a way to effectively extract fault features and realize fault diagnosis. However, when training deep neural networks, in order to enhance the network's feature extraction capability, it is often necessary to set up more convolution channels. However, the proportion of useful fault information in each channel is unbalanced

[11] . Therefore, by adding an attention mechanism module to the deep learning model, the attention to useful information is increased, thereby improving the model's ability to process and learn fault information. The attention mechanism is a plug-and-play module that can be integrated into any layer of any network. It enables the network model to focus on mining the significant features of the image and suppress invalid features, thereby further improving the network's recognition performance. Reference

[12] proposed a bearing fault diagnosis method based on an improved convolutional neural network with an attention mechanism, and verified that the classification accuracy of the improved model was significantly improved. Reference

[13] proposed a convolutional neural network improved with a multi-scale attention mechanism to enhance the network's feature learning capability and effectively realize gearbox fault diagnosis. Reference

[14] proposed a dual-module attention convolutional neural network to extract deep features related to industrial robot reducer faults and improve the accuracy of fault diagnosis. Reference

[15] proposed an improved deep residual network fault diagnosis method, which obtains channel global information and adjusts the weights through the channel attention mechanism to enhance the generalization ability of the network and effectively realize the fault diagnosis of rolling bearings with different loads. Reference

[16] introduced a multi-scale attention mechanism into the convolutional neural network architecture, guiding the model to focus on global and local key fault information, establishing a multi-scale feature enhancement module, eliminating useless features, and realizing the fault diagnosis of gearboxes and bearings under variable working conditions.

[0005] The above-mentioned attention mechanism and deep learning methods have achieved good results in the field of rotating machinery fault diagnosis, but most of them focus on the identification of fault location, and the quantitative characterization of fault severity has not been given sufficient attention and in-depth exploration. However, in actual industrial production, the quantitative analysis of fault severity is more important than fault classification for rotating machinery such as harmonic reducers. In the field of quantitative health status assessment, evaluation indicators are mainly divided into two categories: probability likelihood value indicators and distance type indicators. For example, Gaussian mixture models, continuous hidden Markov models and Bayesian networks are all based on probability likelihood indicators. The probability likelihood value indicator generally ranges between [0,1], but there is an oversaturation phenomenon, that is, when the fault has not reached the maximum severity, the similarity probability between the signal feature to be tested and the signal feature without fault is already 0. At this time, the value of the distance type indicator will no longer change as the fault severity continues to increase, and it cannot well describe the failure state of the rotating machinery

[17] . Distance type indicators show a good monotonic trend with the degradation process of the rotating machinery, such as support vector machines and support vector data description (SVDD). SVDD is widely used in the field of rotating machinery performance degradation assessment. Compared with support vector machines, its single-value classification effect is more significant and accurate. Reference

[18] proposed a degradation index extraction method that considers global and local feature extraction combined with SVDD. The SVDD model is trained using normal state vectors to construct a distance-type degradation index, realizing the performance degradation assessment of rolling bearings. Reference

[19] proposed a rolling bearing performance degradation assessment method based on comprehensive index reduction combined with SVDD. The comprehensive feature index set of bearings is constructed by combining time domain and frequency domain indicators. The local linear embedding algorithm is used to reduce the dimension of the extracted main fault features. The distance from the test sample to the sphere center obtained by training the normal sample is used to describe the degree of performance degradation, to characterize and evaluate the life degradation process of rolling bearings, and objectively reveal the performance degradation process of rolling bearings. Reference

[20] uses a deep belief network to extract the fault features of rolling bearings, and constructs an evaluation model through SVDD to effectively realize the performance degradation assessment of bearings.

[0006] The above-mentioned distance-based evaluation indicators have achieved good results in the field of rotating machinery performance degradation assessment. However, they are limited to performance degradation analysis under specific fault location conditions. Not only do they fail to accurately quantify the fault extent, but they are also unable to quantitatively analyze the fault extent of different fault locations at the same scale, thus limiting the accuracy of quantitative evaluation.

[0007] Based on the above reasons, this paper addresses the difficulties of existing research in quantitatively analyzing the fault severity of harmonic reducers and the inability to evaluate the fault severity at different locations on a unified scale. Taking harmonic reducers as the research object and acoustic emission signals as the signal monitoring source, this paper proposes a new quantitative evaluation method for harmonic reducers in multiple states and at the same scale for industrial robots. This method introduces a convolutional attention module into the original residual network, constructs a feature extraction network that can better focus on important fault information, proposes an MKSVDD evaluation model, constructs a distance indicator that can reflect the degradation law of harmonic reducer faults, proposes a compensation distance based on the relative position of the sensor installation position and the fault, and constructs a multi-state evaluation indicator, ultimately achieving quantitative evaluation of harmonic reducers in multiple states and at the same scale for industrial robots. Summary of the Invention

[0008] The technical problems to be solved by the present invention are:

[0009] In order to solve the problem that existing methods are difficult to quantitatively analyze the fault degree of harmonic reducers and cannot evaluate the fault degree at different locations under a unified scale.

[0010] The present invention is to solve the above technical problems using the following technical solutions:

[0011] The present invention provides a quantitative evaluation method for an industrial robot harmonic reducer in multiple states and at the same scale, comprising the following steps:

[0012] S100: Data collection: collecting one-dimensional acoustic emission signals of harmonic reducers with different fault locations and fault degrees;

[0013] S200, data preprocessing, performing continuous wavelet transform on the one-dimensional acoustic emission signal collected in step S100 to obtain a two-dimensional time-frequency image;

[0014] S300, constructing a sample set, dividing the two-dimensional time-frequency image obtained after the transformation in step S200 for training the feature extraction network and testing the performance of the feature extraction network; collecting data of unknown health status of the harmonic reducer, and constructing a validation set after data preprocessing;

[0015] S400. Construct a fault feature extraction network, introduce a convolutional attention mechanism to improve the residual network, and extract deep features of the training set. After iteration, calculate the model loss function value and use the backpropagation method to calculate the weight of each parameter. Use the gradient descent method to update the model parameters in the direction of minimizing the loss function, and finally make the objective function converge, thus obtaining the constructed harmonic reducer fault feature extraction network.

[0016] S500: Construct a multi-state quantitative evaluation model, input the training set into the fault feature extraction network built in step S400, save and normalize the features corresponding to all correctly classified samples; use the feature vectors in the fault-free state as training samples, train the multi-MKSVDD evaluation model, obtain the radius R of the hypersphere, use the MKSVDD in the fault-free state as a benchmark, calculate the distance from the feature vectors of different fault degrees at a certain position to the center of the normal MKSVDD sphere, obtain a quantitative evaluation curve by fitting the distances of different fault degrees relative to the center of the normal sphere, introduce the relative compensation distance, and obtain a quantitative evaluation curve under the same scale in multiple states;

[0017] S600, quantitative assessment of health status, input the verification set into the multi-state quantitative assessment model trained in step S500, and obtain the distance D between the sample to be tested and the center of the hypersphere. If D≤R, it means that the sample to be tested is in a normal state, otherwise it indicates that the sample to be tested is in a faulty state. Then, the quantitative assessment curve is combined to achieve a quantitative analysis of the degree of fault and obtain the final quantitative assessment result of the health status of the harmonic reducer.

[0018] Furthermore, in step S400, the introduced convolutional attention mechanism includes a channel mechanism and a spatial attention mechanism.

[0019] First, calculate the channel attention of the feature map and obtain the channel attention weighted feature map:

[0020]

[0021] Where: φ is the sigmoid activation function; W0 and W1 are the two hidden layers in the multi-layer perceptron model; Represents the average pooling feature of the channel dimension; Represents the maximum pooling feature of the channel dimension;

[0022] Then, spatial attention calculation is performed on this basis to obtain the spatial attention weighted feature map:

[0023]

[0024] Where: φ is the sigmoid activation function; f 7×7 It is an image convolution operation with a convolution kernel size of 7×7; Represents the average pooled features of the spatial dimension; Represents the maximum pooling feature of the spatial dimension;

[0025] Finally, we get a cascade of channel attention mechanism and spatial attention mechanism, which can be expressed as formula (3) and formula (4):

[0026]

[0027] Where: X represents the input feature map; X1 represents the feature map after channel attention optimization; X2 represents the feature map after channel and spatial attention optimization; M c represents the one-dimensional convolution of the channel attention module; M s Represents the two-dimensional convolution of the spatial attention module; Represents pixel-by-pixel multiplication.

[0028] Furthermore, in step S400, a convolutional attention mechanism is embedded into the tail end of the residual network; information features are extracted through the convolution operation, and the extracted features are passed through the convolutional attention mechanism, allowing the network to adaptively obtain channel information and spatial information related to the fault state of the harmonic reducer to obtain feature representation;

[0029] The output of the residual network based on the channel attention mechanism and the spatial attention mechanism is expressed as:

[0030] H(f(x))+x(5)

[0031] Where: x is the input; f is the residual function; H is the attention function of the convolutional attention mechanism.

[0032] Furthermore, in step S500, when training the MKSVDD evaluation model,

[0033] For the target class sample set M={x1,x2,...,x N}, find the optimal hypersphere, which is described by the center a and radius R, and satisfies the optimization function shown in formula (6):

[0034]

[0035] Where: C is the penalty parameter, which controls the trade-off between hypersphere error and algorithm complexity; ξ i is a slack variable, allowing some training data to be outside the hypersphere; c is the center of the hypersphere;

[0036] Introducing the Lagrange multiplier α i , transform the optimization function of formula (6) into the Lagrangian extreme value problem as shown in formula (7):

[0037]

[0038] Among them, a j is an intermediate variable;

[0039] The kernel function is used to transform the samples from linearly inseparable low-dimensional space to linearly separable high-dimensional space. In the mapping space, the radius of the hypersphere R is determined by any support vector x. k The distance to the center is calculated as:

[0040]

[0041] Where: K(·) is the kernel function, representing the inner product operation;

[0042] For the new sample y, its distance from the center of the sphere is shown in formula (9):

[0043]

[0044] Furthermore, in step S500, multiple kernel functions are combined in a linearly weighted manner. The construction method of the multi-kernel kernel function is shown in formula (10):

[0045]

[0046] Where: K mix Represents the combination of multiple kernel functions, M represents the total number of multi-core kernel functions; μ m Represents the weight coefficient, satisfying μ m ≥0, at the same time

[0047] The Gaussian radial basis kernel function is used to describe the sample distribution. The definition of the Gaussian radial basis kernel function is shown in formula (11):

[0048]

[0049] Where: σ is the kernel function width parameter;

[0050] The multi-kernel kernel function is constructed by weighted summing of two Gaussian radial basis kernel functions with different kernel widths to improve the learning and generalization capabilities of the multi-kernel kernel function. The specific form is shown in formula (12):

[0051]

[0052] Where: K rbf Represents the Gaussian radial basis kernel function; σ1 and σ2 represent two different kernel width parameters of the kernel function; in order to control the weights of the two kernel functions, the weight coefficient μ is taken in the range of 0 to 1, that is, μ∈[0,1].

[0053] Furthermore, when different fault locations are put together, the inner and outer ring signals are compensated, as shown in formula (13):

[0054]

[0055] Where: R OR is the relative radius of the hypersphere containing all fault severity samples in the outer circle, d IR is the relative compensation distance of the inner ring, d ORis the relative compensation distance of the outer ring.

[0056] A quantitative evaluation system for an industrial robot harmonic reducer in multiple states and at the same scale, the system has a program module corresponding to the above steps, and executes the steps in the above-mentioned quantitative evaluation method for an industrial robot harmonic reducer in multiple states and at the same scale during operation.

[0057] A computer-readable storage medium stores a computer program configured to implement, when called by a processor, the steps of a quantitative evaluation method for a harmonic reducer of an industrial robot under multiple states and the same scale.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] (1) Based on existing deep neural networks, this paper proposes a convolutional attention mechanism to enhance the proportion of useful fault information and ignore redundant information. Verified by real-time data collection, the proposed CBAM-ResNet model has higher classification accuracy than the original ResNet model, proving that the improved ResNet network with the CBAM mechanism pays more attention to the important features of the fault and the extracted features are more in line with actual needs.

[0060] (2) The present invention proposes a method for quantitatively evaluating the health status of a harmonic reducer based on MKSVDD. This method constructs a distance-based index based on the MKSVDD model. The MKSVDD is trained using correctly classified normal state sample features. The distance between the sample features of different fault levels and the center of the normal state sphere is then calculated as an evaluation index. This index can clearly distinguish between different fault levels and effectively construct a quantitative evaluation model for the health status of a harmonic reducer.

[0061] (3) Based on the structure and signal propagation mechanism of the flexible thin-walled bearing of the harmonic reducer of the industrial robot, the present invention proposes a relative compensation distance, reveals the characterization relationship between the harmonic reducer and the acoustic emission signal transmission path at different fault locations, and reveals the performance degradation law of the harmonic reducer at different fault locations at the same scale, which facilitates the evaluator to conduct quantitative analysis of the harmonic reducer in different states at the same scale, and provides a new perspective and effective method for the performance evaluation and fault diagnosis of the harmonic reducer. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a diagram of the residual block structure in an embodiment of the present invention;

[0063] Figure 2 This is a structural diagram of the CBAM attention module in an embodiment of the present invention;

[0064] Figure 3Schematic diagram of the improved residual network structure in an embodiment of the present invention;

[0065] Figure 4 This is a two-dimensional space description diagram of SVDD in an embodiment of the present invention;

[0066] Figure 5 This is a flowchart of a quantitative evaluation method for an industrial robot harmonic reducer in multiple states and at the same scale according to an embodiment of the present invention;

[0067] Figure 6 This is a structural diagram of an experimental platform for collecting acoustic emission signals of a harmonic reducer according to an embodiment of the present invention;

[0068] Figure 7 This is a structural diagram of the measurement and control system of the harmonic reducer acoustic emission signal acquisition experimental platform in an embodiment of the present invention;

[0069] Figure 8 1 is a diagram showing the deployment of acoustic emission sensors according to an embodiment of the present invention;

[0070] Figure 9 This is a visualization diagram of the original residual network features in an embodiment of the present invention;

[0071] Figure 10 This is a visualization diagram of improved residual network features in an embodiment of the present invention;

[0072] Figure 11 This is the confusion matrix diagram of the original residual network in an embodiment of the present invention;

[0073] Figure 12 This is a diagram of the improved residual network confusion matrix in an embodiment of the present invention;

[0074] Figure 13 This is the envelope spectrum of the inner race mild fault in the embodiment of the present invention;

[0075] Figure 14 This is a confusion matrix diagram of vibration signal classification in an embodiment of the present invention;

[0076] Figure 15 Schematic diagram of the hyperparameter optimization process in an embodiment of the present invention;

[0077] Figure 16 A visualization diagram of the relative distances of all samples of different fault states in the inner ring according to an embodiment of the present invention;

[0078] Figure 17 This is a visualization diagram of relative distances under different fault levels of the inner ring of the acoustic emission signal in an embodiment of the present invention;

[0079] Figure 18 This is a visualization diagram of relative distances under different fault levels of the inner ring of the vibration signal in an embodiment of the present invention;

[0080] Figure 19 This is a quantitative evaluation curve diagram of the inner ring of the harmonic reducer in an embodiment of the present invention;

[0081] Figure 20 This is a visualization diagram of relative distances under different fault levels of the outer ring in an embodiment of the present invention;

[0082] Figure 21 This is a quantitative evaluation curve diagram of the outer ring of the harmonic reducer in an embodiment of the present invention;

[0083] Figure 22 This is a multi-state same-scale quantitative evaluation curve diagram of a harmonic reducer in an embodiment of the present invention;

[0084] Figure 23 This is a diagram of the quantitative analysis results of the harmonic reducer in multiple states and at the same scale in an embodiment of the present invention. DETAILED DESCRIPTION

[0085] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0086] 1 Improved residual network

[0087] 1.1 Residual Network

[0088] Residual network (ResNet)

[21] It is developed on the basis of convolutional neural network. After the identity mapping operation is performed on the network layer, nonlinear transformation is superimposed as the output, which solves the problem of gradient disappearance caused by the increase of network layers in convolutional neural network. The residual block is the basic unit of the residual network. Its structure is as follows Figure 1 shown.

[0089] The input of the residual block is x, and the output is H(x). The residual is the difference between H(x) and x, that is:

[0090] f(x)=H(x)-x (1)

[0091] The core of the residual network is the learning function f(x). Unlike traditional convolutional neural networks, it only learns the difference between the input and output of the residual block, which reduces the learning difficulty and avoids the problem of network performance degradation caused by the increase in the number of network layers.

[0092] 1.2 Improved Residual Network

[0093] While the original residual network already has feature extraction capabilities, it lacks the ability to prioritize information from different channels, resulting in wasted computing resources and reduced recognition capabilities. The attention mechanism adaptively learns and calculates the weight of the impact of input data on output data, allowing the neural network model to focus on necessary features and suppress unimportant ones, thereby improving recognition accuracy in computer image classification tasks.

[0094] Convolutional block attention module (CBAM)

[22] It consists of a channel attention module and a spatial attention module, such as Figure 2 As shown in the figure, we first calculate the channel attention of the feature map to obtain the feature map weighted by the channel attention. Then, we perform spatial attention calculation on this basis to obtain the feature map weighted by the spatial attention, forming a "cascade" effect.

[0095] The channel attention mechanism can be expressed as:

[0096]

[0097] Where: φ is the sigmoid activation function; W0 and W1 are the two hidden layers in the multi-layer perceptron model; Represents the average pooling feature of the channel dimension; Represents the maximum pooling feature of the channel dimension.

[0098] The spatial attention mechanism can be expressed as:

[0099]

[0100] Where: φ is the sigmoid activation function; f 7×7 It is an image convolution operation with a convolution kernel size of 7×7; Represents the average pooled features of the spatial dimension; Represents the maximum pooling feature of the spatial dimension.

[0101] CBAM consists of a cascade of channel attention mechanism and spatial attention mechanism, expressed as formula (4) and formula (5):

[0102]

[0103] Where: X represents the input feature map; X1 represents the feature map after channel attention optimization; X2 represents the feature map after channel and spatial attention optimization; M c represents the one-dimensional convolution of the channel attention module; M s Represents the two-dimensional convolution of the spatial attention module; Represents pixel-by-pixel multiplication.

[0104] The improved residual network structure is as follows Figure 3 As shown in Figure 1, the CBAM module is embedded at the end of the residual network. Information features can be extracted through convolution operations. Passing these extracted features through the CBAM module allows the network to adaptively acquire channel and spatial information related to the harmonic reducer fault state, resulting in a more refined feature representation. The output of the residual network based on the channel attention and spatial attention mechanisms is represented as:

[0105] H(f(x))+x(6)

[0106] Where: x is the input; f is the residual function; H is the CBAM attention function.

[0107] 2 Multi-core Support Vector Data Description

[0108] 2.1 Support Vector Data Description

[0109] Support Vector Data Description

[23] It is a single-value classification method. Its basic idea is to find a hypersphere with the smallest volume that contains as many target samples as possible through training of target samples. Its two-dimensional description is as follows: Figure 4 shown.

[0110] For the target class sample set M={x1,x2,...,x N}, find an optimal hypersphere that can contain all or most target class samples. The hypersphere is described by center a and radius R, and the optimization function it satisfies is shown in formula (7):

[0111]

[0112] Where: C is the penalty parameter, which controls the trade-off between hypersphere error and algorithm complexity; ξ i is a slack variable that allows some training data to be outside the hypersphere.

[0113] As with traditional support vector machines, the above problem can be solved by introducing the Lagrange multiplier α i , which is transformed into the Lagrangian extreme value problem as shown in formula (8):

[0114]

[0115] Among them, a j is an intermediate variable, which is the result after differentiation and will disappear after the final summation;

[0116] Since the data in the input space is not linearly separable, the kernel function is used to transform the samples from the low-dimensional space linearly inseparable to the high-dimensional space linearly separable. The samples can be better described in the new high-dimensional space. In the mapping space, the radius of the hypersphere R is determined by any support vector x k The distance to the center is calculated as

[0117]

[0118] Where: K(·) is the kernel function, representing the inner product operation.

[0119] For the new sample y, its distance from the center of the sphere is shown in formula (10):

[0120]

[0121] 2.2 Construction of multi-core kernel function

[0122] When designing algorithms, kernel functions can transform problems that are linearly inseparable in low-dimensional space into linearly separable problems in high-dimensional space. A single kernel function is often used in SVDD classification. However, the highly complex fault data resulting from the volatile operating environment of harmonic reducers makes it difficult to achieve effective results with SVDD models constructed using a single kernel function.

[0123] According to the properties of kernel functions, multiple kernel functions are combined in a linearly weighted manner. The constructed multi-core kernel function still satisfies Mercer's theorem. The construction method of the multi-core kernel function is shown in formula (11):

[0124]

[0125] Where: K mix Represents the combination of multiple kernel functions, M represents the total number of multi-core kernel functions; μ m Represents the weight coefficient, satisfying μ m ≥0, at the same time

[0126] The Gaussian radial basis kernel function has a stronger ability to describe sample distribution than other common kernel functions. It can have a greater impact on the input vector in a relatively narrow range and can better adapt to the highly complex situation of harmonic reducer fault data. The definition of the Gaussian radial basis kernel function is shown in formula (12):

[0127]

[0128] Where: σ is the kernel function width parameter.

[0129] The multi-kernel function is constructed by weighted summing of two Gaussian radial basis kernel functions with different kernel widths to improve the learning and generalization capabilities of the multi-kernel function. The specific form is shown in formula (13):

[0130]

[0131] Among them, K rbf represents the Gaussian radial basis kernel function. σ1 and σ2 represent the two different kernel width parameters of the kernel function. To control the weights of the two kernel functions, the weight coefficient μ is set in the range of 0 to 1, that is, μ∈[0,1].

[0132] 3 Quantitative assessment method and process of harmonic reducer health status

[0133] The block diagram of the multi-state and same-scale quantitative evaluation method for harmonic reducers based on improved residual network combined with multi-core support vector data description is as follows: Figure 5 As shown, the specific steps are:

[0134] (1) Data acquisition: Build an experimental platform for collecting acoustic emission signals of harmonic reducers and install harmonic reducers that are either in normal operation or have certain fault types. Use the host computer software to control the drive motor to simulate the actual operating state of the harmonic reducer and complete the acoustic emission signal collection work for harmonic reducers with different fault locations and fault severity.

[0135] (2) Data preprocessing: The one-dimensional acoustic emission signal of the harmonic reducer is transformed into a two-dimensional time-frequency image by continuous wavelet transform to obtain richer time-frequency information of the sample.

[0136] (3) Sample set construction: A dataset of time-frequency graphs of acoustic emission signals from harmonic reducers with different fault locations and fault severity is selected and divided according to a certain ratio. These datasets are used to train the feature extraction network and test its performance. Data of unknown health status of the harmonic reducer is collected and preprocessed to construct a validation set.

[0137] (4) Fault feature extraction network construction: The convolutional attention mechanism is introduced to improve the residual network and extract the deep features of the training set. After a certain number of iterations, the model loss function value is calculated and the weight of each parameter is calculated using the backpropagation method. The model parameters are updated using the gradient descent method to minimize the loss function and finally converge the objective function to build the fault feature extraction network of the harmonic reducer.

[0138] (5) Construction of a multi-state quantitative evaluation model: Input the training set samples into the trained feature extraction network, save the features corresponding to all correctly classified samples, and normalize the features. Use the feature vectors in the normal state as training samples to train the multi-MKSVDD evaluation model, obtain the radius R of the hypersphere, and use the normal MKSVDD as the benchmark to calculate the distance from the feature vectors of different fault levels at a certain location to the center of the normal MKSVDD sphere. By fitting the distances of different fault levels relative to the center of the normal sphere, a quantitative evaluation curve is obtained. By introducing the relative compensation distance, a quantitative evaluation curve under the same scale in multiple states is obtained.

[0139] (6) Quantitative evaluation of health status: The samples of the validation set are input into the trained fault feature extraction network to extract the deep features of the fault data. The features are then normalized and input into the MKSVDD evaluation model to obtain the distance D between the sample to be tested and the center of the hypersphere. If D≤R, it means that the sample to be tested is in a normal state, otherwise it indicates that it is in a fault state. Then, the quantitative evaluation curve is combined to realize the quantitative analysis of the fault degree and obtain the final quantitative evaluation result of the health status of the harmonic reducer.

[0140] 4 Application and Analysis

[0141] 4.1 Experimental bench construction and experimental environment construction

[0142] The experimental data of this invention are all from the harmonic reducer acoustic emission signal acquisition experimental platform built in the laboratory. The actual platform is shown in the figure. Figure 6 shown.

[0143] The measurement and control system of the harmonic reducer acoustic emission signal acquisition experimental platform is mainly composed of a motor control box, a drive motor, a torque and speed sensor, a reducer to be tested, a load motor, an acoustic emission sensor and a data acquisition instrument. The schematic diagram is shown in the figure. Figure 7 As shown. The acoustic emission sensor can well detect the internal deformation process of mechanical materials and is not affected by environmental noise. During the operation of the harmonic reducer, the structure containing the fault will rotate. In order to effectively obtain the complete state information within a cross section of the harmonic reducer, the data acquisition uses the American Physical Acoustics Company (PAC)'s highly sensitive and reliable PAC-R15 acoustic emission sensor, which has a stable frequency response in the range of 400 to 1000kHz, so that the fault frequency characteristics of the acoustic emission signal can be studied more accurately. Use coupling agent to fill the gap between the sensor and the surface of the structure, and fix the acoustic emission sensor with a magnetic clamp. The acoustic emission sensor is deployed at the 12 o'clock position of the harmonic reducer, and the acoustic emission signals of different fault locations and different fault degrees are collected at a sampling frequency of 1000kHz, as shown Figure 8 shown.

[0144] 4.2 Introduction to Experimental Dataset

[0145] Taking the harmonic LCSG-32-50-CI type harmonic reducer as an example, the present invention completes the processing of different fault locations and different fault degrees of the harmonic reducer from two aspects: the structure of the harmonic reducer and the actual failure situation. The data includes 8 types of states, including 1, 5, 10 and 15 laser pitting faults on the inner ring of the flexible bearing, 1, 10 and 15 laser pitting faults on the outer ring of the flexible bearing, and normal state. To facilitate the subsequent presentation of the harmonic reducer data, the 8 types of states of the harmonic reducer are now set, and the specific settings are shown in Table 1. Taking IR5 and OR10 as examples, they represent 5 laser pitting faults on the inner ring of the flexible bearing of the harmonic reducer and 10 laser pitting faults on the outer ring of the flexible bearing of the harmonic reducer, respectively.

[0146] Table 1 Representation of data of different fault states

[0147]

[0148] During data preprocessing, to ensure that the sample contains complete fault information, two revolutions of the harmonic reducer's flexible bearing were used as a dataset sample. Wavelet transform was then performed to generate a time-frequency graph. The 400 time-frequency graphs of each category in the harmonic reducer acoustic emission signal dataset listed in Table 1 were divided into a 3:1 ratio and used to train and test the feature extraction network, respectively. The deep features of all correctly classified samples were used to train the condition assessment model.

[0149] The proposed method adopts the ReLU activation function, the number of iterations epochs in the training process is set to 30, and the learning rate lr is set to 0.0005.

[0150] 4.3 Comparative Experiments before and after Improvement of Residual Network

[0151] In order to explore the effectiveness of introducing the convolutional attention module in the fault diagnosis of harmonic reducer, the original residual network and the improved residual network are used to extract the deep features of the time-frequency graph of the harmonic reducer acoustic emission signal, and the t-distribution neighborhood embedding algorithm is used to perform feature visualization. The feature visualization effect is as follows: Figure 9 and Figure 10 shown.

[0152] By observation Figure 9 and Figure 10As can be seen, both before and after the improvement of the residual network, there are instances of sample misclassification, but there are significant differences in the feature distributions between the two. After visualizing the deep features of the samples extracted by the original residual network, some samples are misclassified in D3, and there is also sample feature aliasing and blurred boundaries in D2 and D3. The improved residual network, on the other hand, only misclassifies a small number of samples in D1, and the category discrimination is relatively clear. This suggests that the improved residual network with the convolutional attention mechanism can better extract the deep features of the harmonic reducer acoustic emission signal, helping to improve the performance of the fault diagnosis model.

[0153] In order to further verify that the improved residual network can better extract deep features for fault diagnosis of harmonic reducer, the diagnosis results are plotted as confusion matrix as follows: Figure 11 and Figure 12 shown.

[0154] By observation Figure 11 and Figure 12 The original residual network was found to have difficulty distinguishing between fault samples B and OR10 when extracting features from harmonic reducer acoustic emission signals, and a small number of misclassified samples were found for OR15 and N. An improved residual network using a convolutional attention mechanism for extracting features from harmonic reducer faults significantly improved classification accuracy, despite a small number of misclassified samples for B and OR15. In summary, the improved residual network can focus on mining salient features in images and suppress ineffective features, thereby improving the accuracy of fault diagnosis for harmonic reducers.

[0155] 4.4 Comparative experiment of vibration signal and acoustic emission signal

[0156] In order to verify that acoustic emission signals are more sensitive to early faults than vibration signals, the vibration signals and acoustic emission signals corresponding to one laser pitting of the inner ring of a flexible thin-walled bearing are selected for analysis. The corresponding envelope spectrum is shown in the figure below. Figure 13 shown.

[0157] Through analysis Figure 13 It can be seen that when the inner ring of the flexible thin-walled bearing fails, both the vibration signal and the acoustic emission signal can extract fault information. However, for a minor pitting fault, the amplitude corresponding to the fault characteristic frequency in the envelope spectrum of the acoustic emission signal is more obvious than that of the vibration signal.

[0158] In order to further prove that the acoustic emission signal has the advantage of high sensitivity compared with the vibration signal and is more sensitive to early faults, the improved residual network is used to diagnose the vibration signal of the harmonic reducer. The diagnosis results are as follows: Figure 14 shown.

[0159] By comparison Figure 12 and Figure 14It was found that the overall accuracy of fault classification increased by 2.4%, among which the accuracy of pitting corrosion fault increased by 6%, which further verified that acoustic emission signals are more sensitive to early harmonic reducer fault information than vibration signals and can keenly capture subtle changes in the early stage of fault.

[0160] 4.5 Establishment of the health status assessment model of harmonic reducer

[0161] The constructed MKSVDD evaluation model needs to include most normal samples, but should not have too many constraints, and the choice of kernel function needs to ensure that the hypersphere is relatively stable. The grid search algorithm is used to optimize the hyperparameters, such as Figure 15 As shown in the figure, the weight coefficient μ is set to 0.3, the kernel width σ1 is set to 0.9 and σ2 is set to 1.5.

[0162] All deep features of 400 samples in the normal state are used to train MKSVDD. The distances of 400 samples of each type with different fault degrees in the inner circle relative to the center of the normal MKSVDD sphere are calculated, and the distance visualization is obtained as follows: Figure 16 shown.

[0163] analyze Figure 16 As can be seen, due to the presence of misclassified samples in each class, the relative distances obtained are aliased, which will lead to large errors in the subsequent evaluation curves and impair quantitative analysis. Therefore, the distance visualization will be performed using correctly classified samples in each state.

[0164] The deep features extracted by the improved residual network are saved in correspondence with the classified labels, and the features are normalized. The features corresponding to all the correctly classified normal state labels in the training set are used to train the MKSVDD evaluation model. Taking the normal MKSVDD model as the benchmark, the distance from the depth features of the vibration signal and acoustic emission signal of the inner ring with different fault degrees to the center of the sphere is calculated, as shown in the figure below: Figure 17 and Figure 18 shown.

[0165] By observation Figure 17 and Figure 18 The researchers found that as the inner race fault severity increased, the distance to the center of the sphere in the normal state exhibited a distinct step-like pattern, indicating that the distance metric can effectively reflect the severity of the inner race fault in harmonic reducer bearings. Furthermore, acoustic emission signals can effectively distinguish between normal and single pitting faults, while vibration signals show no clear distinction between normal and single pitting faults. This further demonstrates that acoustic emission signals are more sensitive to weak signals than vibration signals, making them more suitable for quantitative fault analysis.

[0166] In order to quantitatively evaluate the health status of the harmonic reducer, the relative distance of the inner ring is first averaged, and then a quadratic function is fitted for the distance index of different fault levels of the inner ring to obtain a smooth curve, such as Figure 19 shown.

[0167] Depend on Figure 19 It can be seen that the fitted evaluation curve shows a monotonically increasing trend, and the fault severity is positively correlated with the evaluation index, further demonstrating that the MKSVDD distance-based evaluation index is sensitive to changes in the harmonic reducer's fault severity. The evaluator can determine the harmonic reducer's fault location based on the diagnostic model, calculate the distance between the characteristic vector of the test sample and the center of the normal state sphere based on the evaluation model, and then use the evaluation curve to determine the current fault severity of the reducer, achieving a quantitative analysis of the harmonic reducer's operating status.

[0168] The same experimental steps are carried out using the data of different fault degrees of the outer ring to calculate the distance from the characteristics of different fault degrees of the outer ring to the center of the sphere, such as Figure 20 and Figure 21 shown.

[0169] 4.6 Multi-state and same-scale quantitative evaluation and analysis of harmonic reducers

[0170] from Figure 17-21 It can be seen that different fault severity levels can be distinguished for a single fault location, but if different fault locations are put together, some information is lost due to the propagation mechanism of the acoustic emission signal, making it impossible to uniformly evaluate the fault severity levels of different fault locations of the harmonic reducer.

[0171] Based on the propagation mechanism of the acoustic emission signal of the flexible thin-walled bearing of the harmonic reducer, as well as the structure and fault location of the flexible thin-walled bearing itself, the design uses a reasonable method to compensate for the lost information. Because the sensor is installed on a fixture that wraps the entire outer ring, it is assumed that the signal collected by the outer ring has almost no loss. The signal from the inner ring needs to pass through the outer ring before being transmitted to the acoustic emission sensor, so the relative distance between the outer ring and the inner ring needs to be compensated. The specific compensation scheme is shown in formula (14):

[0172]

[0173] Where: R OR is the relative radius of the hypersphere containing all fault severity samples in the outer circle, d IR is the relative compensation distance of the inner ring, d OR is the relative compensation distance of the outer ring.

[0174] The multi-state evaluation curve after compensation is as follows: Figure 22As shown, it presents an obvious step-like shape. The curve after the relative distance compensation can intuitively feel the multi-state information of the flexible thin-walled bearing of the harmonic reducer, which is convenient for the evaluator to perform visual evaluation at the same scale.

[0175] In order to verify the accuracy and effectiveness of the proposed diagnostic model and evaluation curve, 14 pitting faults of the inner ring and 5 pitting faults of the outer ring of the harmonic reducer were selected for evaluation experiments. First, the one-dimensional fault data was subjected to continuous wavelet transform to obtain a two-dimensional time-frequency graph, and then the two-dimensional time-frequency graph was placed in the trained fault diagnosis model to obtain the deep features of the time-frequency graph. The extracted feature vectors of the IR14 and OR5 faults were normalized and then placed in the normal state MKSVDD model. The distance from the fault vector to the center of the normal state sphere was calculated, and the corresponding distance was processed and the mean was calculated. The corresponding relative compensation distance was added to the inner and outer rings respectively, and then the obtained evaluation distance was quantitatively analyzed according to the multi-state evaluation curve to evaluate the pitting degree of the fault, such as Figure 23 shown.

[0176] observe Figure 23 It can be seen that the fault severity assessed by the model is 13.56 pitting corrosion times for the inner ring and 4.92 pitting corrosion times for the outer ring, respectively. This is close to the actual fault severity of 14 pitting corrosion times for the inner ring and 5 pitting corrosion times for the outer ring, with the maximum assessment error not exceeding 3.2%. This shows that the multi-state unified assessment curve can achieve quantitative analysis of the inner ring of the harmonic reducer, and the results are accurate and effective.

[0177] Although the present invention is disclosed as above, the scope of protection disclosed by the present invention is not limited thereto. Those skilled in the art of the present invention may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the scope of protection of the present invention.

[0178] References

[0179] [1] Zhou Meng, Li Jianyu, Wang Chang, et al. A review of multi-robot collaborative capture methods[J]. Acta Automatica Sinica, 2024, 50(12): 1-34.

[0180] ZHOUM,LIJY,WANGC,etal.Multi-robotcooperativehunting:ASurvey[J].ActaAutomaticaSinica,2024,50(12):1-34.

[0181] [2]JIAYZ,LIYQ,XUMQ,etal.Afaultdiagnosisscheme forharmonicreducerunderpracticaloperatingconditions[J].Measurement,2024,227:114234.

[0182] [3] Zhang Xian, Wang Peng, Zhang Changming, et al. Research on reliability evaluation of precision degradation failure and related competitive failure of harmonic reducer [J]. Vibration and Shock, 2024, 43(15): 277-287.

[0183] ZHANGX,WANGP,ZHANGCM,etal.Reliabilityevaluationforprecisiondegradationfailureanddependentcompetingfailure ofharmonicdrives[J].JournalofVibrationandShock,2024,43(15):277-287.

[0184] [4]ZHIZ,LIUL,LIUD T,etal.FaultdetectionoftheharmonicreducerbasedonCNN-LSTMwithanovel denoisingalgorithm[J].

[0185] IEEE SensorsJournal,2021,22(3):2572-2581.

[0186] [5] Chen Renxiang, Zhang Yong, Hu Xiaolin, et al. Health status assessment of harmonic reducer based on deep fusion of multi-sensor information[J]. Vibration and Shock, 2022, 41(07): 139-144+152.

[0187] CHENRX,ZHANGY,HUXL,etal.Healthstateevaluationofharmonicreducerbasedonmulti-sensorinformationdeep fusion[J].

[0188] Journal of Vibration and Shock,2022,41(07):139-144+152.

[0189] [6]YANG G, ZHONG Y, YANG L, et al. Fault diagnosis of harmonic drive with imbalanced data using generative adversarial network[J].

[0190] IEEE Transactions on Instrumentation and Measurement,2021,70:1-11.

[0191] [7] Chen Renxiang, Zhang Yong, Hu Xiaolin, et al. Health status identification of harmonic reducer based on deep feature learning of voltage signal [J]. Chinese Journal of Scientific Instrument, 2021, 42(07): 234-241.

[0192] CHEN RX,ZHANG Y,HU XL,et al.Health state recognition ofharmonicreducerbased on depth feature learning ofvoltage

[0193] signal[J].Chinese Journal ofScientific Instrument,2021,42(07):234-241.

[0194] [8] Kang Shouqiang, Zhang Weidong, Wang Yujing, et al. Fault diagnosis method of harmonic reducer under different working conditions based on information fusion subdomain adaptation [J]. Chinese Journal of Scientific Instrument, 2024, 45(03): 60-71.

[0195] KANG SQ,ZHANG WD,WANG YJ,et al.A fault diagnosis method for harmonic reducers under different operating conditions

[0196] based on information fusion subdmain adaptation[J].Chinese Journal ofScientific Instrument,2024,45(03):60-71.

[0197] [9]ZHOU X,ZHOU HC,HE YM,et al.Harmonic reducer in-situ faultdiagnosis for industrial robots based on deep learning[J].Science

[0198] China Technological Sciences,2022,65(9):2116-2126.

[0199]

[10] SUN YL,KANG SQ,WANG YJ,et al.Fault diagnosis method forharmonic reducerbased onpersonalized federated aggregation

[0200] strategy with skip cycle weight[J].Measurement,2025,242(PE):116275-116275.

[0201]

[11] Han Zhengjie, Niu Rongjun, Ma Zikui, et al. Bearing fault diagnosis method based on improved residual neural network based on attention mechanism [J]. Vibration and Shock, 2023, 42(16): 82-91.

[0202] HANG ZJ, NIU RJ, MA ZK, et al. Bearing fault diagnosis methods based on an attentional-mechanism-improved residual neural

[0203] network[J].Journal ofVibration and Shock,2023,42(16):82-91.

[0204]

[12] Kang Tao, Duan Rongkai, Yang Lei, et al. Bearing fault diagnosis method based on convolutional neural network integrated with multi-attention mechanism [J]. Journal of Xi'an Jiaotong University, 2022, 56(12): 68-77.

[0205] KANG T, DUAN RK, YANG L, et al. Bearing fault diagnosis using convolutional neural networkbased on a multi-attention

[0206] mechanism[J].Journal ofXi'An Jiaotong University,2022,56(12):68-77.

[0207]

[13] DU XJ, LIU BW, GAI JL, et al.Gear fault diagnosis based on smallchannel convolutional neural networkunder multiscale fusion

[0208] attention mechanism[J].Quality andReliability EngineeringInternational,2024,40(8):4147-4165.

[0209]

[14] LU KJ, CHEN C, WANG T, et al. Fault diagnosis ofindustrialrobotbased on dual-module attention convolutional neural network[J].

[0210] Autonomous Intelligent Systems,2022,2:1-12.

[0211]

[15] Lu Jin, Wang Zhiguo, Liu Fei. Rolling bearing variable load fault diagnosis based on STFT-ECA-ResNet18 network model[J]. Noise and Vibration Control, 2024, 44(02): 122-128.

[0212] LU J,WANG ZG,LIU F.Rollingbearing variable load fault diagnosis based on STFT-ECA-ResNet18 networkmodel[J].Noise and

[0213] Vibration Control,2024,44(02):122-128.

[0214]

[16] XU YD, CHEN YJ, ZHANG HC, et al.Global contextual feature aggregation networks with multiscale attention mechanism for

[0215] mechanical diagnosis fault under non-stationary conditions[J].Mechanical Systems and Signal Processing,2023,203:110724.

[0216]

[17] Yin Wenhao. Rolling bearing performance degradation assessment based on fusion probability modeling and boundary distance[D]. East China Jiaotong University, 2021: 4-5.

[0217] YIN W H.Performance degradation evaluation of rollingbearings based on fusionprobabilistic modeling and boundary distance[D].

[0218] East China Jiaotong University,2021:4-5.

[0219]

[18] Chen Peng, Zhao Xiaoqiang. Rolling bearing performance degradation evaluation method based on GLNPE-SVDD[J]. Journal of Huazhong University of Science and Technology (Natural Science Edition), 2021, 49(01): 12-16.

[0220] CHEN P,ZHAO X Q.Performance degradation evaluation methodofrollingbearingbased on GLNPE-SVDD[J].J.Huazhong Univ.

[0221] ofSci.&Tech.(Natural Science Edition),2021,49(01):12-16.

[0222]

[19] XIN HW, ZHANG HD, YANG YJ, et al.Evaluation of rollingbearing performance degradation based on comprehensive index

[0223] reduction and SVDD[J].Machines,2022,10(8):677-677.

[0224]

[20] Cheng Daolai, Wei Tingting, Pan Yuna, et al. Rolling bearing performance degradation assessment method based on DBN-SVDD[J]. Bearings, 2021, 10:41-46.

[0225] CHEN DL,WEI TT,PAN YN,et al.Performance degradation assessment method for rollingbearings based onDBN-SVDD[J].

[0226] Bearing,2021,10:41-46.

[0227]

[21] Kang Yuxiang, Chen Guo, Wei Xunkai, et al. Deep residual hedging network and its application in rolling bearing fault diagnosis[J]. Acta Aeronautica Sinica, 2022, 43(08): 63-74.

[0228] KANG YX,CHEN G,WEI XK,et al.Deep residual hegingnetwork and its application in fault diagnosis of rollingbearings[J].Acta

[0229] Aeronautica et Astronautica Sinica,2022,43(08):63-74.

[0230]

[22] Qin Guohao, Zhang Kai, Ding Kun, et al. Bearing fault diagnosis method based on dynamic wide convolutional residual network[J]. China Mechanical Engineering, 2023, 34(18): 2212-2221.

[0231] QIN G H,ZHANG K,DING K,et al.Dynamic wide convolutional residualnetwork forbearing fault diagnosis method[J].China

[0232] Mechanical Engineering,2023,34(18):2212-2221.

[0233]

[23] PANG Y N,CHEN D L,WEI T T,et al.Rolling bearingperformancedegradation assessmentbased on deep beliefnetwork and

[0234] improved support vector data description[J].Mechanical Systems andSignal Processing,2022,181:109458.

Claims

1. A quantitative evaluation method for industrial robot harmonic reducers in multiple states and at the same scale, characterized by: The following steps are involved: S100: Data collection: collecting one-dimensional acoustic emission signals of harmonic reducers with different fault locations and fault degrees; S200, data preprocessing, performing continuous wavelet transform on the one-dimensional acoustic emission signal collected in step S100 to obtain a two-dimensional time-frequency image; S300, constructing a sample set, dividing the two-dimensional time-frequency image obtained after the transformation in step S200 for training the feature extraction network and testing the performance of the feature extraction network; collecting data of unknown health status of the harmonic reducer, and constructing a validation set after data preprocessing; S400: Build a fault feature extraction network, introduce a convolutional attention mechanism to improve the residual network, and extract deep features of the training set; After iteration, the model loss function value is calculated and the back propagation method is used to calculate the weight of each parameter. The model parameters are updated using the gradient descent method to minimize the loss function and finally converge the objective function to obtain the established harmonic reducer fault feature extraction network. S500: Construct a multi-state quantitative evaluation model, input the training set into the fault feature extraction network built in step S400, save and normalize the features corresponding to all correctly classified samples; use the feature vectors in the fault-free state as training samples, train the multi-MKSVDD evaluation model, obtain the radius R of the hypersphere, use the MKSVDD in the fault-free state as a benchmark, calculate the distance from the feature vectors of different fault degrees at a certain position to the center of the normal MKSVDD sphere, obtain a quantitative evaluation curve by fitting the distances of different fault degrees relative to the center of the normal sphere, introduce the relative compensation distance, and obtain a quantitative evaluation curve under the same scale in multiple states; S600, quantitative assessment of health status, input the verification set into the multi-state quantitative assessment model trained in step S500, and obtain the distance D between the sample to be tested and the center of the hypersphere. If D≤R, it means that the sample to be tested is in a normal state, otherwise it indicates that the sample to be tested is in a faulty state. Then, the quantitative assessment curve is combined to achieve a quantitative analysis of the degree of fault and obtain the final quantitative assessment result of the health status of the harmonic reducer.

2. The quantitative evaluation method for an industrial robot harmonic reducer in multiple states and at the same scale according to claim 1 is characterized by: In step S400, the introduced convolutional attention mechanism includes channel mechanism and spatial attention mechanism. First, calculate the channel attention of the feature map and obtain the channel attention weighted feature map: Where: φ is the sigmoid activation function; W0 and W1 are the two hidden layers in the multi-layer perceptron model; Represents the average pooling feature of the channel dimension; Represents the maximum pooling feature of the channel dimension; Then, spatial attention calculation is performed on this basis to obtain the spatial attention weighted feature map: Where: φ is the sigmoid activation function; f 7×7 It is an image convolution operation with a convolution kernel size of 7×7; Represents the average pooled features of the spatial dimension; Represents the maximum pooling feature of the spatial dimension; Finally, we get a cascade of channel attention mechanism and spatial attention mechanism, which can be expressed as formula (3) and formula (4): Where: X represents the input feature map; X1 represents the feature map after channel attention optimization; X2 represents the feature map after channel and spatial attention optimization; M c represents the one-dimensional convolution of the channel attention module; M s Represents the two-dimensional convolution of the spatial attention module; Represents pixel-by-pixel multiplication.

3. The quantitative evaluation method for an industrial robot harmonic reducer in multiple states and at the same scale according to claim 1 is characterized by: In step S400, a convolutional attention mechanism is embedded into the tail end of the residual network; information features are extracted through the convolution operation, and the extracted features are passed through the convolutional attention mechanism, allowing the network to adaptively obtain channel information and spatial information related to the fault state of the harmonic reducer to obtain feature representation; The output of the residual network based on the channel attention mechanism and the spatial attention mechanism is expressed as: H(f(x))+x(5) Where: x is the input; f is the residual function; H is the attention function of the convolutional attention mechanism.

4. The quantitative evaluation method for an industrial robot harmonic reducer in multiple states and at the same scale according to claim 1 is characterized by: In step S500, when training the MKSVDD evaluation model, For the target class sample set M={x1,x2,...,x N }, find the optimal hypersphere, which is described by the center a and radius R, and satisfies the optimization function shown in formula (6): Where: C is the penalty parameter, which controls the trade-off between hypersphere error and algorithm complexity; ξ i is a slack variable, allowing some training data to be outside the hypersphere; c is the center of the hypersphere; Introducing the Lagrange multiplier α i , transform the optimization function of formula (6) into the Lagrangian extreme value problem as shown in formula (7): Among them, a j is an intermediate variable; The kernel function is used to transform the samples from linearly inseparable low-dimensional space to linearly separable high-dimensional space. In the mapping space, the radius of the hypersphere R is determined by any support vector x. k The distance to the center is calculated as: Where: K(·) is the kernel function, representing the inner product operation; For the new sample y, its distance from the center of the sphere is shown in formula (9):

5. The quantitative evaluation method for industrial robot harmonic reducers in multiple states and at the same scale according to claim 4 is characterized in that: In step S500, multiple kernel functions are combined in a linear weighted manner. The construction method of the multi-kernel kernel function is shown in formula (10): Where: K mix Represents the combination of multiple kernel functions, M represents the total number of multi-core kernel functions; μ m Represents the weight coefficient, satisfying μ m ≥0, at the same time The Gaussian radial basis kernel function is used to describe the sample distribution. The definition of the Gaussian radial basis kernel function is shown in formula (11): Where: σ is the kernel function width parameter; The multi-kernel kernel function is constructed by weighted summing of two Gaussian radial basis kernel functions with different kernel widths to improve the learning and generalization capabilities of the multi-kernel kernel function. The specific form is shown in formula (12): Where: K rbf Represents the Gaussian radial basis kernel function; σ1 and σ2 represent two different kernel width parameters of the kernel function; in order to control the weights of the two kernel functions, the weight coefficient μ is taken in the range of 0 to 1, that is, μ∈[0,1].

6. The quantitative evaluation method for industrial robot harmonic reducers in multiple states and at the same scale according to claim 1 is characterized by: When different fault locations are put together, the inner and outer ring signals are compensated, as shown in formula (13): Where: R OR is the relative radius of the hypersphere containing all fault severity samples in the outer circle, d IR is the relative compensation distance of the inner ring, d OR is the relative compensation distance of the outer ring.

7. A quantitative evaluation system for industrial robot harmonic reducers in multiple states and at the same scale, characterized by: The system has a program module corresponding to the steps of any one of claims 1 to 6 above, and executes the steps of the above-mentioned quantitative evaluation method for a harmonic reducer of an industrial robot in multiple states and at the same scale during operation.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of a quantitative evaluation method for an industrial robot harmonic reducer in multiple states and at the same scale according to any one of claims 1 to 6 when called by a processor.

Images