Power generation equipment failure risk analysis and prediction system based on reinforcement learning optimization

Abstract

The application discloses a power generation equipment fault risk analysis and prediction system based on reinforcement learning optimization and concretely relates to the power system operation and maintenance technical field, comprising a multi-source data acquisition module, through deploying multi-protocol intelligent sensors at key nodes of power generation equipment, collecting equipment operation state original data flow in real time, and synchronously integrating SCADA system operation logs and environmental working condition data, a full-dimension space-time sequence data set covering equipment physical state and operation environment is constructed, an adaptive sliding window technology is adopted to preliminarily filter and remove abnormal values from the original data, and a multi-modal equipment operation data warehouse with time alignment characteristics is formed; the application generates four-dimensional influence coefficients through an index calculation module, considers multi-factor nonlinear coupling, adopts weighted geometric mean for data fusion, combines dynamic weight reinforcement learning optimization coefficient weight, fits complex fault relations, and achieves the purposes of improving prediction accuracy and solving the limitations of traditional statistical methods.

Description

Technical Field

[0001] This invention relates to the field of power system operation and maintenance technology, and more specifically, to a power generation equipment fault risk analysis and prediction system based on reinforcement learning optimization. Background Technology

[0002] Power generation equipment, such as gas turbines, wind turbines, and steam turbine units, are core assets of the energy system. Their complex structures and continuous operation under harsh conditions such as high loads, high temperatures, and high pressures make key components prone to performance degradation and sudden failures. Unplanned outages not only result in significant power generation losses and high maintenance costs but can also jeopardize grid stability and even trigger safety accidents. Therefore, conducting fault risk analysis and prediction for power generation equipment and implementing predictive maintenance are crucial for ensuring energy security and improving economic efficiency.

[0003] Traditional power generation equipment fault analysis systems include a data acquisition module, a fault diagnosis module, and a results display module. The data acquisition module covers key physical quantities of the core components of the equipment and its function is to obtain the basic operating parameters of the equipment. The fault diagnosis module learns from historical data to identify fault modes and predict the remaining useful life of the equipment. The results display module presents the fault diagnosis results through text reports or simple line graphs, conveying basic information to operation and maintenance personnel.

[0004] However, in practical use, it still has some shortcomings, such as poor real-time performance. Traditional systems rely on historical data and fixed rules to judge faults, which are difficult to adjust with changes in equipment status or environment. Manual intervention is still required for adjustment, and it is impossible to respond to sudden changes in equipment operating conditions in a timely manner. The prediction accuracy is also poor. Traditional systems mostly use simple statistical methods, which can only capture the linear relationship between a single factor and a fault. They cannot fit the nonlinear fault relationship coupled with multiple factors, and the fault generalization ability is weak, so the prediction accuracy is poor. Data analysis is isolated. Traditional systems perform isolated analysis on a single device and cannot handle the chain reaction and collaborative optimization problems between units.

[0005] Therefore, there is an urgent need to provide a power generation equipment failure risk analysis and prediction system based on reinforcement learning optimization to solve the problems of poor real-time performance, poor prediction accuracy, and isolated data analysis in existing power generation equipment failure analysis systems. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a power generation equipment fault risk analysis and prediction system based on reinforcement learning optimization, which solves the problems mentioned in the background art through the following solutions.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a power generation equipment fault risk analysis and prediction system based on reinforcement learning optimization, comprising: The multi-source data acquisition module collects raw data streams of equipment operation status in real time by deploying multi-protocol smart sensors at key nodes of power generation equipment, and synchronously integrates SCADA system operation logs and environmental condition data to build a full-dimensional spatiotemporal sequence dataset covering the physical status and operating environment of the equipment. It uses adaptive sliding window technology to perform preliminary filtering and outlier removal on the raw data, forming a multimodal equipment operation data warehouse with time alignment characteristics. The indicator calculation module, based on the four types of state features in the multimodal equipment operation data warehouse output by the multi-source data acquisition module, generates mechanical data influence coefficients, electrical data influence coefficients, thermal data influence coefficients, and equipment risk loss data influence coefficients through nonlinear relationships, forming a standardized state input vector for the reinforcement learning agent; The data fusion module, based on the four-dimensional influence coefficients of mechanical, electrical, thermal and equipment risk loss, adopts a reinforcement learning agent to dynamically optimize the weight allocation strategy and generate a comprehensive index of equipment health status, which is used to assess the overall health status and risk level of the equipment. The fault risk analysis module, based on the comprehensive health status index of equipment, uses a dynamic weight reinforcement learning algorithm to dynamically allocate weights and optimizes the weight allocation of mechanical, electrical, thermal and loss coefficients in real time through a policy network. The fault risk prediction module calculates the real-time fault probability based on the weighted allocation of the final output equipment health status comprehensive index and mechanical, electrical, thermal and loss coefficients, and then performs fault risk prediction and outputs adaptive maintenance strategies. The data interaction module transmits the obtained comprehensive health status index of the equipment, the weighting of mechanical, electrical, thermal and loss coefficients, real-time failure probability and maintenance strategy to the user information terminal, providing reference data for making adjustment measures.

[0008] Preferably, the multimodal equipment operation data warehouse includes mechanical data impact parameters, electrical data impact parameters, thermal data impact parameters, and equipment risk loss data impact parameters.

[0009] Preferably, the mechanical data influence parameters include the vibration harmonic amplitude, denoted as A; the vibration fundamental frequency amplitude, denoted as A1; and the bearing temperature gradient, denoted as... Electrical data influence parameters include dielectric loss factor, denoted as... The reference dielectric loss factor is denoted as... Partial discharge pulse count, denoted as Np; thermal data influencing parameters include the hottest spot temperature, denoted as T1; cooling medium temperature, denoted as T2; standard deviation of temperature measurement points, denoted as... The rated temperature gradient is denoted as The parameters affecting equipment risk loss data include cumulative operating time, denoted as t; average operating load, denoted as L; oil contamination degree, denoted as C; and effective value of vibration velocity, denoted as V.

[0010] Preferably, the mechanical data influence coefficient characterizes the coupling fault risk of mechanical stress and thermal effect through the product of vibration harmonic energy and temperature gradient. The harmonic energy amplifies the high-frequency fault characteristics, and the relative temperature rise gradient reflects the frictional heat effect. After multiplying the two, the natural logarithm function is used to achieve dimensional normalization and numerical compression, which is used to quantify the overall deterioration degree of rotating machinery.

[0011] Preferably, the electrical data influence coefficient characterizes the degree of insulation degradation by multiplying the dielectric loss factor and the discharge pulse activity. The relative dielectric loss reflects the overall insulation aging, and the logarithmic discharge count amplifies the contribution of local defects. The multiplication of the two reflects the synergistic effect of overall insulation aging and partial discharge.

[0012] Preferably, the thermal data influence coefficient characterizes the thermal stress risk by multiplying the relative superheat and the temperature field non-uniformity. The relative superheat measures the degree of absolute temperature exceeding the limit, the temperature non-uniformity reflects heat dissipation efficiency and internal faults, and the multiplicative relationship reflects the dual influence of the absolute temperature value and the uniformity of distribution.

[0013] Preferably, the impact coefficient of the equipment risk loss data is characterized by the comprehensive aging rate through the multiplication of four factors: time accumulation, load stress, pollution acceleration, and vibration wear. The time factor is linearly accumulated, the load square term emphasizes the overload hazard, the exponential pollution term amplifies the pollution acceleration effect, the vibration factor directly reflects mechanical wear, and the multiplicative coupling reflects the synergistic amplification effect of multiple factors.

[0014] Preferably, the comprehensive health status index of the equipment is fused by weighted geometric average of four types of influence coefficients: mechanical vibration, electrical insulation, thermal distribution and risk loss, through weight parameters dynamically generated by the reinforcement learning agent. The geometric average model strengthens the short-board effect of the deterioration of a single indicator, and the variable weight realizes the adaptive adjustment of the importance of the indicator under different operating conditions. The exponential function naturally eliminates the difference in dimensions and compresses the output value range.

[0015] Preferably, the real-time fault probability is specifically: , EHCI stands for Equipment Health Index. This represents the threshold for the comprehensive health status index of the equipment, which is the average of the historical comprehensive health status index of the equipment under the corresponding weight allocation.

[0016] The technical effects and advantages of this invention are as follows: 1. This invention achieves real-time response to changes in operating conditions by collecting operational data in real time, and optimizes the weight of the four-dimensional influence coefficient in real time, thus solving the problem of poor real-time performance and frequent manual intervention required by traditional systems. 2. This invention generates a four-dimensional influence coefficient through an index calculation module, considers the nonlinear coupling of multiple factors, uses a weighted geometric average for data fusion, and combines dynamic weight reinforcement learning to optimize the coefficient weights, fits complex fault relationships, and achieves the goal of improving prediction accuracy and overcoming the limitations of traditional statistical methods. 3. This invention constructs a full-dimensional spatiotemporal dataset through multi-source data acquisition. It not only collects data from single devices but also integrates system logs and environmental data. The data interaction module outputs comprehensive information, supports collaborative analysis between units, and achieves the goal of solving the problem of isolated data analysis in traditional systems and responding to the chain reaction of units. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall structure of the present invention. Detailed Implementation

[0018] 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.

[0019] As attached Figure 1 The power generation equipment failure risk analysis and prediction system based on reinforcement learning optimization shown includes: The multi-source data acquisition module collects raw data streams of equipment operation status in real time by deploying multi-protocol smart sensors at key nodes of power generation equipment. It also integrates SCADA system operation logs and environmental condition data to construct a full-dimensional spatiotemporal sequence dataset covering the physical status and operating environment of the equipment. Adaptive sliding window technology is used to perform preliminary filtering and outlier removal on the raw data, forming a multimodal equipment operation data warehouse with time alignment characteristics.

[0020] In this embodiment, it should be specifically noted that the multimodal equipment operation data warehouse includes mechanical data influence parameters, electrical data influence parameters, thermal data influence parameters, and equipment risk loss data influence parameters.

[0021] In this embodiment, it should be specifically noted that the mechanical data influence parameters include the vibration harmonic amplitude, denoted as A; the vibration fundamental frequency amplitude, denoted as A1; and the bearing temperature gradient, denoted as... .

[0022] In this embodiment, it should be specifically noted that the vibration harmonic amplitude is obtained by acquiring the original vibration signal at a sampling frequency of 51.2kHz using an ICP accelerometer installed in the bearing housing of the equipment, and the amplitude of the nth harmonic is extracted by FFT spectrum analysis.

[0023] In this embodiment, it should be specifically noted that the fundamental frequency amplitude of the vibration is obtained by acquiring the original vibration signal at a sampling frequency of 51.2kHz using an ICP accelerometer installed in the bearing housing of the equipment, and the fundamental frequency amplitude is extracted by FFT spectrum analysis.

[0024] In this embodiment, it should be specifically noted that the bearing temperature gradient is calculated by collecting temperature values ​​and measuring the installation distance using a PT100 sensor embedded in the bearing outer ring and housing.

[0025] In this embodiment, it should be specifically noted that the electrical data influence parameters include the dielectric loss factor, denoted as... The reference dielectric loss factor is denoted as... The partial discharge pulse count is denoted as Np.

[0026] In this embodiment, it should be specifically noted that the dielectric loss factor is obtained by directly measuring the dielectric loss tangent of the insulating material using a dedicated dielectric loss tester during equipment maintenance by applying a test voltage.

[0027] In this embodiment, it should be specifically noted that the reference dielectric loss factor is the initial value of the dielectric loss factor measured for the first time after the equipment is put into operation, which serves as a health benchmark.

[0028] In this embodiment, it should be specifically noted that the partial discharge pulse count is obtained by collecting the high-frequency pulse signal of the equipment grounding wire through a high-frequency current transformer, and then identifying and counting it by a partial discharge analyzer.

[0029] In this embodiment, it should be specifically noted that the thermal data influence parameters include the temperature of the hottest spot, denoted as T1; the temperature of the cooling medium, denoted as T2; and the standard deviation of the temperature measurement points, denoted as... The rated temperature gradient is denoted as .

[0030] In this embodiment, it should be specifically noted that the temperature of the hottest spot is directly acquired by a fiber optic grating temperature sensor pre-embedded in the winding or core of the equipment.

[0031] In this embodiment, it should be specifically noted that the temperature of the cooling medium is directly acquired by a PT100 temperature sensor installed at the cooling inlet.

[0032] In this embodiment, it should be specifically noted that the standard deviation of the temperature measuring point is obtained by collecting a set of temperature values ​​by arranging thermocouples on the surface of the equipment and calculating their standard deviation.

[0033] In this embodiment, it should be specifically noted that the rated temperature gradient is the normal temperature gradient value calibrated by simulation and experiment under rated operating conditions.

[0034] In this embodiment, it should be specifically noted that the parameters affecting the equipment risk loss data include cumulative operating time, denoted as t; average operating load, denoted as L; oil contamination degree, denoted as C; and effective value of vibration velocity, denoted as V.

[0035] In this embodiment, it should be specifically noted that the cumulative running time is directly read from the cumulative value of the running timer in the equipment control system.

[0036] In this embodiment, it should be specifically noted that the average operating load is obtained by collecting the average active power over a period of time through a smart meter installed in the power circuit of the equipment.

[0037] In this embodiment, it should be specifically noted that the oil contamination level is obtained by directly outputting the number of particles in the oil in real time using an online oil particle counter based on the principle of laser light shielding.

[0038] In this embodiment, it should be specifically noted that the effective value of the vibration velocity is obtained by the data acquisition unit through a magnetoelectric velocity sensor or an acceleration signal that has been integrated.

[0039] The indicator calculation module, based on the four types of state features in the multimodal equipment operation data warehouse output by the multi-source data acquisition module, generates mechanical data influence coefficients, electrical data influence coefficients, thermal data influence coefficients, and equipment risk loss data influence coefficients through nonlinear relationships, forming a standardized state input vector for the reinforcement learning agent.

[0040] In this embodiment, it should be specifically explained that the mechanical data influence coefficient characterizes the coupled fault risk of mechanical stress and thermal effects through the product of vibration harmonic energy and temperature gradient. The harmonic energy amplifies high-frequency fault characteristics, and the relative temperature rise gradient reflects the frictional heat effect. After multiplying the two, the natural logarithm function is used to achieve dimensional normalization and numerical compression, which is used to quantify the overall deterioration degree of rotating machinery. Specifically: , Where An represents the amplitude of the nth vibration harmonic; A1 represents the fundamental frequency amplitude of the vibration. Indicates the bearing temperature gradient. This indicates the rated temperature gradient of the bearing.

[0041] In this embodiment, it should be specifically explained that the electrical data influence coefficient characterizes the degree of insulation degradation through the product of the dielectric loss factor and the discharge pulse activity. The relative dielectric loss reflects the overall insulation aging, and the logarithmic discharge count amplifies the contribution of local defects. The product of the two reflects the synergistic effect of overall insulation aging and partial discharge, achieving dimensional normalization. Specifically: , in Indicates the dielectric loss factor. Np represents the reference dielectric loss factor, and Np represents the partial discharge pulse count. This indicates the rated standard discharge pulse count.

[0042] In this embodiment, it should be specifically explained that the thermal data influence coefficient characterizes the thermal stress risk through the product of relative superheat and temperature field non-uniformity. Relative superheat quantifies the degree of absolute temperature exceeding the limit, while temperature non-uniformity reflects heat dissipation efficiency and internal faults. The multiplicative relationship reflects the dual influence of absolute temperature value and distribution uniformity. Specifically: , Where T1 represents the hottest spot temperature, and T2 represents the cooling medium temperature. Indicates the maximum allowable temperature rise. Indicates the standard deviation of temperature measurement points. This indicates the rated temperature gradient.

[0043] In this embodiment, it should be specifically explained that the influence coefficient of the equipment risk loss data is characterized by the comprehensive aging rate through the multiplication of four factors: time accumulation, load stress, pollution acceleration, and vibration wear. The time factor is linearly accumulated, the load square term emphasizes the overload hazard, the exponential pollution term amplifies the pollution acceleration effect, and the vibration factor directly reflects mechanical wear. The multiplicative coupling reflects the synergistic amplification effect of multiple factors. Specifically: , Where t represents the cumulative operating time, t0 represents the time constant, specifically the normalized time set based on the equipment's design life and maintenance cycle; L represents the average operating load, Lr represents the rated load; C represents the oil contamination degree, C 0表示 The target cleanliness level that the equipment is required to maintain; V represents the effective value of vibration velocity, and Vb represents the vibration baseline value, which is the minimum vibration value established through long-term monitoring and statistics of the equipment under healthy conditions.

[0044] The data fusion module, based on the four-dimensional influence coefficients of mechanical, electrical, thermal, and equipment risk loss, adopts a reinforcement learning agent to dynamically optimize the weight allocation strategy and generate a comprehensive index of equipment health status, which is used to assess the overall health status and risk level of the equipment.

[0045] In this embodiment, it should be specifically explained that the comprehensive health status index of the equipment is fused by weighted geometric mean of four types of influence coefficients: mechanical vibration, electrical insulation, thermal distribution, and risk loss, through weight parameters dynamically generated by the reinforcement learning agent. The geometric mean model strengthens the weak link effect of a single indicator deterioration, and the variable weights enable adaptive adjustment of the importance of the indicators under different operating conditions. The exponential function naturally eliminates the difference in dimensions and compresses the output value range. Specifically: , Where MIC represents the mechanical data influence coefficient, EIC represents the electrical data influence coefficient, TIC represents the thermal data influence coefficient, RLIC represents the equipment risk loss data influence coefficient, ω1 is the weight of the mechanical data influence coefficient, initially 0.3, ω2 is the weight of the electrical data influence coefficient, initially 0.3, ω3 is the weight of the thermal data influence coefficient, initially 0.3, and ω4 is the weight of the equipment risk loss data influence coefficient, initially 0.1.

[0046] The fault risk analysis module, based on the comprehensive health status index of the equipment, uses a dynamic weight reinforcement learning algorithm to dynamically allocate weights and optimizes the weight allocation of mechanical, electrical, thermal and loss coefficients in real time through a policy network.

[0047] In this embodiment, it should be specifically explained that the dynamic weight allocation is as follows: S1. Determine the basic weight and final weight range based on the weight type. The influence coefficient of mechanical data has a basic weight of 0.3 and a final weight range of 0.2-0.4. The influence coefficient of electrical data has a basic weight of 0.3 and a final weight range of 0.22-0.38. The influence coefficient of thermal data has a basic weight of 0.3 and a final weight range of 0.24-0.36. The influence coefficient of mechanical data has a basic weight of 0.1 and a final weight range of 0.06-0.14. S2. Dynamically adjust rules based on data in the equipment operation data warehouse. When the load rate is >90%, it is considered an overload condition, ω1+0.05, ω2+0.03, and the remaining weights are reduced accordingly by increasing the average weight. When the ambient temperature is >35℃, it is considered a high-temperature environment, ω3+0.04, ω4+0.02, and the remaining weights are reduced accordingly by increasing the average weight. If there are more than 5 starts and stops in a day, it is considered frequent starts and stops, ω4+0.03, ω1+0.02, and the remaining weights are reduced accordingly by increasing the average weight. When tanδ > twice the initial value, it is considered as insulation aging, ω2 + 0.05, and the remaining weights are reduced accordingly by increasing the average weight.

[0048] S3, Weight Calculation: Calculate the final weights based on S1 and S2.

[0049] The fault risk prediction module calculates the real-time fault probability based on the weighted allocation of the final output equipment health status comprehensive index and mechanical, electrical, thermal and loss coefficients, and then performs fault risk prediction and outputs adaptive maintenance strategies.

[0050] In this embodiment, it should be specifically noted that the real-time fault probability is as follows: A1. Calculate the real-time failure probability, specifically: , EHCI stands for Equipment Health Index. This represents the threshold for the comprehensive health status index of the equipment, which is the average of the historical comprehensive health status index of the equipment under the corresponding weight allocation.

[0051] A2. Risk Level Determination: The real-time failure probability is compared with a preset threshold range to classify the risk level. Specifically: When P < 0.2, it indicates safety and no action is required.

[0052] When 0.2 ≤ P < 0.4, it indicates that attention is needed and the monitoring frequency should be increased to the next stage.

[0053] When 0.4 ≤ P < 0.6, a warning is issued, the system generates an early warning, and suggests arranging scheduled maintenance.

[0054] When P > 0.6, it indicates danger, and the machine should be stopped immediately for inspection.

[0055] The data interaction module transmits the obtained comprehensive health status index of the equipment, the weighting of mechanical, electrical, thermal and loss coefficients, real-time failure probability and maintenance strategy to the user information terminal, providing reference data for making adjustment measures.

[0056] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power generation equipment fault risk analysis and prediction system based on reinforcement learning optimization, characterized in that, include: The multi-source data acquisition module collects raw data streams of equipment operation status in real time by deploying multi-protocol smart sensors at key nodes of power generation equipment, and synchronously integrates SCADA system operation logs and environmental condition data to build a full-dimensional spatiotemporal sequence dataset covering the physical status and operating environment of the equipment. It uses adaptive sliding window technology to perform preliminary filtering and outlier removal on the raw data, forming a multimodal equipment operation data warehouse with time alignment characteristics. The multimodal equipment operation data warehouse includes parameters affecting mechanical data, electrical data, thermal data, and equipment risk loss data. The mechanical data influencing parameters include vibration harmonic amplitude, denoted as A; vibration fundamental frequency amplitude, denoted as A1; and bearing temperature gradient, denoted as... ; Electrical data influence parameters include dielectric loss factor, denoted as The reference dielectric loss factor is denoted as... Partial discharge pulse count, denoted as Np; thermal data influencing parameters include the hottest spot temperature, denoted as T1; cooling medium temperature, denoted as T2; standard deviation of temperature measurement points, denoted as... The rated temperature gradient is denoted as The parameters affecting equipment risk and loss data include cumulative operating time, denoted as t; average operating load, denoted as L; and oil contamination degree, denoted as C. The effective value of vibration velocity is denoted as V; The indicator calculation module, based on the four types of state features in the multimodal equipment operation data warehouse output by the multi-source data acquisition module, generates mechanical data influence coefficients, electrical data influence coefficients, thermal data influence coefficients, and equipment risk loss data influence coefficients through nonlinear relationships, forming a standardized state input vector for the reinforcement learning agent; The mechanical data influence coefficient characterizes the coupled fault risk of mechanical stress and thermal effects through the product of vibration harmonic energy and temperature gradient. The harmonic energy amplifies high-frequency fault characteristics, and the relative temperature rise gradient reflects the frictional heat effect. The product of the two is then normalized and numerically compressed using the natural logarithm function to quantify the overall deterioration degree of rotating machinery. Specifically: ， Where An represents the amplitude of the nth vibration harmonic; A1 represents the fundamental frequency amplitude of the vibration. Indicates the bearing temperature gradient. Indicates the rated temperature gradient of the bearing; The electrical data influence coefficient characterizes the degree of insulation degradation through the product of the dielectric loss factor and the discharge pulse activity. The relative dielectric loss reflects the overall insulation aging, and the logarithmic discharge count amplifies the contribution of local defects. The product of the two reflects the synergistic effect of overall insulation aging and partial discharge, achieving dimensional normalization. Specifically: ， in Indicates the dielectric loss factor. Np represents the reference dielectric loss factor, and Np represents the partial discharge pulse count. Indicates the rated standard discharge pulse count; The thermal data influence coefficient characterizes the thermal stress risk through the product of relative superheat and temperature field non-uniformity. Relative superheat measures the degree of absolute temperature exceeding the limit, while temperature non-uniformity reflects heat dissipation efficiency and internal faults. The multiplicative relationship reflects the dual influence of absolute temperature value and distribution uniformity. Specifically: ， Where T1 represents the hottest spot temperature, and T2 represents the cooling medium temperature. Indicates the maximum allowable temperature rise. Indicates the standard deviation of temperature measurement points. Indicates the rated temperature gradient; The impact coefficient of equipment risk loss data is characterized by the multiplication of four factors: time accumulation, load stress, pollution acceleration, and vibration wear, representing the comprehensive aging rate. The time factor is linearly accumulated, the load square term emphasizes the harm of overload, the exponential pollution term amplifies the pollution acceleration effect, and the vibration factor directly reflects mechanical wear. The multiplicative coupling reflects the synergistic amplification effect of multiple factors. Specifically: ， Where t represents the cumulative operating time, t0 represents the time constant, specifically the normalized time set based on the equipment's design life and maintenance cycle; L represents the average operating load, Lr represents the rated load; C represents the oil contamination degree, C 0表示 The target cleanliness level that the equipment is required to maintain; V represents the effective value of vibration velocity, and Vb represents the vibration baseline value, which is the minimum vibration value established through long-term monitoring and statistics of the equipment in a healthy state; The data fusion module, based on the four-dimensional influence coefficients of mechanical, electrical, thermal and equipment risk loss, adopts a reinforcement learning agent to dynamically optimize the weight allocation strategy and generate a comprehensive index of equipment health status, which is used to assess the overall health status and risk level of the equipment. The comprehensive equipment health status index uses weighted geometric averages to fuse four types of influence coefficients—mechanical vibration, electrical insulation, thermal distribution, and risk loss—through dynamically generated weight parameters from a reinforcement learning agent. The geometric average model amplifies the weakest link effect caused by the deterioration of a single indicator, while variable weights adaptively adjust the importance of indicators under different operating conditions. The exponential function naturally eliminates dimensional differences and compresses the output value range. Specifically: ， Where MIC represents the mechanical data influence coefficient, EIC represents the electrical data influence coefficient, TIC represents the thermal data influence coefficient, RLIC represents the equipment risk loss data influence coefficient, ω1 is the weight of the mechanical data influence coefficient, initially 0.3, ω2 is the weight of the electrical data influence coefficient, initially 0.3, ω3 is the weight of the thermal data influence coefficient, initially 0.3, and ω4 is the weight of the equipment risk loss data influence coefficient, initially 0.

1. The fault risk analysis module, based on the comprehensive health status index of equipment, uses a dynamic weight reinforcement learning algorithm to dynamically allocate weights and optimizes the weight allocation of mechanical, electrical, thermal and loss coefficients in real time through a policy network. The dynamic weight allocation is specifically as follows: S1. Determine the basic weight and final weight range based on the weight type. The influence coefficient of mechanical data has a basic weight of 0.3 and a final weight range of 0.2-0.

4. The influence coefficient of electrical data has a basic weight of 0.3 and a final weight range of 0.22-0.

38. The influence coefficient of thermal data has a basic weight of 0.3 and a final weight range of 0.24-0.

36. The influence coefficient of mechanical data has a basic weight of 0.1 and a final weight range of 0.06-0.

14. S2. Dynamically adjust rules based on data in the equipment operation data warehouse; When the load rate is >90%, it is considered an overload condition, ω1+0.05, ω2+0.03, and the remaining weights are reduced accordingly by increasing the average weight. When the ambient temperature is >35℃, it is considered a high-temperature environment, ω3+0.04, ω4+0.02, and the remaining weights are reduced accordingly by increasing the average weight. If there are more than 5 starts and stops in a day, it is considered frequent starts and stops, ω4+0.03, ω1+0.02, and the remaining weights are reduced accordingly by increasing the average weight. When tanδ > twice the initial value, it is considered as insulation aging, ω2 + 0.05, and the remaining weights are reduced accordingly by increasing the average weight. S3, Weight Calculation: Calculate the final weights based on S1 and S2; The fault risk prediction module calculates the real-time fault probability based on the weighted allocation of the final output equipment health status comprehensive index and mechanical, electrical, thermal and loss coefficients, and then performs fault risk prediction and outputs adaptive maintenance strategies. The real-time failure probability is specifically: A1. Calculate the real-time failure probability, specifically: ， EHCI stands for Equipment Health Index. This represents the threshold for the comprehensive health status index of the equipment, which is the average of the historical comprehensive health status index of the equipment under the corresponding weight allocation; A2. Risk Level Determination: The real-time failure probability is compared with a preset threshold range to classify the risk level. Specifically: When P < 0.2, it indicates safety and no action is required; When 0.2 ≤ P < 0.4, it indicates that attention is needed and the monitoring frequency should be increased to the next stage; When 0.4 ≤ P < 0.6, a warning is issued, the system generates an early warning, and suggests scheduling maintenance. When P > 0.6, it indicates danger; stop the machine immediately for inspection. The data interaction module transmits the obtained comprehensive health status index of the equipment, the weighting of mechanical, electrical, thermal and loss coefficients, real-time failure probability and maintenance strategy to the user information terminal, providing reference data for making adjustment measures.

Images