Self-adaptive cutting method and cutting pick for mining heading machine

Abstract

The invention discloses a self-adaptive cutting method and cutting teeth for a mining heading machine, and relates to the technical field of coal mine underground coal or rock roadway heading operation. The adaptive cutting method comprises the following steps: constructing a hardness identification source domain model, taking cutting current, voltage, arm vibration and torque signals of a laboratory as source domain signals, extracting features through a multi-layer convolutional neural network, fusing the features with a minimum absolute operator, and training; training in a laboratory by using a back propagation algorithm until the precision reaches the standard; migrating to a tunneling site, performing fine adjustment and updating on site signal input as a target domain model, and outputting hardness of a cutting object; the working modes of the cutting tooth are controlled according to hardness, the cutting tooth comprises a cutting tooth seat, an impact spring, a piston and the like, and an oil way is controlled to switch the working modes through a hydraulic control one-way valve and an electromagnetic reversing valve. The cutting mode can be automatically adjusted according to the coal rock hardness, and compared with a traditional fixed parameter cutting mode, the cutting efficiency can be greatly improved, cutting tooth abrasion is reduced, and energy consumption is reduced.

Description

Technical Field

[0001] The invention belongs to the technical field of coal or rock tunnel excavation operations in underground coal mines, and in particular relates to an adaptive cutting method and a cutting tooth of a mining excavator. Background Art

[0002] In coal mining, the cutting efficiency and durability of the picks of the roadheader have always been the focus of industry attention. Especially when facing hard rock cutting, traditional roadheaders often find it difficult to adapt, resulting in severe wear of the picks, low cutting efficiency, and even equipment failure. Some technologies have attempted to evaluate the hardness of the cutting object by monitoring signals such as current, voltage, and vibration during the tunneling process, but most of these methods are based on simple threshold judgments, with limited recognition accuracy and difficulty adapting to complex and changing mining environments. At the same time, these methods often lack adaptive capabilities and cannot be flexibly adjusted according to the actual hardness of coal or rock, thus affecting the cutting effect and equipment performance. In addition, the existing roadheader cutting teeth have a single working mode and cannot be flexibly adjusted according to the hardness, which limits the improvement of tunneling efficiency. Summary of the invention

[0003] The object of the present invention is to provide an adaptive cutting method and cutting teeth for a mining roadheader, so as to solve the problems mentioned in the background technology.

[0004] In order to solve the above technical problems, the present invention is achieved through the following technical solutions:

[0005] The present invention is an adaptive cutting method for a mining roadheader, comprising:

[0006] S1, construct a hardness recognition source domain model, use the cutting current, cutting voltage, cutting arm vibration, and cutting torque signals tested in the laboratory as the source domain signals for cutting hardness discrimination, use a multi-layer convolutional neural network to extract features from the source domain signals, and use the least absolute operator to achieve feature fusion of multiple source signals, and use the fused source domain signals to train the hardness recognition source domain model;

[0007] S2, in a laboratory environment, through supervised learning and using the back propagation algorithm, the weight parameters of the hardness recognition source domain model are updated until the training accuracy reaches the set target value;

[0008] S3, transfer the hardness recognition source domain model trained in the laboratory to the excavation site, collect the cutting current, cutting voltage, cutting arm vibration and cutting torque signals at the excavation site as target domain signals, input them into the hardness recognition source domain model, and further update the hardness recognition model through parameter fine-tuning. After transfer learning, the hardness recognition source domain model is updated to the hardness recognition target domain model, and the target domain signals collected on site are input into the hardness recognition target domain model, and the hardness of the cutting object is output and obtained;

[0009] S4, based on the determined cutting hardness, controls the working mode of the cutting teeth at the excavation site to achieve adaptive cutting, including ordinary cutting, ordinary vibration cutting, strong vibration cutting and extremely strong vibration cutting.

[0010] As a preferred technical solution of the present invention, in step S1, the multi-layer convolutional neural network includes a normalization layer, a convolution layer, a pooling layer and a fully connected layer respectively;

[0011] The step S1 specifically includes the following sub-steps:

[0012] Step S1.1, using batch normalization algorithm to normalize the signals of each layer;

[0013] Step S1.2, the convolution layer realizes feature extraction through convolution calculation:

[0014] Step S1.3, after the convolution operation, a pooling operation is used to downsample the convolution result;

[0015] Step S1.4: After the features of each signal are obtained, the effective fusion of the feature matrices of different channels is achieved through the least absolute operator;

[0016] Step S1.5, calculate the weight parameter of each feature, and select the number of features to be fused from large to small according to the weight parameter.

[0017] As a preferred technical solution of the present invention, the update optimization target in the S1 model training includes two parts: a cross entropy loss function and a minimum absolute operator regression estimation, and the total loss function is the sum of these two parts.

[0018] As a preferred technical solution of the present invention, in the step S1.4, the minimum absolute operator regression estimate calculated by the classification result is back-propagated to the calculation link of the minimum absolute operator to realize the weight and bias update of the feature fusion and activation function; the loss value calculated by the classification loss is directly back-propagated to the model input to realize the weight update of the structural layers such as convolution, pooling, and normalization.

[0019] As a preferred technical solution of the present invention, after the features of each signal in step S1.4 are extracted and fused, the Tanh function is used as the activation function.

[0020] As a preferred technical solution of the present invention, in step S3, when the source domain model trained in the laboratory is migrated to the excavation site, the low-level weight parameters of the convolutional neural network are kept unchanged, and only the high-level weight parameters of the convolutional layer are fine-tuned.

[0021] As a preferred technical solution of the present invention, in step S3, workers at the cutting site regularly calibrate the hardness of the cut coal and rock, and use the calibration results as label outputs of the model to implement supervised learning, so that the model is continuously iterated and updated to improve recognition accuracy.

[0022] As a preferred technical solution of the present invention, in step S4, after obtaining the hardness of the coal rock, the corresponding cutting mode is activated for the pick according to the hardness grade, specifically:

[0023] When the coal rock hardness grade is soft rock, normal cutting is enabled, the cutting teeth have no vibration, and the cutting arm cuts at the rated cutting speed;

[0024] When the coal rock hardness grade is medium hard rock, the cutting teeth start ordinary vibration cutting, and the cutting speed of the cutting arm is appropriately reduced;

[0025] When the coal rock hardness grade is hard rock, the cutting teeth start to vibrate and cut strongly, while further reducing the cutting speed of the cutting arm;

[0026] When the coal rock hardness grade is extremely hard rock, the cutting teeth start extremely strong vibration cutting and further reduce the cutting speed of the cutting arm.

[0027] The present invention also provides an adaptive cutting pick for a mining tunneling machine, which is used to implement the above-mentioned adaptive cutting method, including a pick seat, an impact spring, a piston, a piston rod, a pick, oil port one, a hydraulically controlled one-way valve, oil port two, oil circuit one, oil circuit two, a hydraulically controlled oil circuit and an electromagnetic reversing valve. The pick seat is used to install the pick on the cutting head; the impact spring is used to generate an impact force on the cutting object during the cutting process; the hydraulically controlled one-way valve is used to control the on-off of the oil circuit, and the electromagnetic reversing valve is used to switch the direction of the oil circuit. By controlling the oil pressure switching between oil port one and oil port two, switching of different cutting working modes is realized.

[0028] The present invention has the following beneficial effects:

[0029] Improve cutting efficiency and safety: The roadheader can achieve adaptive cutting according to the hardness of the cutting object. By automatically identifying the hardness of coal and rock and adaptively adjusting the cutting parameters, it avoids the errors and uncertainties of manual judgment and improves cutting efficiency and safety.

[0030] Reduce pick wear and maintenance costs: Adjust the cutting mode and cutting speed according to the hardness of coal and rock, which effectively reduces the wear of the pick, extends the service life of the tunnel boring machine components, and reduces maintenance costs.

[0031] Enhance the model's generalization ability and field adaptability: By combining laboratory training with field fine-tuning, the model's generalization ability and field adaptability are enhanced, and the recognition accuracy and stability are improved.

[0032] Realize intelligent control: Combining advanced machine learning and artificial intelligence technologies, intelligent control of the cutting process of the roadheader is realized, providing strong support for the intelligent transformation of mining.

[0033] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for describing the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 : Adaptive cutting method for mining roadheaders and simplified process of overall cutting teeth steps;

[0036] Figure 2 :Flowchart for determining cutting hardness characteristics by cutting parameter extraction;

[0037] Figure 3 : Cutting hardness identification model migration flow chart;

[0038] Figure 4 : Hydraulic circuit when the pick works in normal cutting mode;

[0039] Figure 5 : The hydraulic circuit when the pick is working in preparation for vibration cutting.

[0040] In the accompanying drawings, the components represented by the reference numerals are listed as follows:

[0041] 1-pick seat, 2-impact spring, 3-piston, 4-piston rod, 5-pick, 6-oil port 1, 7-hydraulic one-way valve, 8-oil port 2, 9-oil circuit 1, 10-hydraulic oil circuit, 11-oil circuit 2, 12-electromagnetic reversing valve, N1-first normalized layer, C1-first convolutional layer, N2-second normalized layer, P1-first pooling layer, C2-second convolutional layer, N3-third normalized layer, P2-second pooling layer, FC-fully connected layer. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Embodiment 1

[0044] like Figure 1 As shown: The present invention provides an adaptive cutting method for a mining roadheader, comprising:

[0045] S1, construct a hardness recognition source domain model, use the cutting current, cutting voltage, cutting arm vibration, and cutting torque signals tested in the laboratory as the source domain signals for cutting hardness discrimination, use a multi-layer convolutional neural network to extract features from the source domain signals, and use the least absolute operator to achieve feature fusion of multiple source signals, and use the fused source domain signals to train the hardness recognition source domain model;

[0046] S2, in a laboratory environment, through supervised learning and using the back propagation algorithm, the weight parameters of the hardness recognition source domain model are updated until the training accuracy reaches the set target value;

[0047] S3, transfer the hardness recognition source domain model trained in the laboratory to the excavation site, collect the cutting current, cutting voltage, cutting arm vibration and cutting torque signals at the excavation site as target domain signals, input them into the hardness recognition source domain model, and further update the hardness recognition model through parameter fine-tuning. After transfer learning, the hardness recognition source domain model is updated to the hardness recognition target domain model, and the target domain signals collected on site are input into the hardness recognition target domain model, and the hardness of the cutting object is output and obtained;

[0048] S4, based on the determined cutting hardness, controls the working mode of the cutting teeth at the excavation site to achieve adaptive cutting, including ordinary cutting, ordinary vibration cutting, strong vibration cutting and extremely strong vibration cutting.

[0049] This embodiment introduces an adaptive cutting method for a mining roadheader. The specific process is as follows: Step S1: Construct a hardness recognition source domain model:

[0050] Data source selection: Cutting current, cutting voltage, cutting arm vibration, and cutting torque signals tested in a laboratory environment are selected as source domain signals for cutting hardness discrimination. These signals can reflect the physical properties of rocks or coal seams during the cutting process, especially hardness.

[0051] Feature extraction: Use a multi-layer convolutional neural network (CNN) to extract features from these source domain signals. CNN has a wide range of applications in image processing, signal processing, and other fields due to its powerful feature learning capabilities. Here, CNN can automatically learn and extract features related to cutting hardness.

[0052] Feature fusion: The least absolute operator is used to achieve feature fusion of multi-source signals. This means combining the signal features from different sensors (such as cutting current, cutting voltage, etc.) to form a more comprehensive and accurate hardness identification feature.

[0053] Model training: The fused source domain signal is used to train the hardness recognition source domain model. Through repeated iterations and adjustment of model parameters, the model can accurately identify cutting objects of different hardness.

[0054] Step S2: Model optimization in laboratory environment:

[0055] Supervised learning: In a laboratory environment, the model is trained using supervised learning methods using cut objects of known hardness and their corresponding source domain signals.

[0056] Back propagation algorithm: The back propagation algorithm is used to update the weight parameters of the hardness recognition source domain model. The back propagation algorithm is a commonly used neural network training algorithm. It calculates the gradient of the error and back propagates it to each layer of the network to update the weight parameters and make the model's prediction more accurate.

[0057] Training accuracy: Through continuous iterative training, until the training accuracy of the model reaches the set target value, this means that the model has a good hardness recognition ability in the laboratory environment.

[0058] Step S3: Model migration and field application:

[0059] Model transfer: The hardness recognition source domain model trained in the laboratory is transferred to the excavation site. Since the environment and conditions at the excavation site are different from those in the laboratory, the model needs to be fine-tuned to adapt to the on-site environment.

[0060] Parameter fine-tuning: The cutting current, cutting voltage, cutting arm vibration and cutting torque signals at the excavation site are collected as target domain signals, input into the hardness identification source domain model, and further update the hardness identification model through parameter fine-tuning. This process enables the model to better adapt to the actual conditions of the excavation site.

[0061] Hardness identification: The target domain signal collected on site is input into the updated hardness identification target domain model, and the hardness of the cut object is output and obtained.

[0062] Step S4: Adaptive cutting control:

[0063] Working mode selection: Based on the determined cutting hardness, the working mode of the cutting teeth at the excavation site is controlled. These working modes include normal cutting, normal vibration cutting, strong vibration cutting and extremely strong vibration cutting.

[0064] Adaptive cutting: By selecting different working modes, the tunnel boring machine can make adaptive adjustments according to the hardness of the cutting object, thereby improving tunneling efficiency and safety.

[0065] In summary, this embodiment implements an adaptive cutting method for a mining roadheader by building a hardness identification source domain model, optimizing the model in a laboratory environment, implementing model migration and field application, and adaptive cutting control. This method can perform adaptive adjustments according to the hardness of the cutting object, improve the efficiency and safety of the roadheader, and has important application value.

[0066] Embodiment 2

[0067] like Figure 2 As shown, in an optional embodiment, in step S1, the multi-layer convolutional neural network includes a normalization layer, a convolution layer, a pooling layer and a fully connected layer respectively;

[0068] Step S1 specifically includes the following sub-steps:

[0069] Step S1.1, using batch normalization algorithm to normalize the signals of each layer;

[0070] In order to reduce the difference between the source domain and the target domain, the batch normalization algorithm is used to normalize the signals of each layer.

[0071]

[0072] In the formula,

[0073] Input x to the target model t The normalized result of

[0074] y t This is the final target model output after adaptive batch normalization;

[0075] γ s is the scaling of the source model;

[0076] β s is the translation parameter;

[0077]

[0078] In the formula,

[0079] σ t , σ s are the variances of the target model and the source model, respectively;

[0080] κ=μ t -λμ s ;

[0081] In the formula,

[0082] μ t , μ s are the means of the target model and the source model, respectively.

[0083] Step S1.2, the convolution layer realizes feature extraction through convolution calculation:

[0084] The convolution layer realizes feature extraction through convolution calculation. The specific calculation is as follows:

[0085]

[0086] Where: l is the lth layer of convolution; i represents the i-th filter kernel; is the weight coefficient; is the bias; x l (j) is the convolution input; is the result of convolution calculation, that is, the input of the next layer; * indicates convolution calculation.

[0087] Step S1.3, after the convolution operation, a pooling operation is used to downsample the convolution result;

[0088] The calculation is as follows:

[0089]

[0090] Where: t is the tth neuron, t∈[(j-1)W+1,jW]; W is the size of the pooled area; is the output result of the pooling operation; is the value of the neuron.

[0091] Step S1.4: After the features of each signal are obtained, the effective fusion of the feature matrices of different channels is achieved through the least absolute operator;

[0092] The calculation is as follows:

[0093]

[0094] Where: L is the penalty coefficient, which controls the degree of feature fusion shrinkage; c is feature x c The regression coefficient of γ c as a set of feature importance scores (γ1,...,γ c ,...γ m ), let the weight parameter be w c , calculated as follows:

[0095]

[0096] Step S1.5: Calculate the weight parameter w = {w1, w2, ... w c ,...,w m}, when selecting the number of feature fusion, the selection can be made according to the weight parameter from large to small. The larger the weight parameter, the more the feature can reflect the hardness of cutting. According to the claim, after the feature quantity is selected in combination with the computing resources, the weight parameter can be determined in order from large to small.

[0097] At the same time, the hardness of the cutting objects in laboratory model training is divided into soft rock (<30MPa), medium hard rock (30-80MPa), hard rock (80-150MPa), and extremely hard rock (>150MPa). The cutting data of several types of hardness coal and rock are collected as training data, and the hardness grades are numbered as the output of network training to form supervised learning.

[0098] In another optional embodiment, the updated optimization objective in S1 model training includes two parts: a cross entropy loss function and a minimum absolute operator regression estimation, and the total loss function is the sum of these two parts.

[0099] Specific:

[0100] The update optimization objective in model training includes two parts: the cross entropy loss function and the least absolute operator regression estimation. The cross entropy loss can be expressed as follows:

[0101]

[0102] In the formula,

[0103] L s is the classification loss;

[0104] X s , Y s are the fused feature set and label of the source domain respectively.

[0105]

[0106] Where:

[0107] E is the expected value of classification probability;

[0108] P(y|x s ) is the classification probability of the source domain samples.

[0109] The least absolute operator regression estimate is The corresponding loss function is L γ Therefore, the total loss function is as follows:

[0110] L=L s +L γ

[0111] In another optional embodiment, in step S1.4, the least absolute operator regression estimate calculated from the classification result is back-propagated to the calculation link of the least absolute operator to achieve weight and bias updates of feature fusion and activation functions; the loss value calculated from the classification loss is directly back-propagated to the model input to achieve weight updates of structural layers such as convolution, pooling, and normalization.

[0112] After feature extraction and feature fusion of each cut signal in step S1.4, considering that the feature set range of multi-sensor signal extraction is relatively wide, in order to avoid the phenomenon that the calculated value is less than 0 and causes the neuron feature to fail, the Tanh function is used as the activation function, and the calculation is as follows:

[0113]

[0114] Where:

[0115] is the result of the convolution operation;

[0116] is the corresponding activation value.

[0117] In another optional embodiment, in step S3, when the source domain model trained in the laboratory is transferred to the excavation site, the low-level weight parameters of the convolutional neural network are kept unchanged, and only the high-level weight parameters of the convolutional layer are fine-tuned. Specifically: because the hardness of the coal and rock at the working site of the tunnel boring machine cannot be predicted, that is, the training in the target model belongs to unsupervised learning, in order to transfer the source domain model trained in the laboratory to the excavation site, the excavation signals (cutting current, cutting voltage, cutting arm vibration, cutting torque signal) similar to the experimental model are collected at the cutting site as target domain signals and input into the laboratory trained model, and at the same time, the low-level weight parameters of the convolutional neural network are adjusted, and the high-level weight parameters of the convolutional layer are fine-tuned, that is, in the re-training, the weight parameters of C1 and P1 remain unchanged, and only the weight parameters of C2 and P2 are fine-tuned, and the final model output is used as the coal and rock hardness to guide the cutting method.

[0118] In step S3, workers at the cutting site regularly calibrate the hardness of the cut coal and rock, that is, calibrate the hardness of the cut coal and rock into several levels, such as soft rock, medium hard rock, hard rock, and extremely hard rock, and use this classification as the label output of the model, thereby realizing supervised learning, enabling the model to be continuously iterated and updated, improving recognition accuracy, and enabling more accurate selection of excavation modes.

[0119] In step S4, after obtaining the coal rock hardness, the corresponding cutting mode is activated for the pick according to the hardness level, specifically:

[0120] When the coal rock hardness grade is soft rock, normal cutting is enabled, the cutting teeth have no vibration, and the cutting arm cuts at the rated cutting speed;

[0121] When the coal rock hardness grade is medium hard rock, the cutting teeth start ordinary vibration cutting, and the cutting speed of the cutting arm is appropriately reduced;

[0122] When the coal rock hardness grade is hard rock, the cutting teeth start to vibrate and cut strongly, while further reducing the cutting speed of the cutting arm;

[0123] When the coal rock hardness grade is extremely hard rock, the cutting teeth start extremely strong vibration cutting and further reduce the cutting speed of the cutting arm.

[0124] like Figure 4-Figure 5 As shown: The present invention also provides an adaptive cutting pick for a mining roadheader, which is used to implement the above-mentioned adaptive cutting method, including a pick seat 1, an impact spring 2, a piston 3, a piston rod 4, a pick 5, an oil port 1 6, a hydraulically controlled one-way valve 6, an oil port 2 8, an oil circuit 1 9, an oil circuit 2 11, a hydraulically controlled oil circuit 10, and an electromagnetic reversing valve 12;

[0125] The pick seat 1 is used to install the pick on the cutting head, and is provided with a retaining spring groove, which is directly fixed by the retaining spring during installation, so that the pick seat 1 is easy to disassemble;

[0126] The impact spring 2 is used to generate an impact force on the cutting object during the cutting process, and the switching of different cutting working modes is realized by the continuous switching of the oil pressure between the oil port 1 6 and the oil port 2 8;

[0127] The hydraulically controlled one-way valve 6 can be controlled by a cutting control system, and the control signal is the hydraulically controlled oil circuit 10. When the pressure of the hydraulically controlled oil circuit 10 reaches the opening pressure, the one-way valve is opened and called a passage, otherwise the one-way valve is always in a one-way cut-off state;

[0128] When the hardness of the coal rock is identified as soft rock, the pick is in the normal cutting mode, the hydraulic control one-way valve is in the cut-off state, and the oil pressure of the oil port 2 8 is only the residual pressure. At this time, the pick applies cutting force to the coal rock by the spring force and the back pressure at the oil port 1 6. If the coal rock suddenly hardens, the spring is compressed back to protect the pick from being broken by the hard rock.

[0129] When the hardness of the coal rock is identified as medium hard rock, hard rock and extremely hard rock, the electromagnetic reversing valve 12 is in the right working position, and the oil port 28 is connected to high-pressure oil. At this time, the pick is compressed back under the action of hard rock and high-pressure oil, and the cutting force gradually increases. When the piston crosses the oil port 16, the working position of the electromagnetic reversing valve 12 is changed from the right position to the left position. At this time, the hydraulic control oil circuit is connected to high-pressure oil, the hydraulic control check valve is opened, and the pressure oil enters from the oil port 16 and returns from the oil port 28. The pick quickly impacts forward under the action of the impact spring elastic force and the left chamber pressure, so that the cutting teeth have an impact force on the coal rock, and the working position of the electromagnetic reversing valve 12 is switched back and forth, thereby realizing the vibration impact cutting of the tunnel boring machine.

[0130] When the coal rock hardness is identified as soft rock, the electromagnetic reversing valve 12 is in the middle position, and the piston is just in the middle of the oil port 1 6 and the oil port 2 8. At this time, the pick generates a cutting force on the cutting object under the action of the impact spring elastic force. At the same time, there is still a certain back pressure on the left side of the piston to ensure that the pick has sufficient cutting force. When the coal rock hardness suddenly increases, the impact spring contracts to the left to protect the pick.

[0131] The pick can select different impact cutting modes according to the hardness identified by the model. On the premise of meeting the safety regulations of coal mines, the oil pressure is divided into three levels from the maximum working oil pressure. The corresponding pick impact levels are normal vibration, vibration, and extremely strong vibration, which correspond to the cutting of medium hard rock, hard rock, and extremely hard rock respectively.

[0132] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0133] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. An adaptive cutting method for a mining roadheader, characterized in that: include: S1, construct a hardness recognition source domain model, use the cutting current, cutting voltage, cutting arm vibration, and cutting torque signals tested in the laboratory as the source domain signals for cutting hardness discrimination, use a multi-layer convolutional neural network to extract features from the source domain signals, and use the least absolute operator to achieve feature fusion of multiple source signals, and use the fused source domain signals to train the hardness recognition source domain model; S2, in a laboratory environment, through supervised learning and using the back propagation algorithm, the weight parameters of the hardness recognition source domain model are updated until the training accuracy reaches the set target value; S3, transfer the hardness recognition source domain model trained in the laboratory to the excavation site, collect the cutting current, cutting voltage, cutting arm vibration and cutting torque signals at the excavation site as target domain signals, input them into the hardness recognition source domain model, and further update the hardness recognition model through parameter fine-tuning. After transfer learning, the hardness recognition source domain model is updated to the hardness recognition target domain model, and the target domain signals collected on site are input into the hardness recognition target domain model, and the hardness of the cutting object is output and obtained; S4, based on the determined cutting hardness, controls the working mode of the cutting teeth at the excavation site to achieve adaptive cutting, including ordinary cutting, ordinary vibration cutting, strong vibration cutting and extremely strong vibration cutting.

2. The method for adaptive cutting of a mining roadheader according to claim 1, characterized in that: In step S1, the multi-layer convolutional neural network includes a normalization layer, a convolution layer, a pooling layer and a fully connected layer respectively; The step S1 specifically includes the following sub-steps: Step S1.1, using batch normalization algorithm to normalize the signals of each layer; Step S1.2, the convolution layer realizes feature extraction through convolution calculation: step S1.3, after the convolution operation, the convolution result is downsampled by using the pooling operation; Step S1.4: After the features of each signal are obtained, the effective fusion of the feature matrices of different channels is achieved through the least absolute operator; Step S1.5, calculate the weight parameter of each feature, and select the number of features to be fused from large to small according to the weight parameter.

3. The method for adaptive cutting of a mining roadheader according to claim 2, characterized in that: The update optimization target in the S1 model training includes two parts: the cross entropy loss function and the minimum absolute operator regression estimation, and the total loss function is the sum of these two parts.

4. The method for adaptive cutting of a mining roadheader according to claim 3, characterized in that: In the step S1.4, the minimum absolute operator regression estimate calculated by the classification result is back-propagated to the calculation link of the minimum absolute operator to realize the weight and bias update of feature fusion and activation function; the loss value calculated by the classification loss is directly back-propagated to the model input to realize the weight update of structural layers such as convolution, pooling, and normalization.

5. The method for adaptive cutting of a mining roadheader according to claim 4, characterized in that: In step S1.4, after feature extraction and feature fusion of each signal, the Tanh function is used as the activation function.

6. The method for adaptive cutting of a mining roadheader according to claim 1, characterized in that: In step S3, when the source domain model trained in the laboratory is migrated to the excavation site, the low-level weight parameters of the convolutional neural network are kept unchanged, and only the high-level weight parameters of the convolutional layer are fine-tuned.

7. The method for adaptive cutting of a mining roadheader according to claim 6, characterized in that: In step S3, workers at the cutting site regularly calibrate the hardness of the cut coal and rock, and use the calibration results as label outputs of the model to implement supervised learning, so that the model is continuously iterated and updated to improve recognition accuracy.

8. The method for adaptive cutting of a mining roadheader according to claim 1, characterized in that: In step S4, after obtaining the coal rock hardness, the corresponding cutting mode is activated for the pick according to the hardness level, specifically: When the coal rock hardness grade is soft rock, normal cutting is enabled, the cutting teeth have no vibration, and the cutting arm cuts at the rated cutting speed; When the coal rock hardness grade is medium hard rock, the cutting teeth start ordinary vibration cutting, and the cutting speed of the cutting arm is appropriately reduced; When the coal rock hardness grade is hard rock, the cutting teeth start to vibrate and cut strongly, while further reducing the cutting speed of the cutting arm; When the coal rock hardness grade is extremely hard rock, the cutting teeth start extremely strong vibration cutting and further reduce the cutting speed of the cutting arm.

9. An adaptive cutting pick for a mining roadheader, used to implement the adaptive cutting method according to any one of claims 1 to 8, characterized in that: It includes a pick seat, an impact spring, a piston, a piston rod, a pick, an oil port one, a hydraulically controlled one-way valve, an oil port two, an oil circuit one, an oil circuit two, a hydraulically controlled oil circuit and an electromagnetic reversing valve. The pick seat is used for installing the pick on the cutting head; the impact spring is used for generating an impact force on the cutting object during the cutting process; the hydraulically controlled one-way valve is used for controlling the on-off of the oil circuit; the electromagnetic reversing valve is used for switching the direction of the oil circuit. By controlling the oil pressure switching between the oil port one and the oil port two, the switching of different cutting working modes is realized.

Images