A method and device for grinding process fault diagnosis based on convolutional autoencoder

Abstract

The application provides a kind of based on convolutional autoencoder's grinding process fault diagnosis method and device, it is related to fault diagnosis technical field, can diagnose the fault in grinding process in real time. Including offline modeling and online fault diagnosis. Offline modeling uses fault-free data: the frequency of slurry pump, pump pool liquid level, cyclone feed concentration, cyclone feed pressure, cyclone feed flow data establish based on convolutional autoencoder fault diagnosis model. In online fault diagnosis, the above data are collected on line and the model is used to obtain data residual value and monitoring quantity;Judge whether monitoring quantity is greater than or equal to fault detection threshold. If yes, there is a fault, analyze and output the fault type, and the fault type is divided into: sensor fault, equipment fault (the fault is divided into unstable fine ore bin feed, slurry pump feed port blockage, cyclone sand trap blockage). If not, there is no fault, and the next moment fault diagnosis is carried out.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, and in particular to a grinding process fault diagnosis method and device based on a convolutional autoencoder. BACKGROUND

[0002] Ore is the raw material of modern industrial production, and ore dressing is a key link in the entire production process. The preparation work is generally crushing, grinding, screening and separation. The grinding process is one of the important processes of ore dressing, which is to grind the ore more fully after crushing. It has experienced the process from handling coarse materials to fine materials, from handling simple ores to complex ores, and from using only physical methods to using physical and chemical methods. In addition, most of the equipment in the grinding process is expensive and fixed in type, and it is difficult to accurately diagnose the fault after the fault occurs. Due to the heavy load, large impact, large water and gas, and the harsh environment of ore slurry and chemical corrosion in the grinding process, the continuity of the entire process, the mechanical equipment is prone to failure, and if a fault occurs, it will have a serious impact on subsequent production. Grinding is the core process step of ore dressing and the most important link. The grinding process is a very complex industrial system process, which often has uncertainty, nonlinearity, delay, strong coupling and incomplete information, and it is difficult to establish an accurate mathematical model.

[0003] According to the main equipment used in the grinding process: ball mill, pump pool, slurry pump, hydrocyclone. This method focuses on diagnosing the faults related to the slurry pump and cyclone. By analyzing the grinding mechanism, the frequency of the slurry pump, the feed concentration of the cyclone, the feed flow of the cyclone, the feed pressure of the cyclone, and the liquid level of the pump pool are selected to model and realize fault diagnosis, and an alarm is given to avoid subsequent escalation of harm.

[0004] The invention patent "TBS tailings underflow system" with the application number CN201220476852.4 directly sends the tailings underflow discharged by the underflow discharge valve into the medium coal arc screen through a pipeline, cancels the design of the underflow bucket and the slurry pump to reduce the equipment failure rate, and essentially changes the structure of the slurry pump equipment to reduce equipment failure, which cannot detect and diagnose the fault type. The invention patent "Non-pressure feeding three-product dense medium cyclone blockage alarm device" with the application number CN201520526174.1 proposes to use water supply type and drainage type liquid level relays to connect a sound and light alarm, which sends an alarm by detecting the change of water level. The working environment of the relay is harsh, and the relay may fail, causing false alarms. The above two methods can only detect or reduce one kind of fault, which has limitations. SUMMARY

[0005] The technical problem solved by the present application is to provide a grinding process fault diagnosis method and device based on a convolutional autoencoder to solve the problems of the prior art, which can diagnose whether the grinding process has a fault in real time and accurately provide alarm information for production personnel and provide a decision basis for operation adjustment and equipment operation and maintenance of the production process.

[0006] To solve the above technical problems, the technical solution adopted by the present application is:

[0007] In one aspect, the grinding process fault diagnosis method based on a convolutional autoencoder comprises offline modeling and online diagnosis.

[0008] First, the grinding process is introduced: after the ore is crushed, it enters the ball mill of the grinding process through the belt conveyor, and then enters the pump pool after being ground by the ball mill, and the slurry pump extracts the ore slurry in the pump pool and sends it to the cyclone, the ore particles with a particle size of-200 mesh (including-200 mesh) enter the next flotation process, and the ore particles that do not reach the particle size of-200 mesh are then sent to the ball mill, and the above operation is repeated.

[0009] The method focuses on diagnosing the faults related to the slurry pump and the cyclone, and is not applicable to diagnosing the faults related to the ball mill.

[0010] The offline modeling process is based on a convolutional autoencoder to establish a fault diagnosis model, and uses fault-free data: the frequency of the slurry pump, the feed concentration of the cyclone, the feed flow of the cyclone, the feed pressure of the cyclone, and the liquid level of the pump pool. After the fault-free data is initialized and filtered, offline modeling is performed to obtain the convolutional autoencoder model parameters of the grinding process and the fault detection thresholds of each data.

[0011] In the online diagnosis process, the frequency of the slurry pump, the feed concentration of the cyclone, the feed flow of the cyclone, the feed pressure of the cyclone, and the liquid level data of the pump pool at the current time are collected online, and the above data are initialized, filtered, and saved.

[0012] Using the convolutional autoencoder model parameters obtained by offline modeling, the prediction data of the five types of data of the slurry pump frequency, the cyclone feed concentration, the cyclone feed flow, the cyclone feed pressure, and the pump pool liquid level are reconstructed, the residual values of the prediction data and the actual real data at the same time are calculated, and the residual values are filtered; the monitoring quantity is calculated using the filtered residual values.

[0013] Using the fault detection thresholds of each data obtained by offline modeling, the fault is detected by judging whether the monitoring quantity is greater than or equal to the fault detection threshold, and the fault is classified, and the corresponding fault diagnosis result is output.

[0014] The offline modeling initialization and the online diagnosis initialization and filtering processing both include training dimension train_num, setting past step p, and setting modeling data length N; in addition to this, the online diagnosis initialization also needs to load the maximum value max, the minimum value min, the fault detection threshold value Threshold, and the grinding process convolutional autoencoder model parameters obtained by offline modeling; after offline modeling and online diagnosis initialization, the filtering processing of removing null values is performed on each data.

[0015] In the offline modeling process, the convolutional autoencoder is used to establish a fault diagnosis model based on fault-free data, including data preprocessing, establishing a convolutional autoencoder network, training the convolutional autoencoder network to obtain optimal convolutional autoencoder model parameters and each data fault detection threshold value Threshold;

[0016] The data preprocessing includes data normalization, data dimension expansion, and training data division; the data normalization is to normalize the fault-free data, i.e., the frequency of the slurry pump, the feed concentration of the cyclone, the feed flow of the cyclone, the feed pressure of the cyclone, and the liquid level of the pump pool, into the range of 0 to 1, i.e., the frequency of the slurry pump X f , the feed concentration of the cyclone X c , the feed flow of the cyclone X fl , the feed pressure of the cyclone X p , and the liquid level of the pump pool X l , and save the maximum and minimum values of the five normalized data, the maximum value of the frequency of the slurry pump X f _max and the minimum value X f _min, the maximum value of the feed concentration of the cyclone X c _max and the minimum value X c _min, the maximum value of the feed flow of the cyclone X fl _max and the minimum value X fl _min, the maximum value of the feed pressure of the cyclone X p _max and the minimum value X p _min, the maximum value of the liquid level of the pump pool X l _max and the minimum value X l _min, by constructing a matrix form containing the real data x(k) of a certain k moment of the past step p as follows:

[0017]

[0018] The data dimension expansion refers to increasing one dimension on the basis of the two-dimensional data x(k) to convert it into three-dimensional data x(N):

[0019] x(N)=[x(k) x(k+1)... x(k+N-p-1) x(k+N-p)]∈R(N-p+1)×5×p

[0020] Where k is a certain moment, p is the past step, N is the set modeling data length; dividing the training data means that the real data x(N) is sent into the network in the form of sliding window for training;

[0021] The convolutional auto-encoding network is divided into two parts of encoding and decoding, and each of the encoding and decoding is composed of two residual blocks, which contain two layers of convolutional network. The features of data x(N) are learned by the convolutional neural network auto-encoder, and the prediction data at a certain k moment is reconstructed and the overall prediction data

[0022]

[0023]

[0024] Where is the predicted frequency of the slurry pump, is the predicted feed concentration of the cyclone, is the predicted feed flow of the cyclone, is the predicted feed pressure of the cyclone, is the predicted pump pool liquid level;

[0025] The convolutional auto-encoding network is trained, the Adam optimizer is used in the training process, and the MSE is used as the loss function. The residual value at a certain k moment is obtained by subtracting the real data from the prediction data at a certain k moment r(k) is sent into the loss function for training, the model is optimized by changing the network parameters to make the loss converge, so as to obtain the optimal model parameters, and is saved; the last column r(k)_last∈R of the residual value r(k) is taken 5×1 is used to calculate the fault detection threshold of each data, that is, the N residual values of each of the five kinds of data are obtained, and the residual value is squared after being averaged: the slurry pump frequency residual square value R f , the cyclone feed concentration residual square value R c , the cyclone feed flow residual square value R fl , the cyclone feed pressure residual square value R p , and the pump pool liquid level residual square value R l ; On the basis of each residual square value, the adjustment value θ is added to obtain the fault detection threshold of each data: the slurry pump frequency fault detection threshold T f = R f + θ f , the cyclone feed concentration fault detection threshold T c = R c + θ c , the cyclone feed flow fault detection threshold T fl= R fl + θ fl cyclone feed pressure fault detection threshold T p = R p + θ p pump pool liquid level fault detection threshold T l = R l + θ l , wherein θ f , θ c , θ fl , θ p , θ l are the adjustment values of each data respectively.

[0026] In the online diagnosis process, data preprocessing, constructing residual values and filtering, calculating monitoring quantities, diagnosing faults are included.

[0027] Data preprocessing includes data normalization, expanding data dimensions, and dividing test data; loading the maximum value max and the minimum value min of each data when offline modeling, normalizing the data for online diagnosis; expanding data dimensions means adding one dimension to the original two-dimensional data to become three-dimensional data; dividing test data means sending the original data into the network in the form of a sliding window for testing.

[0028] Load the convolutional autoencoder model parameters obtained by offline modeling, and the online acquisition real data x(N')' and the k-time online acquisition real data x(k)' matrix obtained after preprocessing, N' is the number of acquisition data, wherein X f (k)' is the online acquisition frequency of the slurry pump, X c (k)' is the online acquisition cyclone feed concentration, X fl (k)' is the online acquisition cyclone feed flow, X p (k)' is the online acquisition cyclone feed pressure, X l (k)' is the online acquisition pump pool liquid level,

[0029] x(N')' = [x(k)' x(k+1)'... x(k+N'-p-1)' x(k+N'-p)'] ∈ R (N′-p+1)×5×p

[0030]

[0031] Send x(N')' into the network in the form of a sliding window, and reconstruct the corresponding online acquisition prediction data and the k-time online acquisition prediction data wherein is the online acquisition prediction slurry pump frequency, the online collected predicted cyclone feed concentration, the online collected predicted cyclone feed flow rate, the online collected predicted cyclone feed pressure, the online collected predicted pump pool liquid level,

[0032]

[0033]

[0034] Take the last column of x(k)′ as the real data at time k, denoted as x(k)′∈R 5×1 , construct the residual matrix and filter the residual values of each data;

[0035] The specific filtering method is: for each data, filter num residual values, sort the num residual values, remove the c largest values and the c smallest values, and sum the remaining residual values to obtain the last residual value of each data, where r′ f is the residual value of the frequency of the slurry pump, r′ c is the residual value of the cyclone feed concentration, r′ fl is the residual value of the cyclone feed flow rate, r′ p is the residual value of the cyclone feed pressure, and r′ l is the residual value of the pump pool liquid level;

[0036] Calculate each monitoring quantity using the last residual value to realize fault diagnosis. The specific process of calculating the monitoring quantity and performing fault diagnosis is as follows:

[0037] The monitoring quantity J is the square of the residual value r, so the calculation of the monitoring quantity of the slurry pump frequency J f =(r′ f ) 2 , the monitoring quantity of the cyclone feed concentration J c =(r′ c ) 2 , the monitoring quantity of the cyclone feed flow rate J fl =(r′ fl ) 2 , the monitoring quantity of the cyclone feed pressure J p =(r′ p ) 2 , and the monitoring quantity of the pump pool liquid level J l =(r′ l ) 2 ;

[0038] ​Load each data fault detection threshold, compare each monitoring quantity with each fault detection threshold, if the monitoring quantity is less than the fault detection threshold, there is no fault, if the monitoring quantity is greater than or equal to the fault detection threshold, a fault occurs, and then the fault type is analyzed, and the corresponding fault diagnosis result is given.

[0039] The fault diagnosis result can be specifically divided into: sensor fault of collected data, equipment fault;

[0040] Due to the poor grinding environment, the sensor sometimes fails to collect abnormal data, if the collected data has a negative value or a sudden drop or increase in a single data, and other data is normal, it is diagnosed as a sensor fault;

[0041] In addition, according to the system characteristics of the grinding process, each process data has strong coupling, and the equipment fault is analyzed by analyzing the change of each data combined with the production experience on site, and the equipment fault is divided into three types:

[0042] The first type of powder ore bin fault makes the ore feed unstable, resulting in unstable grinding process production, and the fault is called powder ore bin ore feed instability; by checking the ore feed data of the same period when the fault occurs, if a sudden drop occurs, but the grinding process is still normal production, it is diagnosed as a powder ore bin ore feed instability fault;

[0043] The second type is that the pulp concentration at the bottom of the pump pool is too high, causing the slurry pump to encounter a blockage in the upward process, and the fault is called slurry pump ore feeding port blockage; if the slurry pump ore feeding port is blocked, the cyclone ore feeding flow, ore feeding pressure and ore feeding concentration will decrease, and the pump pool liquid level will increase, and the slurry pump frequency is normal, by checking whether the data change in the same period when the fault occurs meets the above conditions, to diagnose whether there is a slurry pump ore feeding port blockage fault;

[0044] The third type is that the cyclone is blocked in the sand setting process, and the fault is called cyclone sand setting port blockage; if the cyclone sand setting port is blocked, the cyclone ore feeding pressure will be high for a long time, and the cyclone ore feeding flow and ore feeding concentration, pump pool liquid level and slurry pump frequency are normal, by checking whether the data change in the same period when the fault occurs meets the above conditions, to diagnose whether there is a cyclone sand setting port blockage fault;

[0045] On the other hand, a kind of grinding process fault detection device based on convolutional autoencoder, the device comprises an offline modeling module and an online diagnosis module;

[0046] The offline modeling module is used for offline establishment of a convolutional autoencoder network model; the module establishes a network model through fault-free data, obtains optimal network model parameters through training, and provides the optimal network model parameters, each data fault detection threshold and maximum and minimum value to the online diagnosis module;

[0047] An online diagnosis module is used to diagnose whether the grinding process is faulty, the module online collects five kinds of data of the frequency of the slag slurry pump, the ore feeding concentration of the cyclone, the ore feeding flow of the cyclone, the ore feeding pressure of the cyclone and the liquid level of the pump pool; the data is initialized: the training dimension is initialized, the modeling data length is set, the null value is removed, the grinding process convolutional autoencoder model parameters obtained by offline modeling are loaded from the data storage module, the data fault detection threshold and the maximum and minimum values are set; the data is input into the network to reconstruct the prediction data, the prediction data and the collected data are used to calculate and filter the residual value; the residual value is used to calculate the monitoring quantity; the fault diagnosis is performed by judging whether the monitoring quantity is greater than or equal to the fault detection threshold, if yes, it indicates that there is a fault, the fault type is diagnosed and output, if not, it indicates that no fault occurs, the next time data is collected for a new round of calculation. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The flow chart of the grinding process fault diagnosis method based on the convolutional autoencoder provided by the embodiment of the present application;

[0049] Figure 2 The grinding process flow chart of the embodiment of the present application;

[0050] Figure 3 The flow chart of the grinding process online fault diagnosis method based on the convolutional autoencoder provided by the embodiment of the present application;

[0051] Figure 4 The structure diagram of the grinding process fault diagnosis device based on the convolutional autoencoder provided by the embodiment of the present application

[0052] Figure 5 The output result of the grinding process in the fault-free state provided by the embodiment of the present application; the data in the figure is the normalized data, figure (5a) is the data graph of the frequency (Frequency) of the slag slurry pump, the ore feeding flow (Flow) of the cyclone, the ore feeding concentration (Concentration) of the cyclone, the liquid level (Level) of the pump pool, the ore feeding pressure (Pressure) of the cyclone and the ore feeding amount (Ore feed) of the powder ore bin in the fault-free state, figure (5b) is the fault detection result graph in the fault-free state, J is the monitoring quantity, and Threshold is the fault detection threshold;

[0053] Figure 6The diagnosis result of abnormal failure of the grinding process sensor provided by the embodiment of the present application is shown in FIG. 6. The data in the figure is normalized data. FIG. 6a is a data graph of the frequency of the slurry pump, the flow of the cyclone feed, the concentration of the cyclone feed, the level of the pump pool, the pressure of the cyclone feed, and the ore feed of the fine ore bin under the failure. FIG. 6b is an enlarged view of the abnormal point of the concentration of the cyclone feed under the failure. FIG. 6c is an enlarged view of the abnormal point of the pressure of the cyclone feed under the failure. FIG. 6d is a failure detection result graph under the failure. J is a monitored quantity, and Threshold is a failure detection threshold. FIG. 6e is a failure diagnosis classification result graph under the failure.

[0054] Figure 7 The diagnosis result of the unstable failure of the ore feed of the fine ore bin in the grinding process equipment failure provided by the embodiment of the present application is shown in FIG. 7. The data in the figure is normalized data. FIG. 7a is a data graph of the frequency of the slurry pump, the flow of the cyclone feed, the concentration of the cyclone feed, the level of the pump pool, the pressure of the cyclone feed, and the ore feed of the fine ore bin under the failure. FIG. 7b is an enlarged view of the abnormal point of the ore feed of the fine ore bin under the failure. FIG. 7c is a failure detection result graph under the failure. J is a monitored quantity, and Threshold is a failure detection threshold. FIG. 7d is a failure diagnosis classification result graph under the failure.

[0055] Figure 8 The diagnosis result of the failure of the blockage of the ore feeding port of the slurry pump and the blockage of the sand discharge port of the cyclone in the grinding process equipment failure provided by the embodiment of the present application is shown in FIG. 8. The data in the figure is normalized data. FIG. 8a is a data graph of the frequency of the slurry pump, the flow of the cyclone feed, the concentration of the cyclone feed, the level of the pump pool, the pressure of the cyclone feed, and the ore feed of the fine ore bin under the failure. FIG. 8b is a failure detection result graph under the failure. J is a monitored quantity, and Threshold is a failure detection threshold. FIG. 8c and FIG. 8d are failure diagnosis classification result graphs under the failure. DETAILED DESCRIPTION

[0056] The specific embodiments of the present application are described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the present application, but are not used to limit the scope of the present application.

[0057] As Figure 1As shown, the embodiment provides a convolutional autoencoder-based grinding process fault diagnosis method, which is specifically described as follows.

[0058] The method focuses on diagnosing faults related to the slurry pump and the cyclone, and is not applicable to diagnosing faults related to the ball mill.

[0059] The grinding process of the embodiment mainly consists of a ball mill, a pump pool, a slurry pump, and a cyclone. The ore reaching the ore bin after the crushing process is sent to the ball mill by a belt conveyor, as shown. Figure 2 In combination with the actual situation of the grinding site and the coupling relationship between the process quantities in the grinding process, five kinds of data of the frequency of the slurry pump, the feed concentration of the cyclone, the feed flow of the cyclone, the feed pressure of the cyclone, and the liquid level of the pump pool are collected for research. The above data are integrated into the factory DCS (Distributed Control System). According to steps S101-S102 in the method, Figure 1 an offline convolutional autoencoder-based fault diagnosis model is established, and the convolutional autoencoder-based fault diagnosis model is established as follows:

[0060] In the offline modeling initialization, the data length is set as N = 94500, the fault-free data of the frequency of the slurry pump, the feed concentration of the cyclone, the feed flow of the cyclone, the feed pressure of the cyclone, and the liquid level of the pump pool are collected, the training dimension train_num = 5 is set, the past step length p = 5 is set, and the data is filtered to remove null values. Then, the data is preprocessed, a convolutional autoencoder network is established, and the convolutional autoencoder network is trained to obtain the optimal convolutional autoencoder model parameters and the fault detection threshold Threshold of each data;

[0061] The data preprocessing includes data normalization, data dimension expansion, and training data division. The fault-free data of the frequency of the slurry pump, the feed concentration of the cyclone, the feed flow of the cyclone, the feed pressure of the cyclone, and the liquid level of the pump pool are normalized to between 0 and 1, i.e., the frequency of the slurry pump X f , the feed concentration of the cyclone X c , the feed flow of the cyclone X fl , the feed pressure of the cyclone X p , and the liquid level of the pump pool X l , and the maximum and minimum values of the five normalized data are saved, the maximum value of the frequency of the slurry pump X f _max = 37.735 (HZ) and the minimum value X f _min = 35.995 (HZ), the maximum value of the feed concentration of the cyclone X c _max = 63.247% and the minimum value X c_min = 57.845%, the feed concentration of the cyclone = (solid mass / total mass)%, the maximum value X fl _max = 689.104 (m 3 / h) and the minimum value X fl _min = 501.025 (m 3 / h), the maximum value X p _max = 0.064 (Mpa) and the minimum value X p _min = 0.05099 (Mpa), the maximum value X l _max = 3.618 (m) and the minimum value X l _min = 2.097 (m), by constructing a matrix form containing the real data x(k) of a certain k moment of the past step 5 as follows:

[0062]

[0063] The extended data dimension means that one dimension is added on the basis of the two-dimensional data x(k) to convert into three-dimensional data x(94500): x(94500) = [x(k) x(k+1)... x(k+94500-5-1) x(k+94500-5)] ∈ R (94500-5+1)×5×5 where k is a certain moment, 5 is the past step, and 94500 is the set modeling data length; the division of training data means that the real data x(94500) is sent into the network in the form of a sliding window for training;

[0064] The convolutional auto-encoder network is divided into two parts of encoding and decoding, and each of the encoding and decoding is composed of two residual blocks, and the residual block contains two layers of convolutional networks. The features of the data x(94500) are learned by the convolutional neural network auto-encoder, and the predicted data of a certain k moment and the overall predicted data

[0065]

[0066]

[0067] wherein is the predicted frequency of the slurry pump, is the predicted feed concentration of the cyclone, is the predicted feed flow rate of the cyclone, is the predicted feed pressure of the cyclone, is the predicted pump pool liquid level;

[0068] The convolutional autoencoder network is trained using the Adam optimizer and MSE as the loss function. The residual value at time k is obtained by subtracting the actual data from the predicted data at time k. The model is trained by feeding r(k) into the loss function and optimizing the model by changing the network parameters until the loss converges, thus obtaining the optimal model parameters, which are then saved. The last column of the residual r(k) is taken as r(k)_last∈R. 5×1 The fault detection threshold for each data point is calculated, resulting in 94,500 residual values ​​for each of the five data types. The average and squared residual values ​​are then used to obtain the squared residual value R for the slurry pump frequency. f =0.001369, the squared value of the residual concentration of the hydrocyclone feed is R c =0.000484, the squared residual value of hydrocyclone feed flow rate Rfl = 0.000289, the squared residual value of hydrocyclone feed pressure R p =0.001764, Squared value of residual liquid level in pump pool R l =0.001089; Based on the squared residual values, add the adjustment value θ to obtain the fault detection threshold for each data point: Slurry pump frequency fault detection threshold T f =R f +θ f =0.001369 + 0.00035 = 0.001719, Hydrocyclone feed concentration fault detection threshold T c =R c +θ c =0.000484+0.000026=0.00051, Hydrocyclone feed flow rate fault detection threshold T fl =R fl +θ fl =0.000289+0.000012=0.000301, Hydrocyclone feed pressure fault detection threshold T p =R p +θ p =0.001764 + 0.000143 = 0.001907, Pump pool level fault detection threshold T l =R l +θ l =0.001089 + 0.000013 = 0.001102, where θ f =0.00035, θ c =0.000026, θ fl =0.000012, θ p =0.000143, θ l =0.000013 represents the adjustment value for each data point.

[0069] See Figure 3The online fault diagnosis flowchart of the convolutional autoencoder-based grinding process fault diagnosis method shown includes the following steps S301 to S313.

[0070] S301 initializes the data length to N'=24, the training dimension train_num=5, and sets the past step length p=5.

[0071] S302 collects the data of the frequency X f of the cyclone feed concentration X c of the cyclone feed flow rate X fl of the cyclone feed pressure X p of the pump pool liquid level X l and performs a null value removal filtering process on the data, always maintaining the data length as 24, when a group of data is collected, the group of data is sent into the data length N', and the first group of data in the data length N' is deleted.

[0072] S303 data preprocessing includes data normalization, data dimension expansion, and test data division; the maximum value max and the minimum value min of each data during offline modeling are loaded, and the data for online diagnosis is normalized; the data dimension expansion means that one dimension is added to the original two-dimensional data to become three-dimensional data; the test data division means that the original data is sent into the network in the form of a sliding window for testing;

[0073] S304 loads the convolutional autoencoder model parameters obtained through offline modeling, and the online collection real data x(24)' obtained after preprocessing and the online collection real data x(k)' matrix at a certain k moment, where X f (k)' is the online collection frequency of the slag slurry pump, X c (k)' is the online collection cyclone feed concentration, X fl (k)' is the online collection cyclone feed flow rate, X p (k)' is the online collection cyclone feed pressure, X l (k)' is the online collection pump pool liquid level,

[0074] x(24)'=[x(k)' x(k+1)'... x(k+24-5-1)' x(k+24-5)']∈R (24-5+1)×5×5

[0075]

[0076] x(24)' is sent into the network in the form of a sliding window, and the corresponding online collection prediction data and the online collection prediction data at a certain k moment are reconstructed through the network, where Predicted frequency of the slurry pump for online collection, Predicted concentration of the cyclone feed for online collection, Predicted flow of the cyclone feed for online collection, Predicted pressure of the cyclone feed for online collection, Predicted pump pool liquid level for online collection,

[0077]

[0078]

[0079] S305 takes the last column of x(k)' as the predicted data at time k, denoted as x(k) S305 takes the last column of x(k)' as the predicted data at time k, denoted as x(k) S305 takes the last column of x(k)' as the predicted data at time k, denoted as x(k) S305 takes the last column of x(k)' as the predicted data at time k, denoted as x(k) S305 takes the last column of x(k)' as the predicted data at time k, denoted as x(k) S305 takes the last column of x(k)' as the predicted data at time k, denoted as x(k)

[0080] The specific filtering method is: for each data, filter every 20 residual values, sort the 20 residual values, remove the 5 largest values and the 5 smallest values, and sum the remaining residual values to obtain the last residual value of each data, where r' f r' is the residual value of the slurry pump frequency, r' c r' is the residual value of the cyclone feed concentration, r' fl r' is the residual value of the cyclone feed flow, r' p r' is the residual value of the cyclone feed pressure, r' l r' is the residual value of the pump pool liquid level;

[0081] S306 calculates each monitoring quantity using the last residual value, and the monitoring quantity J is the square of the residual value r, then the slurry pump frequency monitoring quantity J f f 2 , the cyclone feed concentration monitoring quantity J c c 2 , the cyclone feed flow monitoring quantity J fl fl 2 , the cyclone feed pressure monitoring quantity J p o 2 , the pump pool liquid level monitoring quantity J l l 2 , load the fault detection threshold of each data.

[0082] ​​​​​​​​​S307 compares each monitoring quantity with each fault detection threshold value, if the monitoring quantity is less than the fault detection threshold value, no fault, enter the next moment of diagnosis; if the monitoring quantity is greater than or equal to the fault detection threshold value, fault occurs, and then analyze the fault type, and give the corresponding fault diagnosis result.

[0083] S308 The fault diagnosis result can be divided into: sensor fault of collected data, equipment fault;

[0084] S309 Due to the poor grinding environment, the sensor sometimes fails to cause the collected data to be abnormal, if the collected data appears negative value or single data suddenly drops or increases, and other data is normal, it is diagnosed as sensor fault;

[0085] S310 In addition, according to the system characteristics of the grinding process, each process data has strong coupling, and the equipment fault is divided into three kinds by analyzing the change of each data and combining with the production experience on site:

[0086] S311 The first kind of powder ore bin fault makes the ore feed quantity unstable, which leads to unstable production of the grinding process, and the fault is called powder ore bin ore feed quantity instability; by checking the ore feed quantity data Ore feed in the same period when the fault occurs, if the sudden drop occurs, but the grinding process is still normal production, it is diagnosed as powder ore bin ore feed quantity instability fault;

[0087] S312 The second kind is due to the high concentration of the slurry at the bottom of the pump pool, which causes the slurry pump to encounter blockage in the process of lifting, and the fault is called slurry pump ore feeding port blockage; if the slurry pump ore feeding port is blocked, the cyclone ore feeding flow, ore feeding pressure and ore feeding concentration will decrease, and the pump pool liquid level will increase, and the slurry pump frequency is normal, by checking whether the data change in the same period when the fault occurs meets the above conditions, to diagnose whether there is slurry pump ore feeding port blockage fault;

[0088] S313 The third kind is that the cyclone is blocked in the process of sand setting, and the fault is called cyclone sand setting port blockage; if the cyclone sand setting port is blocked, the cyclone ore feeding pressure will rise and maintain for a long time, the cyclone ore feeding flow and ore feeding concentration, pump pool liquid level and slurry pump frequency are normal, by checking whether the data change in the same period when the fault occurs meets the above conditions, to diagnose whether there is cyclone sand setting port blockage fault;

[0089] The embodiment also provides a mine belt conveyor system fault diagnosis device, as shown in the figure, the device comprises an offline modeling module and an online diagnosis module. Figure 4

[0090] ​An offline modeling module is configured to establish an offline convolutional auto-encoder network model. The module establishes a network model by using fault-free data, obtains optimal network model parameters by training, and provides the optimal network model parameters, data fault detection thresholds, and maximum and minimum values to the online diagnosis module.

[0091] An online diagnosis module is configured to diagnose whether a grinding process is faulty and the fault type. The module collects five types of data, including the frequency of a slurry pump, the ore feeding concentration of a cyclone, the ore feeding flow of the cyclone, the ore feeding pressure of the cyclone, and the liquid level of a pump pool. The data is initialized by initializing the training dimension, setting the modeling data length, removing null values, loading the convolutional auto-encoder model parameters of the grinding process obtained by offline modeling from the data storage module, and setting the data fault detection thresholds and maximum and minimum values. The data is input into the network to reconstruct predicted data. The residual values are calculated and filtered by using the predicted data and the collected data. The monitoring quantity is calculated by using the residual values. The fault is detected by judging whether the monitoring quantity is greater than or equal to the fault detection threshold. If yes, a fault occurs and the fault type is analyzed, and the corresponding fault result is output. If no, no fault occurs, and the next moment data is collected for a new round of calculation.

[0092] The device and method of the embodiment are used to diagnose the fault of a one-stage primary grinding process in a 14000T control system of a certain mine, and three detection results as shown in Figure 5 、 Figure 6 、 Figure 7 、 Figure 8 are obtained. The data in the results are normalized. Since the process data in the grinding process has strong coupling, the fault type can be determined by the change between the data. The model uses five types of data, and the final fault detection result has five corresponding graphs. If a fault is detected in one type of data, the fault is considered to be detected. Therefore, the following fault diagnosis result graphs only show the fault diagnosis result graph of one type of data. In addition, the residual values are filtered before the monitoring quantity is calculated in the online diagnosis. Every 20 residual values are used to calculate one monitoring quantity, so the horizontal coordinates of the fault detection result graphs and the data graphs are inconsistent. In addition, when a fault occurs, it cannot be immediately detected, and there is a time difference. However, this does not affect the overall fault detection.

[0093] As shown in Figure 5As shown, this is the diagnostic result under normal and fault-free conditions. Figure (5a) is a data graph of slurry pump frequency, hydrocyclone feed flow rate, hydrocyclone feed concentration, pump pool level, hydrocyclone feed pressure, and ore feed rate in the fine ore bin. Since the monitored quantity J in Figure (5b) does not exceed the fault detection threshold, it belongs to the case where the monitored quantity is less than the fault detection threshold, which is a fault-free condition.

[0094] like Figure 6 As shown, the diagnostic results of the sensor malfunction are as follows: Figures (6a) and (6b) show that the hydrocyclone feed concentration drops from the normal value to a negative value at positions 11094 and 11095. Figures (6a) and (6c) show that the hydrocyclone feed pressure fluctuates significantly between 6700 and 6725, while other data do not fluctuate significantly at the same time. Moreover, the fault duration is short, so the fault should be a sensor malfunction. Figure (6d) shows that two monitored values ​​exceed the fault detection threshold line. From left to right, these correspond to abnormal hydrocyclone feed pressure data and abnormal hydrocyclone feed concentration data, respectively. The model accurately detected the fault. Figure (6e) shows that the model accurately classifies the fault into two categories: pressure sensor malfunction and concentration sensor malfunction.

[0095] like Figure 7 The results of the fault diagnosis for unstable feed in the fine ore bin are shown in Figure (7a). As can be seen from Figure (7b), the feed rate of the fine ore bin fluctuated between 10,000 and 15,000. As can be seen from Figure (7b), the feed rate of the fine ore bin dropped to 0 at the same position. However, the slurry pump frequency, hydrocyclone feed flow rate, hydrocyclone feed concentration, pump pool level, and hydrocyclone feed pressure were not 0. This indicates that the change in the feed rate of the fine ore bin did not affect the production of the grinding process. Although the data of the grinding process was affected, it was still in normal working condition. The fault should be unstable feed in the fine ore bin. As can be seen from the fault detection results in Figure (7c), the monitored quantity is greater than the fault detection threshold. The model accurately detected the fault. As can be seen from Figure (7d), the model accurately classified the fault into: unstable feed.

[0096] like Figure 8The shown is the slag slurry pump ore feeding port blockage and cyclone grit port blockage fault diagnosis result, it can be known from (8a) that the cyclone ore feeding pressure (Pressure) appears a long time rise around 400 position, but other data normally fluctuate, the position corresponds to the cyclone grit port blockage fault; Between 600-800 position, the cyclone ore feeding flow (Flow), ore feeding pressure (Pressure) and ore feeding concentration (Concentration) appear a long time drop, the pump pool liquid level (Level) appears rise, the slag slurry pump frequency (Frequency) and the data of powder ore bin ore feeding amount (Ore feed) are normal, the fault is slag slurry pump ore feeding port blockage, it can be known from the fault detection result graph of figure (8b) that there are two places of monitoring quantity greater than the fault detection threshold, respectively corresponding to the cyclone grit port blockage fault position (corresponding to the horizontal coordinate 15-20 position) and the slag slurry pump ore feeding port blockage fault position (corresponding to the horizontal coordinate 30-40 position), the model accurately detects the fault, it can be known from figure (8c) and figure (8d) that the model accurately classifies the fault, and the fault is classified as: cyclone grit port blockage fault and slag slurry pump ore feeding port blockage fault.

[0097] The method and device of the embodiment can diagnose whether there is a fault in the grinding process in real time, accurately provide alarm information for on-site staff, and provide a decision basis for operation adjustment and equipment operation and maintenance of the production process.

[0098] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the claims of the present application.

Claims

1. A method for grinding process fault diagnosis based on convolutional autoencoder, characterized in that: The method comprises offline modeling and online diagnosis; The grinding process is that the ore is sent into the ball mill through the belt conveyor after being crushed, and then is sent into the pump pool after being ground by the ball mill, and the slurry pump sends the ore slurry in the pump pool into the cyclone, the particle size of the ore slurry is-200 mesh, the ore particles with the particle size of-200 mesh are sent into the next flotation process, and the ore particles with the particle size not reaching-200 mesh are sent into the ball mill, and the above operation is repeated; The method is used for diagnosing faults related to the slurry pump and the cyclone, and is not applicable to diagnosing faults related to the ball mill; The offline modeling process is to establish a fault diagnosis model based on a convolutional autoencoder, and to use fault-free data, i.e., the frequency of the slurry pump, the feed concentration of the cyclone, the feed flow of the cyclone, the feed pressure of the cyclone, and the liquid level of the pump pool, to perform offline modeling after the fault-free data is initialized and filtered, so as to obtain the convolutional autoencoder model parameters of the grinding process and the fault detection thresholds of the data; In the online diagnosis process, the frequency of the slurry pump, the feed concentration of the cyclone, the feed flow of the cyclone, the feed pressure of the cyclone, and the liquid level of the pump pool are collected, and the data are initialized, filtered, and saved; The convolutional autoencoder model parameters obtained by offline modeling are used to reconstruct the predicted data of the five kinds of data, i.e., the frequency of the slurry pump, the feed concentration of the cyclone, the feed flow of the cyclone, the feed pressure of the cyclone, and the liquid level of the pump pool, the residual values of the predicted data and the actual real data at the same time are calculated, and the residual values are filtered; The residual values after filtering are used to calculate the monitoring quantity; The fault detection thresholds of the data obtained by offline modeling are used to perform fault detection by judging whether the monitoring quantity is greater than or equal to the fault detection threshold, the faults are classified, and the corresponding fault diagnosis results are output.

2. The convolutional autoencoder-based fault diagnosis method for a grinding process according to claim 1, characterized in that: The initialization of offline modeling and the initialization and filtering of online diagnosis both include training dimension train_num, setting past step p, and setting modeling data length N; in addition to the above, the initialization of online diagnosis also needs to load the maximum value max, the minimum value min, the fault detection threshold Threshold, and the convolutional autoencoder model parameters of the grinding process obtained by offline modeling; after offline modeling and online diagnosis initialization, the data are filtered to remove null values.

3. The convolutional autoencoder-based fault diagnosis method for a grinding process according to claim 1, characterized in that: In the offline modeling process, the fault diagnosis model is established based on the convolutional autoencoder using fault-free data, which includes data preprocessing, establishing a convolutional autoencoder network, and training the convolutional autoencoder network to obtain the optimal convolutional autoencoder model parameters and the fault detection threshold Threshold of each data; Data preprocessing includes data normalization, expanding data dimensions, and dividing training data. Data normalization is to normalize the five kinds of data of the slag pump frequency, the cyclone feed concentration, the cyclone feed flow, the cyclone feed pressure, and the pump pool liquid level into 0 to 1, that is, the slag pump frequency X f , the cyclone feed concentration X c , the cyclone feed flow X fl , the cyclone feed pressure X p , and the pump pool liquid level X l , and save the maximum and minimum values of the five kinds of data after normalization. The maximum value of the slag pump frequency is X f _max and the minimum value is X f _min, the maximum value of the cyclone feed concentration is X c _max and the minimum value is X c _min, the maximum value of the cyclone feed flow is X fl _max and the minimum value is X fl _min, the maximum value of the cyclone feed pressure is X p _max and the minimum value is X p _min, and the maximum value of the pump pool liquid level is X l _max and the minimum value is X l _min. The matrix form containing the real data x(k) of a certain k moment of the past step p is constructed as follows: The expanded data dimension means that one dimension is added to the two-dimensional data x(k) to convert it into three-dimensional data x(N): x(N) = [x(k) x(k+1)... x(k+N-p-1) x(k+N-p)] e R (N-p+1)×5×p Where k is a certain time, p is the past step, and N is the set modeling data length; dividing the training data means that the real data x(N) is sent into the network in the form of a sliding window for training; The convolutional auto-encoder network is divided into two parts of encoding and decoding, and each of the encoding and decoding is composed of two residual blocks, and the residual block contains a two-layer convolutional network; the convolutional neural network auto-encoder learns the features of data x(N), and reconstructs the prediction data at a certain k moment and the overall prediction data wherein is the predicted frequency of the slurry pump, is the predicted feed concentration to the cyclone, is the predicted feed flow rate to the cyclone, is the predicted feed pressure to the cyclone, is the predicted pump pit level; The convolutional auto-encoding network is trained, the Adam optimizer is used in the training process, and the MSE is used as the loss function. The residual value at time k is obtained by subtracting the real data from the predicted data at time k r(k) is sent into the loss function for training. By changing the network parameters, the model is optimized, the loss is converged, and the optimal model parameters are obtained and saved. The last column of the residual value r(k)_last∈R is taken 5×1 The fault detection threshold of each data is calculated, that is, the N residual values of each of the five kinds of data are obtained, and the residual values are averaged and squared to obtain: the slag slurry pump frequency residual square value R f , the cyclone feed concentration residual square value R c , the cyclone feed flow residual square value R fl , the cyclone feed pressure residual square value R p , and the pump pool liquid level residual square value R l On the basis of each residual square value, the corresponding adjustment value θ is added to obtain the fault detection threshold of each data: the slag slurry pump frequency fault detection threshold T f = R f + θ f , the cyclone feed concentration fault detection threshold T c = R c + θ c , the cyclone feed flow fault detection threshold T fl = R fl + θ fl , the cyclone feed pressure fault detection threshold T p = R p + θ p , and the pump pool liquid level fault detection threshold T l = R l + θ l , wherein θ f , θ c , θ fl , θ p , and θ l are the adjustment values of each data, respectively.

4. The convolutional autoencoder-based fault diagnosis method of a grinding process according to claim 1, characterized in that: In the online diagnosis process, it includes data preprocessing, constructing residual values and filtering, calculating the monitoring quantity, and diagnosing faults; The data preprocessing includes data normalization, data dimension expansion, and test data division; the maximum value max and the minimum value min of each data are loaded during offline modeling, and the data for online diagnosis are normalized; the data dimension expansion means that one dimension is added to the original two-dimensional data to become three-dimensional data; the test data division means that the original data are sent into the network in the form of sliding window for testing; Load the convolutional auto-encoder model parameters obtained through offline modeling, and the preprocessed online collected real data x(N')', as well as the matrix of a certain k time and online collected real data x(k)', N' is the number of collected data, k is a certain time, p is the past step length, wherein X f (k)' is the online collected slurry pump frequency, X c (k)' is the online collected cyclone feed concentration, X fl (k)' is the online collected cyclone feed flow, X p (k)' is the online collected cyclone feed pressure, X l (k)' is the online collected pump pool liquid level, x(N')' = [x(k)' x(k+1)'... x(k+N'-p-l)' x(k+N'-p)'] e R (N′-p+1)×5×p x(N') is sent into the network in the form of sliding window, and the corresponding online collected prediction data is reconstructed through the network and the prediction data of some k moment collected online wherein is the prediction slurry pump frequency collected online, is the prediction cyclone feed concentration collected online, is the prediction cyclone feed flow collected online, is the prediction cyclone feed pressure collected online, is the prediction pump pool liquid level collected online, Take the last column of the matrix as the prediction data at time k, denoted as Take the last column of the matrix as the real data at time k, denoted as 5×1 , construct the residual matrix and filter the residual values of each data. The filtering method is: for each data, filter every num residual value, sort the num residual values, remove c maximum values and c minimum values, and sum the remaining residual values to obtain the final residual value of each data, wherein r' f is the residual value of the frequency of the slurry pump, r' c is the residual value of the feed concentration of the cyclone, r' fl is the residual value of the feed flow of the cyclone, r' p is the residual value of the feed pressure of the cyclone, r' l is the residual value of the pump pool liquid level; The last residual value is used to calculate each monitoring quantity to realize fault diagnosis.

5. The convolutional autoencoder-based fault diagnosis method of a grinding process according to claim 4, characterized in that: The process of using the last residual value to calculate each monitoring quantity to realize fault diagnosis is as follows: The monitoring quantity J, i.e. the square of the residual value r, is calculated f = (r' f ) 2 , the cyclone feed concentration monitoring quantity J c = (r' c ) 2 , the cyclone feed flow monitoring quantity J fl = (r' fl ) 2 , the cyclone feed pressure monitoring quantity J p = (r' p ) 2 , the pump pool liquid level monitoring quantity J l = (r' l ) 2 ; The fault detection threshold of each data is loaded, each monitoring quantity is compared with each fault detection threshold, if the monitoring quantity is less than the fault detection threshold, there is no fault, if the monitoring quantity is greater than or equal to the fault detection threshold, a fault occurs, and then the fault type is analyzed to give a corresponding fault diagnosis result.

6. The convolutional autoencoder-based fault diagnosis method of a grinding process according to claim 5, characterized in that: The fault diagnosis result can be divided into: sensor fault of collected data, equipment fault; Due to the poor grinding environment, the sensor may fail to collect abnormal data, if the collected data appear negative values or sudden decrease or increase of a single data, and other data are normal, the sensor fault is diagnosed; In addition, according to the system characteristics of the grinding process, there is strong coupling between each process data, the equipment fault is divided into three types by analyzing the change of each data and combining with the production experience on site: The first type is the powder ore bin fault, which makes the ore feed unstable, leading to unstable production of the grinding process, and the fault is called powder ore bin ore feed instability; by checking the ore feed data Orefeed in the same period when the fault occurs, if the sudden increase or decrease occurs, but the grinding process is still normal, the fault of the powder ore bin ore feed instability is diagnosed; The second type is that the pulp concentration at the bottom of the pump pool is too high, causing the slurry pump to be blocked in the upward process, and the fault is called slurry pump ore feeding port blockage; if the slurry pump ore feeding port is blocked, the cyclone ore feeding flow, ore feeding pressure and ore feeding concentration will decrease, and the pump pool liquid level will increase, and the slurry pump frequency is normal, whether the slurry pump ore feeding port blockage fault occurs is diagnosed by checking whether the data change in the same period when the fault occurs meets the above conditions; The third type is that the cyclone is blocked in the sand setting process, and the fault is called cyclone sand setting port blockage; if the cyclone sand setting port is blocked, the cyclone ore feeding pressure will increase for a long time, and the cyclone ore feeding flow and ore feeding concentration, pump pool liquid level and slurry pump frequency are normal, whether the cyclone sand setting port blockage fault occurs is diagnosed by checking whether the data change in the same period when the fault occurs meets the above conditions.

7. A convolutional autoencoder-based fault diagnosis device for a grinding process, the device being configured to implement the convolutional autoencoder-based fault diagnosis method for a grinding process according to claim 1, characterized in that: The device comprises an offline modeling module and an online diagnosis module; The offline modeling module is used for offline establishment of a convolutional autoencoder network model; the module establishes a network model through fault-free data, obtains optimal network model parameters through training, and provides the optimal network model parameters, each data fault detection threshold and the maximum and minimum values to the online diagnosis module; The device comprises an offline modeling module and an online diagnosis module; An online diagnosis module is used to diagnose whether a grinding process has a fault and a fault type, the module online collects five kinds of data of a frequency of a slurry pump, a feed concentration of a cyclone, a feed flow of the cyclone, a feed pressure of the cyclone and a liquid level of a pump pool; the data is initialized: initializing training dimensions, setting modeling data length, removing null values, loading grinding process convolutional autoencoder model parameters, each data fault detection threshold and maximum and minimum values obtained from an offline modeling from a data storage module; the data is input into a network to reconstruct predicted data, and residual values are calculated and filtered by using the predicted data and the collected data; A monitoring quantity is calculated by using the residual values; fault diagnosis is performed by judging whether the monitoring quantity is greater than or equal to a fault detection threshold, if yes, a fault occurs and a fault type is analyzed, and a corresponding fault result is output, if not, no fault occurs, and next time data is collected for a new round of calculation.

Images