CN115235608A - Method for detecting load state of ball mill - Google Patents
Method for detecting load state of ball mill Download PDFInfo
- Publication number
- CN115235608A CN115235608A CN202210716493.3A CN202210716493A CN115235608A CN 115235608 A CN115235608 A CN 115235608A CN 202210716493 A CN202210716493 A CN 202210716493A CN 115235608 A CN115235608 A CN 115235608A
- Authority
- CN
- China
- Prior art keywords
- ball mill
- load state
- layer
- state detection
- input
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01H—MEASUREMENT OF MECHANICAL VIBRATIONS OR ULTRASONIC, SONIC OR INFRASONIC WAVES
- G01H17/00—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves, not provided for in the preceding groups
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Crushing And Grinding (AREA)
Abstract
The invention discloses a method for detecting the load state of a ball mill, which comprises the following steps: collecting vibration signals of a ball mill cylinder; and inputting a vibration signal of the ball mill cylinder into a preset ball mill load state detection model to obtain a ball mill load state detection value. The preset ball mill load state detection model has the function of obtaining a ball mill load state detection value corresponding to an input ball mill cylinder vibration signal according to the input ball mill cylinder vibration signal. The preset ball mill load state detection model comprises a wide convolution neural network, a convolution layer, N residual shrinkage network modules, a pooling layer, a full connection layer and a softmax classifier which are sequentially connected; and the input and output of the residual shrinking network module are merged and then used as the input of the next module or network layer. Compared with the existing method, the method is more efficient, and the load state of the ball mill can be detected more accurately.
Description
Technical Field
The invention relates to the field of detection of load states of ball mills, in particular to a method for detecting the load states of the ball mills.
Background
As indispensable equipment of ore dressing operation, the operating efficiency of the ball mill influences the production efficiency and index evaluation of the whole ore mill production process and even the whole flow. With the digital transformation of mining industry, the problems of low working efficiency, high energy consumption and the like of the ball mill are more and more obvious, the running state of the ball mill is detected in time and adjusted, and the method has important practical significance for reducing the energy consumption and the equipment loss. The load of the ball mill is an important basis for judging the running state of the mill, and represents the total of materials such as water, steel balls, minerals and the like in a cylinder body of the mill. The ideal mill load directly has positive influence on the production efficiency and energy consumption of equipment and a system, but in the actual production process, the ore grinding process environment is complex and the mill runs in a closed mode, so that the components of materials in a cylinder cannot be directly observed, the mill is easily in a non-ideal load running state, namely, under-load or overload running, the under-load can generate electric energy waste, steel consumption is increased, mill equipment is damaged, the serious condition can be in or close to a blank smashing state, and the overload can cause the problems of coarsening, blockage, swelling, interruption of the production process and the like of the outlet granularity of the mill, irreversible damage is caused to the equipment, the maintenance cost is improved, the economic benefit of the mill is reduced, and even the safe production is influenced.
It is necessary to detect the load state of the ball mill and understand the working mechanism of the ball mill. When the ball mill works, the steel balls and materials move and generate strong collision along with the rotation drive of the cylinder body. In the collision process, a part of impact force of the steel ball acts on the material, so that the material is ground into powder; a portion of the kinetic energy is transferred to the cartridge wall and liner. In the operation process of the mill, a cylinder vibration signal generated by the steel ball impacting the cylinder wall can be used for representing the load state of the ball mill, the size of the vibration signal is directly influenced by the load of the mill, but the actually measured cylinder vibration signal is formed by mutually overlapping and coupling sub-signals with different periods, frequencies and amplitudes, the noise content is high, and the vibration signal and the load are difficult to describe by an accurate mathematical model.
In recent years, for load state detection research, a signal decomposition method is mostly adopted to extract principal components for signal reconstruction, characteristics such as a power spectrum, a multi-scale entropy, a singular value entropy, a frequency spectrum and the like are extracted, and finally a mode identification method is used for state detection. Although the influence of the load on the vibration signal characteristics of the cylinder is determined by the methods, the accuracy of deep detection along with algorithm optimization and research is improved, but the methods still have the following defects: 1. in the signal reconstruction process, autonomy and uncertainty exist in the selection of signal components, and part of information which can represent the load state is easy to lose in the process; 2. in the feature extraction, the features extracted by methods such as power spectrum and entropy are single and have certain dependency on expert experience, and particularly the defect of hard signal frequency band division exists in the process of extracting the frequency spectrum features; 3. the detection accuracy needs to be improved, the time consumption of feature extraction is high, and the real-time performance of the reference action on an industrial field is limited.
In view of the above problems in the prior art, a method for detecting the load state of the ball mill needs to be designed to improve the accuracy and efficiency of detecting the load state of the ball mill.
Disclosure of Invention
The present invention aims to provide a method for detecting a load state of a ball mill, which can improve the accuracy and efficiency of detecting the load state of the ball mill.
The technical scheme provided by the invention is as follows:
a method for detecting the load state of a ball mill comprises the following steps:
collecting vibration signals of a ball mill cylinder;
and inputting a vibration signal of the ball mill cylinder into a preset ball mill load state detection model to obtain a ball mill load state detection value.
The preset ball mill load state detection model has the function of obtaining a ball mill load state detection value corresponding to an input ball mill cylinder vibration signal according to the input ball mill cylinder vibration signal.
The preset ball mill load state detection model comprises a wide convolutional neural network, a convolutional layer (a convolutional layer in front of a residual shrinkage network module), N residual shrinkage network modules, a pooling layer, a full connection layer and a softmax classifier which are connected in sequence; and the input and output of the residual shrinkage network module are fused and then used as the input of the next module or network layer. The ball mill load state detection model in the application can also be called a deep wide convolution residual shrinkage network model according to the structural design.
Furthermore, the wide convolution neural network is used for improving the extraction capability of the model on the short-time characteristics of the vibration signals and comprises a convolution layer, a batch normalization layer, an activation function layer and a pooling layer which are connected in sequence.
Further, the pooling layer in the wide convolutional neural network is a maximum pooling (max pooling) layer; and selecting the maximum value of the image area as the pooled value of the area.
Further, each residual shrinkage network module comprises one or more volume blocks which are connected in sequence;
the residual shrinkage network module comprises other convolution blocks except the last convolution block, and each convolution block comprises a convolution layer, a batch normalization layer, an activation function layer and a pooling layer which are sequentially connected;
the last convolution block in the residual shrinkage network module comprises a convolution layer CONV, an attention mechanism module and a soft threshold module; the convolutional layer CONV is used for carrying out feature extraction on input features of the convolutional layer CONV to obtain first features; the attention mechanism module is used for calculating and obtaining an adaptive threshold according to the first characteristic; the soft threshold module is used for performing soft thresholding processing on the first feature according to the self-adaptive threshold to obtain a second feature;
and the input characteristic of the residual error shrinkage network module is fused with the second characteristic and then is used as the output characteristic of the residual error shrinkage network module.
Further, the attention mechanism module comprises an Absolute (Absolute) layer, a pooling layer, a first full-connection layer, a batch normalization layer, a Relu activation function layer, a second full-connection layer and a Sigmoid activation function layer which are sequentially connected; the output of the global average pooling layer is multiplied by the output of the Sigmoid activation function layer to obtain the output of the attention mechanism module, namely the self-adaptive threshold; wherein the Pooling layer adopts a Global Average Pooling (GAP) layer;
it should be understood that the sequential connection described in this application means that the output of the previous module (or network layer) is connected to the input of the next module (or network layer). The connection may be a direct connection, that is, an output signal of a previous module (or network layer) is directly used as an input signal of a next module (or network layer), or an indirect connection, that is, an output signal of a previous module (or network layer) is processed by an intermediate module (or network layer) and then used as an input signal of a next module (or network layer). In the present application, before and after are according to the sequence of signal flow, a module (or network layer) through which a signal passes first is defined as a preceding module (or network layer), and a module (or network layer) through which a signal passes later is defined as a succeeding module (or network layer).
Further, the output of the attention mechanism module is:
τ c ＝α c ·average|x i,j,c |
in the formula, τ c Is the threshold for the first feature x the c-th channel,
the result of the first feature x after being processed by absolute layer and global average pooling in the attention mechanism module is shown, wherein W and H respectively represent the width and height of the first feature x, x i,j,c Representing values of elements of the first feature x having width, height and channel index i, j and c, respectively; z is a radical of c The output of the c-th neuron of the second full connection layer in the attention mechanism module; alpha is alpha c The scaling parameter of the first feature x the c channel output by the Sigmoid activation function layer in the attention mechanism module is obtained.
Further, the soft threshold module is configured to perform soft thresholding on the first feature according to the adaptive threshold to obtain a second feature, and includes:
the output of the attention mechanism module, i.e. the adaptive threshold, is used as the input threshold of the soft threshold module, and the input feature of the soft threshold module, i.e. the first feature x, is subjected to feature mapping, where the feature mapping function (soft threshold function) is:
in the formula, y i,j,c Representing the values of the elements of the second feature y having width, height and channel index i, j and c, respectively.
Further, the ball mill load state detection model is set according to the following conditions:
T≤S (1) (2 n ×3-4)≤L
S (1) |L
wherein S is (1) The step length of the first convolution layer in the ball mill load state detection model is represented, namely the step length of the convolution layer in the wide convolution neural network; n is the total number of the convolution layers in the ball mill load state detection model; t is the length of a vibration signal acquired by one period of rotation of a ball mill cylinder, and L is the length of an input signal of a ball mill load state detection model, namely the length of a wide convolution neural network input signal; wherein S (1) L represents S (1) L can be divided evenly.
Further, the probability that the ball mill cylinder vibration signal S input into the preset ball mill load state detection model corresponds to each category is calculated through a softmax classifier, and the calculation formula is as follows:
in the formula, p j Representing the probability that the ball mill cylinder vibration signal S corresponds to the jth ball mill load state category; q. q.s j The vibration signal S of the ball mill cylinder, which is output for a full connection layer connected in front of the softmax classifier, corresponds to the fraction of the jth ball mill load state category; k is the number of the types of the load states of the ball mill; k =3; the load states of the 3 ball mills are respectively an under-load state of the ball mill, an ideal load state of the ball mill and an overload state of the ball mill.
Further, the preset ball mill load state detection model is obtained by the following method:
constructing a sample data set, wherein the sample data set comprises a plurality of sample data pairs, and each sample data pair comprises a ball mill cylinder vibration signal and a ball mill load state category label (real label) corresponding to the ball mill cylinder vibration signal;
constructing a ball mill load state detection model, wherein the input of the ball mill load state detection model is a ball mill cylinder vibration signal, and the output of the ball mill load state detection model has 3 values which respectively represent the probabilities that the input ball mill cylinder vibration signal corresponds to an under-load state of the ball mill, an ideal load state of the ball mill and an overload state of the ball mill;
training a ball mill load state detection model based on the sample data set; the trained ball mill load state detection model has the function of obtaining a ball mill load state detection value corresponding to the input ball mill cylinder vibration signal according to the input ball mill cylinder vibration signal; the ball mill load state detection value corresponding to the input ball mill cylinder vibration signal is the probability that the input ball mill cylinder vibration signal corresponds to the ball mill underload state, the ball mill ideal load state and the ball mill overload state.
Furthermore, the vibration signal of the ball mill cylinder, which is used as the input of the ball mill load state detection model, is a signal acquired by the acceleration sensor and subjected to down-sampling and normalization preprocessing. That is, firstly, preprocessing a vibration signal of a cylinder body of the ball mill; secondly, the preprocessed signals are used as input of a ball mill load state detection model, vibration features corresponding to different load states are extracted through the ball mill load state detection model, and the extracted vibration features are fused to obtain a trained ball mill load state detection model; and finally, detecting the load state of the ball mill by using the trained ball mill load state detection model.
Has the beneficial effects that:
the invention establishes an end-to-end load state detection method based on the deep neural network aiming at the characteristics of nonlinearity, instability, multi-component, high noise and the like of a cylinder vibration signal, avoids the problem of information loss in the decomposition process of the cylinder vibration signal, can independently learn the characteristics of the vibration signal facing the load state, solves the problem of dependence on expert experience in the characteristic extraction process, avoids the problems of singularity of extracted characteristics, signal transformation and threshold selection, can effectively improve the detection accuracy, and provides accurate and reliable basis for ore grinding optimization control and efficiency improvement.
Compared with the prior art, the method has the following advantages:
1) A wide convolution neural network is fused on the basis of the existing deep residual shrinkage network and is used in the field of ball mill load detection. The upper limit and the lower limit of high-low frequency characteristics obtained by a convolution kernel are improved through the wide convolution neural network, and the capacity of the model for extracting the short-time characteristics of the vibration signals facing the load state of the ball mill is effectively improved, so that the sensing capacity of the model on state information is enhanced through the wide convolution neural network. Meanwhile, a soft threshold function module in the deep residual shrinkage network is used for carrying out nonlinear transformation, and the characteristics irrelevant to the load state are removed. Moreover, the automatic learning of the threshold value is realized through an attention mechanism, and the learned characteristics are transformed, so that the influence of noise is eliminated, and the effect of important characteristics on the result is strengthened. The problem that a signal transformation method and a threshold value need to be redesigned for different signals in the existing method is solved through a soft threshold value module and an attention mechanism automatic learning signal noise reduction threshold value.
2) In the process of designing the wide convolution neural network and the deep residual shrinkage network, the structure of the deep residual shrinkage network is adaptively changed according to the receptive field design principle of the wide convolution neural network, so that the overall detection capability of the model is improved.
3) Experiments carried out by using actually measured data of an industrial field show that the ball mill load state detection method constructed by the invention has higher accuracy and less time consumption.
Compared with the existing method, the method is more efficient, can more accurately detect the load state of the ball mill, and is beneficial to improving the ore grinding efficiency and guiding the ore grinding optimization control.
Drawings
FIG. 1 is a flow chart of a ball mill load detection method disclosed in an embodiment of the present application;
FIG. 2 is an example of vibration signals corresponding to different mill load conditions; wherein, fig. 2 (a) is an under-load state, fig. 2 (b) is an ideal load state, and fig. 2 (c) is an overload state;
FIG. 3 is a schematic diagram of a ball mill load status detection model disclosed in an embodiment of the present application;
FIG. 4 is a diagram of an obfuscation matrix in an embodiment of the present application;
Detailed Description
The invention will be described in detail with reference to the drawings and examples, which are provided for better understanding of the technical solutions of the invention and are not to be construed as limiting the implementation of the invention.
The embodiment of the application provides a ball mill load detection method, as shown in fig. 1, including the following steps:
step S1: and acquiring vibration signals of the ball mill cylinder by using an acceleration sensor to construct a vibration signal data set of the ball mill cylinder.
Illustratively, the ball mill is an overflow ball mill (MQY phi 5.5 x 8.5 m) of a grinding and floating workshop of a gold ore dressing plant of the Shandong gold group coke; the acceleration vibration sensor is arranged on a cylinder body of the ball mill; the ball mill cylinder vibration signal data set is a cylinder vibration signal of a one-month field ball mill in a normal operation state.
Step S2: and (4) performing down-sampling and normalization pretreatment on the vibration signal of the cylinder body of the ball mill.
In this embodiment, the down-sampling process is fractional down-sampling. Illustratively, the length of the vibration signal in each period after the down-sampling process is 32800, the obtained vibration signal includes three states of under-load, ideal load, and overload, and a time domain waveform diagram of the load state is shown in fig. 2.
In this embodiment, the normalization process is to map the data samples to a [0,1] closed interval by a max-min normalization method, so as to avoid overall signal distortion caused by abnormal data acquisition of the acceleration sensor at a certain time, and convert each signal into a standard format:
in the formula, x i ' represents the vibration signal x of the ith ball mill cylinder i And the normalized numerical value Num represents the number of vibration signals of the ball mill cylinder.
And step S3: and constructing a ball mill load state detection model.
The ball mill load state detection model constructed in the embodiment comprises a wide convolution neural network, a convolution layer (a convolution layer in front of a residual contraction network module), N residual contraction network modules, a pooling layer, a full connection layer and a softmax classifier which are sequentially connected; and the input and the output of the residual shrinkage network module are fused and then used as the input of the next module or the network layer.
As shown in fig. 3, the wide convolutional neural network in the present embodiment exemplarily includes a convolutional layer, a batch normalization layer, an activation function layer, and a pooling layer, which are connected in sequence. In the embodiment, the N =3,3 residual error shrinkage network modules are respectively RS-block1, RS-block2 and RS-block3 in the figure. Each residual shrinking network module comprises 3 volume blocks which are connected in sequence. And the residual shrinkage network module comprises other convolution blocks except the last convolution block, and each convolution block comprises a convolution layer, a batch normalization layer, an activation function layer and a pooling layer which are sequentially connected. The last convolution block in the residual contraction network module comprises a convolution layer CONV, an attention mechanism module and a soft threshold module; the convolutional layer CONV is used for carrying out feature extraction on input features of the convolutional layer CONV to obtain first features; the attention mechanism module is used for calculating and obtaining an adaptive threshold according to the first characteristic; the soft threshold module is used for performing soft thresholding processing on the first feature according to the self-adaptive threshold to obtain a second feature. And the input characteristic of the residual error shrinkage network module is fused with the second characteristic and then is used as the output characteristic of the residual error shrinkage network module.
For example, the pooling layer in the wide convolutional neural network in the present embodiment may employ a maximum pooling layer.
The design rule of the wide convolution neural network is receptive field, and the design principle based on the ball mill vibration signal is expressed as follows:
in the embodiment, the design criteria of the ball mill load state detection model are determined based on the characteristics and the receptive field of the ball mill cylinder vibration signal. Specifically, the sensing field of the neuron in the last pooling layer (i.e., the last global average pooling layer in fig. 3) in the ball mill load state detection model is set as R (0) T is the length of the vibration signal acquired by one period of rotation of the ball mill cylinder (the number of sampling points acquired by one period of rotation of the ball mill cylinder), and L is the length of the input signal of the ball mill load state detection model (i.e., input series of the wide convolutional neural network part in fig. 3). Wherein the receptive field refers to the perception range of a neuron in the input layer. The relationship between the receptive field of the l-1 st pooling layer and the receptive field of the l-1 st pooling layer is:
R (l-1) ＝S (l) (P (l) R (l) -1)+W (l)
in the formula, R (l-1) And R (l) Respectively the l-1 st and l-th pooling layer receptive field, S (l) And W (l) Represents the step length and convolution kernel width, P, of the first convolution layer in the ball mill load state detection model (l) For the first pooling layerAnd reducing the number of sampling points. And the total number of the convolution layers in the ball mill load state detection model is recorded as n. The n convolutional layers in the ball mill load state detection model are sequenced according to the transmission sequence of signals in the model, namely the convolutional layers in the width convolutional neural network are the first convolutional layers in the mill load state detection model.
When l is greater than 1, there is S (l) ＝1，W (l) ＝3，P (l) =2, the above formula can therefore be simplified to:
R (l-1) ＝2R (l) +2
when l = n, R (n) =1, receptor field R of neuron of last pooling layer on input signal (0) Can be expressed as:
R (0) ＝S (1) (P (1) -1)+W (1)
≈S (1) (2 n ×3-4)
thus based on T ≦ R (0) L or less design criteria and S (1) The final design criteria can be obtained by dividing L evenly:
T≤S (1) (2 n ×3-4)≤L
S (1) |L
if the signal collected by the whole period of the rotation of the cylinder body of the ball mill is taken as the input signal of the model, namely T = L, T is less than or equal to R (0) Criterion of L is less than or equal to R (0) = T or R (0) = L, the above equation is thus transformed into:
S (1) (2 n ×3-4)≈T。
S (1) |T
in a basic residual error shrinking network module, 3 convolution blocks are usually included, and each convolution block includes one convolution layer, so that 3 convolution layers are usually included in a basic residual error shrinking network module. In this embodiment, according to the above design criteria, i.e. formula:
multiple value taking experiment for variable n representing total number of convolution layers in ball mill load state detection modelBased on the indexes in table 1 below, when the number of convolutional layers n =10, the loss of the test set is higher than that of n =11, but only a short training time is required to achieve fast convergence and achieve a stable effect, and a better result can be obtained in a shorter time, and when n =9 and 8, although the average training time of the model is reduced, because the step length of the l-th convolutional layer in the ball mill load state detection model (i.e. the step length of the convolutional layer in the wide convolutional neural network) needs to be correspondingly increased and the optimal parameters are adjusted, the convergence speed of the model is greatly reduced, and the accuracy and the loss are correspondingly reduced, so that n =10 is selected in the embodiment of the present application. In the load state detection model of the integral ball mill, the wide convolutional neural network comprises a convolutional layer, and the residual contraction network module comprises a convolutional layer, so that 3 convolutional blocks are respectively arranged in RS-block1 and RS-block2 in the embodiment of the application, the convolutional layers correspondingly comprise 3 convolutional layers, 2 convolutional blocks are arranged in RS-block3 and correspondingly comprise 2 convolutional layers, the total number of the convolutional layers in the load state detection model of the ball mill is 10, and the load state detection capability of the ball mill by the model is optimal.
TABLE 1 detection results corresponding to different numbers of convolutional layers in the ball mill load status detection model
The batch normalization layer in the ball mill load state detection model can enable a back propagation algorithm to effectively perform network training, and the main operation steps are similar to a standardized operation and are represented as follows:
in the formula, a i For the value of the element in the ith dimension of the input vector, len is the length of the input vector, mu B Is the mean value, σ, of the input vector B 2 Is the variance of the input vector and,
is an intermediate variable, e is a constant term, b i To output the element values in the ith dimension, γ is the scaling factor and β is the offset.
In particular, the embodiments of the present application introduce a mechanism of attention, setting different thresholds for each channel individually. The importance of a certain region of the channel is represented by the [0,1] interval, so that the threshold value is automatically learned. The attention mechanism is realized by the following steps: simplifying the input features (namely the first feature x) of the attention mechanism module into a one-dimensional vector Odv by using an absolute layer (taking absolute values of all elements of the input feature x) and a global average pooling layer, and propagating the input features (namely the first feature x) to two fully-connected layers, wherein the second fully-connected layer is provided with a plurality of neurons, the number of channels of the neurons is the same as the number of channels of the input features, and the neurons are scaled to a [0,1] interval by a Sigmoid function, so that the obtained scaling parameter can be represented as:
in the formula, z c The output of the c-th neuron of the second full connection layer in the attention mechanism module; alpha (alpha) ("alpha") c The scaling parameter of the first feature x the c channel output by the Sigmoid activation function layer in the attention mechanism module is obtained.
By multiplying the scaling parameter by the one-dimensional vector Odv, the threshold value for each channel is obtained as:
τ c ＝α c ·average|x i,j,c |
in the formula, τ c A threshold value for the first feature xth channel;
the element value of the c-th dimension in a one-dimensional vector Odv obtained after the first feature x is subjected to absolute layer and global average pooling in an attention mechanism module, wherein W and H respectively represent the width and height of the first feature x, and x i,j,c Representing the values of the elements of the first feature x with width, height and channel index i, j and c, respectively.
Specifically, the soft thresholding module applies a separate threshold to each channel of the input feature (i.e., the first feature x) of the soft thresholding module, and the soft threshold function is expressed as:
in the formula, y i,j,c The output characteristic representing the soft threshold module, i.e. the value of the elements of width, height and channel index i, j and c, respectively, in the second characteristic y.
The derivative of the soft threshold function is found by the above equation:
because the derivative of the soft threshold function is only 1 or 0, the soft threshold function is selected as the activation function layer, so that the problems of gradient disappearance, gradient explosion and the like of the model in the training process can be effectively prevented.
In this embodiment, the wide convolutional neural network, the convolutional layer connected after the wide convolutional neural network, and the features extracted by the 3 residual shrinkage network modules, namely RS-block1, RS-block2, and RS-block3, are passed through a global average pooling layer, and then transferred to a third full connection layer (output is 100 × 1) and a fourth full connection layer, and finally state classification is completed by using a softmax classifier, thereby forming the detection process of the ball mill load state detection model.
After multiplying the fourth full-connection layer by the input vector thereof through the weight matrix, adding offset to obtain the fraction of the ball mill cylinder vibration signal S input into the ball mill load state detection model corresponding to a certain ball mill load state category, wherein the fraction is expressed as:
q j ＝w j ·u+b j
＝w j1 ·u 1 +w j2 ·u 2 L+w jm ·u m +b j
wherein in the formula, q j Score, u, for ball mill barrel vibration signal S corresponding to jth ball mill load status category i Is the ith input of the fourth fully connected layer; w is a j A weight vector for the jth neuron; w is a jm A weight corresponding to the mth input quantity of the jth neuron; m is the number of input values of the fourth fully-connected layer and is the number of output values of the third fully-connected layer (in this embodiment, m takes the value of 100); bj is the bias of the jth neuron; j =1,2,3.
And encapsulating the fourth full connection layer in the input of the softmax layer, wherein the fourth full connection layer is used for enabling the input and the output of the softmax layer to be the same dimension.
The score for each category is mapped by the softmax layer to a probability between [0,1], expressed as:
in the formula, p j Representing the probability that the ball mill cylinder vibration signal S corresponds to the jth ball mill load state category; k is the number of the types of the load states of the ball mill; k =3; the load states of the 3 ball mills are respectively an under-load state of the ball mill, an ideal load state of the ball mill and an overload state of the ball mill.
In this embodiment, each layer structure of the ball mill load state detection model is built, and deep learning frame Keras 2.3.1 based on python3.7 can be used for building.
And step S4: and training a ball mill load state detection model, namely taking the preprocessed ball mill cylinder vibration signal as input, and extracting the characteristic of the ball mill cylinder vibration signal facing the ball mill load state through the ball mill load state detection model to obtain the trained ball mill load state detection model.
And obtaining the probability that the sample belongs to each load state through the softmax layer, wherein the maximum probability is the final detection result. And setting a standard through one-hot label setting, comparing the obtained load state probability of each ball mill, and setting a ball mill load state category label corresponding to the sample for finishing the training of the ball mill load state detection model. Wherein, one-hot coding is also called unique hot coding and one-bit effective coding. The method is to use a multi-bit status register to encode multiple states, each having its own independent register bit and only one of which is active at any one time.
Specifically, the ball mill load state detection model is trained, RS-block3 is set to be two convolution blocks, a Relu function is adopted as an activation function layer, a regularizer is set to be L2 (0.0001), the learning rate of each parameter is dynamically adjusted by an Adam optimization algorithm, and the batch size BatchSize is set to be 20.
Step S5: and detecting the load state of the ball mill by using the trained ball mill load state detection model.
Exemplarily, fig. 4 is a confusion matrix result diagram for engineering experimental verification by using operation data of cylinder vibration signals of one month in a field normal operation state of an overflow type ball mill (MQY Φ 5.5 × 8.5 m) in a grinding and floating workshop of a gold ore dressing plant of the jia gold group of Shandong. The prediction label is a detection result obtained by detecting the load state of the ball mill through a trained ball mill load state detection model. The detection effect accuracy of the model is very high as can be seen from the results of the confusion matrix.
Experimental verification
The method provided by the embodiment of the application is compared with various other methods for detecting the load state of the ball mill based on deep learning, and the results are shown in table 1 respectively by a ball mill load state detection method based on a deep convolutional neural network (method 1), a ball mill load state detection method based on a residual error network (method 2), a ball mill load state detection method based on a deep residual error shrinkage network (method 3), a ball mill load state detection method based on spectrum analysis (kernel principal component analysis (KPCA) and an error minimization limit learning machine (EM _ ELM), abbreviated as KPCA-EM _ ELM) (method 4), a ball mill load state detection method based on adaptive variational modal decomposition-improved power spectrum (AVMD-improved power spectrum-GA _ BP) (method 5), and a ball mill load state detection method based on a complete empirical mode decomposition (CEEMDAN) -multi-scale fuzzy entropy (MFE) -improved stacked cyclic neural network (ISRNN) (method 6).
TABLE 1 test results of different methods
According to analysis, the accuracy of the ball mill load state detection method provided by the embodiment of the application can reach 99.7%, and the ball mill load state detection method has very good ball mill load state detection capability.
In summary, the embodiment of the application provides a method capable of accurately detecting the load state of a ball mill in an ore grinding process. Theoretical analysis and actual measurement results show that: the wide convolution kernel neural network has better short-time feature extraction capability, and effectively improves the network performance; attention mechanism and soft thresholding can strengthen important features, a feature extraction method with human brain function is constructed, and the feature extraction and selection capability of the model is obviously improved; compared with the prior art, the method provided by the embodiment of the application improves the detection accuracy, reduces the time required by detection, and can provide accurate and reliable criteria for improving the ore grinding efficiency and optimizing and controlling the process.
Although the present invention has been described in terms of the preferred embodiment, it is not intended that the invention be limited to the embodiment. Any equivalent changes or modifications made without departing from the spirit and scope of the present invention also belong to the protection scope of the present invention. The scope of the invention should therefore be determined with reference to the appended claims.
Claims (10)
1. A method for detecting a load state of a ball mill is characterized by comprising the following steps:
collecting vibration signals of a ball mill cylinder;
inputting a vibration signal of a ball mill cylinder into a preset ball mill load state detection model to obtain a ball mill load state detection value; the preset ball mill load state detection model has the function of obtaining a ball mill load state detection value corresponding to an input ball mill cylinder vibration signal according to the input ball mill cylinder vibration signal; the preset ball mill load state detection model comprises a wide convolution neural network, a convolution layer, N residual shrinkage network modules, a pooling layer, a full connection layer and a softmax classifier which are sequentially connected; and the input and output of the residual shrinkage network module are fused and then used as the input of the next module or network layer.
2. The ball mill load state detection method according to claim 1, wherein the wide convolutional neural network includes a convolutional layer, a batch normalization layer, an activation function layer, and a pooling layer, which are connected in sequence.
3. A ball mill load status detection method according to claim 2, characterized in that each residual shrinkage network module comprises one or more volume blocks connected in sequence;
the residual shrinkage network module comprises other convolution blocks except the last convolution block, and each convolution block comprises a convolution layer, a batch normalization layer, an activation function layer and a pooling layer which are sequentially connected;
the last convolution block in the residual shrinkage network module comprises a convolution layer CONV, an attention mechanism module and a soft threshold module; the convolutional layer CONV is used for carrying out feature extraction on input features of the convolutional layer CONV to obtain first features; the attention mechanism module is used for calculating and obtaining an adaptive threshold according to the first characteristic; the soft threshold module is used for performing soft thresholding processing on the first feature according to the self-adaptive threshold to obtain a second feature;
and the input characteristic of the residual error shrinkage network module is fused with the second characteristic and then is used as the output characteristic of the residual error shrinkage network module.
4. The ball mill load state detection method according to claim 3, wherein the attention mechanism module comprises an absolute layer, a pooling layer, a first full connection layer, a batch normalization layer, a Relu activation function layer, a second full connection layer, a Sigmoid activation function layer, which are connected in sequence; the output of the pooling layer is multiplied by the output of the Sigmoid activation function layer to obtain the output of the attention mechanism module, namely the self-adaptive threshold; wherein the pooling layer employs a global average pooling layer.
5. The ball mill load status detection method according to claim 4, wherein the output of the attention mechanism module is:
τ c ＝α c ·average|x i,j,c |
in the formula, τ c Is the threshold for the first feature x the c-th channel,
the result of the first feature x after being processed by the absolute layer and the global average pooling in the attention mechanism module is shown, wherein W and H respectively represent the width and the height of the first feature x, and x i,j,c Representing the values of the elements of the first feature x with width, height and channel index i, j and c, respectively; z is a radical of formula c The output of the c-th neuron of the second full connection layer in the attention mechanism module; alpha (alpha) ("alpha") c The scaling parameter of the first feature x the c channel output by the Sigmoid activation function layer in the attention mechanism module is obtained.
6. The ball mill load state detection method according to claim 5, wherein the soft threshold module is configured to perform soft thresholding on the first feature according to the adaptive threshold to obtain a second feature, and comprises:
the output of the attention mechanism module, i.e. the adaptive threshold, is used as the input threshold of the soft threshold module, and the input feature of the soft threshold module, i.e. the first feature x, is subjected to feature mapping, where the feature mapping function is:
in the formula, y i,j,c Representing the values of the elements of the second feature y having width, height and channel index i, j and c, respectively.
7. A ball mill load status detection method according to any one of claims 1 to 6, characterized in that the ball mill load status detection model is set according to the following conditions:
T≤S (1) (2 n ×3-4)≤L
S (1) |L
wherein S is (1) Representing the step length of a first convolution layer in the ball mill load state detection model; n is the total number of the convolution layers in the ball mill load state detection model; t is the length of a vibration signal acquired by one period of rotation of a ball mill cylinder, and L is the length of an input signal of a ball mill load state detection model; wherein S (1) L represents S (1) L can be divided evenly.
8. A ball mill load state detection method according to claim 7, characterized in that the probability that the ball mill cylinder vibration signal S inputted into the preset ball mill load state detection model corresponds to each category is calculated by a softmax classifier, and the calculation formula is:
in the formula, p j Representing the probability that the ball mill cylinder vibration signal S corresponds to the jth ball mill load state category; q. q of j The vibration signal S of the ball mill cylinder, which is output for a full connection layer connected in front of the softmax classifier, corresponds to the fraction of the jth ball mill load state category; k is the number of the types of the load states of the ball mill; k =3; the load states of the 3 ball mills are respectively an under-load state of the ball mill, an ideal load state of the ball mill and an overload state of the ball mill.
9. A ball mill load status detection method according to claim 8, characterized in that the preset ball mill load status detection model is obtained by:
constructing a sample data set, wherein the sample data set comprises a plurality of sample data pairs, and each sample data pair comprises a ball mill cylinder vibration signal and a ball mill load state category label corresponding to the ball mill cylinder vibration signal;
constructing a ball mill load state detection model, wherein an input of the ball mill load state detection model is a ball mill cylinder vibration signal, and 3 values are output and respectively represent the probabilities that the input ball mill cylinder vibration signal corresponds to an under-load state, an ideal load state and an overload state of the ball mill;
training a ball mill load state detection model based on the sample data set; the trained ball mill load state detection model has the function of obtaining a ball mill load state detection value corresponding to the input ball mill cylinder vibration signal according to the input ball mill cylinder vibration signal; the ball mill load state detection value corresponding to the input ball mill cylinder vibration signal is the probability that the input ball mill cylinder vibration signal corresponds to the ball mill under-load state, the ball mill ideal load state and the ball mill overload state.
10. A ball mill load state detection method according to claim 9, characterized in that the ball mill cylinder vibration signal as an input of the ball mill load state detection model is a signal collected by an acceleration sensor and subjected to down-sampling and normalization preprocessing.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210716493.3A CN115235608A (en) | 2022-06-22 | 2022-06-22 | Method for detecting load state of ball mill |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210716493.3A CN115235608A (en) | 2022-06-22 | 2022-06-22 | Method for detecting load state of ball mill |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115235608A true CN115235608A (en) | 2022-10-25 |
Family
ID=83668952
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210716493.3A Pending CN115235608A (en) | 2022-06-22 | 2022-06-22 | Method for detecting load state of ball mill |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115235608A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117839819A (en) * | 2024-03-07 | 2024-04-09 | 太原理工大学 | Online multitasking mill load prediction method based on physical information neural network |
-
2022
- 2022-06-22 CN CN202210716493.3A patent/CN115235608A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117839819A (en) * | 2024-03-07 | 2024-04-09 | 太原理工大学 | Online multitasking mill load prediction method based on physical information neural network |
| CN117839819B (en) * | 2024-03-07 | 2024-05-14 | 太原理工大学 | Online multitasking mill load prediction method based on physical information neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110414788B (en) | Electric energy quality prediction method based on similar days and improved LSTM | |
| Xia et al. | Multi-stage fault diagnosis framework for rolling bearing based on OHF Elman AdaBoost-Bagging algorithm | |
| CN107092743B (en) | Blast furnace smelting process data-driven modeling method based on rule assistance | |
| CN106529814A (en) | Distributed photovoltaic ultra-short-term forecasting method based on Adaboost clustering and Markov chain | |
| CN112541532B (en) | Target detection method based on dense connection structure | |
| Xu et al. | A novel fault diagnosis method of bearing based on improved fuzzy ARTMAP and modified distance discriminant technique | |
| CN107025468A (en) | Highway congestion recognition methods based on PCA GA SVM algorithms | |
| CN108994088B (en) | HAGC cylinder fault diagnosis method and device based on PCA dimension reduction and DBN network | |
| Yun | Prediction model of algal blooms using logistic regression and confusion matrix | |
| CN111865932B (en) | Intrusion detection method based on context correlation attention mechanism and simplified LSTM network | |
| CN114676742A (en) | Power grid abnormal electricity utilization detection method based on attention mechanism and residual error network | |
| CN111046961B (en) | Fault classification method based on bidirectional long-time and short-time memory unit and capsule network | |
| CN115235608A (en) | Method for detecting load state of ball mill | |
| CN113344288A (en) | Method and device for predicting water level of cascade hydropower station group and computer readable storage medium | |
| CN116738868B (en) | Rolling bearing residual life prediction method | |
| CN115841278B (en) | Method, system, equipment and medium for evaluating running error state of electric energy metering device | |
| CN111695413A (en) | Signal first arrival pickup method and device combining U-Net and Temporal encoding | |
| CN114818817A (en) | Weak fault recognition system and method for capacitive voltage transformer | |
| CN114152442A (en) | Rolling bearing cross-working condition fault detection method based on migration convolutional neural network | |
| CN113858566A (en) | Injection molding machine energy consumption prediction method and system based on machine learning | |
| CN112614122A (en) | Element analysis method based on ore grinding image | |
| CN115271154B (en) | Nonlinear regression flood element prediction method based on polynomial and partial least square coupling | |
| CN116707918A (en) | Network security situation assessment method based on CBAM-EfficientNet anomaly detection | |
| CN116244625A (en) | Overflow type mill load indirect forecasting method based on multi-feature fusion neural network | |
| CN115034268A (en) | Bearing sub-health recognition algorithm with bidirectional LSTM fusion multi-scale convolution |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |