Intelligent edge side state data acquisition system, method and equipment for robot

Abstract

The invention discloses an intelligent edge side state data acquisition system, method and device for a robot, and relates to the technical field of edge computing and industrial big data. The intelligent edge side state data acquisition system comprises a device end, the device end is in butt joint with an edge end, the edge end is in butt joint with a cloud end, and the cloud end is in butt joint with the edge end. According to the intelligent edge side state data acquisition system, method and equipment for the robot, while the problems of data real-time processing and equipment monitoring on industrial big data are effectively solved, the data acquisition speed is further improved through a hardware enhancement mode, and the advantages of a cloud end and an edge end can be combined through a cloud-edge cooperation technology, so that the data acquisition efficiency is improved. According to the method, real-time data processing and equipment monitoring can be achieved, the capacity of large-scale calculation is achieved, data can be fully utilized, model migration can be achieved, and the high accuracy and the strong generalization capacity of the model are guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of edge computing and industrial big data technology, and specifically to an intelligent edge-side state data acquisition system, method and device for a robot. Background Art

[0002] With the rapid development of edge computing devices in the industrial Internet of Things, data and computing applications are gradually increasing. It is necessary to work together through the computing performance of the cloud and the real-time characteristics of the edge to maximize the use of data and improve efficiency. Cloud-edge collaboration uses cloud platforms to provide computing, storage and network resources for edge devices to achieve collaborative work between edge devices and cloud platforms. It combines the elasticity of cloud computing with the real-time nature of edge computing to better meet the needs of different scenarios. Specifically, cloud computing usually refers to storing data and applications on servers in remote data centers and providing services over the network; while edge computing pushes data processing to edge devices that are closer to the data source, such as sensors, embedded systems, etc.

[0003] At present, edge computing devices for industrial big data are mostly concentrated in intelligent transportation, security monitoring, and industrial Internet of Things. The main purpose is to process data in real time, alleviate cloud pressure, reduce costs and energy consumption, make full use of the advantages of both, and realize fault diagnosis based on transfer learning.

[0004] However, most of the existing cloud-edge collaborative architectures are often designed to solve the problems of real-time data processing and alleviating cloud pressure. They do not fully utilize the advantages of the cloud-edge collaborative joint architecture, and the data collection speed at the edge is slow. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides an intelligent edge-side state data acquisition system, method and device for a robot, which solves the problems raised in the above-mentioned background technology.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0007] On the one hand, an intelligent edge-side state data acquisition system for a robot includes:

[0008] The device end is connected to the edge end, and the device end is used to collect robot operation data and transmit the robot operation data to the edge end;

[0009] The cloud is connected to the edge, and the cloud is used to train public data, obtain a cloud state recognition model, perform universal processing, obtain a universal state recognition model and migrate it to the edge;

[0010] The edge end is connected to the cloud. The edge end is used to obtain edge data after preprocessing and spectrum analysis of the robot operation data, upload it to the cloud, and use a universal state recognition model to diagnose faults on the edge data.

[0011] The present invention is further configured as follows: the public data includes a cloud data set and an edge data set, wherein the edge data set is data obtained after the edge performs preprocessing and spectrum analysis on the robot operation data;

[0012] After receiving the edge data, the cloud stores it in the edge data center and performs data preprocessing and spectrum analysis on the public data.

[0013] The present invention is further configured as follows: the spectrum analysis includes time domain analysis, frequency domain analysis and time-frequency domain analysis;

[0014] The time domain analysis is used to obtain time domain dimensioned features and time domain dimensionless features;

[0015] The frequency domain analysis is used to obtain frequency domain features: centroid frequency, mean square frequency and frequency variance;

[0016] The time-frequency domain analysis is used to obtain time-frequency domain features: three-layer wavelet packet energy.

[0017] The present invention is further configured as follows: the time domain dimensioned features include: mean, mean square error, root mean square and peak-to-peak value;

[0018] The time-domain dimensionless characteristics include: skewness, kurtosis, waveform factor, peak factor, impulse factor and margin factor.

[0019] The present invention is further configured as follows: the state recognition model is used to extract features from the spectrum analysis results according to a one-dimensional convolutional neural network, and classify the extracted features by a least squares support vector machine;

[0020] The one-dimensional convolutional neural network includes a convolution layer, a pooling layer, an activation layer and a fully connected layer.

[0021] The present invention is further configured as follows: the convolutional layer calculation expression includes:

[0022]

[0023] In the formula, y i l-1 is the input of the jth feature map of the l-1th layer, y i l is the output of the i-th feature of the l-th layer, w j，i l is the i-th convolution kernel of the j-th feature map of the l-th convolution layer, * is the convolution calculation, f is the activation function, yi l represents the output value of the i-th neuron in the l-th layer, It means summing the neurons in layer l-1;

[0024] The pooling layer calculation expression includes:

[0025]

[0026] In the formula, the function down() is the downsampling operation, β is the multiplicative bias of the j-th output feature block, b is the additive bias of the j-th output feature block, and X j L is the output of the jth neuron in layer l, and f() is the activation function used to perform nonlinear transformation on the input;

[0027] The activation layer calculation expression includes:

[0028]

[0029] Relu(x)=max(0,x)

[0030] In the formula, Sigmoid() is the activation function, e x To calculate the exponential function of the input independent variable x, e -x To calculate the inverse of the exponential of the input independent variable x, x is the independent variable, ReLu() is the activation function, and max is the operator, which takes the maximum value from 0 to x;

[0031] The fully connected layer calculation expression includes:

[0032]

[0033] In the formula, For the input z j The output after the Softmax function transformation, z j is the jth component in the input vector, usually the output of a layer in the neural network, To sum the indices of all categories k, the sum of the probability values ​​output by Softmax is 1. Enter z for calculation j The exponential of , Softmax() is the activation function used for multi-classification problems.

[0034] The present invention is further configured such that the decision function of the least squares support vector machine includes:

[0035]

[0036] Where f(x) is the classification result of the input sample x, sgn() is the symbol function used to determine the classification result, and α k is the Lagrange multiplier, y k is the kth training sample x k The category label, K(x k， , x) is the kernel function, and b is the bias term.

[0037] On the other hand, the present invention also discloses a method for collecting state data of an intelligent edge side of a robot, which specifically comprises the following steps:

[0038] S1. After the cloud preprocesses and performs spectrum analysis on the public data set, it uses a one-dimensional convolutional neural network to extract features from the spectrum analysis results, and classifies the extracted features through a least squares support vector machine to form a cloud state recognition model. After universal processing, the universal state recognition model is obtained and migrated to the edge.

[0039] S2. The device side collects the robot operation data. After the edge side preprocesses and performs spectrum analysis on the robot operation data, it obtains the edge data, uses the universal state recognition model to diagnose the edge data faults, and transmits the diagnosis results to the edge monitoring platform for visual display.

[0040] The present invention is further configured as follows: the edge end uploads the edge data to the cloud, updates the public data set, the cloud uses the updated public data set to update the cloud state recognition model, and after universal processing, obtains the updated universal state recognition model and migrates it to the edge end.

[0041] In yet another aspect, the present invention further discloses an intelligent edge-side state data acquisition device for a robot, comprising FPGA hardware, a level acquisition card, a nine-axis attitude sensor, a vibration sensor, an edge-end processor, an edge-end memory, an edge-end monitoring platform, a cloud-end processor, and a cloud-end memory;

[0042] The nine-axis attitude sensor and vibration sensor are used to collect the operation data of the robot;

[0043] The FPGA hardware and level acquisition card are used to obtain operation data and upload the operation data to the edge processor;

[0044] The edge processor is used to receive operation data, pre-process and perform frequency domain analysis on the operation data, and perform fault diagnosis;

[0045] The edge end memory is used to store the universal state recognition model and the frequency domain analysis results;

[0046] The edge monitoring platform is used to visualize the fault diagnosis results;

[0047] The cloud processor is used to pre-process and perform frequency domain analysis on the public data, generate a cloud state recognition model, and then perform universal processing to obtain a universal state recognition model;

[0048] The cloud storage is used to store public data and a cloud status recognition model.

[0049] The present invention provides an intelligent edge-side state data acquisition system, method and device for a robot, which has the following beneficial effects:

[0050] While effectively solving the problems of real-time data processing and equipment monitoring on industrial big data, the present invention further improves the data collection speed through hardware enhancement. Through cloud-edge collaborative technology, it can combine the respective advantages of the cloud and edge, and achieve real-time data processing and equipment monitoring, as well as the ability to perform large-scale calculations. It can make full use of data and realize model migration, ensuring the high accuracy and strong generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the system architecture of the present invention;

[0052] Figure 2 The figure is a flow chart of the method for data collection of the present invention. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0054] See also Figure 1-2 , the embodiment of the present invention provides the following technical solutions:

[0055] Embodiment 1: An intelligent edge-side state data acquisition system for a robot includes a device end, an edge end and a cloud end. The device end is connected to the edge end. The device end is used to collect robot operation data and transmit the robot operation data to the edge end. In order to improve data acquisition efficiency, a nine-axis attitude sensor and a vibration sensor are deployed on the edge end to collect the operation data of the industrial robot. Multi-channel synchronous data acquisition is performed through the cooperation of FPGA hardware and a level acquisition card to improve data acquisition efficiency. The level acquisition card is an MCC118 level acquisition card.

[0056] As a preferred solution, in order to realize the diagnosis and processing of operation data, the cloud and the edge are connected. The cloud is used to build a cloud status recognition model, pre-process and spectrally analyze the cloud data set, obtain the cloud status recognition model, and after universal processing, obtain the universal status recognition model and migrate it to the edge to provide guarantee for data processing. As a further explanation, universal processing does not change the overall framework of the model, but only modifies the network parameters so that the large model of the cloud status recognition model becomes a small model that can be run by edge computing devices with weak computing power, namely the universal status recognition model.

[0057] As a preferred solution, in order to achieve an intuitive display of the diagnosis results, the edge is connected to the cloud. The edge is used to obtain edge data after preprocessing and spectral analysis of the robot operation data, and use a universal state recognition model to perform fault diagnosis on the edge data.

[0058] As a detailed description, spectrum analysis includes time domain analysis, frequency domain analysis and time-frequency domain analysis to obtain corresponding features. Specifically, time domain analysis is used to obtain time domain dimensional features and time domain dimensionless features, among which time domain dimensional features include: mean, mean square error, root mean square and peak-to-peak value:

[0059] The mean is the average value within the signal period:

[0060]

[0061] In the formula, F 1 and are the mean values ​​of the signals, N is the total number of samples, X i is the i-th sample value;

[0062] The mean square error is used to reflect the degree of dispersion of data within the signal period:

[0063]

[0064] In the formula, F 2 With X std The mean square error is the mean of the signal, which is used to measure the degree of deviation between the signal data points and the mean;

[0065] Root mean square is also called effective value:

[0066]

[0067] In the formula, F 3 With X RMS are the root mean square of the signal;

[0068] The peak-to-peak value is the difference between the maximum and minimum values ​​of the data within the signal period.

[0069] F 4=X PtP =max(X i )-min(X i )

[0070] In the formula, F 4 With X PtP Both are the peak-to-peak values ​​of the signal.

[0071] The dimensionless characteristics in the time domain include: skewness, kurtosis, shape factor, crest factor, impulse factor, and margin factor:

[0072] Skewness is the asymmetry of the probability distribution of the input signal:

[0073]

[0074] In the formula, F 5 is the skewness of the signal;

[0075] Kurtosis is the height of the peak of the probability density distribution curve of the input signal at the mean:

[0076]

[0077] In the formula, F 6 is the kurtosis of the signal;

[0078] The form factor is used to characterize the degree to which a waveform deviates from a sine wave:

[0079]

[0080] In the formula, F 7 is the waveform factor of the signal;

[0081] The crest factor is how extreme the peaks in a waveform appear:

[0082]

[0083] In the formula, F 8 is the peak factor of the signal, X RMS is the root mean square of the signal, X peak is the peak value of the signal;

[0084] The impulse factor can be used to determine whether the signal has an impact factor:

[0085]

[0086] In the formula, F 9 is the impulse factor of the signal;

[0087] The margin factor can be used to determine the wear and tear of the signal:

[0088]

[0089] In the formula, F 10 is the signal margin factor, X R is the average amplitude;

[0090] Frequency domain analysis is used to obtain frequency domain features: centroid frequency, mean square frequency, and frequency variance:

[0091] Center of gravity frequency:

[0092]

[0093] In the formula, F 11 With FC as the center of gravity frequency, f i is the ith frequency point, S(f i ) is the frequency f i The frequency amplitude at , N is the Nth input value;

[0094] Mean square frequency:

[0095]

[0096] In the formula, F 12 and MSF is the mean square frequency;

[0097] Frequency variance:

[0098]

[0099] In the formula, F 13 and VF is the mean square frequency.

[0100] Time-frequency domain analysis is used to obtain time-frequency domain features: three-layer wavelet packet energy. Wavelet transform analysis is used in time-frequency domain analysis. When the frequency is obtained, the time can also be located. The infinite-length triangular basis function in Fourier transform is replaced by the finite-length attenuated wavelet basis. It divides the original signal into multiple sub-bands, and analyzes the signal locally by analyzing the time domain components on each sub-band. Specifically:

[0101] Fourier Transform:

[0102]

[0103] Where F(w) is the frequency domain signal, f(t) is the time-frequency signal, w is the angular frequency, e -iwt To find the reciprocal of the logarithm of the input value;

[0104] The wavelet fundamental wave expression is as follows:

[0105]

[0106] Where: s is the scale parameter; t is the displacement parameter, Ψ(t) is the wavelet mother function, Ψ st(t) is the wavelet basis function after scaling and translation;

[0107] In discrete wavelet transform, the wavelet basis function can be obtained:

[0108]

[0109] In the formula, Ψ j，k (t) is the wavelet function at scale j and translation k, Ψ(t) is the wavelet mother function, s 0 is the basic scale, t 0 is the basic translation, t is the displacement parameter, j is the scale index, and k is the translation index.

[0110] As a preferred solution, after completing the spectrum analysis, a one-dimensional convolutional neural network is used to extract features from the spectrum analysis results. The one-dimensional convolutional neural network, i.e., 1DCNN, includes a convolutional layer, a pooling layer, an activation layer, and a fully connected layer. Specifically:

[0111] The convolutional layer calculation expressions include:

[0112]

[0113] In the formula, y i l-1 is the input of the jth feature map of the l-1th layer, y i l is the output of the i-th feature of the l-th layer, w j，i l is the i-th convolution kernel of the j-th feature map of the l-th convolution layer, * is the convolution calculation, f is the activation function, y i l represents the output value of the i-th neuron in the l-th layer, It means summing the neurons in layer l-1;

[0114] The pooling layer calculation expression includes:

[0115]

[0116] In the formula, the function down() is the downsampling operation, β is the multiplicative bias of the j-th output feature block, b is the additive bias of the j-th output feature block, and X j L is the output of the jth neuron in layer l, and f() is the activation function used to perform nonlinear transformation on the input;

[0117] The activation layer calculation expression includes:

[0118]

[0119] Relu(x)=max(0,x)

[0120] In the formula, Sigmoid() is the activation function, e x To calculate the exponential function of the input independent variable x, e -x To calculate the inverse of the exponential of the input independent variable x, x is the independent variable, ReLu() is the activation function, and max is the operator, which takes the maximum value from 0 to x;

[0121] The fully connected layer calculation expression includes:

[0122]

[0123] In the formula, For the input z j The output after the Softmax function transformation, z j is the jth component in the input vector, usually the output of a layer in the neural network, To sum the indices of all categories k, the sum of the probability values ​​output by Softmax is 1. Enter z for calculation j The exponential of , Softmax() is the activation function used for multi-classification problems.

[0124] The decision functions of the least squares support vector machine include:

[0125]

[0126] Where f(x) is the classification result of the input sample x, sgn() is the symbol function used to determine the classification result, and α k is the Lagrange multiplier, y k is the kth training sample x k The category label, K(x k， , x) is the kernel function, and b is the bias term.

[0127] Embodiment 2. This embodiment is an improvement of the previous embodiment and is an intelligent edge-side state data acquisition system for a robot. It also includes the edge end uploading edge data to the cloud and storing it in an edge end data set. The edge end data set and the cloud end data set constitute public data, which serves as training data for a cloud end state recognition model to train and update the cloud end state recognition model.

[0128] The advantage of the second embodiment over the first embodiment is that the model training and use are separated, and the effective application of the updated universal state recognition model by the edge end is ensured through data migration.

[0129] The equipment used in the above system includes FPGA hardware, level acquisition card, nine-axis attitude sensor, vibration sensor, edge processor, edge memory, edge monitoring platform, cloud processor and cloud memory;

[0130] The nine-axis attitude sensor and vibration sensor are used to collect the robot's operating data. Figure 2 As shown, the sensor industrial robot is a robot equipped with a nine-axis attitude sensor and a vibration sensor;

[0131] FPGA hardware and level acquisition card are used to obtain operation data and upload the above operation data to the edge processor. Among them, FPGA hardware and level acquisition card are installed on Xilinx AX7010 development board, expansion board MCC118 level acquisition card, provides 8 channels of analog voltage input, maximum sampling rate of 100kS / s, supports up to 8 identical modules, up to 64 channels and maximum sampling rate of up to 320kS / s, can collect single-point voltage and voltage waveform, Xilinx AX7010 development board consists of processor system and programmable logic, equipped with rich peripherals and interfaces, equipped with USB Cable circuit, using the USB interface to download programs and debug, convenient for use and functional verification, especially the acquisition end adds the AD7606 module as a data acquisition expansion module, which is an 8-channel data acquisition module, +5V single power supply, does not need to drive the operational amplifier and external bipolar power supply, can handle ±10V and ±5V true bipolar input signals, all channels can be sampled at the maximum sampling rate of 200KSPS at the same time, digital filter, input amplifier, analog multiplexer and overvoltage protection circuit are integrated on the chip;

[0132] The edge processor is used to receive operation data, pre-process and analyze the operation data in the frequency domain, and perform fault diagnosis. Specifically, Raspberry Pi 3B+ is used as the main processor, and the official Raspbian OS is installed. The main parameters are 1.4GHz 64-bit quad-core ARM Cortex-A53 CPU, dual-band wireless LAN (5G / 2.4G WIFI), Gigabit Ethernet, Bluetooth 4.2, 1G LPDDR2, 4 USB interfaces, 40 GPIO ports, etc., supporting SPI, I2C, USB, IO, Ethernet communication, etc.

[0133] The edge end memory is used to store the above universal state recognition model and the above frequency domain analysis results;

[0134] The edge monitoring platform is used to visualize the fault diagnosis results of the universal state recognition model;

[0135] The cloud processor is used to pre-process and perform frequency domain analysis on the above public data, generate a cloud state recognition model, and then perform universal processing to obtain a universal state recognition model;

[0136] The cloud storage is used to store the above public data and cloud status recognition model.

[0137] Further description, the intelligent edge-side state data collection method for a robot specifically includes the following steps:

[0138] S1. After the cloud performs preprocessing and spectrum analysis on the public data set stored in the cloud storage, the mean, mean square error, root mean square, peak, skewness, kurtosis, waveform factor, peak factor, pulse factor, margin factor, centroid frequency, mean square frequency, frequency variance and three-layer wavelet packet energy are obtained. A one-dimensional convolutional neural network is used to extract features from the spectrum analysis results, and the extracted features are classified by the least squares support vector machine to form a cloud state recognition model. After universal processing, the universal state recognition model is obtained and migrated to the edge.

[0139] S2. The device collects the robot operation data. After the edge performs preprocessing and spectrum analysis on the robot operation data, the mean, mean square error, root mean square, peak, skewness, kurtosis, waveform factor, peak factor, pulse factor, margin factor, center of gravity frequency, mean square frequency, frequency variance and three-layer wavelet packet energy are obtained as edge data. The universal state recognition model is used to diagnose the edge data for faults, and the diagnosis results are transmitted to the edge monitoring platform for visual display.

[0140] S3. The edge uploads the edge data to the cloud and updates the public data set. The cloud uses the updated public data set to update the cloud state recognition model. After universal processing, the updated universal state recognition model is obtained and migrated to the edge.

[0141] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0142] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent edge-side state data acquisition system for robots, characterized in that: include: The device end is connected to the edge end, and the device end is used to collect robot operation data and transmit the robot operation data to the edge end; The cloud is connected to the edge, and the cloud is used to train public data, obtain a cloud state recognition model, perform universal processing, obtain a universal state recognition model and migrate it to the edge; The edge end is connected to the cloud. The edge end is used to obtain edge data after preprocessing and spectrum analysis of the robot operation data, upload it to the cloud, and use a universal state recognition model to diagnose faults on the edge data.

2. The intelligent edge-side state data acquisition system for a robot according to claim 1, characterized in that: The public data includes a cloud data set and an edge data set, wherein the edge data set is data obtained after the edge performs preprocessing and spectrum analysis on the robot operation data; After receiving the edge data, the cloud stores it in the edge data center and performs data preprocessing and spectrum analysis on the public data.

3. The intelligent edge-side state data acquisition system for a robot according to claim 2, characterized in that: The spectrum analysis includes time domain analysis, frequency domain analysis and time-frequency domain analysis; The time domain analysis is used to obtain time domain dimensioned features and time domain dimensionless features; The frequency domain analysis is used to obtain frequency domain features: centroid frequency, mean square frequency and frequency variance; The time-frequency domain analysis is used to obtain time-frequency domain features: three-layer wavelet packet energy.

4. The intelligent edge-side state data acquisition system for a robot according to claim 3, characterized in that: The time domain dimensioned features include: mean, mean square error, root mean square and peak-to-peak value; The time-domain dimensionless characteristics include: skewness, kurtosis, waveform factor, peak factor, impulse factor and margin factor.

5. The intelligent edge-side state data acquisition system for a robot according to claim 4, characterized in that: The state recognition model is used to extract features from the spectrum analysis results according to a one-dimensional convolutional neural network, and classify the extracted features through a least squares support vector machine; The one-dimensional convolutional neural network includes a convolution layer, a pooling layer, an activation layer and a fully connected layer.

6. The intelligent edge-side state data acquisition system for a robot according to claim 5, characterized in that: The convolutional layer calculation expression includes: In the formula, is the input of the jth feature map of the l-1th layer, is the output of the i-th feature of the l-th layer, is the i-th convolution kernel of the j-th feature map of the l-th convolution layer, * is the convolution calculation, f is the activation function, represents the output value of the i-th neuron in the l-th layer, It means summing the neurons in layer l-1; The pooling layer calculation expression includes: In the formula, the function down() is the downsampling operation, β is the multiplicative bias of the j-th output feature block, and b is the additive bias of the j-th output feature block. is the output of the jth neuron in layer l, and f() is the activation function used to perform nonlinear transformation on the input; The activation layer calculation expression includes: Relu(x)=max(0,x) In the formula, Sigmoid() is the activation function, e x To calculate the exponential function of the input independent variable x, e -x To calculate the inverse of the exponential of the input independent variable x, x is the independent variable, ReLu() is the activation function, and max is the operator, which takes the maximum value from 0 to x; The fully connected layer calculation expression includes: In the formula, For the input z j The output after the Softmax function transformation, z j is the jth component in the input vector, usually the output of a layer in the neural network, To sum the indices of all categories k, the sum of the probability values ​​output by Softmax is 1. Enter z for calculation j The exponential of , Softmax() is the activation function used for multi-classification problems.

7. The intelligent edge-side state data acquisition system for a robot according to claim 6, characterized in that: The decision function of the least squares support vector machine includes: Where f(x) is the classification result of the input sample x, sgn() is the symbol function used to determine the classification result, and α k is the Lagrange multiplier, y k is the kth training sample x k The category label, K(x k， , x) is the kernel function, and b is the bias term.

8. The data collection method for the intelligent edge-side state data collection system for a robot according to claim 7, characterized in that: The specific steps include: S1. After the cloud preprocesses and performs spectrum analysis on the public data set, it uses a one-dimensional convolutional neural network to extract features from the spectrum analysis results, and classifies the extracted features through a least squares support vector machine to form a cloud state recognition model. After universal processing, the universal state recognition model is obtained and migrated to the edge. S2. The device side collects the robot operation data. After the edge side preprocesses and performs spectrum analysis on the robot operation data, it obtains the edge data, uses the universal state recognition model to diagnose the edge data faults, and transmits the diagnosis results to the edge monitoring platform for visual display.

9. The data collection method of the intelligent edge-side state data collection system for a robot according to claim 8, characterized in that: The edge uploads the edge data to the cloud, updates the public data set, and the cloud uses the updated public data set to update the cloud state recognition model. After universal processing, the updated universal state recognition model is obtained and migrated to the edge.

10. The device used in the intelligent edge-side state data acquisition system for a robot according to claim 7, characterized in that: Including FPGA hardware, level acquisition card, nine-axis attitude sensor, vibration sensor, edge processor, edge memory, edge monitoring platform, cloud processor and cloud memory; The nine-axis attitude sensor and vibration sensor are used to collect the operation data of the robot; The FPGA hardware and level acquisition card are used to obtain operation data and upload the operation data to the edge processor; The edge processor is used to receive operation data, pre-process and perform frequency domain analysis on the operation data, and perform fault diagnosis; The edge end memory is used to store the universal state recognition model and the frequency domain analysis results; The edge monitoring platform is used to visualize the fault diagnosis results; The cloud processor is used to pre-process and perform frequency domain analysis on the public data, generate a cloud state recognition model, and then perform universal processing to obtain a universal state recognition model; The cloud storage is used to store public data and a cloud status recognition model.

Images