Current signal-based method for evaluating health state of industrial robot

Abstract

The present invention relates to a current signal-based method for evaluating the health state of an industrial robot. A current sensor is used to detect total life current data of each axis of an industrial robot, the data is preprocessed and arranged in time sequence, the arranged data is reconstructed, interference signals are removed, the reconstructed data is fed into a model for identifying health data and non-health data, the identified data is respectively fed into models for training, a degradation baseline and a fault baseline are acquired, and then the baselines are introduced into an online model to determine the health state of the industrial robot in real time. By means of the method of the present invention, the health state of an industrial robot can be detected in real time and an alarm is given, so that a field technician can grasp the operating state of the industrial robot conveniently and in a timely manner, thereby promoting the safe and smooth progress of an engineering project.

Description

Industrial robot health status evaluation method based on current signal Technical Field

[0001] The present invention relates to an industrial robot fault technology, and in particular to an industrial robot health status evaluation method based on current signals. Background Art

[0002] With the advancement of industrial production technology, industrial robots are becoming increasingly common in production. In highly automated production systems, failure of industrial robots can lead to economic losses and even endanger human safety. Therefore, using models to monitor the operating status of industrial robots in real time can eliminate potential risks and prevent major safety incidents.

[0003] Due to the complexity of current industrial robot-based production systems, it is difficult to establish accurate mathematical models. However, with the rapid advancement of sensing technology in recent years, sensors can accurately capture large amounts of industrial robot operating data. By deeply mining these data features, we can obtain information about equipment degradation and failures.

[0004] Currently, companies generally require regular manual inspection and maintenance of industrial robots during normal operation. This manual method requires specialized technicians to assess on-site conditions, which is time-consuming, labor-intensive, and inefficient. Industrial robot malfunctions are often accompanied by fluctuations in signals such as current, vibration, and acoustic emission. Therefore, modeling based on sensor data from these sensors is common. However, varying installation locations for vibration and acoustic emission sensors can cause signal variations, making it difficult to determine the optimal placement. Mounting vibration and acoustic emission sensors on the surface of industrial robots is intrusive, and their signal cables can affect the robot's normal operation. Furthermore, ambient noise in the production environment can introduce unpredictable noise into vibration and acoustic emission signals. In contrast, current sensors are inexpensive, non-invasive, and unaffected by ambient noise when installed in the robot's control cabinet. Therefore, current signals are chosen as the raw data source for modeling.

[0005] A method based on PCA-T is currently disclosed. 2 The health score calculation method of industrial robots is proposed. This method reconstructs the current signal of industrial robots by PCA and calculates T 2 Statistics are used as control limits. The robot fault detection and health score calculation are well implemented, but there is still some room for improvement.

[0006] Summary of the Invention

[0007] To address the issue of health status evaluation of industrial robots, manual inspection is time-consuming, labor-intensive and inefficient. Based on the fact that traditional vibration signal analysis is not suitable for the actual operating conditions of industrial robots, a health status evaluation method for industrial robots based on current signals is proposed to overcome the shortcomings of manual and vibration and acoustic emission signal analysis, and improve the accuracy and real-time performance of health status evaluation.

[0008] The technical solution of the present invention is: a method for evaluating the health status of an industrial robot based on current signals, which specifically includes the following steps:

[0009] 1) Use current sensors to detect the full life current data of each axis of the industrial robot. The data includes multiple non-fault cycles and fault cycles. After the data is collected, it is pre-processed;

[0010] 2) Using the period of current data as the window and step length, extract the time domain, frequency domain, and time-frequency domain features of each window respectively, and combine them into 32-dimensional original time series features;

[0011] 3) Using the feature reconstruction model to reconstruct the original time series features obtained in step 2), the reconstruction is to retain the essential characteristics of the original data while removing the low-dimensional signals of the interference signal in the high-latitude signal as the input features of the degradation model;

[0012] 4) Based on the input features obtained in step 3), an industrial robot degradation model is constructed, and the results are divided into three stage types: healthy, degraded, and faulty. The model is trained using healthy data and non-fault data including healthy and degraded data, respectively. The trained model is tested using full lifecycle data to obtain degradation baselines and fault baselines, thereby generating the industrial robot health status and fault alarm information for industrial robot health status evaluation.

[0013] Furthermore, the pretreatment in step 1) includes:

[0014] 1.1) Use the autocorrelation method to calculate its period, and the formula satisfies:

[0015] Where R x (τ) is mainly used to describe the similarity between the current signal x(t) and its own x(t+τ) after a delay of τ. The higher the similarity, the higher the correlation value. The coordinate with the highest correlation value is the periodic point of the signal, and the corresponding delay time τ is the period. Because the sampling value is discrete, the approximate calculation formula satisfies:

[0016] Where M is the maximum computational delay and N is the number of data points.

[0017] 1.2) The acquisition of industrial robot current signals is discontinuous, and the initial phases of data collected at different time periods may not be consistent. Therefore, it is necessary to phase align the data at all acquisition points. The cross-correlation method is used for alignment, and the formula satisfies:

[0018] Where R xy (τ) is mainly used to describe the similarity between the current signal x(t) and the current signal y(t) delayed by τ (y(t+τ). The higher the similarity, the higher the cross-correlation value. The coordinate with the highest cross-correlation value is the number of phase points that need to be moved. Because the sampling value is discrete, the following formula is used for approximate calculation:

[0019] Where M is the maximum computational delay and N is the number of data points.

[0020] Furthermore, the time domain and frequency domain features extracted in step 2) include mean, absolute mean, variance, root mean square value, root square amplitude, absolute maximum, maximum, minimum, peak-to-peak value, skewness, kurtosis, crest factor, margin factor, shape factor, pulse factor, average frequency, center frequency, root mean square frequency, standard deviation frequency, frequency concentration, and frequency kurtosis.

[0021] Furthermore, in step 2), the time-frequency domain feature extraction uses an ensemble empirical mode decomposition algorithm to decompose the current signal of each cycle; the ensemble empirical mode decomposition algorithm is based on the empirical mode decomposition algorithm, and then Gaussian white noise with zero mean and equal variance is added to perform several modal decompositions, and the intrinsic mode component IMF obtained each time is averaged to improve the modal aliasing phenomenon of the EMD algorithm;

[0022] For the decomposed IMF components, the first three are taken to extract energy entropy, kurtosis, and variance contribution rate features, forming a total of 9-dimensional features; the calculation method of kurtosis is consistent with that of time domain signals, and the variance contribution rate is the ratio of the variance of the IMF component to the variance of the original signal. The calculation formula of energy entropy satisfies:

[0023] Where |imf i (t)| represents the absolute value of the i-th IMF component.

[0024] Furthermore, in the reconstruction method in step 3), the model used for feature reconstruction is a sparse stacked autoencoder SSAE, and its training calculation formula satisfies:

[0025] Where w represents the neuron weight of each layer of the SSAE network, b is the bias of the neuron; the first term is the mean squared error, which describes the input x i and reconstructed input The difference between the two, the number of input data is k; in the second term, m is the number of hidden layer units, α is the sparsity penalty factor, and ρ is the sparsity parameter. is the average activation of the j-th hidden layer neuron, represents the KL divergence; the third term in the formula is the weight decay term, and β is the coefficient of weight W decay, which helps prevent overfitting of SSAE;

[0026] Model training method: the dimension of the original feature is 32 feature dimensions. The original feature is used as the encoder input of SSAE, and the encoder output of SSAE is taken as the input feature of the subsequent degradation model; SSAE has a total of 5 encoders and 5 decoders, and the number of neurons in each layer from front to back is set to 32-16-8-4-2-1-2-4-8-16-32, that is, the first layer encoder compresses the original feature to 16 dimensions in feature dimension, the second layer encoder compresses the first layer output to 8 dimensions, the third layer encoder compresses the second layer output to 4 dimensions, the fourth layer encoder compresses the third layer output to 2 dimensions, and the fifth layer encoder compresses the fourth layer output to 1 dimension, that is, the final output feature is 1 dimension; the activation function of the last layer uses the Sigmoid function, and the remaining layers use the Relu activation function; the gradient descent algorithm is used to perform greedy unsupervised training on SSAE layer by layer, the mean square error function is selected as the loss function, and L2 regularization is added to improve the robustness of the model.

[0027] Furthermore, the specific implementation method of step 4) is as follows:

[0028] 4.1) After obtaining the input features, a model is constructed to distinguish different health states. The degradation model used is Support Vector Data Description (SVDD). SVDD builds a minimal hypersphere that contains as much non-faulty data as possible.

[0029] 4.2) Use healthy data features for training and full life data features for testing to build model 1; then use non-fault data features for training and full life data features for testing to build model 2;

[0030] Obtain the degradation and fault baselines and calculate the distance from the feature point to the sphere center. If the following formula is satisfied, the data is considered abnormal; otherwise, it is considered normal data:

[0031] Where z is the new sample point; x (i) 、x (j) are the training samples, which are all characteristic components of the sphere center a; i is the Lagrangian operator, a is the center position of the hypersphere, R 2 is the radius of the hypersphere;

[0032] 4.3) The degradation process of the industrial robot is measured using the distance from the feature point to the center of the hypersphere as a health indicator. The degradation baseline and fault baseline obtained in Model 1 and Model 2 are introduced into the online model to judge the health status of the industrial robot in real time.

[0033] Furthermore, the current sensor is installed on the signal line between the industrial robot control cabinet and the body to collect current signals in non-fault cycles and current signals in fault cycles.

[0034] A method for confirming degradation points and fault points of an industrial robot is disclosed. The method uses full-life current data of different machines to process and train according to the current signal-based industrial robot health status evaluation method, simulates the degradation trend of the industrial robot, and confirms the degradation baseline and fault baseline. The intersection of the degradation baseline and the degradation curve of the operating industrial robot is the degradation starting point; the intersection of the fault baseline and the degradation curve of the operating industrial robot is the fault starting point.

[0035] The beneficial effects of the present invention are that the industrial robot health status evaluation method based on current signals can detect and alarm the health status of the industrial robot in real time, so that on-site technicians can grasp the operating status of the industrial robot in a timely and convenient manner, which is conducive to promoting the safe and smooth progress of engineering projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] FIG1 is a flow chart of an implementation of a method for evaluating the health status of an industrial robot based on current signals according to the present invention;

[0037] FIG2 is a diagram showing the effect of the degradation curve of Model 1 using the implementation method of the present invention;

[0038] FIG3 is a diagram showing the effect of the degradation curve of Model 2 using the implementation method of the present invention. DETAILED DESCRIPTION

[0039] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0040] FIG1 is a flow chart of the implementation of the method for evaluating the health status of an industrial robot based on current signals according to the present invention. As shown in FIG1 , the health status evaluation method according to this example is divided into the following steps:

[0041] Step S1: Use a current sensor to detect the full life current data of each axis of the industrial robot. The data includes multiple non-fault cycles and fault cycles. After the data is collected, pre-process the data.

[0042] The current sensor is installed on the signal line between the control cabinet and the body of the industrial robot. In this embodiment, a total of 250 current signals of non-fault cycles and 37 current signals of fault cycles are collected.

[0043] The data preprocessing method specifically comprises the following steps:

[0044] Step S11: Since the industrial robot performs the same action repeatedly on the production line, the collected current signal is a periodic signal. When preprocessing the data, it is necessary to first calculate the period of the signal. The autocorrelation method is used to calculate its period, and the formula satisfies:

[0045] Where R x (τ) is mainly used to describe the similarity between the current signal x(t) and its own x(t+τ) after a delay of τ. The higher the similarity, the higher the correlation value. Generally, the coordinate with the highest correlation value is the periodic point of the signal, and the corresponding delay time τ is the period. Because the sampling value is discrete, the approximate calculation formula satisfies:

[0046] Where M is the maximum computational delay and N is the number of data points.

[0047] Step S12: The acquisition of the industrial robot current signal is discontinuous, and the initial phases of the data collected in different time periods are not necessarily consistent. Therefore, it is necessary to perform phase alignment on the data of all acquisition points. The cross-correlation method is used for alignment, and the formula satisfies:

[0048] Where R xy (τ) is mainly used to describe the similarity between the current signal x(t) and the current signal y(t) delayed by τ (y(t+τ). The higher the similarity, the higher the cross-correlation value. Generally, the coordinate with the highest cross-correlation value is the number of phase points that need to be moved. Because the sampling values ​​are discrete, the following formula is generally used for approximate calculation:

[0049] Where M is the maximum computational delay and N is the number of data points.

[0050] By calculation, we can know how many data points of the two current signals x(t) and y(t) need to differ in order for them to completely overlap.

[0051] Step S2: Using the period of the current data as the window and step length, extract the time domain, frequency domain, and time-frequency domain features of each window and combine them into 32-dimensional original time series features. This specifically includes the following steps:

[0052] Step S21: Extraction of time domain and frequency domain features includes mean, absolute mean, variance, root mean square value, root square amplitude, absolute maximum, maximum, minimum, peak-to-peak value, skewness, kurtosis, crest factor, margin factor, shape factor, impulse factor, average frequency, center frequency, root mean square frequency, standard deviation frequency, frequency concentration, and frequency kurtosis;

[0053] Step S22: Time-frequency domain feature extraction uses the Ensemble Empirical Mode Decomposition (EEMD) algorithm to decompose the current signal of each cycle. This algorithm adds Gaussian white noise with zero mean and equal variance to the Empirical Mode Decomposition (EMD) algorithm, performs several modal decompositions, and averages the intrinsic mode function (IMF) obtained each time. This can effectively improve the modal aliasing phenomenon of the EMD algorithm.

[0054] For the decomposed IMF components, the first three are taken to extract energy entropy, kurtosis, and variance contribution rate features, forming a total of 9-dimensional features. The calculation method of kurtosis is the same as that of time domain signals. The variance contribution rate is the ratio of the variance of the IMF component to the variance of the original signal. The calculation formula of energy entropy satisfies:

[0055] Where |imf i (t)| represents the absolute value of the i-th IMF component.

[0056] Step S3: The original features contain a lot of redundant information. The feature reconstruction model is used to reconstruct the original features into a suitable dimension as the input features of the degradation model. The specific steps include:

[0057] Step S31: The model used for feature reconstruction is the Sparse Stacked Autoencoder (SSAE). By taking the secular information as the learning target, the input information is represented and learned. It is an unsupervised learning algorithm used to learn the feature representation of high-dimensional data. The input layer information is encoded and sent to the hidden layer for training and fitting, and then decoded and sent to the output layer. The output low-dimensional signal includes the essential features of the original data while removing the interference part of the high-dimensional signal. Its training calculation formula satisfies:

[0058] Where w is the weight of each layer of SSAE network neurons, b is the bias of the neurons; the first term is the mean square error, which describes the input x i and reconstructed input The difference between the two, the number of input data is k; in the second term, m is the number of hidden layer units, α is the sparsity penalty factor, and ρ is the sparsity parameter. is the average activation of the j-th hidden layer neuron, represents the KL divergence; the third term in the formula is the weight decay term, and β is the coefficient of weight W decay, which helps prevent overfitting of SSAE;

[0059] Step S32: The original features have a dimension of 287*32, where 287 represents the sample dimension and 32 represents the feature dimension. The original features are used as the encoder input for the SSAE, and the encoder outputs are used as the input features for the subsequent degradation model. The SSAE has five encoders and five decoders, with the number of neurons in each layer from front to back set to 32-16-8-4-2-1-2-4-8-16-32. Specifically, the first encoder layer compresses the original features to 16 dimensions, the second encoder layer compresses the first layer output to 8 dimensions, the third encoder layer compresses the second layer output to 4 dimensions, the fourth encoder layer compresses the third layer output to 2 dimensions, and the fifth encoder layer compresses the fourth layer output to 1 dimension, resulting in a final output feature of 1 dimension. The activation function for the last layer is the Sigmoid function, and the remaining layers use the Reinforced Luminance (ReLU) activation function. The SSAE is trained layer-by-layer with greedy unsupervised gradient descent. The mean squared error (MSE) loss function is used, with L2 regularization added to enhance model robustness.

[0060] Step S4: Based on the input features obtained in step S3, an industrial robot degradation model is constructed. The results are divided into three stages: healthy, degraded, and faulty. The model is trained using healthy data and non-faulty data (including healthy and degraded data). The trained model is tested using full lifecycle data to obtain degradation baselines and fault baselines, thereby generating information such as the industrial robot's health status and fault alarms for industrial robot health status evaluation. The specific steps include:

[0061] Step S41: After obtaining the final input features, a model needs to be constructed to distinguish different health states. The degradation model used is Support Vector Data Description (SVDD). SVDD is a classic single-classification model suitable for actual industrial conditions with an imbalance between normal and abnormal data. Its basic idea is to establish a minimal hypersphere that contains as much non-fault data as possible.

[0062] Step S42: Use healthy and full-life data features to build Model 1; then use non-fault and full-life data features to build Model 2; Model 1 is trained with healthy data and tested with full-life data; Model 2 is trained with healthy + degradation data and tested with full-life data.

[0063] Obtain the degradation and fault baselines and calculate the distance from the feature point to the center of the sphere. If the following formula is satisfied, the data is considered abnormal; otherwise, it is considered normal data:

[0064] Where z is the new sample point; x (i) 、x (j) are the training samples, which are all characteristic components of the sphere center a; i is the Lagrangian operator; a is the center position of the hypersphere; R 2 is the radius of the hypersphere;

[0065] The degradation process of the industrial robot is measured using the distance between the feature point and the center of the hypersphere as a health indicator. Introducing the degradation baseline fault baseline obtained in Model 1 and Model 2 into the online model can determine the health status of the industrial robot in real time.

[0066] The method of the present invention can combine the expert knowledge of on-site technicians to determine the operating time range of health data, and use the health data of different machines to train model 1 in large quantities to obtain similar effects as shown in Figure 2; use the non-fault (including healthy and degraded) data of different machines to train model 2 in large quantities to obtain similar effects as shown in Figure 3.

[0067] As shown in Figure 2, this technology can effectively simulate the degradation trend of industrial robots. Based on the training results of healthy data, the degradation starting point of the device can be determined, namely the intersection of the degradation baseline and the degradation curve in Figure 2. The degradation starting point can also be flexibly adjusted according to actual field conditions. Based on the training results of non-fault data, the failure starting point of the device can be determined, namely the intersection of the failure baseline and the degradation curve in Figure 3. The failure starting point can also be flexibly adjusted according to actual field conditions.

[0068] The method of the present invention can provide a great reference for on-site technicians. When the model indicates that the equipment is running in the healthy area, there is no need to inspect and maintain the equipment. When the model indicates that the equipment is running in the degraded area, reasonable inspection and maintenance should be arranged for the equipment based on the application scenario, indicator characteristics and expert experience. When the model curve enters the fault area, it is very likely that an early fault has occurred, and on-site technicians need to inspect and repair the equipment.

[0069] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A method for evaluating the health status of an industrial robot based on current signals, characterized in that: The specific steps include: 1) Use current sensors to detect the full life current data of each axis of the industrial robot. The data includes multiple non-fault cycles and fault cycles. After the data is collected, pre-process the data; 2) Taking the period of current data as the window and step length, extract the time domain, frequency domain, and time-frequency domain features of each window respectively and combine them into 32-dimensional original time series features; 3) Using the feature reconstruction model to reconstruct the original time series features obtained in step 2), the reconstruction is to retain the essential features of the original data, while removing the low-dimensional signals of the interference signals in the high-latitude signals as the input features of the degradation model; 4) Based on the input features obtained in step 3), an industrial robot degradation model is constructed, and the results are divided into three stage types: healthy, degraded, and faulty. The model is trained using healthy data and non-fault data including healthy and degraded data, respectively. The trained model is tested using full life data to obtain degradation baselines and fault baselines, thereby generating the health status of the industrial robot and fault alarm information for the health status evaluation of the industrial robot.

2. The method for evaluating the health status of an industrial robot based on current signals according to claim 1, characterized in that: In the step 1), pretreatment include: 1.1) Use the autocorrelation method to calculate its period, and the formula satisfies: Where R x (τ) is mainly used to describe the similarity between the current signal x(t) and its own x(t+τ) after a delay of τ. The higher the similarity, the higher the correlation value. The coordinate with the highest correlation value is the periodic point of the signal, and the corresponding delay time τ is the period. Because the sampling value is discrete, the approximate calculation formula satisfies: In the formula, M is the maximum calculation delay, and N is the number of data points; 1.2) The acquisition of industrial robot current signals is discontinuous, and the initial phases of data collected in different time periods are not necessarily consistent, so it is necessary to phase align the data of all acquisition points. The cross-correlation method is used for alignment, and the formula satisfies: Where R xy (τ) is mainly used to describe the similarity between the current signal x(t) and the current signal y(t) delayed by τ (t+τ). The higher the similarity, the higher the cross-correlation value. The coordinate with the highest cross-correlation value is the number of phase points that need to be moved. Because the sampling value is discrete, the following formula is used for approximate calculation: Where M is the maximum calculation delay and N is the number of data points.

3. [Corrected 25.09.2024 according to Rule 26] The method for evaluating the health status of an industrial robot based on current signals according to claim 2, characterized in that: The extraction of time domain and frequency domain features in the step 2) includes mean, absolute mean, variance, root mean square value, root amplitude, absolute maximum value, maximum value, minimum value, peak-to-peak value, skewness, kurtosis, peak factor, margin factor, waveform factor, pulse factor, average frequency, center frequency, root mean square frequency, standard deviation frequency, frequency concentration, and frequency kurtosis.

4. The method for evaluating the health status of an industrial robot based on current signals according to claim 2 or 3, characterized in that: In the step 2), the time-frequency domain feature extraction uses an ensemble empirical mode decomposition algorithm to decompose the current signal of each cycle; the ensemble empirical mode decomposition algorithm is to perform several modal decompositions after adding Gaussian white noise with zero mean and equal variance on the basis of the empirical mode decomposition algorithm, and average the intrinsic mode component IMF obtained each time to improve the modal aliasing phenomenon of the EMD algorithm; For the decomposed IMF components, the first three are taken to extract the energy entropy, kurtosis, and variance contribution rate features, forming a total of 9-dimensional features; The calculation method of kurtosis is consistent with that of time domain signals. The variance contribution rate is the ratio of the IMF component variance to the original signal variance. The calculation formula of energy entropy satisfies: Where |imf i (t)| represents the absolute value of the i-th IMF component.

5. The method for evaluating the health status of an industrial robot based on current signals according to claim 4, characterized in that: Reconstruction method in step 3): The model used for feature reconstruction is a sparse stacked autoencoder SSAE, and its training calculation formula satisfies: Where w represents the neuron weight of each layer of the SSAE network, b is the bias of the neuron; the first term is the mean square error, which describes the input x i With refactored input The difference between the two, the number of input data is k; in the second term, m is the number of hidden layer units, α is the sparsity penalty factor, and ρ is the sparsity parameter. is the average activation of the jth hidden layer neuron, represents KL divergence; the third term in the formula is the weight decay term, and β is the coefficient of weight W decay, which helps prevent overfitting of SSAE; Model training method: The dimension of the original feature is 32 feature dimensions. The original feature is used as the encoder input of SSAE, and the encoder output of SSAE is taken as the input feature of the subsequent degradation model; SSAE has a total of 5 encoders and 5 decoders. The number of neurons in each layer from front to back is set to 32-16-8-4-2-1-2-4-8-16-32, that is, the first layer encoder compresses the original features to 16 dimensions in feature dimension, the second layer encoder compresses the first layer output to 8 dimensions, the third layer encoder compresses the second layer output to 4 dimensions, the fourth layer encoder compresses the third layer output to 2 dimensions, and the fifth layer encoder compresses the fourth layer output to 1 dimension, that is, the final output feature is 1 dimension; the activation function of the last layer uses Sigmoid function, and the remaining layers use Relu activation function; the gradient descent algorithm is used to perform greedy unsupervised training on SSAE layer by layer, the loss function selects the mean square error function and adds L2 regularization to improve the robustness of the model.

6. The method for evaluating the health status of an industrial robot based on current signals according to claim 5, characterized in that: The specific implementation method of step 4) is as follows: 4.1) After obtaining the input features, a model is constructed to distinguish different health states. The degradation model used is support vector data representation / description SVDD. SVDD is to establish a minimum hypersphere and use this hypersphere to contain all non-fault data as much as possible; 4.2) Use healthy data features for training and full life data features for testing to build model 1; then use non-fault data features for training and full life data features for testing to build model 2; Obtain the degradation and fault baselines, calculate the distance from the feature point to the center of the sphere, and if the following formula is satisfied, it is abnormal data, otherwise it is normal data: Where z is the new sample point; x (i) 、x (j) are the training samples, which are all characteristic components of the sphere center a; α i is the Lagrangian operator, a is the center position of the hypersphere, R 2 is the radius of the hypersphere; 4.3) The distance from the feature point to the center of the hypersphere is used as a health indicator to measure the degradation process of the industrial robot. The degradation baseline and fault baseline obtained in Model 1 and Model 2 are introduced into the online model to judge the health status of the industrial robot in real time.

7. The method for evaluating the health status of an industrial robot based on current signals according to claim 1, characterized in that: The current sensor is installed on the signal line between the industrial robot control cabinet and the body to collect current signals of non-fault cycles and current signals of fault cycles.

8. A method for confirming degradation points and fault points of an industrial robot, characterized in that: The full-life current data of different machines are processed and trained according to the industrial robot health status evaluation method based on current signals, the degradation trend of the industrial robot is simulated, the degradation baseline and the fault baseline are confirmed, and the intersection of the degradation baseline and the degradation curve of the running industrial robot is the degradation starting point; The intersection of the fault baseline and the degradation curve of the running industrial robot is the starting point of the fault.

Images