Edge end coal cutter cutting part fault diagnosis system for low-quality data

Abstract

The invention discloses an edge end coal cutter cutting unit fault diagnosis system for low-quality data. The system comprises the following steps: collecting an actual working condition sample and a simulation working condition sample on a simulation experiment table; and constructing a field fault data virtual sample generation model based on the adversarial neural network, and obtaining a balanced actual working condition sample data set. And establishing an experiment table fault simulation working condition sample screening model based on the probability mass function, and selecting an experiment table migratable simulation working condition sample. And performing autocorrelation spectrum kurtosis diagram conversion on the two groups of samples, building a deep convolutional neural network, calculating a feature distribution difference of each layer by using maximum mean difference MMD, and determining a network parameter sharing boundary and a coal mining machine cutting unit fault diagnosis model. And predicting the health state of the cutting unit of the coal mining machine by using the fault diagnosis model of the cutting unit of the coal mining machine. According to the invention, the fault diagnosis of the cutting unit of the coal mining machine under low-quality data is realized, and the system has low degree of dependence on the quality of fault samples and has good universality for cutting units of different models of coal mining machines.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mine machinery health monitoring, and particularly relates to a method for building an edge-end coal mining machine cutting part fault diagnosis system for low-quality data. BACKGROUND

[0002] The cutting part of a coal mining machine is a main component of power transmission of the coal mining machine, and its health status seriously affects the overall operation efficiency of the coal mining machine. However, the high temperature, humidity and heavy load operation conditions make the cutting part of the coal mining machine fail from time to time, and in severe cases, may even cause mechanical failure and even safety accidents. Real-time monitoring of the health status of the transmission part during operation, combined with the maintenance cycle, timely replacement of faulty parts, is of great significance to ensure processing quality and safety.

[0003] At present, the fault diagnosis of the cutting part of the coal mining machine is mainly through current, sound, vibration and other ways to collect and analyze data. Compared with current, sound and other monitoring methods, vibration signals can monitor the vibration of the equipment caused by early faults based on the vibration characteristics of rotating machinery. At the same time, with the development of artificial intelligence and big data technology, data-driven vibration signal analysis methods have gradually become mainstream. Therefore, based on the data-driven method, the research on the fault diagnosis of the cutting part of the coal mining machine is also an important direction of the current intelligent operation and maintenance of the cutting part of the coal mining machine.

[0004] However, the data-driven fault diagnosis technology is highly dependent on the quality of the sample set, and the data-driven fault diagnosis of the cutting part of the coal mining machine mainly has the following problems: (1) the existing coal mining machine is low in intelligence, and most of the cutting parts of the coal mining machine do not configure vibration monitoring devices, resulting in the problem of small sample set of fault data sample set; (2) the cutting part of the coal mining machine has the characteristics of occasional occurrence, and it is impossible to comprehensively collect fault samples of various fault types, resulting in the problem of sample imbalance of the fault data sample set; (3) there are many production manufacturers and types of coal mining machines, and the sizes and structures of the transmission systems also have certain differences, resulting in the problem of poor universality of the fault data sample set. Therefore, building a coal mining machine cutting part fault diagnosis system and jointly constructing a high-precision multi-source fusion deep migration diagnosis model under low-quality data (small sample size, imbalance and poor universality) is a difficult point of the fault diagnosis of the cutting part of the coal mining machine. However, the existing work often only solves a single problem in low-quality data, and cannot effectively realize the fault diagnosis of the cutting part of the coal mining machine under low-quality sample data.

[0005] In addition, the vibration information sample collected during the working process of the cutting unit of the coal mining machine is large in quantity and high in sampling frequency, and the transmission of a large amount of sensing data to the upper computer will cause congestion of the transmission channel, and the massive data will increase the operation load of the upper computer, which seriously affects the speed of diagnosis and cannot meet the real-time diagnosis requirement of the cutting unit fault of the coal mining machine. SUMMARY

[0006] In view of the above problems, the purpose of the present application is to provide an edge-end coal mining machine cutting unit fault diagnosis system for low-quality data, and a multi-source fusion multi-order transfer coal mining machine cutting unit fault diagnosis algorithm is proposed, an embedded model of the coal mining machine cutting unit fault diagnosis is built based on an edge computing processor, and the fault diagnosis and state evaluation of different types of coal mining machine cutting units are realized by using multi-source sample data, which provides important technical support for the fault diagnosis and health monitoring of the coal mining machine cutting unit in the field of intelligent mine mechanical equipment.

[0007] The present application provides an edge-end coal mining machine cutting unit fault diagnosis system for low-quality data, and the specific steps are as follows: Low-quality data

[0008] Step one, using a vibration sensor to collect vibration signals of the cutting unit of the coal mining machine at multiple points, obtaining actual working condition fault sample data set, using a vibration sensor to collect vibration signals of the cutting unit of the coal mining machine fault simulation test bench, obtaining simulation working condition fault sample data set;

[0009] Step two, according to the different actual working condition sample balance degree in the actual working condition fault sample data set, respectively designing the generator and the discriminator, generating virtual samples, and obtaining balanced real working condition sample data set;

[0010] Step three, based on the distribution of the actual working condition samples in the actual working condition fault sample data set and the simulation working condition samples in the simulation working condition fault sample data set, the probability mass function is used to calculate the probability distribution of the actual working condition fault sample data set and the simulation working condition fault sample data set respectively, and the simulation working condition samples in the simulation working condition sample data set that meet the migration condition are selected;

[0011] Step four, autocorrelation spectrum kurtosis diagram conversion is performed on the actual working condition fault sample data set, the simulation working condition fault sample data set and the balanced real working condition sample data set, and the spectrum kurtosis diagram is obtained; a deep feature extraction model based on deep convolutional neural network is constructed; based on the size of the spectrum kurtosis diagram and the number of class labels, the size and structure parameters of the deep convolutional neural network are determined, and the high-order deep features in the spectrum kurtosis diagram are gradually extracted through convolution and pooling processing;

[0012] ​Step five, using the actual working condition fault sample data set, the balanced real working condition sample data set and the simulated working condition fault sample data set, the deep feature extraction model is trained respectively, the vibration signal of the cutting part of the coal mining machine is taken as the input of the deep feature extraction model, and the health state of the cutting part of the coal mining machine is taken as the output of the deep feature extraction model, two parameter optimal classification networks are obtained, the feature distribution gap of each layer of the two parameter optimal classification networks is calculated, and the network parameters are obtained to obtain the fault diagnosis model of the cutting part of the coal mining machine.

[0013] Preferably, step six, using the constructed fault diagnosis model of the cutting part of the coal mining machine to predict the newly collected vibration signal sample of the cutting part of the coal mining machine, the health state of the cutting part of the coal mining machine is obtained.

[0014] Preferably, the health state of the cutting part of the coal mining machine is displayed on the edge display screen in real time; if the health state of the cutting part of the coal mining machine is fault, the alarm is given through the sound and light alarm or the screen text prompt mode, and the control instruction is sent to guide the coal mining machine to stop.

[0015] Preferably, the step one, the vibration signals of the cutting part of the coal mining machine are collected by multiple vibration sensors, and the actual working condition fault sample data set is obtained, the vibration signals of the cutting part of the coal mining machine fault simulation test bench are collected by vibration sensors, and the simulated working condition fault sample data set is obtained, including:

[0016] The vibration signals of the cutting part of the coal mining machine are collected by multiple vibration sensors, and the actual working condition fault sample data set is obtained, the vibration signals of the cutting part of the coal mining machine fault simulation test bench are collected by vibration sensors, and the simulated working condition fault sample data set is obtained, including:

[0017] Preferably, the step two, according to the balance degree of different actual working condition samples in the actual working condition fault sample data set, the generator and the discriminator are designed and generated respectively, the virtual samples are generated, and the balanced real working condition sample data set is obtained, including:

[0018] According to the different actual working condition sample equalization degrees in the actual working condition fault sample data set, the generator and the discriminator are respectively designed to generate virtual samples and obtain an equalized real working condition sample data set; the Wassertein distance is used to measure the data distribution distance between the virtual samples and the actual working condition samples in the adversarial generation network; a Lipschitz continuous restriction is introduced to set a fixed constant and truncate the absolute value of the parameter update; the RMSProp optimization algorithm is used to improve the stability of gradient descent and balance the training degrees of the generator and the discriminator to obtain an equalized actual working condition sample data set.

[0019] Preferably, in the step three, based on the distribution of the actual working condition samples in the actual working condition fault sample data set and the simulated working condition samples in the simulated working condition fault sample data set, the probability mass function is used to calculate the probability distribution of the actual working condition fault sample data set and the simulated working condition fault sample data set respectively, and the simulated working condition samples in the simulated working condition fault sample data set that meet the migration condition are screened out, including:

[0020] Based on the distribution of the actual working condition samples and the simulated fault samples, the probability mass function is used to calculate the probability distribution of the actual working condition samples and the probability distribution of the simulated working condition samples respectively; the KL divergence is introduced to measure the non-symmetry of the difference between the probability distribution of the actual working condition samples and the probability distribution of the simulated fault samples; the training data set and the test data set are constructed according to the distribution of the actual working condition samples and the simulated fault samples, and the Tradaboost algorithm is used to solve the multi-domain non-independent and identically distributed problem; the error rate of the weak classifier on the training data set is calculated, and the sample weight parameter is adjusted according to the distribution of the actual working condition samples and the simulated working condition samples using different update criteria; the migration condition includes screening the simulated working condition samples higher than the preset weight threshold to obtain the simulated working condition samples.

[0021] Preferably, in the step four, the actual working condition fault sample data set, the simulated working condition fault sample data set and the equalized real working condition sample data set are subjected to autocorrelation spectrum kurtosis map conversion to obtain a spectrum kurtosis map; a deep feature extraction model based on a deep convolutional neural network is constructed; the size and structure parameters of the deep convolutional neural network are determined based on the size of the spectrum kurtosis map and the number of class labels, and high-order deep features in the spectrum kurtosis map are gradually extracted through convolution and pooling processing, specifically:

[0022] The autocorrelation spectrum kurtosis diagram of the actual working condition fault sample data set, the simulated working condition fault sample data set and the balanced real working condition sample data set is converted to obtain a spectrum kurtosis diagram; a deep feature extraction model based on a deep convolutional neural network is constructed; the size and structure parameters of the deep convolutional neural network are determined based on the size of the spectrum kurtosis diagram and the number of category labels, and high-order deep features in the spectrum kurtosis diagram are gradually extracted through convolution and pooling processing; a linear correction unit Relu function is selected as an activation function to improve the convergence speed of the deep feature extraction model by using the sparse activation characteristics thereof; L2 regularization is introduced to limit the model parameters of the deep feature extraction model; a Momentum optimizer gradient descent optimization algorithm is selected to accelerate the learning rate and optimize the parameters of the deep convolutional neural network, so as to obtain the deep feature extraction model.

[0023] Preferably, the step five, using the actual working condition fault sample data set, the balanced real working condition sample data set and the simulated working condition fault sample data set respectively trains the deep feature extraction model, takes the vibration signal of the cutting unit of the coal mining machine as the input of the deep feature extraction model, takes the health state of the cutting unit of the coal mining machine as the output of the deep feature extraction model, obtains two-parameter optimal classification networks, calculates the feature distribution gap of each layer of the two-parameter optimal classification networks, determines the network parameters to obtain the fault diagnosis model of the cutting unit of the coal mining machine, comprising:

[0024] The balanced real working condition sample and the simulated working condition sample are used to train the deep feature extraction model to obtain two-parameter optimal classification networks; the maximum mean difference MMD is used to calculate the feature distribution gap of each layer to determine the network parameter sharing limit; the Coral Loss is introduced into the full connection layer in the deep convolutional neural network to linearly change the features of the full connection layer, and the fault diagnosis model of the cutting unit of the coal mining machine is obtained.

[0025] Compared with the prior art, the present application has the following advantages:

[0026] 1、The present application realizes the fault diagnosis of the cutting unit of the coal mining machine based on low-quality samples, fully considers the problems of small quantity, imbalance and poor universality of the cutting unit data of the coal mining machine, constructs a multi-source fusion multi-order transfer fault diagnosis model of the cutting unit of the coal mining machine for low-quality data, and realizes the construction of a high-precision and high-applicability fault diagnosis model; compared with the traditional fault diagnosis method, the fault diagnosis method of the cutting unit of the coal mining machine proposed by the present application has low dependence on sample quality and high intelligence;

[0027] 2、The coal mining machine cutting part fault diagnosis model based on the edge computing processor is designed, the edge computing processor near the coal mining machine cutting part carries out real-time processing on the multi-path vibration data, the data processing speed and the timeliness of diagnosis are significantly improved. The traditional signal acquisition and analysis system only uploads the fault diagnosis result and the original data corresponding to the fault sample to the centralized control room on the well, and the data transmission quantity and the data transmission pressure of the underground ring network are effectively reduced. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The low-quality data-oriented edge end coal mining machine cutting part fault diagnosis system flowchart is provided in the application.

[0029] Figure 2 The multi-source fusion multi-order migration diagnosis algorithm flowchart. DETAILED DESCRIPTION

[0030] The application will be further described below.

[0031] As shown in the low-quality data-oriented edge end coal mining machine cutting part fault diagnosis system flowchart provided by the application, the specific steps are as follows: Figure 1

[0032] Step one, the vibration signals of the coal mining machine cutting part are collected by using vibration sensors, including one vibration sensor arranged in the axial direction of the cutting transmission part and four vibration sensors arranged in the radial direction of the cutting transmission part, the signal cable is connected to the edge computing processor, and the actual working condition fault sample data set of the coal mining machine cutting part is obtained; the vibration signals of the coal mining machine cutting part fault simulation test bench are collected by using vibration sensors, including one vibration sensor arranged in the axial direction of the cutting transmission part and four vibration sensors arranged in the radial direction of the cutting transmission part, the signal cable is connected to the edge computing processor, and the simulation working condition fault sample data set of the coal mining machine cutting part is obtained.

[0033] The parameters of the edge computing processor selected by the application are as follows: MCU: ARM Cortex-A9 dual-core 1.0 GHZ; compact size: 110.5 (L) x 40 (W) x 126.5 (H) mm; communication interface: Ethernet, wireless, supporting multi-device cascading; running Linux operating system, independent running, built-in web console; 4-channel synchronous sampling analog input, the highest sampling rate can reach 128 KS / S. 24-bit resolution, 95 dB dynamic range, gain error 0.05%; DC or AC coupling, software selectable. IEPE current excitation, software selectable; analog or digital trigger; 4-channel programmable 10: digital input, digital output, tachometer input.

[0034] ​Among them, the multi-point vibration sensor is selected from an explosion-proof vibration sensor, and main technical indexes are as follows:

[0035] Working voltage: (10-24) V, working current: ≤30mA.

[0036] Measurement range: (0~±80) Error: ±1.6 .

[0037] Frequency: 80Hz (calibration frequency).

[0038] Output signal system: two-wire analog signal type: 4mA~12mA~20mA, corresponding display range: (-80~0~+80) .

[0039] As Figure 2 shown, the present application comprises the steps of:

[0040] Step two, according to the balance degree of different working conditions and fault type field data samples, the generator and discriminator are designed and generated respectively for analog working condition samples and actual working condition samples, and virtual field samples are generated. The Wassertein distance is used to measure the distribution distance between the virtual sample and the actual working condition sample in the adversarial generation network, and a fixed constant is introduced to limit the Lipschitz continuity to truncate the absolute value of parameter update, so as to ensure the boundedness of weight parameters. The RMSProp optimization algorithm is used to improve the stability of gradient descent, balance the training degree of generator and discriminator, ensure the diversity of virtual sample generation, and obtain balanced real working condition sample data set.

[0041] Wassertein distance is a method for measuring the distance between two probability distributions, defined as:

[0042] (1)

[0043] Where, represents the distribution and combination of all possible joint distributions, represents each possible joint distribution, represents a pair of samples sampled in , represents the distance of the sample pair, and represents the expected value of the sample pair distance.

[0044] Lipschitz continuity is a strong type of continuity concept in mathematical functions, defined as:

[0045] For a function on a subset of the real numbers, if there exists a constant such that:

[0046] (2)

[0047] then is said to be Lipschitz continuous.

[0048] RMSProp is an optimization algorithm in deep learning that adjusts the learning rate of each parameter based on the history of its gradients. The specific steps are as follows:

[0049] ① Use an exponentially weighted average to calculate the moving average of the squared gradient. For each parameter's gradient , calculate the moving average of the squared gradient ;

[0050] (3)

[0051] where β represents a decay factor that controls the weight of historical gradients on the current gradient, generally taking a value between 0 and 1.

[0052] ② For each parameter's learning rate , calculate the adjusted learning rate ;

[0053] (4)

[0054] where ε represents a very small constant for numerical stability.

[0055] ③ Update the parameter using the adjusted learning rate to the parameter :

[0056] (5)

[0057] ④ Repeat the above steps, in each iteration, the learning rate will be adjusted according to the information of historical gradients, so that the learning rate of parameters with larger gradients is smaller, and the learning rate of parameters with smaller gradients is larger, so as to better adapt to the updating needs of different parameters.

[0058] Step three, based on the distribution of field data samples and experimental data samples, the probability distribution is calculated by using the probability mass function. The KL divergence is introduced to measure the non-symmetry of the difference between the two probability distributions. According to the distribution of the sample, the training data set and the test data set are constructed, and the Tradaboost algorithm is used to solve the problem of multi-field non-independent and homogeneous distribution. The error rate of the weak classifier on the training data set is calculated, and different updating criteria are used to adjust the sample weight parameters and set the weight threshold to filter the experimental data samples. The transferable experimental samples are obtained.

[0059] The KL divergence is used to measure the difference between two probability distributions, defined as:

[0060] For two given probability distributions and , their corresponding probability density functions are and , and approximates , the expression of KL divergence is:

[0061] (6)

[0062] From the above formula, it can be seen that the KL divergence can be used to measure the non-symmetry, and only when , , otherwise the KL divergence of is not equal to the KL divergence of .

[0063] The Tradaboost algorithm can train a classifier using multiple training data with different distributions. The specific steps are as follows:

[0064] ①, define different distribution training data sets and , the combined training data set , a test data set , a classification algorithm Learner, and the number of iterations ;

[0065] ②, the initial weight vector :

[0066] (7)

[0067] ③, set :

[0068] : (8)

[0069] IV. For , set the weight distribution on the training data to satisfy:

[0070] : (9)

[0071] V. Call the Learner to obtain a classifier on according to the merged training data and the weight distribution on the test data ;

[0072] (10)

[0073] (11)

[0074] VI. Set :

[0075] (12)

[0076] VII. Set a new weight vector :

[0077] (13)

[0078] VIII. Obtain the final classifier :

[0079] (14)

[0080] Step four, perform autocorrelation spectrum kurtosis conversion on all samples and construct a deep feature extraction model based on deep convolutional neural network. Determine the size and structure parameters of the deep convolutional neural network based on the size of the spectrum kurtosis and the number of class labels, and gradually extract high-order deep features in the spectrum kurtosis through convolution and pooling processing. Select the linear correction unit Relu function as the activation function to improve the convergence speed by using its sparse activation characteristics. Introduce L2 regularization to limit model parameters and reduce overfitting in deep neural networks. Select the Momentum optimizer gradient descent optimization algorithm to accelerate the learning rate and optimize the parameters of the deep convolutional neural network.

[0081] The Relu function introduces non-linear characteristics in the deep learning network, and the expression is:

[0082] (15)

[0083] ​​​L2 regularization prevents the model from overfitting the training data by adding the L2 norm of the weights to the model's loss function. It is defined as:

[0084] (15)

[0085] in, Represents the original loss function. Represents the weight vector. The L2 norm of the weight vector is represented. The hyperparameter representing the strength of regularization is used to control the magnitude of the influence of the regularization term. It is the loss function after regularization.

[0086] The Momentum optimizer is used to accelerate the convergence speed of the gradient descent algorithm. It uses past gradient information to update parameters, and the update rule is as follows:

[0087] (16)

[0088] (17)

[0089] in, Indicates time step momentum, Indicates the momentum coefficient. Indicates the learning rate. It is a loss function Regarding parameters gradient, These are parameters that need to be updated.

[0090] Step 5: Train the deep convolutional neural network using both actual and simulated working condition samples to obtain two optimal classification networks. Calculate the feature distribution gap between each layer using the Maximum Mean Difference (MMD) method to determine the parameter sharing limits. Transfer the fault diagnosis model based on the deep convolutional neural network by introducing Coral Loss into the fully connected layer to linearly transform the features of the fully connected layer, thus obtaining a fault diagnosis model for the coal mining machine's cutting section.

[0091] MMD is used to measure two sets of features. and The distribution difference is expressed as:

[0092] (18)

[0093] in, Represents the regenerated Hilbert space. , , .

[0094] Coral Loss is used to reduce the size and range of feature values learned from source and target domains in network training, and the calculation method is as follows:

[0095] First, the covariance matrices of the source domain dataset and the target domain dataset are calculated respectively and :

[0096] (19)

[0097] (20)

[0098] Then the Coral Loss of the source and target domain features can be expressed as:

[0099] (21)

[0100] Wherein, denotes the Frobenius norm of the square matrix, and d denotes the dimension of the sample.

[0101] Step six, compile the fault diagnosis model of the cutting unit of the coal mining machine constructed in the edge program, embed it in the edge computing processor, analyze the newly collected vibration signal samples of the cutting unit of the coal mining machine, and diagnose the health state of the cutting unit of the coal mining machine. Further, the health state information of the cutting unit of the coal mining machine is displayed on the edge display screen in real time, and the analysis result is uploaded to the control center on the ground by using the underground ring network. When a fault occurs, the alarm is sounded through the audible and visual alarm, the screen text prompt mode is used to issue an alarm, and a control command is sent to guide the coal mining machine to stop and protect. The health state of the cutting unit of the coal mining machine includes tooth surface pitting, cracking, tooth breaking and the like of the transmission gear of the cutting unit.

[0102] Table 1, audible and visual alarm parameters

[0103]

[0104] Among them, the edge computing processor selected by the application is Jianyi MCM-6204, the display screen selected is Kunlun Tong State TPC7032Ki, the specific parameters are shown in Table 2, the industrial audible and visual alarm selected is Hangya YS-01H, and the specific parameters are shown in Figure 1 .

[0105] Table 2, display screen parameters

[0106]

[0107] The above merely describes the preferred embodiments of the present application, and it should be pointed out that those skilled in the art can make several improvements and refinements without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A low-quality data-oriented edge-end coal cutter cutting unit fault diagnosis system, characterized in that, The specific steps are: Step one, using vibration sensor multi-point acquisition of the vibration signal of the cutting part of the coal mining machine, obtaining the actual working condition fault sample data set, using the vibration sensor to collect the vibration signal of the cutting part of the coal mining machine fault simulation test bench, obtaining the simulation working condition fault sample data set; Step two, according to the different actual working condition sample balance degree in the actual working condition fault sample data set, respectively design the generator and the discriminator, generate virtual samples, and obtain the balanced real working condition sample data set; Step three, based on the distribution of the actual working condition sample in the actual working condition fault sample data set and the simulation working condition sample in the simulation working condition fault sample data set, the probability mass function is used to calculate the probability distribution of the actual working condition fault sample data set and the simulation working condition fault sample data set respectively, and the simulation working condition sample in the simulation working condition sample data set that meets the migration condition is screened out; Step four, autocorrelation spectrum kurtosis diagram conversion is performed on the actual working condition fault sample data set, the simulation working condition fault sample data set and the balanced real working condition sample data set, and the spectrum kurtosis diagram is obtained; a deep feature extraction model based on deep convolutional neural network is constructed; Based on the size and category label number of the spectrum kurtosis diagram, the size and structure parameters of the deep convolutional neural network are determined, and the high-order deep features in the spectrum kurtosis diagram are gradually extracted through convolution and pooling processing; Step five, the actual working condition fault sample data set, the balanced real working condition sample data set and the simulation working condition fault sample data set are used to train the deep feature extraction model, the vibration signal of the cutting part of the coal mining machine is used as the input of the deep feature extraction model, and the health status of the cutting part of the coal mining machine is used as the output of the deep feature extraction model, two parameter optimal classification networks are obtained, the feature distribution difference of each layer of the two parameter optimal classification networks is calculated, and the network parameters are determined to obtain the coal mining machine cutting part fault diagnosis model.

2. The edge-end coal mining machine cutting part fault diagnosis system facing low-quality data according to claim 1, wherein Step six, using the constructed coal mining machine cutting part fault diagnosis model to predict the newly collected coal mining machine cutting part vibration signal sample, and obtaining the health status of the coal mining machine cutting part.

3. The edge-end coal mining machine cutting part fault diagnosis system facing low-quality data according to claim 2, wherein The health status of the cutting part of the coal mining machine is displayed in real time on the edge-end display screen; if the health status of the cutting part of the coal mining machine is fault, an alarm is given through the sound-light alarm or the screen text prompt mode, and a control instruction is sent to guide the coal mining machine to stop.

4. The low-quality data-oriented edge-end coal cutter cutting unit fault diagnosis system according to claim 1, characterized in that, The step one, using vibration sensor multi-point acquisition of the vibration signal of the cutting part of the coal mining machine, obtaining the actual working condition fault sample data set, using the vibration sensor to collect the vibration signal of the cutting part of the coal mining machine fault simulation test bench, obtaining the simulation working condition fault sample data set, comprising: The vibration signals of the cutting part of the coal mining machine are collected by multiple vibration sensors, the vibration sensors include one first vibration sensor arranged in the axial direction of the cutting transmission part and four second vibration sensors arranged in the radial direction of the cutting transmission part, the vibration sensors are connected to an edge computing processor through signal cables, and actual working condition fault sample data sets of the cutting part of the coal mining machine are obtained; the vibration signals of the cutting part fault simulation test bench of the coal mining machine are collected by multiple vibration sensors, the vibration sensors include one third vibration sensor arranged in the axial direction of the cutting transmission part and four fourth vibration sensors arranged in the radial direction of the cutting transmission part, the vibration sensors are connected to an edge computing processor through signal cables, and simulated working condition fault sample data sets of the cutting part of the coal mining machine are obtained.

5. The low-quality data-oriented edge-end coal cutter cutting unit fault diagnosis system according to claim 1, characterized in that, In the step two, the generator and the discriminator are respectively designed according to the balancing degree of different actual working condition samples in the actual working condition fault sample data set, virtual samples are generated, and balanced real working condition sample data sets are obtained, including: In the step two, the generator and the discriminator are respectively designed according to the balancing degree of different actual working condition samples in the actual working condition fault sample data set, virtual samples are generated, and balanced real working condition sample data sets are obtained; the Wassertein distance is used to measure the data distribution distance between the virtual samples and the actual working condition samples in the adversarial generative network; a fixed constant is set by introducing the Lipschitz continuous restriction to truncate the absolute value of the parameter update; the gradient descent stability is improved by using the RMSProp optimization algorithm, the training degree of the generator and the discriminator is balanced, and balanced actual working condition sample data sets are obtained.

6. The low-quality data-oriented edge-end coal cutter cutting unit fault diagnosis system according to claim 1, characterized in that, In the step three, the probability distributions of the actual working condition fault sample data set and the simulated working condition fault sample data set are calculated by using the probability mass function based on the distribution of the actual working condition samples in the actual working condition fault sample data set and the simulated working condition samples in the simulated working condition fault sample data set, and the simulated working condition samples in the simulated working condition sample data set that meet the migration condition are screened out, including: In the step three, the probability distributions of the actual working condition fault sample data set and the simulated working condition fault sample data set are calculated by using the probability mass function based on the distribution of the actual working condition samples in the actual working condition fault sample data set and the simulated working condition samples in the simulated working condition fault sample data set, and the simulated working condition samples in the simulated working condition sample data set that meet the migration condition are screened out, including:

7. The low-quality data-oriented edge-end coal cutter cutting unit fault diagnosis system according to claim 1, characterized in that, In the step four, the autocorrelation spectrum kurtosis diagrams of the actual working condition fault sample data set, the simulated working condition fault sample data set and the balanced real working condition sample data set are converted to obtain spectrum kurtosis diagrams; a deep feature extraction model based on a deep convolutional neural network is constructed; Based on the size of the spectral kurtosis map and the number of category labels, the size and structure parameters of the deep convolutional neural network are determined, and high-order deep features in the spectral kurtosis map are gradually extracted through convolution and pooling processing. The autocorrelation spectral kurtosis map is converted based on the actual working condition fault sample data set, the simulated working condition fault sample data set and the balanced real working condition sample data set to obtain the spectral kurtosis map; a deep feature extraction model based on a deep convolutional neural network is constructed; Based on the size of the spectral kurtosis map and the number of category labels, the size and structure parameters of the deep convolutional neural network are determined, and high-order deep features in the spectral kurtosis map are gradually extracted through convolution and pooling processing; the linear correction unit Relu function is selected as the activation function to improve the convergence speed of the deep feature extraction model by using its sparse activation characteristics; L2 regularization is introduced to limit the model parameters of the deep feature extraction model; the Momentum optimizer gradient descent optimization algorithm is selected to accelerate the learning rate and optimize the parameters of the deep convolutional neural network to obtain the deep feature extraction model.

8. The low-quality data-oriented edge-end coal cutter cutting unit fault diagnosis system according to claim 1, characterized in that, In the fifth step, the actual working condition fault sample data set, the balanced real working condition sample data set and the simulated working condition fault sample data set are used to train the deep feature extraction model, the vibration signal of the cutting part of the coal mining machine is used as the input of the deep feature extraction model, and the health status of the cutting part of the coal mining machine is used as the output of the deep feature extraction model, two parameter optimal classification networks are obtained, the feature distribution difference of each layer of the two parameter optimal classification networks is calculated, the network parameters are determined to obtain the fault diagnosis model of the cutting part of the coal mining machine, including: The balanced real working condition sample and the simulated working condition sample are used to train the deep feature extraction model to obtain two parameter optimal classification networks; the maximum mean difference MMD is used to calculate the feature distribution difference of each layer to determine the network parameter sharing limit; the Coral Loss is introduced in the full connection layer of the deep convolutional neural network to linearly change the features of the full connection layer to obtain the fault diagnosis model of the cutting part of the coal mining machine.

Images