Method for ore recognition of a mining eccentric roll crusher

Abstract

The application provides a kind of ore identification method of mining eccentric roll crusher.The method comprises: collecting the signal of the crusher when various ores cannot be broken;The signal includes the vibration acceleration signal of the main shaft support frame, the real-time speed of the main shaft, and the stress signal of the lower jaw plate;Pretreatment is carried out on the collected signal;The preprocessed data is converted into an RGB image based on the MTF algorithm, and the RGB image corresponding to the vibration acceleration signal is used as the training data set;A DenseNet-SE ore type identification model is constructed and the model is trained;According to the ore type preliminarily confirmed by the model, the real-time speed of the main shaft and the stress signal of the lower jaw plate are combined to judge the ore type;The second confirmed ore type and the credibility of the second confirmed ore type are output, and the instruction to stop the work of the crusher is executed.The application identifies the ore that is difficult to break in the mining eccentric roll crusher, and the DenseNet-SE ore type identification model designed can improve the extraction of important features.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent ore identification in mines, and more specifically, to a method for ore identification in a mining eccentric roller crusher. Background Technology

[0002] Currently, crushers are crucial crushing equipment in the mining industry. Eccentric roller crushers, with their advantages of high crushing ratio and compact structure, have become core crushing equipment in mines. They achieve efficient crushing of ore by applying compression and impact force through the rotational motion of the eccentric rollers. However, in actual production processes, the physical properties of ores vary significantly. Traditional crushing processes often cannot determine the crushability of the ore in real time. When uncrushable materials (such as metallic foreign objects or ultra-hard ores) enter the crushing chamber, the equipment continues to operate, resulting in ineffective crushing and energy waste. This situation not only affects mining efficiency but also causes serious economic losses. Existing technologies largely rely on manual inspection to determine equipment status, which is lagging and makes it difficult to provide timely warnings of mechanical failures, potentially leading to serious accidents such as shaft breakage and bearing seizure.

[0003] Therefore, developing a detection method that can effectively identify difficult-to-crush ores in eccentric roller crushers for mining has become an urgent need in the mineral processing field. Although deep learning-based fault diagnosis technology has made significant progress in recent years, especially the fusion method combining time-frequency signal imaging with convolutional neural networks (CNNs) has shown high precision advantages in mechanical condition monitoring, there is still a significant research gap in the application of this technology paradigm under ore crushing conditions. The core challenges are concentrated in two aspects: First, the vibration signals of crushers have strong nonlinear and transient impact characteristics, requiring the design of a signal-image coding method that balances time-frequency resolution and physical interpretability; Second, the complex noise interference in industrial sites (such as ore collision noise and equipment coupling vibration) requires the classification model to have strong robust feature extraction capabilities to achieve accurate identification of abnormal conditions such as "ore jamming" and "overload". Summary of the Invention

[0004] This invention aims to solve the problems of eccentric roller crushers in mining operations in harsh environments. Based on the MTF algorithm and DenseNet-SE network, this invention proposes an ore identification method for eccentric roller crushers, which is used to identify ores that are difficult to crush by the eccentric roller crusher in real time, and prevent them from falling into abnormal states and affecting production efficiency.

[0005] Therefore, the purpose of this invention is to propose a method for ore identification in a mining eccentric roller crusher.

[0006] To achieve the above objectives, the present invention provides a method for identifying ore in a mining eccentric roller crusher. The mining eccentric roller crusher includes: a roller body, a main shaft, a main shaft support frame, a frame, and a lower jaw plate; the main shaft is connected to the roller body via bearings, and a motor drives the main shaft, which in turn drives the roller body to operate the mining eccentric roller crusher; the lower jaw plate is directly fixed to the frame with bolts, and the ore is crushed in the area near the lower jaw plate; the main shaft support frame is connected to the frame with bolts, and the main shaft support frame supports the main shaft and bearings; the ore identification method includes: Step S1: ... The eccentric roller crusher collects signals when various ores cannot be crushed, and uses these signals as a reference dataset for uncrushable ores. The collected signals include: the vibration acceleration signal of the main shaft support frame, the real-time rotational speed of the main shaft, and the force signal of the lower jaw plate. The reference dataset is labeled with the type of ore. Step S2: Preprocess the collected signals. Step S3: Convert the preprocessed time-series data into RGB images based on the MTF algorithm. Step S4: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] The RGB image corresponding to the signal is used as the training dataset; Step S5: Construct a DenseNet-SE ore type recognition model and train the DenseNet-SE ore type recognition model based on the training dataset; wherein, the DenseNet-SE ore type recognition model uses DenseNet as the backbone network and combines the SE attention mechanism to extract features from the training dataset; the output of the DenseNet-SE ore type recognition model is the preliminarily confirmed ore type; Step S6: Based on the identification result of the ore type preliminarily confirmed by the DenseNet-SE ore type recognition model, and combined with the real-time rotation speed of the main shaft and the force signal of the lower jaw plate, the ore type is comprehensively judged to perform a secondary confirmation of the ore type in the training dataset; Step S7: Directly output the secondary confirmed ore type in the training dataset, and the confidence level corresponding to the secondary confirmed ore type, and directly execute step S8; Step S8: Issue a command to stop the operation of the mining eccentric roller crusher.

[0007] Preferably, the ore identification method for the mining eccentric roller crusher further includes: Step S9: acquiring the signal of the eccentric roller crusher during the crushing of the ore to be identified, and sequentially executing steps S2 and S3; Step S10: inputting the RGB image corresponding to the vibration acceleration signal during the crushing of the ore to be identified into the trained DenseNet-SE ore type identification model, wherein the DenseNet-SE ore type identification model directly outputs the preliminary confirmed ore type of the ore to be identified; Step S11: based on the identification result of the preliminary confirmed ore type of the DenseNet-SE ore type identification model, and combined with the real-time rotational speed of the main shaft... The force signal from the lower jaw plate is used to comprehensively determine the type of ore to be identified, and to perform secondary confirmation of the type of ore to be identified; Step S12: Directly output the type of ore confirmed in step S11 and the corresponding confidence level of the type of ore confirmed in step S11; Step S13: Based on the type of ore confirmed in step S11 and the corresponding confidence level, determine whether the ore to be identified can be crushed; If the determination result is yes, execute step S14; If the determination result is no, execute step S8, and add the collected signal of the eccentric roller crusher when the ore to be identified is crushed to the reference dataset; Step S14: Issue a command for the mining eccentric roller crusher to work normally.

[0008] Preferably, in step S1, a vibration sensor is used to collect the vibration acceleration signal of the spindle support frame, a rotary encoder is used to collect the real-time rotational speed of the spindle, and a pressure sensor is used to collect the force on the lower jaw plate.

[0009] Preferably, step S2 specifically includes: step S2.1: using a wavelet adaptive threshold noise reduction algorithm to exclude useless signals from the collected signals.

[0010] Preferably, in step S4, the DenseNet-SE ore type identification model is an improved DenseNet121 convolutional network; the improved DenseNet121 convolutional network specifically includes: DenseBlock1, Transition Layer1, Dense Block2, Transition Layer2, ..., Dense BlockN connected in sequence; N Dense Blocks and (N-1) Transition Layers are alternately connected; Dense Blocks are composed of a certain number of stacked DenseLayer layers; each DenseBlockN is followed by an SE module; where N is a positive integer greater than 1.

[0011] The beneficial effects of this invention are:

[0012] The ore identification method for mining eccentric roller crushers provided by this invention can identify ores that are difficult for the crusher to break. This method can significantly reduce the damage and destruction caused by the mixing of difficult-to-crush ores during operation. Furthermore, because convolutional neural networks (CNNs) have excellent data feature extraction capabilities, they can significantly reduce the randomness of relying on manual feature extraction. Moreover, the feature extraction process of CNNs, through the combination of convolutional and pooling layers, can effectively capture information in the input data, helping the model to achieve a more abstract and higher-level understanding of the data. Therefore, the DenseNet-SE ore type identification model (a modified DenseNet121 convolutional network) designed in this invention can improve the extraction of important features and determine information weights based on the target. This characteristic not only helps improve the model's generalization ability but also reduces the cost of algorithm design.

[0013] Additional aspects and advantages of the invention will become apparent from the description which follows, or may be learned by practice of the invention. Attached Figure Description

[0014] Figure 1 A schematic flowchart of an ore identification method for a mining eccentric roller crusher according to an embodiment of the present invention is shown.

[0015] Figure 2 A schematic diagram of wavelet threshold denoising according to an embodiment of the present invention is shown;

[0016] Figure 3 A schematic diagram of the structure of a DenseNet121-SE ore type identification model according to an embodiment of the present invention is shown;

[0017] Figure 4 A schematic flowchart illustrating the training process of the DenseNet-SE ore type identification model according to an embodiment of the present invention is shown. Detailed Implementation

[0018] To better understand the above-mentioned objects, features, and advantages of the present invention, such as Figures 1 to 4 As shown in the accompanying drawings and specific embodiments, the present invention will be further described in detail below. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0020] Figure 1 A schematic flowchart illustrating an embodiment of the ore identification method for a mining eccentric roller crusher according to the present invention is shown. The mining eccentric roller crusher includes: a roller body, a main shaft, a main shaft support frame, a frame, and a lower jaw plate; the main shaft is connected to the roller body via bearings, and a motor drives the main shaft, which in turn drives the roller body to operate the mining eccentric roller crusher; the lower jaw plate is directly fixed to the frame by bolts, and the ore is crushed in the area near the lower jaw plate; the main shaft support frame is connected to the frame by bolts, and the main shaft support frame supports the main shaft and bearings; as shown... Figure 1 As shown, the ore identification method of the eccentric roller crusher for mining includes:

[0021] Step S1: Collect signals from the eccentric roller crusher when various ores cannot be crushed, and use the collected signals as a reference dataset for those that cannot be crushed; wherein, the collected signals include: the vibration acceleration signal of the main shaft support frame, the real-time rotational speed of the main shaft, and the force signal of the lower jaw plate; the reference dataset is labeled with the type of ore;

[0022] Step S2: Preprocess the acquired signal; Step S2 specifically includes: Step S2.1: Use wavelet adaptive threshold noise reduction algorithm to remove useless signals from the acquired signal;

[0023] Step S3: Convert the preprocessed time series data into RGB images based on the MTF algorithm;

[0024] Step S4: Use the RGB image corresponding to the vibration acceleration signal as the training dataset;

[0025] Step S5: Construct a DenseNet-SE ore type identification model and train the DenseNet-SE ore type identification model based on the training dataset; wherein, the DenseNet-SE ore type identification model uses DenseNet as the backbone network and combines the SE attention mechanism to extract features from the training dataset; the output of the DenseNet-SE ore type identification model is the preliminarily confirmed ore type.

[0026] Step S6: Based on the preliminary identification results of the ore types confirmed by the DenseNet-SE ore type identification model, and combined with the real-time rotation speed of the spindle and the force signal of the jaw plate, the ore types are comprehensively judged to make a secondary confirmation of the ore types in the training dataset.

[0027] Step S7: Directly output the types of ore confirmed by the second confirmation in the training dataset, as well as the confidence level corresponding to the types of ore confirmed by the second confirmation, and directly execute step S8;

[0028] Step S8: Issue a command to stop the operation of the mining eccentric roller crusher.

[0029] In this embodiment, the ore identification method for a mining eccentric roller crusher provided by the present invention can identify ores that are difficult to crush by the mining eccentric roller crusher. This method can significantly reduce the damage and destruction caused to the mining eccentric roller crusher during operation due to the mixing of difficult-to-crush ores. In addition, since convolutional neural networks have excellent data feature extraction capabilities, they can significantly reduce the randomness of relying on manual feature extraction. Moreover, the feature extraction process of convolutional neural networks, through the combination of convolutional layers and pooling layers, can effectively capture information in the input data, which helps the model to have a more abstract and higher-level understanding of the data. Therefore, the DenseNet-SE ore type identification model (DenseNet-SE ore type identification model is an improved DenseNet121 convolutional network) designed in this invention can improve the extraction of important features and determine information weights according to the target. This feature not only helps to improve the generalization ability of the model, but also reduces the cost of algorithm design.

[0030] In one embodiment of the present invention, such as Figure 1 As shown, the ore identification method for the mining eccentric roller crusher further includes:

[0031] Step S9: Collect the signal from the eccentric roller crusher during the crushing of the ore to be identified, and execute steps S2 and S3 in sequence;

[0032] Step S10: Input the RGB image corresponding to the vibration acceleration signal of the ore to be identified during crushing into the trained DenseNet-SE ore type recognition model. The DenseNet-SE ore type recognition model directly outputs the preliminary confirmed ore type of the ore to be identified.

[0033] Step S11: Based on the preliminary identification results of the ore type to be identified by the DenseNet-SE ore type identification model, and combined with the real-time rotation speed of the spindle and the force signal of the lower jaw plate, the ore type to be identified is comprehensively judged to perform a second confirmation of the ore type to be identified.

[0034] Step S12: Directly output the type of ore confirmed in step S11 and the corresponding confidence level of the type of ore confirmed in step S11.

[0035] Step S13: Based on the type of ore confirmed twice in step S11 and the corresponding confidence level, determine whether the ore to be identified can be crushed; if the determination result is yes, proceed to step S14; if the determination result is no, proceed to step S8, and add the collected signal of the eccentric roller crusher when the ore to be identified is crushed to the reference dataset.

[0036] Step S14: Issue a command to enable the mining eccentric roller crusher to operate normally.

[0037] In one embodiment of the present invention, in step S1, a vibration sensor is used to collect the vibration acceleration signal of the spindle support frame, a rotary encoder is used to collect the real-time rotational speed of the spindle, and a pressure sensor is used to collect the force on the lower jaw plate.

[0038] In one embodiment of the present invention, in step S4, the DenseNet-SE ore type identification model is an improved DenseNet121 convolutional network; the improved DenseNet121 convolutional network specifically includes: Dense Block1, Transition Layer1, Dense Block2, Transition Layer2, ..., DenseBlockN connected in sequence; N Dense Blocks and (N-1) Transition Layers are alternately connected; a Dense Block is composed of a certain number of stacked Dense Layers; each DenseBlockN is followed by an SE module; where N is a positive integer greater than 1.

[0039] In one embodiment of the present invention, an identification system for the crushed ore identification method applied to the above-mentioned eccentric roller crusher is also disclosed, such as... Figure 1 As shown, the system includes:

[0040] Data acquisition subsystem: It collects crushing status signals of the crusher by fusing information from multiple sensors;

[0041] Information processing subsystem: responsible for processing and converting the signals from the sensing subsystem, and transmitting the processed images to the decision-making subsystem;

[0042] Decision Subsystem: The system uses the trained network model to extract and analyze features from the obtained image modules, evaluates the type of ore to be crushed by the crusher based on the extracted information, and compares the information with other auxiliary signals to determine the next step of the crusher.

[0043] The following specific embodiment illustrates the ore identification method of the mining eccentric roller crusher of the present invention. The mining eccentric roller crusher includes: a roller body, a main shaft, a main shaft support frame, a frame, and a lower jaw plate; the main shaft is connected to the roller body via bearings, and a motor drives the main shaft, which in turn drives the roller body to operate the mining eccentric roller crusher; the lower jaw plate is directly fixed to the frame by bolts, and the ore is crushed in the area near the lower jaw plate; the main shaft support frame is connected to the frame by bolts, and the main shaft support frame supports the main shaft and the bearings. The ore identification method of the mining eccentric roller crusher in this specific embodiment is as follows... Figure 1 As shown, this can be achieved through the following steps:

[0044] (1) Step S1: Collect signals from the eccentric roller crusher when various ores cannot be crushed, and use the collected signals as a reference dataset for those that cannot be crushed; wherein, the collected signals include: the vibration acceleration signal of the main shaft support frame, the real-time rotational speed of the main shaft, and the force signal of the lower jaw plate; the reference dataset is labeled with the type of ore; in step S1, the vibration acceleration signal of the main shaft support frame during the movement of the mechanism is collected using a vibration sensor, the signal is integrated using a data acquisition card and a host computer, the real-time rotational speed of the main shaft is collected using a rotary encoder (for auxiliary acquisition), and the force on the lower jaw plate during crushing is collected using a pressure sensor (for auxiliary acquisition).

[0045] In step S1, two sets of vibration sensors are vertically installed on both sides of the main shaft support frame of the eccentric roller crusher using a rotating magnet adsorption method. The installation positions of the sensors in the same set require that the axes of the two sensors are orthogonal to the axis of the main shaft. Each set of vibration sensors collects two sets of vibration data. A data acquisition card collects the vibration analog signals from the vibration sensors and converts them into electrical signals, realizing the acquisition of high-frequency analog quantities. The host computer receives data from the data acquisition card. A rotary encoder is installed on the bracket in contact with the flywheel, and a pressure sensor is installed between the lower jaw plate and the support device.

[0046] In step S1, common hard-to-break ores include: granite, basalt, quartzite, copper-bearing and other dense massive ores. Four labels can be established for each type as a reference. At the same time, in order to support subsequent training, an additional set of hard-to-break ores is added to store the hard-to-break ores identified in the work, for a total of five reference sets.

[0047] In this specific embodiment, a laboratory eccentric roller crusher was used to collect data and create a dataset for training the DenseNet-SE ore type identification model that was subsequently constructed.

[0048] Specifically, in this embodiment, the vibration sensor used for acquiring vibration signals is a YD25 YD25 piezoelectric accelerometer (IEPE type). The data acquisition card used is an Altair Technology USB bus data acquisition card, providing 8 RSE / NRSE channels for analog input. The host computer runs Windows 10 with 16GB of RAM and 128GB of storage. The pressure sensor used for acquiring pressure signals is a JLBU-1 spoke-type tension / compression sensor (Junwan Jinno type). An Omron rotary encoder is used to acquire the spindle speed. Four vibration sensors are installed vertically on both sides of the eccentric roller crusher's spindle using a rotating magnet adsorption method, with the axes of the two sensors in the same group being orthogonal to the spindle axis. Each group of vibration sensors acquires two sets of vibration data. The data acquisition card acquires the analog vibration signals from the sensors and converts them into electrical signals, achieving high-frequency analog signal acquisition. The host computer receives data from the data acquisition card. The pressure sensor is installed between the lower jaw plate and the support device, and the rotary encoder is installed on the bracket in contact with the flywheel.

[0049] (2) Step S2: Preprocess the acquired signal; Step S2 specifically includes: Step S2.1: Use wavelet adaptive threshold noise reduction algorithm to remove useless signals in the acquired signal.

[0050] Because the sensor is affected by the complex working environment of the eccentric roller crusher, which can interfere with the accuracy of subsequent feature extraction, this specific embodiment uses a wavelet adaptive threshold noise reduction algorithm to eliminate useless signals.

[0051] In step S2, wavelet adaptive threshold denoising is a common filtering method. Its main principle is to decompose the noisy signal layer by layer into approximate and detail components at each different scale. The approximate components represent the low-frequency part of the signal, while the detail components represent the high-frequency part, i.e., the noisy part. Figure 2 As shown, this is the basic process of wavelet thresholding denoising. Wavelet decomposition requires a suitable wavelet function. In this specific example, db14 is chosen as the wavelet basis function based on the collected vibration signal. The basic process involves selecting a threshold based on the signal characteristics, dividing the signal into two parts according to the threshold, retaining the values ​​greater than the threshold, and removing the values ​​less than the threshold. Then, the decomposed signal is reconstructed using inverse wavelet transform to obtain a noise-free signal (i.e., the denoised signal). The most crucial aspect of this algorithm is selecting a suitable threshold. To adapt the threshold to different working conditions, this specific embodiment uses an adaptive threshold, the specific mathematical expression of which is as follows:

[0052]

[0053] In equation (1): N is the signal length; l k The adaptive threshold for layer-by-layer decomposition; s k The specific formula for calculating the noise standard deviation is as follows:

[0054]

[0055] (3) Step S3: Convert the preprocessed time series data into an RGB image based on the MTF algorithm; Since deep learning models are mostly built on two-dimensional data and two-dimensional data can provide more comprehensive information, this specific embodiment uses MTF (Markov Transfer Field) to convert the obtained one-dimensional time series signal (preprocessed time series data) into a two-dimensional RGB image as the model input.

[0056] In step S3, MTF is an effective method for converting one-dimensional signals into two-dimensional images. The basic idea is to capture the Markov transition probabilities between different time points in the time series and arrange them into a transition probability matrix. The aggregated image is obtained by MTF matrix encoding, and the pressure signal and rotation speed signal are directly transmitted to the decision system. The vibration information in the vertical and horizontal directions of the spindle plumb surface contains effective information. After MTF conversion, the vibration data in these two directions are used to generate grayscale images using MTF. The vertical MTF image is assigned to the blue channel (R), the horizontal MTF grayscale image is assigned to the green channel (G), and the red channel (B) is assigned a value of 0, thereby realizing the generation of a color RGB image of the vibration data of the spindle in a specific time period.

[0057] The MTF transformation process has low computational cost, and the introduction of quantiles enhances its robustness against interference. By using MTF to change the data dimensionality, one-dimensional time data is transformed into a two-dimensional image, which is then used as input for model training. The temporal information is entirely transformed into information at the image pixels, avoiding the loss of important information. Markov chains are memoryless; the transition probability of the current state depends only on its previous state. Therefore, the Markov transition probability matrix constructed based on this property is shown in the following formula:

[0058]

[0059] In equation (3), m ij quantile q j Transition to quantile q i The transition probability, i.e., m ij =P(q) i →q j ), where q i ,q j ∈[1,H]. Matrix M is an n*n matrix, where n is the time series length.

[0060] Due to the versatility of RGB images, models pre-trained on large-scale datasets can be easily fine-tuned for specific tasks. In this example, since MTF generates grayscale images, in order to convert them into RGB images, based on the fact that the vertical and horizontal axes contain more information, the vertical MTF image is assigned to the blue channel (R), the horizontal MTF grayscale image is assigned to the green channel (G), and the red channel (B) is assigned a value of 0, thereby generating an RGB image as input.

[0061] (4) When performing step S3, step S4 is performed simultaneously; Step S4: Use the RGB image corresponding to the vibration acceleration signal as the training dataset;

[0062] (5) Step S5: Construct a DenseNet-SE ore type identification model and train the DenseNet-SE ore type identification model based on the training dataset; wherein, the DenseNet-SE ore type identification model uses DenseNet as the backbone network and combines the SE attention mechanism to extract features from the training dataset; the output of the DenseNet-SE ore type identification model is the preliminarily confirmed ore type;

[0063] Specifically, the types of minerals were initially determined by comparing the established DenseNet-SE mineral type identification model with the reference dataset.

[0064] Specifically, Figure 4 The network training process for the DenseNet-SE ore type identification model, such as Figure 4 As shown, after the designed ore identification model for the eccentric roller crusher is formed, the model is trained and iterated using the dataset of difficult-to-crush ores from the eccentric roller crusher. The PyTorch framework is used to load and process the data, training the designed MFT-DenseNet121-SE-based eccentric roller crusher state identification model. In step S5, the DenseNet-SE ore type identification model is an improved DenseNet121 convolutional network. This DenseNet-SE ore type identification model uses DenseNet121 as the overall framework, combining the SE attention mechanism to extract features from the obtained RGB two-dimensional images. DenseNet consists of 4 Dense Blocks and 3 Transition Layers. An SE module is added after the Cov3*3 of each Dense Layer, using SE to adjust the original feature map according to the weights. Its specific expression is as follows:

[0065]

[0066] The model's output layer uses the softmax activation function. The softmax function standardizes the output features and outputs the probabilities, thus showing the probability that a sample feature belongs to each category. This output layer uses the softmax activation function.

[0067] The mathematical expression for the softmax function is:

[0068]

[0069] In equation (5), P(j) represents the value of the network input data to be classified as the j-th (j = 1, 2, ..., n)-th. f The probability of class n f The total number of categories, Let be the input to the j-th neuron in the m-th fully connected layer, and its functional formula is:

[0070]

[0071] In equation (6), The output of the k-th neuron in the (m-1)-th layer of the network can be expressed as the formula:

[0072]

[0073] In equation (7), fr() is the ReLU activation function.

[0074] That is, the improved DenseNet121 convolutional network specifically includes: Dense Block1, Transition Layer1, Dense Block2, Transition Layer2, ..., Dense BlockN connected in sequence; N Dense Blocks and (N-1) Transition Layers are alternately connected; Dense Block is composed of a certain number of stacked Dense Layers; each Dense BlockN is followed by an SE module; where N is a positive integer greater than 1.

[0075] The DenseNet121 convolutional network consists of an input layer, convolutional layers, pooling layers, dense blocks, and fully connected layers. Compared to other convolutional networks, DenseNet121 has a more aggressive dense connection mechanism; each layer is connected to all channels of all preceding layers and serves as the input to the next layer. Its mathematical model is as follows:

[0076] x i =H i ([x0,x1,…,xi-1 (8)

[0077] In equation (8), x i This is the output for the i-th layer; [x0, x1, ..., x i-1 ] represents the feature map connections for layers 0, 1, ..., i-1; H i It is a non-linear transformation function.

[0078] DenseNet continuously stitches together feature maps from previous layers, allowing subsequent layers to directly reuse features from earlier layers. This enables richer feature fusion and reduces repetitive learning. However, under such a comprehensive approach, DenseNet struggles to distinguish the importance of features at each layer. To overcome this issue, this specific implementation introduces the Squeeze-and-Excitation (SE) attention mechanism. SE allows the model to focus more on important parts of the input feature map, improving its feature learning ability and reducing computational intensity. Its basic idea is to dynamically adjust the importance of each channel in the feature map using global information.

[0079] In step S5, the core idea of ​​DenseNet is dense connectivity, ensuring that each layer is directly connected to all previous layers. This means that the input of a layer includes not only the output of the previous layer but also the outputs of all preceding layers. Internally, it consists of multiple Dense Blocks and Transition Layers stacked alternately. A Dense Block is essentially a stack of Dense Layers, which are the most basic atomic units of the entire model, performing a basic feature extraction. It typically consists of BN + ReLU + 1*1 Conv + BN + ReLU + 3*3 Conv. Dense connections occur between different Dense Layers within a Dense Block. The Transition module contains BN + ReLU + 1*1 Conv + 2*2 AvgPool. The 1*1 Conv reduces the number of channels, and the 2*2 AvgPool reduces the feature layer width to half its original width. The Transition module connects different Dense Blocks. Establishing direct connections across all layers within a dense block effectively improves training efficiency and alleviates gradient vanishing compared to traditional convolutional networks. This helps extract richer and more representative features, mitigating the vanishing and exploding gradient problems, thus enabling deeper networks to be trained. However, the Dense Block model has a weak ability to capture features at different scales. Therefore, to enhance the feature capture capability of the Dense Block, the core of this specific implementation is to introduce an SE attention mechanism into the Dense Block: Specifically, 1. An SE Block is inserted after the 3*3 convolution of each Dense Layer. In the Squeeze part, global pooling is used to convert the feature map into a 1*1*C tensor. In the Excitation part, a two-layer fully connected gate mechanism is used. The first fully connected layer compresses it into C / r channels for easier subsequent calculations. Then, the ReLU activation function is used, and the second fully connected layer restores the channels to C. The Sigmoid activation is then used to obtain the weights s. Finally, a Scale operation is used to obtain the weighted features. 2. An additional SE module is added after the final Dense Block.

[0080] like Figure 3 As shown below, the network processing flow of the DenseNet-SE ore type identification model proposed in this specific embodiment will be described in detail:

[0081] A. Input the RGB image converted from MTF, and output a size of 3*224*224.

[0082] B. After 7*7 convolution + BN + ReLU + 3*3 pooling, the output size is 64*56*56.

[0083] C. After passing through Dense Block 1, which contains 6 Dense Layers including SE modules, the output size is 256*56*56.

[0084] D. After BN+ReLU+1*1 convolution (128 channels)+2*2 pooling, the output is 128*28*28.

[0085] E. After passing through Dense Block 2, which contains 12 Dense Layers including SE modules, the output size is 512*28*28.

[0086] F. After BN+ReLU+1*1 convolution (256 channels)+2*2 pooling, the output is 256*14*14.

[0087] G. After passing through Dense Block 3, which contains 24 Dense Layers including SE modules, the output size is 1024*14*14.

[0088] H. After BN+ReLU+1*1 convolution (512 channels)+2*2 pooling, the output is 512*7*7.

[0089] I. After passing through Dense Block 4, which contains 16 Dense Layers including SE modules, the output size is 1024*7*7.

[0090] J. Perform global average pooling (7x7 GAP) to output 1024x1x1.

[0091] K. The results are output to 5 layers through a fully connected layer.

[0092] (6) Step S6: Based on the identification results of the ore type initially confirmed by the DenseNet-SE ore type identification model, and combined with the real-time rotation speed of the spindle and the force signal of the jaw plate, the ore type is comprehensively judged to make a second confirmation of the ore type in the training dataset.

[0093] (7) Step S7: Directly output the second-confirmed ore type of the training dataset and the confidence level corresponding to the second-confirmed ore type, and directly execute step S8;

[0094] (8) Step S8: Issue a command to stop the operation of the mining eccentric roller crusher.

[0095] (9) Step S9: Collect the signal of the eccentric roller crusher when the ore to be identified is crushed, and execute steps S2 and S3 in sequence.

[0096] (10) Step S10: Input the RGB image corresponding to the vibration acceleration signal when the ore to be identified is crushed into the trained DenseNet-SE ore type identification model. The DenseNet-SE ore type identification model directly outputs the preliminary ore type of the ore to be identified.

[0097] (11) Step S11: Based on the identification results of the ore type to be identified initially confirmed by the DenseNet-SE ore type identification model, and combined with the real-time rotation speed of the spindle and the force signal of the lower jaw plate, the ore type to be identified is comprehensively judged in order to make a second confirmation of the ore type to be identified.

[0098] (12) Step S12: Directly output the type of ore confirmed twice in step S11, and the credibility of the type of ore confirmed twice.

[0099] (13) Step S13: Based on the type of ore confirmed twice in step S11 and the corresponding credibility, determine whether the ore to be identified can be crushed; if the determination result is yes, execute step S14; if the determination result is no, execute step S8, and add the collected signal of the eccentric roller crusher when the ore to be identified is crushed to the reference dataset.

[0100] Specifically, the crushability is judged. If the crushability limit is exceeded, the crusher stops working and the signal of the eccentric roller crusher during the crushing of the ore to be identified is returned to the reference dataset of uncrushable ore established in step S1 and marked as uncrushable ore. Then, the next step of model training is carried out. By continuously looping, it can accurately determine whether it is crushable. If it is crushable, the crushing work is completed.

[0101] (14) Step S14: Issue the instruction for the mining eccentric roller crusher to work normally.

[0102] In summary, in this specific embodiment, the ore identification method for a mining eccentric roller crusher provided by the present invention can identify difficult-to-crush ores in the mining eccentric roller crusher. Specifically, by installing multiple sensors such as vibration acceleration sensors, pressure sensors, and rotary encoders on the mining eccentric roller crusher, relevant data is collected, the data is processed, and the processed data is used for model training to construct an eccentric roller crusher ore type identification model. The vibration data to be identified is preprocessed and transformed, and then input into the trained eccentric roller crusher ore type identification model. The model is compared with an existing reference dataset for similarity and with the simultaneously extracted pressure and spindle speed data to identify difficult-to-crush ores.

[0103] In this specific embodiment, time-series signal-image conversion technology is combined with deep learning for the first time in the field of material characteristic identification. This is expected to break through the perception bottleneck of traditional monitoring technology, provide core criteria for intelligent crushing systems, and promote the upgrading of mining equipment towards self-adaptation and low energy consumption.

[0104] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying ore using a mining eccentric roller crusher, the mining eccentric roller crusher comprising: The eccentric roller crusher comprises a roller body, a main shaft, a main shaft support frame, a frame, and a lower jaw plate. The main shaft is connected to the roller body via bearings, and a motor drives the main shaft, which in turn drives the roller body to operate. The lower jaw plate is directly fixed to the frame with bolts, and the ore is crushed in the area near the lower jaw plate. The main shaft support frame is connected to the frame with bolts, and the main shaft support frame supports the main shaft and bearings. The ore identification method includes: Step S1: Collect signals from the eccentric roller crusher when various ores cannot be crushed, and use the collected signals as a reference dataset for those that cannot be crushed; wherein, the collected signals include: the vibration acceleration signal of the main shaft support frame, the real-time rotational speed of the main shaft, and the force signal of the lower jaw plate; the reference dataset is labeled with the type of ore; Step S2: Preprocess the acquired signals; Step S3: Convert the preprocessed time series data into RGB images based on the MTF algorithm; Step S4: Use the RGB image corresponding to the vibration acceleration signal as the training dataset; Step S5: Construct a DenseNet-SE ore type identification model and train the DenseNet-SE ore type identification model based on the training dataset; wherein, the DenseNet-SE ore type identification model uses DenseNet as the backbone network and combines the SE attention mechanism to extract features from the training dataset; the output of the DenseNet-SE ore type identification model is the preliminarily confirmed ore type. Step S6: Based on the preliminary identification results of the ore types confirmed by the DenseNet-SE ore type identification model, and combined with the real-time rotation speed of the spindle and the force signal of the jaw plate, the ore types are comprehensively judged to make a secondary confirmation of the ore types in the training dataset. Step S7: Directly output the types of ore confirmed by the second confirmation in the training dataset, as well as the confidence level corresponding to the types of ore confirmed by the second confirmation, and directly execute step S8; Step S8: Issue a command to stop the operation of the mining eccentric roller crusher.

2. The ore identification method for a mining eccentric roller crusher according to claim 1, characterized in that, Also includes: Step S9: Collect the signal from the eccentric roller crusher during the crushing of the ore to be identified, and execute steps S2 and S3 in sequence; Step S10: Input the RGB image corresponding to the vibration acceleration signal of the ore to be identified during crushing into the trained DenseNet-SE ore type recognition model. The DenseNet-SE ore type recognition model directly outputs the preliminary confirmed ore type of the ore to be identified. Step S11: Based on the preliminary identification results of the ore type to be identified by the DenseNet-SE ore type identification model, and combined with the real-time rotation speed of the spindle and the force signal of the lower jaw plate, the ore type to be identified is comprehensively judged to perform a second confirmation of the ore type to be identified. Step S12: Directly output the type of ore confirmed in step S11 and the corresponding confidence level of the type of ore confirmed in step S11. Step S13: Based on the type of ore confirmed twice in step S11 and the corresponding credibility, determine whether the ore to be identified can be broken. If the judgment result is yes, proceed to step S14; if the judgment result is no, proceed to step S8, and add the collected signal of the eccentric roller crusher during the crushing of the ore to be identified to the reference dataset. Step S14: Issue a command to enable the mining eccentric roller crusher to operate normally.

3. The ore identification method for a mining eccentric roller crusher according to claim 1, characterized in that, In step S1, a vibration sensor is used to collect the vibration acceleration signal of the spindle support frame, a rotary encoder is used to collect the real-time rotational speed of the spindle, and a pressure sensor is used to collect the force on the lower jaw plate.

4. The ore identification method for a mining eccentric roller crusher according to claim 1, characterized in that, Step S2 specifically includes: Step S2.1: Use wavelet adaptive threshold denoising algorithm to eliminate useless signals in the collected signals.

5. The ore identification method for a mining eccentric roller crusher according to any one of claims 1 to 4, characterized in that, In step S4, the DenseNet-SE ore type identification model is an improved DenseNet121 convolutional network. The improved DenseNet121 convolutional network specifically includes: Dense Block1, TransitionLayer1, Dense Block2, Transition Layer2, ..., Dense BlockN connected in sequence; N Dense Blocks and (N-1) Transition Layers are alternately connected; Dense Blocks are composed of a certain number of stacked Dense Layers; each DenseBlockN is followed by an SE module; where N is a positive integer greater than 1.

Images