Method and system for intelligently adjusting lubrication state of coal mine equipment based on deep learning

Abstract

The invention relates to the technical field of operation and maintenance control of coal mine equipment, in particular to an intelligent adjustment method and system for the lubrication state of the coal mine equipment based on deep learning, and the method comprises the steps: firstly collecting pressure difference signals of an oil supply pipeline and an oil return pipeline, and extracting time domain non-stationary fluctuation and frequency domain energy distribution characteristics through a dual-channel feature fusion network; generating an oil film thickness estimation value and a pollution degree index, when the oil film thickness is lower than a dynamic maintenance threshold value, dynamically generating the duration and interval of oil injection pulses through a reinforcement learning strategy network of a double-branch competitive architecture in combination with an equipment wear coefficient and historical oil injection feedback, and triggering online self-cleaning circulation when the pollution degree exceeds the limit; pollutants are removed through the high-gradient magnetic filter element, three-section type high-flow-speed washing and magnetic pulse desorption, a pollution degree base line is updated, the lubrication adjusting accuracy is improved, and stable operation of equipment is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of coal mine equipment operation and maintenance control technology, and more specifically, to a method and system for intelligent adjustment of lubrication status of coal mine equipment based on deep learning. Background Technology

[0002] Operation and maintenance control technology for coal mine equipment is an important technology. In coal mine production, the lubrication status of equipment directly affects operational stability and service life. Effective lubrication adjustment is the core means to reduce mechanical wear, lower downtime rates, and ensure continuous production.

[0003] However, traditional lubrication regulation technology for coal mine equipment suffers from core problems: incomplete state perception and a lack of adaptability in control strategies. Existing solutions judge lubrication status solely based on a single pressure signal, failing to deeply analyze the oil film thickness and contamination characteristics implicit in the pressure difference signal. This makes it impossible to capture instantaneous changes in oil film load corresponding to pressure surges, and also difficult to distinguish between high-frequency noise and characteristic frequency bands reflecting contaminant particle concentration. This leads to misjudgments of the actual oil film condition; when the oil film is too thin, timely oil replenishment is not possible, or when contamination exceeds standards, cleaning is not triggered, directly causing insufficient lubrication or accelerated oil deterioration. Furthermore, traditional oil injection uses fixed parameter modes, failing to dynamically adjust based on the wear coefficient formed by the equipment's cumulative operating time and historical oil injection effects. When equipment load or operating conditions change, the injection... The duration and interval of the oil pulse cannot match the oil film decay rate, resulting in either excessive oil injection leading to increased energy consumption or untimely oil replenishment exacerbating mechanical wear. During the cleaning process, the ferromagnetic particles adsorbed by the filter element are difficult to completely desorb, and the fixed baseline for judging the contamination level causes a delay in the triggering of the cleaning cycle, further worsening the oil contamination state. These problems combined lead to the equipment being in a suboptimal lubrication state for a long time, accelerating the wear of mechanical parts, increasing the frequency of downtime due to malfunctions, not only increasing maintenance costs, but also potentially affecting the safety of coal mine production due to sudden failures of critical equipment. It is difficult to meet the demand for precise and intelligent lubrication control under complex underground working conditions. To solve this technical problem, we provide a method and system for intelligent adjustment of lubrication status of coal mine equipment based on deep learning. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for intelligent adjustment of lubrication status of coal mine equipment based on deep learning, so as to solve the problems mentioned in the background art.

[0005] 1. Because traditional state perception is one-sided, relying solely on a single pressure signal can lead to misjudgments of oil film state and contamination level, resulting in insufficient lubrication or oil deterioration. Therefore, this case uses a dual-channel feature fusion network to extract the time and frequency domain features of the pressure difference signal, outputting an estimated oil film thickness and contamination level index, which can accurately perceive the lubrication state and avoid misjudgments.

[0006] 2. Because traditional control strategies lack adaptability, fixed oil injection parameters and incomplete cleaning exacerbate wear or increase energy consumption, this case study uses a reinforcement learning strategy network to dynamically adjust the oil injection pulse parameters. Combined with a high-gradient magnetic filter and a magnetic pulse cleaning module, it achieves efficient cleaning, which can match actual needs and reduce wear and energy consumption.

[0007] To achieve the above objectives, one of the objectives of this invention is to provide a method for intelligent adjustment of lubrication status of coal mine equipment based on deep learning, comprising the following steps: S1. Real-time acquisition of pressure difference signals between the oil supply line and the return line of the lubrication system, extraction of pressure fluctuation characteristics through a pre-trained deep learning model, and output of the current lubricating oil film thickness estimate and lubricating oil contamination index. S2. When the estimated value of the lubricating oil film thickness is lower than the dynamic maintenance threshold, the oil injection pulse is activated. The duration and pulse interval of the oil injection pulse are dynamically generated by the reinforcement learning strategy network. The input of the reinforcement learning strategy network is the current estimated value of the lubricating oil film thickness, the cumulative wear coefficient of the equipment running time, and the feedback of historical oil injection effectiveness. S3. When the lubricating oil contamination index exceeds the downhole working condition adaptive threshold, the online self-cleaning cycle is triggered. The online self-cleaning cycle includes: switching the bypass oil valve to allow the lubricating oil to flow through the high gradient magnetic filter element, and controlling the oil pump to enter a short-term high flow rate flushing mode for a preset period. After cleaning is completed, the system is reset to the normal oil injection mode and the contamination baseline is updated.

[0008] The second objective of this invention is to provide a system for implementing a deep learning-based intelligent adjustment method for the lubrication status of coal mine equipment, including any of the above-mentioned methods, comprising: The oil film status sensing unit integrates differential pressure sensors for the oil supply and return pipelines with an embedded dual-channel feature fusion module. It extracts temporal abrupt change features through one-dimensional dilatational convolution and separates feature frequency bands through wavelet packet decomposition, and outputs an estimated oil film thickness and a contamination index. The oil injection decision unit deploys a dual-branch competitive architecture reinforcement learning network, inputs oil film thickness, wear coefficient and oil injection effectiveness feedback, dynamically generates oil injection pulse parameters, and connects to a pulse width modulation controller to perform linear interpolation duration calculation and sliding window interval adjustment; The self-cleaning collaborative execution unit includes an axial gradient magnetic field filter element, a three-stage variable frequency flushing module, and a magnetic pulse desorption mechanism. After the bypass oil valve is switched, the cleaning cycle is triggered, and the baseline value of the contamination degree of the sliding window is updated when the cycle is completed.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: In terms of oil injection adjustment, the duration and interval of the oil injection pulse are dynamically generated by the reinforcement learning strategy network. Combined with the cumulative wear coefficient of the equipment during operation and the feedback of historical oil injection effectiveness, the oil injection parameters can be accurately matched with the oil film decay rate and the actual working conditions of the equipment. When the oil film thickness is lower than the threshold, the pulse parameters can be flexibly adjusted according to the deviation. This avoids the problem of excessive or insufficient oil injection caused by fixed parameters, reduces lubricating oil waste, and replenishes the oil film in time, effectively suppressing mechanical wear and improving the stability of equipment operation. In terms of cleaning control, when the online self-cleaning cycle is triggered, the axial gradient magnetic field and swirling effect of the high-gradient magnetic filter element enhance the adsorption of ferromagnetic particles. Combined with short-time high-flow-rate three-stage flushing and magnetic pulse desorption, it can thoroughly remove the pollutants adsorbed by the filter element. At the same time, it dynamically updates the contamination baseline to ensure that the timing of the cleaning cycle is accurate, avoids the deterioration of oil contamination, extends the service life of the oil, reduces the risk of equipment failure due to contamination, and ensures the continuity of coal mine production. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the overall workflow of the present invention; Figure 2 This is a schematic diagram of the overall structure of the present invention; The meanings of the labels in the diagram are as follows: 1. Oil film status sensing unit; 2. Oil injection decision unit; 3. Self-cleaning collaborative execution unit. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Please see Figure 1 As shown, one of the objectives of this embodiment is to provide a method for intelligent adjustment of lubrication status of coal mine equipment based on deep learning. The method is characterized by: sensing the state of the lubricating oil film and adaptively adjusting oil injection and cleaning behaviors, including the following steps: S1. Real-time acquisition of pressure difference signals between the oil supply line and the return line of the lubrication system, extraction of pressure fluctuation characteristics through a pre-trained deep learning model, and output of the current lubricating oil film thickness estimate and lubricating oil contamination index. To accurately analyze the oil film state and contamination information hidden in the pressure difference signal of the lubrication system, a deep learning model employs a dual-channel feature fusion network. Through the collaborative extraction and fusion of time-domain and frequency-domain features, it achieves accurate assessment of oil film thickness and contamination level. The specific implementation method is as follows: The deep learning model is a dual-channel feature fusion network. The first channel extracts the time-domain non-stationary fluctuation features of the pressure difference signal between the supply and return oil pipelines through one-dimensional dilated convolution. Based on the time-domain non-stationary fluctuation features, it captures the time-domain feature vector, including the instantaneous load-bearing state changes of the oil film corresponding to pressure abrupt changes. The first channel focuses on the variation law of the pressure difference signal over time. First, the real-time acquired pressure difference signals from the supply and return oil pipelines are preprocessed to remove baseline drift, such as overall pressure shifts caused by slow temperature changes, while retaining the true pressure fluctuation components. The one-dimensional dilated convolution, by setting intervals in the convolution kernel (e.g., with an expansion rate of 2, the kernel elements are spaced one data point apart), can achieve this without increasing parameters. In the case of a large number of cases, the receptive field is expanded to capture the fluctuation characteristics at different time scales. For example, a small expansion rate convolution kernel extracts high-frequency small-amplitude fluctuations, such as the pressure change caused by the instantaneous rupture of the oil film, while a large expansion rate convolution kernel extracts low-frequency trend changes, such as the pressure decreases slowly as the oil film gradually thins. Based on the extracted time-domain non-stationary fluctuation characteristics, the time-domain feature vector is further captured, focusing on pressure change points: when the equipment load suddenly increases or the oil film ruptures locally, the pressure difference will change instantaneously, such as the pressure difference rising from 0.5 MPa to 1.2 MPa within 0.1 seconds. These change points directly correspond to the change in the instantaneous load-bearing state of the oil film. The time-domain feature vector integrates the time position, amplitude change and fluctuation trend of these change points into quantitative data to form a time-domain feature set that reflects the dynamic load-bearing capacity of the oil film. The second channel reconstructs the frequency domain energy distribution of the pressure difference signal and separates the frequency domain feature vectors through wavelet packet decomposition. Wavelet packet decomposition can decompose the signal into multiple sub-bands of different frequencies. The energy of each sub-band reflects the signal strength within that frequency range. Based on the frequency domain energy distribution, frequency domain feature vectors are separated and integrated into a frequency domain feature vector to achieve a frequency dimension description of oil film thickness and contamination status.

[0013] The temporal and frequency domain feature vectors are input into the feature fusion gating module of the dual-channel feature fusion network. A fused feature vector is generated by cross-channel attention weight allocation. The attention weights are dynamically adjusted according to the importance of the features. Through this dynamic allocation, a fused feature vector is generated, which retains the instantaneous change information in the temporal domain and integrates the stable feature patterns in the frequency domain. Finally, it is mapped to the lubricating oil film thickness estimate and lubricating oil contamination index through a fully connected layer. The contamination index is associated with the proportion of pulses with energy higher than the benchmark energy threshold in the feature frequency band. The calculation of the contamination index is associated with the proportion of pulses with energy higher than the benchmark energy threshold in the feature frequency band that reflects the concentration of pollutant particles. For example, if there are 100 energy pulses in the frequency band, and 60 of them exceed the benchmark threshold, the contamination index is generated based on this proportion (60%). The higher the proportion, the larger the contamination index, which intuitively reflects the density of pollutants in the oil.

[0014] S2. When the estimated value of the lubricating oil film thickness is lower than the dynamic maintenance threshold, the oil injection pulse is activated. The duration and pulse interval of the oil injection pulse are dynamically generated by the reinforcement learning strategy network. The input of the reinforcement learning strategy network is the current estimated value of the lubricating oil film thickness, the cumulative wear coefficient of the equipment running time, and the feedback of historical oil injection effectiveness. To dynamically optimize the oil injection pulse parameters based on the real-time lubrication status of coal mine equipment, a reinforcement learning strategy network adopts a dual-branch competitive architecture. Through collaborative decision-making of state evaluation and action optimization, it generates an oil injection scheme adapted to the actual operating conditions of the equipment. The specific implementation method is as follows: The dual-branch competitive architecture of the reinforcement learning policy network includes a state value branch and an action advantage branch. The two branches work together to optimize decision-making. The state value branch evaluates the current lubrication state of the equipment by analyzing parameters such as oil film thickness and wear degree, and outputs a comprehensive score. The higher the score, the better the current lubrication state. For example, if the oil film thickness is sufficient and the wear rate is slow, this score serves as the basis for subsequent action selection to ensure that the decision is in line with the overall operating goals of the equipment. The action advantage branch evaluates the optimization potential of different oil injection pulse parameters, that is, the degree of improvement in lubrication state after adopting the parameter. The higher the advantage value of the parameter, the more suitable it is for the current working condition.

[0015] The network input layer receives three types of key parameters, providing a data foundation for the two-branch decision-making process. These parameters include the estimated lubricating oil film thickness, the cumulative wear coefficient over equipment runtime, and historical lubrication effectiveness feedback. The estimated lubricating oil film thickness is directly taken from the output of the dual-channel feature fusion network, reflecting the actual thickness of the current oil film and serving as the core basis for determining whether lubrication is needed. The cumulative wear coefficient over equipment runtime is generated through a nonlinear function of the equipment's cumulative runtime and a preset lifespan curve. The preset lifespan curve is set according to the equipment's factory standards, while the nonlinear function is dynamically adjusted based on actual operating conditions. For example, for equipment that has accumulated 3000 hours of operation and has been running under high load for a long period, the wear coefficient may increase from a baseline value of 1.0 to 1.5, directly reflecting the change in lubrication requirements caused by wear. The more severe the wear, the more lubrication is required. The historical lubrication effectiveness feedback is defined as the feedback from the first three lubrication cycles. The mean deviation between the post-oil film thickness recovery rate and the expected recovery rate is calculated. For example, if the expected recovery rate for the first oil injection is 80% but the actual recovery rate is 70% (deviation -10%), the expected recovery rate for the second injection is 80% but the actual recovery rate is 90% (deviation +10%), and the expected recovery rate for the third injection is 80% but the actual recovery rate is 75% (deviation -5%), then the mean deviation is (-10% + 10% - 5%) / 3 ≈ -1.7%. The smaller the feedback value, the higher the matching degree between the historical oil injection parameters and the actual demand, providing an empirical reference for the generation of new parameters. The input layer standardizes the three types of parameters, eliminates the difference in dimensions, and simultaneously inputs the state value branch and the action advantage branch. Based on the evaluation results of the two branches, the output layer generates a discrete action space for the duration and interval of the oil injection pulse. The discrete action space contains multiple sets of preset parameter combinations, each corresponding to a specific action. The generation process follows the principle of "state value guidance + action advantage screening". After the State Value branch outputs the current state score, if the score is lower than the threshold, it indicates that the lubrication state needs to be significantly improved. The Action Advantage branch prioritizes the selection of parameter combinations with high advantage values. If the status score is high (e.g., 8 points), then the parameter combination with a moderate advantage value is selected first. For example, if the current oil film thickness is low (15 micrometers), the wear coefficient is 1.3 (moderate wear), and the historical feedback deviation average is -2% (the previous lubrication effect was good), the status value branch score is 4 points (needs improvement). The action advantage branch evaluates the parameter combination of "2 seconds of continuous operation + 6 seconds of interval" as having the highest advantage value. Then, the output layer incorporates this combination into the discrete action space as the current optimal lubrication parameter, so that the generated lubrication pulse parameter can dynamically match the actual needs of the equipment, avoiding the problem of insufficient or excessive lubrication caused by traditional fixed parameters, and reducing energy consumption and wear while ensuring the lubrication effect.

[0016] To ensure that the duration and interval of the oil injection pulses precisely match the real-time lubrication needs of coal mining equipment, the duration of the oil injection pulses is dynamically adjusted by the pulse width modulation controller based on the action commands output by the reinforcement learning strategy network. Specifically: After receiving the action command output by the reinforcement learning strategy network, the pulse width modulation controller adjusts the duration of the oil injection pulse according to the deviation between the estimated oil film thickness and the dynamic maintenance threshold, ensuring that the oil replenishment rate matches the degree of oil film damage. When the estimated oil film thickness is lower than the dynamic maintenance threshold by more than a preset amplitude threshold, the maximum duration pulse is triggered to increase the oil replenishment rate. For example, if the dynamic maintenance threshold is 20 micrometers and the preset amplitude threshold is 8 micrometers, and the actual oil film thickness is 10 micrometers, the deviation is 10 micrometers, exceeding 8 micrometers, indicating severe oil film damage, requiring rapid oil replenishment to prevent further wear on the equipment. In this case, the controller triggers the maximum duration pulse, increasing the lubricant supply in a short time by extending the oil injection time, quickly improving the oil film thickness (e.g., 3-second oil injection can restore the oil film from 10 micrometers to 18 micrometers). When the amplitude is within the preset amplitude threshold range, the duration is dynamically calculated dynamically within the preset time range using linear interpolation (e.g., a deviation of 5 micrometers, within the 3-8 micrometer range), indicating that the oil film has some damage but does not require emergency oil replenishment. The controller uses linear interpolation to dynamically calculate the duration within a preset time range: the larger the deviation, the closer the duration is to 3 seconds; the smaller the deviation, the closer the duration is to 1 second. For example, a deviation of 5 micrometers is in the middle of the range (the middle value for 3-8 micrometers is 5.5 micrometers), so the duration is calculated to be 2 seconds through linear interpolation. This ensures that the oil film can gradually recover while avoiding over-injection and waste. The pulse interval uses a sliding window adaptive mechanism. The core is to predict the next injection time based on the oil film decay rate after the most recent preset number of injections, so that the interval time matches the oil film consumption rate. The sliding window size is set to the most recent 3 injection records, that is, the window contains 3 sets of data: "oil film thickness after injection - time to decay to threshold". For example: After the first oil injection, the oil film thickness recovered from 15 micrometers to 22 micrometers, and then decayed to 20 micrometers after 8 minutes (dynamic maintenance threshold). After the second oil injection, the oil film thickness recovered from 16 micrometers to 23 micrometers, and then decreased to 20 micrometers after 7 minutes; After the third oil injection, the oil film thickness recovered from 14 micrometers to 21 micrometers, and then decreased to 20 micrometers after 9 minutes.

[0017] The controller calculates the average decay time of these three cycles to be 8 minutes, using this average time as the base interval. It then fine-tunes the interval based on the current oil film thickness decay trend: if the current oil film decay rate is faster than the average rate within the window (e.g., the film has decreased from 22 micrometers to 21 micrometers in 5 minutes), the interval is shortened; if the decay rate is slower, the interval is appropriately extended. This dynamic adjustment ensures that each oil injection occurs just before the oil film falls below the threshold, avoiding frequent oil injections that increase energy consumption and preventing excessively thin oil films that lead to wear. This allows the oil injection operation to precisely match the actual needs of the oil film and adapt to changes in the oil film state during equipment operation, achieving energy saving and consumption reduction while ensuring lubrication effectiveness.

[0018] To ensure that the lubrication pulse parameters generated by the reinforcement learning policy network can guarantee lubrication effectiveness while reducing energy consumption and equipment wear, the training of the reinforcement learning policy network adopts a multi-objective reward function. The network's optimization decisions are guided by the synergistic effect of positive incentives and negative penalties. The specific implementation method is as follows: The dynamic maintenance threshold is the core reference benchmark of the reward function, and its value is dynamically updated based on the real-time rotational speed and load pressure of the equipment through a pre-calibrated three-dimensional mapping table. The construction process of the three-dimensional mapping table is as follows: Before the equipment leaves the factory, bench tests are conducted to simulate different combinations of speed and load pressure. The critical value of oil film thickness under each condition is measured, which is the thickness below which wear will be aggravated. For example, the threshold for "speed 500 rpm + load 10 MPa" is 20 micrometers, and the threshold for "speed 1000 rpm + load 20 MPa" is 25 micrometers. These data are stored in a three-dimensional relationship of "speed-load-threshold" to form a mapping table. During real-time updates, the equipment sensors continuously collect the current speed and load pressure, and the network matches the corresponding threshold from the mapping table. If the current speed is 800 rpm and the load is 15 MPa, the dynamic maintenance threshold of 22 micrometers is obtained by interpolation through table lookup, and a range of ±2 micrometers is allowed, that is, 20-24 micrometers is the normal range, providing a clear standard for judging the oil film maintenance bonus.

[0019] (a) Oil film maintenance reward: A positive reward is given when the oil film thickness remains within the allowable range of the dynamic maintenance threshold after oil injection. The oil film maintenance reward is used to incentivize the network to generate oil injection parameters that ensure the oil film remains stable within a reasonable range over the long term. In practice, the change in oil film thickness is continuously monitored after each oil injection. If the oil film thickness remains within the dynamic maintenance threshold range (e.g., 20-24 micrometers) for 10 minutes after oil injection, a positive reward (e.g., a reward value of 5) will be given. If the oil film thickness briefly meets the target after oiling but quickly exceeds the range (e.g., drops to 19 micrometers after 3 minutes), the bonus value will be reduced (e.g., the bonus value will be 2). If the oil film thickness consistently fails to reach the specified range, there will be no reward (reward value is 0). For example, after a certain oil injection, the oil film rises from 18 micrometers to 23 micrometers and stabilizes at 22-23 micrometers within 10 minutes, meeting the maintenance requirements and earning a 5-point reward. This guides the network to tend to choose oil injection parameters that can achieve long-term oil film stability.

[0020] (b) Energy Consumption Penalty: The negative energy reward is calculated by multiplying the oil injection duration by the system's rated power. This penalty is used to suppress the network from selecting parameters for excessive oil injection, thus avoiding energy waste. The calculation method is as follows: the energy consumed during the oil injection process is converted into a negative reward based on the product of the oil injection pulse duration and the system's rated power. The higher the energy consumption, the larger the penalty value (the larger the absolute value of the negative reward). If the oil injection duration is 3 seconds, the energy consumption per unit time corresponding to the system's rated power is 2 units / second, then the total energy consumption is 6 units, and the penalty value is -3. If the duration is 1 second, the total energy consumption is 2 units, and the penalty value is -1.

[0021] For example, the network generates two parameters, "lasts for 3 seconds" and "lasts for 1 second", under a certain state. Although the former can quickly restore the oil film, the energy consumption penalty is higher (-3), and the overall reward may be lower than the latter (e.g., the latter rewards 5-1=4, while the former rewards 5-3=2), prompting the network to prioritize the parameter with lower energy consumption.

[0022] (c) Wear Suppression Reward: An incremental reward is given when the rate of increase of the cumulative wear coefficient decreases. The dynamic maintenance threshold is dynamically updated based on the real-time speed and load pressure of the equipment through a pre-calibrated three-dimensional mapping table. The wear suppression reward is used to incentivize the network to generate lubrication parameters that can slow down equipment wear, which is achieved by monitoring the rate of increase of the cumulative wear coefficient. The cumulative wear coefficient is calculated based on the equipment's operating time and the preset life curve (e.g., the coefficient is 0.1 for a new machine running for 100 hours and 0.5 for running for 500 hours). The growth rate is the difference between two consecutive calculated coefficients (e.g., the growth rate from 0.5 to 0.6 is 0.1). If, after oiling, the wear coefficient growth rate decreases from 0.1 to 0.05 (a 50% decrease in growth rate), an incremental bonus (e.g., a bonus value of 4) will be given. If the growth rate does not change significantly (e.g., it remains at 0.1), the reward value is 1; If the growth rate increases (e.g., to 0.12), there will be no reward.

[0023] For example, if the wear coefficient of a certain piece of equipment increases by 0.08 before oiling, and the increase drops to 0.03 within 1 hour after oiling, which is a significant decrease, it will receive a reward of 4 points, guiding the network to pay attention to the effect of oiling parameters on inhibiting long-term wear of equipment.

[0024] During network training, the three reward values ​​for each decision are summed (e.g., oil film maintenance reward 5 + energy consumption penalty -1 + wear inhibition reward 4 = total reward 8). The higher the total reward value, the better the lubrication parameter. Through a large number of sample iterations, the network gradually learns the optimal parameter selection rules under different scenarios, guiding the reinforcement learning strategy network to generate optimal lubrication parameters that take into account multiple objectives, making lubrication adjustment both precise and economical, and extending the service life of equipment.

[0025] S3. When the lubricating oil contamination index exceeds the downhole working condition adaptive threshold, the online self-cleaning cycle is triggered. The online self-cleaning cycle includes: switching the bypass oil valve to allow the lubricating oil to flow through the high gradient magnetic filter element, and controlling the oil pump to enter a short-term high flow rate flushing mode for a preset period. After cleaning is completed, the system is reset to the normal oil injection mode and the contamination baseline is updated.

[0026] To efficiently adsorb ferromagnetic contaminants in lubricating oil, the high-gradient magnetic filter element adopts an axial gradient magnetic field design. This design enhances adsorption force through magnetic field gradient, optimizes oil flow path through a guiding structure, and improves adsorption efficiency through a coating. Simultaneously, it monitors the contaminant load status in real time. The specific implementation method is as follows: Four sets of NdFeB magnetic rings are arranged along the direction of lubricating oil flow (e.g., axially from the filter element inlet to the outlet). Each set of rings is a ring structure, fitted onto the outside of the filter element skeleton. Adjacent sets of rings are arranged with opposite polarities to form a magnetic field gradient. For example, the first set of rings is oriented with the N pole facing outwards and the S pole facing inwards, the second set with the S pole facing outwards and the N pole facing inwards, the third set again with the N pole facing outwards and the S pole facing inwards, and the fourth set with the S pole facing outwards and the N pole facing inwards. This reverse arrangement can create a continuously changing magnetic field gradient along the axial direction: from the inlet to the outlet, the magnetic field strength first increases and then decreases, or changes in a stepwise manner (e.g., the first set of magnetic field strength is 2000 Gauss, the second set increases to 3000 Gauss, the third set decreases to 2500 Gauss, and the fourth set decreases to 1500 Gauss). The presence of a magnetic field gradient causes ferromagnetic particles to experience a gradually changing magnetic force in the oil flow, avoiding the problem of weak particle adsorption caused by a uniform magnetic field. For example, small particles are initially adsorbed by a weaker magnetic field at the inlet, and are more firmly captured after entering the stronger magnetic field region with the oil flow, preventing the particles from falling off again. The skeleton surface of the high-gradient magnetic filter element is covered with honeycomb-shaped guide channels, which create a swirling effect when lubricating oil containing contaminants flows through it. These guide channels are distributed in a hexagonal grid, with the channel direction forming an angle of 30-45 degrees with the axial direction. When lubricating oil containing contaminants flows through the filter element, the guide channels guide the oil flow to generate a rotational motion, i.e., a swirling effect: the oil flow moves in a circle around the central axis of the filter element while advancing along the axial direction. The swirling flow allows ferromagnetic particles in the oil flow to fully contact different areas of the magnetic ring assembly. For example, particles originally located in the center of the oil flow are thrown towards the inner wall of the filter element by the swirling flow, contacting the magnetic field area on the surface of the magnetic ring assembly and thus being adsorbed. Particles close to the inner wall are carried by the swirling flow across more magnetic rings, increasing the probability of being captured. At the same time, the honeycomb structure can also disperse the oil flow, avoiding adsorption dead zones caused by excessively high local flow rates, ensuring uniform adsorption efficiency of the entire filter element. The four-stage NdFeB magnetic ring assembly is wrapped with an oleophilic and dust-repellent coating to improve the adsorption efficiency of ferromagnetic particles. This coating has the characteristics of having an affinity for lubricating oil, allowing the oil film to adhere evenly, but having a repulsive force on dust and particles. When the lubricating oil flows through the magnetic ring assembly, a thin oil film is formed on the surface of the coating. Ferromagnetic particles pass through the oil film under the action of magnetic field force and are adsorbed onto the surface of the magnetic rings. The dust-repellent properties of the coating can reduce the adhesion of non-magnetic impurities on the surface of the magnetic rings, preventing these impurities from blocking the magnetic field and reducing the adsorption force on the ferromagnetic particles.For example, after long-term use, the surface of an uncoated magnetic ring may accumulate a large amount of mixed impurities, resulting in a 30% decrease in magnetic field strength. In contrast, the coated magnetic ring surface primarily adsorbs ferromagnetic particles with fewer impurities, leading to a mere 10% decrease in magnetic field strength. This significantly improves adsorption stability during long-term use. The contamination load status of the high-gradient magnetic filter cartridge is monitored in real-time by a built-in Hall sensor that monitors the magnetic field strength attenuation rate. The sensor is installed on the outlet side of the fourth set of magnetic rings, continuously measuring the current magnetic field strength and comparing it with the initial magnetic field strength (e.g., 3000 Gauss at the factory) to calculate the attenuation rate (attenuation rate = (initial strength - current strength) / initial strength × 100%). When a large number of ferromagnetic particles are adsorbed on the magnetic ring surface, the magnetic field strength decreases; the more particles adsorbed, the higher the attenuation rate. For example, a new filter cartridge with a magnetic field strength of 3000 Gauss may drop to 2400 Gauss after a period of use, representing a 20% attenuation rate, indicating that the filter cartridge has adsorbed a certain amount of contaminants. When the attenuation rate reaches 50% (e.g., dropping to 1500 Gauss), it indicates that the filter cartridge is nearing saturation and a cleaning cycle needs to be triggered. Through this real-time monitoring, the system can accurately grasp the degree of filter contamination, avoiding resource waste caused by premature cleaning or filtration failure caused by late cleaning. It achieves efficient capture and status perception of ferromagnetic contaminants in lubricating oil, providing a precise basis for triggering online self-cleaning cycles and ensuring the long-term cleanliness and stability of the lubrication system.

[0027] To efficiently remove contaminants adsorbed on the high-gradient magnetic filter element during online self-cleaning circulation, while avoiding sudden changes in flow rate that could impact the lubrication system, a short-duration high-flow-rate flushing mode employs a three-stage flow rate modulation. The smooth transition and precise control of the flow rate are achieved by adjusting the oil pump motor speed via a frequency converter. The specific implementation method is as follows: The short-duration high-flow-rate flushing mode adjusts the flow rate in stages according to the sequence of "start-core-decay". The flow rate and duration of each stage are set according to the cleaning needs and system stability requirements. The overall logic is "smooth start-powerful flushing-buffered finish", which ensures the cleaning effect while protecting the pipeline and equipment. In the start-up stage, the oil flow is stabilized by maintaining the basic flow rate for n seconds. The core of the start-up stage is to allow the oil flow to smoothly transition from static or normal flow rate to high flow rate, avoiding pressure fluctuations caused by sudden acceleration. In actual operation, after the system switches the bypass oil valve to allow lubricating oil to enter the filter element pipeline, the frequency converter controls the oil pump motor to run at the base speed, corresponding to the base flow rate. This flow rate is slightly higher than the flow rate in the normal oil filling mode, but much lower than the subsequent peak flow rate. The duration of n seconds is set according to the pipeline length and oil flow inertia to ensure that the oil flow can fill the entire filter element pipeline within 5 seconds and the flow rate is stable. For example, if the total pipeline length is 10 meters and the base flow rate is 1 m / s, the oil flow can complete the flow from the inlet to the outlet within 5 seconds, and the flow rate fluctuation is controlled within ±0.1 m / s, laying a stable foundation for the subsequent high flow rate flushing. If the start-up phase is too short (e.g., 2 seconds), the oil flow may not stabilize before entering a high-velocity flow, potentially causing a sudden increase in localized pressure in the pipeline and leading to joint leaks. Therefore, a 5-second duration is a choice that balances efficiency and safety. The core flushing phase increases the flow rate to the peak value for a preset period, which is dynamically calculated based on the extent to which the contamination index exceeds the standard. The core flushing phase is crucial for removing contaminants. By increasing the flow rate to the peak value, a strong impact force is generated, washing away ferromagnetic particles (especially stubbornly attached particles) adsorbed on the filter element. The frequency converter controls the oil pump motor to reach its peak speed, corresponding to a peak flow rate (e.g., 3 m / s, three times the base flow rate). This flow rate creates turbulence on the filter element surface, and the resulting shear force can overcome the adsorption force between the particles and the magnetic ring. The preset period is dynamically calculated based on the extent to which the contamination index exceeds the standard. If the pollution index slightly exceeds the threshold, it indicates that there are few pollutants. The preset cycle is set to 10 seconds, and the pollutants can be removed by a high flow rate in a short period of time. If the exceedance is significant (e.g., 50%), it indicates a large amount of pollutants that may be deeply attached. The preset cycle is extended to 20 seconds to ensure sufficient time to flush away stubborn particles.

[0028] For example, if the contamination index of a certain filter element exceeds the standard by 40%, the system calculates that the preset cycle is 15 seconds. Within 15 seconds, the oil flow at the peak velocity continuously washes the filter element, and the ferromagnetic particles adsorbed on the surface of the magnetic ring gradually fall off under the impact of turbulence and flow with the oil flow to the subsequent sedimentation chamber.

[0029] The attenuation phase reduces the flow rate to the base velocity for m seconds to avoid pressure surges. Flow rate control is achieved by adjusting the oil pump motor speed via a frequency converter. The purpose of the attenuation phase is to smoothly reduce the oil flow rate from the peak velocity to the normal flow rate, avoiding sudden deceleration that could cause a sharp drop in pressure (potentially creating negative pressure and drawing in air). The frequency converter controls the oil pump motor to gradually reduce from the peak speed to the base speed, corresponding to a slow decrease in flow rate from 3 m / s to 1 m / s over a period of m seconds (e.g., m=3 seconds). This ensures that the rate of change in flow rate is controlled within 1 m / s² (i.e., a decrease of 0.67 m / s per second). During this 3-second duration, the pressure of the oil flow in the pipeline gradually decreases from the peak pressure (e.g., 3 MPa) to the base pressure (e.g., 1.2 MPa), preventing sudden pressure changes from damaging the filter element and pipeline joints. For example, if the flow rate were to stop abruptly from the peak velocity, the pressure might drop instantly to 0.5 MPa, creating negative pressure and causing the pipeline to collapse. The buffering effect of the attenuation phase allows for a smooth pressure transition, eventually restoring the pressure level to the normal oil injection mode.

[0030] The entire three-stage flow rate modulation is achieved by adjusting the oil pump motor speed in real time through a frequency converter: the motor speed is directly proportional to the flow rate (the higher the speed, the faster the flow rate). The frequency converter receives speed commands from the system (e.g., 500 rpm for the start-up phase, 1500 rpm for the core phase, and gradually decreasing from 1500 rpm to 500 rpm for the decay phase). By changing the power supply frequency, the frequency converter precisely controls the motor speed, thereby achieving dynamic adjustment of the flow rate. For example, the start-up phase command is 500 rpm, and the motor runs stably for 5 seconds; the core phase command rises to 1500 rpm and lasts for 15 seconds; the decay phase command linearly decreases from 1500 rpm to 500 rpm within 3 seconds. The flow rate changes smoothly throughout the process without significant fluctuations, allowing contaminants on the filter element to be efficiently removed while minimizing the impact on the lubrication system, creating favorable conditions for subsequent magnetic pulse desorption.

[0031] To thoroughly remove ferromagnetic particles adsorbed on the high-gradient magnetic filter element through short-duration high-flow-rate flushing, the magnetic pulse cleaning module is activated simultaneously with switching the bypass oil valve to start the self-cleaning cycle. Through the synergistic effect of alternating magnetic field pulses and high-flow-rate flushing, efficient desorption and collection of pollutants are achieved. The specific implementation method is as follows: When the system detects that the lubricating oil contamination index exceeds the threshold, it switches the bypass oil valve to allow lubricating oil to enter the filter element cleaning pipeline. At the same time as switching the bypass oil valve, the magnetic pulse cleaning module is activated. The magnetic pulse cleaning module applies a sequence of alternating magnetic field pulses with gradually increasing amplitude to the high-gradient magnetic filter element, causing the adsorbed ferromagnetic particles to vibrate slightly and detach from the magnetic poles. The detached particles are flushed by the oil flow under short-term high flow rate and enter the sedimentation chamber. The electromagnetic discharge valve at the bottom of the sedimentation chamber automatically opens to discharge sewage at the end of the cleaning cycle, and the magnetic pulse cleaning module is immediately activated. This module includes a pulse generator and an electromagnetic coil wound around the outside of a high-gradient magnetic filter. The pulse generator generates an alternating magnetic field pulse sequence with gradually increasing amplitude, meaning the magnetic field strength gradually increases from low to high, and the direction changes alternately (e.g., first increasing in the positive direction, then increasing in the negative direction, and repeating). For example, the initial amplitude of the pulse sequence is set to generate a magnetic field of 1000 Gauss, and then the amplitude of each pulse increases by 500 Gauss, successively to 1000 Gauss, 1500 Gauss, 2000 Gauss, and so on, until the preset maximum amplitude (e.g., 3000 Gauss) is reached. The duration of each pulse is fixed at 0.5 seconds, and the interval between adjacent pulses is 0.2 seconds, forming a progressive mode of "low amplitude short pulse - high amplitude long pulse". This gradually increasing alternating magnetic field causes the ferromagnetic particles adsorbed on the filter element to be subjected to a gradually increasing alternating magnetic force. First, loose small particles are detached. Then, a higher amplitude pulse causes stubbornly attached large particles to vibrate slightly, preventing damage to the filter element surface caused by forcibly pulling the particles due to an excessively high initial amplitude. When the alternating magnetic field pulse acts on the high-gradient magnetic filter element, the ferromagnetic particles inside the filter element repeatedly change their magnetization direction as the magnetic field direction changes. Tiny repulsive and attractive forces occur between the particles and between the particles and the filter element surface, forming high-frequency micro-vibrations (e.g., 50 vibrations per second). This vibration breaks down the static friction and magnetic attraction between the particles and the filter element, causing the particles to gradually loosen and detach from the magnetic pole surface. Simultaneously, a short-duration high-velocity flush (e.g., a peak flow rate of 3 m / s) is running, and the detached particles are quickly carried away from the filter element area by the high-speed oil flow, preventing the particles from re-adsorbing onto the filter element.For example, a particle that was originally firmly adsorbed on the second-stage magnetic ring of the filter element by magnetic force will vibrate slightly under the action of a 1500 Gauss alternating pulse. Combined with the impact force of the high-speed oil flow, the particle will detach from the magnetic ring within 0.3 seconds and move downstream with the oil flow. This ensures that the desorption effect is not negated by the particle's re-attachment. The particle that detaches from the filter element enters the sedimentation chamber located downstream of the filter element with the oil flow. This chamber has a large volume (e.g., it can hold 10 liters of lubricating oil) and is equipped with a guide plate inside, which makes the flow rate of the high-speed oil flow drop sharply (from 3 m / s to 0.2 m / s). Gravity will be used to allow the particles to settle. Large particles (e.g., those with a diameter of 50 micrometers or more) will settle to the bottom of the chamber within 10 seconds, while small particles (e.g., 20-50 micrometers) will gradually settle within 30 seconds, preventing the particles from re-entering the lubrication system with the oil flow. After the cleaning cycle is completed, that is, the three-stage high-flow-rate flushing is completed, and the system is ready to reset to the normal oil injection mode, the electromagnetic drain valve at the bottom of the sedimentation chamber will automatically open. The electromagnetic discharge valve is triggered by a command from the control system, and its opening time is set according to the amount of sediment. It discharges the particulate-containing oily waste at the bottom of the chamber to the collection container. For example, if the cleaning cycle lasts for 30 seconds, the discharge valve will open for 5 seconds after the cycle ends to discharge about 2 liters of oily waste, and then close to ensure that the residual cleaning lubricating oil in the chamber can re-enter the cycle. The working duration of the magnetic pulse cleaning module is consistent with the core phase of the short-term high-flow-rate flushing: the pulse sequence begins to output when the start-up phase ends and the core flushing phase begins; the pulse sequence stops when the core flushing phase ends and the decay phase begins, avoiding unnecessary energy consumption. For example, the core flushing phase lasts for 15 seconds, during which the pulse sequence outputs 10 sets of pulses with gradually increasing amplitude. This forms a complete chain of "vibration desorption - flushing removal - sedimentation collection" with the high-flow-rate flushing. After the entire cleaning cycle is completed, the magnetic pulse cleaning module is turned off, the bypass oil valve is switched back to the normal position, and the system resumes normal oil injection mode. At this time, the high-gradient magnetic filter element has removed most of the contaminants, the magnetic field strength is restored, and it can re-adsorb new particles efficiently, significantly improving the effect of online self-cleaning cycle and extending the service life of lubricating oil.

[0032] To ensure the contamination baseline dynamically adapts to changes in lubricating oil quality during long-term equipment operation and guarantees the accuracy of self-cleaning cycle triggering, a sliding window weighted mechanism is used for contamination baseline updates. This mechanism achieves reasonable baseline adjustments through dynamic analysis and weighted fusion of historical cleaning data. The specific implementation method is as follows: The average contamination index at the moment the most recent k self-cleaning cycles are completed is recorded as the new baseline reference value. When the deviation between the new baseline reference value and the historical baseline reference value is less than a preset deviation threshold for p consecutive cycles, the original baseline is maintained. When the deviation exceeds the preset deviation threshold, the new baseline reference value and the historical baseline reference value are merged and updated according to a preset weight ratio. The updated baseline reference value is used as the comparison reference value for the next self-cleaning cycle.

[0033] The core of the sliding window weighted mechanism is to continuously track the effects of the most recent self-cleaning cycles to determine whether the contamination baseline needs to be updated, thus avoiding a disconnect between the baseline and the actual situation caused by long-term factors such as equipment aging and oil deterioration. Key parameters include: window size k (e.g., k=5, i.e., tracking data from the most recent 5 self-cleaning cycles), number of consecutive verifications p (e.g., p=3, i.e., maintaining the original baseline only after 3 consecutive deviations meet the standard), preset deviation threshold (e.g., 5%, i.e., the difference between the new and old baselines within 5% is considered acceptable), preset weight ratio (e.g., the new baseline value accounts for 70%, and the historical baseline value accounts for 30%, used for fusion and updating). At the moment each self-cleaning cycle is completed (i.e., when the three-stage flushing and magnetic pulse cleaning end and the system is ready to reset), the oil film status sensing unit will collect the current contamination index in real time (e.g., 30, the higher the value, the more severe the contamination). The sliding window automatically records the index for the most recent k (5) times and calculates its average as the new baseline. For example, if the contamination index at the completion of the last 5 self-cleaning cycles was 30, 28, 32, 29, and 31, the average is (30+28+32+29+31)÷5=30, meaning the new baseline is 30. This value reflects the baseline level of contamination after cleaning under the current equipment condition. If the equipment oil quality gradually deteriorates, this average may gradually increase (e.g., if the subsequent indices are 32, 33, and 34, the average rises to 33). The new baseline is compared with the historical baseline (i.e., the previously determined baseline, such as an initial baseline of 25), and the decision to update is made based on the magnitude of the deviation. If the deviation of the new baseline value calculated consecutively for p times from the historical baseline value is less than 5%, the current baseline is considered still applicable and does not need to be updated. For example, if the historical baseline is 25, the first new baseline is 30 (deviation 20%, exceeding the threshold), the second new baseline is 29 (deviation 16%, exceeding the threshold), the third new baseline is 26 (deviation 4%, less than the threshold), and the fourth new baseline is 27 (deviation 8%, exceeding the threshold). Since there are no three consecutive deviations less than the threshold, the original baseline is not maintained. If the subsequent three consecutive new baselines are 26, 25.5, and 26.2 (deviations from the historical baseline 25 of 4%, 2%, and 5% respectively), all less than 5%, then the original baseline 25 is maintained.

[0034] If the deviation between the new baseline value and the historical baseline value exceeds 5% (e.g., historical baseline 25, new baseline 30, deviation 20%), an update process is initiated, and the new baseline is calculated by merging the two values ​​according to a preset weight ratio (new baseline 70%, historical baseline 30%). For example, 30 × 70% + 25 × 30% = 21 + 7.5 = 28.5. This 28.5 is set as the new historical baseline value and used as the comparison benchmark for triggering the next self-cleaning cycle (i.e., cleaning is triggered when the real-time contamination index exceeds 28.5).

[0035] The updated baseline value immediately replaces the original baseline to determine whether subsequent contamination levels exceed the limit. For example, after the new baseline 28.5 takes effect, if the real-time contamination index rises to 29, exceeding the baseline, the system triggers a self-cleaning cycle; if the index is 28, it is not triggered, and the sliding window continues to record new self-cleaning cycle data, entering the next baseline judgment cycle. For example, after the new baseline 28.5 has been running for a period of time, the average contamination index after the last 5 cleanings is 29, with a deviation of approximately 1.7% (less than 5%) from 28.5. If this is the case for 3 consecutive times, it will remain at 28.5; if the subsequent average rises to 32 (deviation of 12.3%), it will be updated again by weighted fusion to ensure that the baseline is always synchronized with the actual contamination status of the equipment, allowing the contamination baseline to dynamically adapt to long-term changes in the equipment, and ensuring that the triggering time of the self-cleaning cycle is always accurate, neither wasting energy too early nor causing oil quality deterioration too late.

[0036] Please see Figure 2 As shown, a second objective of this invention is to provide a system for implementing a deep learning-based intelligent adjustment method for the lubrication status of coal mine equipment, including any of the above-mentioned methods, comprising: The oil film status sensing unit 1 integrates the differential pressure sensor of the oil supply line and the oil return line with the embedded dual-channel feature fusion module. It extracts the time-domain abrupt change features through one-dimensional dilatation convolution and separates the feature frequency bands through wavelet packet decomposition, and outputs the oil film thickness estimate and the contamination index. The oil injection decision unit 2 deploys a dual-branch competitive architecture reinforcement learning network, inputs oil film thickness, wear coefficient and oil injection effectiveness feedback, dynamically generates oil injection pulse parameters, and connects to a pulse width modulation controller to perform linear interpolation duration calculation and sliding window interval adjustment; The self-cleaning collaborative execution unit 3 includes an axial gradient magnetic field filter element, a three-stage variable frequency flushing module, and a magnetic pulse desorption mechanism. After the bypass oil valve is switched, the cleaning cycle is triggered, and the baseline value of the contamination degree of the sliding window is updated when the cycle is completed.

[0037] This invention first collects the pressure difference signal between the oil supply and return pipelines. Then, it extracts the time-domain non-stationary fluctuations and frequency-domain energy distribution features through a dual-channel feature fusion network to generate an estimated oil film thickness and a contamination index. When the oil film thickness is lower than the dynamic maintenance threshold, a reinforcement learning strategy network with a dual-branch competitive architecture, combined with the equipment wear coefficient and historical oil injection feedback, dynamically generates the duration and interval of the oil injection pulse. When the contamination exceeds the limit, an online self-cleaning cycle is triggered. Contaminants are removed through a high-gradient magnetic filter, a three-stage high-flow-rate flushing, and magnetic pulse desorption, and the contamination baseline is updated, improving the accuracy of lubrication adjustment and ensuring stable equipment operation.

[0038] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent adjustment of lubrication status of coal mine equipment based on deep learning, characterized in that: By sensing the state of the lubricating oil film and adaptively adjusting the oiling and cleaning behavior, the following steps are included: S1. Real-time acquisition of pressure difference signals between the oil supply line and the return line of the lubrication system, extraction of pressure fluctuation characteristics through a pre-trained deep learning model, and output of the current lubricating oil film thickness estimate and lubricating oil contamination index. S2. When the estimated value of the lubricating oil film thickness is lower than the dynamic maintenance threshold, an oil injection pulse is initiated. The duration and pulse interval of the oil injection pulse are dynamically generated by the reinforcement learning strategy network. The input of the reinforcement learning strategy network is the current estimated value of the lubricating oil film thickness, the cumulative wear coefficient of the equipment during operation, and the feedback on the effectiveness of historical oil injection. S3. When the lubricating oil contamination index exceeds the downhole working condition adaptive threshold, an online self-cleaning cycle is triggered. The online self-cleaning cycle includes: switching the bypass oil valve to allow the lubricating oil to flow through the high gradient magnetic filter element, and controlling the oil pump to enter a short-time high flow rate flushing mode for a preset period. After cleaning is completed, the system is reset to the normal oil injection mode and the contamination baseline is updated.

2. The intelligent adjustment method for lubrication status of coal mine equipment based on deep learning according to claim 1, characterized in that, The deep learning model is a dual-channel feature fusion network. The first channel extracts the time-domain non-stationary fluctuation features of the pressure difference signal between the oil supply pipeline and the return pipeline through one-dimensional dilated convolution. Based on the time-domain non-stationary fluctuation features, the time-domain feature vector is captured, including the instantaneous bearing state change of the oil film corresponding to the pressure change point. The second channel reconstructs the frequency domain energy distribution of the pressure difference signal through wavelet packet decomposition, and separates the frequency domain feature vector based on the frequency domain energy distribution, including high-frequency noise, feature frequency bands reflecting oil film thickness, and feature frequency bands reflecting the concentration of pollutant particles. The time-domain feature vector and the frequency-domain feature vector are input into the feature fusion gating module of the dual-channel feature fusion network. The fused feature vector is generated by cross-channel attention weight allocation. Finally, it is mapped to the lubricating oil film thickness estimate and the lubricating oil contamination index through the fully connected layer. The contamination index is associated with the proportion of pulses with a value higher than the baseline energy threshold in the feature frequency band.

3. The intelligent adjustment method for lubrication status of coal mine equipment based on deep learning according to claim 2, characterized in that, The reinforcement learning policy network adopts a two-branch competition architecture: The State Value branch evaluates the current lubrication state value of the equipment, while the Action Advantage branch evaluates the optimization potential of different injection pulse parameters. The network input layer receives the estimated value of the lubricating oil film thickness, the cumulative wear coefficient of the equipment running time, and the feedback on the effectiveness of historical oil injection. The cumulative wear coefficient of the equipment running time is generated by a nonlinear function of the cumulative running time of the equipment and the preset life curve. The feedback on the effectiveness of historical oil injection is defined as the average deviation between the oil film thickness recovery rate and the expected recovery rate after the first three oil injections. The output layer generates a discrete action space of the duration and interval of the oil injection pulse.

4. The intelligent adjustment method for lubrication status of coal mine equipment based on deep learning according to claim 3, characterized in that, The duration of the oil injection pulse is dynamically adjusted by the pulse width modulation controller according to the action command output by the reinforcement learning policy network, specifically as follows: When the estimated oil film thickness is lower than the dynamic maintenance threshold by more than the preset amplitude threshold, a maximum duration pulse is triggered to increase the oil replenishment rate. When the amplitude is within the preset amplitude threshold range, the duration is dynamically calculated within the preset time range by linear interpolation. The pulse interval adopts a sliding window adaptive mechanism to predict the next oil replenishment time based on the oil film decay rate after the most recent preset number of oil injections.

5. The intelligent adjustment method for lubrication status of coal mine equipment based on deep learning according to claim 1, characterized in that, The reinforcement learning policy network is trained using a multi-objective reward function, including: (a) Oil film maintenance reward: When the oil film thickness is maintained within the range allowed by the dynamic maintenance threshold after oil injection, a positive reward is given. (b) Energy consumption penalty item: the negative energy consumption reward is calculated by multiplying the oil injection duration by the system's rated power; (c) Wear suppression reward: When the cumulative wear coefficient growth rate decreases, an incremental reward is given. The dynamic maintenance threshold is dynamically updated based on the real-time speed and load pressure of the equipment through a pre-calibrated three-dimensional mapping table.

6. The intelligent adjustment method for lubrication status of coal mine equipment based on deep learning according to claim 1, characterized in that, The high-gradient magnetic filter element adopts an axial gradient magnetic field design, as detailed below: A four-stage NdFeB magnetic ring assembly is arranged along the direction of lubricating oil flow. The adjacent magnetic ring assemblies are arranged with opposite polarities to form a magnetic field gradient. The skeleton surface of the high-gradient magnetic filter element is covered with honeycomb-shaped guide grooves, which generate a swirling effect when the lubricating oil containing contaminant particles flows through it. The four-stage NdFeB magnetic ring assembly is wrapped with an oleophilic and dust-repellent coating to improve the adsorption efficiency of ferromagnetic particles. The contamination load status of the high-gradient magnetic filter element is monitored in real time by a built-in Hall sensor to assess the magnetic field strength attenuation rate.

7. The intelligent adjustment method for lubrication status of coal mine equipment based on deep learning according to claim 1, characterized in that, The short-duration high-velocity flushing mode employs a three-stage velocity modulation, as detailed below: During the start-up phase, the oil flow is stabilized by maintaining a base flow rate for n seconds. During the core flushing phase, the flow rate is increased to the peak flow rate for the preset period, which is dynamically calculated based on the extent of exceedance of the pollution index. During the decay phase, the flow rate is reduced to the base flow rate for m seconds to avoid pressure shock. Flow rate control is achieved by adjusting the speed of the oil pump motor through a frequency converter.

8. The intelligent adjustment method for lubrication status of coal mine equipment based on deep learning according to claim 7, characterized in that, When switching the bypass oil valve, the magnetic pulse cleaning module is activated simultaneously. The magnetic pulse cleaning module applies an alternating magnetic field pulse sequence with gradually increasing amplitude to the high gradient magnetic filter element, causing the adsorbed ferromagnetic particles to vibrate slightly and detach from the magnetic pole. The detached particles are flushed by the oil flow in a short time under high flow rate and enter the sedimentation chamber. An electromagnetic discharge valve is set at the bottom of the sedimentation chamber and automatically opens to discharge sewage at the end of the cleaning cycle.

9. The intelligent adjustment method for lubrication status of coal mine equipment based on deep learning according to claim 1, characterized in that, The pollution baseline update employs a sliding window weighted mechanism, as detailed below: The average contamination index at the moment the most recent k self-cleaning cycles are completed is recorded as the new baseline reference value. When the deviation between the new baseline reference value and the historical baseline reference value is less than a preset deviation threshold for p consecutive cycles, the original baseline is maintained. When the deviation exceeds the preset deviation threshold, the new baseline reference value and the historical baseline reference value are merged and updated according to a preset weight ratio. The updated baseline reference value is used as the comparison reference value for the next self-cleaning cycle.

10. A system for implementing the intelligent adjustment method for lubrication status of coal mine equipment based on deep learning as described in any one of claims 1-9, characterized in that, include: The oil film state sensing unit (1) integrates the differential pressure sensor of the oil supply pipeline and the return oil pipeline with the embedded dual-channel feature fusion module. It extracts the time-domain abrupt change features through one-dimensional dilatation convolution and separates the feature frequency bands through wavelet packet decomposition, and outputs the oil film thickness estimate and the contamination index. The oil injection decision unit (2) deploys a dual-branch competitive architecture reinforcement learning network, inputs oil film thickness, wear coefficient and oil injection effectiveness feedback, dynamically generates oil injection pulse parameters, and connects to a pulse width modulation controller to perform linear interpolation duration calculation and sliding window interval adjustment; The self-cleaning collaborative execution unit (3) includes an axial gradient magnetic field filter element, a three-stage variable frequency flushing module and a magnetic pulse desorption mechanism. After the bypass oil valve is switched, the cleaning cycle is triggered, and the baseline value of the contamination degree of the sliding window is updated when the cycle is completed.

Images