Efficient rock breaking and slag discharging cooperative control system of mining crossing bottom angle drilling machine

Abstract

The invention relates to the technical field of drilling machine cooperative control, and discloses an efficient rock breaking and slag discharging cooperative control system of a mining crossing bottom angle drilling machine. The system is integrated with a data acquisition assembly, an intelligent analysis assembly and a regulation and control assembly. The data acquisition assembly collects vibration signals and deslagging parameters in the rock breaking process of the drilling machine in real time, and a rock breaking oscillogram and a state description text are generated based on the signals and the parameters. A multi-modal recognition model and a mechanism simulation model are arranged in the intelligent analysis assembly, the multi-modal recognition model jointly processes the rock breaking oscillogram and the state description text to output rock breaking working condition classification, and the mechanism simulation model analyzes characteristic parameters from the rock breaking oscillogram, matches the characteristic parameters with a standard mechanism library and outputs a working condition matching result. And the regulation and control assembly performs comprehensive evaluation according to output classification and matching results of the intelligent analysis assembly, generates a drilling machine control command and optimizes the rock breaking and slag discharging process. According to the system, intelligent cooperative control over rock breaking and slag discharging is achieved.

Description

Technical Field

[0001] This invention relates to the field of drilling rig collaborative control technology, specifically to a high-efficiency rock breaking and slag removal collaborative control system for mining cross-layer bottom corner drilling rigs. Background Technology

[0002] Mining drilling rigs are key engineering equipment in mining, tunneling, and geological exploration. Their core function is to break up rock strata through the rotation and advancement of the drill bit, simultaneously removing the resulting rock cuttings to form a borehole that meets design requirements. In practical engineering applications, rock breaking and cuttings removal are two closely coupled and mutually restrictive core processes. Traditional drilling rig control systems typically manage the rock breaking and cuttings removal processes as relatively independent stages. Rock breaking parameters are mainly preset based on operator experience or a rough description in a geological survey report, while the cuttings removal system often adopts a fixed operating mode. This control strategy based on experience-based presets and fixed modes shows significant limitations when dealing with complex and variable underground rock conditions.

[0003] Underground rock strata are heterogeneous and anisotropic. Even along the same drilling path, complex geological structures such as faults, fracture zones, and alternations between soft and hard rocks may be encountered. When the drill bit moves from one type of rock to another, or encounters areas with well-developed joints and fractures, the rock-machine interaction undergoes drastic changes. These changes are directly reflected in the drilling rig's operating parameters: the spectrum and amplitude of vibration signals generated during rock breaking will change, and the particle size distribution, generation rate, and morphological characteristics of rock cuttings will also change accordingly. If the rock breaking parameters fail to adapt to the changes in rock strata in a timely manner, it may lead to excessive wear of the drill bit, stuck drill bit, or even breakage. If the cuttings removal parameters are not adjusted synchronously, it may cause poor cuttings removal, resulting in rock cuttings accumulating in the borehole, exacerbating drill bit wear, increasing rotational torque, and in severe cases, potentially causing drill bit burial accidents, which not only affect construction efficiency but also pose significant safety hazards.

[0004] In existing technologies, some improved control systems attempt to indirectly determine the operating condition and trigger alarms or simple adjustments by monitoring a single or a few key parameters. However, this simple feedback mechanism based on threshold alarms has significant shortcomings. First, a single parameter cannot comprehensively and accurately reflect the complex rock breaking state and rock strata characteristics. For example, an increase in motor current could be due to encountering hard rock strata or increased resistance caused by poor slag removal, making it difficult for the system to distinguish the root cause. Second, the response is lag-dependent, usually intervening only after an anomaly occurs, which is a passive response rather than proactive prevention. Furthermore, the status monitoring of the slag removal process is often neglected, and its parameter adjustments lack an effective coordination mechanism with the rock breaking process. A deeper problem is that existing systems lack a deep understanding of the rock breaking mechanism and the ability to fuse and analyze multi-source information. Vibration signals contain rich information about the rock mass breaking mechanism, and slag removal parameters directly reflect the working conditions inside the borehole, but these multimodal data have not been effectively correlated and deeply mined, failing to form a systematic understanding of the operating conditions, thus restricting the level of intelligence in the control strategy. Therefore, there is an urgent need for a control system that can sense multi-source information on rock breaking and slag removal in real time, intelligently identify changes in working conditions, and achieve coordinated optimization of rock breaking and slag removal, so as to improve the working efficiency, adaptability and reliability of drilling rigs under complex geological conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a high-efficiency rock breaking and slag removal coordinated control system for mining cross-layer bottom angle drilling rigs, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a high-efficiency rock breaking and slag removal coordinated control system for a mining bottom-angle drilling rig, the system comprising: Data acquisition components, intelligent analysis components, and control components, among which: The data acquisition component is configured to collect vibration signals and slag discharge parameters during the rock breaking process of the drilling rig in real time, and generate rock breaking waveform diagrams and status description text based on these signals and parameters; The intelligent analysis component has a built-in multimodal recognition model and a mechanism simulation model. The multimodal recognition model is used to jointly process the rock breaking waveform diagram and the state description text to output the rock breaking condition classification. The mechanism simulation model is used to parse the feature parameters from the rock breaking waveform diagram, match them with the standard mechanism library, and output the condition matching result. The control component performs a comprehensive evaluation based on the output classification and matching results of the intelligent analysis component, generates drilling rig control commands, and optimizes the rock breaking and slag removal process.

[0007] Preferably, while generating the rock-breaking waveform, the data acquisition component performs a feature extraction operation to derive multiple mechanistic feature indicators from the vibration signal. These indicators include amplitude fluctuation, frequency domain concentration, pressure change rate, vibration mode, rotational speed consistency, low-frequency energy, and high-frequency energy. This component also calculates the correlation strength between rock breaking parameters and slag discharge parameters, and derives a slag discharge correction factor through correlation deduction. Subsequently, the mechanistic characteristic indicators and slag discharge correction factors were numerically normalized and text-mapped, and compiled into a state description text with a key-value structure.

[0008] Preferably, the slag discharge parameters include slag discharge flow rate readings and slag discharge consistency readings; The correlation derivation includes: calculating the correlation between rock-breaking vibration signals and slag discharge velocity readings at the same time interval to obtain a velocity correlation factor; calculating the correlation between rock-breaking vibration signals and slag discharge consistency readings to obtain a consistency correlation factor; and then synthesizing the velocity correlation factor and the consistency correlation factor according to a predetermined ratio to generate a slag discharge correction factor.

[0009] Preferably, the multimodal recognition model includes a visual coding unit, a semantic coding unit, a multimodal fusion unit, and a multimodal parsing unit; When the multimodal recognition model processes the rock-breaking waveform and the state description text, it performs feature encoding on the rock-breaking waveform through a visual encoding unit to obtain a waveform feature vector, and performs feature encoding on the state description text through a semantic encoding unit to obtain a text feature vector. Based on cross-modal alignment loss and image-text consistency loss, waveform feature vectors and text feature vectors are input into a multimodal fusion unit for feature integration to obtain fusion coding results; The fusion encoding result is input into the multimodal parsing unit for decoding, and the rock breaking condition classification is output.

[0010] Preferably, when analyzing the characteristic parameters, the mechanism simulation model quantifies multiple characteristic dimensions of the rock-breaking waveform, including amplitude level, event frequency, time interval, low-frequency to high-frequency energy ratio, waveform symmetry, number of peaks, distribution pattern, acoustic detection probability, and phase characteristics. The similarity of multiple feature dimensions with pre-stored multi-type working condition mechanism models is compared to obtain each similarity score, and the working condition corresponding to the highest score is taken as the working condition matching result.

[0011] Preferably, the rock-breaking condition classification output by the multimodal recognition model includes the condition type and its confidence level; When the control component performs a comprehensive evaluation, it assigns multimodal weight values ​​based on confidence level and mechanism weight values ​​based on similarity score. The multimodal weight values ​​and mechanism weight values ​​are standardized to obtain a comprehensive weight value. The final working condition type is then determined based on the comprehensive weight value, rock breaking condition classification, and working condition matching results.

[0012] Preferably, the system further includes a security monitoring component; The safety monitoring component establishes a set of drilling rig fault characteristics, including drill bit wear, increased rock mass resistance, and slag discharge blockage. Obtain the drilling rig equipment identification and combine it with real-time vibration signals to determine whether it meets the fault characteristic set, calculate the probability of occurrence of each fault characteristic, and generate a risk report; Early warning signals are generated based on risk reports.

[0013] Preferably, when the security monitoring component generates an early warning signal, it performs the following operations: Obtain the drilling rig control commands generated by the control component; Based on the aforementioned risk report and drilling rig control commands, determine whether safety intervention is required; If necessary, a safety instruction containing the intervention level and recommended action is generated and sent to the control component; The control component is configured to prioritize responding to the safety instructions and adjust the drilling rig control commands.

[0014] Preferably, when the multimodal fusion unit performs feature integration, it employs a fusion mechanism based on cross-attention, specifically including: Using the waveform feature vector as the query vector and the text feature vector as the key vector and value vector, cross-attention calculation is performed to obtain the first fusion vector; The text feature vector is used as the query vector, and the waveform feature vector is used as the key vector and value vector. Cross-attention calculation is performed to obtain the second fusion vector. The first fusion vector and the second fusion vector are concatenated, and then dimensionality is reduced through a fully connected layer to obtain the fusion encoding result.

[0015] Preferably, the mechanism simulation model is further configured with a self-updating mechanism, the self-updating mechanism including: Record the feature parameters obtained from each analysis and the final working condition matching result; When the rock breaking condition classification output by the multimodal recognition model differs from the condition matching result by a preset number of times, the model optimization process is triggered. Based on the recorded feature parameters and the rock-breaking condition classification, the parameters of the pre-stored condition mechanism model are corrected.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention's system achieves intelligent collaborative control of the rock breaking and slag removal processes by constructing an integrated architecture encompassing data acquisition, intelligent analysis, and control execution. The data acquisition component synchronously collects vibration signals and slag removal parameters, generating intuitive rock breaking waveforms and state description texts, providing a rich and structured data foundation for subsequent analysis. The intelligent analysis component employs a multimodal recognition model to jointly process the waveforms and text descriptions, enabling a comprehensive perception of the rock breaking conditions from different dimensions and improving the accuracy of condition classification. Simultaneously, the mechanism simulation model analyzes waveform feature parameters and matches them with a standard mechanism library, providing a deeper understanding of the rock breaking process from a rock mechanics perspective, making condition judgments more theoretically grounded.

[0017] The analysis method combining multimodal recognition and mechanism simulation considers both the advantages of data-driven pattern recognition and the explanatory power of physical mechanisms, making the intelligent analysis results more comprehensive and reliable. The control component performs a comprehensive evaluation based on the working condition classification and matching results output by the intelligent analysis, generating collaborative control commands. This control strategy, based on a deep understanding of the working conditions, can adaptively optimize the drilling rig's rock-breaking and slag-discharge parameters for different rock strata conditions and rock-breaking states, achieving a balance between rock-breaking efficiency and slag-discharge effect. Through closed-loop control of real-time monitoring, intelligent analysis, and dynamic regulation, the system ensures that the drilling rig always maintains a highly efficient and safe working state. This not only improves the overall efficiency of rock-breaking and slag-discharge and reduces energy consumption, but also reduces equipment wear and failure risks by adapting to changes in working conditions in a timely manner, providing technical support for safe and efficient mining. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the working principle of the high-efficiency rock breaking and slag removal collaborative control system of the mining bottom angle drilling rig described in this invention; Figure 2 A flowchart illustrating the workflow for generating status description text for the data acquisition component; Figure 3 Workflow diagram for processing rock breaking waveforms and state description text for multimodal recognition models; Figure 4 This is a comparison chart of the processing times of each component in the multimodal recognition model. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1This invention provides a high-efficiency rock breaking and slag removal collaborative control system for a mining drilling rig with bottom corners and cross-layers. The system includes an integrated component architecture. At the hardware level, a sensor network is deployed for real-time monitoring of the drilling rig's operating status. At the software level, a data processing and analysis module is constructed. The data acquisition component acquires raw signals through vibration sensors and slag removal monitoring devices. The vibration signals include acceleration and frequency data, and the slag removal parameters involve fluid dynamics readings. The data acquisition component preprocesses the raw signals, including filtering and normalization, to generate a time-series rock breaking waveform; simultaneously, it extracts key indicators and converts them into structured text to form a state description text. The intelligent analysis component loads a multimodal recognition model and a mechanism simulation model. The multimodal recognition model uses a deep learning framework to process multi-source data, and the mechanism simulation model constructs a simulation environment based on physical laws. The multimodal recognition model receives the rock breaking waveform and state description text as input, and outputs a rock breaking condition classification through feature extraction and fusion. The condition classification includes categories such as normal rock breaking, rock mass variation, or equipment anomaly. The mechanism simulation model analyzes the numerical characteristics of the rock-breaking waveform and performs pattern matching with a pre-stored standard mechanism library, which stores waveform characteristics and rock mass properties from historical working conditions. The control component integrates control algorithms, receives the output results from the intelligent analysis component, and uses the rock-breaking condition classification and condition matching results as input variables. The control component performs weighted evaluation to generate drilling rig control commands, which adjust the drill bit speed, feed pressure, or slag discharge valve opening. The system optimizes rock-breaking efficiency through feedback loops. The data acquisition component continuously updates signals, the intelligent analysis component dynamically adapts model parameters, and the control component sends commands to the drilling rig's actuators in real time.

[0021] Example 1: See Figure 2The data acquisition component in the system is responsible for collecting and initially processing raw signals. Its hardware includes a vibration sensor array located at key parts of the drilling rig and a fluid parameter monitoring unit integrated into the slag discharge pipe. The vibration sensors are piezoelectric accelerometers, installed close to the drill pipe joint or drill bit carrier, with a sampling frequency set to 20kHz to capture high-frequency impact signals. The fluid parameter monitoring unit includes an electromagnetic flowmeter and a microwave consistency meter, used to collect slag discharge velocity and consistency readings, respectively. The software module of the data acquisition component runs on an embedded signal processor, implementing signal preprocessing, feature extraction, and data fusion algorithms. The process of generating the rock-breaking waveform begins with the analog-to-digital conversion of the vibration signal. The continuous voltage signal is discretized into a time series by a high-speed acquisition card after anti-aliasing filtering. The rock-breaking waveform is plotted with time on the horizontal axis and amplitude on the vertical axis. The image resolution is set to 1024×768 pixels, and grayscale gradient is used for color mapping to enhance visual contrast. The rock-breaking waveform is updated in real time, with a sliding time window width of 1 second and an overlap rate of 50% to ensure continuity. Feature extraction is performed simultaneously with waveform generation, calculating seven mechanistic characteristic indices from the vibration signal sequence. Amplitude fluctuation is obtained by calculating the standard deviation of amplitude values ​​within a sliding window with a width of 100 sampling points. Frequency domain clustering is extracted by performing a Fast Fourier Transform on the signal to obtain the proportion of energy in the main frequency band, which is set to 50-200Hz based on the rated speed of the drilling rig.

[0022] The pressure change rate is derived from the pressure sensor readings of the drilling rig's hydraulic system, calculating the pressure difference per unit time. Vibration pattern identification employs a waveform decomposition algorithm, decomposing the signal into steady-state and transient components, and classifying it into uniform vibration, impact vibration, or a mixed mode based on the envelope shape. Speed ​​consistency is compared by the absolute value of the error between the actual speed fed back by the spindle encoder and the speed set by the control system. Low-frequency and high-frequency energy are achieved through digital filters; the low-pass filter has a cutoff frequency of 100Hz, and the high-pass filter has a cutoff frequency of 1000Hz. The energy value is the sum of squares of the filtered signal within the time window. The data acquisition component synchronously acquires slag discharge parameters. The slag discharge velocity reading is derived from the pulse frequency output of the electromagnetic flowmeter, converted into a volumetric flow rate value after calibration, and the slag discharge consistency reading is measured using microwave phase difference to measure the solid particle concentration. The correlation derivation module analyzes the statistical correlation between rock-breaking parameters and slag discharge parameters; the rock-breaking parameters are derived from the root mean square value sequence of the vibration signal. The correlation derivation calculates the Pearson correlation coefficient for rock-breaking vibration signals and slag discharge velocity readings at the same time interval. The time interval is 1 second, consistent with the waveform window. The calculation result is quantified as a velocity correlation factor. The same operation is performed on the rock-breaking vibration signal and slag discharge consistency reading to obtain the consistency correlation factor. The velocity correlation factor and consistency correlation factor are synthesized according to a predetermined ratio using fixed weighting coefficients: the velocity correlation factor has a weight of 0.7, and the consistency correlation factor has a weight of 0.3. The weighted sum is linearly scaled to generate a slag discharge correction factor. The slag discharge correction factor is specified to be between 0 and 1; a larger value indicates a higher degree of coordination between the slag discharge state and the rock-breaking action.

[0023] Mechanistic characteristic indicators and slag discharge correction factors are numerically normalized using a minimum-maximum normalization method, mapping each characteristic value to the [0,1] interval. The mapping formula is the characteristic value minus the historical minimum value, divided by the historical range. Historical statistical data comes from a sample library from the drilling rig commissioning phase. Text mapping converts the normalized values ​​into key-value pairs, with keys using English abbreviations, such as "AMP_VAR" for amplitude fluctuation, and values ​​retained to three decimal places. The status description text is compiled in JSON format, with key-value pairs arranged in a fixed order, starting with a timestamp field followed by seven mechanistic characteristic indicators and slag discharge correction factors. The status description text uses UTF-8 character encoding, with a maximum length limit of 512 bytes to ensure transmission efficiency. The real-time performance of the data acquisition component is ensured through a pipelined architecture, with signal acquisition, feature calculation, and data encapsulation executed in parallel in independent threads. The hardware interface uses the industrial Ethernet protocol, with vibration sensors connected via an IEPE interface, and the slag discharge monitoring equipment supports 4-20mA analog signal output. The embedded software uses a real-time operating system, with data acquisition set as the highest priority in task scheduling. The data caching mechanism employs a dual-buffer alternating read / write approach to prevent data loss. The computational complexity of the feature extraction algorithm has been optimized, and the Fast Fourier Transform uses a lookup table method to reduce computation. Historical statistical values ​​with normalized values ​​are updated every 24 hours to adapt to long-term changes in drilling rig operating conditions.

[0024] The interfaces between the data acquisition component and other system components are clearly defined. Rock breaking waveforms are transmitted via shared memory in bitmap format, and status description text is transmitted using TCP socket streams. Time synchronization uses the NTP protocol, and all data is timestamped. Fault tolerance design includes sensor disconnection detection and data validity verification; outliers are replaced with forward padding. Configuration parameters for the data acquisition component can be adjusted via configuration files, such as feature extraction window size, correlation derivation time length, and normalization parameters. Calibration procedures are performed regularly; vibration sensors are calibrated using a standard vibration table, and the slag discharge monitoring unit is verified using a calibration device. Power consumption management of the data acquisition component meets intrinsic safety requirements for mining applications, and the circuit design includes overvoltage and overcurrent protection. Environmental adaptability meets the high temperature and humidity conditions underground, with a protection level of IP67. Electromagnetic compatibility measures include shielded cables and filtering circuits to reduce underground electrical interference. The operating status of the data acquisition component can be remotely monitored, and a heartbeat mechanism reports component health. Logs record detailed operation events and error codes for easy maintenance and diagnosis.

[0025] The feature extraction algorithm of the data acquisition component is based on rock mechanics principles. Amplitude fluctuation reflects the degree of rock mass inhomogeneity, frequency domain aggregation indicates drill bit wear status, pressure change rate reflects hydraulic system response speed, vibration morphology distinguishes between shear and impact rock breaking mechanisms, rotational speed consistency detects power transmission efficiency, low-frequency energy correlates with rock layer plastic deformation, and high-frequency energy corresponds to brittle fracture processes. A cuttings removal correction factor establishes a quantitative relationship between rock breaking action and cuttings removal effect; a high value indicates timely removal of cuttings. The output of the data acquisition component provides multimodal input to the intelligent analysis component. The rock breaking waveform retains the original signal morphology characteristics, and the state description text provides a structured parameter summary. The data acquisition component is tested under various working conditions, including soft rock drilling, hard rock fracturing, and fault crossing. Vibration signal simulation tests inject standard waveforms to verify the accuracy of feature extraction, and cuttings removal parameter tests use slurry of different concentrations and flow rates. Performance indicators include a data packet loss rate of less than 0.1% and latency of less than 10 milliseconds. The hardware selection of the data acquisition component considers a balance between cost and reliability, and the sensor lifespan is matched with the drilling rig overhaul cycle. The software modules adopt a modular design, and the feature extraction algorithm and correlation derivation can be upgraded independently.

[0026] Example 2: See Figure 3 The multimodal recognition model, as the core module of the intelligent analysis component, is responsible for processing multi-source data from the data acquisition component. The multimodal recognition model employs a deep learning architecture to achieve cross-modal information fusion. The model structure comprises four main parts: a visual encoding unit, a semantic encoding unit, a multimodal fusion unit, and a multimodal parsing unit. The visual encoding unit specifically processes image data such as rock-breaking waveforms; the semantic encoding unit focuses on parsing the semantic information of the state description text; the multimodal fusion unit effectively integrates features from different modalities; and the multimodal parsing unit ultimately outputs the rock-breaking condition classification result. The visual encoding unit uses a deep convolutional neural network structure to process the rock-breaking waveforms. The waveforms are input in single-channel grayscale format, with the image size normalized to 224×224 pixels. The convolutional neural network contains five convolutional layers and three max-pooling layers. The number of filters in the convolutional layers are 64, 128, 256, 512, and 512 respectively, with a filter size of 3×3 and a stride of 1 pixel. Each convolutional layer is followed by a ReLU activation function. The max pooling layer uses a 2×2 window with a stride of 2 to downsample the feature maps. The convolutional neural network is then connected to a global average pooling layer to convert the feature maps into 512-dimensional feature vectors, which are then further reduced to 256-dimensional waveform feature vectors via fully connected layers. The visual encoding unit is pre-trained on the ImageNet dataset and fine-tuned on a rock-breaking waveform dataset to adapt to the specific feature extraction task.

[0027] The semantic encoding unit processes state description text based on the Transformer architecture. The state description text is first segmented to generate a word sequence. The segmented sequence is then input into a word embedding layer to be converted into 128-dimensional word vectors. The word embedding layer uses a trainable parameter matrix. The Transformer encoder contains six attention layers, each with eight attention heads, and the feedforward network has a dimension of 512. Positional encoding uses sine and cosine functions to add positional information to the input sequence. The final output of the semantic encoding unit is a 256-dimensional text feature vector obtained through pooling operations. This text feature vector captures the semantic features and numerical relationships of the state description text. The multimodal fusion unit achieves deep fusion of waveform feature vectors and text feature vectors, employing a cross-attention-based fusion mechanism. Cross-attention calculation is performed in two directions: the waveform feature vector is used as the query vector, and the text feature vector is used as the key and value vectors to calculate cross-attention. Attention calculation uses the scaled dot product attention formula. The query vector and key vector are used to calculate a similarity score, which is then normalized using softmax and weighted and summed on the value vector. The text feature vector is used as the query vector, and the waveform feature vector is used as the key and value vectors for inverse attention computation. The attention outputs in both directions produce 256-dimensional feature vectors respectively. The two feature vectors are concatenated into a 512-dimensional composite vector, which is then dimensionality-reduced through a fully connected layer to obtain a 256-dimensional fused encoding result.

[0028] The multimodal parsing unit receives the fused encoding results and performs working condition classification. The multimodal parsing unit comprises a three-layer fully connected neural network. The first fully connected layer maps the 256-dimensional input to a 128-dimensional feature space, the second fully connected layer further compresses it to 64 dimensions, and the final output layer corresponds to the preset number of working condition categories. The output layer uses the softmax activation function to generate the category probability distribution. The rock breaking working condition classification includes four categories: normal rock breaking, rock mass variation, drill bit wear, and abnormal slag discharge. Each category outputs a confidence score, which represents the model's certainty about the classification result. The training process of the multimodal recognition model uses a labeled dataset containing paired samples of rock breaking waveforms and state description text. The loss function combines classification cross-entropy loss and modality alignment loss. Classification cross-entropy loss optimizes classification accuracy, while modality alignment loss constrains the consistency of different modal features in the latent space. Model optimization uses the Adam optimizer with an initial learning rate of 0.001 and a batch size of 32. The training process employs an early stopping strategy to prevent overfitting; training terminates when the validation set loss no longer decreases for 10 consecutive epochs.

[0029] The inference process of the multimodal recognition model is processed in real time, with rock breaking waveform images and state description text being input into the model simultaneously. Visual encoding units and semantic encoding units perform feature extraction in parallel, a multimodal fusion unit performs feature interaction, and a multimodal parsing unit outputs classification results. The entire inference process runs in a GPU-accelerated environment, with an average processing time controlled within 50 milliseconds to meet real-time requirements. TensorRT is used to optimize inference speed, and quantization techniques are employed to reduce model size and improve inference efficiency. The multimodal recognition model interfaces with other system components using standard data formats: rock breaking waveform images are input as JPEG compressed images, and state description text is input as JSON strings. The model outputs rock breaking condition classification results, including category labels and confidence scores, and the output data is passed to the control component via shared memory. The model's running status is monitored in real time, and a heartbeat mechanism periodically reports the model's health status. Model version management records each update, supporting rapid rollback and A / B testing. The multimodal recognition model's anomaly handling mechanism includes input data validation and model self-checking; input data validation checks whether image size and text format conform to specifications. The model self-checks periodically run diagnostic tests to detect abnormal model weights and performance degradation. When a performance decline is detected, a retraining process is automatically triggered to update the model parameters with the latest data to maintain classification accuracy.

[0030] The multimodal recognition model's resource configuration is optimized for the downhole environment, with dynamic adjustment of GPU memory allocation to prevent memory overflow. The model supports hot updates, replacing model files without interrupting system operation. A logging system records detailed inference processes and performance metrics, facilitating troubleshooting and performance analysis. The multimodal recognition model works collaboratively with the mechanism simulation model; their outputs are comprehensively evaluated by the control component, improving the accuracy and reliability of condition recognition. The visual encoding unit of the multimodal recognition model can accept multi-resolution input, adapting to different sizes of rock-breaking waveforms through bilinear interpolation. The semantic encoding unit supports variable-length text input, processing state description text of varying lengths through attention masks. The cross-attention mechanism of the multimodal fusion unit allows the model to adaptively focus on important features of different modalities, enhancing its adaptability to complex working conditions. The fully connected network parameters of the multimodal parsing unit are adjusted according to the actual application scenario, balancing model complexity and inference speed through network depth and width.

[0031] The training data augmentation strategies for the multimodal recognition model include image rotation, color jitter, and text synonym replacement to improve the model's generalization ability. The cross-modal alignment loss function employs a contrastive learning approach, narrowing the distance between features of different modalities within the same sample and widening the distance between features of different samples. Model regularization uses Dropout technology, with the Dropout ratio in fully connected layers set to 0.5 to prevent overfitting. Batch normalization layers accelerate model convergence and improve training stability. The deployment environment of the multimodal recognition model meets mining explosion-proof requirements, and the hardware platform uses intrinsically safe GPU computing equipment. The software environment is deployed using Docker containers, with fixed dependency library versions to ensure consistency. The model service provides remote call support via the gRPC interface, allowing other system components to access model functions in a distributed manner. Performance monitoring of the multimodal recognition model includes throughput, latency, and accuracy metrics, with monitoring data displayed in real-time on the system management interface. The continuous learning mechanism of the multimodal recognition model supports online updates, with new labeled data automatically added to the training set for incremental model updates. Model quality evaluation uses a confusion matrix and F1 score to ensure that model updates do not cause performance degradation. The version control system records the model iteration history, saving complete training configurations and evaluation reports for each version. The time synchronization mechanism between the multimodal recognition model and the data acquisition components ensures multimodal data alignment, with timestamp errors controlled within milliseconds.

[0032] See Figure 4 This paper presents a multi-dimensional comparison of the processing time of the four core components of a multimodal recognition model: the visual encoding unit, semantic encoding unit, multimodal fusion unit, and multimodal parsing unit. It covers three key indicators: average processing time (ms), maximum processing time (ms), and minimum processing time (ms). As shown in the figure, the visual encoding unit has the highest maximum processing time at 22.1ms, an average processing time of 15.2ms, and a minimum processing time of 8.7ms, indicating significant time fluctuations when processing rock breaking waveforms. The semantic encoding unit has a maximum processing time of 18.5ms, an average processing time of 12.8ms, and a minimum processing time of 7.3ms. The multimodal fusion unit has a maximum processing time of 15.3ms, an average processing time of 10.5ms, and a minimum processing time of 6.2ms. The multimodal parsing unit has a maximum processing time of 12.7ms, an average processing time of 8.3ms, and a minimum processing time of 4.5ms. From visual encoding and semantic encoding to multimodal fusion and multimodal parsing, the average, maximum, and minimum processing times of each component show a gradual decreasing trend, reflecting the time efficiency hierarchy of the multimodal recognition model in component design. This also provides professional quantitative basis for subsequent performance optimization, resource allocation, and overall real-time improvement of each component, contributing to the efficient operation of the intelligent analysis component in the collaborative control system for rock breaking and slag removal of mining bottom corner drilling rigs.

[0033] Example 3: The mechanism simulation model, as the core analysis module of the intelligent analysis component, plays a crucial role in analyzing the rock-breaking process from a physical mechanism perspective. The mechanism simulation model constructs a mathematical model based on rock mechanics principles and drilling rig operating characteristics. The model input is the rock-breaking waveform generated by the data acquisition component, and the output is the working condition matching result. The control component receives the rock-breaking working condition classification output by the multi-modal recognition model and the working condition matching result output by the mechanism simulation model, and generates drilling rig control commands through comprehensive evaluation. When analyzing the characteristic parameters of the rock-breaking waveform, the mechanism simulation model employs a multi-dimensional quantitative analysis method. The characteristic dimensions cover nine aspects: amplitude level, event frequency, time interval, low-frequency to high-frequency energy ratio, waveform symmetry, peak count, distribution pattern, acoustic detection probability, and phase characteristics. The amplitude level is obtained by calculating the root mean square (RMS) value and peak factor of the waveform signal. The RMS value reflects the average energy of the signal, and the peak factor characterizes the impact intensity. The calculation process uses a sliding window with a window size of 256 sampling points. The event frequency is counted to determine the number of waveform events exceeding a threshold amplitude per unit time. The threshold is dynamically adjusted based on the 85th percentile of historical data. The time interval is measured to determine the time difference between consecutive waveform events, and a peak detection algorithm is used to identify the starting point of the waveform events. The low-frequency to high-frequency energy ratio is calculated by performing a 5-level wavelet packet decomposition on the signal, using the db4 wavelet basis function. Waveform symmetry is quantified by calculating the Pearson correlation coefficient between the positive and negative half-cycles of the waveform. The number of peaks is statistically determined using an algorithm based on local extremum detection, and the extrema must satisfy a minimum peak height condition. The distribution pattern is classified by fitting a Gaussian mixture model of the amplitude distribution. The acoustic wave detection probability is calculated based on the time-domain correlation between the acoustic emission sensor signal and the vibration signal. Phase characteristics are extracted using Hilbert transform to extract instantaneous phase information, and the phase lock value is calculated.

[0034] The mechanism simulation model compares the similarity of nine feature dimensions with pre-stored multi-condition mechanism models, which contain a standard feature vector library of typical working conditions. The similarity comparison employs an improved cosine similarity calculation method.

[0035] in: This indicates the similarity score of the working conditions. This represents the weighting coefficient for the d-th feature dimension. Indicates the input feature parameter values. Indicates the template feature parameter value. This represents the historical mean of the d-th feature dimension. Weighting coefficients are assigned based on feature importance: amplitude level weighting coefficient 0.15, event frequency weighting coefficient 0.14, time interval weighting coefficient 0.12, low-frequency to high-frequency energy ratio weighting coefficient 0.11, waveform symmetry weighting coefficient 0.11, peak count weighting coefficient 0.10, distribution pattern weighting coefficient 0.10, acoustic detection probability weighting coefficient 0.09, and phase feature weighting coefficient 0.08. After similarity scoring is completed, the mechanism simulation model selects the working condition corresponding to the highest score as the working condition matching result. When the highest score is below a preset threshold of 0.7, an unknown working condition identifier is output.

[0036] The multimodal recognition model outputs a rock-breaking condition classification, including the condition type and its confidence level. The confidence level indicates the model's certainty regarding the classification result. A weight allocation mechanism is established during the comprehensive evaluation by the control component, assigning multimodal weight values ​​based on the confidence level; higher confidence levels result in larger multimodal weight values. The rock-breaking condition classification and condition matching results are weighted and fused based on the comprehensive weight values ​​to determine the final condition type. The control component generates drilling rig control commands based on the final condition type. These commands include instructions for drill bit speed adjustment, feed pressure setting, and cuttings removal parameter optimization. Drill bit speed adjustment is determined based on the rock mass drillability level, which is divided into six levels, each corresponding to a different speed range. Feed pressure setting considers the rock mass compressive strength characteristics and is determined using a lookup table method. Cuttings removal parameter optimization is calculated based on the cuttings generation rate, estimated by multiplying the drilling speed by the borehole cross-sectional area. Control commands are transmitted to the drilling rig actuators via a PROFIBUS fieldbus with a transmission cycle of 100 milliseconds.

[0037] The mechanism simulation model is configured with a self-updating mechanism to improve model adaptability. This mechanism continuously records the feature parameters obtained from each analysis and the final determined working condition matching result. The recorded data includes timestamps, feature parameter vectors, and working condition matching result labels, stored in a circular database with a capacity of 10,000 records. When the rock breaking working condition classification output by the multimodal recognition model differs from the working condition matching result five consecutive times, a model optimization process is triggered. The model optimization process, based on the recorded feature parameters and rock breaking working condition classification data, performs parameter correction on the pre-stored working condition mechanism model. Parameter correction uses the momentum gradient descent algorithm, with a learning rate of 0.001, a momentum parameter of 0.9, and a batch size of 32. The mechanism simulation model and the multimodal recognition model establish a data exchange channel, exchanging feature extraction methods and classification result information every 10 minutes. The mechanism simulation model provides feature importance ranking to the multimodal recognition model, and the multimodal recognition model feeds back classification confidence information to the mechanism simulation model. Data exchange uses JSON format and achieves high-speed communication through shared memory. The control component implements a safety-first control strategy, automatically adjusting the weight allocation scheme upon receiving an early warning signal from the safety monitoring component. Under safety warning conditions, the weight value of the mechanism simulation model increases to 0.8, while the weight value of the multimodal recognition model decreases to 0.2. The control component maintains a priority list for control commands, with safety-related commands having the highest execution priority, followed by normal optimization commands, with priority values ​​ranging from 0 to 255.

[0038] The real-time performance of the mechanism simulation model is guaranteed through algorithm optimization. Feature parameter parsing employs a 9-thread parallel computing architecture, with each thread processing one feature dimension. The similarity comparison process is accelerated using SIMD vectorization, and pre-stored standard feature vectors are loaded into the L2 cache. Model updates are executed in a background thread with a priority lower than the real-time control thread. A feedback loop is established between the mechanism simulation model and the data acquisition component, with feature parameter parsing results fed back to the data acquisition component to optimize acquisition parameters. When poor data quality is detected in certain feature dimensions, the mechanism simulation model suggests that the data acquisition component adjust the sampling frequency or filtering parameters. Data quality assessment is based on signal-to-noise ratio (SNR) and signal integrity metrics, with an SNR threshold set at 20 dB. The long-term learning capability of the mechanism simulation model is enhanced through knowledge graph technology. Recorded operating condition data is constructed into a knowledge graph containing nodes and relationships. The knowledge graph contains the relationships between operating condition features, matching results, control effects, and other elements, and deep patterns are mined using graph convolutional networks. The knowledge graph is fully updated weekly and incrementally updated daily. The control component implements multi-objective optimization control, comprehensively considering three optimization objectives: rock-breaking efficiency, drill bit wear, and energy consumption. The multi-objective optimization employs a constrained Pareto front search algorithm, with constraints including equipment safety limits and process requirements. The control component evaluates the control effect every 30 minutes, adjusting the optimization objective weights based on the actual rock-breaking and slag-removal results.

[0039] Example 4: The safety monitoring component, as an independent protection module of the system, focuses on real-time monitoring and early warning of the drilling rig's operational safety status. By establishing a complete drilling rig fault feature library and developing a multi-parameter fusion analysis algorithm, the safety monitoring component constructs a multi-layered protection system from status monitoring to risk warning and safety intervention. The safety monitoring component establishes a close data interaction mechanism with the data acquisition and control components, forming a complete monitoring closed loop. The drilling rig fault feature set established by the safety monitoring component includes three main fault types: drill bit wear, enhanced rock resistance, and slag blockage. Drill bit wear fault features are identified by analyzing the variation patterns of high-frequency components in vibration signals. Wavelet packet decomposition technology is used to extract energy feature values ​​in the 8000-12000Hz frequency band. The wavelet packet decomposition has 5 layers, using the db4 wavelet basis function, and the frequency band energy percentage is calculated every 100 milliseconds. A wear quantification model is established based on the cumulative working time of the drill bit, taking into account factors such as drill bit type, rock strength, and drilling depth. The characteristics of rock mass resistance enhancement faults are identified by monitoring the amplitude increase rate and energy accumulation trend of the rock breaking waveform. A time series analysis method is used to detect the monotonically increasing amplitude, and an amplitude gradient threshold is set as the judgment condition. This threshold is dynamically adjusted according to the drilling rig model. The characteristics of slag discharge blockage faults are identified by analyzing the dynamic changes in slag discharge parameters. Abrupt changes in slag discharge velocity and a continuous upward trend in slag discharge consistency are monitored. A sliding window statistical method is used to detect abnormal fluctuations, with a sliding window size of 10 sampling points.

[0040] The safety monitoring component acquires the drilling rig equipment identification and correlates it with real-time vibration signals. The drilling rig equipment identification uses a unique coding rule and includes equipment model, factory serial number, and installation location information. The probability of fault occurrence is calculated using a multi-sensor data fusion algorithm. The probability of drill string wear is calculated by weighting three parameters: high-frequency vibration energy, drilling speed fluctuation rate, and torque change value, with weighting coefficients of 0.4, 0.3, and 0.3, respectively. The probability of enhanced rock mass resistance is determined by comprehensively considering three indicators: amplitude growth rate, energy accumulation value, and propulsion pressure rise rate, with weighting coefficients of 0.5, 0.3, and 0.2, respectively. The probability of slag discharge blockage is analyzed by jointly considering three parameters: flow rate decrease rate, consistency increase value, and pump pressure fluctuation value, with weighting coefficients of 0.4, 0.4, and 0.2, respectively. The probability calculation uses a Bayesian inference framework. Prior probabilities are derived from historical statistical data, and the likelihood function is calculated based on the deviation of the current feature value from the normal range. The risk report generated by the safety monitoring component uses a standardized data structure, including five fields: fault type, probability value, risk level, timestamp, and recommended measures. Risk levels are categorized into three levels—Attention, Warning, and Emergency—based on probability values. A probability value of 0.3-0.6 indicates the Attention level, 0.6-0.8 indicates the Warning level, and a probability value above 0.8 indicates the Emergency level. Risk reports are updated every 5 seconds, storing the 100 most recent records in a circular buffer, supporting historical data queries and trend analysis.

[0041] The logic for generating early warning signals by the safety monitoring component is based on multi-dimensional analysis of risk reports. A primary early warning signal is triggered when the risk level of any fault type reaches the warning level, and a higher-level early warning signal is triggered when it reaches the emergency level. The early warning signal includes four parts: warning type, risk level, trigger time, and handling recommendations, and is output simultaneously via digital and visual signals. The digital signal is transmitted using the Modbus RTU protocol, and the visual signal is displayed as a red flashing warning on the HMI interface. When generating an early warning signal, the safety monitoring component simultaneously acquires the drilling rig control commands generated by the control component and establishes a safety intervention decision model. The safety intervention decision model assesses the degree of matching between the current risk status and the control commands, using a combination of rule-based reasoning and fuzzy logic. Rule-based reasoning is based on a pre-set safety rule base, containing the types of operations allowed under different risk levels. Fuzzy logic handles boundary cases, evaluating the severity of the risk and the aggressiveness of the control commands through membership functions. The safety instructions generated by the safety monitoring component are structured, including five elements: instruction number, intervention level, effective time, duration, and recommended operation. Intervention levels are divided into three levels: observation, restriction, and stop. The observation level maintains monitoring without intervention, the restriction level adjusts control parameters, and the stop level interrupts the current operation. Recommended operations are carefully designed based on the type of fault. For drill bit wear, it is recommended to reduce the rotation speed and feed pressure. For increased rock resistance, it is recommended to adjust the rock breaking parameters. For slag discharge blockage, it is recommended to clean the pipeline and optimize the slurry parameters.

[0042] Safety commands are sent to the control component via a priority management mechanism. The control component establishes a safety command processing queue, prioritizing the execution of higher-level safety commands. The execution results of the safety commands are fed back to the safety monitoring component, forming a closed-loop verification mechanism. When a safety command fails to effectively reduce the risk, the safety monitoring component escalates the intervention level and generates a new safety command. Specific parameters for the drilling rig fault characteristic set are shown in Table 1. Table 1: Drilling Rig Fault Characteristic Parameters Table

[0043] The hardware platform of the safety monitoring component adopts an industrial-grade embedded system, equipped with a multi-channel data acquisition card and an isolated digital output module. The system software is developed based on a real-time operating system and employs a modular design, comprising four core modules: data acquisition, feature extraction, probability calculation, and decision generation. The data acquisition module is responsible for reading and preprocessing sensor data; the feature extraction module performs real-time calculation of fault characteristics; the probability calculation module completes multi-parameter fusion analysis; and the decision generation module produces early warning signals and safety commands.

[0044] The safety monitoring component communicates with other system components using the industrial Ethernet protocol, with data transmission following a defined cycle and frame format. The data exchange cycle with the data acquisition component is 10 milliseconds, and the command transmission cycle with the control component is 20 milliseconds. The communication protocol includes a complete data verification mechanism to ensure transmission reliability. The safety monitoring component's self-diagnostic function periodically checks the system's operating status, including sensor connection status, computing resource utilization, and communication link quality. Self-diagnostic results are recorded in the system log; abnormal states trigger local alarms and are reported to the monitoring center. The system maintenance interface supports parameter configuration, data export, and fault injection testing, facilitating on-site debugging and performance optimization. The safety monitoring component's historical data storage uses a time-series database, storing complete operating data for the most recent 30 days. Data compression algorithms reduce storage space usage, and the fast retrieval function supports multi-condition queries. Data analysis tools provide trend charts and statistical reports to help managers understand equipment health status. The safety monitoring component's parameter tuning interface provides a visual operation method, allowing authorized users to adjust fault judgment thresholds, risk level boundaries, and safety command parameters. Parameter modifications are logged in an audit log to prevent unauthorized changes. The system supports saving and switching between multiple sets of parameter configurations to adapt to different working conditions and equipment statuses.

[0045] The performance indicators of the safety monitoring component include three dimensions: detection accuracy, response time, and reliability. Detection accuracy is verified by comparing actual faults with early warning records. Response time is measured from the appearance of a feature to the issuance of an early warning. Reliability is evaluated using mean time between failures (MTBF). The system design meets the requirements for long-term stable operation in a mining environment, and its protection level reaches IP65 standard. The safety monitoring component establishes data interface with the mine's centralized monitoring system, uploading safety status and early warning information in real time. The remote monitoring center can view the safety status of any drilling rig and issue control commands to adjust safety strategies. This distributed monitoring architecture achieves unified management of individual machine safety and system safety, improving the overall safety level of the mine production system. The redundant design of the safety monitoring component ensures high system availability. Key sensors adopt a dual-redundant configuration, and the computing module has a hot backup mechanism. When the main system fails, the backup system switches over within 50 milliseconds, ensuring the continuity of safety monitoring. The system periodically performs automatic fault switching tests to verify the reliability of the redundant system. The clock synchronization mechanism of the safety monitoring component is synchronized with the mine's time server, with a time error controlled within 1 millisecond. Precise timestamps ensure the correctness of the event sequence in a distributed system, providing a reliable basis for incident analysis and liability determination. System logs contain complete timestamp sequences, supporting post-incident reconstruction of the system's operational state.

[0046] Example 5: The multimodal fusion unit employs a cross-attention fusion mechanism to process waveform feature vectors and text feature vectors. The waveform feature vectors are derived from the feature extraction results of the rock-breaking waveform diagram by the visual encoding unit, while the text feature vectors are derived from the encoded output of the state description text by the semantic encoding unit. During the cross-attention calculation process, the waveform feature vector is used as the query vector and interacts with the key vector of the text feature vector to calculate similarity, generating an attention weight distribution. The value vector of the text feature vector is then weighted and summed to produce the first fusion vector. The same calculation process is executed synchronously in the reverse path, where the text feature vector interacts with the key vector of the waveform feature vector to generate the second fusion vector. After concatenating the two fusion vectors, dimensionality reduction is achieved through a fully connected layer, and the fusion encoding result is output to the multimodal parsing unit. The self-updating mechanism of the mechanism simulation model records system operation data through a circular buffer. The capacity of the circular buffer is set to 1000 records, each containing a timestamp, feature parameter vector, working condition matching result, and rock-breaking working condition classification result. The feature parameter vector stores normalized values ​​for nine feature dimensions, including amplitude level, event frequency, and time interval, with a data recording interval of 500 milliseconds. When the rock breaking condition classification output by the multimodal recognition model is found to be inconsistent with the condition matching result output by the mechanism simulation model for five consecutive times, the self-update mechanism triggers the model optimization process. The optimization process extracts the most recent 50 inconsistent records from the circular buffer as training samples, and the sample data is input into the incremental learning algorithm after standardization.

[0047] Model optimization employs stochastic gradient descent to adjust the parameters of the mechanistic simulation model, updating the weight coefficients and standard feature vector values ​​for the nine feature dimensions. Training uses mini-batch data input with a batch size of 10 samples and a learning rate of 0.001. The attention weight parameters of the multimodal fusion unit are simultaneously optimized during system maintenance, using high-quality, consistently labeled samples from a circular buffer. A learning rate decay strategy is employed to prevent overfitting. The self-updating mechanism includes a model performance evaluation module that periodically calculates the consistency ratio between the mechanistic simulation model and the multimodal recognition model. When the consistency ratio of the most recent 100 outputs falls below 85%, a reinforcement optimization process is initiated. This process uses the most recent 500 records as training data, extending the training period to 100 iterations, and adjusting the learning rate to 0.0005. During optimization, a parameter freeze command is sent to the multimodal fusion unit via a message queue to ensure stable feature extraction. Parameter freeze is lifted upon optimization completion, with an average execution time of 3.2 seconds.

[0048] Taking drilling rig ZK-2023-078 drilling through granite strata as an example, the system uses vibration sensors to collect impact signals with an amplitude of ±5V. The data acquisition component generates a rock-breaking waveform diagram containing periodic spikes. Cross-attention calculation by the multimodal fusion unit shows that the waveform features have an attention weight of 0.78 in the 4kHz-6kHz frequency band, and the attention weight for the rotational speed fluctuation parameter in the text features is 0.65. The fusion encoding result is output as a rock mass variation classification by the multimodal analysis unit, with a confidence level of 0.87. The mechanism simulation model detects an increase in event frequency from 10 times / second to 25 times / second, a decrease in the waveform symmetry coefficient from 0.9 to 0.7, and a matching degree of 0.92 with the standard features of rock mass variation. The self-updating mechanism records the feature parameter vector, including data such as amplitude level 0.8 and event frequency 25. After detecting five consecutive consistent outputs within the next two minutes, the self-updating mechanism updated the standard feature vector of the rock mass variation condition, adjusting the event frequency threshold from 20 times / second to 22 times / second, and increasing the waveform symmetry coefficient weight from 0.1 to 0.15. When the drilling rig entered the fault fracture zone, the rock fracture waveform diagram showed increased amplitude fluctuations. The multimodal recognition model output a slag discharge anomaly classification, and the mechanism simulation model matched the drill bit wear condition based on waveform symmetry of 0.3 and acoustic detection probability of 0.2. The self-updating mechanism detected the output difference and recalibrated the feature weights using 30 sets of fault zone data. After optimization, the mechanism simulation model output a slag discharge anomaly classification under similar conditions, and the consistency was improved. The multimodal fusion unit adjusted the attention weight distribution, increasing the waveform symmetry attention weight from 0.2 to 0.35, the acoustic detection feature weight from 0.15 to 0.25, and the slag discharge parameter weight in the text features to 0.4. After 10 running cycles of optimization, the model's consistency rate in the fault zone increased from 70% to 92%.

[0049] The quality control module of the self-updating mechanism verifies the effectiveness of parameter updates, using a validation set of 200 samples to test the model accuracy before and after optimization. An update is accepted when the accuracy improvement exceeds 3%; otherwise, it rolls back to the previous version within 0.5 seconds. After three months of system operation, the mechanistic simulation model showed a 12% improvement in accuracy in hard rock formations and an 18% improvement in fractured zone formations. The self-updating mechanism completed 47 parameter adjustments, with the model consistency ratio consistently above 95%. The collaborative work of the multimodal fusion unit and the self-updating mechanism significantly improved the quality of drilling rig control command generation, increased average drilling efficiency in complex formations by 15%, and shortened fault warning time by 20%. False alarms from the safety monitoring component decreased by 30%, resulting in a comprehensive improvement in system reliability. This self-improvement mechanism ensures the system continuously adapts to changes in geological conditions and maintains optimal operating conditions.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A collaborative control system for rock breaking and slag removal in a mining bottom-angle drilling rig, characterized in that, The system integrates data acquisition components, intelligent analysis components, and control components, wherein: The data acquisition component is configured to collect vibration signals and slag discharge parameters during the rock breaking process of the drilling rig in real time, and generate rock breaking waveform diagrams and status description text based on these signals and parameters; The intelligent analysis component has a built-in multimodal recognition model and a mechanism simulation model. The multimodal recognition model is used to jointly process the rock breaking waveform diagram and the state description text to output the rock breaking condition classification. The mechanism simulation model is used to parse the feature parameters from the rock breaking waveform diagram, match them with the standard mechanism library, and output the condition matching result. The control component performs a comprehensive evaluation based on the output classification and matching results of the intelligent analysis component, generates drilling rig control commands, and optimizes the rock breaking and slag removal process.

2. The collaborative control system for rock breaking and slag removal of a mining bottom-angle drilling rig according to claim 1, characterized in that, While generating the rock-breaking waveform, the data acquisition component performs feature extraction operations to derive multiple mechanistic feature indicators from the vibration signal. These indicators include amplitude fluctuation, frequency domain concentration, pressure change rate, vibration mode, rotational speed consistency, low-frequency energy, and high-frequency energy. This component also calculates the correlation strength between rock breaking parameters and slag discharge parameters, and derives a slag discharge correction factor through correlation deduction. Subsequently, the mechanistic characteristic indicators and slag discharge correction factors were numerically normalized and text-mapped, and compiled into a state description text with a key-value structure.

3. The collaborative control system for rock breaking and slag removal of a mining bottom-angle drilling rig according to claim 2, characterized in that, The slag discharge parameters include slag discharge flow rate readings and slag discharge consistency readings; The correlation derivation includes: calculating the correlation between rock-breaking vibration signals and slag discharge velocity readings at the same time interval to obtain a velocity correlation factor; calculating the correlation between rock-breaking vibration signals and slag discharge consistency readings to obtain a consistency correlation factor; and then synthesizing the velocity correlation factor and the consistency correlation factor according to a predetermined ratio to generate a slag discharge correction factor.

4. The collaborative control system for rock breaking and slag removal of a mining bottom-angle drilling rig according to claim 1, characterized in that, The multimodal recognition model includes a visual coding unit, a semantic coding unit, a multimodal fusion unit, and a multimodal parsing unit; When the multimodal recognition model processes the rock-breaking waveform and the state description text, it performs feature encoding on the rock-breaking waveform through a visual encoding unit to obtain a waveform feature vector, and performs feature encoding on the state description text through a semantic encoding unit to obtain a text feature vector. Based on cross-modal alignment loss and image-text consistency loss, waveform feature vectors and text feature vectors are input into a multimodal fusion unit for feature integration to obtain fusion coding results; The fusion encoding result is input into the multimodal parsing unit for decoding, and the rock breaking condition classification is output.

5. The collaborative control system for rock breaking and slag removal of a mining bottom-angle drilling rig according to claim 1, characterized in that, When analyzing characteristic parameters, the mechanism simulation model quantifies multiple characteristic dimensions of the rock-breaking waveform, including amplitude level, event frequency, time interval, low-frequency to high-frequency energy ratio, waveform symmetry, number of peaks, distribution pattern, acoustic detection probability, and phase characteristics. The similarity of multiple feature dimensions with pre-stored multi-type working condition mechanism models is compared to obtain each similarity score, and the working condition corresponding to the highest score is taken as the working condition matching result.

6. The collaborative control system for rock breaking and slag removal of a mining bottom-angle drilling rig according to claim 1, characterized in that, The rock-breaking condition classification output by the multimodal recognition model includes the condition type and its confidence level; When the control component performs a comprehensive evaluation, it assigns multimodal weight values ​​based on confidence level and mechanism weight values ​​based on similarity score. The multimodal weight values ​​and mechanism weight values ​​are standardized to obtain a comprehensive weight value. The final working condition type is then determined based on the comprehensive weight value, rock breaking condition classification, and working condition matching results.

7. The collaborative control system for rock breaking and slag removal of a mining bottom-angle drilling rig according to claim 1, characterized in that, It also includes security monitoring components; The safety monitoring component establishes a set of drilling rig fault characteristics, including drill bit wear, increased rock mass resistance, and slag discharge blockage. Obtain the drilling rig equipment identification and combine it with real-time vibration signals to determine whether it meets the fault characteristic set, calculate the probability of occurrence of each fault characteristic, and generate a risk report; Early warning signals are generated based on risk reports.

8. The collaborative control system for rock breaking and slag removal of a mining bottom-angle drilling rig according to claim 7, characterized in that, When the security monitoring component generates an early warning signal, it performs the following operations: Obtain the drilling rig control commands generated by the control component; Based on the aforementioned risk report and drilling rig control commands, determine whether safety intervention is required; If necessary, a safety instruction containing the intervention level and recommended action is generated and sent to the control component; The control component is configured to prioritize responding to the safety instructions and adjust the drilling rig control commands.

9. The collaborative control system for rock breaking and slag removal of a mining bottom-angle drilling rig according to claim 4, characterized in that, When performing feature integration, the multimodal fusion unit employs a cross-attention-based fusion mechanism, specifically including: Using the waveform feature vector as the query vector and the text feature vector as the key vector and value vector, cross-attention calculation is performed to obtain the first fusion vector; The text feature vector is used as the query vector, and the waveform feature vector is used as the key vector and value vector. Cross-attention calculation is performed to obtain the second fusion vector. The first fusion vector and the second fusion vector are concatenated, and then dimensionality is reduced through a fully connected layer to obtain the fusion encoding result.

10. The collaborative control system for rock breaking and slag removal of a mining bottom-angle drilling rig according to claim 5, characterized in that, The mechanism simulation model is also equipped with a self-updating mechanism, which includes: Record the feature parameters obtained from each analysis and the final working condition matching result; When the rock breaking condition classification output by the multimodal recognition model differs from the condition matching result by a preset number of times, the model optimization process is triggered. Based on the recorded feature parameters and the rock-breaking condition classification, the parameters of the pre-stored condition mechanism model are corrected.

Images