Ball mill grinding condition detection method, storage medium and electronic equipment

Abstract

The invention provides a ball mill grinding condition detection method, a storage medium and electronic equipment. The method comprises the steps that a three-dimensional vibration signal of a ball mill cylinder is collected; the three-dimensional vibration signal of the ball mill cylinder is subjected to filtering and frequency domain transformation and then input into a preset ball mill grinding condition detection model, and a grinding condition detection value is obtained; wherein the preset ball mill grinding condition detection model has a function of obtaining a corresponding ball mill grinding condition detection value according to the input three-dimensional vibration frequency domain signal data; the preset grinding condition detection model comprises a convolution layer, a deep convolution layer, a point-by-point convolution layer, an expansion convolution layer, a pooling layer, a random inactivation layer, a multi-head attention mechanism layer, a full connection layer and a softmax classifier layer which are connected in sequence. According to the scheme, the grinding condition state detection accuracy can be improved, and the detection complexity can be reduced.

Description

Technical Field

[0001] The invention relates to the field of cement ball mill grinding condition detection, and in particular to a ball mill grinding condition detection method, a storage medium and an electronic device. Background Art

[0002] As an indispensable part of building materials, the production quality and efficiency of cement directly affect the structural strength and durability of buildings. Cement grinding, as a key link in the cement preparation process, determines the particle size and performance of the final product. In recent years, with the development of industrial automation and data-driven technology, mill grinding condition detection, as one of the key technologies for optimizing the cement production process, has attracted extensive research. By accurately predicting and real-time monitoring of mill load changes, not only can energy efficiency be significantly improved, but also the uniformity and stability of cement products can be ensured.

[0003] The ball mill is a kind of grinding equipment for cement combined grinding process. Its main body is a rotating cylinder, which is usually divided into two or more silos by partitions, and different grinding bodies are installed in each silo. The working principle of the ball mill is to first feed the cement particles into the first section of the cylinder from the feed port. Since the cylinder keeps rotating at a certain speed, the grinding body will first rise upward along the inner wall of the cylinder due to the combined effect of inertia and centrifugal force. When it reaches a certain height, the inertia cannot overcome gravity, and it will fall under the action of gravity, smashing the cement particles in the cylinder. After that, the grinding body will continue to roll, further crushing the particles, and then continue to rise and fall, repeating the above cycle. Under the repeated action of the grinding body on the cement particles, the particle size of the cement particles gradually becomes smaller, and under the action of the horizontal circulating wind, the finer particles are blown into the second section of the cylinder and continue to be ground until the particle size is small enough, and then blown out of the ball machine cylinder and enter the subsequent production link. In this process, the mill will generate vibrations of different frequencies, and these vibration frequency information are of great significance to reflecting the mill load.

[0004] There are various methods for detecting wear conditions, which are suitable for different production environments and monitoring needs. Electric power measurement is one of the most common methods. It is easy to operate, but its accuracy is greatly affected by motor efficiency and load characteristics. The acoustic signal analysis method obtains the internal state information of the mill from the sound spectrum, but it is greatly interfered by environmental noise and has high requirements for signal processing technology. The mill vibration signal is an important indicator of the mill's wear condition. Collecting and processing vibration signals to reflect the load state is an effective way to identify the ball mill's wear condition. However, the traditional vibration method has the problem of a single signal acquisition channel during the detection process and cannot characterize the mill's vibration frequency characteristics from multiple angles. In addition, neural network modeling has a high demand for computing resources, and other methods need to be introduced to reduce the amount of calculation. In view of the above problems, it is necessary to design a ball mill wear condition detection method to improve the accuracy of wear condition detection and reduce detection complexity. Summary of the invention

[0005] The purpose of the present invention is to address the problems of long time lag, noise interference, etc. of traditional grinding condition detection methods such as electric power and grinding sound, and to propose a ball mill grinding condition detection method, storage medium and electronic equipment based on the three-dimensional vibration signal of the industrial ball mill cylinder collected by a three-axis wireless vibration sensor combined with an improved convolutional neural network.

[0006] The present invention provides a method for detecting the wear condition of a ball mill, comprising:

[0007] Collect three-dimensional vibration signals of ball mill cylinder;

[0008] The three-dimensional vibration signal of the ball mill barrel is filtered and frequency-domain transformed, and then input into a preset ball mill wear condition detection model to obtain a wear condition detection value; wherein, the preset ball mill wear condition detection model has the function of obtaining the corresponding ball mill wear condition detection value according to the input three-dimensional vibration frequency-domain signal data; the preset wear condition detection model includes sequentially connected convolution layers, depth convolution layers, point-by-point convolution layers, dilated convolution layers, pooling layers, random inactivation layers, multi-head attention mechanism layers, fully connected layers and softmax classifier layers.

[0009] In some schemes of the ball mill wear condition detection method, the deep convolution layer includes one; the deep convolution layer performs convolution operation on each channel of the input feature map separately without performing weighted summation between channels; the input feature map is obtained by processing the three-dimensional vibration frequency domain signal by the convolution layer.

[0010] In the ball mill wear condition detection method described in some schemes, each convolution kernel size of the point-by-point convolution layer is 1×1, and the point-by-point convolution layer extracts the local features of the three-dimensional vibration signal of the ball mill cylinder to assist the model in identifying the signal mode under different mill wear conditions; the output feature map calculation formula of the point-by-point convolution layer is:

[0011]

[0012] The physical meanings of the above parameters are:

[0013] X h,w,c' is the value of the input feature map at position (h, w) and channel c;

[0014] W c,c' is the weight of the point-by-point convolution, representing the convolution kernel from the input channel c to the output channel c';

[0015] h, w, c and c' represent the input feature map height, input feature map width, input feature map channel number and output feature map channel number respectively.

[0016] In some embodiments of the ball mill grinding condition detection method, the dilated convolution layer includes a dilated convolution, an activation function layer, a pooling layer, and a random inactivation layer connected in sequence, and holes are inserted between elements of the convolution kernel;

[0017] The number of holes inserted is controlled by the dilation rate parameter; if the dilation rate parameter is set to d, the receptive field of the dilated convolutional layer is:

[0018] Receptive Field=K d =K+(K-1)×(d-1)

[0019] Among them, K d is the receptive field size, K is the convolution kernel size;

[0020] The calculation formula of dilated convolution is as follows:

[0021]

[0022] Among them, Y(i,j) is the value of the output feature map at position (i,j), W(m,n) is the weight of the convolution kernel, X(i,j) is the value of the input feature map at position (i,j), and (m,n) is the number of rows and columns of the W matrix.

[0023] In the ball mill wear condition detection method described in some schemes, the multi-head attention mechanism of the multi-head attention mechanism layer is designed as follows:

[0024] Independently learn h different sets of linear projections to transform Q, K and V;

[0025] Among them, Q is the query vector, which represents the input features that need to be paid attention to at present; K is the key vector, which represents the features of each position in the input sequence; V is the value vector, which represents the feature information corresponding to the key;

[0026] The h groups of transformed Q, K and V will be sent to the attention pool in parallel;

[0027] The outputs of the h attention pools are concatenated together and transformed through a learnable linear projection to produce the final output.

[0028] In some embodiments of the ball mill wear condition detection method, the implementation steps of the multi-head attention mechanism include:

[0029] According to the parameter value h, multiple groups of different weight matrices of Q, K and V are defined, denoted as Wq, Wk and Wv. Each group calculates and generates different Q, K, V, and finally learns different parameters;

[0030] According to the set number h, self-attention is performed separately to obtain h outputs Z0-Z h ;

[0031] Before outputting to the next layer, Z0-Z h Splice, multiply by W 0 The matrix is ​​transformed linearly to reduce its dimension and we get Z.

[0032] In the ball mill wear condition detection method described in some schemes, the softmax classifier layer calculates through the softmax classifier to obtain the probability of the ball mill cylinder vibration signal S in each category in the preset ball mill load state detection model input, and the calculation formula is:

[0033]

[0034] In the formula, P(y t ) represents the probability that the ball mill cylinder vibration signal corresponds to the grinding condition of the uth ball mill, s t It is the probability score of the vibration signal output by the fully connected layer before the softmax classifier corresponding to the u-th wear condition category. The wear condition category is divided into three categories, namely underload wear condition, normal load wear condition and overload wear condition.

[0035] In some schemes of the ball mill wear condition detection method, the three-dimensional vibration signal of the ball mill cylinder, which is used as the input of the ball mill wear condition detection model, is collected by a three-dimensional vibration sensor and is processed by filtering and frequency domain transformation.

[0036] The present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon, which implements the steps of any of the above methods when executed by a processor.

[0037] The present invention also provides an electronic system, comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of any one of the above methods.

[0038] The technical solution of the present invention collects the three-dimensional vibration signal of the ball mill cylinder; the three-dimensional vibration signal of the ball mill cylinder is input into a preset ball mill wear condition detection model after filtering and frequency domain transformation to obtain a wear condition detection value, which can improve the accuracy of wear condition detection and reduce detection complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of ball mill grinding condition detection in the implementation case of this application;

[0040] Figure 2(a)-Figure 2(c) The three-dimensional vibration signal time domain information corresponding to different wear conditions; among them, Figure 2 (a) is the underload wear condition, Figure 2 (b) is the normal load wear condition, and Figure 2 (c) is the overload wear condition;

[0041] Figure 3(a)-Figure 3(c) The frequency domain information of three-dimensional vibration signals corresponding to different wear conditions; among them, Figure 3(a) is the underload wear condition, Figure 3(b) is the normal load wear condition, and Figure 3(c) is the overload wear condition;

[0042] Figure 4(a)-Figure 4(c) 4(a) is a diagram of the model classification accuracy results; wherein, FIG4(a) is a traditional convolutional neural network denoted as CNN, FIG4(b) is a depthwise separable dilated convolution denoted as DP-CNN, and FIG4(c) is a depthwise separable dilated convolution combined with a multi-head attention mechanism modeling algorithm proposed in this application denoted as MD-CNN;

[0043] Figure 5 The figure is a schematic diagram of the hardware connection relationship of an electronic device for executing a ball mill wear condition detection method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be described in detail below with reference to the accompanying drawings and examples. Figure 1 As shown, the embodiment of the present application provides a ball mill wear condition detection model, comprising the following steps:

[0045] Step S1: collecting the three-dimensional vibration signal of the ball mill cylinder.

[0046] Specifically, a three-axis vibration sensor is used to collect the vibration signal of the ball mill cylinder, and a vibration signal data set is constructed by collecting signals of different grinding conditions.

[0047] The ball mill is a cement ball mill in a cement grinding workshop of a cement plant in Shandong. The three-axis vibration sensor is installed on the mill cylinder through a switching tool. The three-dimensional vibration data set is the vibration signal of three grinding conditions including underload, normal load and overload collected from the on-site ball mill during normal operation for one month. The collected time domain vibration information is as follows: Figure 2(a)-Figure 2(c) shown.

[0048] Step S2: After filtering and frequency domain transformation of the three-dimensional vibration signal of the ball mill barrel, the signal is input into a preset ball mill wear condition detection model to obtain a wear condition detection value; wherein, the preset ball mill wear condition detection model has the function of obtaining the corresponding ball mill wear condition detection value according to the input three-dimensional vibration frequency domain signal data; the preset wear condition detection model includes sequentially connected convolution layers, depth convolution layers, point-by-point convolution layers (the point-by-point convolution layers help the model to identify signal patterns under different mill wear conditions by extracting local features of the vibration signal, which can reduce data dimensions and enhance computational efficiency, making the overall model more accurate in analyzing mill wear conditions), dilated convolution layers (the dilated convolution layers can more effectively learn long-distance dependencies in vibration signals, thereby more accurately identifying complex patterns related to mill states and improving the model's ability to predict mill wear conditions), pooling layers, random inactivation layers, multi-head attention mechanism layers (the multi-head attention mechanism layers dynamically capture the relationship between various features by focusing on different parts of the input vibration signal, thereby effectively extracting key information related to mill states), fully connected layers and softmax classifier layers.

[0049] Specifically, the collected vibration signal is filtered and frequency-domain transformed, including:

[0050] The time domain vibration signal received from the sensor contains a lot of vibration information about the grinding condition of the ball mill. However, the time domain vibration signal contains noise and has a certain frequency range, so it is difficult to interpret. Therefore, the signal can be interpreted in the frequency domain after appropriate transformation. The frequency content in the sampled signal x(n) can be estimated using discrete Fourier transform:

[0051]

[0052] Where n is the number of samples, f is the sampling frequency, e (.) is the exponential operator. In the following formula, X(f) is a complex vector of length N, and the sum of the real part and the imaginary part is represented by X R (f)+jX I (f). The absolute value of this complex vector is Where |X(f)| includes the amplitude of positive and negative sampling frequencies. The first N / 2 points of |X(f)| with a scaling factor of (2 / N) represent the corresponding spectrum of x(n). Therefore, the frequency domain amplitude of the sampled vibration signal can be calculated as follows:

[0053]

[0054] X in the formula AThe calculation of (f) can be done efficiently using the Fast Fourier Transform (FFT). In this way, all distributions of the total energy of the time domain signal can be obtained in the frequency domain. The frequency domain information of each working condition after the transformation is as follows: Figure 3(a)-Figure 3(c) shown.

[0055] Construct a grinding mill wear condition detection model.

[0056] The wear detection model constructed in this example includes sequentially connected convolutional layers, depth convolutional layers, point-by-point convolutional layers, dilated convolutional layers, pooling layers, random inactivation layers (Dropout), multi-head attention mechanism layers, fully connected layers, and softmax classifier layers.

[0057] The deep convolution layer performs convolution operations on each channel of the input feature map separately, without performing weighted summation between channels. Assuming that the size of the input feature map is H×W×C and the size of the convolution kernel is K×K, the computational complexity of the traditional convolution is:

[0058] Computation Complexity=H×W×C×K 2 ×C'

[0059] Where C' is the number of channels of the output feature map. For depthwise convolution, the computational complexity is reduced by C' times, as shown below:

[0060] Computation Complexity=H×W×C×K 2

[0061] The size of each convolution kernel in the point-by-point convolution layer is 1×1, that is, for each position of the input feature map, only the corresponding channel information is processed without involving the neighborhood information in the spatial dimension. It is mainly used to adjust the channel dimension.

[0062] The calculation formula for the output feature map of point-by-point convolution is:

[0063]

[0064] The dilated convolution layer inserts holes between the elements of the convolution kernel, allowing the convolution kernel to cover a larger range of input areas, thereby increasing the receptive field. The number of holes inserted is controlled by a dilation rate parameter, which determines the spacing between the convolution kernel elements.

[0065] Assuming the dilation rate is d, the receptive field of the dilated convolution is:

[0066] Receptive Field=K d =K+(K-1)×(d-1)

[0067] The calculation formula of dilated convolution is as follows:

[0068]

[0069] Among them, Y(i,j) is the value of the output feature map at position (i,j), W(m,n) is the weight of the convolution kernel, X(i,j) is the value of the input feature map at position (i,j), and d is the expansion rate.

[0070] Furthermore, the multi-head attention mechanism layer first transforms Q, K, and V using h sets of different linear projections learned independently. Then, these h sets of transformed Q, K, and V are sent to the attention pool in parallel. Finally, the outputs of the h attention pools are concatenated together and transformed by another learnable linear projection to produce the final output. This design is called a multi-head attention mechanism. The specific steps are as follows:

[0071] Step 1. Define multiple different sets of W q , W k , W v , each group calculates and generates different Q, K, and V, and finally learns different parameters.

[0072] Step 2. Perform self-attention according to the set number h, and get h outputs Z0-Z h .

[0073] Step 3. Before outputting to the next layer, Z0-Z h Splice, multiply by W 0 The matrix is ​​transformed linearly to reduce its dimension and we get Z.

[0074] The softmax classifier is used to calculate the probability of the ball mill cylinder vibration signal S in each category in the preset ball mill load state detection model. The calculation formula is:

[0075]

[0076] In the formula, P(y i ) represents the probability that the ball mill cylinder vibration signal corresponds to the jth type of ball mill grinding condition, s i It is the probability score of the vibration signal output by the fully connected layer before the softmax classifier corresponding to the jth wear condition category. The present invention divides the wear condition category into three categories, namely underload wear condition, normal load wear condition and overload wear condition.

[0077] In this embodiment, the structure of each layer of the ball mill wear condition detection model can be built using the deep learning framework PyTorch 2.0.0 based on Python3.9.

[0078] Step S4: ball mill wear condition detection model training, taking the preprocessed frequency domain vibration signal as input, extracting the ball mill cylinder vibration signal characteristics through the ball mill wear condition detection model, and obtaining a trained ball mill wear condition detection model.

[0079] The probability of the sample belonging to each wear condition is obtained through the softmax layer, and the maximum probability is the final detection result.

[0080] Specifically, the ball mill wear condition detection model training adopts the Relu function as the activation function layer, the learning rate of each parameter is dynamically adjusted by the Adam optimization algorithm, and the batch size BatchSize is set to 16.

[0081] Step S5: Detect the wear condition of the ball mill using the trained ball mill wear condition detection model.

[0082] Experimental verification

[0083] Firstly, the proposed method for detecting mill wear condition based on three-dimensional vibration signal is compared with the traditional one-dimensional vibration signal detection method to prove the comprehensiveness and accuracy of three-dimensional vibration signal detection of mill wear condition. The acquired three-dimensional vibration signal data set is divided into three categories according to the axial direction, recorded as x-axis, y-axis and z-axis. At this time, the data set becomes one-dimensional data, which is used to simulate the traditional one-dimensional vibration signal data set. The proposed model is trained and verified respectively to evaluate the accuracy of predicting wear condition. The results are shown in Table 1.

[0084] Table 1 Detailed classification accuracy statistics of each data set

[0085]

[0086] It can be seen from the table that the x-axis data has a poor classification effect on underload and overload, the y-axis data has a poor classification effect on overload wear conditions, the z-axis data has a good classification effect on normal load, but poor classification on underload and overload, and the three-dimensional data set has a good classification effect on all types of wear conditions.

[0087] Secondly, the classification accuracy of the proposed modeling method is compared with that of the traditional convolutional neural network to evaluate the improvement of the new method in terms of model complexity, robustness, and convergence speed compared with the traditional method. The traditional convolutional neural network is denoted as CNN, the deep separable dilated convolution is denoted as DP-CNN, and the deep separable dilated convolution combined with multi-head attention mechanism modeling algorithm proposed in this application is denoted as MD-CNN. In terms of model hyperparameter settings, except for the deep convolution layer and the multi-head attention layer, other parameters remain the same. The experimental results are shown in Figure 4(a)-Figure 4(c) shown.

[0088] Among them, the CNN model failed to classify overload wear more times, the DP-CNN model made more mistakes in classifying underload wear, and the MD-CNN model performed better in classifying the three types of wear. Although the DP-CNN algorithm reduced the complexity of the CNN model and accelerated convergence, it did not have an advantage in training accuracy. The MD-CNN algorithm, thanks to the multi-head attention mechanism, can effectively improve training accuracy and enhance model prediction capabilities by learning the spatial feature relationship between data.

[0089] In summary, this application proposes a model that uses a three-axis wireless vibration sensor installed on the cylinder of an industrial mill to obtain the three-dimensional vibration signal of the mill and combines the proposed deep separable dilated convolution with a multi-head attention mechanism modeling method to detect the mill's wear condition. Through multiple experimental tests, the following conclusions are drawn:

[0090] 1. The vibration data obtained in the three-dimensional motion direction of the mill is more comprehensive than the single-dimensional vibration data (such as mill bearings) in characterizing the mill's wear condition, and the accuracy is improved by 10%;

[0091] 2. Compared with the traditional CNN algorithm, the proposed deep separable dilated convolution combined with multi-head attention mechanism modeling method reduces the model parameters by about 9.41% by adding a deep separable dilation layer; the introduction of the multi-head attention mechanism enables the model to better learn the relationship structure between data, reduces the number of iterations required for model convergence by about 30%, and improves the wear condition classification accuracy by 6.66%.

[0092] Although the present invention has been disclosed as above with preferred embodiments, the embodiments do not limit the present invention. 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. Therefore, the protection scope of the present invention should be based on the content defined by the claims of this application.

[0093] The solution provided in this application first filters and cleans the time domain data of the three-dimensional vibration signal, and then transforms it to the frequency domain through the FFT algorithm, and then uses the improved convolutional neural network prediction model to achieve training and prediction of the wear condition of the mill. The improved convolutional neural network solves the problems of large number of training parameters, weak generalization ability and inability to learn the relationship structure between data in traditional neural networks by adding a deep expansion separable module and a multi-head attention module. Experimental results show that the accuracy of three-dimensional vibration signal characterization of wear condition is 10% higher than that of one-dimensional one, and the proposed modeling method has a 13.8% faster convergence speed than the traditional convolutional neural network model, and the classification accuracy is increased by 6.7%. The proposed method provides a new idea for mill wear condition feature extraction and wear condition prediction.

[0094] An embodiment of the present application also provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the steps of the ball mill wear condition detection method described in any one of the aforementioned embodiments are implemented.

[0095] An embodiment of the present application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the ball mill wear condition detection method described in any of the aforementioned embodiments.

[0096] The present application also provides an electronic device, such as Figure 5 As shown, the electronic device includes at least one processor 51 and at least one memory 52, at least one of the memories 52 stores program information, and at least one of the processors 51 reads the program information and executes the ball mill wear condition detection method described in any of the above method embodiments. The device may also include: an input device 53 and an output device 54. The processor 51, the memory 52, the input device 53 and the output device 54 can be connected in communication. The memory 52, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The processor 51 executes various functional applications and data processing by running the non-volatile software programs, instructions and modules stored in the memory 52, that is, the ball mill wear condition detection method provided by any of the above schemes is realized. The memory 52 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required by at least one function; the data storage area may store data created according to the use of the ball mill wear condition detection method, etc. In addition, the memory 52 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 52 may optionally include a memory remotely arranged relative to the processor 51, and these remote memories may be connected to the device for executing the ball mill wear condition detection method through a network. Examples of the above-mentioned network include but are not limited to the Internet, corporate intranet, local area network, mobile communication network and a combination thereof. The input device 53 may receive input user clicks, and generate signal inputs related to user settings and function controls of the ball mill wear condition detection method. The output device 54 may include a display device such as a display screen. When the one or more modules are stored in the memory 52, when they are run by the one or more processors 51, the ball mill wear condition detection method in any of the above-mentioned method embodiments is executed.

[0097] As needed, the above technical solutions can be combined to achieve the best technical effect.

[0098] The above are only the principles and preferred embodiments of the present application. It should be noted that, for ordinary technicians in this field, on the basis of the principles of the present application, several other modifications can be made, which should also be considered as the protection scope of the present application.

Claims

1. A method for detecting the grinding condition of a ball mill, characterized in that: include: Collect three-dimensional vibration signals of ball mill cylinder; The three-dimensional vibration signal of the ball mill barrel is filtered and frequency-domain transformed, and then input into a preset ball mill wear condition detection model to obtain a wear condition detection value; wherein, the preset ball mill wear condition detection model has the function of obtaining the corresponding ball mill wear condition detection value according to the input three-dimensional vibration frequency-domain signal data; the preset wear condition detection model includes sequentially connected convolution layers, depth convolution layers, point-by-point convolution layers, dilated convolution layers, pooling layers, random inactivation layers, multi-head attention mechanism layers, fully connected layers and softmax classifier layers.

2. The ball mill wear condition detection method according to claim 1, characterized in that: The deep convolution layer includes one; the deep convolution layer performs convolution operation on each channel of the input feature map respectively without performing weighted summation between channels; the input feature map is obtained by processing the three-dimensional vibration frequency domain signal by the convolution layer.

3. The ball mill wear condition detection method according to claim 2, characterized in that: The size of each convolution kernel of the point-by-point convolution layer is 1×1. The point-by-point convolution layer extracts the local features of the three-dimensional vibration signal of the ball mill cylinder to assist the model in identifying the signal mode under different grinding conditions of the mill. The calculation formula of the output feature map of the point-by-point convolution layer is: The physical meanings of the above parameters are: X h,w,c' is the value of the input feature map at position (h, w) and channel c; W c,c' is the weight of the point-by-point convolution, representing the convolution kernel from the input channel c to the output channel c'; h, w, c and c' represent: input feature map height, input feature map width, input feature map channel number and output feature map channel number respectively.

4. The method for detecting the wear condition of a ball mill according to claim 3, characterized in that: The dilated convolution layer includes a dilated convolution, an activation function layer, a pooling layer and a random dropout layer connected in sequence, and inserts holes between elements of the convolution kernel; The number of holes inserted is controlled by the dilation rate parameter; if the dilation rate parameter is set to d, the receptive field of the dilated convolutional layer is: Receptive Field=K d =K+(K-1)×(d-1) Among them, K d is the receptive field size, K is the convolution kernel size; The calculation formula of dilated convolution is as follows: Among them, Y(i,j) is the value of the output feature map at position (i,j), W(m,n) is the weight of the convolution kernel, X(i,j) is the value of the input feature map at position (i,j), and (m,n) is the number of rows and columns of the W matrix.

5. The method for detecting the wear condition of a ball mill according to claim 4, characterized in that: The multi-head attention mechanism design of the multi-head attention mechanism layer is: Independently learn h different sets of linear projections to transform Q, K and V; Among them, Q is the query vector, which represents the input features that need to be paid attention to at present; K is the key vector, which represents the features of each position in the input sequence; V is the value vector, which represents the feature information corresponding to the key; The h groups of transformed Q, K and V will be sent to the attention pool in parallel; The outputs of the h attention pools are concatenated together and transformed through a learnable linear projection to produce the final output.

6. The method for detecting the wear condition of a ball mill according to claim 5, characterized in that: The implementation steps of the multi-head attention mechanism include: According to the parameter value h, multiple groups of different weight matrices of Q, K and V are defined, denoted as Wq, Wk and Wv. Each group calculates and generates different Q, K, V, and finally learns different parameters; According to the set number h, self-attention is performed separately to obtain h outputs Z0-Z h ; Before outputting to the next layer, Z0-Z h Splice, multiply by W 0 The matrix is ​​transformed linearly to reduce its dimension and we get Z.

7. The method for detecting the wear condition of a ball mill according to claim 6, characterized in that: The softmax classifier layer calculates the probability of the ball mill cylinder vibration signal S in each category in the preset ball mill load state detection model input through the softmax classifier, and the calculation formula is: In the formula, P(y t ) represents the probability that the ball mill cylinder vibration signal corresponds to the grinding condition of the uth ball mill, s t It is the probability score of the vibration signal output by the fully connected layer before the softmax classifier corresponding to the u-th wear condition category. The wear condition category is divided into three categories, namely underload wear condition, normal load wear condition and overload wear condition.

8. The method for detecting the wear condition of a ball mill according to claim 7, characterized in that: The three-dimensional vibration signal of the ball mill cylinder, which is the input of the ball mill grinding condition detection model, is collected by a three-dimensional vibration sensor and processed by filtering and frequency domain transformation.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

10. An electronic system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.

Images