A deep neural network-based black-lamp mine multi-dimensional running state perception method

Abstract

The application relates to the field of industrial automatic control system device manufacturing, and discloses a black-lamp mine multi-dimensional running state sensing method based on a deep neural network, which comprises the following steps: collecting real-time control instructions and multi-dimensional observation data of mine equipment; generating a prior state prediction value by using a preset neural dynamics prediction model, wherein the model is embedded with a physical constraint term based on a nonlinear dynamics equation, the evolution track is limited to satisfy the energy conservation and the kinematics continuity boundary; calculating a multi-dimensional residual error vector between the observation data and the prediction value and a statistical variance thereof; generating dynamic credibility weights of each sensing channel by using a nonlinear attenuation strategy, and weighting and fusing the prediction value and the observation data to generate a posterior state estimation value to drive a closed-loop control; and the neural observer embedded with the physical constraint realizes decoupling determination of sensor failure and equipment physical failure, and ensures the physical authenticity and stability of the control loop under the condition of strong interference and partial sensing failure.

Description

Technical Field

[0001] This invention relates to a method for multi-dimensional operational status perception in dark mines based on deep neural networks, belonging to the field of industrial automatic control system device manufacturing technology. Background Technology

[0002] Currently, automated control of fully mechanized longwall mining faces in dark mines relies on multi-dimensional sensing feedback from coal mining machines and hydraulic supports to maintain closed-loop operation. The mainstream control strategy adopts a multi-sensor fusion architecture to collect motor current, hydraulic pressure, and visual image signals. After logical judgment, control commands are generated to realize cutting trajectory planning and overload protection. Such systems assume that sensor data directly corresponds to the physical state of the equipment and are scheduled based on preset static thresholds or statistical feature models.

[0003] When underground fully mechanized mining environments are accompanied by high-frequency mechanical vibrations, water mist and dust, and sudden changes in rock hardness, severe vibrations cause distortion of inertial sensor signals, and mud splashes obstruct the visual channel. Under these non-ideal observation conditions, existing control systems lack a self-checking mechanism based on the dynamic mechanism of the controlled object, making it impossible to distinguish between sensor signal failure and equipment physical faults. Data-driven identification algorithms lack physical constraints in real-time control loops, and input signals are disturbed, resulting in non-physical divergence and causing the controller to issue incorrect shutdown commands or abnormal actions. Simply relying on multi-dimensional correlation at the algorithm level cannot compensate for the lack of physical mechanism constraints. For example, Chinese invention patent CN120540181A discloses a method for monitoring unit operation status based on multi-dimensional model prediction. Although the scheme constructs an equipment operation network and parameter correlation group, attempting to improve the monitoring dimensions by using causal, temporal, and operational correlation functions, the core logic is still based on historical data. Based on statistical regression of operational data, this method is essentially a pure data-driven mapping relationship. However, in unstructured environments, the lack of rigid body dynamics and energy conservation physical mechanisms within the model makes it prone to model mismatch when the operating conditions exceed the statistical range covered by the training samples. It is difficult to decouple the drift of the sensor itself from the actual physical evolution of the equipment, resulting in the control system being unable to confirm the physical reality when faced with data anomalies. Existing passive observation modes have defects in steady-state system operation. When the coal mining machine is cutting at a constant traction speed or steady rotation speed, the lack of frequency excitation in the control input leads to a decrease in the observability of system parameters. Relying solely on passive feedback makes it difficult to decouple the relationship between load characteristic changes and actuator wear, resulting in a lag in the control system's perception of the wear of the cutting teeth or the evolution of the physical properties of coal seam interbedded with gangue. In addition, a significant temperature rise underground causes thermo-mechanical coupling drift, and fixed model parameters cannot adapt to the evolution of physical parameters, leading to prediction bias.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a physical dynamic constraint state observation mechanism to maintain the physical authenticity of the control loop under strong interference and partial sensing failure conditions, and how to solve the problem of steady-state parameter identification through active excitation. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A method for multi-dimensional operation status perception in a dark mine based on deep neural networks, which is applied to an industrial automatic control system including a multi-dimensional sensing unit and a closed-loop controller, the method comprising the following steps: The system collects real-time control command data from mining equipment and real-time observation data from multi-dimensional sensing units, including one-dimensional time-series data reflecting the physical properties of the equipment and visual feature data reflecting the spatial characteristics of the equipment. The posterior state estimate from the previous moment and the real-time control command data are input into a pre-set neurodynamic prediction model. The neurodynamic prediction model is used to generate the prior state prediction for the current moment. The neurodynamic prediction model is constructed based on the nonlinear dynamic equations of the mining equipment, and its loss function includes physical constraint terms. The physical constraint terms force the evolution trajectory of the prior state prediction to satisfy energy conservation and kinematic continuity boundaries during the model training and inference stages. Calculate the multidimensional residual vector between real-time observation data and prior state prediction values, and use channel weighted logic to analyze the statistical distribution characteristics of the multidimensional residual vector in each sensing channel, thereby generating dynamic confidence weights for each sensing channel. Based on dynamic confidence weights, a weighted fusion calculation is performed on the prior state prediction value and the real-time observation data to generate the posterior state estimate value at the current moment. The posterior state estimate is input to the closed-loop controller, which calculates the deviation between the posterior state estimate and the preset safe operating boundary, and generates equipment control commands based on the deviation.

[0006] Preferably, the method further includes a steady-state parameter identification step based on active excitation: real-time monitoring of the time series variance of the posterior state estimate; when the time series variance is lower than a preset activity threshold, superimposing a micro-perturbation excitation signal with a preset frequency and amplitude into the real-time control command data; setting the amplitude of the micro-perturbation excitation signal to be lower than the dead zone threshold for driving the mining equipment to produce effective mechanical action; acquiring the transient response characteristic data of the mining equipment to the micro-perturbation excitation signal, and converting the transient response characteristic data into dynamic load impedance parameters; inputting the dynamic load impedance parameters as constraints into the neurodynamic prediction model, and using the neurodynamic prediction model to correct the prior state prediction value based on the dynamic load impedance parameters.

[0007] Preferably, the method further includes a thermal drift compensation step based on temperature monitoring: acquiring real-time temperature data of the actuator of the mining equipment; converting the real-time temperature data into dynamic boundary parameters of physical constraint terms using a preset temperature rise parameter compensation model; the temperature rise parameter compensation model characterizes the evolution law of the physical property parameters of the mining equipment with temperature changes; and adjusting the constraint tolerance range of the physical constraint terms in real time using the dynamic boundary parameters during the operation of the neurodynamic prediction model.

[0008] Preferably, the step of generating dynamic reliability weights using channel weighted logic includes: for each sensing channel, calculating the statistical variance of its corresponding residual component within a preset sliding time window; comparing the statistical variance with a preset reference noise amplitude; when the statistical variance exceeds the reference noise amplitude, performing a decay operation on the weight value corresponding to the sensing channel in the dynamic reliability weights using a nonlinear decay strategy; the nonlinear decay strategy is configured to isolate sensing channel data that undergoes abrupt changes and does not conform to physical correlation, while maintaining the control loop closure.

[0009] Preferably, the method further includes a physical mechanism-based working condition adaptation step: real-time monitoring of the overall modulus of the multidimensional residual vector; when the overall modulus continuously exceeds a preset working condition change threshold, acquiring the running data segment within a preset time window at the current moment, and using the least squares method to identify key physical attribute parameters under the current working condition based on a simplified physical model; generating a linear gain adjustment factor based on the change in the key physical attribute parameters, and using the linear gain adjustment factor to weight and correct the output layer of the neurodynamic prediction model, so that the response characteristics of the neurodynamic prediction model are adapted to the current geological working condition.

[0010] Preferably, in the step of generating dynamic credibility weights, for any given... The perception channel and its corresponding dynamic reliability weight Calculate according to the following formula: ,in, For the first Dynamic reliability weights of the perception channel For the corresponding number in the multidimensional residual vector The residual modulus of the sensing channel, The preset residual tolerance threshold, This is a preset sensitivity coefficient used to adjust the weight decay rate; the formula is used to suppress the channel weight when the residual magnitude exceeds the residual tolerance threshold.

[0011] Preferably, the method further includes an active calibration step for minute drifts: during the steady-state operation of the mining equipment, a test signal is generated and superimposed onto the selected sensing channel data; the response sensitivity characteristic data of the neurodynamic prediction model to the test signal is obtained; the sensor's inherent drift component is separated from the raw data of the sensing channel using the response sensitivity characteristic data, and a negative feedback compensation loop is constructed; the sensor's inherent drift component is used to perform online zero-point calibration on subsequent real-time observation data of the sensing channel.

[0012] Preferably, the method further includes a rule-based redundancy control step: running a set of boundary protection procedures based on deterministic logic rules in parallel; monitoring the overall confidence level of dynamic confidence weights in real time; when the overall confidence level is lower than a preset safety threshold or when the posterior state estimate indicates that the device state is close to the limit value of the safe operation boundary, blocking the output of the posterior state estimate and directly switching to the boundary protection procedure to take over the right to generate device control commands.

[0013] Preferably, the calculation steps of the multidimensional residual vector include: converting the visual feature data into a feature vector of the same dimension as the one-dimensional time series data; calculating the Euclidean distance or cosine similarity between the real-time observation data and the prior state prediction value in the feature vector space to generate a residual component representing the consistency of the visual channel; and combining the residual component with the algebraic difference of the one-dimensional time series data to construct a multidimensional residual vector.

[0014] Preferably, the method further includes an offline update step for model parameters: storing historical running data and corresponding fault markers on the server side; periodically training the neurodynamic prediction model using the historical running data, with the optimization objective being to minimize the long-term statistical mean of the multidimensional residual vector; and sending the updated model parameters to the edge controller to replace the current model parameters used to generate prior state prediction values.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In the multi-dimensional operation of a mine, a neurodynamic prediction model with a physical consistency penalty term is constructed. A state reconstruction mechanism with the mechanical kinematic limit of the equipment as the boundary is established. The control command is used as a priori input. A virtual reference trajectory is generated using physical conservation laws and compared with the measured data of the multi-dimensional sensing unit in real time. When high-frequency vibration or water mist obstruction in the mine causes the signal of a single sensor to be distorted or interrupted, the inertial characteristics of the physical model are used to fill the sensing blind zone. The output state vector is forced to converge within the physically permissible solution space. This eliminates the non-physical divergence generated when the input signal is decoupled in a pure data-driven model. It ensures that the state estimate fed back to the main controller has the continuity of dynamic constraints and physical reality, and maintains the stability of the closed-loop control system.

[0016] 2. Establish a dynamic confidence arbitration mechanism based on multidimensional residual distribution characteristics to decouple sensor failure from equipment physical faults. Calculate the innovation vector between measured data and prior predicted states in real time, and adjust the weight gain of each sensing channel in the state fusion stage accordingly. When a specific channel residual changes abruptly and does not conform to the physical correlation between multiple variables, automatically suppress the corresponding weight and shield the influence of false error signals on the posterior state estimation. Isolate local anomalies in the sensing link without cutting off the control loop, avoiding the traditional control logic from triggering defensive shutdowns of the entire system due to disturbances of a single sensor, and ensuring continuous operation of the automated production line in harsh environments.

[0017] 3. An active micro-disturbance detection mechanism for steady-state conditions is introduced to solve the problem of decreased observability of system parameters when the input command of the passive observer is constant for a long time. When the system is in a steady-state range with constant speed or rotation speed, a micro-excitation signal below the mechanical action dead zone is superimposed on the command to extract the system's transient impedance response characteristics to the excitation. This directly reflects the real-time coupling properties between the cutting component and the load object, enabling the observer to identify the wear of the cutting teeth or the drift of surface physical parameters due to sudden changes in coal and rock hardness even when the overall action has not changed. This improves the steady-state active flaw detection capability of the control system and eliminates the ambiguity of state estimation and response lag caused by the steady-state dead zone. Attached Figure Description

[0018] Figure 1 This is a flowchart of the multidimensional state perception and closed-loop control of the physical constraint neurodynamics of the present invention. Figure 2 This is a comparison curve of the state estimation error of different sensing strategies under the combined interference conditions of the present invention; Figure 3 This is a schematic diagram of the system hardware deployment architecture and data interaction topology for cloud-edge collaboration according to the present invention. Detailed Implementation

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

[0020] This invention provides a method for multi-dimensional operational status perception in a dark mine based on deep neural networks. The method comprises an industrial automatic control system with multi-dimensional sensing units and a closed-loop controller, deployed at an edge computing node in the longwall mining face, such as an intrinsically safe industrial control computer. This method collects real-time control command data from mining equipment. and real-time observation data from the multi-dimensional sensing unit. Among them, real-time control command data Including but not limited to the traction speed command given by the frequency converter, the torque setpoint of the cutting motor, and the action command of the height adjustment cylinder, and real-time observation data. The system contains two types of signals: the first type is one-dimensional time-series data reflecting the physical properties of the equipment, such as motor current, stator temperature, hydraulic system pressure, and machine body vibration acceleration; the second type is visual feature data reflecting the spatial characteristics of the equipment, such as video stream data from explosion-proof cameras or infrared thermal imagers. For the visual feature data, the system converts it into feature vectors of the same dimension as the one-dimensional time-series data. The system then uses a pre-trained convolutional neural network encoder to... Video frame mapping as The low-dimensional pose feature vector is obtained, and its Euclidean distance with a preset standard pose reference is calculated to generate a residual component representing visual channel consistency. This residual component is then combined with the algebraic difference of one-dimensional temporal data to construct a unified multi-dimensional residual vector. To implement a hard real-time spatiotemporal alignment procedure for heterogeneous data: configure the FPGA front-end to synchronize the clocks of all network acquisition nodes based on the IEEE 1588PTP protocol, and label all data packets with nanosecond-level global timestamps; construct a dual-port RAM ring buffer on the edge computing node with a high-frequency one-dimensional time-series data sampling rate. Focusing on low-frequency visual feature data Perform zero-order hold interpolation synchronization; set the maximum allowable delay jitter window. Any data frames whose timestamps fall outside the window are discarded directly. Only the aligned synchronization feature vector sequence is written to the neural network input layer register, preventing phase aliasing of the state observer caused by asynchronous sampling. This method uses the posterior state estimate from the previous time step. and real-time control command data The input is fed into a pre-defined neurodynamic prediction model, which is built on a recurrent neural network architecture and is used to generate a prediction of the prior state at the current time step. This neurodynamic prediction model is constructed based on the nonlinear dynamic equations of mining equipment, and its loss function... Includes physical constraint terms The physical constraint term Simplified nonlinear dynamic equations based on the coal mining machine traction unit The physical constraint term is constructed and implemented during model training and real-time inference. Limited prior state prediction values The evolution trajectory satisfies energy conservation and kinematic continuity boundaries. If the model predicts the rate of change of velocity... If the predicted value exceeds the maximum acceleration boundary of the mechanical system, the constraint layer will project the predicted value back to the feasible region boundary to ensure that the predicted state conforms to physical reality.

[0021] After obtaining the predicted values, the system calculates the real-time observation data. Compared with the predicted value of the prior state The multidimensional residual vector between And use channel-weighted logic to analyze multidimensional residual vectors Based on the statistical distribution characteristics within each sensing channel, dynamic reliability weights are generated for each sensing channel. For any one The perception channel and its corresponding dynamic reliability weight Calculate according to the following formula: ,in, For the first Dynamic reliability weights of the perception channel For the corresponding number in the multidimensional residual vector The residual modulus of the sensing channel, The preset residual tolerance threshold, This is a preset sensitivity coefficient used to adjust the weight decay rate; the formula is used to adjust the residual modulus. Exceeding the residual tolerance threshold Suppress the channel weight at the time Secondly, for each sensing channel, the system calculates the statistical variance of its corresponding residual component within a preset sliding time window and compares this statistical variance with a preset reference noise amplitude. When the statistical variance exceeds the reference noise amplitude, the system uses the aforementioned nonlinear attenuation strategy to perform attenuation calculations on the weight value corresponding to the sensing channel in the dynamic reliability weight. This strategy can isolate sensing channel data that has undergone abrupt changes and does not conform to physical correlation while maintaining the control loop closure. Register-level physical blocking logic is embedded: real-time comparison of dynamic reliability weights. Hardware cutoff threshold ;when When this occurs, a low-level memory protection interrupt is triggered, and a channel mask is written to the relevant configuration register of the direct memory access controller (DMA), forcibly suspending the data transfer instruction for the corresponding sensing address. In the digital signal processor (DSP) pipeline, the input feature component corresponding to the failed channel is replaced with an all-zero vector or a no-operation instruction (NOP) is written, physically cutting off the electrical link for high-amplitude noise to propagate to subsequent neurons, preventing numerical overflow or gradient explosion during the forward inference process of the computation graph.

[0022] Based on the above dynamic credibility weight The system predicts the value of the prior state. With real-time observation data A weighted fusion computation is performed to generate the posterior state estimate for the current time step. The posterior state estimate The input is fed into the closed-loop controller, which calculates the posterior state estimate. The method also includes a steady-state parameter identification step based on active excitation. The system monitors the time series variance of the posterior state estimate in real time. When the time series variance is lower than a preset activity threshold, such as the rated fluctuation, the system will determine the deviation from the preset safe operating boundary and generate equipment control commands to drive the mining equipment to perform corresponding adjustment actions. When this occurs, it indicates that the system has entered the steady-state dead zone, at which point the system is in real-time control command data. A micro-perturbation excitation signal with a preset frequency and amplitude is superimposed on the system. The amplitude of the micro-perturbation excitation signal is set below the dead zone threshold required to drive the mining equipment to produce effective mechanical action, ensuring that normal operation is not affected. The system acquires the transient response characteristic data of the mining equipment to the micro-perturbation excitation signal and converts this transient response characteristic data into dynamic load impedance parameters. These dynamic load impedance parameters are input as constraints into the neurodynamic prediction model, which then uses these dynamic load impedance parameters to correct the prior state prediction values. This eliminates ambiguities in steady-state state estimation. Furthermore, to address model mismatch issues caused by equipment temperature rise, this method includes a temperature-monitored thermal drift compensation step. The system acquires real-time temperature data of the mining equipment's actuators and converts this data into physical constraint terms using a pre-set temperature rise parameter compensation model. The dynamic boundary parameter, which is used in the temperature rise parameter compensation model, characterizes the physical property parameters of mining equipment, such as the evolution of damping coefficient and friction coefficient with temperature. During the operation of the neurodynamic prediction model, the system uses this dynamic boundary parameter to adjust the physical constraint terms in real time. The constraint tolerance range is used to compensate for parameter drift caused by thermo-mechanical coupling effects.

[0023] The neurodynamic prediction model employs a two-layer gated recurrent unit network architecture. The input layer receives the posterior state estimation vector from the previous time step and real-time control command data. This information is then sequentially input into two hidden layers, each configured with... The model employs a hyperbolic tangent activation function and includes residual connections across gated recurrent unit layers. These connections add a linearly weighted combination of the input vectors to the output of the first hidden layer to enhance gradient stability and information flow. The final output layer is a fully connected layer that uses a linear activation function to map the hidden state to the current prior state prediction vector. The rule redundancy control strategy uses two quantitative indicators as switching conditions: one is the overall confidence level, i.e., when the dynamic confidence weights of all perceptual channels are... average continuous The control cycle is lower than The first trigger is a handover, and the second is when the physical security boundary is exceeded, i.e., when the posterior state estimate is... With respect to the preset safety limits of the equipment's mechanical system between margin ratio continuous The control cycle is lower than In such cases, the system executes an emergency stop command. Furthermore, to ensure absolute system safety under extreme anomalies, this embodiment also runs a rule-based redundancy control strategy in parallel, with the system monitoring the dynamic reliability weights of all sensing channels in real time. The overall mean, when the overall confidence level is below a preset safety threshold (e.g. This means that most sensors are unreliable or the data is severely conflicting, or when the posterior state estimate... Indicator device is about to cross physical safety boundary, such as when the tilt angle exceeds [a certain value]. When the system reaches a certain threshold, it immediately blocks the output of the neural network and directly switches to a deterministic boundary protection procedure, such as performing an emergency stop or interlocking operation, taking over the generation of control commands to meet the rigid requirements of industrial control systems for fault-oriented safety. For the problem of operating condition migration caused by changes in geological conditions, such as moving from soft coal seams to hard rock layers, this method includes a physical mechanism-based operating condition adaptation step. When the overall magnitude of the multidimensional residual vector continuously exceeds a certain threshold... When the time exceeds the operating condition change threshold, the system triggers an online identification mechanism, using the most recent time window, such as... The input and output data in seconds are processed using the recursive least squares (RLS) method based on a simplified single-degree-of-freedom physical model. Quickly estimate the current equivalent stiffness and equivalent damping Based on the changes in these key physical property parameters, a linear gain adjustment factor is generated and directly applied to the output layer of the neurodynamic prediction model to scale and correct its output amplitude.

[0024] Example 1: In the actual mining process of a fully mechanized longwall face in a mine with a capacity of tens of millions of tons, when the coal mining machine advances at a constant traction speed to a geological area containing hard interbedded rock, the cutting resistance increases sharply, causing the machine body to generate severe non-periodic vibrations. At this time, the vibration acceleration sensor installed on the machine body outputs a signal with a clipping distortion, and the high-pressure spray system, which is automatically activated to suppress cutting dust, causes a large amount of water mist and mud to splash, partially obstructing the lens of the explosion-proof camera and resulting in a large loss of effective data from the visual observation channel. Faced with this extreme condition of dual perception failure, traditional threshold-based control systems are often forced to trigger defensive shutdowns due to the inability to obtain reliable state feedback, leading to production interruptions. This example uses the neurodynamic state perception system constructed in the aforementioned specific implementation to address the above challenges. The system utilizes a neurodynamic prediction model with embedded physical constraints to estimate the posterior state value of the previous moment. and the control commands at the current moment As input, generate prior state predictions. Even with real-time observation data The vibration component is distorted due to clipping, and the visual component is abnormal due to occlusion; physical constraint terms. Still limited to the predicted state The system evolves along a trajectory that conforms to the inertial characteristics of the coal mining machine and the law of conservation of energy, filling the perception blind spots caused by sensor failure.

[0025] The system then calculates the real-time observation data. Compared with the predicted value of the prior state The multidimensional residual vector between Because the visual channel is blocked, its corresponding residual magnitude is... The increase is inconsistent with physical correlation, and the channel weighting logic is then based on the formula. Dynamic credibility weights of the visual channel The data is attenuated to near zero, thereby automatically isolating the data from the failed channel during state fusion and preventing spurious error signals from influencing the posterior state estimate. This mechanism decouples sensor failure from equipment physical malfunctions. Addressing the steady-state dead zone problem in hard rock cutting, when the system detects that the time series variance of the posterior state estimate is below the activity threshold, it actively superimposes a micro-perturbation excitation signal with an amplitude below the mechanical action dead zone threshold into the control command. For example... The system captures the transient response characteristics of the equipment to the micro-disturbance using the PRBS signal at rated speed, and calculates the current dynamic load impedance parameters. These impedance parameters directly characterize the true physical state of the contact between the cutting teeth and hard rock, eliminating the ambiguity that the steady-state current value alone cannot distinguish between constant load and slippage. This method reconstructs the true operating state of the coal mining machine under hard rock cutting without interrupting production, and drives the closed-loop controller to dynamically adjust the cutting height and traction speed, achieving adaptive control for complex geological conditions.

[0026] Example 2: This example objectively verifies the effectiveness and stability of the aforementioned deep neural network-based multi-dimensional operation status perception method for dark mines under real and complex working conditions through rigorous comparative experiments. The test environment is built on a mine fully mechanized mining simulation platform with high-fidelity physical simulation capabilities. It integrates dynamic entity models of heavy coal mining machines and hydraulic support groups, as well as multi-dimensional perception simulation units, and is equipped with a programmable environmental disturbance generator to reproduce extreme working conditions such as strong underground vibration, water mist obstruction, and sudden changes in geological conditions.

[0027] The experimental design included three parallel control groups: Control group A adopted a common industrial logic judgment strategy based on a fixed threshold, which lacked multi-source information fusion and physical constraint mechanisms; Control group B adopted the same neurodynamic prediction model architecture as this invention, but removed the physical constraint term from the loss function. Furthermore, no channel attention mechanism was introduced; the experimental group fully adopted the embedded physical constraint neurodynamic prediction model, channel attention mechanism, and active incentive strategy of this invention; the experimental process simulated the coal mining machine in... to The continuous cutting operation within the time window is interrupted by two typical disturbance conditions introduced sequentially. to The simulated strong vibration caused clipping distortion in the accelerometer signal, with the amplitude being truncated. Simultaneously, the simulated high-pressure spray caused a visual channel occlusion rate of 100%. ;exist to The simulated coal mining machine in the interval is in a constant-speed traction steady state, but the coal seam hardness changes from... Mutation The system runs in real time, recording all data throughout the process, including raw observation data, prior prediction states, residual vector magnitudes, and final posterior state estimates.

[0028] Table 1: Key Data Recording Table for Multidimensional Perception and State Reconstruction

[0029] Table 2: Comparison of Key Performance Indicators under Different Sensing Strategies

[0030] Data shows that under condition one (composite sensing failure), the state estimation of the experimental group... Only Furthermore, no accidental shutdowns occurred, unlike control group B. Gundam This indicates the physical constraint terms. The evolution trajectory of the forced-constraint state prediction value effectively fills the perception blind spot. Under condition two (steady-state dead zone), the experimental group actively superimposed the data. The micro-perturbation excitation signal at rated speed, The internal calculation of the dynamic load impedance change enables rapid adaptive identification of the interbedded rock condition.

[0031] Example 3: This example combines Figures 1 to 3 This paper describes a method for multi-dimensional operational status perception in a dark mine based on deep neural networks, such as... Figure 1As shown, the system acquires real-time control command data of mining equipment and real-time observation data, including one-dimensional time-series data reflecting physical properties and visual feature data reflecting spatial characteristics, through a data acquisition module. These data are fed into the multi-dimensional sensing unit data stream to provide a real-time observation benchmark for residual calculation and fusion. Simultaneously, the real-time control command data is input to a neurodynamic prediction model with embedded energy conservation and motion continuity physical constraints to generate prior state prediction values. The system uses channel weighted logic to calculate the multi-dimensional residual vector between the real-time observation data and the prior state prediction values, and analyzes its statistical distribution characteristics to generate dynamic credibility weights for each sensing channel. Finally, the system performs weighted fusion calculation, fusing the prior state prediction values ​​and real-time observation data based on the dynamic credibility weights to generate posterior state estimates, which are then input to the closed-loop controller. The closed-loop controller calculates the deviation between the posterior state estimates and the safe operating boundary and generates equipment control commands.

[0032] like Figure 2 As shown in the figure, the root mean square error (RMSE) curves of the method of the present invention, the unconstrained neural network method, and the traditional thresholding method are compared within a time window of 0 to 60 seconds. The horizontal axis represents time (s), and the vertical axis represents RMSE%. The figure shows that in the intervals of 15 to 25 seconds and 36 to 45 seconds, the dashed line representing the unconstrained neural network method and the dotted line representing the traditional thresholding method both show error peaks, while the solid line representing the method of the present invention maintains a low and stable error level during these periods, indicating that the physical constraint mechanism limits the fluctuations in state estimation. Figure 3 As shown, the system uses an intrinsically safe industrial control computer as its core, which connects to the equipment group at the fully mechanized mining face. This equipment group includes vibration sensors, current / temperature probes, explosion-proof cameras, traction frequency converters, and cutting actuators. It also transmits multi-dimensional data streams to the edge computing center. The edge computing center is equipped with a neurodynamic prediction model, channel credibility weighted logic, and redundant safety closed-loop control modules. The control commands generated by the model are directly fed back to the control equipment group. At the same time, the edge computing center maintains a connection with the ground data center. The ground data center is responsible for offline training and parameter updates of the model using historical data, and then sends the updated model parameters to the edge computing center to complete the online update of the neurodynamic prediction model.

[0033] Example 4: This example clarifies the channel attention weight attenuation strategy used to isolate failed sensors in the aforementioned multi-dimensional operational status perception method, as well as the objective determination procedure for key parameters. When a sensing channel experiences a non-physical abrupt change, how to adaptively adjust its weight in state fusion based on quantitative logic rather than empirical judgment is demonstrated. This method constructs key parameters by collecting historical residual data of the device under normal operating conditions. (Sensitivity coefficient) and An adaptive calibration procedure for the residual tolerance threshold is used. The system continuously monitors the residual samples of each sensing channel during stable operation and calculates the standard deviation of the residual sequence. Residual tolerance threshold Set as This setting is based on statistics. The principle is to ensure that any residual fluctuations used to trigger the attenuation mechanism exceed the normal system noise range. .

[0034] Based on this, the sensitivity coefficient The value of depends on the system's response speed requirements to fault events. To balance the technical trade-off between fault isolation speed and false trigger rate, this invention determines by setting control theory boundary conditions. The value of the boundary condition is specified by the condition that when the residual modulus is... achieve At that time, the corresponding dynamic weight Decay to within the sampling period The following section substitutes this boundary condition into the channel weight calculation formula. In this way, the minimum value required to satisfy the response characteristic can be calculated. Secondly, regarding the calculation and processing flow of multidimensional residual vectors, this embodiment further discloses the specific algorithm path for residual confidence assessment. The system does not simply calculate the Euclidean distance, but introduces a temporal consistency verification step, calculating the current time step. instantaneous residual Then, the system will backtrack to the most recent The residual history at each time step is calculated. The deviation between the magnitude of the change and its historical trend is considered. When the sudden change in the instantaneous residual exceeds the preset dynamic margin, the residual is identified as a signal with an abnormal trend and input to the weight attenuation logic. This logic effectively filters out misjudgments caused by transient spike noise from the sensor, ensuring the smoothness of weight adjustment.

[0035] Example 5: Before deploying this sensing method on new equipment or in a fully mechanized mining face after major overhaul, an offline calibration procedure must be performed to construct the core parameters and dynamic boundary mapping required for the physical constraint model. The system is tested on a no-load test bench at a constant ambient temperature. Next, step response and frequency sweep tests were conducted on the traction system of the coal mining machine, and control inputs were collected. and the resulting state of motion The data is used to identify and determine the equivalent mass of the traction unit through a least squares system. Equivalent damping and equivalent stiffness The initial values ​​are used to initialize the physical constraints in the neurodynamic prediction model, and the system is subjected to different temperature gradients. to Repeated load tests were conducted, and temperature data and mechanical parameter evolution data were recorded simultaneously. A nonlinear mapping function describing the changes of physical properties such as damping coefficient and friction coefficient with temperature was generated. This nonlinear mapping function is the temperature rise parameter compensation model, which is used to adjust the boundary tolerance of physical constraint terms online.

[0036] When a critical sensor is first connected to or replaced in the system, a standardized field calibration procedure is performed to determine critical operational thresholds. This procedure includes continuously collecting data while the equipment is running unloaded and in steady state at its rated traction speed. Observe data in real time, calculate the statistical variance of the signal residuals of each sensing channel, and then calculate the variance. A multiple as the residual tolerance threshold in the channel attention mechanism The initial value is used to determine the precise amplitude of the active excitation signal. The system inputs an amplitude to the traction motor from... Starting from rated torque, Gradient-increasing excitation signals are recorded and used to determine the motion state of the equipment. The variance exceeded for the first time The minimum excitation amplitude, which is defined as the dead zone threshold of the active excitation. The amplitude of the actual micro-perturbation excitation signal is set to This allows micro-perturbations to maintain the observability of system parameters without causing overall mechanical movement.

[0037] Example 6: The system executes the neurodynamic prediction model construction procedure, through embedded physical constraint terms. Forced guarantee of the state prediction values ​​output by the model To satisfy the device's dynamic characteristics, the training process of the neurodynamic prediction model employs minimizing the total loss function. The strategy is that the total loss function is a data-driven standard loss term. Physical consistency penalty item It is composed of weighted combinations, including a physical consistency penalty term. Simplified nonlinear dynamic equations based on the coal mining machine traction unit Build, in For equivalent quality, For equivalent damping, For equivalent stiffness, , and These are acceleration, velocity, and displacement, respectively. For real-time control commands The traction of the decision, For load reaction force, physical consistency penalty term Calculated as the state predicted by the model The mean square error of the residuals obtained after substituting into the simplified dynamic equation is used to quantify the degree of deviation between the model's predicted values ​​and the physical laws, and the total loss function. The calculation relationship is expressed as follows ,in This is a weighting factor for the physical constraint term, used to adjust the degree to which physical constraints enforce model training.

[0038] The model training procedure incorporates weight factors. The system optimization calibration, the optimization method is: in the preset... On a training dataset containing partially failed sensors, a grid search method was used to... exist to Scope The gradient is iterated to determine the state estimation bias of the model on the failure samples. Minimum and on normal samples Increase not exceeding Best The system will determine the weighting factor by taking the value. After substituting the total loss function, the recurrent neural network architecture is trained offline, and an early stopping mechanism is employed, i.e., when the model's loss value on the dedicated validation dataset is continuously... The training cycle failed to reduce by more than Training terminates at this point, ensuring that the model generates state predictions. While meeting the data fitting requirements, the device dynamics constraints must be followed.

[0039] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for multi-dimensional operational status perception in a dark mine based on deep neural networks, applied to an industrial automatic control system including a multi-dimensional sensing unit and a closed-loop controller, characterized in that... The method includes the following steps: The system collects real-time control command data from mining equipment and real-time observation data from multi-dimensional sensing units, including one-dimensional time-series data reflecting the physical properties of the equipment and visual feature data reflecting the spatial characteristics of the equipment. The posterior state estimate from the previous moment and the real-time control command data are input into a pre-set neurodynamic prediction model. The neurodynamic prediction model is used to generate the prior state prediction for the current moment. The neurodynamic prediction model is constructed based on the nonlinear dynamic equations of the mining equipment, and its loss function includes physical constraint terms. The physical constraint terms force the evolution trajectory of the prior state prediction to satisfy energy conservation and kinematic continuity boundaries during the model training and inference stages. Calculate the multidimensional residual vector between real-time observation data and prior state prediction values, and use channel weighted logic to analyze the statistical distribution characteristics of the multidimensional residual vector in each sensing channel, thereby generating dynamic confidence weights for each sensing channel. Based on dynamic confidence weights, a weighted fusion calculation is performed on the prior state prediction value and the real-time observation data to generate the posterior state estimate value at the current moment. The posterior state estimate is input to the closed-loop controller, which calculates the deviation between the posterior state estimate and the preset safe operating boundary, and generates equipment control commands based on the deviation.

2. The method for multi-dimensional operational status perception in a dark mine based on a deep neural network according to claim 1, characterized in that, The method also includes a steady-state parameter identification step based on active excitation: real-time monitoring of the time series variance of the posterior state estimate; when the time series variance is lower than a preset activity threshold, superimposing a micro-perturbation excitation signal with a preset frequency and amplitude into the real-time control command data; setting the amplitude of the micro-perturbation excitation signal to be lower than the dead zone threshold for driving the mining equipment to produce effective mechanical action; acquiring the transient response characteristic data of the mining equipment to the micro-perturbation excitation signal, and converting the transient response characteristic data into dynamic load impedance parameters; The dynamic load impedance parameter is input as a constraint into the neurodynamic prediction model, and the neurodynamic prediction model is used to correct the prior state prediction value based on the dynamic load impedance parameter.

3. The method for multi-dimensional operational status perception in a dark mine based on a deep neural network according to claim 1, characterized in that, The method also includes a thermal drift compensation step based on temperature monitoring: acquiring real-time temperature data of the actuator of the mining equipment; converting the real-time temperature data into dynamic boundary parameters of physical constraint terms using a preset temperature rise parameter compensation model; the temperature rise parameter compensation model characterizes the evolution law of the physical property parameters of the mining equipment with temperature changes; and adjusting the constraint tolerance range of the physical constraint terms in real time using dynamic boundary parameters during the operation of the neurodynamic prediction model.

4. The method for multi-dimensional operational status perception in a dark mine based on a deep neural network according to claim 1, characterized in that, The steps for generating dynamic reliability weights using channel weighting logic include: for each sensing channel, calculating the statistical variance of its corresponding residual component within a preset sliding time window; comparing the statistical variance with a preset reference noise amplitude; when the statistical variance exceeds the reference noise amplitude, using a nonlinear attenuation strategy to perform attenuation operations on the weight value corresponding to the sensing channel in the dynamic reliability weights; the nonlinear attenuation strategy is configured to isolate sensing channel data that undergoes abrupt changes and does not conform to physical correlations, while maintaining the closure of the control loop.

5. The method for multi-dimensional operational status perception in a dark mine based on a deep neural network according to claim 1, characterized in that, The method also includes a physical mechanism-based working condition adaptation step: real-time monitoring of the overall modulus of the multidimensional residual vector; when the overall modulus continuously exceeds the preset working condition change threshold, obtaining the running data segment within the preset time window at the current moment, and using the least squares method to identify the key physical attribute parameters under the current working condition based on the simplified physical model; A linear gain adjustment factor is generated based on the changes in key physical property parameters, and the output layer of the neurodynamic prediction model is weighted and modified using the linear gain adjustment factor to adapt the response characteristics of the neurodynamic prediction model to the current geological conditions.

6. The method for multi-dimensional operational status perception in a dark mine based on a deep neural network according to claim 1, characterized in that, In the step of generating dynamic credibility weights, for any given... The perception channel and its corresponding dynamic reliability weight Calculate according to the following formula: ,in, For the first Dynamic reliability weights of the perception channel For the corresponding number in the multidimensional residual vector The residual modulus of the sensing channel, The preset residual tolerance threshold, This is a preset sensitivity coefficient used to adjust the weight decay rate; the formula is used to suppress the channel weight when the residual magnitude exceeds the residual tolerance threshold.

7. The method for multi-dimensional operational status perception in a dark mine based on a deep neural network according to claim 1, characterized in that, The method also includes an active calibration step for minute drifts: during the steady-state operation of the mining equipment, a test signal is generated and superimposed onto the selected sensing channel data; the response sensitivity characteristic data of the neurodynamic prediction model to the test signal is obtained; the sensor's inherent drift component is separated from the raw data of the sensing channel using the response sensitivity characteristic data, and a negative feedback compensation loop is constructed; the sensor's inherent drift component is used to perform online zero-point calibration on the subsequent real-time observation data of the sensing channel.

8. The method for multi-dimensional operational status perception in a dark mine based on a deep neural network according to claim 1, characterized in that, The method also includes a rule-based redundancy control step: running a set of boundary protection procedures based on deterministic logic rules in parallel; monitoring the overall confidence level of dynamic confidence weights in real time; when the overall confidence level is lower than a preset safety threshold or when the posterior state estimate indicates that the device state is close to the limit value of the safe operation boundary, blocking the output of the posterior state estimate and directly switching to the boundary protection procedure to take over the right to generate device control commands.

9. A method for multi-dimensional operational status perception in a dark mine based on a deep neural network according to claim 1, characterized in that, The calculation steps of the multidimensional residual vector include: converting visual feature data into feature vectors of the same dimension as one-dimensional time series data; calculating the Euclidean distance or cosine similarity between real-time observation data and prior state prediction values ​​in the feature vector space to generate residual components that characterize visual channel consistency; and combining the residual components with the algebraic difference of one-dimensional time series data to construct a multidimensional residual vector.

10. A method for multi-dimensional operational status perception in a dark mine based on a deep neural network according to claim 1, characterized in that, The method also includes an offline update step for model parameters: storing historical running data and corresponding fault markers on the server side; periodically training the neurodynamic prediction model using the historical running data, with the optimization objective being to minimize the long-term statistical mean of the multidimensional residual vector; and sending the updated model parameters to the edge controller to replace the current model parameters used to generate prior state prediction values.

Images