High-speed rail turnout fault diagnosis system and method based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a high-speed rail turnout fault diagnosis system and method based on artificial intelligence, and the system comprises a data feature extraction unit which is used for collecting multi-source heterogeneous data and extracting fault features and working condition compensation coefficients based on the multi-source heterogeneous data; the association graph construction unit is used for constructing and dynamically updating a component association graph; the association fusion analysis unit is used for carrying out iterative evidence reasoning and fusion analysis on the fault features based on a component association map, and outputting final fault probabilities of all components; and the fault positioning output unit is used for generating a maintenance guidance report. According to the method, the specific characteristics of the fault can be reflected in real time, and the reasoning intensity and range of different propagation paths are dynamically modulated according to the specific characteristics, so that the diagnosis focus can be intelligently adjusted according to the fault mode, and the positioning pertinence and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a high-speed railway turnout fault diagnosis system and method based on artificial intelligence. Background Technology

[0002] High-speed railway turnouts are critical switching devices on railway tracks, and their condition directly affects train operation safety and efficiency. Turnouts have a complex structure, containing a large number of mechanical and electrical components, and are prone to various failure modes (such as jamming, asynchrony, and wear).

[0003] Traditional and existing machine learning-based diagnostic methods can be mainly categorized as follows: first, single-signal alarms based on fixed thresholds (e.g., operating current exceeding limits); second, data-driven single-classification models (e.g., training a classifier using vibration signals); and third, simple association reasoning based on static knowledge graphs. These methods have gradually revealed the following inherent shortcomings in practical applications:

[0004] 1. Using static models or fixed parameters makes it impossible to dynamically adjust the reasoning logic based on the specific characteristics of the fault (such as the severity of synchronization error or the concentration of vibration energy distribution), resulting in a lack of targeted diagnosis for faults of different degrees and natures, which can easily lead to false alarms or missed alarms.

[0005] 2. Existing methods often perform "one-off" or superficial analyses, lacking a closed-loop reasoning mechanism that allows initial evidence to deeply interact and integrate with the system's physical correlation model and real-time environmental factors through multiple iterations. This makes it difficult to correct initial misjudgments during the diagnostic process, and it is impossible to effectively synthesize conflicting evidence from multiple sources. The final output of fault confidence often has uncertainty, making it difficult to support accurate component-level maintenance decisions.

[0006] Therefore, an artificial intelligence-based high-speed railway turnout fault diagnosis system and method are provided. Summary of the Invention

[0007] The purpose of this invention is to provide a high-speed railway turnout fault diagnosis system and method based on artificial intelligence, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, on the one hand, the present invention aims to provide a high-speed railway turnout fault diagnosis system based on artificial intelligence, comprising:

[0009] A data feature extraction unit is used to synchronously collect multi-source heterogeneous data during the turnout operation process, and extract fault features and working condition compensation coefficients based on the multi-source heterogeneous data.

[0010] A component association graph construction unit is used to construct and dynamically update a component association graph.

[0011] The correlation fusion analysis unit is used to perform iterative evidence reasoning and fusion analysis on the fault characteristics based on the component correlation map. By integrating the initial diagnostic evidence, the correlation between components and environmental factors in multiple iterations, the unit outputs the final failure probability of all components.

[0012] A fault location output unit is used to receive the final fault probability from the correlation fusion analysis unit and generate a maintenance guidance report based on the correlation map.

[0013] As a further improvement to this technical solution, the multi-source heterogeneous data includes action current curves, vibration acceleration signals, ambient temperature and humidity data, and action timing signals of traction points.

[0014] As a further improvement to this technical solution, the fault characteristics and operating condition compensation coefficients include:

[0015] Macroscopic resistance time-series characteristics extracted from the action current curve ;

[0016] The component physical state spectrum features extracted from the vibration acceleration signal after short-time Fourier transform. ;

[0017] Synchronization dynamic error characteristics calculated based on the action timing signals of each traction point ;

[0018] The operating condition compensation coefficient is obtained by querying a pre-set temperature-friction coefficient lookup table based on ambient temperature. .

[0019] As a further improvement to this technical solution, in the association map construction unit, the component association map includes a set of nodes for turnout mechanical components. A set of edges used to represent the physical connections between components. and the weight matrix used to represent the strength of the coupling relationship The specific construction steps are as follows:

[0020] S21. Based on the turnout mechanical structure, a unique node identifier is defined for each monitorable component. The node set V includes slide plate nodes, locking point nodes, action rod nodes, and indicator rod nodes.

[0021] S22. Based on kinematic analysis and dynamic simulation, establish different types of edges between nodes. Specifically, it includes:

[0022] Physical contact edges are used to characterize the support and contact relationships between components;

[0023] Motion transmission edges are used to characterize motion driving relationships;

[0024] Force transmission edge, used to characterize the stress transmission path;

[0025] Synchronization constraint edges are used to characterize the requirements for synchronized actions;

[0026] S23. For each edge, calculate the weight of the coupled edge using a hybrid model:

[0027] ;

[0028] in, For nodes and nodes Edge weights; Physical weights based on multibody dynamics simulation; Statistical weights are based on historical fault data; This is the preset fusion coefficient.

[0029] As a further improvement to this technical solution, the correlation fusion analysis unit includes an initial diagnosis module, a multi-modal evidence propagation module, an environmental compensation module, and a global convergence decision module;

[0030] The initial diagnosis module is used to receive the fault features and process them through the deep learning classifier. The initial failure probabilities of all components are generated, and simultaneously, based on the aforementioned synchronous dynamic error characteristics... and the spectral characteristics of the physical state of components Generate a global correction factor;

[0031] In each iteration, the multi-mode evidence propagation module is used to update the failure probability of all nodes based on the probability updated in the previous round, the global correction factor, and the association graph, according to the predefined propagation strategy corresponding to the edge type, so as to obtain the propagated failure probability.

[0032] The environmental compensation module is used to adjust the compensation coefficient based on the operating conditions in each iteration. The failure probability after propagation is corrected to obtain the failure probability after environmental compensation.

[0033] The global convergence decision module receives the fault probability after environmental compensation after each iteration and judges whether the iteration has converged according to preset logic. Based on the convergence judgment result, it decides whether to continue the next round of iteration or terminate the iteration process. After the iteration ends, it outputs the final fault probability.

[0034] As a further improvement to this technical solution, the global correction factor is a synchronization influence factor. and vibration evidence concentration factor .

[0035] As a further improvement to this technical solution, in the multi-mode evidence propagation module, based on the edge set... Different types of [the virus / organization] employ different propagation strategies modulated by a global correction factor, specifically:

[0036] For the motion transmission edge, its propagation intensity is affected by the synchronization factor. Modulation, employing a fault tracing mode;

[0037] For the force transmission edge and the physical contact edge, the diffusion range is affected by the concentration factor of vibration evidence. Constraints are applied, and a fault propagation mode is adopted;

[0038] For synchronously constrained edges, their balance strength is simultaneously affected by the synchronicity factor. and vibration evidence concentration factor The combined effects of these factors necessitate a two-way balance model.

[0039] As a further improvement to this technical solution, the global convergence decision module determines whether the iteration has converged according to preset logic, and decides whether to continue to the next round of iteration or terminate the iteration process based on the convergence determination result. The specific logic judgment process is as follows:

[0040] S341. Convergence Judgment: Based on the fault probability distribution after environmental compensation and the fault probability distribution of the previous iteration, calculate the overall uncertainty measure of the system. ;

[0041] ;

[0042] In the formula, For the current iteration, the fault probability distribution is based on environmental compensation. For the previous iteration, the fault probability distribution was based on environmental compensation. For nodes ;

[0043] S342, Iterative Control: If ,in If the system has converged to a preset convergence threshold or the number of iterations reaches a preset maximum value, then the final decision is triggered.

[0044] S343, Final Decision: When the system converges, the current... Determined as the final failure probability distribution and based on Generate diagnostic results;

[0045] S344, Circular Feedback: If the system does not converge, then... The current failure probability distribution, used as the basis for the next iteration, is fed back to the multi-mode evidence propagation module to begin a new iteration.

[0046] As a further improvement to this technical solution, the fault location output unit generates a maintenance guidance report specifically as follows:

[0047] S41. Faulty Component Location: The probability value of each node in the final fault probability distribution is compared with a preset first threshold. Second threshold Comparison, among which ;

[0048] All satisfied nodes Mark the component as a confirmed fault and add it to the confirmed fault set. ;

[0049] All satisfied nodes Mark the component as a fault to be confirmed and add it to the set of faults to be confirmed. ;

[0050] S42. Impact Scope Analysis: Based on the aforementioned correlation map, For each confirmed faulty component, its neighbor nodes in the graph are traversed, and related components with a fault propagation risk higher than a preset level are identified based on edge weights, generating a list of secondary affected components.

[0051] S43. Report Structure Generation: Integrating the aforementioned confirmed fault set Set of faults to be confirmed The system automatically generates a maintenance guidance report based on the list of secondary affected components. The maintenance guidance report includes at least the following: a summary of the fault location, an explanation of the associated impacts, suggestions for maintenance priorities, and recommended testing items.

[0052] On the other hand, the present invention provides an artificial intelligence-based high-speed railway turnout fault diagnosis method, used in any one of the above-mentioned artificial intelligence-based high-speed railway turnout fault diagnosis systems, comprising the following steps:

[0053] S1. Synchronously collect multi-source heterogeneous data during a single complete operation of the turnout, and extract fault characteristics and working condition compensation coefficients based on the multi-source heterogeneous data.

[0054] S2. Construct and maintain the association map of turnout components, and update the edge weights in the weight matrix based on historical fault data;

[0055] S3. Input the fault features into the trained deep learning classifier to obtain the initial fault probability distribution of all components. At the same time, obtain the global correction factor based on the fault features.

[0056] S4. Starting from the initial failure probability distribution, based on the component association map and global correction factor, perform multiple rounds of iterative evidence reasoning and fusion analysis until the preset convergence condition is met, and output the final failure probability distribution.

[0057] S5. Based on the final failure probability distribution and combined with preset thresholds, locate the faulty component and generate a structured maintenance guidance report.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] 1. In a high-speed railway turnout fault diagnosis system and method based on artificial intelligence, a global correction factor dynamically generated by synchronization error and vibration spectrum characteristics is introduced, which can reflect the specific characteristics of the fault in real time and dynamically modulate the inference intensity and range of different propagation paths accordingly. This realizes the leap from "static graph inference" to "evidence feature-driven dynamic inference", which enables the diagnostic focus to be intelligently adjusted according to the fault mode, thereby improving the pertinence and accuracy of the location.

[0060] 2. In a high-speed railway turnout fault diagnosis system and method based on artificial intelligence, an iterative evidence reasoning and fusion analysis mechanism is adopted to effectively correct the deviation of the initial diagnosis, eliminate the uncertainty of single-source evidence, and finally output a stable and consistent final fault probability distribution. It simulates the meticulous thinking process of human experts who "boldly hypothesize and verify from multiple sources" and realizes the continuous self-optimization and verification of the diagnostic conclusion in an algorithmic form, fundamentally improving the reliability and confidence of fault location in complex systems. Attached Figure Description

[0061] Figure 1 This is an overall flowchart of the present invention;

[0062] Figure 2 This is a flowchart illustrating the overall method of the present invention;

[0063] The meanings of the labels in the diagram are as follows:

[0064] 1. Data feature extraction unit; 2. Association graph construction unit; 3. Association fusion analysis unit; 31. Initial diagnosis module; 32. Multimodal evidence propagation module; 33. Environmental compensation module; 34. Global convergence decision module; 4. Fault location output unit. Detailed Implementation

[0065] 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.

[0066] Example 1: Please refer to Figure 1 As shown, an artificial intelligence-based high-speed railway turnout fault diagnosis system is provided, including a data feature extraction unit 1, an association map construction unit 2, an association fusion analysis unit 3, and a fault location output unit 4.

[0067] Among them, the data feature extraction unit 1 is used to synchronously collect multi-source heterogeneous data during the turnout operation process, and extract fault features and working condition compensation coefficients based on the multi-source heterogeneous data; the multi-source heterogeneous data includes operating current curves, vibration acceleration signals, ambient temperature and humidity data, and operating timing signals of traction points.

[0068] These data are collected synchronously by installing sensors under the turnout control cabinet, each traction point, and key slide plates. Specifically, the operating current curve is obtained through the turnout control cabinet at a sampling frequency of 1kHz. Six triaxial accelerometers are installed on three slide plates and three locking points to obtain vibration acceleration signals at a sampling frequency of 5kHz. The timing signal of the traction point is obtained by recording the precise timestamps of the start and end of the action at each traction point. Environmental temperature and humidity data are obtained from the field environmental sensors.

[0069] Fault characteristics and operating condition compensation coefficients include:

[0070] Macroscopic resistance time-series characteristics extracted from the action current curve Specifically, after denoising the operating current, its effective value, peak value, and dynamic time warping (DTW) distance from the standard curve are calculated.

[0071] The component physical state spectrum features extracted from the vibration acceleration signal after short-time Fourier transform. Specifically, a short-time Fourier transform (STFT) with a Hanning window is performed on each vibration acceleration signal to extract the amplitude energy of each octave within the 200Hz-1500Hz frequency band as a feature.

[0072] Synchronization dynamic error characteristics calculated based on the action timing signals of each traction point Specifically, the standard deviation of the start time of each traction point (target <10ms) and the maximum difference in the duration of the action are calculated.

[0073] The operating condition compensation coefficient is obtained by querying a pre-set temperature-friction coefficient lookup table based on ambient temperature. Specifically: (T<0℃) (0℃≤T≤30℃), This is used to compensate for the effect of changes in the friction coefficient on resistance;

[0074] Association Graph Construction Unit 2 is used to construct and dynamically update the component association graph;

[0075] In the association graph construction unit 2, the component association graph includes the node set of turnout mechanical components. A set of edges used to represent the physical connections between components. and the weight matrix used to represent the strength of the coupling relationship The coupling weight matrix of the edge In the matrix, elements Based on a joint calibration using historical maintenance data and physical simulation analysis, the components are characterized. Fault conditions of components The intensity of the impact, and the specific construction steps are as follows:

[0076] S21. Based on the turnout mechanical structure, define a unique node identifier for each monitorable component. The node set V includes slide plate nodes, locking point nodes, action rod nodes, and indicator rod nodes.

[0077] In this embodiment, 18 nodes are defined: 6 slide plate nodes (S1-S6), 6 locking point nodes (L1-L6), 3 action lever nodes (A1-A3), and 3 indicator lever nodes (R1-R3);

[0078] S22. Based on kinematic analysis and dynamic simulation, establish different types of edges between nodes. Specifically, it includes:

[0079] Physical contact edges are used to characterize the support and contact relationships between components, such as the relationship between slide plate node S1 and switch rail component;

[0080] Motion transmission edges are used to characterize motion driving relationships, such as the connection from the action lever node A1 to the locking point nodes L1 and L2.

[0081] Force transmission edge is used to characterize the stress transmission path, such as the force transmitted from the locking point node L1 to the slide plate node S1.

[0082] Synchronization constraint edges are used to characterize the requirements for synchronous actions and are established between nodes that need to synchronize actions, such as L1 and L2, L3 and L4.

[0083] S23. For each edge, calculate the weight of the coupled edge using a hybrid model. :

[0084] ;

[0085] in, For nodes and nodes The edge weights are elements in the weight matrix; To establish physical weights based on multibody dynamics simulation, a multibody dynamics simulation model of a turnout is built, simulating its operation under typical working conditions, and statistically analyzing the nodes. and nodes The average interaction force or energy transfer efficiency between them is calculated. Based on the statistical weights of historical fault data, it analyzes the historical fault data to determine the weights of nodes. When a failure occurs, the node This is obtained by normalizing the conditional probability of associated faults occurring. The preset fusion coefficient, and .

[0086] In the weight matrix calculation step, different weight calculation strategies are used for different types of edges:

[0087] For the motion transfer edge Proportional to the efficiency coefficient of motion transmission;

[0088] For the force transmission edge Proportional to the stress concentration factor;

[0089] For synchronous constraint edges It is inversely proportional to the allowable synchronization error tolerance.

[0090] The dynamic updating of the component association map adopts a confidence-based weight adjustment strategy;

[0091] For the motion transfer edge The motion transmission efficiency was obtained through multibody dynamics simulation. Normalization .

[0092] Analyzing historical data, when When a fault occurs, The probability of a cascading failure is Therefore .

[0093] Pick Then the edge weight .

[0094] The correlation fusion analysis unit 3 is used to perform iterative evidence reasoning and fusion analysis on fault characteristics based on the component correlation map. By integrating the initial diagnostic evidence, the correlation between components and environmental factors in multiple iterations, it outputs the final failure probability of all components.

[0095] The correlation fusion analysis unit 3 includes an initial diagnosis module 31, a multimodal evidence propagation module 32, an environmental compensation module 33, and a global convergence decision module 34;

[0096] The initial diagnosis module 31 is used to receive fault features and process them through a deep learning classifier. Generate the initial failure probabilities for all components, and simultaneously, based on the synchronous dynamic error characteristics... and the spectral characteristics of the physical state of components Generate a global correction factor;

[0097] It is a three-layer fully connected neural network with input as... , , The concatenated feature vectors are output as an 18-dimensional vector, which is then processed by Softmax to obtain the initial failure probability of each node.

[0098] The global correction factor is a synchronous influence factor. and vibration evidence concentration factor ;

[0099] Among them, the synchronicity influence factor Specifically:

[0100]

[0101] in , is a proportionality coefficient used to quantify the severity of overall system synchronization misalignment, and is set to 0.5; if If the vector norm is large, then Increase the intensity of source tracing.

[0102] Concentration factor of vibration evidence Specifically:

[0103]

[0104] In the formula, A simple two-layer perceptron is used to map spectral features to scalar evidence values; For the first The spectral characteristics of the physical state of each component; used to characterize the clarity and concentration of vibration evidence pointing to a specific faulty component; The higher the value, the more concentrated the vibration evidence is at a particular node.

[0105] The multi-mode evidence propagation module 32 is used to update the failure probability of all nodes in each iteration based on the probability updated in the previous round, the global correction factor, and the association graph, according to the predefined propagation strategy corresponding to the edge type, so as to obtain the propagated failure probability.

[0106] In the multi-modal evidence propagation module 32, based on the edge set Different types of [the virus / organization] employ different propagation strategies modulated by a global correction factor, specifically:

[0107] For the motion transmission edge, its propagation intensity is affected by the synchronization factor. Modulation employs a fault tracing mode; the greater the synchronization error, the higher the weight of the tracing reasoning.

[0108]

[0109] In the formula, For nodes The updated failure probability value after this round of evidence dissemination; For nodes The failure probability value at the start of the current iteration round. For the first iteration round, ; upstream node The probability of failure at the start of the current iteration round; upstream node The initial failure probability is determined by the deep learning classifier. Generate directly; For nodes ; For chain multiplication, on nodes All upstream nodes Multiply the contribution values ​​together; This difference represents the upstream node's value after iterative propagation. The change in the probability of failure relative to the initial diagnosis. A positive value indicates strengthening evidence, and a negative value indicates weakening evidence.

[0110] This rule simulates the reverse reasoning process of "following the vine to find the melon". When a downstream component When fault symptoms appear, the system will trace back along the motion transmission chain (from the driven component to the driving component) to check its upstream driving component. Does the failure probability of an upstream component increase significantly during iteration? If the probability of an upstream component increases (difference > 0), the failure probability of downstream components that have a direct motion connection with it will be amplified accordingly. The factor plays a modulating role here: the greater the synchronization error, the higher the weight of this tracing reasoning.

[0111] For the force transmission edge and the physical contact edge, the diffusion range is affected by the concentration factor of vibration evidence. Under the constraints of the fault propagation mode, the more concentrated the vibration evidence, the more conservative the assessment of fault propagation.

[0112]

[0113] In the formula, For nodes The updated failure probability value after this round of evidence dissemination; A fixed intensity factor (set to 0.5), a preset constant, is used to control the baseline intensity of fault propagation; For the summation operator, on the node All neighboring nodes The diffusion contributions are summed. Neighbors are defined as those directly connected via force-transmitting edges or physical contact edges. The node.

[0114] This rule simulates the diffusion process of "collateral damage." When a component fails (such as jamming or deformation), its physically connected or stress-bearing neighboring components are likely to be affected, experiencing abnormal stress or changes in condition. In the rule, the failure probabilities of neighboring nodes are weighted according to their connection weights, summed, and then multiplied by a coefficient. As a target node The potential increase in the probability of failure. The factor plays a constraining role here: if the vibration evidence is highly concentrated in a specific component, it indicates that the fault is likely limited to that component, and the system will suppress the spread of the fault probability to other components, thus making the diagnostic conclusion more focused.

[0115] For synchronously constrained edges, their balance strength is simultaneously affected by the synchronicity factor. and vibration evidence concentration factor The combined effects of these factors necessitate a two-way balance model.

[0116]

[0117] In the formula, This is a function for taking the average value;

[0118] This rule simulates the balancing process of "coordinated failures." For components with strict synchronization requirements (such as multiple locking points under the same traction point), a failure in one component almost inevitably leads to asynchrony or failure in another component. Therefore, the system should not view their failure probabilities independently, but rather force their probability values ​​to approach each other and tend towards balance. This avoids misjudging one component as the primary cause of failure and another as normal due to minor fluctuations in evidence, and is more consistent with the reality of their physical coupling.

[0119] In summary, the system no longer treats all connections equally, but adopts completely different propagation directions and update rules based on the physical meaning of the edges (motion transmission, force transmission, synchronization constraints), which enhances the physical rationality and accuracy of positioning. The tracing mode is used to accurately locate the root cause (e.g., the problem of the action rod); the diffusion mode is used to evaluate the impact (e.g., the impact of the slide plate jamming on the switch rail); and the balance mode is used to diagnose coordination failures (e.g., multiple locking points are not synchronized).

[0120] The environmental compensation module 33 is used to, in each iteration, calculate the compensation based on the operating condition coefficient. The failure probability after propagation is corrected to obtain the failure probability after environmental compensation.

[0121]

[0122] In the formula, This represents the failure probability after environmental compensation.

[0123] The global convergence decision module 34 is used to receive the fault probability after environmental compensation after each iteration and judge whether the iteration has converged according to the preset logic. Based on the convergence judgment result, it decides whether to continue the next round of iteration or terminate the iteration process. After the iteration ends, it outputs the final fault probability.

[0124] In the global convergence decision module 34, the iteration is judged according to preset logic to determine whether it has converged, and based on the convergence judgment result, it is decided whether to continue to the next round of iteration or terminate the iteration process. The specific logic judgment process is as follows:

[0125] To prevent repeated changes in diagnostic conclusions due to random fluctuations or oscillations during the iteration process, and to ensure that the final output fault probability distribution is stable and reliable, the overall change in the probability distribution after each iteration (uncertainty measurement) is monitored. Only when the change is small enough ( Convergence is determined only when the iteration is complete, to avoid terminating the iteration too early or too late.

[0126] S341. Convergence Judgment: Based on the fault probability distribution after environmental compensation and the fault probability distribution of the previous iteration, calculate the overall uncertainty measure of the system. ;

[0127] ;

[0128] In the formula, For the current iteration, the fault probability distribution is based on environmental compensation. For the previous iteration, the fault probability distribution was based on environmental compensation. For nodes ;

[0129] S342, Iterative Control: If ,in If the system reaches a preset convergence threshold, or if the number of iterations reaches a preset maximum value, then the system is considered to have converged, and the final decision is triggered; convergence threshold Maximum number of iterations ;

[0130] S343, Final Decision: When the system converges, the current... Determined as the final failure probability distribution and based on Generate diagnostic results;

[0131] S344, Circular Feedback: If the system does not converge, then... The current failure probability distribution, used as the basis for the next iteration, is fed back to the multimodal evidence propagation module 32 to begin a new iteration.

[0132] This convergence decision mechanism prevents oscillation divergence and corrects early errors. Without convergence control, certain complex failure modes may cause the probability distribution to oscillate repeatedly among multiple possible failure hypotheses. The convergence decision mechanism can identify such oscillations. (The value may increase instead of decrease or remain at a high level), and the current optimal solution is given when the maximum number of iterations is reached. At the same time, the initial diagnosis may be biased due to noisy data. Through multiple rounds of iterative propagation and convergence judgment, the system can gradually correct early errors and eventually converge to the correct fault mode.

[0133] Furthermore, the iterative process is visualized, allowing engineers to view the U-value change curve after each iteration and intuitively understand the convergence process of the diagnosis.

[0134] Ideal convergence: The value monotonically decreases with each iteration, eventually falling below... .

[0135] Fast convergence: first 2-3 rounds A significant drop in the value indicates a clear fault characteristic.

[0136] Slow convergence: A slow decrease in the value indicates a complex failure mode, requiring further reasoning.

[0137] Non-convergence: A value fluctuating at a high level suggests a possible anomaly (such as sensor failure or a new, unknown fault).

[0138] The fault location output unit 4 is used to receive the final fault probability from the correlation fusion analysis unit 3 and generate a maintenance guidance report based on the correlation map.

[0139] The maintenance guidance report generated in fault location output unit 4 is as follows:

[0140] S41. Faulty Component Location: The probability value of each node in the final fault probability distribution is compared with a preset first threshold. Second threshold Comparison, among which In this embodiment, , ;

[0141] All satisfied nodes Mark the component as a confirmed fault and add it to the confirmed fault set. ;

[0142] All satisfied nodes Mark the component as a fault to be confirmed and add it to the set of faults to be confirmed. ;

[0143] S42. Impact Scope Analysis: Based on the correlation map, For each confirmed faulty component, its neighbor nodes in the graph are traversed, and related components with a fault propagation risk higher than a preset level are identified based on edge weights, generating a list of secondary affected components.

[0144] S43. Report Structure Generation: Comprehensive Confirmation of Fault Sets Set of faults to be confirmed The system also generates a list of secondary affected components and automatically produces a maintenance guidance report. The maintenance guidance report should include at least the following: a summary of the fault location, an explanation of the associated effects, recommendations for maintenance priorities, and recommended testing items.

[0145] Among them, the fault location summary is used to clearly list the confirmed and unconfirmed faulty components; the related impact description is used to explain the possible propagation path of the fault and the affected components; the maintenance priority suggestion is used to give a maintenance sequence suggestion based on the fault probability value, component criticality and impact range; and the recommended test items are used to recommend key physical characteristics or test items for the unconfirmed faulty components and secondary affected components.

[0146] The specific implementation process after reporting a "blockage" alarm during turnout switching in this embodiment is as follows:

[0147] System startup diagnostics: Extract current and other data from this operation, and calculate... (Friction increases at low temperatures) The display shows a delay of 120ms for the second traction point.

[0148] Initial diagnosis: Output show (Second traction point action lever) probability , , (Corresponding to the locking point) probability approximately .

[0149] Iterative fusion:

[0150] Round 1: The increase is due to large synchronization errors. In the source tracing mode, evidence comes from... , Towards Convergence, The probability rises to Environmental compensation Slightly reduce all probabilities.

[0151] Round Two: Probability Continues to Advance concentrated, Probability reached Others decreased. Changes It is already less than ;

[0152] Output result: System determination To confirm the fault (probability greater than 0.8), the report states "Second traction point action lever ( The fault caused synchronization failure, which, combined with the low ambient temperature, led to jamming. It is recommended to prioritize checking... Lubrication and mechanical connection.

[0153] Example 2: Please refer to Figure 2 As shown, an artificial intelligence-based high-speed railway turnout fault diagnosis method is provided, which is used in any of the above-mentioned artificial intelligence-based high-speed railway turnout fault diagnosis systems, and includes the following steps:

[0154] S1. Synchronously collect multi-source heterogeneous data during a single complete operation of the turnout, and extract fault characteristics and working condition compensation coefficients based on the multi-source heterogeneous data.

[0155] S2. Construct and maintain the association map of turnout components, and update the edge weights in the weight matrix based on historical fault data;

[0156] S3. Input the fault features into the trained deep learning classifier to obtain the initial fault probability distribution of all components. At the same time, obtain the global correction factor based on the fault features.

[0157] S4. Starting from the initial failure probability distribution, based on the component association map and global correction factor, perform multiple rounds of iterative evidence reasoning and fusion analysis until the preset convergence condition is met, and output the final failure probability distribution.

[0158] S5. Based on the final failure probability distribution and combined with preset thresholds, locate the faulty component and generate a structured maintenance guidance report.

[0159] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A high-speed railway turnout fault diagnosis system based on artificial intelligence, characterized in that, include: Data feature extraction unit (1), the data feature extraction unit (1) is used to synchronously collect multi-source heterogeneous data during the turnout operation process, and extract fault features and working condition compensation coefficients based on the multi-source heterogeneous data; Association graph construction unit (2), which is used to construct and dynamically update the component association graph; The correlation fusion analysis unit (3) is used to perform iterative evidence reasoning and fusion analysis on the fault characteristics based on the component correlation map. By integrating the initial diagnostic evidence, the correlation between components and environmental factors in multiple iterations, the final failure probability of all components is output. Fault location output unit (4) is used to receive the final fault probability from the correlation fusion analysis unit (3) and generate a maintenance guidance report based on the correlation map.

2. The high-speed railway turnout fault diagnosis system based on artificial intelligence according to claim 1, characterized in that: The multi-source heterogeneous data includes action current curves, vibration acceleration signals, ambient temperature and humidity data, and action timing signals of traction points.

3. The high-speed railway turnout fault diagnosis system based on artificial intelligence according to claim 2, characterized in that: The fault characteristics and operating condition compensation coefficients include: Macroscopic resistance time-series characteristics extracted from the action current curve ; The component physical state spectrum features extracted from the vibration acceleration signal after short-time Fourier transform. ; Synchronization dynamic error characteristics calculated based on the action timing signals of each traction point ; The operating condition compensation coefficient is obtained by querying a pre-set temperature-friction coefficient lookup table based on ambient temperature. .

4. The high-speed railway turnout fault diagnosis system based on artificial intelligence according to claim 3, characterized in that: In the association graph construction unit (2), the component association graph includes a set of nodes for the turnout mechanical components. A set of edges used to represent the physical connections between components. and the weight matrix used to represent the strength of the coupling relationship The specific construction steps are as follows: S21. Based on the turnout mechanical structure, a unique node identifier is defined for each monitorable component. The node set V includes slide plate nodes, locking point nodes, action rod nodes, and indicator rod nodes. S22. Based on kinematic analysis and dynamic simulation, establish different types of edges between nodes. Specifically, it includes: Physical contact edges are used to characterize the support and contact relationships between components; Motion transmission edges are used to characterize motion driving relationships; Force transmission edge, used to characterize the stress transmission path; Synchronization constraint edges are used to characterize the requirements for synchronized actions; S23. For each edge, calculate the weight of the coupled edge using a hybrid model: ； in, For nodes and nodes Edge weights; Physical weights based on multibody dynamics simulation; Statistical weights are based on historical fault data; This is the preset fusion coefficient.

5. The high-speed railway turnout fault diagnosis system based on artificial intelligence according to claim 4, characterized in that: The correlation fusion analysis unit (3) includes an initial diagnosis module (31), a multimodal evidence propagation module (32), an environmental compensation module (33), and a global convergence decision module (34). The initial diagnosis module (31) is used to receive the fault features and process them through the deep learning classifier. The initial failure probabilities of all components are generated, and simultaneously, based on the aforementioned synchronous dynamic error characteristics... and the spectral characteristics of the physical state of components Generate a global correction factor; In each iteration, the multi-mode evidence propagation module (32) is used to update the failure probability of all nodes based on the probability updated in the previous round, the global correction factor and the association graph, according to the predefined propagation strategy corresponding to the edge type, so as to obtain the propagated failure probability. The environmental compensation module (33) is used to perform compensation based on the operating condition coefficient in each iteration. The failure probability after propagation is corrected to obtain the failure probability after environmental compensation. The global convergence decision module (34) is used to receive the fault probability after environmental compensation after each iteration and judge whether the iteration has converged according to the preset logic. Based on the convergence judgment result, it decides whether to continue the next round of iteration or terminate the iteration process. After the iteration ends, it outputs the final fault probability.

6. The high-speed railway turnout fault diagnosis system based on artificial intelligence according to claim 5, characterized in that: The global correction factor is a synchronous influence factor. and vibration evidence concentration factor .

7. The high-speed railway turnout fault diagnosis system based on artificial intelligence according to claim 6, characterized in that: In the multi-mode evidence propagation module (32), based on the edge set Different types of [the virus / organization] employ different propagation strategies modulated by a global correction factor, specifically: For the motion transmission edge, its propagation intensity is affected by the synchronization factor. Modulation, employing a fault tracing mode; For the force transmission edge and the physical contact edge, the diffusion range is affected by the concentration factor of vibration evidence. Constraints are applied, and a fault propagation mode is adopted; For synchronously constrained edges, their balance strength is simultaneously affected by the synchronicity factor. and vibration evidence concentration factor The combined effects of these factors necessitate a two-way balance model.

8. The high-speed railway turnout fault diagnosis system based on artificial intelligence according to claim 7, characterized in that: In the global convergence decision module (34), the iteration is judged according to the preset logic to determine whether it has converged, and the decision is made on whether to continue the next round of iteration or terminate the iteration process based on the convergence judgment result. The specific logic judgment process is as follows: S341. Convergence Judgment: Based on the fault probability distribution after environmental compensation and the fault probability distribution of the previous iteration, calculate the overall uncertainty measure of the system. ; ； In the formula, For the current iteration, the fault probability distribution is based on environmental compensation. For the previous iteration, the fault probability distribution was based on environmental compensation. For nodes ; S342, Iterative Control: If ,in If the system has converged to a preset convergence threshold or the number of iterations reaches a preset maximum value, then the final decision is triggered. S343, Final Decision: When the system converges, the current... Determined as the final failure probability distribution and based on Generate diagnostic results; S344, Circular Feedback: If the system does not converge, then... The current failure probability distribution, as the basis for the next iteration, is fed back to the multimodal evidence propagation module (32) to begin a new iteration.

9. The high-speed railway turnout fault diagnosis system based on artificial intelligence according to claim 8, characterized in that: The fault location output unit (4) generates a maintenance guidance report as follows: S41. Faulty Component Location: The probability value of each node in the final fault probability distribution is compared with a preset first threshold. Second threshold Comparison, among which ; All satisfied nodes Mark the component as a confirmed fault and add it to the confirmed fault set. ; All satisfied nodes Mark the component as a fault to be confirmed and add it to the set of faults to be confirmed. ; S42. Impact Scope Analysis: Based on the aforementioned correlation map, For each confirmed faulty component, its neighbor nodes in the graph are traversed, and related components with a fault propagation risk higher than a preset level are identified based on edge weights, generating a list of secondary affected components. S43. Report Structure Generation: Integrating the aforementioned confirmed fault set Set of faults to be confirmed The system automatically generates a maintenance guidance report based on the list of secondary affected components. The maintenance guidance report includes at least the following: a summary of the fault location, an explanation of the associated impacts, suggestions for maintenance priorities, and recommended testing items.

10. A method for diagnosing high-speed railway turnout faults based on artificial intelligence, used in a high-speed railway turnout fault diagnosis system based on artificial intelligence as described in any one of claims 1-9, characterized in that: Includes the following steps: S1. Synchronously collect multi-source heterogeneous data during a single complete operation of the turnout, and extract fault characteristics and working condition compensation coefficients based on the multi-source heterogeneous data. S2. Construct and maintain the association map of turnout components, and update the edge weights in the weight matrix based on historical fault data; S3. Input the fault features into the trained deep learning classifier to obtain the initial fault probability distribution of all components. At the same time, obtain the global correction factor based on the fault features. S4. Starting from the initial failure probability distribution, based on the component association map and global correction factor, perform multiple rounds of iterative evidence reasoning and fusion analysis until the preset convergence condition is met, and output the final failure probability distribution. S5. Based on the final failure probability distribution and combined with preset thresholds, locate the faulty component and generate a structured maintenance guidance report.

Images