Fracturing whole pump reliability repairing method based on dynamic fault analysis

Abstract

The invention discloses a method for repairing reliability of a whole fracturing pump based on dynamic fault analysis, which comprises the following steps of: S1, acquiring operation data of the whole fracturing pump, and preprocessing and storing the operation data by utilizing edge computing equipment; s2, extracting short-term features and long-term features of the operation data by adopting an adaptive segmentation method of a dynamic window, and carrying out anomaly detection by combining an isolated forest and a local anomaly factor method; s3, constructing a fracturing whole pump fault propagation model, updating the fault propagation probability by using a Markov chain Monte Carlo method, and generating a fault propagation influence matrix; s4, evaluating a fault mode by adopting a dynamic fault tree-Markov chain conjoint analysis method to form a fault influence degree matrix; s5, potential failure point prediction is carried out, and a fault evolution trend chart is generated; and S6, constructing a three-layer optimal scheduling framework by adopting an optimal scheduling method of a game theory, and generating an optimal repair strategy. According to the method, dynamic fault analysis, game theory optimization and the like are combined, and accurate prediction and intelligent repair of the whole fracturing pump are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic fault analysis, and particularly to a reliability repair method for a fracturing pump based on dynamic fault analysis. Background Art

[0002] During the process of oil and gas extraction, a fracturing pump is an important high-pressure conveying device, and its core function is to inject fracturing fluid into the formation at high pressure to improve the oil and gas recovery rate. However, due to long-term operation in extreme environments of high load, high pressure, high temperature, and high corrosion, the fracturing pump is extremely vulnerable to factors such as component fatigue, fluid erosion, vibration shock, and temperature change, resulting in frequent equipment failures. The existing maintenance methods for fracturing pumps mainly rely on regular inspections and manual patrols. However, such traditional maintenance methods have many problems, such as low fault prediction accuracy, difficult-to-trace fault propagation paths, repair strategies relying on experience, high maintenance costs, and untimely responses, which cannot meet the requirements for high reliability and low downtime of equipment in modern oil and gas extraction processes.

[0003] The existing fault diagnosis methods for fracturing pumps mainly rely on means such as rule-based expert systems, vibration analysis, and thermal imaging monitoring. These methods can identify equipment abnormalities to a certain extent, but often have problems of poor real-time performance and insufficient adaptability. For example, traditional vibration analysis methods usually extract statistical features based on time domain and frequency domain, and judge the operating state of the equipment through thresholds, but this method is difficult to process complex fault data with nonlinearity and multiple modes. In addition, although rule-based expert systems can use historical experience and preset rules to identify faults, due to the complex operating environment of the fracturing pump and the highly dynamic fault modes, traditional rule-based methods are difficult to handle sudden faults and multi-component coupling faults. In addition, traditional fault detection systems usually use independent sensors for data collection and transmit the data to a central server for centralized processing, but this method has problems such as high data transmission delay and large bandwidth occupancy, resulting in limited real-time performance of fault prediction.

[0004] To improve the accuracy of fault prediction, in recent years, some studies have adopted data-driven intelligent fault diagnosis methods, such as machine learning and deep learning techniques. These methods automatically learn the change patterns of the device operation status by training models and predict the occurrence time of potential faults. However, the existing intelligent fault diagnosis methods still have deficiencies in practical applications. On the one hand, most data-driven methods rely on high-quality fault data sets. However, in the actual engineering environment, the fault data of the fracturing pump unit is often limited, making it difficult to support the effective training of deep models. On the other hand, traditional deep learning methods mainly rely on static feature extraction and ignore the dynamics of fault propagation, making it difficult for the model to adapt to the real-time changes of the device operation status. In addition, most existing methods only focus on the state judgment at a single time step and cannot predict the fault risks at multiple future time steps, thus limiting their application value in long-term fault management.

[0005] In addition, the problem of tracing the fault propagation path is also one of the challenges that are difficult to effectively solve by the existing technologies. The inside of the fracturing pump unit consists of multiple interrelated components, such as pump shafts, impellers, seals, bearings, etc. Once a component fails, it may affect other components through mechanical vibrations, fluid impacts, etc. However, the traditional fault tree analysis (FTA) method is difficult to model such complex dynamic fault propagation relationships and cannot handle the fault impacts that change over time. In recent years, some studies have introduced Bayesian networks and Markov chain methods to analyze the fault propagation path, but there are still problems such as high computational complexity and difficulty in real-time reasoning. In addition, traditional fault prediction methods usually only focus on the state of a single component and ignore the interactions between components, making the health assessment of the overall system insufficiently comprehensive and resulting in inaccurate prediction results in practical applications.

[0006] The existing maintenance strategies mainly make decisions based on the experience of engineers or preset maintenance rules. However, due to the complexity of the device operation status and the uncertainty of the environment, empirical repair methods often lead to too high maintenance costs or unreasonable allocation of maintenance resources. For example, over-maintenance may result in unnecessary downtime and waste of spare parts, while under-maintenance may cause more serious fault consequences. In recent years, some studies have tried to introduce optimization scheduling algorithms, such as heuristic optimization, reinforcement learning, etc., to balance costs and reliability. However, traditional optimization methods usually rely on static optimization models and lack the ability to adaptively adjust to the system operation status and cannot optimize the repair plan in real time.

[0007] Therefore, how to provide a reliability repair method for the fracturing pump unit based on dynamic fault analysis is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose a reliability repair method for a fracturing pump based on dynamic fault analysis. The present invention combines dynamic fault analysis, Bayesian causal reasoning, deep graph attention mechanism, game theory optimization and deep reinforcement learning to construct a reliability repair method for a fracturing pump. Through multi-modal data acquisition, fault propagation modeling, dynamic fault tree analysis and optimal scheduling, accurate fault prediction, propagation path tracking and intelligent repair are realized. Compared with traditional methods, the present invention has the advantages of high prediction accuracy, optimal scheduling, reasonable resource allocation, adaptive maintenance strategy and low maintenance cost, etc., and can effectively improve the operation stability of the fracturing pump, reduce unplanned downtime, and improve the long-term reliability and service life of the equipment.

[0009] The reliability repair method for a fracturing pump based on dynamic fault analysis according to an embodiment of the present invention includes the following steps:

[0010] S1. Collect the operation data of the fracturing pump through a distributed intelligent sensor network. The operation data includes structural parameters, fluid parameters, vibration parameters and electrical parameters, and use edge computing devices for preprocessing and storage;

[0011] S2. Adopt an adaptive segmentation method with a dynamic window to extract the short-term features and long-term features of the operation data, and combine the isolation forest and local outlier factor methods for anomaly detection to generate an abnormal state data set;

[0012] S3. Based on the abnormal state data set, construct a fault propagation model for the fracturing pump, use Bayesian dynamic causal reasoning combined with a deep graph attention mechanism for modeling, and use the Markov chain Monte Carlo method to update the fault propagation probability to generate a fault propagation influence matrix;

[0013] S4. Based on the fault propagation influence matrix, adopt a dynamic fault tree-Markov chain joint analysis method to evaluate the fault mode, calculate the fault probability of different fault modes, and form a fault impact degree matrix;

[0014] S5. Based on the fault impact degree matrix, predict potential failure points, and calculate the occurrence probability of the failure mode in the next few time steps to generate a fault evolution trend diagram;

[0015] S6. Based on the fault evolution trend diagram, adopt an optimal scheduling method of game theory to establish an optimal repair strategy, and construct a three-layer optimal scheduling framework. The first layer calculates the fault severity based on the fault evolution trend diagram, the second layer uses the Stackelberg game model to perform dynamic optimization between cost and reliability, and calculates the Nash equilibrium point to select the optimal repair plan. The third layer selects the optimal repair path based on the deep Q network, inputs the fault mode, potential failure points and historical repair data for optimization training, and generates the best repair path, including repair steps, required materials and time cost information.

[0016] Optionally, S2 specifically includes:

[0017] S21. Set a group of dynamic windows with variable lengths according to the operation cycle and load variation of the fracturing pump unit. After the operation data acquisition is completed, divide the operation data into the corresponding windows in chronological order to form a segmented sequence:

[0018] D i = {x t |t ∈ [t start , t end};

[0019] Among them, D i represents the i-th segment of the operation data sequence, x t represents the discrete signal at time step t, t start represents the start time, t end represents the end time;

[0020] S22. Perform a fast Fourier transform with a weighted window function on each segmented sequence to extract short-term features. The windowed signal is:

[0021]

[0022] Among them, represents the windowed discrete signal, and w(t) represents the Hamming window function;

[0023] The discrete Fourier transform is:

[0024]

[0025] Among them, X win (k) represents the frequency domain transformation result, k represents the frequency index, T represents the length of the segmented sequence, and j represents the imaginary unit;

[0026] S23. Apply wavelet multi-scale analysis to each segmented sequence to extract long-term features. The discrete wavelet transform can be expressed as:

[0027]

[0028] Among them, a represents the scale factor, b represents the translation factor, ψ * represents the complex conjugate function of the mother wavelet function ψ, and W(a, b) represents the wavelet transform coefficient. By selecting the low-frequency coefficient components and performing multi-scale reconstruction to obtain the trend part, it is used to analyze the overall change trend of the fracturing pump unit;

[0029] S24. Align the short-term features and long-term features in the time dimension and merge them into a feature set:

[0030] Θ = {θ1 , θ 2 , …, θ m};

[0031] Among them, Θ represents the feature set, m represents the total number of features, and θ i represents the feature vector obtained after fast Fourier transform and wavelet transform in the i-th segment, including the frequency features corresponding to the transient fault signal and the long-term operation trend information;

[0032] S25. Perform preliminary anomaly detection on the feature set Θ using the isolation forest model. Assume that the isolation forest contains r random isolation trees:

[0033]

[0034] Among them, represents the average path length, and h u (θ i ) represents the path length of θ i in the u-th tree;

[0035] And define the anomaly score function:

[0036]

[0037] Among them, c(N) represents the normalization coefficient set based on the total number of features m, and score IF (θ i ) represents the anomaly score function. When score IF (θ i ) exceeds the set threshold Γ 1 , then θ i is regarded as a suspicious feature point;

[0038] S26. Perform secondary detection on the suspicious feature points:

[0039]

[0040] Among them, lrd k (θ i ) represents the local reachability density of θ i , N k (θ i ) represents the k-nearest neighbor set, |N k (θ i )| represents the total number of elements in the k-nearest neighbor set, d(θ i , θ j ) represents the distance function between θ i and θ j in the feature space, and θ j represents the feature vector obtained after fast Fourier transform and wavelet transform in the j-th segment;

[0041] Calculate the local outlier factor:

[0042]

[0043] Among them, LOF k (θ i ) represents the local outlier factor of θ, and lrd i (θ k ) represents the local reachability density of θ; j ) represents θ j The local reachability density;

[0044] When LOF k (θ i ) is greater than the pre-set decision threshold Γ 2 , it is marked as abnormal data, and all the detected abnormal data and the corresponding feature vectors are included in the abnormal state dataset.

[0045] Optionally, the S3 specifically includes:

[0046] S31. Based on the abnormal state dataset, construct a fracturing pump failure propagation model, and define the fracturing pump component set as:

[0047] V = {v 1 , v 2 , …, v n};

[0048] Among them, V represents the fracturing pump component set, n represents the total number of fracturing pump components, and construct a dynamic directed graph G t = (V, E t ), where E t = {(v i , v j , p ij (t))∣i≠j} represents the fault propagation edge set that changes with time, and p ij (t) represents the fault influence probability of component v i on component v j , and adjust the topological structure using the time-weighted attention mechanism;

[0049] S32. Construct an asymmetric Bayesian causal inference model, and define the component fault state vector as:

[0050] Z t = (Z t,1 , Z t,2 , …, Z t,n );

[0051] Among them, Z t represents the component fault state vector at time step t, and Z t,i∈{0,1} represents the component v i At time step t, the binary fault state defines a non-uniform Bayesian causal inference process:

[0052]

[0053] Among them, p(Z t ∣Z t-1 ) represents the transition probability of the fracturing pump unit fault state in the time dimension. Z t-1 represents the component fault state vector at time step t - 1, D t represents the fault propagation historical data matrix at time step t, and β i represents the dynamic influence factor of component v i ;

[0054] S33. Optimize the fault propagation probability using adaptive Markov chain Monte Carlo, and define the fault propagation probability matrix as:

[0055] T t ={T ij (t)∣i,j∈[1,n]};

[0056] Among them, T t represents the fault propagation probability matrix at time step t, and T ij (t) represents the transition probability that the fault propagates from component v i to component v j at time step t. Define the dynamic Bayesian update mechanism:

[0057]

[0058] Among them, T ij (t + 1) represents the transition probability that the fault propagates from component v i to component v j at time step t + 1. λ represents the learning rate, I(·) represents the exponential function, γ represents the time discount factor, t represents the current time step, s represents the historical time step, Z s,i represents the fault state of component v i at time step s. If Z s,i =1, it means that component v i sends a fault at time step s, otherwise it means normal operation; Z s,j represents the fault state of component v j at time step s. If Z s,j =1, it means that component v j sends a fault at time step s, otherwise it means normal operation;

[0059] S34. Integrate Bayesian causal inference and temporal graph attention mechanism to generate a fault propagation influence matrix:

[0060]

[0061] Among them, M t represents the fault propagation influence matrix at time step t, and T s represents the fault propagation probability matrix at time step s, and A s represents the fault propagation adjacency matrix at time step s, and ⊙ represents the element-by-element product operator.

[0062] Optionally, the S4 specifically includes:

[0063] S41. Obtain the fault propagation influence matrix, determine the fracturing pump unit set and the top event, establish a dynamic fault tree structure, correspond each basic event to the component, and mark the failure path in the dynamic fault tree;

[0064] S42. Construct a Markov chain state description according to the dynamic fault tree, and define the fault state probability vector at time t as:

[0065] P(t) = (P 1 (t), P 2 (t), …, P N (t));

[0066] Among them, P i (t) represents the probability of the i-th state, N represents the total number of states, and construct a state transition matrix Q according to the transition rate between failure events, satisfying:

[0067]

[0068] S43. Analyze the multiple modes of the dynamic fault tree, distinguish the parallel gates, priority gates and multi-input gates, extract the combined failure event sets under different logic gates, form a fault mode list, and dynamically track the fault modes in combination with the component failure correlation information;

[0069] S44. Calculate the failure rates of each fault mode at different time steps:

[0070]

[0071] Among them, represents the Markov failure rate function of the fault mode F k , ξ i represents the weighting coefficient of the component v i in the fault mode F k , and φ i (t) represents the fault propagation intensity of the component v i at time step t;

[0072] S45. Based on the Markov chain simulation results, combined with the failure rate function of the failure mode, calculate the failure probability step by step over time, fuse the dynamic fault tree logical relationship to generate the failure probability distribution, and sort the probability values of each failure mode to identify the main sources of failure risks;

[0073] S46. Generate a failure impact matrix according to the cumulative impact degree of different failure modes at each time step.

[0074] Optionally, the specific steps of S5 are as follows:

[0075] S51. Expand the failure impact matrix in time series, construct a time series dataset according to the historical failure propagation mode of different components, extract the impact factors at each time step, and combine historical failure events to construct a preliminary screening set of potential failure points;

[0076] S52. Segment the time series data using the adaptive sliding window method, extract multi-scale features of the failure impact mode at different time scales, and adaptively adjust the window size according to the characteristics of different failure modes;

[0077] S53. Construct a time evolution model of the failure mode based on the failure impact matrix:

[0078]

[0079] where P(Q k ∣t) represents the probability of the failure mode Q occurring at time step t k , E k represents the set of components of the failure mode Q k , w i represents the influence weight of component v i in the failure mode Q k , Z represents the normalization constant, exp represents the exponential function, ζ i represents the failure attenuation factor of component v i , represents the most recent time point when component v i is in an abnormal state;

[0080] S54. Based on the multi-step prediction method, perform rolling calculations on the failure probability of future time steps, adopt the Bayesian time update strategy to dynamically correct the prediction results, adaptively adjust the occurrence probability of the failure mode, and perform statistics on the time distribution of different failure modes to generate the failure probability trend of future time steps;

[0081] S55. Generate a failure evolution trend graph according to the failure probability trend of future time steps.

[0082] Optionally, the specific steps of S6 are as follows:

[0083] S61. Analyze the influence scope and fault propagation intensity of different failure modes according to the fault evolution trend diagram, perform time-weighted calculation on each failure mode, and obtain the fault severity.

[0084] S62. Based on the Stackelberg game theory, construct a two-step optimization solution process. In the first step, dynamically adjust the repair priority based on the global repair goal of the system manager. In the second step, based on the cost constraint conditions of the repair strategy executor, perform game calculation among different repair schemes, and use the iterative method to solve the Nash equilibrium point, so that the final repair scheme achieves an optimal trade-off between cost and reliability.

[0085] S63. Use the deep Q-network to optimize the repair path, construct a training framework based on the experience replay mechanism, input the fault mode, potential failure points and historical repair data, and search for the best repair path:

[0086] Q * (s,a) = (1 - η)Q(s,a) + η[ρ + ρmax a' Q(s′,a′) - τlog 2 π(a′∣s′)];

[0087] Among them, Q * (s,a) represents the expected return after updating the execution of action a in the current state s, η represents the learning rate, Q(s,a) represents the expected return before updating the execution of action a in the current state s, ρ represents the current immediate reward, max a' Q(s',a') represents the maximum expected return of executing action a' in the next state s', τ represents the temperature parameter, and π(a'∣s') represents the probability of selecting action a' in the next state s'.

[0088] The beneficial effects of the present invention are as follows:

[0089] First of all, the present invention uses a distributed intelligent sensor network to realize the real-time acquisition of multi-modal operation data of the fracturing pump unit, and combines edge computing for data preprocessing, reduces data transmission delay, improves data integrity and timeliness, and provides accurate and reliable input data for subsequent fault analysis. Compared with the traditional data acquisition method based on rules or single sensors, the present invention can comprehensively obtain the structural parameters, fluid parameters, vibration parameters and electrical parameters during the operation of the fracturing pump unit, and extract short-term features and long-term features through the dynamic window adaptive segmentation method, so as to take into account instantaneous fault signals and long-term operation trends and improve the accuracy of anomaly detection.

[0090] Secondly, in terms of fault propagation modeling, the present invention combines Bayesian dynamic causal reasoning with a deep graph attention mechanism to construct a fault propagation model that better conforms to the dynamic association characteristics among components of a fracturing pump unit. Traditional methods are difficult to handle the temporal evolution of faults and the coupling relationships among components in complex systems. However, the present invention optimizes the fault propagation probability by introducing the Markov chain Monte Carlo method to generate a fault propagation influence matrix, making the fault propagation path clearer and enabling accurate prediction of the fault influence range among different components. In addition, through the combined analysis method of dynamic fault tree - Markov chain, the present invention can calculate the occurrence probabilities of different fault modes, form a fault influence degree matrix, and achieve a quantitative assessment of the health state of the fracturing pump unit system, thereby improving the reliability of fault diagnosis.

[0091] In addition, the present invention breaks through the traditional maintenance strategy based on experience or fixed rules and innovatively introduces a game theory - optimized scheduling method to achieve intelligent optimization of the optimal repair strategy. In terms of optimization scheduling, the present invention adopts a Stackelberg game model to establish a master - slave game relationship, enabling the system manager and the maintenance decision - making party to weigh between repair costs and reliability, and calculating the Nash equilibrium point to ensure the global optimality of the repair plan, thus solving the problem of unreasonable allocation of maintenance resources in traditional methods. At the same time, in terms of optimizing the repair path, the present invention is based on the deep Q - network (DQN) reinforcement learning method, dynamically adjusts the maintenance strategy during multiple rounds of game optimization, and combines an experience replay mechanism for intelligent training, enabling the maintenance strategy to adaptively adjust according to the equipment state and the trend of fault evolution, thereby achieving optimal repair path planning and improving the maintenance efficiency.

[0092] Furthermore, in order to improve the timeliness of fault prediction, the present invention adopts a multi - step - length rolling prediction method to calculate the occurrence probabilities of failure modes within several future time steps, and combines a time - weighted mechanism to generate a fault evolution trend graph. Compared with traditional methods that can only passively analyze based on static data, the multi - step - length prediction and dynamic adjustment mechanism of the present invention enable the system to perceive possible future fault modes in advance and dynamically adjust the maintenance strategy according to the fault evolution trend, ensuring that maintenance decisions intervene in advance and effectively reducing the equipment downtime risk. Finally, the present invention can output a complete optimal repair path, including maintenance steps, required materials, and time - cost information, and synchronize the optimized maintenance plan to the remote monitoring system to ensure the feasibility and reliability of the repair strategy in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0094] Figure 1This is the overall flowchart of the reliability repair method for the fracturing pump based on dynamic fault analysis proposed by the present invention. Specific embodiments

[0095] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0096] Reference Figure 1 , the reliability repair method for the fracturing pump based on dynamic fault analysis includes the following steps:

[0097] S1. Collect the operation data of the fracturing pump through a distributed intelligent sensor network. The operation data includes structural parameters, fluid parameters, vibration parameters, and electrical parameters, and preprocess and store them using edge computing devices;

[0098] S2. Extract the short-term features and long-term features of the operation data using an adaptive segmentation method with a dynamic window, and combine the isolation forest and local outlier factor methods for anomaly detection to generate an abnormal state dataset;

[0099] S3. Based on the abnormal state dataset, construct a fault propagation model for the fracturing pump, use Bayesian dynamic causal inference combined with a deep graph attention mechanism for modeling, and use the Markov chain Monte Carlo method to update the fault propagation probability to generate a fault propagation influence matrix;

[0100] S4. Based on the fault propagation influence matrix, use a combined analysis method of dynamic fault tree - Markov chain to evaluate the fault mode, calculate the fault probability of different fault modes, and form a fault impact degree matrix;

[0101] S5. Based on the fault impact degree matrix, predict potential failure points, and calculate the occurrence probability of the failure mode in the next several time steps to generate a fault evolution trend graph;

[0102] S6. Based on the fault evolution trend graph, use an optimization scheduling method based on game theory to establish an optimal repair strategy, and construct a three-layer optimization scheduling framework. The first layer calculates the fault severity based on the fault evolution trend graph, the second layer uses the Stackelberg game model to perform dynamic optimization between cost and reliability, and calculates the Nash equilibrium point to select the optimal repair plan. The third layer selects the optimal repair path based on the deep Q network, inputs the fault mode, potential failure points, and historical repair data for optimization training to generate the best repair path, including repair steps, required materials, and time cost information.

[0103] In this embodiment, the specific content of S2 includes:

[0104] S21. Set a group of dynamic windows with variable lengths according to the operation cycle and load variation of the fracturing pump. After the operation data acquisition is completed, divide the operation data into corresponding windows in chronological order to form segmented sequences:

[0105] D i ={x t |t∈[t start ,t end};

[0106] Among them, D i represents the i-th segment of operation data sequence, x t represents the discrete signal at time step t, t start represents the start time, t end represents the end time;

[0107] S22. Perform fast Fourier transform with a weighted window function on each segmented sequence to extract short-term features. The windowed signal is:

[0108]

[0109] Among them, represents the windowed discrete signal, and w(t) represents the Hamming window function;

[0110] The discrete Fourier transform is:

[0111]

[0112] Among them, X win (k) represents the frequency domain transformation result, k represents the frequency index, T represents the length of the segmented sequence, and j represents the imaginary unit;

[0113] S23. Apply wavelet multi-scale analysis to each segmented sequence to extract long-term features. The discrete wavelet transform can be expressed as:

[0114]

[0115] Among them, a represents the scale factor, b represents the translation factor, ψ * represents the complex conjugate function of the mother wavelet function ψ, and W(a,b) represents the wavelet transform coefficient. The trend part is obtained by selecting the low-frequency coefficient components and performing multi-scale reconstruction, which is used to analyze the overall change trend of the fracturing pump;

[0116] S24. Align the short-term features and long-term features in the time dimension and merge them into a feature set:

[0117] Θ={θ 1 ,θ 2 ,…,θ m};

[0118] Among them, Θ represents the feature set, m represents the total number of features, and θ i represents the feature vector obtained after fast Fourier transform and wavelet transform in the i-th segment, including the frequency features corresponding to the transient fault signal and the long-term operation trend information;

[0119] S25. Perform preliminary anomaly detection on the feature set Θ using the isolation forest model. Assume that the isolation forest contains r randomly isolated trees:

[0120]

[0121] Among them, represents the average path length, and h u (θ i ) represents the path length of θ i in the u-th tree;

[0122] And define the anomaly score function:

[0123]

[0124] Among them, c(N) represents the normalization coefficient set based on the total number of features m, and score IF (θ i ) represents the anomaly score function. When score IF (θ i ) exceeds the set threshold Γ 1 , then θ i is regarded as a suspicious feature point;

[0125] S26. Perform secondary detection on the suspicious feature points:

[0126]

[0127] Among them, lrd k (θ i ) represents the local reachability density of θ i , N k (θ i ) represents the k-nearest neighbor set, |N k (θ i )| represents the total number of elements in the k-nearest neighbor set, and d(θ i , θ j ) represents the distance function between θ i and θ j in the feature space. θ j represents the feature vector obtained after fast Fourier transform and wavelet transform in the j-th segment;

[0128] Calculate the local anomaly factor:

[0129]

[0130] Among them, LOF k (θ i ) represents the local outlier factor of θ i , and lrd k (θ j ) represents the local reachability density of θ j ;

[0131] When LOF k (θ i ) is greater than the preset determination threshold Γ 2 , it is marked as abnormal data, and all the detected abnormal data and corresponding feature vectors are included in the abnormal state data set.

[0132] In this embodiment, S3 specifically includes:

[0133] S31. Based on the abnormal state data set, construct a fracturing pump unit fault propagation model, and define the fracturing pump unit component set as:

[0134] V = {v 1 , v 2 , …, v n};

[0135] Among them, V represents the fracturing pump unit component set, n represents the total number of fracturing pump unit components, construct a dynamic directed graph G t = (V, E t ), where E t = {(v i , v j , p ij (t))∣i≠j} represents the fault propagation edge set that changes with time, and p ij (t) represents the fault influence probability of component v i on component v j , and adjust the topological structure using the time-weighted attention mechanism;

[0136] S32. Construct an asymmetric Bayesian causal inference model, and define the component fault state vector as:

[0137] Z t = (Z t,1 , Z t,2 , …, Z t,n );

[0138] Among them, Z t represents the component fault state vector at time step t, Z t,i ∈ {0, 1} represents the binary fault state of component v i at time step t, and define the non-uniform Bayesian causal inference process:

[0139]

[0140] Among them, p(Z t ∣Z t-1 ) represents the transition probability of the fracturing pump failure state in the time dimension. Z t-1 represents the component failure state vector at time step t - 1, D t represents the failure propagation historical data matrix at time step t, and β i represents the dynamic influence factor of component v i .

[0141] S33. Optimize the failure propagation probability using adaptive Markov chain Monte Carlo, and define the failure propagation probability matrix as:

[0142] T t ={T ij (t)∣i,j∈[1,n]};

[0143] Among them, T t represents the failure propagation probability matrix at time step t, and T ij (t) represents the transition probability of the failure from component v i to component v j at time step t. Define the dynamic Bayesian update mechanism:

[0144]

[0145] Among them, T ij (t + 1) represents the transition probability of the failure from component v i to component v j at time step t + 1. λ represents the learning rate, I(·) represents the exponential function, γ represents the time discount factor, t represents the current time step, s represents the historical time step, Z s,i represents the failure state of component v i at time step s. If Z s,i = 1, it means that component v i sends a failure at time step s, otherwise it means normal operation; Z s,j represents the failure state of component v j at time step s. If Z s,j = 1, it means that component v j sends a failure at time step s, otherwise it means normal operation;

[0146] S34. Integrate Bayesian causal reasoning and temporal graph attention mechanism to generate the failure propagation influence matrix:

[0147]

[0148] Among them, M t represents the fault propagation influence matrix at time step t, and T s represents the fault propagation probability matrix at time step s, and A s represents the fault propagation adjacency matrix at time step s, and ⊙ represents the element-wise product operator.

[0149] In this embodiment, the S4 specifically includes:

[0150] S41. Obtain the fault propagation influence matrix, determine the fracturing pump component set and the top event, establish a dynamic fault tree structure, correspond each basic event to a component, and mark the failure path in the dynamic fault tree;

[0151] S42. Construct a Markov chain state description according to the dynamic fault tree, and define the fault state probability vector at time t as:

[0152] P(t) = (P 1 (t), P 2 (t), …, P N (t));

[0153] Among them, P i (t) represents the probability of the i-th state, N represents the total number of states, and construct a state transition matrix Q according to the transition rate between failure events, satisfying:

[0154]

[0155] S43. Analyze the multiple modes of the dynamic fault tree, distinguish parallel gates, priority gates and multi-input gates, extract the combined failure event sets under different logic gates, form a fault mode list, and dynamically track the fault modes in combination with component failure correlation information;

[0156] S44. Calculate the failure rates of each fault mode at different time steps:

[0157]

[0158] Among them, represents the Markov failure rate function of the fault mode F k , ξ i represents the weighted coefficient of the component v i in the fault mode F k , and φ i (t) represents the fault propagation intensity of the component v i at time step t;

[0159] S45. Based on the simulation results of the Markov chain, combined with the failure rate function of the failure mode, calculate the failure probability step by step in time, fuse the logical relationship of the dynamic fault tree to generate the failure probability distribution, and sort the probability values of each failure mode to identify the main sources of failure risks;

[0160] S46. Generate a failure impact matrix according to the cumulative impact degree of different failure modes at each time step.

[0161] In this embodiment, the specific steps of S5 are as follows:

[0162] S51. Expand the failure impact matrix in time series, construct a time series data set according to the historical failure propagation mode of different components, extract the influence factors at each time step, and combine the historical failure events to construct a preliminary screening set of potential failure points;

[0163] S52. Segment the time series data by using the adaptive sliding window method, extract multi-scale features of the failure impact modes at different time scales, and adaptively adjust the window size according to the characteristics of different failure modes;

[0164] S53. Build a time evolution model of the failure mode based on the failure impact matrix:

[0165]

[0166] where P(Q k ∣t) represents the probability of the failure mode Q occurring at time step t k , E k represents the set of components of the failure mode Q k , w i represents the influence weight of the component v i in the failure mode Q k , Z represents the normalization constant, exp represents the exponential function, ζ i represents the failure attenuation factor of the component v i , represents the most recent time point when the component v i is in an abnormal state;

[0167] S54. Based on the multi-step prediction method, perform rolling calculations on the failure probability of future time steps, adopt the Bayesian time update strategy to dynamically correct the prediction results, adaptively adjust the occurrence probability of the failure mode, and perform statistics on the time distribution of different failure modes to generate the failure probability trend of future time steps;

[0168] S55. Generate a failure evolution trend chart according to the failure probability trend of future time steps.

[0169] In this embodiment, the specific steps of S6 are as follows:

[0170] S61. Analyze the influence scope and fault propagation intensity of different failure modes according to the fault evolution trend diagram, perform time-weighted calculation on each failure mode, and obtain the fault severity.

[0171] S62. Based on the Stackelberg game theory, construct a two-step optimization solution process. In the first step, dynamically adjust the repair priority based on the global repair goal of the system manager. In the second step, based on the cost constraint conditions of the repair strategy executor, perform game calculation among different repair plans, and use the iterative method to solve the Nash equilibrium point, so that the final repair plan achieves an optimal trade-off between cost and reliability.

[0172] S63. Optimize the repair path using a deep Q-network, construct a training framework based on the experience replay mechanism, input the fault mode, potential failure points, and historical repair data, and search for the best repair path:

[0173] Q * (s,a) = (1 - η)Q(s,a) + η[ρ + ρmax a 'Q(s′,a′) - τlog 2 π(a∣s′)];

[0174] Where Q * (s,a) represents the updated expected return when performing action a in the current state s, η represents the learning rate, Q(s,a) represents the expected return before updating when performing action a in the current state s, ρ represents the current immediate reward, max a 'Q(s',a') represents the maximum expected return when performing action a' in the next state s', τ represents the temperature parameter, and π(a'∣s') represents the probability of selecting action a' in the next state s'.

[0175] Example 1:

[0176] To verify the feasibility of the present invention in implementation, the present invention is applied to the fracturing construction site of a large shale gas field. The fracturing pump equipment of this gas field has been operating under high load and high pressure for a long time. Affected by geological conditions and working methods, the wear and fatigue failure of internal components of the pump body such as pump shafts, impellers, seals, bearings and other components are relatively serious. Due to the lag of existing fault prediction methods, there are still problems such as frequent sudden failures, untimely maintenance responses, and high maintenance costs during the fracturing construction process. The traditional equipment maintenance mode mainly relies on fixed-time interval inspections and maintenance plans, and it is difficult to accurately predict and actively repair before equipment failure, resulting in long downtime for repair after sudden failures, seriously affecting the construction progress.

[0177] In this embodiment, the reliability repair method of the fracturing pump based on dynamic fault analysis of the present invention is introduced for equipment health management. First, a distributed intelligent sensor network is installed at the construction site, and a variety of sensors such as temperature, flow, pressure, vibration, current, and oil pressure are arranged at the key parts of each fracturing pump. The real-time operation data is transmitted to the edge computing device on-site through the 5G industrial Internet for local preprocessing and storage.

[0178] In the data processing stage, the present invention adopts an adaptive segmentation method with a dynamic window to extract short-term features and long-term features from the collected operation data. For short-term features, Fourier transform is used to analyze the transient vibration signal of the equipment, and the Isolation Forest method is combined for preliminary anomaly detection; for long-term features, wavelet transform is used to extract trend information, and the Local Outlier Factor (LOF) method is used for further screening to ensure that fault symptoms can be identified at an early stage.

[0179] In terms of fault propagation modeling, a fault propagation network is established using Bayesian dynamic causal inference and a deep graph attention mechanism. The Markov Chain Monte Carlo method is used to calculate the fault propagation probability between different components, and a fault propagation impact matrix is dynamically generated. For example, during the test of a certain fracturing pump, the system monitored that the vibration amplitude of its impeller increased by 23% within 72 hours, the power consumption of the pump shaft increased abnormally by 12.5%, and the temperature of the seal increased by 4.7°C in a short time. The system automatically analyzed that the seal might fail after 72 hours and predicted that the probability of this failure causing further damage to the impeller was as high as 82%. In the traditional maintenance mode, the seal is only replaced when it completely fails, while the present invention can predict possible faults in advance, enabling the maintenance team to carry out preventive maintenance before it completely fails, thus avoiding a larger-scale fault expansion.

[0180] In terms of optimizing the maintenance strategy, the present invention constructs a maintenance optimization scheduling framework based on the Stackelberg game model. In a certain actual test, the system calculated according to the fault impact degree matrix that if the repair was carried out 48 hours in advance before the seal failed, the maintenance cost was about 12,000 yuan and the equipment downtime was only 3 hours. However, if the repair was carried out after the seal was completely damaged, the maintenance cost increased to 38,000 yuan and the downtime was as high as 15 hours, which had a great impact on the construction progress. Finally, the system calculated the optimal maintenance time through the game optimization method, enabling the maintenance team to complete the maintenance during the low-load period of construction (3:00 - 6:00 in the morning), minimizing the production loss to the greatest extent.

[0181] In addition, the present invention also uses deep Q-network (DQN) reinforcement learning to optimize the optimal maintenance path. In the analysis of long-term operation data, the system gradually adjusts the maintenance decision and continuously optimizes the maintenance plan in combination with the experience replay mechanism. In the continuous test for 3 months, after adopting the present invention, the average failure rate of the fracturing pump unit decreased by 47%, the unplanned downtime decreased by 38%, and the overall maintenance cost decreased by 29%. Compared with the traditional maintenance mode, the present invention significantly improves the reliability of the equipment and effectively reduces the maintenance cost while ensuring the normal production of the oil and gas field.

[0182] Table 1 Comparison table of operation and maintenance data of fracturing pump unit

[0183]

[0184] As can be seen from Table 1 above, the optimization scheme of the present invention has significant advantages in the operation and maintenance of the fracturing pump unit. First of all, in terms of the operation time, the operation time of each device basically remains between 1050 and 1250 hours, which indicates that the test environment and working conditions are relatively consistent, providing a good comparison basis for fault prediction and maintenance optimization. By comparing different devices of the traditional scheme and the optimization scheme of the present invention, it can be found that the present invention can significantly reduce the number of unplanned shutdowns and the overall shutdown duration within a similar operation time.

[0185] In terms of the predicted number of faults, the optimization scheme of the present invention can effectively improve the fault prediction ability. The predicted number of faults of the devices adopting the present invention is mostly between 6 and 10 times, while the predicted number of faults of the devices adopting the traditional scheme is relatively high. For example, the predicted number of faults of PUMP-04 device reaches 11 times, and that of PUMP-08 device even reaches 12 times, which indicates that the traditional scheme fails to accurately identify potential failure points, resulting in more frequent faults. In addition, in terms of the number of preventive maintenance, the devices of the present invention have carried out more preventive maintenance. For example, PUMP-02 has carried out 7 times of preventive maintenance, and PUMP-03 and PUMP-07 have carried out 5 times respectively, while the devices of the traditional scheme have less preventive maintenance. PUMP-06 and PUMP-08 even have only 2-3 times, which means that the traditional scheme cannot take effective intervention measures before the occurrence of faults.

[0186] The comparison of the number of unplanned shutdowns more intuitively demonstrates the advantages of the present invention. For the devices PUMP-04 and PUMP-08 using the traditional solution, there were 8 and 9 unplanned shutdowns respectively, while for the devices controlled by the optimized solution of the present invention, such as PUMP-01, PUMP-02, and PUMP-05, there were only 3 unplanned shutdowns, and PUMP-07 had only 1 unplanned shutdown, indicating that the present invention plays an important role in early intervention of faults and reduction of sudden shutdowns. In addition, in terms of the overall shutdown duration, the shutdown duration of the devices controlled by the optimized solution of the present invention has been significantly reduced. The shutdown duration of PUMP-07 is only 5 hours, while that of PUMP-08 of the traditional solution is as high as 30 hours, and the shutdown duration of PUMP-04 also reaches 28 hours, indicating that the present invention can significantly shorten the shutdown time and improve the availability of the device.

[0187] In terms of maintenance costs, the maintenance costs of the devices of the optimized solution of the present invention are generally low, basically maintained between 28,000 yuan and 45,000 yuan, while the maintenance costs of the devices of the traditional solution are high. For example, the maintenance cost of PUMP-08 reaches 81,000 yuan, and that of PUMP-06 also reaches 68,000 yuan. This shows that within the same or longer operation time, the present invention can reduce maintenance costs, optimize the allocation of maintenance resources, and improve the economy of maintenance.

[0188] In summary, after adopting the optimized solution of the present invention, the number of unplanned shutdowns of the fracturing pump unit has been reduced by more than 50%, the overall shutdown duration has been reduced by nearly 40%, and at the same time, the maintenance costs have been reduced by an average of 30%-40%, proving the significant advantages of the present invention in improving the accuracy of fault prediction, optimizing maintenance scheduling, and reducing maintenance costs. Through dynamic fault analysis, optimized scheduling, and intelligent repair strategies, the present invention makes the device operation more stable and reliable, showing strong engineering value and economic benefits in practical applications.

[0189] The above is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A reliability repair method for a fracturing pump based on dynamic fault analysis, characterized in that: The steps include: S1. Collect the operation data of the whole fracturing pump through a distributed intelligent sensor network, wherein the operation data includes structural parameters, fluid parameters, vibration parameters and electrical parameters, and use edge computing equipment for preprocessing and storage; S2, using the adaptive segmentation method of dynamic window to extract the short-term and long-term features of the operation data, and combining the isolation forest and local anomaly factor methods to perform anomaly detection and generate an abnormal state data set; S3. Based on the abnormal state data set, a fracturing pump fault propagation model is constructed. Bayesian dynamic causal reasoning combined with deep graph attention mechanism is used for modeling. The Markov chain Monte Carlo method is used to update the fault propagation probability and generate the fault propagation impact matrix. S4. Based on the fault propagation influence matrix, the dynamic fault tree-Markov chain joint analysis method is used to evaluate the fault mode, calculate the fault probability of different fault modes, and form a fault influence matrix; S5. Based on the fault impact matrix, predict potential failure points, calculate the probability of occurrence of failure modes in several future time steps, and generate a fault evolution trend diagram; S6. Based on the fault evolution trend graph, the optimal repair strategy is established by using the optimization scheduling method of game theory, and a three-layer optimization scheduling framework is constructed. The first layer calculates the fault severity based on the fault evolution trend graph. The second layer uses the Stackelberg game model to dynamically optimize between cost and reliability, and calculates the Nash equilibrium point to select the optimal repair solution. The third layer selects the optimal repair path based on the deep Q network, inputs the fault mode, potential failure point and historical repair data for optimization training, and generates the best repair path, including maintenance steps, required materials and time cost information.

2. The reliability repair method for the whole fracturing pump based on dynamic fault analysis according to claim 1 is characterized in that: The S2 specifically includes: S21. According to the operation cycle and load variation of the whole fracturing pump, a group of dynamic windows with variable lengths are set, and after the operation data collection is completed, the operation data are divided into corresponding windows in chronological order to form a segmented sequence: D i ={x t ∣t∈[t start ,t end ]}; Among them, D i represents the i-th segment running data sequence, x t represents the discrete signal at time step t, t start Indicates the start time, t end Indicates the end time; S22, performing a fast Fourier transform with a weighted window function on each segment sequence to extract short-term features, and the windowed signal is: in, represents the discrete signal after windowing, w(t) represents the Hamming window function; The discrete Fourier transform is: Among them, X win (k) represents the frequency domain transformation result, k represents the frequency index, T represents the length of the segment sequence, and j represents the imaginary unit; S23. Apply wavelet multi-scale analysis to each segmented sequence to extract long-term features. The discrete wavelet transform can be expressed as: Among them, a represents the scale factor, b represents the translation factor, ψ * represents the complex conjugate function of the mother wavelet function ψ, W(a,b) represents the wavelet transform coefficient, and the trend part is obtained by selecting the low-frequency coefficient component and performing multi-scale reconstruction to analyze the overall change trend of the whole fracturing pump; S24. Align the indexes of short-term features and long-term features in the time dimension and merge them into a feature set: Θ={θ1,θ2,…,θ m }; Among them, Θ represents the feature set, m represents the total number of features, and θ i It represents the feature vector obtained after fast Fourier transform and wavelet transform in the i-th segment, which contains the frequency characteristics and long-term operation trend information corresponding to the transient fault signal; S25. Use the isolation forest model to perform preliminary anomaly detection on the feature set Θ. Assume that the isolation forest contains r random isolated trees: in, represents the average path length, h u (θ i ) represents θ i The length of the path in the u-th tree; And define the anomaly score function: Among them, c(N) represents the normalization coefficient set based on the total number of features m, score IF (θ i ) represents the abnormal score function, when score IF (θ i ) exceeds the set threshold Γ1, it is considered that θ i is a suspicious feature point; S26, perform secondary detection on suspicious feature points: Among them, lrd k (θ i ) represents θ i The local reachable density, N k (θ i ) represents the k nearest neighbor set, |N k (θ i )| represents the total number of elements in the k nearest neighbor set, d(θ i ,θ j ) represents θ i With θ j The distance function in feature space, θ j represents the feature vector obtained after fast Fourier transform and wavelet transform in the jth segment; Calculate the local outlier factor: Among them, LOF k (θ i ) represents θ i The local outlier factor, lrd k (θ j ) represents θ j The local reachable density of When LOF k (θ i ) is greater than the preset judgment threshold Γ2, it is marked as abnormal data, and all the detected abnormal data and the corresponding feature vectors are included in the abnormal state data set.

3. The reliability repair method for the whole fracturing pump based on dynamic fault analysis according to claim 1 is characterized in that: The S3 specifically includes: S31. Based on the abnormal state data set, a fracturing pump fault propagation model is constructed, and the fracturing pump component set is defined as: V={v1,v2,…,v n }; Among them, V represents the set of fracturing pump components, n represents the total number of fracturing pump components, and a dynamic directed graph G is constructed. t =(V,E t ), where E t ={(v i ,v j ,p ij (t))|i≠j} represents the fault propagation edge set that changes over time, p ij (t) represents the component v at time step t i For components v j The failure impact probability is calculated and the topology is adjusted using the time-weighted attention mechanism; S32. Construct an asymmetric Bayesian causal reasoning model and define the component failure state vector as: WITH t =(From t,1 ,WITH t,2 ,…,WITH t,n ); Among them, Z t represents the component failure state vector at time step t, Z t,i ∈{0,1} represents component v i In the binary fault state at time step t, the non-uniform Bayesian causal inference process is defined as: Among them, p(Z t ∣Z t-1 ) represents the transition probability of the whole fracturing pump failure state in the time dimension, Z t-1 represents the component failure state vector at time step t-1, D t represents the fault propagation history data matrix at time step t, β i Represents component v i Dynamic impact factor of S33, using adaptive Markov chain Monte Carlo to optimize the fault propagation probability, the fault propagation probability matrix is ​​defined as: T t ={T ij (t)∣i,j∈[1,n]}; Among them, T t represents the fault propagation probability matrix at time step t, T ij (t) represents the fault from component v at time step t i Propagate to component v j The transition probability defines the dynamic Bayesian update mechanism: Among them, T ij (t+1) indicates that the fault is from component v at time step t+1 i Propagate to component v j The transition probability, λ represents the learning rate, I(·) represents the exponential function, γ represents the time discount factor, t represents the current time step, s represents the historical time step, and Z s,i Represents component v at time step s i Fault state, if Z s,i =1 means component v i Send a fault at time step s, otherwise it indicates normal operation; Z s,j Represents component v at time step s j Fault state, if Z s,j =1 means component v j Send failure at time step s, otherwise it indicates normal operation; S34. Integrate Bayesian causal reasoning and the time series graph attention mechanism to generate the fault propagation impact matrix: Among them, M t represents the fault propagation impact matrix at time step t, T s represents the fault propagation probability matrix at time step s, A s represents the fault propagation adjacency matrix at time step s, and ⊙ represents the element-wise product operator.

4. The reliability repair method for the whole fracturing pump based on dynamic fault analysis according to claim 1 is characterized in that: The S4 specifically includes: S41, obtaining a fault propagation influence matrix, determining a set of components of the entire fracturing pump and a top event, establishing a dynamic fault tree structure, corresponding each basic event to a component, and marking a failure path in the dynamic fault tree; S42. Construct a Markov chain state description based on the dynamic fault tree, and define the fault state probability vector at time t as: P(t)=(P1(t),P2(t),…,P N (t)); Among them, P i (t) represents the probability of the i-th state, N represents the total number of states, and the state transfer matrix Q is constructed according to the transition rate between failure events, satisfying: S43, analyzing multiple modes of the dynamic fault tree, distinguishing parallel gates, priority gates and multi-input gates, extracting the combined failure event sets under different logic gates, forming a fault mode list, and dynamically tracking the fault mode in combination with component failure association information; S44. Calculate the failure rate of each fault mode at different time steps: in, Indicates failure mode F k The Markov failure rate function, ξ i Represents component v i In failure mode F k The weighting coefficient in i (t) represents component v i The fault propagation intensity at time step t; S45. Based on the Markov chain simulation results, combined with the failure rate function of the fault mode, the fault probability is accumulated and calculated step by step, the dynamic fault tree logical relationship is integrated to generate the fault probability distribution, and the probability values ​​of each fault mode are sorted to identify the main sources of fault risk; S46. Generate a fault impact matrix according to the cumulative impact of different fault modes at each time step.

5. The reliability repair method for the whole fracturing pump based on dynamic fault analysis according to claim 1 is characterized in that: The S5 specifically includes: S51, perform time series expansion on the fault impact matrix, construct a time series data set according to the historical fault propagation mode of different components, extract the influencing factors of each time step, and construct a preliminary screening set of potential failure points in combination with historical failure events; S52, segmenting the time series data using an adaptive sliding window method, extracting multi-scale features of fault impact modes at different time scales, and adaptively adjusting the window size according to the characteristics of different failure modes; S53. Constructing a time evolution model of failure mode based on the fault impact matrix: Among them, P(Q k |t) indicates that failure mode Q occurs at time step t k The probability of E k Indicates failure mode Q k The component set, w i Represents component v i In failure mode Q k The influence weight in , Z represents the normalization constant, exp represents the exponential function, ζ i Represents component v i The failure attenuation factor, Represents component v i The most recent time point when the abnormal state occurred; S54. Based on the multi-step prediction method, the failure probability of the future time step is calculated in a rolling manner, and the prediction result is dynamically corrected using the Bayesian time update strategy, so that the probability of occurrence of the failure mode is adaptively adjusted, and the time distribution of different failure modes is statistically analyzed to generate the failure probability trend of the future time step; S55. Generate a fault evolution trend diagram according to the failure probability trend of future time steps.

6. The reliability repair method for the whole fracturing pump based on dynamic fault analysis according to claim 1 is characterized in that: The S6 specifically includes: S61. Analyze the impact range and fault propagation intensity of different failure modes according to the fault evolution trend diagram, perform time-weighted calculation on each failure mode, and obtain the fault severity; S62. Based on Stackelberg game theory, a two-step optimization solution process is constructed. The first step is to dynamically adjust the repair priority based on the global repair goal of the system manager. The second step is to perform game calculations between different repair plans based on the cost constraints of the repair strategy executor, and use an iterative method to solve the Nash equilibrium point, so that the final repair plan achieves the optimal trade-off between cost and reliability. S63. Use the deep Q network to optimize the repair path, build a training framework based on the experience playback mechanism, input the fault mode, potential failure point and historical repair data, and search for the best repair path: Q * (s,a)=(1-η)Q(s,a)+η[ρ+ρmax a' Q(s′,a′)-τlog2π(a′∣s′)]; Among them, Q * (s,a) represents the expected return after executing action a in the current state s, η represents the learning rate, Q(s,a) represents the expected return before executing action a in the current state s, ρ represents the current instant reward, max a' Q(s',a') represents the maximum expected return of performing action a' in the next state s', τ represents the temperature parameter, and π(a'|s') represents the probability of selecting action a' in the next state s'.

Images