A multi-modal perception and decision optimization method for an intelligent mine lamp system of a coal mine

Abstract

This invention belongs to the field of artificial intelligence technology and relates to a multimodal perception and decision optimization method for a coal mine intelligent mine lamp system. The method includes: acquisition of multimodal monitoring data and construction of modal state values ​​for the mine lamp; multimodal reliability assessment and construction of a multi-objective cost function for mine lamp operation and maintenance conditions; a two-layer particle swarm optimization driven by reliability and operational complexity; and online adaptive operation and maintenance execution for the mine lamp. This invention quantifies the degree of sensor contamination and drift by utilizing median spatial reference, absolute deviation scale, and Thales-Sen slope estimation; constructs an environmental complexity index that integrates modal reliability fluctuations and video motion entropy to achieve accurate operational condition perception; employs a two-layer particle swarm optimization mechanism driven by reliability and operational complexity to adaptively adjust parameters, directionally perturb low-reliability dimensions, and ensure gas sensing sensitivity with safety constraints; and constructs a multi-objective cost function that coordinates safety risks, lighting power consumption, and communication performance to complete online adaptive operation and maintenance decision-making for the mine lamp system.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a multimodal perception and decision optimization method for a coal mine intelligent mine lamp system. Background Technology

[0002] Intelligent mine lamps in coal mines simultaneously perform multiple functions during underground operations, including lighting, environmental sensing, personnel status recognition, and wireless communication. Their integrated multimodal sensors, encompassing gas, temperature and humidity, dust, inertial measurement, video, and wireless communication, exhibit varying degradation characteristics across different modes due to factors such as coal dust adhesion, water vapor condensation, mechanical impact, battery degradation, and lens contamination. Furthermore, safety risk identification, lighting energy consumption control, and communication reliability scheduling are interdependent. Under complex and changing underground conditions, the sensing fusion strategy, alarm decision criteria, and lighting and communication resource allocation must be dynamically adjusted based on the reliability status of each mode and environmental complexity to achieve long-term online adaptive and efficient operation and maintenance of the mine lamps.

[0003] However, existing methods often employ fixed alarm baselines and fixed fusion weights, failing to consider the differential degradation of each modal sensor caused by factors such as coal dust adhesion, water vapor condensation, and impact drift. This makes them prone to false alarms or missed alarms when sensors are contaminated or drift. Current technologies also lack mechanisms to distinguish between changes in the actual underground environment and sensor malfunctions using spatial references from nearby mine lamps and long-term drift analysis, making it difficult to dynamically quantify the reliability of modal state values. Existing optimization algorithms typically use fixed inertial weights and uniform update speeds, failing to dynamically adjust search behavior according to the complexity of underground conditions. This can lead to insufficient global exploration and getting trapped in local optima under complex conditions, or slow convergence and over-search under stable conditions. Furthermore, existing technologies do not incorporate safety risks, lighting energy consumption, and communication reliability into multi-objective collaborative optimization, and do not perform targeted corrections to relevant decision variables based on modal reliability. This results in optimization results that are difficult to adapt to changes in sensor states, affecting the robustness of online mine lamp operation and maintenance.

[0004] Therefore, there is an urgent need to develop a multimodal perception and decision optimization method for intelligent mine lamp systems in coal mines to effectively solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a multimodal perception and decision optimization method for intelligent mine lamp systems in coal mines, in order to solve the problems of existing methods, such as fixed baselines, weights, and parameters, failure to take into account sensor operating condition degradation and dynamic quantification of state reliability, poor algorithm adaptability, lack of multi-objective collaborative optimization and decision correction, easy occurrence of alarm deviations and optimization defects, and insufficient overall operational robustness.

[0006] This invention is achieved through the following technical solution:

[0007] A multimodal perception and decision optimization method for a coal mine intelligent mine lamp system includes the following steps:

[0008] S1. Data acquisition and modal state value construction for multimodal monitoring of mine lamps:

[0009] Collect multimodal monitoring data of the target mine lamp, and convert sensor outputs of different dimensions and data forms into modal state values ​​for subsequent reliability assessment and optimization calculation;

[0010] S2. Multimodal reliability assessment and multi-objective cost function construction for mine lamp operation and maintenance:

[0011] By utilizing the spatial consistency of adjacent miner lamps, the long-term drift characteristics of the target miner lamp, and the short-term operating condition fluctuation characteristics, modal reliability, environmental complexity indicators, and multi-objective cost functions are constructed, enabling the subsequent optimization process to distinguish between sensor degradation and changes in the actual underground environment.

[0012] S3. Two-layer particle swarm optimization driven by credibility and operational complexity:

[0013] By embedding the environmental complexity index into the inertia weight adjustment process and the modal credibility into the velocity update process, the particle swarm can enhance global search under complex conditions, enhance local convergence under stable conditions, and perform targeted perturbation on decision variables related to low credibility modes.

[0014] S4, Mining Lamp Online Adaptive Operation and Maintenance Execution:

[0015] After completing modal reliability assessment and particle swarm optimization, the system periodically updates the operation and maintenance strategy according to the same data acquisition, reliability assessment and constraint correction rules as in the optimization phase, and the target mine lamp directly applies the optimal decision vector to the underground online operation.

[0016] Further, step S1 specifically includes the following steps:

[0017] S11. Arrange gas sensors, temperature and humidity sensors, dust sensors, inertial measurement sensors, video sensors, light-emitting diode lighting drive sampling circuits and wireless communication status sampling units on the target mine lamp, and collect raw mine lamp multimodal monitoring data according to a fixed sampling period.

[0018] S12. Multimodal monitoring data of the original mine lamp Perform time alignment and validity checks to obtain multimodal monitoring data of synchronized mining lamps;

[0019] S13. Convert the multimodal monitoring data of the synchronous mine lamp into modal state values ​​to obtain the first... Within the sampling period, the first Modal state values ​​of each mode ;in, This represents the modal index, with values ​​ranging from 1 to 2. , This represents the total number of modes involved in the credibility assessment and optimization calculation. It refers to the first Within the sampling period, the first The scalar state values ​​obtained after preprocessing each mode;

[0020] S14. Arrange the modal state values ​​within the continuous sampling period into a target miner lamp modal state sequence in chronological order. , This refers to a data sequence consisting of all modal state values ​​across multiple sampling periods, used for subsequent calculations of modal reliability, environmental complexity indices, and multi-objective cost functions. It is a data sequence with dimensions of [missing information - likely a typology or specific dimension]. The matrix; where the first... The row corresponds to the first The sampling period, the first Column corresponding to the first One modality.

[0021] Further, step S2 specifically includes the following steps:

[0022] S21. Modal confidence calculation based on spatial consistency and long-term drift characteristics: Combine the spatial reference value of the neighboring miner lamp with the long-term drift of the target miner lamp to obtain the confidence of each mode, so as to quantify the confidence of each mode in the target miner lamp not being affected by contamination, drift or sensitivity degradation in the current sampling period.

[0023] S22. Calculation of environmental complexity index based on credible modal fluctuations and video action entropy: The environmental complexity index is obtained by weighting multimodal short-term fluctuations by modal credibility and combining them with video action entropy.

[0024] S23. Construction of a multi-objective cost function for security risks, lighting power consumption and communication vulnerability: The multi-modal fusion weights, alarm baseline values, sensitivity index, LED channel drive current and wireless communication transmission power are uniformly encoded into a decision vector, and a multi-objective cost function is constructed.

[0025] Furthermore, step S21 specifically includes the following steps:

[0026] S211, in the Within each sampling period, the target miner lamp obtains the identification of neighboring miner lamps, received signal strength, communication distance, and same-mode state value through wireless broadcasting or underground gateway, and establishes a set of neighboring miner lamps based on the communication distance and received signal strength;

[0027] S212, Collect from nearby miners' lamps Extract the first The modal state values ​​of the neighboring miner lamps for the mode are calculated, and the ... Within the sampling period, the first Spatial reference values ​​for each mode and spatial discrete scale ;

[0028] S213, Calculate the target miner's lamp number within a long-term sliding window. The relative deviation sequence of the nth mode relative to the spatial reference value, and the estimation of the nth mode based on the relative deviation sequence. Within the sampling period, the first Long-term drift of each mode ;

[0029] S214, According to the target miner's lamp Calculation of spatial deviation and long-term drift of the first mode. Within the sampling period, the first Modal reliability of each mode ;

[0030] S215, the first The modal confidence scores within each sampling period are combined in modal index order to obtain the modal confidence score vector. .

[0031] Furthermore, step S22 specifically includes the following steps:

[0032] S221, with the first The sampling period is the current sampling period, and the sampling is extracted within the short-time sliding window. The historical modal state values ​​of the mode are calculated, and the ... Within the sampling period, the first Short-time mean of each modality and short-term standard deviation ;

[0033] S222. Based on the deviation of the current modal state value from the short-time mean, and in conjunction with the modal reliability, calculate the first... Multimodal reliable fluctuation intensity within each sampling period ;

[0034] S223. Extract the person's action state from the video frame data and calculate the first... Video motion entropy within a sampling period ;

[0035] S224, Combining Multimodal Reliable Fluctuation Intensity and video motion entropy , obtained the Environmental complexity index within each sampling period .

[0036] Furthermore, step S23 specifically includes the following steps:

[0037] S231, Constructing the first Decision vector within each sampling period ;

[0038] S232, according to the... Modal state values Alarm baseline value and sensitivity index Calculate the first Risk response for each modality, combined with modality fusion weights and modal credibility Obtain the credible risk impact value;

[0039] S233. Calculate the lighting power consumption based on the driving current of each LED channel.

[0040] S234. Based on wireless communication transmission power Calculate the communication vulnerability item based on the current link state;

[0041] S235. The trusted security risk item, lighting power consumption item, and communication vulnerability item are weighted and summed to obtain the... Multi-objective cost function within each sampling period .

[0042] Further, step S3 specifically includes the following steps:

[0043] S31, Particle coding and constraint-aware initialization;

[0044] An initial population satisfying boundary constraints is generated using chaotic sequences, and the initialization coverage is expanded to include low-confidence mode-related dimensions by combining modal confidence.

[0045] S32. Heterogeneous velocity update that integrates environmental complexity and modal credibility: Differentiated velocity update is achieved by driving Cauchy mutation through adaptive inertia weights and credibility.

[0046] S33, Constraint Projection, Fitness Update and Optimal Decision Output: Boundary truncation, weight normalization and safety retention constraint correction are performed on different types of decision variables to ensure that the optimization results can be directly deployed to the mining lamp embedded system.

[0047] Furthermore, step S31 specifically includes the following steps:

[0048] S311, Decision vector As a particle position vector, and maintaining a fixed encoding order;

[0049] S312. Set physical feasible boundaries for each decision dimension and establish constraint rules according to variable types;

[0050] S313. Generate a Tent chaotic sequence value for each particle and map the Tent chaotic sequence value to the physical feasible boundary of each decision dimension.

[0051] S314. Perform constraint correction on the initial particle positions to ensure that each initial particle satisfies the deployable conditions.

[0052] S315. Divide the particle swarm into a perception strategy subgroup and a resource scheduling subgroup, and make the two subgroups share the group's optimal position.

[0053] Furthermore, step S32 specifically includes the following steps:

[0054] S321. Based on environmental complexity indicators Calculate the first Within the sampling period, the first Adaptive inertia weights of the generation ;

[0055] S322. Establish the mapping relationship between decision dimensions and modalities. , used to determine the first Does each decision dimension correspond to a certain modality?

[0056] S323. Perform velocity updates for each particle and each decision dimension to obtain the... Sub-particle velocity;

[0057] S324. Based on modal mapping Control the effectiveness of confidence-driven variants;

[0058] S325. Limit the updated speed to obtain a controllable speed;

[0059] Step S33 specifically includes the following steps:

[0060] S331. Calculate the temporary position of the particle based on the updated velocity;

[0061] S332. Perform boundary truncation and weight normalization on the temporary locations to obtain preliminary feasible locations;

[0062] S333, Perform safety retention constraint corrections on gas safety-related modes;

[0063] S334. Utilizing a multi-objective cost function Calculate the fitness value for each particle;

[0064] S335. Update the individual's historical best position and the group's best position based on the fitness value;

[0065] S336. Determine whether the iteration termination condition is met, and output the iteration term. The optimal decision vector corresponding to each sampling period .

[0066] Furthermore, step S4 specifically includes the following steps:

[0067] S41, the target miner's lamp is in Raw mine lamp multimodal monitoring data were collected in one sampling period. And obtain the modal state values ​​according to step S1. ;

[0068] S42, Target miner's lamp is based on the set of nearby miner's lamps. Calculate the modal confidence vector And calculate the environment complexity index based on the short-time sliding window. ;

[0069] S43. Construct a multi-objective cost function for the current sampling period using the target miner's lamp. Then, step S3 is executed to obtain the optimal decision vector. ;

[0070] S44, The target miner's lamp is determined according to the optimal decision vector. Update the modal fusion weights, alarm baseline values, sensitivity index, LED channel drive current, and wireless communication transmit power;

[0071] S45. The target miner's lamp performs safety hard threshold judgment and continuous confirmation judgment, and outputs alarm result or low confidence prompt.

[0072] S46, the target miner's lamp will be the first Modal state values ​​and modal confidence vector for each sampling period Environmental complexity index Optimal decision vector The alarm results are written to the operation and maintenance log and then proceed to the next step. One sampling period.

[0073] Compared with the prior art, the beneficial effects of the present invention are:

[0074] 1. This invention proposes a multimodal reliability assessment method based on the spatial consistency and long-term drift characteristics of adjacent miners' lamps. It quantifies the degree of contamination and drift influence on each mode through median spatial reference, median absolute deviation scale, and Thiel-Sen slope estimation.

[0075] 2. This invention constructs an environmental complexity index that integrates multimodal reliable fluctuations and video action entropy. It uses reliability-weighted short-term deviation to reflect the dynamics on the sensor side and action entropy to reflect the degree of disorder on the personnel behavior side, providing a basis for working condition perception for optimization algorithms.

[0076] 3. This invention designs a two-layer particle swarm collaborative optimization mechanism driven by credibility and operating condition complexity. The environmental complexity adaptively adjusts the inertia weight, the modal credibility drives the Cauchy mutation directional perturbation of the low-credibility modal correlation dimension, and combines the safety retention constraint to ensure gas sensing sensitivity, thereby realizing the collaborative search of sensing strategy and resource scheduling.

[0077] 4. This invention establishes a multi-objective cost function that integrates unified coding mode fusion weights, alarm baseline values, sensitivity index, LED channel drive current, and wireless communication transmission power. It incorporates security risks, lighting power consumption, and communication vulnerability into the same optimization framework to achieve online adaptive operation and maintenance decision-making. Attached Figure Description

[0078] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0079] Figure 1 Plot showing the relative deviation sequence and long-term drift trend of gas concentration modes;

[0080] Figure 2 A short-time sliding window historical value graph of gas concentration modes;

[0081] Figure 3 This is a graph showing the magnitude of the deviation of each mode from the short-time mean in the current sampling period;

[0082] Figure 4 A graph showing how environmental complexity indicators change over time;

[0083] Figure 5 This is a distribution diagram of the modal fusion weights in the optimal decision vector for the 51st sampling period;

[0084] Figure 6 This is a flowchart illustrating the steps of the multimodal perception and decision optimization method for the intelligent mine lamp system of the present invention. Detailed Implementation

[0085] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0086] like Figure 6 As shown, the invention relates to a multimodal perception and decision optimization method for an intelligent mine lamp system, comprising the following steps:

[0087] S1. Data acquisition and modal state value construction for multimodal monitoring of mine lamps;

[0088] Intelligent mine lamps in coal mines simultaneously perform functions such as lighting, environmental sensing, personnel status recognition, and wireless communication during underground operations. However, long-term operation is affected by factors such as coal dust adhesion, water vapor condensation, mechanical impact, battery degradation, lens contamination, and wireless signal obstruction. This invention first collects multimodal monitoring data from the target mine lamp and then uniformly converts sensor outputs of different dimensions and data formats into modal state values ​​that can participate in subsequent reliability assessment and optimization calculations. The specific steps are as follows:

[0089] S11. Arrange gas sensors, temperature and humidity sensors, dust sensors, inertial measurement sensors, video sensors, LED lighting drive sampling circuits, and wireless communication status sampling units on the target mine lamp to collect raw mine lamp multimodal monitoring data according to a fixed sampling period.

[0090] Specifically, the first The raw mine lamp multimodal monitoring data collected within each sampling period are denoted as follows: , This refers to the target miner's lamp in the first... The combination of gas concentration data, temperature and humidity data, dust concentration data, inertial measurement data, video frame data, lighting drive data, and wireless communication status data collected within a sampling period; This represents the sampling period index, with a value of ; This indicates the total number of sampling periods, which can be determined based on the online runtime; the sampling interval can be from 1 second to 5 seconds, with 3 seconds being the preferred option.

[0091] Furthermore, the gas concentration data includes methane concentration, carbon monoxide concentration, and oxygen concentration; the temperature and humidity data includes ambient temperature and relative humidity; the dust concentration data includes total dust concentration; the inertial measurement data includes 3-axis acceleration, 3-axis angular velocity, and attitude angle; the video frame data includes continuous video frames of the working area in front of the target miner's lamp; the lighting drive data includes the drive current and forward voltage drop of each LED channel; and the wireless communication status data includes received signal strength, packet loss rate, retransmission count, co-channel interference power, and noise power.

[0092] In one embodiment, for example, in the first One sampling period, raw multimodal monitoring data of the target miner lamp collected. Specifically: methane concentration 0.42% VOL, carbon monoxide concentration 5 ppm, oxygen concentration 20.5% VOL; ambient temperature 28.3℃, relative humidity 85% RH; total dust concentration 2.6 mg / m³. 3 The 3-axis accelerations are (-0.05, 0.01, 0.96)g, the 3-axis angular velocities are (0.02, -0.01, 0.03)rad / s, and the attitude angles are (roll -2°, pitch 5°, yaw -15°); a 640×480 infrared video frame image at the current moment; the driving currents of the 3 LEDs are (160, 150, 0)mA, with corresponding forward voltage drops of (3.2, 3.2, 3.1)V; the wireless communication received signal strength is -68dBm, the packet loss rate is 2%, the retransmission count is 1, the co-channel interference power is -95dBm, and the noise power is -102dBm.

[0093] S12. Multimodal monitoring data of the original mine lamp Perform time alignment and validity checks to obtain multimodal monitoring data of synchronized mining lamps.

[0094] In practical implementation, the gas concentration data, temperature and humidity data, dust concentration data, inertial measurement data, lighting drive data, and wireless communication status data are aligned to the target miner's lamp's master control clock as a reference. The sampling period; for video frame data, in the first sampling period; Within each sampling period, the video frame closest to the sampling time is selected, or several frames within the sampling period are selected to form a short video clip. If a certain mode experiences communication loss, sensor self-test anomaly, or data exceeding the hardware range within the current sampling period, the mode is marked as an invalid mode, and its reliability is reduced in subsequent mode reliability calculations.

[0095] In one embodiment, for example, if the sampling period The starting time is The timestamp of the sensor data packet and If the deviations are all within the preset synchronization tolerance (e.g., ±50ms), then these data are directly aligned to that period; for video frames, select and The frame with the closest time frame is used as the video frame data for that period.

[0096] Furthermore, data that clearly does not conform to the sensor's physical measurement range is truncated to prevent abnormal sampling values ​​from directly entering the subsequent optimization process and causing numerical instability. For example, readings with a gas concentration below 0 are truncated to 0, readings with an oxygen concentration above the sensor's full scale are truncated to the full scale value, and readings with an LED channel drive current higher than the maximum allowable current are truncated to the maximum allowable current.

[0097] S13. Convert the multimodal monitoring data of the synchronous mine lamp into modal state values ​​to obtain the first... Within the sampling period, the first Modal state values ​​of each mode .in, This represents the modal index, with values ​​ranging from 1 to 2. ; This represents the total number of modes participating in the credibility assessment and optimization calculations, which can be taken as 7, corresponding to the following modes: gas concentration (%VOL), carbon monoxide concentration (ppm), oxygen concentration (%VOL), temperature and humidity (dimensionless), and dust concentration (mg / m³). 3 Inertial measurement mode (dimensionless) and video motion mode (dimensionless); It refers to the first Within the sampling period, the first The scalar state values ​​obtained after preprocessing each mode are used to participate in subsequent spatial consistency comparisons, long-term drift analysis, and safety risk calculations.

[0098] In practical implementation, for the gas concentration mode, carbon monoxide concentration mode, and dust concentration mode, the concentration values ​​after range checking and filtering are directly used as the corresponding mode state values. For the oxygen concentration mode, the deviation of the oxygen concentration from the safe center value (e.g., 21% VOL) ​​can be used as the mode state value; the greater the deviation from low or rich oxygen, the larger the mode state value. For the temperature and humidity mode, the thermal and humidity pressure index can be calculated based on the ambient temperature and relative humidity, and used as the temperature and humidity mode state value. For the inertial measurement mode, the changes in acceleration modulus, angular velocity, and attitude change amplitude within a short time window can be calculated to obtain the inertial measurement mode state values ​​related to personnel falling or violent shaking. For the video motion mode, the probability of abnormal motion can be obtained through human key point detection, and the probability of abnormal motion can be used as the video motion mode state value.

[0099] In one embodiment, for example, in the first Within each sampling period, the modal state values ​​for that period are calculated. for: =1 (gas concentration). = ; =2 (carbon monoxide concentration). = ; =3 (oxygen concentration). = ; =4 (temperature and humidity) = ; =5 (dust concentration) = ; =6 (Inertial Measurement). =0.05; =7 (video action) =0.02.

[0100] It should be noted that the human key point detection method is a computer vision technology used to locate the two-dimensional pixel coordinates of the main joints of the human body in video frames. By analyzing the displacement, velocity and relative position changes of the same joint between adjacent frames, the current action state of the person can be identified and the probability of abnormal action can be calculated.

[0101] S14. Arrange the modal state values ​​within the continuous sampling period into a target miner lamp modal state sequence in chronological order. , This refers to a data sequence consisting of all modal state values ​​across multiple sampling periods, used for subsequent calculations of modal reliability, environmental complexity indices, and multi-objective cost functions. It is a data sequence with dimensions of [missing information - likely a typology or specific dimension]. The matrix; where the first... The row corresponds to the first The sampling period, the first Column corresponding to the first One modality.

[0102] It should be noted that modal state values ​​do not require all modes to have the same physical dimensions. Subsequent steps can combine spatial reference values, short-time statistics, and full-scale reference values ​​for scale normalization. By first constructing modal state values ​​and then performing reliability assessment, heterogeneous data such as gas, dust, inertial measurement, and video motion can be incorporated into the same optimization framework. This avoids the problem of excessive computational burden caused by directly including raw video frames or high-frequency inertial measurement sequences in particle swarm optimization.

[0103] S2. Multimodal reliability assessment and multi-objective cost function construction for mine lamp operation and maintenance conditions;

[0104] When target miner lamps operate underground for extended periods, the degradation patterns of different sensor modalities vary. Gas sensors are prone to zero-point drift, dust sensors are susceptible to coal dust adhesion, video sensors are affected by lens contamination and light obstruction, and inertial measurement sensors are prone to loose installation or impact-induced displacement. Using fixed alarm baselines and fixed fusion weights can easily lead to false alarms, missed alarms, or unstable resource scheduling. This invention utilizes the spatial consistency of neighboring miner lamps, the long-term drift characteristics of the target miner lamp, and short-term operating condition fluctuation characteristics to construct modal reliability, environmental complexity indices, and a multi-objective cost function. This enables the subsequent optimization process to distinguish between sensor degradation and changes in the actual underground environment. The specific steps are as follows:

[0105] S21. Modal reliability calculation based on spatial consistency and long-term drift characteristics;

[0106] For multiple miners' lamps in the same area, if the modal state value of a certain mode of the target miner's lamp deviates significantly from the modal state value of the neighboring miners' lamps, and this deviation continues to change unidirectionally over a long time window, then that mode is more likely to have sensor contamination, zero-point drift, or sensitivity decay.

[0107] This step combines the spatial reference values ​​of neighboring miner lamps with the long-term drift of the target miner lamp to obtain the confidence level of each mode. This can quantify the confidence level of each mode in the target miner lamp within the current sampling period, indicating that it has not been affected by contamination, drift, or sensitivity degradation. The specific steps are as follows:

[0108] S211, in the Within each sampling period, the target miner lamp obtains the identification of neighboring miner lamps, received signal strength, communication distance, and same-mode state value through wireless broadcasting or underground gateway, and establishes a set of neighboring miner lamps based on the communication distance and received signal strength.

[0109] Specifically, the set of nearby miners' lamps is denoted as , It refers to the first The set of mine lamps that can provide spatial reference for the target mine lamp within a sampling period can be determined based on the condition that the communication distance is less than 15 meters and the received signal strength is higher than -75dBm; when the roadway structure is narrow or there are metal obstructions, the set of nearby mine lamps can also be determined by combining mine lamp positioning tags.

[0110] In practical implementation, the target miner lamp receives modal state values ​​reported by neighboring miner lamps within the same sampling period, and removes data with abnormal self-tests of neighboring miner lamps, severe missing communication packets, or significantly abnormal positioning distances. If the number of neighboring miner lamps is less than the preset number (for example, 3), the spatial reference confidence level is marked as low, and the spatial deviation penalty coefficient is reduced in subsequent modal confidence calculations to avoid drastic fluctuations in the confidence level of the target miner lamp due to insufficient neighboring samples.

[0111] S212, Collect from nearby miners' lamps Extract the first The modal state values ​​of the neighboring miner lamps for the mode are calculated, and the ... Within the sampling period, the first Spatial reference values ​​for each mode and spatial discrete scale .

[0112] Specifically, It refers to the first Set of neighboring miner lamps within a sampling period The Middle The median of the modal state values ​​is used to characterize the th modal state value within the same region. The stability level of each mode in the current sampling period; It refers to the first Set of neighboring miner lamps within a sampling period The Middle Each modal state value relative to the spatial reference value The median absolute deviation is used to characterize the normal dispersion range of state values ​​of the same mode within a neighborhood. Using the median and median absolute deviation can reduce the impact of anomalous readings from individual neighboring miners' lamps on spatial reference results.

[0113] In the actual implementation, the adjacent miners' lamps are first grouped together. All valid first The modal state values ​​are sorted by numerical value, and the median is taken as the spatial reference value. Then calculate the number of each adjacent miner's lamp. Modal state values ​​and spatial reference values The absolute differences between the values ​​are calculated, and the median of these absolute differences is taken as the median absolute deviation. Finally, the median absolute deviation is multiplied by 1.4826 to obtain the spatial discrete scale. .

[0114] In one embodiment, for example, the first The modal values ​​of gas concentration in neighboring miners' lamps within each sampling period were 0.41, 0.43, 0.44, 0.80, and 0.42 (in %VOL). Since 0.80 significantly deviates from the readings of other neighboring miners' lamps, using the arithmetic mean would inflate the spatial reference level. Using the median method provides a more stable representation of the actual gas concentration level in the current area. The sorted modal values ​​of gas concentration in neighboring miners' lamps were 0.41, 0.42, 0.43, 0.44, and 0.80, with spatial reference values... %VOL; the absolute differences are |0.41-0.43|=0.02, |0.42-0.43|=0.01, |0.43-0.43|=0, |0.44-0.43|=0.01, and |0.80-0.43|=0.37, respectively. The absolute deviations are ranked as 0, 0.01, 0.01, 0.02, and 0.37. The median absolute deviation is 0.01, therefore the spatial dispersion scale is... %VOL.

[0115] S213, Calculate the target miner's lamp number within a long-term sliding window. The relative deviation sequence of the nth mode relative to the spatial reference value, and the estimation of the nth mode based on the relative deviation sequence. Within the sampling period, the first Long-term drift of each mode .

[0116] Specifically, a long-term sliding window refers to the historical time range used to analyze the slow drift of a sensor. The window length can be 30 minutes. If the sampling interval is 3 seconds, then the long-term sliding window contains 600 sampling periods. It refers to the first Within the sampling period, the first The long-term offset of a mode relative to the spatial reference value of a neighboring miner's lamp is used to characterize whether the mode has a continuous unidirectional drift.

[0117] In the specific implementation, for each historical sampling period within the long-term sliding window... First calculate the target miner's lamp number 1 Modal state values With corresponding spatial reference value The difference between them yields the relative deviation value. ;

[0118] Right now ;in, This represents the index of historical sampling periods within a long-term sliding window. Indicates the target miner's lamp number The first historical sampling period The degree of deviation of a mode from the horizontal level of its neighborhood space. Indicates the target miner's lamp number The first historical sampling period Modal state values ​​of each mode. Indicates the target miner's lamp number The first historical sampling period Spatial reference values ​​for each modality are then used. Next, the Theil-Sen slope estimation method (a robust linear trend estimation method that effectively resists the interference of outliers on trend judgment by taking the median of the slopes calculated for all data point pairs) is applied to the relative deviation sequence within the long-term sliding window. This method calculates the ratio of the change in deviation between any two historical sampling periods to the time interval, and the median of all ratios is taken as the long-term trend slope. Finally, the absolute value of the long-term trend slope is multiplied by the length of the long-term sliding window to obtain the long-term drift. .

[0119] like Figure 1 As shown in one embodiment, the relative deviation sequence and long-term drift trend of the gas concentration mode are analyzed to show the deviation sequence of the target mine lamp gas concentration mode state value relative to the adjacent spatial reference value and the long-term trend of the Thales-Sen slope estimation; the horizontal axis is the sampling period number (dimensionless), and the vertical axis is the relative deviation value (unit: %VOL); the experiment shows that the long-term trend of the Thales-Sen slope estimation shows a continuous unidirectional increasing trend. Based on this, the long-term drift amount can be used to calculate the mode confidence corresponding to the gas mode, thereby distinguishing the long-term drift of the sensor from the overall change of the actual underground gas concentration.

[0120] In one embodiment, as an example, if the relative deviation value sequence is (0.01, 0.03, 0.02, 0.05, 0.04, ... ); Calculate the slope for all data pairs. For example, the slope for the 1st and 2nd cycles is (0.03-0.01) / 1=0.02, and the slope for the 1st and 5th cycles is (0.04-0.01) / 4=0.0075, etc. If the median of all slopes is 0.0001 and the window length is 600, then the long-term drift is... .

[0121] It should be noted that using a relative deviation sequence instead of the target lamp's own modal state value sequence can reduce the interference of overall changes in the underground environment on drift judgment. For example, when the gas concentration in a certain area actually increases, the gas concentration modal state values ​​of the target lamp and neighboring lamps will increase synchronously. At this time, the relative deviation of the target lamp relative to the spatial reference value may not necessarily increase. When the gas sensor of the target lamp experiences zero-point drift, the gas concentration modal state value of the target lamp will continuously deviate from the spatial reference value of neighboring lamps, and the relative deviation sequence will show a more obvious unidirectional change.

[0122] S214, According to the target miner's lamp Calculation of spatial deviation and long-term drift of the first mode. Within the sampling period, the first Modal reliability of each mode .

[0123] Specifically, modal credibility It refers to the first Within the sampling period, the first The confidence level of a mode unaffected by contamination, drift, or sensitivity degradation, dimensionless, with a value range of [value missing]. The closer the value is to 1, the more reliable the mode is; the closer the value is to 0, the more significantly the mode is affected by contamination, drift, or sensitivity degradation. In one implementation, the modal reliability is calculated based on the current spatial bias and long-term drift. The product of the two factors allows both to independently reduce modal confidence, and the calculation method is expressed as follows:

[0124] ;

[0125] in, This represents the spatial deviation penalty coefficient, used to control the decay rate of modal reliability when the target mine lamp modal state value deviates from the spatial reference value. It can be taken as 1.0 to 1.5. This represents the long-term drift penalty coefficient, used to control the intensity of the influence of long-term drift on modal reliability, and can be taken from 1.5 to 3.0; This represents a very small constant, used to prevent the denominator from being zero; it can be taken as... ; Indicates the target miner's lamp number The sampling period The absolute deviation between a modal state value and its spatial reference value, i.e., the spatial deviation; Indicates the first The full-scale reference value for each mode can be determined by the sensor calibration range and used to convert long-term drift to a comparable scale. This represents the natural exponential function.

[0126] It should be noted that this calculation method consists of two parts. The first part reduces the modal confidence based on the degree of deviation (i.e., spatial deviation) between the current modal state value of the target miner lamp and the reference value in the neighborhood space. The second part reduces the modal confidence based on the long-term drift. The two parts are fused by product to enable both short-term sudden deviations and long-term slow drifts to be identified.

[0127] S215, the first The modal confidence scores within each sampling period are combined in modal index order to obtain the modal confidence score vector. .

[0128] Specifically, modal credibility vector It refers to the first The vector composed of the confidence scores of all modes within each sampling period can be represented as: , is a dimension of a column vector, where, This represents the transpose operation of a vector.

[0129] It should be noted that modal confidence is used to characterize the reliability of modal state values ​​when participating in judgment and optimization search within the current sampling period. For gas safety-related modes such as methane concentration, carbon monoxide concentration, and oxygen concentration, by setting safety hard thresholds and minimum effective constraints, the ability to perceive gas hazards is avoided from being completely suppressed due to a decrease in modal confidence.

[0130] S22. Calculation of environmental complexity index based on trusted modal fluctuation and video motion entropy;

[0131] In actual coal mine operations, the more complex the underground working conditions, the more unstable the sensor status values ​​and personnel behavior are, requiring the optimization algorithm to have a stronger global search capability; the more stable the underground working conditions, the more stable the optimization algorithm needs to have a stronger local convergence capability. If particle swarm optimization always uses fixed inertial weights and fixed disturbance intensity, it is easy to get stuck in local optima under complex working conditions, or to over-search under stable working conditions.

[0132] This step weights the short-term fluctuations of multimodal data using modal credibility and combines it with video action entropy to obtain an environmental complexity index. The specific steps are as follows:

[0133] S221, with the first The sampling period is the current sampling period, and the sampling is extracted within the short-time sliding window. The historical modal state values ​​of the mode are calculated, and the ... Within the sampling period, the first Short-time mean of each modality and short-term standard deviation .

[0134] Specifically, a short-time sliding window refers to the time range used to describe the fluctuations of the current local operating conditions. The window length can be 60 seconds. If the sampling interval is 3 seconds, then the short-time sliding window contains 20 sampling periods. It refers to the first The sampling period corresponds to the first sampling period within the short-time sliding window. The mean of the modal state values ​​is used to characterize the stability level of the mode within the current local time range; It refers to the first The sampling period corresponds to the first sampling period within the short-time sliding window. The standard deviation of each modal state value is used to characterize the normal fluctuation scale of the mode within the current local time range.

[0135] In one embodiment, as an example, for the gas concentration mode, within a short-time sliding window containing 20 sampling periods, the historical state value sequence is (0.40, 0.42, 0.41, ..., 0.44)%VOL, and its mean is calculated. %VOL, standard deviation %VOL; Current period state value If %VOL is used, then the current deviation is |0.45-0.42|=0.03%VOL.

[0136] like Figure 2 and Figure 3 As shown in one embodiment, the historical values ​​of the gas concentration mode in the short-time sliding window and the magnitude of deviation of each mode from the short-time mean in the current sampling period are analyzed. The first graph shows the sampling period index (dimensionless) within the short-time sliding window on the horizontal axis and the gas concentration mode state value (unit: %VOL) on the vertical axis, where the red dashed line is the short-time mean and the shaded area is the short-time standard deviation range. The second graph shows the mode name (dimensionless) on the horizontal axis and the current deviation magnitude (unit: unit of the corresponding mode state value) on the vertical axis. Experiments show that the current value of the gas concentration mode is fluctuating around the mean. Most state values ​​have small deviations from the mean, but some state values ​​still have relatively large deviations. The deviation magnitude of the temperature and humidity mode is large, and the deviation magnitudes of the dust concentration mode and the carbon monoxide concentration mode are also relatively large. Therefore, the overall fluctuation level of the current mode state value can be measured by calculating the multimodal reliable fluctuation intensity.

[0137] In practical implementation, for modes with short-term missing values, the modal state value from the most recent valid sampling period can be used to temporarily fill in the missing values. If the duration of continuous missing values ​​exceeds a preset upper limit (e.g., 30 seconds), the modal reliability of that mode in the current sampling period is reduced, and that mode is no longer used to calculate short-term fluctuations. Based on this, it is possible to avoid the erroneous amplification of environmental complexity indicators caused by communication packet loss or short-term sensor interruptions.

[0138] S222. Based on the deviation of the current modal state value from the short-time mean, and in conjunction with the modal reliability, calculate the first... Multimodal reliable fluctuation intensity within each sampling period .

[0139] Specifically, multimodal reliable fluctuation intensity It refers to the first The overall deviation of the target mine lamp's multimodal state from its short-term steady state within a sampling period is dimensionless. Higher modal reliability means the short-term deviation of that mode better reflects the true changes in the underground environment; lower modal reliability means the short-term deviation is more likely to originate from sensor contamination, lens obstruction, or device drift, thus reducing its impact on the multimodal reliability fluctuation intensity. In one implementation, the multimodal reliability fluctuation intensity... By using modal confidence weighting, the degree to which the current state of each mode deviates from its short-term mean is calculated, and then normalized by dividing by the weighted standard deviation to obtain the overall volatility level. A value greater than 1 indicates that the current operating condition fluctuation exceeds the recent normal range. The calculation method is as follows:

[0140] ;

[0141] in, Indicates the first Multimodal reliable fluctuation intensity within each sampling period; Indicates the first Within the sampling period, the first The deviation of the current state of a modality from its short-time mean; This means that, while avoiding a denominator of zero, the current deviation of the mode is compared with its own local fluctuation level to achieve scale normalization of modes with different dimensions. This represents a very small constant, used to prevent the denominator from being zero; it can be taken as... .

[0142] S223. Extract the person's action state from the video frame data and calculate the first... Video motion entropy within a sampling period .

[0143] Specifically, video action entropy It is a dimensionless index used to characterize the movements, posture changes, and visual disturbances of personnel underground. Its value range can be normalized to [value range missing]. When people's movements are simple and the scene is stable, the video motion entropy is low; when people move densely, their postures change frequently, there are falls, or there is significant screen disturbance, the video motion entropy is high.

[0144] In one implementation, human keypoint detection is first performed on video frames within a short sliding window to obtain the velocity of keypoints, the amplitude of human posture changes, and the distribution of movement directions. Then, human actions are categorized into stillness, normal walking, bending over to work, rapid movement, and abnormal falls, etc. Action states; then calculate the probability of each action state occurring within a short sliding window. ,in, This represents the index of the action state category, with a value of [value]. , Indicates the first The probability of an action state occurring within a short sliding window is calculated; finally, the normalized entropy value is calculated based on the action state probability and used as the video action entropy. The calculation method is expressed as follows:

[0145] ;

[0146] in, This represents the logarithmic function with base 2.

[0147] In one embodiment, for example, if the actions are divided into Classes (stationary / normal walking / bending over, rapid movement, abnormal fall) are identified within a short sliding window of 60 seconds, spanning 20 video frames. The frequency of each action is recorded, and its probability is calculated.

[0148] ;

[0149] Information entropy ;

[0150] Normalize it to : .

[0151] S224, Combining Multimodal Reliable Fluctuation Intensity and video motion entropy , obtained the Environmental complexity index within each sampling period .

[0152] Specifically, environmental complexity index This refers to a comprehensive index used to characterize the intensity of dynamic changes and the complexity of potential hazards in a current downhole scenario; a higher value indicates a more complex downhole working condition. In one implementation, the environmental complexity index is calculated by linearly weighted combination of multimodal reliable fluctuation intensity and video motion entropy. This is used to comprehensively assess the complexity of the downhole scenario, and the calculation method is expressed as follows:

[0153] ;

[0154] in, This represents the multimodal reliable fluctuation weight, used to adjust the proportion of sensor fluctuation information and video motion information in the environmental complexity index, with a value range of [value range missing]. A value of 0.6 to 0.8 is acceptable. This represents the video motion entropy weight, used to supplement information on personnel actions and scene disturbances.

[0155] In practical implementation, when the modal confidence level corresponding to a video action modality is lower than a preset threshold (e.g., 0.4), it indicates that the video frame may be affected by lens occlusion, coal dust coverage, strong reflection, or low illumination. The video action entropy can then be adjusted. Replace it with the average video action entropy of the most recent few effective short-term sliding windows, or reduce the weight of video action entropy to prevent the environmental complexity index from being incorrectly amplified by distorted video information.

[0156] like Figure 4 As shown in one embodiment, the environmental complexity index changes over time, displaying the curve of environmental complexity index change that integrates multimodal reliable fluctuation intensity and video action entropy; the horizontal axis is the sampling period number (dimensionless), and the vertical axis is the environmental complexity index (dimensionless); experiments show that the environmental complexity index can dynamically reflect the non-stationarity of the underground working conditions, and the index increases when sensor fluctuations increase or personnel actions become chaotic, providing a basis for subsequent particle swarm optimization to adaptively adjust the inertial weight.

[0157] It should be noted that the environmental complexity index can be used as a search state adjustment parameter for subsequent particle swarm optimization, allowing the optimization process to be dynamically adjusted according to downhole operating conditions. When the environmental complexity index is high, particle swarm optimization can enhance global search capabilities; when the environmental complexity index is low, particle swarm optimization can enhance local convergence capabilities.

[0158] S23. Construction of a multi-objective cost function considering security risks, lighting power consumption, and communication vulnerability;

[0159] Online operation and maintenance optimization of mining lamps needs to consider safety risk identification, lighting energy consumption control, and wireless communication reliability simultaneously. If only safety risks are reduced, it may lead to excessive lighting and communication power consumption. If only power consumption is reduced, it may lead to insufficient hazard perception or alarm uploading capabilities.

[0160] This step encodes the multimodal fusion weights, alarm baseline values, sensitivity index, LED channel drive current, and wireless communication transmit power into a unified decision vector, and constructs a multi-objective cost function. The specific steps are as follows:

[0161] S231, Constructing the first Decision vector within each sampling period .

[0162] Specifically, decision vector This refers to the vector of all continuous decision variables for the online optimization of the target miner's lamp, which can be represented as:

[0163] ;

[0164] in, Indicates the first The modal fusion weights of the i-th modality are used to control the modal fusion weights of the i-th modality. The proportion of the influence of each modality in the safety risk assessment, dimensionless, satisfies... and ; Indicates the first The alarm baseline value of the first modality is used to determine the first... Whether a modal state value has entered the abnormal range, the unit should be consistent with the unit of the corresponding modal state value; Indicates the first The sensitivity index for each modality is used to adjust the rate of risk response growth after exceeding the alarm baseline value. It is dimensionless and can be taken from 0.5 to 3.0. This represents the total number of LED channels, which can be set to 3. Indicates the first The driving current of the LED channel, measured in mA. This represents the LED channel index, with a value of [value missing]. ; This indicates the wireless communication transmission power, used to control the link margin when the target mine lamp uploads alarm and status data, and is measured in dBm.

[0165] It should be noted that the decision vector The variables in the Particle Swarm Optimization (PSO) algorithm are used as decision variables to be optimized in the PSO algorithm. They are obtained through iterative search. The PSO algorithm evaluates particles representing different combinations of decision variables through a multi-objective cost function and finally outputs the optimal decision variable that minimizes the cost function value as the value of each variable in the current period.

[0166] S232, according to the... Modal state values Alarm baseline value and sensitivity index Calculate the first Risk response for each modality, combined with modality fusion weights and modal credibility Obtain the credible risk impact value.

[0167] In practical implementation, if the first Modal state values Not higher than the alarm baseline value Then the first The first mode does not produce a positive risk response; if the second mode does not produce a positive risk response. Modal state values Higher than the alarm baseline value First, calculate the relative proportion exceeding the alarm baseline value, then use the sensitivity index. Nonlinear amplification is applied. The larger the sensitivity index, the faster the risk response increases after exceeding the limit; the smaller the sensitivity index, the more gradual the risk response increases after exceeding the limit.

[0168] Furthermore, the first The risk response of each modality is multiplied by the modality fusion weight. and modal credibility Modal fusion weights Reflecting the first The importance of each modality in security risk assessment, and modality credibility. Reflecting the first The reliability of the current readings for each mode. Based on this, even if an abnormal spike occurs in a mode affected by contamination or drift, its impact on the overall safety risk will be suppressed, thereby reducing the probability of false alarms.

[0169] In one embodiment, as an example, if the gas concentration mode state value Alarm baseline value ,because Therefore, the risk response is 0, and this mode has no risk. If the gas concentration mode state value... Sensitivity Index ,at this time First, calculate the relative proportion exceeding the alarm baseline value. Then, the risk response is nonlinearly amplified through the sensitivity index. If modal fusion weights Modal reliability The credible risk impact of this mode on the total cost is then... .

[0170] It should be noted that the gas safety-related modes also have a safety hard threshold set independently of the multi-objective cost function. (The safety hard threshold refers to a fixed alarm limit set directly according to mandatory standards. As long as the mode state value reaches this limit, regardless of the mode's reliability or optimization result, an alarm will be forcibly triggered.) If the gas concentration, carbon monoxide concentration, or oxygen concentration meets the direct alarm conditions specified in the underground safety regulations, the target mine lamp will directly trigger a safety alarm and increase the wireless communication transmission power. The alarm will not be canceled due to a decrease in mode reliability, thus avoiding safety omissions caused by sensor reliability adjustment.

[0171] S233. Calculate the lighting power consumption based on the driving current of each LED channel.

[0172] Specifically, for the first LED channel, drive current Used to control the brightness and energy consumption of the channel's lighting, with a value range of [value range missing]. , This indicates the maximum allowable current of the LED channel, determined by the LED's rated parameters, the lamp's heat dissipation capacity, and the battery capacity; a value of 350mA can be taken. The forward voltage drop of the LED channel is denoted as... , The value can be obtained from the parameters of the LED device or the online sampling circuit, and can be taken as 3.2V.

[0173] In practical implementation, the channels of each light-emitting diode... Summing yields an estimated DC power consumption for lighting, which is used in subsequent calculations of the lighting power consumption term. If the target mine lamp has a low beam channel, a high beam channel, and a warning flashing channel, different current boundaries can be set for each of the three channels, and the brightness distribution can be dynamically adjusted according to the underground working conditions during subsequent optimization.

[0174] In one embodiment, for example, if the target miner's lamp has three LED channels (L=3) for low beam, high beam, and warning flashing, their driving currents are respectively , , The corresponding positive pressure drop Since both are 3.2V, the estimated DC power consumption for lighting is: .

[0175] S234. Based on wireless communication transmission power Calculate the communication vulnerability item based on the current link state.

[0176] Specifically, wireless communication transmission power The range of values ​​is , The maximum allowable transmission power for wireless communication is determined by the rated power of the communication module and the requirements for underground wireless management, and can be taken as 20dBm; the communication vulnerability item is used to characterize the risk of alarm data upload failure or increased delay of the target mine lamp under the current transmission power.

[0177] In practical implementation, the signal-to-interference-plus-noise ratio (SIR) is calculated based on the received signal strength of the target miner's lamp, path loss estimation, co-channel interference power, and noise power. , This refers to the transmission power being The quality index of the wireless link of the target miner's lamp; the higher the value, the more stable the link. The maximum signal-to-interference-plus-noise ratio (SIR) represents the ideal low-interference condition and can be determined by historical communication calibration data or downhole communication system configuration.

[0178] Furthermore, first calculate the signal-to-interference-plus-noise ratio at the current transmit power. Compared to the ideal maximum signal-to-interference-plus-noise ratio The ratio ( This represents a very small constant, used to prevent the denominator from being zero; it can be taken as... Then, subtracting this ratio from 1 yields the difference in link quality relative to the ideal state, resulting in the normalized signal-to-interference-plus-noise ratio (SIR) defect. The normalized SIR defect is used to calculate the communication vulnerability term in subsequent calculations.

[0179] S235. The trusted security risk item, lighting power consumption item, and communication vulnerability item are weighted and summed to obtain the... Multi-objective cost function within each sampling period .

[0180] Specifically, multi-objective cost function Refers to the decision vector used for evaluation The comprehensive objective function assesses the overall effectiveness of the current decision configuration in balancing security risk, lighting energy consumption, and communication reliability. A smaller value indicates a better overall effect. In one implementation, the first part is a trusted security risk term, summing the risk responses of each mode modulated by trustworthiness and weights, representing the overall risk cost under the current configuration. The second part is a lighting power consumption term, calculating the normalized power consumption of all LED channels, representing the energy consumption cost. The third part is a communication vulnerability term, quantifying the difference between the current signal-to-interference-plus-noise ratio (SNR) and the ideal value. A squared term is used to more quickly penalize low SNR states, representing the cost of unreliable communication links. Multi-objective cost function. The calculation method is expressed as follows:

[0181] ;

[0182] in, This represents the typical forward voltage drop of the LED, which can be taken as 3.2V; This represents the lighting power consumption penalty coefficient, used to balance the optimization tendency between safety risks and lighting energy consumption, and can be taken from 0.1 to 1.0; This represents the communication vulnerability penalty coefficient, used to control the penalty strength on the multi-objective cost function when the link quality deteriorates, and can be taken from 0.1 to 0.3; This represents a very small constant, used to prevent the denominator from being zero; it can be taken as... ; This represents the maximum value function.

[0183] It should be noted that the output of step S2 includes the modal confidence vector. Environmental complexity index and multi-objective cost function Modal credibility vector Used to guide subsequent particle swarm optimization in applying directional perturbations to decision variables related to low-confidence modes; environmental complexity index Used to adjust the overall search scale of subsequent particle swarm optimization; multi-objective cost function Used to evaluate the quality of candidate decision vectors in a particle swarm.

[0184] S3. Two-layer particle swarm optimization driven by credibility and operational complexity;

[0185] Given the high dimensionality, nonlinearity, multiple constraints, and non-stationary operating conditions of the constructed multi-objective cost function, conventional particle swarm optimization typically employs fixed inertial weights and a uniform velocity update method, which makes it difficult to simultaneously consider the large-scale search under complex operating conditions, the rapid convergence under stationary operating conditions, and the directional correction of variables related to low-confidence modes.

[0186] This step embeds the environmental complexity index into the inertia weight adjustment process and the modal confidence into the velocity update process, enabling the particle swarm to enhance global search under complex conditions and local convergence under stable conditions. It also applies targeted perturbations to decision variables related to low-confidence modes. The specific steps are as follows:

[0187] S31, Particle coding and constraint-aware initialization;

[0188] By employing particle encoding, the multi-objective decision-making problem for miners' lamps can be transformed into a particle swarm optimization problem. Each particle represents a candidate miners' lamp decision configuration, its position represents the decision vector, and its velocity represents the update direction and magnitude of the candidate decision configuration. This step uses chaotic sequences to generate an initial population that satisfies boundary constraints, and combines modal confidence to expand the initialization coverage of low-confidence modal relevance dimensions. The specific steps are as follows:

[0189] S311, Decision vector It serves as the particle position vector, while maintaining a fixed encoding order.

[0190] In a specific implementation, the particle position vector includes, in sequence: Modal fusion weights, Each modal alarm baseline value, Modal sensitivity index, One LED channel drive current and one wireless communication transmit power; the total dimension of the decision vector is denoted as... , The number of variables contained in the particle position vector, satisfying... .

[0191] Specifically, the first The particle in the first The position vector of the algebra is denoted as , It refers to the first The particle in the first The corresponding candidate mining lamp decision configuration; Represents the particle index, with a value of ; This represents the total number of particles, which can be between 30 and 60. This represents the particle swarm iteration algebra index, with values ​​ranging from 1 to 2. ; This represents the maximum number of iterations, which can range from 50 to 120. The particle in the first The generation The position of each dimension is denoted as , This represents the decision dimension index, with values... .

[0192] S312. Set physical feasible boundaries for each decision dimension and establish constraint rules based on variable types.

[0193] Specifically, the first The lower bound of each dimension is denoted as , No. The upper bound of each dimension is denoted as For modal fusion weights, the boundary is... Furthermore, all modal fusion weights must be non-negative and sum to 1; for alarm baseline values, the boundaries are jointly determined by the corresponding sensor range and downhole safety regulations; for sensitivity indices, the boundaries can be taken as... For the LED channel drive current, the boundary is For wireless communication transmit power, the boundary is .

[0194] In practice, the baseline value for the gas concentration mode alarm shall not exceed the upper limit of the warning allowed in the downhole safety regulations; the baseline value for the carbon monoxide concentration mode alarm shall not exceed the corresponding upper limit of the safety warning; the oxygen concentration mode can be set with low oxygen deviation alarm baseline value and oxygen enrichment deviation alarm baseline value, or the magnitude of oxygen deviation from the safety center value can be used as the mode state value and then a unified alarm baseline value can be used.

[0195] In one embodiment, for example, the lower limit of the gas concentration modal alarm baseline value can be set to 0.1%VOL, and the upper limit must be strictly less than the warning upper limit specified in the "Coal Mine Safety Regulations", such as 1.0%VOL, which can be set to 0.8%VOL; the upper limit of the carbon monoxide concentration modal alarm baseline value can be set to 24ppm; the oxygen concentration modal can directly use the "deviation from the safety center value" as the modal state value, and at this time the unified alarm baseline value can be set to the deviation degree of 2.0%VOL.

[0196] S313. Generate a Tent chaotic sequence value for each particle and map the Tent chaotic sequence value to the physical feasible boundary of each decision dimension.

[0197] Specifically, the Tent chaotic sequence value is denoted as , Indicates the first The particle in the first The chaotic sequence values ​​in each dimension range from 1 to 1. In one implementation, each particle is first randomly generated. Then, other dimensions are generated recursively according to the Tent mapping. Then Linear mapping to the first Within the boundary range of each dimension, the initial position is obtained. .

[0198] In one embodiment, for example, for particles First, generate randomly. Mapping via Tent It is obtained by recursion. , And so on; if the dimension The corresponding gas concentration alarm baseline value has a boundary of [0.1, 0.8]; the chaos value of this dimension Mapped to this interval, initial position .

[0199] It should be noted that, compared with completely uniform random initialization, Tent chaotic sequences can enhance the ergodicity of initial particles in the search space and reduce the probability of particles concentrating in local areas. This is especially suitable for high-dimensional scenarios where modal fusion weights, alarm baseline values, sensitivity indices, lighting currents, and communication transmission power are all involved in optimization.

[0200] S314. Perform constraint correction on the initial particle positions to ensure that each initial particle meets the deployable conditions.

[0201] Specifically, for modal fusion weights, negative values ​​are first truncated to 0, and then the sum of all modal fusion weights is calculated. If the sum of weights is greater than 0, each modal fusion weight is divided by the sum of weights to make all modal fusion weights non-negative and summed to 1. If the sum of weights is equal to 0, each modal fusion weight is set to... For alarm baseline values, sensitivity index, LED channel drive current, and wireless communication transmission power, truncation is performed according to their respective physical feasible boundaries.

[0202] In one embodiment, as an example, if the modal fusion weights of a certain particle before correction are [0.4, -0.1, 0.2]; first, truncate the negative values ​​to 0, obtaining [0.4, 0, 0.2], and the total weights are 0.6; normalize by dividing each weight by the total, obtaining the final weights as: [0.4 / 0.6, 0, 0.2 / 0.6] ≈ [0.667, 0, 0.333]. For scalar variables such as alarm baseline values, if a certain It is 0.4A but If so, it is directly truncated to 0.35A.

[0203] Furthermore, based on the modal confidence vector Expanding the initial perturbation range for low-confidence modal correlation dimensions allows the initial population to cover more potentially effective candidate configurations under sensor degradation states, preventing the particle swarm from starting its search from overly ideal healthy sensor configurations. In specific implementations, when the... Modal reliability of each mode Below the preset confidence threshold When (for example, 0.4) is taken, the initial coverage of the modality fusion weight, alarm baseline value and sensitivity index corresponding to this modality is expanded.

[0204] In one embodiment, for example, if the dust concentration mode confidence level If it is below the threshold of 0.4, its alarm baseline value The normal initialization range is [0.1, 0.8]%VOL; to expand the coverage, the initialization range is adjusted to [0.05, 0.95]%VOL, and the Tent chaotic sequence values ​​in this dimension are... A linear mapping from [0, 1] to the %VOL interval [0.05, 0.95].

[0205] S315. Divide the particle swarm into a perception strategy subgroup and a resource scheduling subgroup, and make the two subgroups share the group's optimal position.

[0206] Specifically, the perception strategy subgroup focuses on searching for modal fusion weights, alarm baseline values, and sensitivity indices, while the resource scheduling subgroup focuses on searching for LED channel drive current and wireless communication transmission power. Both subgroups retain complete decision vectors. The perception strategy subgroup uses a higher update probability for perception-related dimensions, while the resource scheduling subgroup uses a higher update probability for resource-related dimensions. The two subgroups achieve collaborative optimization by sharing the group's optimal position.

[0207] In one implementation, when updating particle velocities, the perception strategy subgroup applies full individual and group learning factors to the dimensions of "modal fusion weights, alarm baseline values, and sensitivity index" in the decision vector; for the dimensions of "LED current and communication transmission power," the individual and group learning factors are set to a very small value (e.g., 0.001), making their update probability extremely low. The resource scheduling subgroup does the opposite. After each iteration, both subgroups compare their respective group-optimal positions and share the position with the lower fitness as the global group-optimal position with all particles.

[0208] It should be noted that the perception strategy subgroup can quickly adapt to sensor contamination, drift, and scene errors, while the resource scheduling subgroup can quickly adapt to changes in lighting energy consumption and communication links. The two subgroups update collaboratively under the same multi-objective cost function, ensuring that security risks, lighting power consumption, and communication reliability are not separated into unrelated local optimization problems.

[0209] S32, Heterogeneous speed update that integrates environmental complexity and modal reliability;

[0210] Differentiated velocity updates are used to dynamically adjust particle search behavior based on downhole operating condition complexity and modal reliability. Higher environmental complexity requires particles to retain historical velocities more effectively to expand the search range; lower modal reliability necessitates stronger directional perturbations to the corresponding modal fusion weights, alarm baseline values, and sensitivity indices to quickly escape unreliable configurations. This step achieves differentiated velocity updates through adaptive inertial weights and reliability-driven Cauchy mutation, as detailed below:

[0211] S321. Based on environmental complexity indicators Calculate the first Within the sampling period, the first Adaptive inertia weights of the generation .

[0212] Specifically, adaptive inertia weights This value controls the degree to which particles inherit historical speed. A larger value indicates that the particle is more likely to maintain its original search direction and perform global exploration, while a smaller value indicates that the particle is more likely to reduce its step size and perform local convergence. In one implementation, adaptive inertia weights are used. The calculation method is expressed as follows:

[0213] ;

[0214] in, This represents the minimum inertia weight, used to ensure local development capability under stable operating conditions, and can be taken as 0.4; This represents the maximum inertia weight, used to ensure global exploration capability under complex working conditions, and can be set to 0.9; This represents the environmental complexity mapping coefficient, used to adjust the influence of the environmental complexity index on the adaptive inertia weight, and can be taken from 1.0 to 2.0; This represents the hyperbolic tangent function, used to smoothly map environmental complexity indicators to a finite range.

[0215] It should be noted that during stable inspections, the environmental complexity index is low, the adaptive inertia weight is close to 0.4 to 0.5, and the particle swarm converges to a local optimum configuration with low energy consumption and stable communication more quickly. When there are rapid changes in dust concentration, rapid movement of personnel, and communication jitter, the environmental complexity index increases, the adaptive inertia weight increases, and the particle swarm expands its search range to adapt to non-stationary operating conditions.

[0216] S322. Establish the mapping relationship between decision dimensions and modalities. , used to determine the first Does each decision dimension correspond to a certain modality?

[0217] Specifically, modal mapping It refers to the first The modality number corresponding to the decision dimension. When the first... The dimension corresponds to the first When setting the modal fusion weights, alarm baseline values, or sensitivity indices for each modality, let When the first When a dimension corresponds to the LED channel drive current or wireless communication transmit power, it indicates that the dimension does not belong to a specific sensing mode. A value of 0 is acceptable; credibility-driven variants do not apply to this dimension.

[0218] It should be noted that, through modal mapping, particle swarm optimization can apply additional perturbations only to sensing policy variables related to low-confidence modes, without directly perturbing the lighting current or wireless communication transmission power due to contamination by a single sensor, thus avoiding unnecessary coupling interference between different types of variables.

[0219] S323. Perform velocity updates for each particle and each decision dimension to obtain the... Particle velocity.

[0220] Specifically, the first The particle in the first The generation The velocity in each dimension is denoted as , Used to characterize the particle's historical search direction and search step size in this dimension; The particle up to the [number]th The optimal historical position of an individual in the 20th generation The values ​​of each dimension are denoted as ; up to the The optimal position of the group in the th generation The values ​​of each dimension are denoted as In one implementation, a credibility-driven mutation term is introduced. This term applies additional perturbations to the dimensions of decision variables associated with a modality when the credibility of that modality is low, in order to escape local optima. The velocity update method is expressed as follows:

[0221] ;

[0222] in, Indicates the first The particle in the first The generation Speed ​​in each dimension; This represents the individual learning factor, used to control how close a particle gets to its historical best position, and can be taken from 1.4 to 1.8. This represents the group learning factor, used to control the degree to which particles move closer to the optimal position of the group, and can be taken from 1.6 to 2.0; Represents the individual random factor, by Obtained by uniformly distributed sampling within the interval; Represents the population random factor, by Obtained by uniformly distributed sampling within the interval; This represents the variation intensity coefficient, used to control the additional perturbation amplitude triggered by modal confidence, and can be taken from 0.03 to 0.08; This represents a Cauchy random perturbation, used to generate a small number of large step jumps; Used to scale the variation magnitude to the 1st Within the physical dimensions of each dimension; Indicates the first Within the sampling period, the first Each dimension is mapped through modality. The associated first Modal reliability of each modality.

[0223] It should be noted that Cauchy random perturbation It can be obtained by random sampling from the standard Cauchy distribution; to avoid extreme random values ​​causing particles to exceed the limit, it can be... Cut off to Within the range, the speed update calculation is then performed.

[0224] S324. Based on modal mapping Control whether credibility-driven variants take effect.

[0225] Specifically, when the first When each dimension corresponds to the LED channel drive current or wireless communication transmission power, Processed as 1, making the confidence-driven variance term 0; when the... The dimension corresponds to the first When determining the modality fusion weights, alarm baseline values, or sensitivity indices for a given modality, the corresponding modality confidence level should be used. Controlling the amplitude of variation. Modal reliability. The lower, The larger the value, the greater the variation in the modality-related decision variable.

[0226] In one embodiment, for example, if the target mine lamp gas sensor experiences zero-point drift due to a high-humidity environment, and the modal confidence level corresponding to the gas concentration mode decreases from 0.9 to 0.3, then the modal fusion weight, alarm baseline value, and sensitivity index corresponding to the gas concentration mode will experience a larger dimensional perturbation, enabling the optimization process to quickly find a more robust gas alarm configuration; however, the LED channel drive current and wireless communication transmission power are not affected by the direct perturbation caused by the decrease in the confidence level of the gas concentration mode.

[0227] S325. Limit the updated speed to obtain a controllable speed.

[0228] Specifically, for the first Each dimension, with the maximum speed denoted as... It can be set to ;in, This represents the speed limiting ratio, which can be between 0.1 and 0.3. If... Greater than Then Set as ;like Less than Then Set as .

[0229] In one embodiment, for example, if a certain dimension has a lower bound Upper Realm Speed ​​limiting ratio is taken Then the maximum speed If the speed after the update If it exceeds the upper limit, its amplitude will be limited to 0.07.

[0230] It should be noted that adaptive inertia weights address the question of "whether the particle swarm should expand the search or accelerate convergence under the current operating conditions," while credibility-driven Cauchy mutation addresses the question of "which modal-related variables need to be prioritized for perturbation." Based on this, the combination of the two allows the particle swarm to adjust the overall search scale according to the complexity of the downhole conditions, and to perform targeted optimization of low-credibility modal-related variables based on the modal credibility state.

[0231] S33, Constrained Projection, Fitness Update and Optimal Decision Output;

[0232] After completing the velocity update, the candidate particle positions need to be corrected to a physically feasible range, and fitness evaluation is performed using a multi-objective cost function. This step performs boundary truncation, weight normalization, and safety retention constraint corrections on different types of decision variables to ensure that the optimization results can be directly deployed to the mining lamp embedded system. The specific steps are as follows:

[0233] S331. Calculate the temporary position of the particle based on the updated velocity.

[0234] Specifically, the first The particle in the first The generation Temporary positions in each dimension are denoted as , This refers to the candidate positions before constraint correction, used as input for subsequent feasibility correction; temporary positions are obtained by... The position and the first The result is obtained by adding the speeds together.

[0235] In practical implementation, for the perception strategy subgroup, a larger update probability (e.g., 0.8 to 1.0) can be set for the modality fusion weight, alarm baseline value, and sensitivity index dimensions, while a smaller update probability (e.g., 0.1 to 0.3) can be set for other dimensions. For the resource scheduling subgroup, a larger update probability can be set for the LED channel drive current and wireless communication transmission power dimensions, while a smaller update probability can be set for other dimensions. Dimensions that are not prioritized for updating still retain complete position values, thereby ensuring that each particle always corresponds to a complete candidate decision vector.

[0236] S332. Perform boundary truncation and weight normalization on the temporary location to obtain a preliminary feasible location.

[0237] In practical implementation, for modal fusion weights, all negative weights are first truncated to 0. Then, the sum of the truncated weights is calculated. If the sum of weights is greater than 0, each modal fusion weight is divided by the sum of weights. If the sum of weights is equal to 0, each modal fusion weight is set to zero. For alarm baseline values, the boundary is truncated according to the allowable range of the corresponding sensor; for sensitivity indices, the threshold is truncated to... For the LED channel drive current, cut off to For wireless communication transmission power, truncation to .

[0238] In one embodiment, as an example, for a decision vector before particle constraint correction, if the modal fusion weight is [0.6, -0.1, 0.2], all negative weights are first truncated to 0, and then normalized to obtain [0.75, 0, 0.25]. If the alarm baseline value is 1.5%VOL, but the upper bound is 0.8%VOL, it is truncated to 0.8%VOL. If the sensitivity index is 5.0, it is truncated to 3.0. If the LED current is -0.05A, it is truncated to 0A. If the wireless communication transmission power is -5dBm, it is truncated to 0dBm.

[0239] S333, Perform safety retention constraint corrections on gas safety-related modes.

[0240] Specifically, the set of gas safety-related modes is denoted as... , It refers to the set of gas modes that are directly related to safety, including methane concentration mode, carbon monoxide concentration mode and oxygen concentration mode; the safety retention constraint is used to ensure that the target miner lamp does not completely lose its gas hazard perception capability due to the decrease in the credibility of multiple gas safety-related modes.

[0241] In practical implementation, it is based on the set of gas safety-related modes. Select the gas safety-related mode with the highest modal confidence. and check Is it below the minimum effective impact threshold? .in, This represents the set of gas safety-related modes within the current sampling period. The modality index with the highest confidence level in the middle modality; Indicates the first The modality fusion weights corresponding to each modality; Indicates the first Within the sampling period, the first Modal reliability corresponding to each mode; This represents the minimum effective impact threshold for gas safety-related modes, and can be taken as 0.03 to 0.08. If... Then improve ,make And proportionally reduce the fusion weights of other non-critical modalities so that all fusion weights are still non-negative and sum to 1.

[0242] In one embodiment, for example, if the gas safety-related mode is methane ( ) and oxygen ( Of these, gas has the highest reliability, therefore ;like , ,but ;like ,because , need to Increase to If the original dust weight is 0.2, the original video weight is 0.1, and the sum of the weights for other non-gas modes is 0.5, then the additional weight of 0.04 will be calculated as follows: and The proportion of deduction, i.e., the reduction in dust weight. Video weight reduced .

[0243] S334. Utilizing a multi-objective cost function Calculate the fitness value for each particle.

[0244] Specifically, the first The first particle The feasible locations after constraint correction form the candidate decision vector. , and the candidate decision vector Substitute into the multi-objective cost function , obtain fitness The smaller the fitness value, the better the overall effect of the candidate decision vector in terms of security risk, lighting energy consumption, and communication reliability.

[0245] In practical implementation, if a candidate decision vector violates the safety hard threshold constraint for a certain mode, but the candidate decision vector suppresses the mode fusion weight of the corresponding mode too low, a penalty term much larger than the normal value can be added to the fitness value (for example, the fitness value can be set to a lower value). This ensures that the optimization direction of the multi-objective cost function will not conflict with downhole safety regulations.

[0246] S335. Update the individual's historical best position and the group's best position based on the fitness value.

[0247] Specifically, if the first If the fitness value of the current candidate decision vector of a particle is less than the fitness value corresponding to the particle's historical best position, then the current candidate decision vector is updated to the value of the particle with the best position in history. The individual best position of each particle is determined; if the best fitness value among all particles is less than the fitness value corresponding to the current best position in the population, then the best position in the population is updated.

[0248] Furthermore, the perception strategy subgroup and the resource scheduling subgroup share their respective optimal candidate decision vectors after each iteration, and select the one with the smaller fitness value from the optimal candidate decision vectors of the two subgroups as the global group optimal position. This collaborative approach can avoid the isolation between perception strategy optimization and resource scheduling optimization, enabling the target mine lamp to make unified decisions in reducing false alarms, reducing lighting power consumption, and improving communication reliability.

[0249] S336. Determine whether the iteration termination condition is met, and output the iteration term. The optimal decision vector corresponding to each sampling period .

[0250] In practical implementation, the iteration termination condition includes reaching the maximum number of iterations. Or, the change in the optimal fitness value of the population over several consecutive generations is less than a preset threshold (for example, a threshold value can be taken as 0). If the iteration termination condition is not met, the speed update, position correction, and fitness evaluation will continue; if the iteration termination condition is met, the optimal decision vector will be output. .

[0251] Specifically, the optimal decision vector It refers to the first The optimal mine lamp decision configuration obtained by particle swarm search within a sampling period includes the fusion weights of each mode, the alarm baseline values ​​of each mode, the sensitivity index of each mode, the drive current of each LED channel, and the wireless communication transmission power; the target mine lamp embedded system is based on the optimal decision vector Real-time updates of anomaly detection, alarm determination, lighting adjustment, and communication transmission strategies.

[0252] In one embodiment, for example, the first After optimization for each sampling period, the optimal decision vector is output. The vector indicates that the target miner lamp should currently set the fusion weight of the gas concentration mode to 0.35, the alarm baseline value to 0.45%VOL, the sensitivity index to 1.8, the drive current of the high beam channel to 120mA, and the wireless transmission power to 15dBm.

[0253] like Figure 5 As shown in one embodiment, the modal fusion weight distribution in the optimal decision vector of the 51st sampling period is analyzed; the horizontal axis represents the modal name (dimensionless), and the vertical axis represents the modal fusion weight (dimensionless). The bar chart shows that the gas concentration modal has the highest fusion weight, followed by the oxygen concentration modal and the carbon monoxide concentration modal, while the weights of other modalities are relatively low. Experiments show that the particle swarm optimization algorithm automatically adjusts the fusion weight according to the confidence level of each modality, suppresses low-confidence modes, and thus reduces the risk of false alarms.

[0254] It should be noted that the core of step S3 is to continuously introduce downhole operating condition complexity and modal reliability information during the particle swarm search process. The environmental complexity index is used to control the overall search scale of the particle swarm, modal reliability is used to control the directional perturbations of modal-related dimensions, and safety retention constraints are used to ensure that the gas hazard detection capability is not completely suppressed. Based on this, the target mine lamp can achieve more robust online operation and maintenance decisions even under conditions of simultaneous sensor contamination, limited lighting resources, and communication link fluctuations.

[0255] S4, online adaptive operation and maintenance execution of mining lamps;

[0256] After completing modal reliability assessment and particle swarm optimization, the target mine lamp can directly apply the optimal decision vector to its online operation underground. During online execution, the system periodically updates the operation and maintenance strategy according to the same data acquisition, reliability assessment, and constraint correction rules as in the optimization phase. The specific steps are as follows:

[0257] S41, the target miner's lamp is in Raw mine lamp multimodal monitoring data were collected in one sampling period. And obtain the modal state values ​​according to step S1. .

[0258] In practical implementation, the target mine lamp performs synchronization, validity checks, and modal state value conversion on gas concentration data, temperature and humidity data, dust concentration data, inertial measurement data, video frame data, lighting drive data, and wireless communication status data, forming the first... The set of modal state values ​​within a sampling period; if a certain mode is missing for multiple consecutive sampling periods, the target miner lamp marks the mode as a low-confidence mode and increases the search perturbation of the decision variables related to the mode in subsequent optimization.

[0259] S42, Target miner's lamp is based on the set of nearby miner's lamps. Calculate the modal confidence vector And calculate the environment complexity index based on the short-time sliding window. .

[0260] In practical implementation, the target miner lamp obtains the modal state values ​​of neighboring miner lamps via wireless broadcast or underground gateway, calculates spatial reference values, spatial discrete scale, and long-term drift, and then obtains the confidence level of each modality. Simultaneously, it calculates the environmental complexity index based on the fluctuation of modal state values ​​within a short-term sliding window and video motion entropy. If the number of neighboring miner lamps is insufficient, the target miner lamp reduces its spatial deviation penalty coefficient and uses the median of its most recent stable time period as a temporary spatial reference value to avoid sudden decision-making due to missing spatial references.

[0261] S43. Construct a multi-objective cost function for the current sampling period using the target miner's lamp. Then, step S3 is executed to obtain the optimal decision vector. .

[0262] Specifically, when embedded computing resources are limited, the number of particles can be... Setting it to 20 or 30 will increase the maximum number of iterations. Setting the number to 30 to 50 allows for increasing the particle count and iteration count when downhole gateways or edge servers are involved in the computation, resulting in more comprehensive search results. Constraint projection, safety retention constraints, and safety hard thresholds remain consistent regardless of whether local computation or edge server collaborative computation is used.

[0263] S44, The target miner's lamp is determined according to the optimal decision vector. Update the modal fusion weights, alarm baseline values, sensitivity index, LED channel drive current, and wireless communication transmit power.

[0264] In practical implementation, the target miner's lamp first calculates the credible security risk based on the optimal modal fusion weight, optimal alarm baseline value, and optimal sensitivity index; then, it adjusts the brightness of the low beam, high beam, or warning LED channels based on the optimal LED channel drive current; finally, it uploads the target miner's lamp status data, credibility data, and alarm data based on the optimal wireless communication transmission power. When the communication link quality is poor, the optimal wireless communication transmission power is increased to improve the reliability of alarm upload; when the environment is stable and the link quality is good, the optimal wireless communication transmission power is reduced to save battery energy.

[0265] S45. The target miner's lamp performs a safety hard threshold judgment and a continuous confirmation judgment, and outputs an alarm result or a low confidence prompt.

[0266] In practical implementation, for gas safety-related modes such as methane concentration, carbon monoxide concentration, and oxygen concentration, if the mode state value reaches the direct alarm condition corresponding to the safety hard threshold, the target miner's lamp immediately outputs a safety alarm and uploads the alarm information with a high wireless communication transmission power. For situations where the safety hard threshold has not been reached but the credible safety risk continues to increase, the target miner's lamp performs continuous confirmation judgment, and the number of continuous confirmation windows can be set. The value is 3 to 5; when the same risk type occurs consecutively If the warning conditions are exceeded within each sampling period, and the corresponding modality confidence level is not lower than the preset minimum confidence threshold, a formal warning will be output. If the risk type changes frequently or is mainly triggered by low-confidence modalities, a low-confidence alert will be output, and subsequent sampling periods will continue to be observed.

[0267] In one embodiment, for example, if the gas concentration state value is 1.6%VOL, reaching the safety hard threshold direct alarm condition, the system immediately issues an audible and visual alarm and uploads the alarm information at maximum transmission power; if the dust concentration state value is continuously for 3 sampling periods ( If all three dust sensor modal confidence values ​​exceed their optimal alarm baseline values ​​and the modal confidence of the dust sensor is greater than 0.6 in all three cycles, the system will issue a formal warning of dust concentration exceeding the limit and upload the warning information according to the optimal transmission power. If the video motion modality detects "abnormal falls" for two consecutive cycles, but the modal confidence is only 0.3 due to lens contamination, the system will not issue a formal personnel fall alarm. Instead, it will record and output a "video modal low confidence event" prompt locally, waiting for the confidence to recover or for further confirmation.

[0268] S46, the target miner's lamp will be the first Modal state values ​​and modal confidence vector for each sampling period Environmental complexity index Optimal decision vector The alarm results are written to the operation and maintenance log and then proceed to the next step. One sampling period.

[0269] Specifically, the operation and maintenance logs are used for subsequent sensor maintenance, mine lamp status tracking, and model parameter calibration. If the modality confidence level of a certain mode is lower than the maintenance threshold (e.g., 0.3) for more than 30 minutes in consecutive sampling periods, the target mine lamp or underground gateway will generate a maintenance prompt, suggesting that the corresponding sensor be cleaned, calibrated, or replaced.

[0270] It should be noted that, through the above-described online execution process, the target mine lamp can continuously perform multimodal reliable sensing, lighting resource adjustment, and communication resource scheduling in non-stationary underground environments. Simultaneously, it can reduce the probability of false alarms when sensor contamination or drift exists, retain direct alarm capabilities when real dangers occur in gas safety-related modes, and achieve more robust operation and maintenance decisions through adaptive particle swarm optimization under complex operating conditions.

[0271] In one embodiment, for example, in a fully mechanized coal mining face, a smart mine lamp has been running continuously for 15 days. Due to the high level of coal dust, its dust sensor window has gradually become contaminated, resulting in a persistently high dust concentration modal state value, and the confidence level of this mode rapidly decreased from 0.8 within 30 minutes. The system recognizes the increased long-term drift and spatial deviation of the dust mode and reduces its mode confidence level to 0.25. During particle swarm optimization, the alarm baseline value and fusion weight related to the dust mode are subjected to stronger confidence-driven mutation perturbations. After optimization calculation, the dust alarm baseline value of the output new decision vector is reduced from 5 mg / m³. 3 Automatically increased to 15mg / m 3The fusion weight was reduced from 0.2 to 0.05, effectively avoiding continuous false alarms caused by sensor contamination. Simultaneously, due to normal personnel operation within the roadway, video motion entropy was low, and the environmental complexity index remained at 0.3. The adaptive inertia weight of particle swarm optimization was reduced to 0.5, enhancing local convergence capability and driving the lighting current and transmission power towards a more refined search for lower power consumption. Ultimately, the near-light channel current was reduced from 200mA to 150mA, and the transmission power from 15dBm to 10dBm, saving battery power. During the optimization process, the gas concentration mode maintained a confidence level of 0.95 due to the good sensor condition, and the safety retention constraint forced its effective influence to be above the minimum threshold of 0.05, ensuring sensitivity to gas hazard perception. When the gas concentration at the working face increased to 0.6% VOL, although it did not reach the 1.0% VOL safety hard threshold, the system could still immediately trigger an early warning because the fusion weight and sensitivity index in the risk response calculation remained effective.

[0272] It will be understood by those skilled in the art that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments, and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A multi-modal perception and decision optimization method for an intelligent mine lamp system of a coal mine, characterized in that, Includes the following steps: S1. Data acquisition and modal state value construction for multimodal monitoring of mine lamps: Collect multimodal monitoring data of the target mine lamp, and convert sensor outputs of different dimensions and data forms into modal state values ​​for subsequent reliability assessment and optimization calculation; S2. Multimodal reliability assessment and multi-objective cost function construction for mine lamp operation and maintenance: By utilizing the spatial consistency of adjacent miner lamps, the long-term drift characteristics of the target miner lamp, and the short-term operating condition fluctuation characteristics, modal reliability, environmental complexity indicators, and multi-objective cost functions are constructed, enabling the subsequent optimization process to distinguish between sensor degradation and changes in the actual underground environment. S3. Two-layer particle swarm optimization driven by credibility and operational complexity: By embedding the environmental complexity index into the inertia weight adjustment process and the modal credibility into the velocity update process, the particle swarm can enhance global search under complex conditions, enhance local convergence under stable conditions, and perform targeted perturbation on decision variables related to low credibility modes. S4, Mining Lamp Online Adaptive Operation and Maintenance Execution: After completing modal reliability assessment and particle swarm optimization, the system periodically updates the operation and maintenance strategy according to the same data acquisition, reliability assessment and constraint correction rules as in the optimization phase, and the target mine lamp directly applies the optimal decision vector to the underground online operation.

2. The multi-modal perception and decision optimization method of the intelligent mine lamp system of the coal mine according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11. Arrange gas sensors, temperature and humidity sensors, dust sensors, inertial measurement sensors, video sensors, light-emitting diode lighting drive sampling circuits and wireless communication status sampling units on the target mine lamp, and collect raw mine lamp multimodal monitoring data according to a fixed sampling period. S12. Multimodal monitoring data of the original mine lamp Perform time alignment and validity checks to obtain multimodal monitoring data of synchronized mining lamps; S13, convert the multi-modal monitoring data of the synchronous mining lamp into modal state values to obtain the modal state value of the mth modal in the nth sampling period ; wherein, indicates a modal index, and takes a value of , indicates the total number of modes participating in the credibility evaluation and optimization calculation, refers to the scalar state value of the mth modal after preprocessing in the nth sampling period .​​​ S14. Arrange the modal state values ​​within the continuous sampling period into a target miner lamp modal state sequence in chronological order. , This refers to a data sequence consisting of all modal state values ​​across multiple sampling periods, used for subsequent calculations of modal reliability, environmental complexity indices, and multi-objective cost functions. It is a data sequence with dimensions of [missing information - likely a typology or specific dimension]. The matrix; where the first... The row corresponds to the first The sampling period, the first Column corresponding to the first One modality.

3. The multimodal perception and decision optimization method for a coal mine intelligent mine lamp system according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21. Modal confidence calculation based on spatial consistency and long-term drift characteristics: Combine the spatial reference value of the neighboring miner lamp with the long-term drift of the target miner lamp to obtain the confidence of each mode, so as to quantify the confidence of each mode in the target miner lamp not being affected by contamination, drift or sensitivity degradation in the current sampling period. S22. Calculation of environmental complexity index based on credible modal fluctuations and video action entropy: The environmental complexity index is obtained by weighting multimodal short-term fluctuations by modal credibility and combining them with video action entropy. S23. Construction of a multi-objective cost function for security risks, lighting power consumption and communication vulnerability: The multi-modal fusion weights, alarm baseline values, sensitivity index, LED channel drive current and wireless communication transmission power are uniformly encoded into a decision vector, and a multi-objective cost function is constructed.

4. The multimodal perception and decision optimization method for a coal mine intelligent mine lamp system according to claim 3, characterized in that, Step S21 specifically includes the following steps: S211, in the Within each sampling period, the target miner lamp obtains the identification of neighboring miner lamps, received signal strength, communication distance, and same-mode state value through wireless broadcasting or underground gateway, and establishes a set of neighboring miner lamps based on the communication distance and received signal strength; S212, Collect from nearby miners' lamps Extract the first The modal state values ​​of the neighboring miner lamps for the mode are calculated, and the ... Within the sampling period, the first Spatial reference values ​​for each mode and spatial discrete scale ; S213, Calculate the target miner's lamp number within a long-term sliding window. The relative deviation sequence of the nth mode relative to the spatial reference value, and the estimation of the nth mode based on the relative deviation sequence. Within the sampling period, the first Long-term drift of each mode ; S214, According to the target miner's lamp Calculation of spatial deviation and long-term drift of the first mode. Within the sampling period, the first Modal reliability of each mode ; S215, the first The modal confidence scores within each sampling period are combined in modal index order to obtain the modal confidence score vector. .

5. The multimodal perception and decision optimization method for a coal mine intelligent mine lamp system according to claim 4, characterized in that, Step S22 specifically includes the following steps: S221, with the first The sampling period is the current sampling period, and the sampling is extracted within the short-time sliding window. The historical modal state values ​​of the mode are calculated, and the ... Within the sampling period, the first Short-time mean of each modality and short-term standard deviation ; S222. Based on the deviation of the current modal state value from the short-time mean, and in conjunction with the modal reliability, calculate the first... Multimodal reliable fluctuation intensity within each sampling period ; S223. Extract the person's action state from the video frame data and calculate the first... Video motion entropy within a sampling period ; S224, Combining Multimodal Reliable Fluctuation Intensity and video motion entropy , obtained the Environmental complexity index within each sampling period .

6. The multimodal perception and decision optimization method for a coal mine intelligent mine lamp system according to claim 5, characterized in that, Step S23 specifically includes the following steps: S231, Constructing the first Decision vector within each sampling period ; S232, according to the... Modal state values Alarm baseline value and sensitivity index Calculate the first Risk response for each modality, combined with modality fusion weights and modal credibility Obtain the credible risk impact value; S233. Calculate the lighting power consumption based on the driving current of each LED channel. S234. Based on wireless communication transmission power Calculate the communication vulnerability item based on the current link state; S235. The trusted security risk item, lighting power consumption item, and communication vulnerability item are weighted and summed to obtain the... Multi-objective cost function within each sampling period .

7. The multimodal perception and decision optimization method for a coal mine intelligent mine lamp system according to claim 1, characterized in that, Step S3, in detail Includes the following steps: S31, Particle coding and constraint-aware initialization; An initial population satisfying boundary constraints is generated using chaotic sequences, and the initialization coverage is expanded to include low-confidence mode-related dimensions by combining modal confidence. S32. Heterogeneous velocity update that integrates environmental complexity and modal credibility: Differentiated velocity update is achieved by driving Cauchy mutation through adaptive inertia weights and credibility. S33, Constraint Projection, Fitness Update and Optimal Decision Output: Boundary truncation, weight normalization and safety retention constraint correction are performed on different types of decision variables to ensure that the optimization results can be directly deployed to the mining lamp embedded system.

8. The multimodal perception and decision optimization method for a coal mine intelligent mine lamp system according to claim 7, characterized in that, Step S31 specifically includes the following steps: S311, Decision vector As a particle position vector, and maintaining a fixed encoding order; S312. Set physical feasible boundaries for each decision dimension and establish constraint rules according to variable types; S313. Generate a Tent chaotic sequence value for each particle and map the Tent chaotic sequence value to the physical feasible boundary of each decision dimension. S314. Perform constraint correction on the initial particle positions to ensure that each initial particle satisfies the deployable conditions. S315. Divide the particle swarm into a perception strategy subgroup and a resource scheduling subgroup, and make the two subgroups share the group's optimal position.

9. A multimodal perception and decision optimization method for a coal mine intelligent mine lamp system according to claim 8, characterized in that, Step S32 specifically includes the following steps: S321. Based on environmental complexity indicators Calculate the first Within the sampling period, the first Adaptive inertia weights of the generation ; S322. Establish the mapping relationship between decision dimensions and modalities. , used to determine the first Does each decision dimension correspond to a certain modality? S323. Perform velocity updates for each particle and each decision dimension to obtain the... Sub-particle velocity; S324. Based on modal mapping Control the effectiveness of confidence-driven variants; S325. Limit the updated speed to obtain a controllable speed; Step S33 specifically includes the following steps: S331. Calculate the temporary position of the particle based on the updated velocity; S332. Perform boundary truncation and weight normalization on the temporary locations to obtain preliminary feasible locations; S333, Perform safety retention constraint corrections on gas safety-related modes; S334. Utilizing a multi-objective cost function Calculate the fitness value for each particle; S335. Update the individual's historical best position and the group's best position based on the fitness value; S336. Determine whether the iteration termination condition is met, and output the iteration term. The optimal decision vector corresponding to each sampling period .

10. A multimodal perception and decision optimization method for a coal mine intelligent mine lamp system according to claim 6, characterized in that, Step S4 specifically includes the following steps: S41, the target miner's lamp is in Raw mine lamp multimodal monitoring data were collected in one sampling period. And obtain the modal state values ​​according to step S1. ; S42, Target miner's lamp is based on the set of nearby miner's lamps. Calculate the modal confidence vector And calculate the environment complexity index based on the short-time sliding window. ; S43. Construct a multi-objective cost function for the current sampling period using the target miner's lamp. Then, step S3 is executed to obtain the optimal decision vector. ; S44, The target miner's lamp is determined according to the optimal decision vector. Update the modal fusion weights, alarm baseline values, sensitivity index, LED channel drive current, and wireless communication transmit power; S45. The target miner's lamp performs safety hard threshold judgment and continuous confirmation judgment, and outputs alarm result or low confidence prompt. S46, the target miner's lamp will be the first Modal state values ​​and modal confidence vector for each sampling period Environmental complexity index Optimal decision vector The alarm results are written to the operation and maintenance log and then proceed to the next step. One sampling period.

Images