AI Fault Diagnosis Method and System for Suspended Rail Transit System

Abstract

The present invention provides an AI fault diagnosis method and system for a suspended rail transit system. By combining supervised learning and unsupervised learning strategies, the accuracy and generalization ability of fault diagnosis are significantly improved. The fault diagnosis neural network is initially trained using supervised training data to ensure the effective identification of known fault types by the model. At the same time, unsupervised training data is introduced and optimized through a shared fault diagnosis network, which not only enhances the model's diagnostic ability for unknown or rare fault types but also realizes cross-domain diagnosis of cross-type faults, greatly improving the comprehensiveness and robustness of fault diagnosis. By continuously iteratively optimizing the fault diagnosis neural network, the target fault state path diagnosis results corresponding to the operation data of the target rail transit system can be accurately generated, ensuring the safe and efficient operation of the suspended rail transit system and effectively reducing the misdiagnosis rate and missed diagnosis rate.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and more particularly, to an AI fault diagnosis method and system for a suspended rail transit system. Background Art

[0002] As a new type of urban transportation mode, the suspended rail transit system has been widely used in more and more cities due to its advantages such as small floor area, high operation efficiency, and strong environmental adaptability. However, with the long-term operation of the suspended rail transit system, various faults will inevitably occur in its complex mechanical structure and electrical system. These faults will not only affect the normal operation of the system but also pose a threat to the safety of passengers. Therefore, it is particularly important to perform timely and accurate fault diagnosis on the suspended rail transit system.

[0003] Traditional fault diagnosis methods often rely on manual experience and simple rule judgments. In the face of complex and changing fault situations, it is often difficult to ensure the accuracy and efficiency of diagnosis. With the rapid development of artificial intelligence technology, especially the remarkable achievements of deep learning technology in various fields, applying AI technology to the fault diagnosis of suspended rail transit systems has become an important research direction.

[0004] However, most of the existing AI fault diagnosis methods only rely on supervised learning, that is, using a large amount of labeled fault data to train the model. Although this method can improve the accuracy of fault diagnosis to a certain extent, limited by the quantity and variety of the labeled data, the generalization ability of the model is often insufficient, and it is difficult to handle various unknown or rare faults that occur in actual operation. In addition, there may be correlations between different fault types, but the existing methods often ignore the diagnostic ability of such cross-type faults, resulting in poor diagnostic effects when facing complex faults. Summary of the Invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, embodiments of the present invention provide an AI fault diagnosis method for a suspended rail transit system, and the method includes:

[0006] Obtain supervised training data and unsupervised training data for the suspended rail transit system, where the supervised training data is sample rail transit system operation data carrying fault state path annotation data;

[0007] Load the supervised training data into a fault diagnosis neural network, generate a first fault state path diagnosis result of the supervised training data, and determine a first fault diagnosis error based on the first fault state path diagnosis result and the fault state path annotation data;

[0008] Load the unsupervised training data into the fault diagnosis neural network to generate the second fault state path diagnosis result of the unsupervised training data. Optimize the second fault state path diagnosis result using the shared fault diagnosis network to generate the optimized second fault state path diagnosis result. The shared fault diagnosis network has cross-domain fault diagnosis performance for cross-type faults, and determine the second fault diagnosis error based on the optimized second fault state path diagnosis result.

[0009] Optimize the fault diagnosis neural network based on the first fault diagnosis error and the second fault diagnosis error, and perform fault diagnosis on any input target rail transit system operation data based on the optimized fault diagnosis neural network to generate the target fault state path diagnosis result corresponding to the target rail transit system operation data.

[0010] In another aspect, an AI fault diagnosis system for a suspended rail transit system according to an embodiment of the present invention includes a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiments of the present application significantly improve the accuracy and generalization ability of fault diagnosis by combining supervised learning and unsupervised learning strategies. The fault diagnosis neural network is initially trained using supervised training data to ensure the effective identification of known fault types by the model. At the same time, the introduction of unsupervised training data and optimization through the shared fault diagnosis network not only enhances the model's diagnostic ability for unknown or rare fault types but also realizes cross-domain diagnosis of cross-type faults, greatly improving the comprehensiveness and robustness of fault diagnosis. By continuously iteratively optimizing the fault diagnosis neural network, the target fault state path diagnosis result corresponding to the target rail transit system operation data can be accurately generated, ensuring the safe and efficient operation of the suspended rail transit system, effectively reducing the misdiagnosis rate and missed diagnosis rate, and improving the overall operation and maintenance efficiency and reliability of the suspended rail transit system. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a schematic execution flow diagram of an AI fault diagnosis method for a suspended rail transit system provided by an embodiment of the present invention.

[0013] Figure 2 is a schematic hardware architecture diagram of an AI fault diagnosis system for a suspended rail transit system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be specifically described below with reference to the accompanying drawings of the specification.Figure 1 It is a schematic flowchart of an AI fault diagnosis method for a suspended rail transit system provided by an embodiment of the present invention. The AI fault diagnosis method for the suspended rail transit system will be introduced in detail below.

[0015] Step S110: Obtain supervised training data and unsupervised training data for the suspended rail transit system. The supervised training data is sample rail transit system operation data carrying fault state path annotation data.

[0016] In this embodiment, the server can obtain training data for the suspended rail transit system, including supervised training data and unsupervised training data.

[0017] For the supervised training data, the server collects sample rail transit system operation data from multiple sources. For example, the server obtains data from the on-vehicle monitoring systems of each train in the suspended rail transit system. These on-vehicle monitoring systems continuously record various operation parameters of the train, such as the train speed, acceleration, braking pressure, motor current, voltage, connection status of each carriage, shock absorption parameters of the suspension system, etc. At the same time, during long-term work, operation and maintenance personnel have conducted detailed fault analysis and annotation on some historical operation data. For example, when a braking fault occurs in the train, a series of operation data before, during, and after the braking fault will be recorded in detail, and the fault state path will be marked. This fault state path annotation data may include the changes in various operation parameters during the whole process from the normal operation of the braking system to a slight decrease in braking efficiency (such as the initial stage of brake pad wear), and then to a significantly longer braking distance (severe braking fault). The server integrates these sample rail transit system operation data with detailed fault state path annotation data as the supervised training data.

[0018] For the unsupervised training data, the server also obtains it from multiple channels. On the one hand, the server obtains data from sensors along the line of the suspended rail transit system. These sensors are distributed in various parts such as the track, platform, and power supply system, and monitor information such as the vibration frequency of the track, the change in the distance between the platform and the train, and the stability of the power supply voltage. On the other hand, the server also obtains weather data from the meteorological department because weather conditions may affect the operation of the suspended rail transit system. For example, strong wind weather may affect the suspension stability of the train. These multi-dimensional data from different channels, without pre-set fault state path annotation, are collected by the server as unsupervised training data.

[0019] Step S120: Load the supervised training data into the fault diagnosis neural network, generate the first fault state path diagnosis result of the supervised training data, and determine the first fault diagnosis error based on the first fault state path diagnosis result and the fault state path annotation data.

[0020] In this embodiment, the fault diagnosis neural network includes a teaching deep learning network and a learning deep learning network. The server first loads the collected supervised training data into this fault diagnosis neural network.

[0021] Taking the fault diagnosis of the train braking system as an example, when a set of data in the supervised training data representing the relevant parameters of the train braking system during an operation is loaded into the fault diagnosis neural network, the teaching deep learning network will preliminarily process these data according to its existing knowledge structure (model parameters obtained through previous learning or pre-training), and then transfer the processing results to the learning deep learning network. The learning deep learning network further analyzes these data to generate the first fault state path diagnosis result for this set of data. Suppose the actual fault corresponding to this set of data is the reduction of braking efficiency caused by brake pad wear. The learning deep learning network may judge that there is an abnormality in the braking system based on factors such as the change trend of braking pressure and the fluctuation of motor current in the data, and infer that the fault may be in the stage where the brake pads start to wear, thus giving a fault state path diagnosis result from normal to slight wear of the brake pads.

[0022] Then, the server determines the first fault diagnosis error based on this first fault state path diagnosis result and the pre-stored fault state path annotation data. The server will compare the fault state path output by the learning deep learning network with the actual annotated fault state path. For example, in the actual annotated fault state path, the reduction of braking efficiency caused by brake pad wear should have reached a relatively serious level at a certain specific moment, while the diagnosis result of the learning deep learning network shows a less serious fault degree. The server determines the first fault diagnosis error of the learning deep learning network through a specific error calculation method, such as calculating the weighted sum of the differences between the two at each node on the fault state path. This first fault diagnosis error reflects the deviation degree of the learning deep learning network from the actual situation when performing fault diagnosis based on the supervised training data.

[0023] Step S130: Load the unsupervised training data into the fault diagnosis neural network to generate a second fault state path diagnosis result of the unsupervised training data, optimize the second fault state path diagnosis result by using a shared fault diagnosis network to generate an optimized second fault state path diagnosis result. The shared fault diagnosis network has cross-domain fault diagnosis performance for cross-type faults, and determine a second fault diagnosis error based on the optimized second fault state path diagnosis result.

[0024] In this embodiment, the server loads the unsupervised training data into the fault diagnosis neural network. For example, a set of data in the unsupervised training data is about the vibration frequency of the track along the suspended rail transit system and the stability of the power supply voltage under specific weather conditions (such as heavy rain). The teaching deep learning network and the learning deep learning network will process these data to generate a second fault state path diagnosis result. Suppose that in this process, the teaching deep learning network preliminarily judges that there may be a change in the friction force between the track and the train caused by water accumulation on the track according to the data pattern in previous similar weather conditions, and the learning deep learning network speculates that there may be a local moisture problem in the power supply line based on the fluctuation of the power supply voltage stability, thus forming a preliminary second fault state path diagnosis result.

[0025] Next, the server uses the shared fault diagnosis network to optimize this second fault state path diagnosis result. The shared fault diagnosis network has cross-domain fault diagnosis performance for cross-type faults, and it can integrate the correlation information between different types of faults and different regions (such as faults inside the train and faults outside the track).

[0026] For the case where the unsupervised training data is multi-dimensional rail transit operation data, the server extracts state knowledge graph data of multiple key state knowledge dimensions from the unsupervised training data. Taking the track and train power supply system as an example, analyze the characteristic structure of the unsupervised training data, and identify a list of key state knowledge dimensions related to the fault state path, such as the humidity state of the track, the insulation state of the power supply line, etc. For each key state knowledge dimension, create a corresponding state knowledge graph framework. For the track humidity state, the nodes in the state knowledge graph framework may include states such as dry, slightly wet, and severely waterlogged, and the edges represent the conversion conditions from the dry state to the slightly wet state (such as the amount of rainfall, the working condition of the drainage system, etc.). Map each data point in the unsupervised training data into the corresponding state knowledge graph framework according to the key state knowledge dimension. For each data point, determine its position in the state knowledge graph according to its value in the key state knowledge dimension of the track humidity state (such as the value measured by the humidity sensor), and generate a preliminarily mapped state knowledge graph. Define and label the nodes and edges in the preliminarily mapped state knowledge graph according to the operation rules and fault modes of the rail transit system. For the dry node in the track humidity state, label its normal operation range (humidity below a certain value), critical state range (humidity in a certain interval), and fault state range (humidity above a certain value). For the edge (such as the conversion edge from dry to slightly wet), label the conditional probability of this state conversion under different rainfall conditions. And based on the interaction between multiple key state knowledge dimensions (such as an increase in track humidity may affect the insulation performance of the power supply line), add cross-dimensional connections and interaction relationships in the preliminarily mapped state knowledge graph to generate the target state knowledge graph data. Extract the state knowledge graph data corresponding to each key state knowledge dimension from the target state knowledge graph data.

[0027] Taking the state knowledge graph data of multiple key state knowledge dimensions as the to-be-diagnosed operation data, and using the diagnosis result of the second fault state path as the training supervision data, a shared fault diagnosis network is utilized to output the fuzzy fault diagnosis results on multiple key state knowledge dimensions. For example, on the key state knowledge dimension of the track humidity state, the fuzzy fault diagnosis result may be the diagnosis result of the fault state path where track waterlogging exists with a certain probability, leading to an increase in the train running resistance; on the key state knowledge dimension of the power supply line insulation state, it may be the diagnosis result of the fault state path where the insulation performance decreases to a certain extent due to moisture. According to the diagnosis fusion mechanism, the fuzzy fault diagnosis results on multiple key state knowledge dimensions are fused to generate the optimized second fault state path diagnosis result. For each diagnosis node in the unsupervised training data, when the fuzzy fault diagnosis results on multiple key state knowledge dimensions are determined to match the generated fault state knowledge graph of the diagnosis node, the optimized second fault state path diagnosis result is generated according to the fault state knowledge graph of each diagnosis node in the unsupervised training data. For example, if the diagnosis results on multiple key state knowledge dimensions all indicate that there is a safety risk when the train runs in heavy rain weather (such as the combined effect of track waterlogging and power supply line moisture), then the optimized second fault state path diagnosis result will clearly reflect the formation path of this risk state.

[0028] Finally, the server determines the second fault diagnosis error of the taught deep learning network based on the taught fault state path diagnosis result and the optimized taught fault state path diagnosis result. Suppose the initially diagnosed result of the taught deep learning network for the train in heavy rain weather is only a minor power supply problem, while the optimized taught fault state path diagnosis result after passing through the shared fault diagnosis network shows that the train has both track waterlogging and power supply problems leading to running risks. The server calculates the fuzzy fault diagnosis error based on the taught fault state path diagnosis result and the optimized taught fault state path diagnosis result. For example, the fuzzy fault diagnosis error is determined by calculating the degree of difference between the two at the key nodes of the fault state path. At the same time, the transfer learning fault diagnosis error is calculated based on the taught fault state path diagnosis result and the taught fault state path diagnosis result. According to the current iteration round and the maximum iteration round of the fault diagnosis learning process, the first influence factor and the second influence factor are determined. Suppose the current iteration round is 3 and the maximum iteration round is 10, and the first influence factor is determined to be 0.3 and the second influence factor is determined to be 0.7 according to a specific calculation rule. Calculate the first weighted result of the first influence factor and the fuzzy fault diagnosis error, and the second weighted result of the second influence factor and the transfer learning fault diagnosis error. The sum of the first weighted result and the second weighted result is output as the second fault diagnosis error of the taught deep learning network.

[0029] Step S140: Optimize the fault diagnosis neural network according to the first fault diagnosis error and the second fault diagnosis error, and perform fault diagnosis on the arbitrarily input target rail transit system operation data based on the optimized fault diagnosis neural network to generate a target fault state path diagnosis result corresponding to the target rail transit system operation data.

[0030] In this embodiment, the fault diagnosis neural network is optimized according to the first fault diagnosis error and the second fault diagnosis error. For example, if the first fault diagnosis error indicates that there is a large deviation in the diagnosis of train braking system faults by the learning deep learning network under supervised training data, and the second fault diagnosis error indicates that there are also certain problems in the diagnosis of train faults under special weather conditions in unsupervised training data, the server will adjust the neuron weight parameters in the learning process of the learning deep learning network in the a-th iteration round (assuming a = 5) according to these errors. The server adjusts the connection weights between neurons according to a specific optimization algorithm, such as the backpropagation algorithm, according to the magnitude and direction of the errors, so that the learning deep learning network can more accurately judge the fault state path in subsequent diagnoses.

[0031] Then, optimize the neuron weight parameters of the teaching deep learning network according to the neuron weight parameters before the learning process of the learning deep learning network in the 5th iteration round and the neuron weight parameters in the learning process of the 5th iteration round. The teaching deep learning network will adjust its own knowledge structure according to the improvement of the learning deep learning network, so as to better guide the learning deep learning network to perform fault diagnosis in subsequent diagnosis processes.

[0032] When the server receives arbitrarily input target rail transit system operation data, such as a set of data on the suspension system and traction system of a train during high-speed driving, perform fault diagnosis on it based on the optimized fault diagnosis neural network. The teaching deep learning network first performs a preliminary analysis on the data and extracts key features in the data, such as the damping frequency of the suspension system and the motor torque of the traction system. Then these features are passed to the learning deep learning network, and the learning deep learning network generates a target fault state path diagnosis result corresponding to the target rail transit system operation data according to the optimized neuron weight parameters and the learned fault diagnosis knowledge. Assuming that according to the abnormal increase in the damping frequency of the suspension system and the instability of the motor torque of the traction system in the data, the learning deep learning network judges that there may be a shock absorber fault in the suspension system, and due to the abnormality of the suspension system, the shaking of the train during driving may be aggravated, which in turn affects the normal operation of the traction system, thus generating a target fault state path diagnosis result starting from the shock absorber of the suspension system, to the decrease in train driving stability, and then to the traction system being affected.

[0033] In addition, during the learning process of fault diagnosis that includes multiple iterative rounds, the server will also select candidate training data from the unsupervised training data based on the diagnostic results of the second fault state path of the unsupervised training data during the learning process of the a-th iterative round (e.g., a = 5). For example, based on the diagnostic results of the second fault state path of the unsupervised training data during the learning process of the 5th iterative round, the unsupervised training data that meets the target requirements is labeled as candidate training data. The unsupervised training data that meets the target requirements may include the following situations: the first X unsupervised training data (assuming X = 10) arranged in ascending order according to the diagnostic probability values based on the diagnostic results of the second fault state path, and these data may be the cases that were considered most likely to have misjudgments or unclear diagnoses in previous diagnoses; the first Y unsupervised training data (assuming Y = 5) arranged in descending order according to the average diagnostic likelihood ratio based on the diagnostic results of the second fault state path, where the average diagnostic likelihood ratio is the average of the diagnostic likelihood ratios of all diagnostic nodes in the diagnostic results of the second fault state path, and these data may be the cases that were considered most likely to have faults but need further confirmation in previous diagnoses; the first Z unsupervised training data (assuming Z = 8) arranged in descending order according to the likelihood ratio proportion based on the diagnostic results of the second fault state path, and the likelihood ratio proportion is the proportion of diagnostic nodes greater than the preset threshold in all diagnostic nodes in the diagnostic results of the second fault state path, and these data may be the cases that were considered to have obvious fault characteristics but need more in-depth analysis in previous diagnoses.

[0034] The server obtains the fault state path annotation data for adding training supervision data to the candidate training data, generates the active supervision training data during the learning process of the a-th iterative round (the 5th iterative round), and loads the active supervision training data into the supervision training data sequence called during the learning process of the (a + 1)-th (the 6th) iterative round. This can continuously improve the accuracy and reliability of the fault diagnosis neural network for the fault diagnosis of the suspended rail transit system.

[0035] Based on the above steps, the embodiments of the present application significantly improve the accuracy and generalization ability of fault diagnosis by combining supervised learning and unsupervised learning strategies. The fault diagnosis neural network is preliminarily trained using supervised training data to ensure the effective recognition of known fault types by the model; at the same time, unsupervised training data is introduced and optimized through the shared fault diagnosis network, which not only enhances the model's diagnostic ability for unknown or rare fault types but also realizes cross-domain diagnosis of cross-type faults, greatly improving the comprehensiveness and robustness of fault diagnosis. By continuously iteratively optimizing the fault diagnosis neural network, the target fault state path diagnosis result corresponding to the target rail transit system operation data can be accurately generated, ensuring the safe and efficient operation of the suspended rail transit system, effectively reducing the misdiagnosis rate and missed diagnosis rate, and improving the overall operation and maintenance efficiency and reliability of the suspended rail transit system.

[0036] In a possible implementation manner, step S130 includes:

[0037] Using the unsupervised training data as the operation data to be diagnosed and the second fault state path diagnosis result as the training supervision data, and generating the optimized second fault state path diagnosis result by using the shared fault diagnosis network.

[0038] In a possible implementation manner, the unsupervised training data is multi-dimensional rail transit operation data. Using the unsupervised training data as the operation data to be diagnosed and the second fault state path diagnosis result as the training supervision data, and generating the optimized second fault state path diagnosis result by using the shared fault diagnosis network includes:

[0039] Step S131, extracting multiple state knowledge graph data of multiple key state knowledge dimensions from the unsupervised training data, where the multiple key state knowledge dimensions correspond to multiple feature mapping spaces of the multi-dimensional rail transit operation data.

[0040] Step S132, using the multiple state knowledge graph data of the multiple key state knowledge dimensions as the operation data to be diagnosed and the second fault state path diagnosis result as the training supervision data, and outputting the fuzzy fault diagnosis results on the multiple key state knowledge dimensions by using the shared fault diagnosis network. The fuzzy fault diagnosis result on each key state knowledge dimension includes the fault state path diagnosis results of multiple state knowledge graph data that match the same key state knowledge dimension.

[0041] Step S133, fusing the fuzzy fault diagnosis results on the multiple key state knowledge dimensions according to the diagnosis fusion mechanism to generate the optimized second fault state path diagnosis result.

[0042] In a possible implementation manner, step S133 includes:

[0043] For each diagnostic node in the unsupervised training data, when the fuzzy fault diagnosis result on the multiple key state knowledge dimensions is determined to match the fault state knowledge graph generated by the diagnostic node, the optimized second fault state path diagnosis result is generated according to the fault state knowledge graphs of the respective diagnostic nodes in the unsupervised training data.

[0044] In this embodiment, the unsupervised training data is used as the operation data to be diagnosed, and at the same time, the previously obtained second fault state path diagnosis result is used as the training supervision data, and then the optimized second fault state path diagnosis result is generated by using the shared fault diagnosis network. The unsupervised training data here is multi-dimensional rail transit operation data, such as the operation parameter data of multiple subsystems of the train (including the power system, suspension system, braking system, etc.) and track-related equipment (such as track circuits, switches, etc.). These data have multiple feature mapping spaces, such as the mapping space between the train speed and the power output, the mapping space between the shock absorption effect of the suspension system and the track flatness, etc.

[0045] First, the server extracts state knowledge graph data for multiple key state knowledge dimensions from unsupervised training data. Taking the train power system as an example, the server analyzes the feature structure of the unsupervised training data and identifies key state knowledge dimensions related to the fault state path, such as motor speed, motor temperature, current stability, etc. For the key state knowledge dimension of motor speed, a corresponding state knowledge graph framework is created. In this framework, the nodes may include different intervals of the normal speed range (such as low-speed stable interval, high-speed stable interval), abnormal speed intervals (such as excessive speed fluctuations, speed below the lowest normal threshold, etc.), and the edges represent the conditions for transitioning from one interval to another, which may involve changes in the power load, instructions from the control system, etc. Then, each data point in the unsupervised training data is mapped to the corresponding state knowledge graph framework according to the key state knowledge dimension. For each data point, based on its value in the key state knowledge dimension of motor speed, the position of the data point in the state knowledge graph is determined, thus generating a preliminarily mapped state knowledge graph. Next, according to the operating rules and fault modes of the rail transit system, the nodes and edges in the preliminarily mapped state knowledge graph are defined and labeled. For the nodes in the normal speed range, the upper and lower limits of the normal operating speed are labeled, and for the edge from normal speed to excessive speed fluctuations, the factors that may cause this transition are labeled, such as the probability of motor failure, the state of the cooling system, etc. And based on the interaction between multiple key state knowledge dimensions (such as abnormal motor speed may affect motor temperature), cross-dimensional connections and interaction relationships are added to the preliminarily mapped state knowledge graph, thus generating the target state knowledge graph data. The state knowledge graph data corresponding to the key state knowledge dimension of motor speed is extracted from this target state knowledge graph data. Using the same method, operations are performed on other key state knowledge dimensions such as motor temperature and current stability to obtain state knowledge graph data for multiple key state knowledge dimensions.

[0046] After that, the server uses the state knowledge graph data for multiple key state knowledge dimensions as the data to be diagnosed during operation, and the second fault state path diagnosis result as the training supervision data, and uses the shared fault diagnosis network to output fuzzy fault diagnosis results on multiple key state knowledge dimensions. On the key state knowledge dimension of motor speed, based on the state knowledge graph data and the training supervision data, the shared fault diagnosis network may obtain a fault state path diagnosis result that the excessive motor speed fluctuations are caused by motor bearing wear with a certain probability, which is the fuzzy fault diagnosis result on this key state knowledge dimension. For the key state knowledge dimension of motor temperature, a fault state path diagnosis result that the increase in motor temperature may be due to a cooling system failure or a potential internal short circuit in the motor may be obtained. The fuzzy fault diagnosis result on each key state knowledge dimension includes the fault state path diagnosis results of multiple state knowledge graph data that match the same key state knowledge dimension.

[0047] Finally, the server fuses the fuzzy fault diagnosis results on multiple key state knowledge dimensions according to the diagnosis fusion mechanism to generate an optimized second fault state path diagnosis result. For each diagnosis node in the unsupervised training data, when the fuzzy fault diagnosis results on multiple key state knowledge dimensions are determined to match the fault state knowledge graph generated by the diagnosis node, an optimized result is generated according to the fault state knowledge graph of each diagnosis node in the unsupervised training data. For example, for a diagnosis node in the train power system, if the fuzzy fault diagnosis results on several key state knowledge dimensions such as motor speed, motor temperature, and current stability all indicate that there is a fault risk in the power system (such as excessive motor speed fluctuations, rising motor temperature, and unstable current existing simultaneously), then the server generates an optimized second fault state path diagnosis result according to the complete fault state knowledge graph of this diagnosis node in the unsupervised training data, comprehensively considering the mutual relationship between each key state knowledge dimension. This result may detail the complete fault development path from the initial slight abnormality of the motor (such as a small fluctuation in motor speed detected by a certain sensor), to the gradual increase in motor temperature over time, which in turn affects the current stability, and ultimately leads to the overall performance degradation of the power system. Through such a process, the server optimizes the second fault state path diagnosis result using the shared fault diagnosis network, improving the accuracy and comprehensiveness of fault diagnosis.

[0048] For example, in a possible implementation manner, step S131 includes:

[0049] Step S1311, analyze the feature structure of the unsupervised training data and identify a list of key state knowledge dimensions related to the fault state path.

[0050] Step S1312, for each key state knowledge dimension in the list of key state knowledge dimensions, create a corresponding state knowledge graph framework. The state knowledge graph framework defines the basic structure of the state knowledge graph. The basic structure includes the types of nodes and edges. The nodes will represent different states or sub-states, and the edges represent the transition relationships or mutual influence relationships between states.

[0051] Step S1313, map each data point in the unsupervised training data into the corresponding state knowledge graph framework according to the key state knowledge dimensions. For each data point, determine the position of the data point in the state knowledge graph according to the value of the data point on the key state knowledge dimension, and generate a preliminarily mapped state knowledge graph.

[0052] Step S1314: According to the operation rules and fault modes of the rail transit system, define and label the nodes and edges in the initially mapped state knowledge graph. Specifically, for each node in the initially mapped state knowledge graph, label its normal operation range, critical state range, and fault state range. For the edges in the initially mapped state knowledge graph, label the conditional probability or triggering mechanism during different state transitions, and based on the interaction between multiple key state knowledge dimensions, add cross-dimensional connections and interaction relationships in the initially mapped state knowledge graph to generate target state knowledge graph data.

[0053] Step S1315: Extract the state knowledge graph data corresponding to each key state knowledge dimension from the target state knowledge graph data.

[0054] For example, in a possible implementation manner, Step S1314 includes:

[0055] Step S1314-1: Analyze each key state knowledge dimension separately to determine the role and function of this key state knowledge dimension in the operation of the rail transit system. Based on the role and function of this key state knowledge dimension in the operation of the rail transit system and combined with the logical connections existing between different key state knowledge dimensions, determine the set of interaction relationships existing between key state knowledge dimensions. The interaction relationships include causal relationships, collaborative relationships, and restrictive relationships.

[0056] Step S1314-2: Define the cross-dimensional connection types according to the set of interaction relationships. Specifically, for causal relationships, define a connection type representing the causal flow direction to determine the direction identifier from the cause dimension to the result dimension. For collaborative relationships, define a connection type representing the collaborative effect. For restrictive relationships, define a connection type representing the restrictive direction and degree, thereby generating the defined cross-dimensional connection types.

[0057] Step S1314-3: In the initially mapped state knowledge graph, for each key state knowledge dimension, find the target nodes that establish cross-dimensional connections with other key state knowledge dimensions, and determine these target nodes as cross-dimensional connection nodes, thereby generating a set of cross-dimensional connection nodes. Each target node in the set of cross-dimensional connection nodes will serve as the starting point or ending point for establishing cross-dimensional connections.

[0058] Step S1314-4: Based on the defined cross-dimensional connection types and the determined set of cross-dimensional connection nodes, construct cross-dimensional connection mapping rules. Specifically, for each cross-dimensional connection type, define the mapping method between the cross-dimensional connection nodes of different key state knowledge dimensions for this cross-dimensional connection type.

[0059] Step S1314-5: Add cross-dimensional connections to the initially mapped state knowledge graph according to the cross-dimensional connection mapping rules. Specifically, for each cross-dimensional connection mapping rule, starting from the corresponding cross-dimensional connection node, add connection edges between the state knowledge graphs of different key state knowledge dimensions according to the cross-dimensional connection mapping rules. The connection edges represent cross-dimensional relationships and are marked with connection types according to the rules.

[0060] Step S1314-6: Perform attribute annotation on the cross-dimensional connections in the state knowledge graph after initially adding cross-dimensional connections according to the set of interaction relationships between key state knowledge dimensions. Among them, for cross-dimensional connections of the causal relationship type, annotate the intensity of the causal relationship; for cross-dimensional connections of the collaborative relationship type, annotate the collaborative effect or degree; for cross-dimensional connections of the restrictive relationship type, annotate the specific restrictive factors and the degree of restriction.

[0061] Step S1314-7: Conduct logical verification on the cross-dimensional connection state knowledge graph after attribute annotation, and use the cross-dimensional connection state knowledge graph that has passed the logical verification as the final state knowledge graph to generate target state knowledge graph data. The target state knowledge graph data includes all node information, edge information, and relevant information of cross-dimensional connections in the state knowledge graph.

[0062] In this embodiment, first, an in-depth analysis is performed on the feature structure of unsupervised training data. This unsupervised training data contains operation information of various aspects of the suspended rail transit system, such as the power system, braking system, suspension device, and track condition monitoring data of the train. The server identifies a list of key state knowledge dimensions related to the fault state path through the structural analysis of this massive data. Taking the power system of the train as an example, the key state knowledge dimensions may include the input current, output torque, and rotational speed stability of the motor; for the braking system, the key state knowledge dimensions may be braking pressure, brake pad wear degree, and braking response time; in terms of the track condition, the track flatness, the gap size at the track connection, etc. may be identified as key state knowledge dimensions.

[0063] For each key state knowledge dimension in the identified list of key state knowledge dimensions, the server begins to create a corresponding state knowledge graph framework. Take the key state knowledge dimension of the input current of the motor as an example. The state knowledge graph framework defines its basic structure. In this state knowledge graph framework, nodes represent different states or sub-states. For example, nodes can include different intervals within the normal input current range, such as low current stable intervals, normal operating current intervals, high current critical intervals, etc., and also include current intervals under specific fault states, such as overcurrent fault intervals. Edges represent the conversion relationships or mutual influence relationships between states. For example, the edge from the normal operating current interval to the high current critical interval may represent the relationship that the motor load suddenly increases, resulting in an increase in current. The edge from the high current critical interval to the overcurrent fault interval may represent the conversion relationship that factors such as a malfunction in the heat dissipation system or a short circuit in the circuit cause the current to get out of control. For the key state knowledge dimension of the braking pressure in the braking system, the nodes in the state knowledge graph framework may include subdivided intervals within the normal braking pressure range, intervals with insufficient braking pressure, and intervals with abnormally high braking pressure, etc. The edges represent conversion relationships such as the braking pressure dropping from the normal interval to the insufficient braking pressure interval due to a leak in the hydraulic pipeline of the braking system.

[0064] Then, the server maps each data point in the unsupervised training data to the corresponding state knowledge graph framework according to the key state knowledge dimension. For the key state knowledge dimension of the motor input current, when the server processes each data point, it determines the position of the data point in the state knowledge graph based on the value of the data point in the key state knowledge dimension of the motor input current. For example, if the motor input current value corresponding to a data point is within a specific value in the normal operating current interval, then this data point is located at the position corresponding to the normal operating current interval in the state knowledge graph, thus generating a preliminarily mapped state knowledge graph. The same is true for the key state knowledge dimension of the braking pressure. If the braking pressure value corresponding to a data point is within a value in the interval with insufficient braking pressure, the data point is mapped to the corresponding position in the interval with insufficient braking pressure in the state knowledge graph.

[0065] Next, according to the operation rules and fault modes of the rail transit system, the server defines and annotates the nodes and edges in the preliminarily mapped state knowledge graph. For each node in the motor input current state knowledge graph, the server annotates the normal operating range of the node, such as the upper and lower limit values of the normal operating current range; the critical state range, like the upper and lower limit values of the high current critical range; the fault state range, which is the specific numerical range of the overcurrent fault range. For the edges in the state knowledge graph, such as the edge from the normal operating current range to the high current critical range, the server annotates the conditional probability or triggering mechanism during different state transitions. For example, when the motor load increases beyond a certain value, there is an 80% probability that the current will rise from the normal operating current range to the high current critical range, and the value of the motor load increase here is part of the triggering mechanism. For the nodes in the brake pressure state knowledge graph, the normal operating range may be the normal pressure fluctuation range of the brake determined according to the train design standard, the critical state range is the range where the brake pressure starts to show abnormalities but has not completely failed, and the fault state range is the range where the brake pressure is severely insufficient or too high, resulting in the failure of the braking system. For the edges, such as the edge from the normal brake pressure range to the insufficient brake pressure range, the annotated triggering mechanism may be the probability of pressure drop caused by a certain component in the braking system (such as the brake pads wearing to a certain extent or a small leak in the hydraulic pipeline).

[0066] Moreover, based on the interactions between multiple key state knowledge dimensions, the server adds cross-dimensional connections and interaction relationships in the preliminarily mapped state knowledge graph to generate the target state knowledge graph data. The server analyzes each key state knowledge dimension separately to determine its role and function in the operation of the rail transit system. For example, for the key state knowledge dimension of motor input current, its role in the operation of the rail transit system is an electrical input index that provides power for the motor, and its function is to directly affect the output torque and speed of the motor, and thus affect the running speed of the train. For the key state knowledge dimension of brake pressure, its role is to provide an index of the magnitude of the braking force, and its function is to ensure that the train can stop safely when needed. Combining the logical connections existing between different key state knowledge dimensions, a set of interaction relationships between key state knowledge dimensions is determined. Taking motor input current and brake pressure as an example, when the motor input current suddenly drops, it may cause the train speed to decrease, thereby reducing the braking pressure requirement of the braking system, which is a causal relationship; there is a restrictive relationship between the motor output torque and the brake pressure, because too much brake pressure may cause excessive impact on the train's suspension system and the track when the train is running at high speed and the motor output torque is large.

[0067] Based on the determined set of interaction relationships, the server defines cross-dimensional connection types. For the causal relationship between the motor input current and the braking pressure, the server defines a connection type representing the causal flow direction, using an arrow to indicate the direction from the cause dimension of the motor input current to the result dimension of the braking pressure, and the direction of the arrow identifies the direction of the causal relationship. For the constraint relationship between the motor output torque and the braking pressure, a connection type representing the constraint direction and degree is defined. For example, a connection line with a numerical value and a symbol is used to represent the constraint relationship. The numerical value represents the degree of constraint (such as the braking pressure cannot exceed a certain multiple of the motor output torque), and the symbol represents the constraint direction (the braking pressure is constrained by the motor output torque).

[0068] In the initially mapped state knowledge graph, for each key state knowledge dimension, the server searches for target nodes that establish cross-dimensional connections with other key state knowledge dimensions, and determines these target nodes as cross-dimensional connection nodes, thereby generating a set of cross-dimensional connection nodes. For example, in the motor input current state knowledge graph, the current value node that has a causal relationship with the braking pressure (such as when the motor input current drops to a certain value, it will affect the braking pressure) is a cross-dimensional connection node; in the braking pressure state knowledge graph, the braking pressure value node that has a causal relationship with the motor input current (such as the value to which the braking pressure drops due to the decrease in the motor input current) is also a cross-dimensional connection node.

[0069] Based on the defined cross-dimensional connection types and the determined set of cross-dimensional connection nodes, the server constructs cross-dimensional connection mapping rules. For the cross-dimensional connection of the causal relationship type, the mapping method from the motor input current to the braking pressure is defined. For example, when the motor input current drops to a specific numerical value, the numerical range of the possible decrease in the braking pressure is calculated according to a certain functional relationship, and then a connection edge is added between the state knowledge graphs. For the cross-dimensional connection of the constraint relationship type, according to the constraint relationship between the motor output torque and the braking pressure, the allowable range of the braking pressure is specified under different motor output torque numerical values, and a connection edge is added between the state knowledge graphs according to this rule.

[0070] According to the cross-dimensional connection mapping rules, the server adds cross-dimensional connections to the initially mapped state knowledge graph. For each cross-dimensional connection mapping rule, starting from the corresponding cross-dimensional connection node, a connection edge is added between the state knowledge graphs of different key state knowledge dimensions according to the cross-dimensional connection mapping rule. For example, starting from the cross-dimensional connection node in the motor input current state knowledge graph, according to the causal relationship mapping rule between the motor input current and the braking pressure, a connection edge is added between the motor input current state knowledge graph and the braking pressure state knowledge graph. This connection edge represents the cross-dimensional relationship and is marked with the connection type (the arrow identifier of the causal relationship) according to the rule.

[0071] Based on the set of interaction relationships between the key state knowledge dimensions, the server performs attribute annotation on the cross-dimensional connections in the state knowledge graph after initially adding cross-dimensional connections. For the cross-dimensional connection of the causal relationship type between the motor input current and the braking pressure, the strength of the causal relationship is annotated. For example, if the braking pressure will necessarily decrease when the motor input current drops by a certain value, then the strength of the causal relationship is strong; if the braking pressure has a certain probability of decreasing when the motor input current drops, then the strength of the causal relationship is weak. For the cross-dimensional connection of the restriction relationship type between the motor output torque and the braking pressure, the specific limiting factors and the degree of restriction are annotated. For example, it is annotated that when the motor output torque is a certain value, the braking pressure cannot exceed a certain specific value, which are the specific limiting factors and the degree of restriction.

[0072] Finally, the server performs logical verification on the cross-dimensional connection state knowledge graph after annotating the attributes, checking whether the cross-dimensional connections conform to the actual operation logic and fault logic of the rail transit system. For example, verifying whether the logic that the decrease in the motor input current leads to the decrease in the braking pressure is reasonable, and whether there are other factors that may affect this relationship. The cross-dimensional connection state knowledge graph after logical verification is used as the final state knowledge graph to generate target state knowledge graph data, which contains all node information, edge information, and relevant information of the cross-dimensional connections in the state knowledge graph, providing a comprehensive and accurate basis for subsequent fault diagnosis.

[0073] In a possible implementation manner, the fault diagnosis neural network includes a teaching deep learning network and a taught deep learning network. The first fault state path diagnosis result is the fault state path diagnosis result output by the taught deep learning network.

[0074] Step S120 includes: determining the first fault diagnosis error of the taught deep learning network according to the first fault state path diagnosis result and the fault state path annotation data.

[0075] In a possible implementation manner, the fault diagnosis neural network includes a teaching deep learning network and a taught deep learning network. The second fault state path diagnosis result includes the taught fault state path diagnosis result output by the teaching deep learning network and the taught fault state path diagnosis result output by the taught deep learning network, and the optimized second fault state path diagnosis result includes the optimized taught fault state path diagnosis result.

[0076] Step S130 includes: determining the second fault diagnosis error of the taught deep learning network according to the taught fault state path diagnosis result and the optimized taught fault state path diagnosis result.

[0077] In a possible implementation manner, determining the second fault diagnosis error of the taught deep learning network according to the taught fault status path diagnosis result and the optimized taught fault status path diagnosis result includes:

[0078] Calculating a fuzzy fault diagnosis error according to the taught fault status path diagnosis result and the optimized taught fault status path diagnosis result, and calculating a transfer learning fault diagnosis error according to the taught fault status path diagnosis result and the taught fault status path diagnosis result.

[0079] Determining a first influence factor and a second influence factor according to the current iteration round and the maximum iteration round of the fault diagnosis learning process.

[0080] Calculating a first weighted result of the first influence factor and the fuzzy fault diagnosis error, and a second weighted result of the second influence factor and the transfer learning fault diagnosis error.

[0081] Outputting the sum of the first weighted result and the second weighted result as the second fault diagnosis error of the taught deep learning network.

[0082] In this embodiment, the fault diagnosis neural network includes a taught deep learning network and a taught deep learning network. When processing supervised training data, the taught deep learning network outputs a first fault status path diagnosis result. The server determines the first fault diagnosis error of the taught deep learning network according to the first fault status path diagnosis result and the fault status path annotation data.

[0083] Taking the fault diagnosis of the train's braking system as an example, it is assumed that the supervised training data obtained by the server is the operation data related to the braking operation of the train over a period of time, including braking pressure, braking distance, braking response time, etc., and all these data are accompanied by accurate fault status path annotation data, such as the data annotation of each stage from normal wear of the brake pads to abnormal wear and then to a significant decrease in braking efficiency. After processing these supervised training data, the taught deep learning network outputs a first fault status path diagnosis result of the braking system. For example, based on factors such as the change trend of the input braking pressure and the fluctuation of the braking distance, the taught deep learning network determines that the braking system may be in the stage of moderate wear of the brake pads, and gives the fault status path diagnosis result of the change of each relevant operation parameter from the start of wear of the brake pads to the current moderate wear state.

[0084] The server compares this first fault state path diagnosis result with the pre-stored fault state path annotation data. For example, in the annotation data, the brake pads should have been in a severely worn state at a certain specific moment, while the taught deep learning network diagnoses them as being in a moderately worn state. The server determines the difference between the two through a specific algorithm, thereby determining the first fault diagnosis error of the taught deep learning network. This algorithm may involve comparing each key node on the fault state path (such as the key operating parameter values corresponding to different wear degrees of the brake pads), calculating the deviation between the diagnosis result and the annotation result at each node, and then determining the first fault diagnosis error based on the comprehensive situation of these deviations. The calculation of this error is very meticulous and needs to consider the weights of each relevant operating parameter in different fault stages to accurately reflect the deviation degree of the taught deep learning network from the actual situation when performing fault diagnosis based on supervised training data.

[0085] When dealing with unsupervised training data, the taught deep learning network outputs a taught fault state path diagnosis result, and the taught deep learning network outputs a taught fault state path diagnosis result. The two together constitute the second fault state path diagnosis result. Subsequently, the shared fault diagnosis network is used to optimize the second fault state path diagnosis result to obtain an optimized taught fault state path diagnosis result. The optimized second fault state path diagnosis result includes the optimized taught fault state path diagnosis result.

[0086] Continuing with the example of the train braking system, assume that the unsupervised training data includes braking-related data of the train in complex environments (such as different weather conditions and different track conditions). The taught deep learning network analyzes this unsupervised training data based on its existing knowledge and pattern recognition capabilities and outputs a taught fault state path diagnosis result. For example, the taught deep learning network may judge that there is a trend of reduced braking efficiency due to the brake pads being affected by moisture and rusting based on the impact of humid weather on the brake pads, and give the corresponding fault state path diagnosis result. The taught deep learning network will also output a taught fault state path diagnosis result based on the input unsupervised training data. For example, the taught deep learning network judges that there may be certain instability factors in the braking system based on the fluctuations of the braking pressure in complex environments, but the judgment of the specific fault causes and development paths may not be as accurate as that of the taught deep learning network.

[0087] The server determines the second fault diagnosis error of the taught deep learning network based on the diagnosis results of the taught fault status path and the optimized diagnosis results of the taught fault status path. First, the server calculates the fuzzy fault diagnosis error based on the diagnosis results of the taught fault status path and the optimized diagnosis results of the taught fault status path. For example, when calculating the fuzzy fault diagnosis error, the server compares the differences in the judgments of the occurrence probabilities of braking system faults in the diagnosis results of the taught fault status path and the optimized diagnosis results of the taught fault status path. If the diagnosis result of the taught fault status path believes that the probability of a decrease in braking efficiency is 30%, while the optimized diagnosis result of the taught fault status path believes that this probability is 50%, the server will calculate the fuzzy fault diagnosis error based on the difference between the two and other relevant factors on the fault status path (such as the weights of different environmental factors on braking efficiency).

[0088] Meanwhile, the server calculates the transfer learning fault diagnosis error based on the diagnosis results of the taught fault status path and the diagnosis results of the teaching fault status path. For example, the taught deep learning network may have accumulated knowledge about fault diagnosis of other similar rail transit systems in similar environments during previous learning processes. When diagnosing the current unsupervised training data, this knowledge may affect the diagnosis results of the teaching fault status path. The taught deep learning network will attempt to transfer this knowledge from the taught deep learning network during the learning process, and the transfer learning fault diagnosis error is a measure of the difference between the diagnosis results of the taught fault status path and the diagnosis results of the teaching fault status path during the transfer learning process. Suppose the taught deep learning network determines through transfer learning that the possibility of the braking system getting damp and rusty is relatively high, while the taught deep learning network has a lower judgment of this possibility due to incomplete or inaccurate transfer learning. The server will calculate the transfer learning fault diagnosis error based on the difference between the two and relevant factors.

[0089] Then, the server determines the first influence factor and the second influence factor based on the current iteration number and the maximum iteration number of the fault diagnosis learning process. Suppose the maximum iteration number of the fault diagnosis learning process is 100 times, and the current iteration number is 30 times. According to a specific calculation rule (this rule may be based on the optimization strategy of the learning process, for example, paying more attention to the fuzzy fault diagnosis error in the early stage of learning and more attention to the transfer learning fault diagnosis error in the later stage), the first influence factor is calculated to be 0.4, and the second influence factor is calculated to be 0.6.

[0090] Next, the server calculates the first weighted result of the first influence factor and the fuzzy fault diagnosis error, and the second weighted result of the second influence factor and the transfer learning fault diagnosis error. For example, if the fuzzy fault diagnosis error is calculated as 0.2, then the first weighted result is 0.4 * 0.2 = 0.08; if the transfer learning fault diagnosis error is calculated as 0.3, then the second weighted result is 0.6 * 0.3 = 0.18.

[0091] Finally, the server outputs the sum of the first weighted result and the second weighted result, that is, 0.08 + 0.18 = 0.26, as the second fault diagnosis error of the taught deep learning network. This second fault diagnosis error will serve as an important basis for adjusting the parameters of the taught deep learning network to improve its accuracy in fault diagnosis under unsupervised training data, thereby enhancing the performance of the entire fault diagnosis neural network. Throughout the process, the accurate calculation of the fault diagnosis error is ensured, providing a reliable basis for subsequent neural network optimization.

[0092] In a possible implementation manner, the fault diagnosis learning process includes a learning process of multiple iterative rounds, and the fault diagnosis neural network includes a taught deep learning network and a taught deep learning network.

[0093] Step S140 includes:

[0094] Step S141, according to the first fault diagnosis error and the second fault diagnosis error, optimize the neuron weight parameters of the taught deep learning network in the learning process of the a-th iterative round, where a is a positive integer.

[0095] Step S142, according to the neuron weight parameters of the taught deep learning network before the learning process of the a-th iterative round and the neuron weight parameters of the learning process of the a-th iterative round, optimize the neuron weight parameters of the teaching deep learning network.

[0096] In this embodiment, when it comes to optimizing the fault diagnosis neural network according to the first fault diagnosis error and the second fault diagnosis error, the server will operate according to a specific process. First, the server optimizes the neuron weight parameters of the taught deep learning network in the learning process of the a-th iterative round according to the first fault diagnosis error and the second fault diagnosis error (where a is a positive integer).

[0097] Taking the fault diagnosis of the suspension system of a train as an example, it is assumed that the ath iteration round is the fifth iteration round. In the previous operation, the taught deep learning network processed the supervised training data and unsupervised training data containing the relevant data of the train suspension system (such as the expansion and contraction frequency of the shock absorber, the elastic coefficient of the suspension spring, the shaking amplitude of the train during driving, etc.), and obtained the corresponding first fault state path diagnosis results and second fault state path diagnosis results, and then calculated the first fault diagnosis error and the second fault diagnosis error.

[0098] The server will adjust the neuron weight parameters of the taught deep learning network in the fifth iteration of learning according to these errors. For example, if it is found in the first fault diagnosis error that the taught deep learning network has a large deviation in its judgment on the relationship between the change in the elastic coefficient of the suspension spring and the train sway amplitude, and similar problems are also reflected in the second fault diagnosis error, this indicates that the neuron connection weights related to this part of the relationship in the taught deep learning network may be unreasonable. The server will use a suitable optimization algorithm, such as a gradient descent-based algorithm, to adjust the weight parameters of these neurons according to the size and direction of the error. Specifically, if the error indicates that the taught deep learning network is too optimistic about the increase in the elastic coefficient of the suspension spring leading to a decrease in the train sway amplitude (that is, the actual reduction in the sway amplitude is not as large as the network judgment), then the server will appropriately reduce the weights of the neuron connections related to the elastic coefficient of the suspension spring and the train sway amplitude, so that the taught deep learning network can more accurately judge this relationship in subsequent diagnosis. This adjustment is very delicate and needs to take into account the position and role of each neuron in the entire network structure, as well as the relationship between different neurons, to ensure that the adjusted network can better fit the actual fault state path.

[0099] After optimizing the neuron weight parameters of the taught deep learning network in the a-th iteration learning process, the server will optimize the neuron weight parameters of the teaching deep learning network based on the neuron weight parameters of the taught deep learning network before the a-th iteration learning process and the neuron weight parameters of the a-th iteration learning process.

[0100] Still taking the fault diagnosis of the train suspension system as an example, it is taught that the teaching deep learning network plays a guiding and knowledge transfer role in the previous learning process of the taught deep learning network. Before the 5th iteration round, the teaching deep learning network has a set of neuron weight parameters, which determine its processing method for the data related to the train suspension system and the judgment logic for the fault state path. After the taught deep learning network completes the adjustment of the neuron weight parameters in the 5th iteration round, the server will analyze the change situation of the weight parameters of the taught deep learning network. If the taught deep learning network makes a large adjustment to the weights of some key factors (such as the relationship between the shock absorber telescopic frequency and the train running stability) in the 5th iteration round, and this adjustment is proven to be effective in subsequent diagnoses (for example, the adjusted taught deep learning network can more accurately diagnose the fault state path when processing new suspension system data), then the server will feedback this adjustment information to the teaching deep learning network.

[0101] The server will adopt a specific strategy to optimize the neuron weight parameters of the teaching deep learning network according to the change situation of the weight parameters of the taught deep learning network. For example, if the taught deep learning network increases the connection weight between the neurons related to the shock absorber telescopic frequency and the neurons related to the train running stability and achieves a better diagnosis effect, the server may increase the connection weight between the neurons related to these two factors in the teaching deep learning network according to a certain proportion. At the same time, the server also needs to consider the structure and characteristics of the teaching deep learning network itself to avoid over-adjustment resulting in the teaching deep learning network losing the effective knowledge accumulated before. For example, the teaching deep learning network may also contain other knowledge structures related to the suspension system (such as the relationship between the suspension system and the track flatness). When adjusting the neuron weight parameters related to the shock absorber telescopic frequency and the train running stability, the existing knowledge structures and relationships cannot be damaged.

[0102] Through such an optimization process, the parameters of the fault diagnosis neural network are continuously adjusted in the learning process of multiple iteration rounds, so that the teaching deep learning network and the taught deep learning network can work better together, improving the accuracy and efficiency of the fault diagnosis of the suspended rail transit system. In the whole process, it is necessary to accurately analyze the error information, reasonably adjust the neuron weight parameters of the taught deep learning network, and feedback these adjustments to the teaching deep learning network to realize the gradual optimization of the whole fault diagnosis neural network, ensuring that the fault state path can be accurately diagnosed in the face of various complex operation data of the rail transit system.

[0103] In a possible implementation manner, the fault diagnosis learning process includes a learning process of multiple iteration rounds, and the method further includes:

[0104] Step A110: Select candidate training data from the unsupervised training data according to the second fault state path diagnosis result of the unsupervised training data in the learning process of the a-th iteration round, where a is a positive integer.

[0105] Step A120: Obtain the fault state path annotation data for adding training supervision data to the candidate training data, generate the active supervision training data in the learning process of the a-th iteration round, and load the active supervision training data into the supervision training data sequence called in the learning process of the (a + 1)-th iteration round.

[0106] Among them, Step A110 includes:

[0107] According to the second fault state path diagnosis result of the unsupervised training data in the learning process of the a-th iteration round, label the unsupervised training data that meets the target requirements as the candidate training data.

[0108] Among them, the unsupervised training data that meets the target requirements includes at least one of the following data:

[0109] A. The first X unsupervised training data arranged in ascending order according to the diagnosis probability values based on the second fault state path diagnosis result.

[0110] B. The first Y unsupervised training data arranged in descending order according to the average diagnosis likelihood ratio based on the second fault state path diagnosis result, where the average diagnosis likelihood ratio is the average diagnosis likelihood ratio of all diagnosis nodes in the second fault state path diagnosis result.

[0111] C. The first Z unsupervised training data arranged in descending order according to the likelihood ratio proportion based on the second fault state path diagnosis result, where the likelihood ratio proportion is the proportion of diagnosis nodes greater than the preset threshold in all diagnosis nodes in the second fault state path diagnosis result. Here, X, Y, and Z are all positive integers.

[0112] In this embodiment, in the learning process of the a-th iteration round (where a is a positive integer), the server will select candidate training data from the unsupervised training data according to the second fault state path diagnosis result of the unsupervised training data. Taking the fault diagnosis of the train's power system as an example, the unsupervised training data includes various parameters of the train's power system under different operating conditions, such as the current, voltage, and speed of the motor, as well as the oil temperature and pressure of the transmission system. In the a-th iteration round, the fault diagnosis neural network (including the teaching deep learning network and the taught deep learning network) processes these unsupervised training data to obtain the second fault state path diagnosis result.

[0113] The server selects unsupervised training data that meets the target requirements as candidate training data based on the diagnostic results of this second fault status path. For the top X unsupervised training data (assuming X = 10) with the diagnostic probability values arranged in ascending order based on the diagnostic results of the second fault status path, the server will analyze the diagnostic probability values of each data point in detail. For example, in the diagnosis of a train power system, a certain unsupervised training data point corresponds to the operating parameters of the motor when the train is running at high speed and with a large load, and its diagnostic probability value is relatively low, indicating that under the current fault diagnosis neural network, the possibility of this data point being determined to have a fault is relatively small. However, the server still pays attention to such data points because they may represent some special situations or situations that the current diagnosis model has not fully understood. After arranging them in ascending order of the diagnostic probability values, the server selects the top 10 such unsupervised training data as part of the candidate training data.

[0114] For the top Y unsupervised training data (assuming Y = 5) with the mean diagnostic likelihood ratios arranged in descending order based on the diagnostic results of the second fault status path, the server will first calculate the mean diagnostic likelihood ratio of each unsupervised training data. The mean diagnostic likelihood ratio is the mean of the diagnostic likelihood ratios of all diagnostic nodes in the diagnostic results of the second fault status path. Taking the train power system as an example, the diagnostic nodes corresponding to an unsupervised training data may include multiple aspects such as motor faults and transmission system faults. For a certain unsupervised training data, the server calculates the mean based on the diagnostic likelihood ratios of each diagnostic node. For example, for a data point, the diagnostic likelihood ratio of the motor fault is relatively high, and the diagnostic likelihood ratio of the transmission system fault is relatively low, and a mean is calculated comprehensively. Then, they are arranged in descending order according to this mean, and the top 5 such unsupervised training data are selected as another part of the candidate training data. These data points have relatively high overall mean diagnostic likelihood ratios, perhaps because they have relatively high credibility or representativeness in some key fault diagnosis aspects and are worthy of further in-depth study and use for subsequent training.

[0115] For the top Z unsupervised training data (assuming Z = 8) sorted in descending order according to the likelihood ratio proportion based on the diagnostic results of the second fault state path, the likelihood ratio proportion is the proportion of diagnostic nodes greater than the preset threshold in all diagnostic nodes in the diagnostic results of the second fault state path. In the diagnosis of the train power system, the preset threshold may be a boundary for judging the possibility of a fault set according to a large amount of historical data and practical experience. For each unsupervised training data, the server calculates the proportion of the number of diagnostic nodes greater than this preset threshold in the total number of diagnostic nodes. For example, in an unsupervised training data, the diagnostic likelihood ratios of multiple diagnostic nodes (such as motor overheating, abnormal transmission system pressure, etc.) exceed the preset threshold, and the calculated likelihood ratio proportion is relatively high. After the server sorts the likelihood ratio proportions in descending order, it selects the top 8 such unsupervised training data as another part of the candidate training data. These data points are outstanding in the proportion of key diagnostic nodes that meet the possibility of fault diagnosis and may contain very valuable information for improving the fault diagnosis model.

[0116] After the server selects the candidate training data, it will obtain the fault state path annotation data for adding training supervision data to these candidate training data. Continuing with the example of the train power system, for those unsupervised training data selected previously, the server will search for the corresponding fault state path annotation data. These annotation data may come from the records of actual fault cases or the results obtained through in-depth analysis of historical data by experts. For example, for a certain candidate training data point (the operation data of the train power system under specific working conditions), the corresponding fault state path annotation data may detail the complete fault development path from the initial slight abnormality of the motor (such as a slight fluctuation in current), to the motor overheating over time, to the possible abnormal transmission system pressure, and finally resulting in a decrease in the train power.

[0117] The server generates active supervised training data during the learning process of the a-th iteration round using these candidate training data with labeled data. Then, these active supervised training data are loaded into the sequence of supervised training data called during the learning process of the (a + 1)-th iteration round. The purpose of doing this is to enable the fault diagnosis neural network to better learn these special data situations in subsequent iteration rounds, further optimize the parameters of the model, and improve the accuracy of fault diagnosis. In the (a + 1)-th iteration round, when the fault diagnosis neural network learns again, these newly added active supervised training data will be used together with the original supervised training data to adjust the parameters of the neural network. For example, the teaching deep learning network will re-adjust the judgment logic for different fault modes of the train power system based on these new data. The taught deep learning network will also, under the guidance of the teaching deep learning network, more precisely learn the fault features in these data, thereby enhancing the ability of the entire fault diagnosis neural network to diagnose faults in the train power system.

[0118] Throughout the process, the server selects candidate training data, obtains accurate labeled data to generate active supervised training data, and reasonably integrates it into the learning process of subsequent iteration rounds, continuously optimizing the fault diagnosis ability of the fault diagnosis neural network for the suspended rail transit system to adapt to various complex operating conditions and fault modes.

[0119] Figure 2 The hardware structure diagram of the AI fault diagnosis system 100 for the suspended rail transit system provided by the embodiment of the present invention to implement the above-mentioned AI fault diagnosis method for the suspended rail transit system is shown, as Figure 2 shown, the AI fault diagnosis system 100 for the suspended rail transit system may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0120] The machine-readable storage medium 120 can store data and / or instructions. In some embodiments, the machine-readable storage medium 120 can store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 can store the data and / or instructions used by the AI fault diagnosis system 100 for the suspended rail transit system to execute or use to complete the exemplary methods described in the present invention.

[0121] In the specific implementation process, one or more processors 110 execute the computer-executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the AI fault diagnosis method for the suspended rail transit system as described in the above method embodiments. The processor 110, the machine-readable storage medium 120, and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the transceiver actions of the communication unit 140.

[0122] For the specific implementation process of the processor 110, reference can be made to the various method embodiments executed by the AI fault diagnosis system 100 for the suspended rail transit system described above. Their implementation principles and technical effects are similar, and thus will not be elaborated herein in this embodiment.

[0123] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the AI fault diagnosis method for the suspended rail transit system as described above is implemented.

[0124] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are incorporated into one embodiment, drawing, or description thereof.

Claims

1. An AI fault diagnosis method for a suspended rail transit system, characterized in that: The method comprises: Obtain supervised training data and unsupervised training data for the suspended rail transit system, wherein the supervised training data is sample rail transit system operation data carrying fault state path annotation data; Loading the supervised training data into a fault diagnosis neural network, generating a first fault state path diagnosis result of the supervised training data, and determining a first fault diagnosis error based on the first fault state path diagnosis result and the fault state path annotation data; The unsupervised training data is loaded into the fault diagnosis neural network to generate a second fault state path diagnosis result of the unsupervised training data, the second fault state path diagnosis result is optimized by using a shared fault diagnosis network to generate an optimized second fault state path diagnosis result, the shared fault diagnosis network has a cross-domain fault diagnosis performance across fault types, and a second fault diagnosis error is determined based on the optimized second fault state path diagnosis result; According to the first fault diagnosis error and the second fault diagnosis error, optimizing the fault diagnosis neural network, and performing fault diagnosis on any input target rail transit system operation data based on the optimized fault diagnosis neural network to generate a target fault state path diagnosis result corresponding to the target rail transit system operation data; The step of optimizing the second fault state path diagnosis result by using the shared fault diagnosis network to generate the optimized second fault state path diagnosis result includes: Using the unsupervised training data as the operation data to be diagnosed, and using the second fault state path diagnosis result as the training supervision data, and using the shared fault diagnosis network to generate the optimized second fault state path diagnosis result; The fault diagnosis neural network includes a teaching deep learning network and a taught deep learning network; the second fault state path diagnosis result includes a teaching fault state path diagnosis result output by the teaching deep learning network and a taught fault state path diagnosis result output by the taught deep learning network, and the optimized second fault state path diagnosis result includes an optimized teaching fault state path diagnosis result; Determining a second fault diagnosis error according to the optimized second fault state path diagnosis result includes: The second fault diagnosis error of the taught deep learning network is determined according to the taught fault state path diagnosis result and the optimized teaching fault state path diagnosis result.

2. The AI ​​fault diagnosis method for a suspended rail transit system according to claim 1, characterized in that: The unsupervised training data is multi-dimensional rail transit operation data, and the unsupervised training data is used as the operation data to be diagnosed, and the second fault state path diagnosis result is used as the training supervision data, and the shared fault diagnosis network is used to generate the optimized second fault state path diagnosis result, including: Extracting a plurality of state knowledge graph data of a plurality of key state knowledge dimensions from the unsupervised training data, wherein the plurality of key state knowledge dimensions correspond to a plurality of feature mapping spaces of the multi-dimensional rail transit operation data; Using the multiple state knowledge graph data of the multiple key state knowledge dimensions as the to-be-diagnosed operating data, and using the second fault state path diagnosis result as the training supervision data, and using the shared fault diagnosis network to output fuzzy fault diagnosis results on the multiple key state knowledge dimensions, wherein the fuzzy fault diagnosis result on each key state knowledge dimension includes a fault state path diagnosis result matching the multiple state knowledge graph data of the same key state knowledge dimension; The fuzzy fault diagnosis results on the multiple key state knowledge dimensions are fused according to the diagnosis fusion mechanism to generate the optimized second fault state path diagnosis result.

3. The AI ​​fault diagnosis method for a suspended rail transit system according to claim 2, characterized in that: The step of fusing the fuzzy fault diagnosis results on the multiple key state knowledge dimensions according to the diagnosis fusion mechanism to generate the optimized second fault state path diagnosis result includes: For each diagnostic node in the unsupervised training data, when the fuzzy fault diagnosis results on the multiple key state knowledge dimensions are determined to be the fault state knowledge graph generated by matching the diagnostic node, the optimized second fault state path diagnosis result is generated based on the fault state knowledge graphs of each diagnostic node in the unsupervised training data.

4. The AI ​​fault diagnosis method for a suspended rail transit system according to any one of claims 1 to 3, characterized in that: The fault diagnosis neural network includes a teaching deep learning network and a taught deep learning network; the first fault state path diagnosis result is a fault state path diagnosis result output by the taught deep learning network; The determining a first fault diagnosis error according to the first fault state path diagnosis result and the fault state path marking data includes: The first fault diagnosis error of the taught deep learning network is determined according to the first fault state path diagnosis result and the fault state path labeling data.

5. The AI ​​fault diagnosis method for a suspended rail transit system according to claim 1, characterized in that: The determining the second fault diagnosis error of the taught deep learning network according to the taught fault state path diagnosis result and the optimized teaching fault state path diagnosis result comprises: Calculating a fuzzy fault diagnosis error based on the taught fault state path diagnosis result and the optimized teaching fault state path diagnosis result; and calculating a transfer learning fault diagnosis error based on the taught fault state path diagnosis result and the teaching fault state path diagnosis result; Determining a first influencing factor and a second influencing factor according to a current iteration round and a maximum iteration round of a fault diagnosis learning process; Calculating a first weighted result of the first influencing factor and the fuzzy fault diagnosis error, and a second weighted result of the second influencing factor and the transfer learning fault diagnosis error; The sum of the first weighted result and the second weighted result is output as the second fault diagnosis error of the taught deep learning network.

6. The AI ​​fault diagnosis method for a suspended rail transit system according to any one of claims 1 to 3, characterized in that: The fault diagnosis learning process includes a learning process of multiple iterative rounds, and the fault diagnosis neural network includes a teaching deep learning network and a taught deep learning network; The step of optimizing the fault diagnosis neural network according to the first fault diagnosis error and the second fault diagnosis error includes: According to the first fault diagnosis error and the second fault diagnosis error, optimizing the neuron weight parameters of the learning process of the taught deep learning network in the ath iteration round, where a is a positive integer; The neuron weight parameters of the teaching deep learning network are optimized according to the neuron weight parameters of the taught deep learning network before the a-th iteration round of learning process and the neuron weight parameters of the a-th iteration round of learning process.

7. The AI ​​fault diagnosis method for a suspended rail transit system according to any one of claims 1 to 3, characterized in that: The fault diagnosis learning process includes a learning process of multiple iterative rounds, and the method further includes: Selecting candidate training data from the unsupervised training data according to the second fault state path diagnosis result of the unsupervised training data in the a-th iteration round of learning, where a is a positive integer; Acquire the fault state path labeling data for adding training supervision data to the candidate training data, generate active supervision training data in the learning process of the a-th iteration round, and load the active supervision training data into the supervision training data sequence called by the learning process of the a+1-th iteration round; The step of selecting candidate training data from the unsupervised training data according to the second fault state path diagnosis result of the unsupervised training data in the a-th iteration round of learning process includes: According to the second fault state path diagnosis result of the unsupervised training data in the learning process of the a-th iteration round, marking the unsupervised training data that meets the target requirement as the candidate training data; The unsupervised training data that meets the target requirements includes at least one of the following data: The first X unsupervised training data are arranged in ascending order based on the diagnosis probability values ​​of the second fault state path diagnosis result; The first Y unsupervised training data are arranged in descending order based on the mean diagnostic likelihood of the second fault state path diagnostic result, wherein the mean diagnostic likelihood is the mean diagnostic likelihood of all diagnostic nodes in the second fault state path diagnostic result; The first Z unsupervised training data are arranged in descending order based on the likelihood ratio of the second fault state path diagnosis result, and the likelihood ratio is the ratio of the diagnosis nodes greater than a preset threshold in the second fault state path diagnosis result to all the diagnosis nodes; wherein X, Y and Z are all positive integers.

8. An AI fault diagnosis system for suspended rail transit systems, characterized in that: The AI ​​fault diagnosis system for the suspended rail transit system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the AI ​​fault diagnosis method for the suspended rail transit system described in any one of claims 1 to 7.

Images