Railway facility state prediction method based on working condition recognition and multi-task learning

Abstract

The present application provides a railway facility state prediction method based on working condition recognition and multi-task learning, belonging to the technical field of railway safe operation, comprising: collecting multi-source data in the process of railway facility operation by using sensor network and monitoring equipment and preprocessing; adopting the combination of convolutional neural network and long short-term memory network to construct a working condition feature extraction model, and extracting features from the preprocessed data to obtain working condition features; correlating and mining the working condition features to determine the number of task branches and the state prediction target of each task branch, and constructing a multi-task learning neural network model; inputting the real-time collected railway facility operation data and current working condition information into the trained multi-task learning prediction model to output the state prediction results of the railway facility in different directions; adjusting the structure and parameters of the working condition feature extraction model and the multi-task learning model according to the actual working condition change and the prediction effect feedback to ensure the accuracy of the state prediction.

Description

Technical Field

[0001] This invention relates to the field of railway safety operation technology, and in particular to a method for predicting the condition of railway facilities based on operating condition identification and multi-task learning. Background Technology

[0002] In railway transportation systems, the normal operation of railway facilities is crucial for ensuring transportation safety and efficiency. Traditional methods for predicting the condition of railway facilities mainly rely on manual inspections and simple data analysis. Manual inspections involve staff periodically checking the appearance of railway tracks, bridges, and signaling equipment, and measuring basic parameters such as track dimensions and signaling equipment voltage and current. This approach has significant drawbacks: firstly, manual inspections are inefficient, limited by manpower and time, making it difficult to conduct high-frequency, comprehensive inspections of railway facilities and failing to promptly detect potential minor damage or performance degradation; secondly, human judgment is subjective, with differences in experience and standards among inspectors leading to inconsistent results and affecting prediction accuracy. Furthermore, analysis methods based on simple statistical data can only handle single types of data, failing to uncover complex relationships between facility operation data, and are ill-suited to the changing operating conditions of railway facilities, thus unable to effectively predict changes in facility condition under different conditions. With the development of railway transportation towards high speed and heavy loads, the operating environment of railway facilities is becoming increasingly complex, and traditional prediction methods are no longer sufficient to meet the needs of ensuring safe and efficient railway operation. A new prediction method is urgently needed to improve the accuracy and reliability of railway facility condition predictions.

[0003] Therefore, this invention proposes a railway facility condition prediction method based on operating condition identification and multi-task learning. Summary of the Invention

[0004] This invention provides a method for predicting the state of railway facilities based on operating condition identification and multi-task learning, in order to solve the aforementioned technical problems.

[0005] This invention proposes a method for predicting the state of railway facilities based on operating condition identification and multi-task learning, comprising:

[0006] Step 1: Collect multi-source data during the operation of railway facilities using sensor networks and monitoring equipment, and perform preprocessing.

[0007] Step 2: Construct a working condition feature extraction model by combining convolutional neural networks and long short-term memory networks, and extract features from the preprocessed data to obtain working condition features;

[0008] Step 3: Perform correlation mining on the working condition features, determine the number of task branches and the state prediction target of each task branch, and construct a multi-task learning neural network model;

[0009] Step 4: Input the real-time collected railway facility operation data and current operating condition information into the trained multi-task learning prediction model, and output the state prediction results of the railway facility in different directions;

[0010] Step 5: Adjust the structure and parameters of the working condition feature extraction model and the multi-task learning model based on actual working condition changes and prediction effect feedback.

[0011] Preferably, multi-source data is collected and preprocessed using sensor networks and monitoring equipment during the operation of railway facilities, including:

[0012] Based on the distribution characteristics and operational features of railway facilities, a sensor network deployment plan was developed;

[0013] The sensor network is activated, and each sensor collects railway facility operation data in real time according to the set sampling frequency. During train operation, the on-board sensors continuously collect data on train speed, load and route, and transmit the data to the ground data aggregation node through the railway dedicated communication network. The data aggregation node performs preliminary integration and encoding of the received multi-source data.

[0014] The data collected from the data aggregation nodes is transmitted to the data processing center, where data cleaning algorithms are used to clean the collected data.

[0015] For different types of data with different dimensions and numerical ranges, a normalization algorithm is used to process the cleaned data, and a one-hot encoding method is used to convert categorical data.

[0016] Through statistical analysis of data, missing values ​​in the data are identified and filled in.

[0017] Preferably, a sensor network deployment plan is formulated based on the distribution characteristics and operational features of railway facilities, including:

[0018] Collect basic information on railway facilities and use data analysis tools to conduct in-depth analysis of the basic information to determine the key monitoring points and data requirements for each part of the railway facilities;

[0019] Based on the key monitoring points and data requirements, and combined with the operating environment characteristics of each key monitoring point, select the matching sensor type;

[0020] Based on the distribution of railway facilities and monitoring needs, geographic information system technology is used to conduct preliminary planning of sensor networks on electronic maps of railway lines. According to the key parts of the track and the characteristics of operation data, the deployment locations of sensors matched with the key monitoring points involved in the corresponding key parts are determined, thus forming a sensor network deployment scheme.

[0021] Preferably, step 2 includes:

[0022] Construct a CNN network and extract local features from the preprocessed data to obtain preliminary feature representations;

[0023] The preliminary feature representations are used as input to the Long Short-Term Memory Network for learning, to explore the dynamic changes of the preprocessed data in the time dimension, and to extract the temporal features and contextual information in the preprocessed data.

[0024] The features output by the LSTM network are fused with some local features extracted by the CNN network in the intermediate layer to form a complete working condition feature vector containing spatiotemporal information and dynamic change patterns of the data. The working condition feature vector is then output as the final extraction result.

[0025] Preferably, step 3 includes:

[0026] The association rule mining algorithm is used to analyze the working condition features, calculate the support and confidence between different working condition features, and find feature combinations with strong correlation.

[0027] The number of combinations of the aforementioned feature combinations is counted, and the number of task branches and the state prediction target of each task branch are determined in combination with the actual needs of railway facility status prediction.

[0028] Based on a shared underlying network, common features are extracted from all operating condition characteristics;

[0029] Connect the task branch to the shared underlying network;

[0030] For each task branch, the branch network is selected according to the state prediction target. The general features are processed to extract specific features related to the corresponding task branch, thus completing the state prediction task.

[0031] Preferably, the number of task branches is determined, including:

[0032] The total number of single working condition characteristics that exist in a strong correlation;

[0033] Calculate the frequency of occurrence and strong correlation coefficient of each working condition feature with strong correlation, and set the first coefficient for the corresponding working condition feature;

[0034] Assign a representation target to the corresponding feature combination based on the feature representation of each feature combination, and perform cluster analysis on all representation targets to obtain the number of clusters;

[0035] The comparison quantity is calculated based on all the first coefficients and the single total quantity;

[0036]

[0037] Where Nb represents the number of comparisons; N max N represents the maximum frequency of occurrence. max-1 Indicates the second highest frequency of occurrence; Indicates the rounding up symbol;

[0038] If the number of comparisons is less than or equal to the number of clusters, the number of clusters is set to the number of task branches;

[0039] If the number of comparisons is greater than the number of clusters, refine the precision of the cluster analysis, and re-cluster all representation targets, setting the new number of clusters as the number of task branches.

[0040] Preferably, the calculation formula for setting the first coefficient for the corresponding working condition characteristic is as follows:

[0041]

[0042] Where X represents the first coefficient of the corresponding working condition characteristic; Nc represents the frequency of occurrence of the corresponding working condition characteristic; Nd represents the single total quantity; g i1 It represents the strong correlation coefficient of the feature combination that appears for the i1th time in the corresponding working condition feature.

[0043] Preferably, the multi-source data includes: vibration data, temperature data, and geometric dimension data of railway tracks; stress-strain data and displacement data of bridges; current and voltage data and operating frequency data of signaling equipment; train speed, load, and route.

[0044] Preferably, the basic information includes: topographic data of the railway line, track design parameters, bridge structure information, and signal equipment layout diagram.

[0045] Preferably, based on key monitoring points and data requirements, and considering the operating environment characteristics of each key monitoring point, suitable sensor types are selected, including:

[0046] Analyze the data types and accuracy requirements that need to be monitored at each key monitoring point;

[0047] Conduct on-site surveys and data collection of the operating environment of each key monitoring point, analyze the environmental characteristics of the corresponding key monitoring points, and collect historical environmental data of each key monitoring point to analyze the patterns of environmental change.

[0048] Fuzzy comprehensive evaluation method is introduced to handle the uncertainty factors in the evaluation indicators. Based on the data requirements, environmental characteristics and environmental change patterns of each key monitoring point, sensor types that meet the basic measurement requirements are selected from the sensor product library.

[0049] Compared with the prior art, the beneficial effects of this application are as follows:

[0050] Preprocessing ensures data accuracy, providing a reliable data foundation for subsequent operating condition feature extraction and state prediction. By extracting operating condition features containing spatiotemporal information and dynamic change patterns from the preprocessed multi-source data, the key features of railway facility operation conditions can be captured more comprehensively and accurately. The multi-task learning neural network model can make full use of the correlation between operating condition features to achieve multi-task collaborative learning. By feeding back errors, the operating condition feature extraction model and the multi-task learning model can adapt to the dynamic changes of actual operating conditions and continuously optimize model performance. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0052] Figure 1 This is a flowchart of a railway facility condition prediction method based on working condition identification and multi-task learning provided by an embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0054] This invention proposes a method for predicting the state of railway facilities based on operating condition identification and multi-task learning, such as... Figure 1 As shown, it includes:

[0055] Step 1: Collect multi-source data during the operation of railway facilities using sensor networks and monitoring equipment, and perform preprocessing.

[0056] Step 2: Construct a working condition feature extraction model by combining convolutional neural networks and long short-term memory networks, and extract features from the preprocessed data to obtain working condition features;

[0057] Step 3: Perform correlation mining on the working condition features, determine the number of task branches and the state prediction target of each task branch, and construct a multi-task learning neural network model;

[0058] Step 4: Input the real-time collected railway facility operation data and current operating condition information into the trained multi-task learning prediction model, and output the state prediction results of the railway facility in different directions;

[0059] Step 5: Adjust the structure and parameters of the working condition feature extraction model and the multi-task learning model based on actual working condition changes and prediction effect feedback.

[0060] Preferably, the multi-source data includes: vibration data, temperature data, and geometric dimension data of railway tracks; stress-strain data and displacement data of bridges; current and voltage data and operating frequency data of signaling equipment; train speed, load, and route.

[0061] In this embodiment, vibration sensors and temperature sensors are installed at regular intervals along the railway track, while stress-strain sensors and displacement sensors are installed at key stress points such as bridge piers and main beams. These sensors are connected wirelessly or via wired means to form a network capable of real-time data acquisition. Monitoring equipment refers to devices used to assist sensors in data acquisition or to monitor specific facilities. For example, a track gauge measuring instrument can measure the geometric dimensions of the track, and a signal monitoring instrument can collect current, voltage, and operating frequency data from signal equipment.

[0062] Multi-source data refers to data from multiple sources and of multiple types. Taking railway facilities as an example, it includes vibration data of railway tracks (reflecting the vibration of the tracks when trains pass), temperature data (showing the changes of the tracks under different ambient temperatures), and geometric dimension data (such as track gauge and track elevation); stress and strain data of bridges (showing the stress on the bridge structure) and displacement data (monitoring whether the bridge has experienced displacement or deformation); current and voltage data of signaling equipment (determining whether the power supply to the signaling equipment is normal) and operating frequency data (ensuring stable signal transmission); train speed, load (understanding the impact of train operating conditions on facilities), and running lines (analyzing the state of facilities under different line conditions), etc.

[0063] Operating condition characteristics are key features reflecting the operating conditions of railway facilities. For example, abnormal vibration frequency characteristics extracted from track vibration data, stress change characteristics extracted from bridge stress and strain data, and comprehensive operating condition characteristics obtained by combining factors such as train speed and load can characterize the current operating status of railway facilities.

[0064] In this embodiment, correlation mining is the process of discovering potential correlations between features from a large amount of operational feature data. For example, analysis reveals that when a train passes through a railway curve at a high speed and with a heavy load, there is a significant correlation between the vibration frequency characteristics of the track and the stress-strain characteristics of the bridge piers. This correlation mining helps to gain a deeper understanding of the mutual influence between different parts of the railway facility.

[0065] In a multi-task learning neural network model, a task branch is an independent task unit divided to target different prediction objectives. For example, track wear prediction can be one task branch, and bridge structural health prediction can be another task branch, with each branch focusing on the state prediction of a specific facility or aspect.

[0066] In this embodiment, the operating condition information reflects the current operating condition of the railway facilities, including the current train speed, load, and route, as well as environmental factors (such as weather conditions). This information is closely related to the condition of the facilities.

[0067] Condition prediction results include, for example, predicting the wear and tear of tracks over the next week, stress changes in bridge structures when the next train passes, and the likelihood of signal equipment malfunctioning within the next month.

[0068] In this embodiment, the prediction performance feedback is obtained by comparing the model's prediction results with the actual situation, providing feedback information on the accuracy and effectiveness of the prediction. For example, if the model predicts that the wear of a certain section of track this week will be 5mm, while the actual measured wear is 8mm, the difference between the two is part of the prediction performance feedback.

[0069] For the working condition feature extraction model, adjust the network structure of CNN and LSTM (such as increasing the number of convolutional layers or adjusting the number of LSTM units) or parameters (such as learning rate and convolutional kernel size); for the multi-task learning model, modify the structure of the shared underlying network and task branch network, readjust the task branch weights, and retrain the model using a new dataset.

[0070] The beneficial effects of the above technical solution are as follows: Preprocessing ensures the accuracy of the data, providing a reliable data foundation for subsequent operating condition feature extraction and state prediction; extracting operating condition features containing spatiotemporal information and dynamic change patterns from the preprocessed multi-source data can more comprehensively and accurately capture the key features of railway facility operation conditions; the multi-task learning neural network model can make full use of the correlation between operating condition features to achieve multi-task collaborative learning; and through feedback error, the operating condition feature extraction model and the multi-task learning model can adapt to the dynamic changes of actual operating conditions and continuously optimize model performance.

[0071] This invention proposes a railway facility condition prediction method based on operating condition identification and multi-task learning. It utilizes sensor networks and monitoring equipment to collect multi-source data during railway facility operation and performs preprocessing, including:

[0072] Based on the distribution characteristics and operational features of railway facilities, a sensor network deployment plan was developed;

[0073] The sensor network is activated, and each sensor collects railway facility operation data in real time according to the set sampling frequency. During train operation, the on-board sensors continuously collect data on train speed, load and route, and transmit the data to the ground data aggregation node through the railway dedicated communication network. The data aggregation node performs preliminary integration and encoding of the received multi-source data.

[0074] The data collected from the data aggregation nodes is transmitted to the data processing center, where data cleaning algorithms are used to clean the collected data.

[0075] For different types of data with different dimensions and numerical ranges, a normalization algorithm is used to process the cleaned data, and a one-hot encoding method is used to convert categorical data.

[0076] Through statistical analysis of data, missing values ​​in the data are identified and filled in.

[0077] In this embodiment, railway facilities include railway tracks, bridges, tunnels, signaling equipment, power supply equipment, etc., and the distribution characteristics refer to the layout of railway facilities in geographical space.

[0078] Operational characteristics involve the impact of factors such as train speed, load, frequency of operation, and track gradient on railway facilities.

[0079] In this embodiment, the on-board sensors are sensors installed on the train to collect the train's own operating parameters. For example, speed sensors measure the train's speed; load sensors detect the weight of the cargo loaded on the train; and positioning sensors acquire information about the train's route.

[0080] In this embodiment, the dedicated railway communication network is a communication network specifically built for railway data transmission. It can adopt wireless communication (such as 4G / 5G private network) or wired communication (such as fiber optic). The data aggregation node is a device used to receive and aggregate data transmitted from multiple sensors.

[0081] In this embodiment, the data cleaning algorithm is an algorithm used to identify and process problems such as noise, outliers, and duplicate data in the data.

[0082] Categorical data: Data representing different categories or attributes. Examples include train route names (such as the Beijing-Shanghai line and the Beijing-Guangzhou line) and signal equipment types (such as traffic lights and turnout controllers).

[0083] One-hot encoding: a method of encoding that converts categorical data into binary vectors. For example, for train routes such as "Beijing-Shanghai Line", "Beijing-Guangzhou Line", and "Beijing-Kowloon Line", after one-hot encoding, "Beijing-Shanghai Line" can be represented as [1,0,0], "Beijing-Guangzhou Line" as [0,1,0], and "Beijing-Kowloon Line" as [0,0,1].

[0084] The beneficial effects of the above technical solution are as follows: The sensor network deployment scheme can ensure that sensors fully cover key railway facilities and parts, collect key data reflecting the operating status of railway facilities, and after preliminary integration and encoding, ensure that the data format is uniform and the structure is clear, thereby improving the stability and efficiency of data transmission. After data cleaning, outliers, duplicate data and noise in the data are effectively removed, and the accuracy and reliability of the data are significantly improved. After normalization and one-hot encoding, different types of data have a unified format and scale, eliminating the impact of differences in dimensions and numerical ranges on data analysis, and filling in missing values ​​in the data, thereby restoring the integrity of the data.

[0085] This invention proposes a railway facility condition prediction method based on operating condition identification and multi-task learning. Based on the distribution characteristics and operational features of railway facilities, a sensor network deployment scheme is formulated, including:

[0086] Collect basic information on railway facilities and use data analysis tools to conduct in-depth analysis of the basic information to determine the key monitoring points and data requirements for each part of the railway facilities;

[0087] Based on the key monitoring points and data requirements, and combined with the operating environment characteristics of each key monitoring point, select the matching sensor type;

[0088] Based on the distribution of railway facilities and monitoring needs, geographic information system technology is used to conduct preliminary planning of sensor networks on electronic maps of railway lines. According to the key parts of the track and the characteristics of operation data, the deployment locations of sensors matched with the key monitoring points involved in the corresponding key parts are determined, thus forming a sensor network deployment scheme.

[0089] The basic information includes: topographic data of the railway line, track design parameters, bridge structure information, and signal equipment layout diagram.

[0090] In this embodiment, the basic information of railway facilities is the basic data for the construction and operation of railway facilities. The topographic data describes the natural terrain features of the areas through which the railway line passes, such as plains, mountains, and river distribution. The track design parameters cover track type (such as ballastless track and ballasted track), track gauge, curve radius, etc. The bridge structure data includes bridge type (such as beam bridge, arch bridge, and cable-stayed bridge), span, load-bearing design, etc. The signal equipment layout diagram shows the installation location and connection relationship of equipment such as signal lights, turnout controllers, and communication base stations.

[0091] Data analysis tools are software or algorithms used to process and analyze data, such as Excel, Python's Pandas library, SPSS and other statistical analysis software, which can perform data cleaning, statistical calculations, visualization and other operations.

[0092] Key monitoring points include, for example, turnout areas, rail joints, and curves in the track; bridge piers, towers, and key stress points of the main beam; and the core control boxes of signaling equipment.

[0093] Data requirements specify the types of data and accuracy requirements to be collected at key monitoring points. For example, key track monitoring points may require the collection of vibration, temperature, and geometric dimension data, with vibration sensor accuracy required to reach ±0.1 m / s². 2 Key monitoring points on the bridge need to monitor stress, strain, and displacement data, and the accuracy of stress and strain sensors needs to reach ±1με.

[0094] Sensor types are devices used to sense different physical quantities, such as vibration sensors, temperature sensors, stress and strain sensors, displacement sensors, current and voltage sensors, etc.

[0095] Railway electronic maps are maps that present information such as railway line routes, station distribution, and topography in digital form. They are the result of the application of GIS technology in the railway field.

[0096] The beneficial effects of the above technical solution are: by analyzing the basic information, key monitoring points such as track curves and bridge main spans are accurately located, and the data monitoring requirements for vibration, stress and strain at each point are determined.

[0097] This invention proposes a method for predicting the state of railway facilities based on operating condition identification and multi-task learning. Step 2 includes:

[0098] Construct a CNN network and extract local features from the preprocessed data to obtain preliminary feature representations;

[0099] The preliminary feature representations are used as input to the Long Short-Term Memory Network for learning, to explore the dynamic changes of the preprocessed data in the time dimension, and to extract the temporal features and contextual information in the preprocessed data.

[0100] The features output by the LSTM network are fused with some local features extracted by the CNN network in the intermediate layer to form a complete working condition feature vector containing spatiotemporal information and dynamic change patterns of the data. The working condition feature vector is then output as the final extraction result.

[0101] In this embodiment, representative local features are selected from the intermediate layers of the CNN network. These features reflect the local details and structural information of the data. The temporal features and contextual information output by the LSTM network are fused with the selected local features from the intermediate layers of the CNN network using methods such as concatenation and weighted summation. For example, the feature vector output by the LSTM network and the feature map of the intermediate layer of the CNN network are unfolded and concatenated to form a new feature vector; or different features are assigned weights according to their importance, and the fused feature vector is obtained by weighted summation. The fused feature vector is then normalized to form the final working condition feature vector, which is used as the output of the entire feature extraction process.

[0102] The beneficial effect of the above technical solution is that the fused feature vector obtained by using the features extracted by CNN network and LSTM network can more comprehensively describe the operating conditions of railway facilities.

[0103] This invention proposes a method for predicting the state of railway facilities based on operating condition identification and multi-task learning. Step 3 includes:

[0104] The association rule mining algorithm is used to analyze the working condition features, calculate the support and confidence between different working condition features, and find feature combinations with strong correlation.

[0105] The number of combinations of the aforementioned feature combinations is counted, and the number of task branches and the state prediction target of each task branch are determined in combination with the actual needs of railway facility status prediction.

[0106] Based on a shared underlying network, common features are extracted from all operating condition characteristics;

[0107] Connect the task branch to the shared underlying network;

[0108] For each task branch, the branch network is selected according to the state prediction target. The general features are processed to extract specific features related to the corresponding task branch, thus completing the state prediction task.

[0109] In this embodiment, the operating condition characteristics reflect the key characteristics of the railway facility's operating conditions, such as the vibration frequency and temperature changes of the track, the stress and strain values ​​and displacement of the bridge, and the characteristics reflected by data such as the train's operating speed and load.

[0110] Support: This represents the proportion of all data sets that simultaneously contain certain combinations of features, measuring the frequency of these combinations. For example, if 200 out of 1000 sets of railway operating condition data simultaneously contain the features "high-speed train operation" and "high track vibration frequency," then the support for this feature combination is 20%.

[0111] Confidence level: Given a given feature, the probability of another feature appearing reflects the reliability of the association between features. For example, in 500 sets of data where "train is running at high speed" appears, 400 sets also show "high track vibration frequency". Therefore, the confidence level of the association rule "train running at high speed → high track vibration frequency" is 80%.

[0112] Feature combination: A collection of multiple operating condition features, such as "high-speed train operation + heavy load + high track vibration frequency" is a feature combination. The operating condition feature data of railway facilities obtained through the operating condition feature extraction model is organized into a format suitable for algorithm processing, such as a transaction dataset. The Apriori algorithm or FP-Growth algorithm is used to analyze the operating condition feature data, setting support thresholds (e.g., 15%) and confidence thresholds (e.g., 70%). The algorithm first finds all individual features (frequent 1-itemsets) that meet the minimum support, and then progressively generates higher-order frequent itemsets, calculating the support and confidence of each feature combination. Feature combinations with both support and confidence higher than the threshold are selected; these are feature combinations with strong correlations.

[0113] The state prediction objective is the specific prediction content that each task branch needs to achieve, such as predicting the degree of track wear, the probability of bridge structure damage, and the failure time of signal equipment.

[0114] General features are valuable across multiple task branches, reflecting commonalities in the data. Examples include time information and basic train operation parameters (speed, load, etc.) in railway condition data. These features provide valuable reference for predicting the condition of different facilities such as tracks, bridges, and signaling equipment. Specific features, on the other hand, are features closely related to the condition prediction objective of each task branch. They are obtained by further extraction and transformation of general features through a branch network and can be directly used for condition prediction. For instance, in the track wear prediction task branch, the features related to track material properties and train wheel-rail forces obtained through branch network processing are specific features.

[0115] The beneficial effects of the above technical solution are: by statistically combining the number of features, it is easy to reasonably determine the number of task branches and clarify the state prediction target; based on the shared underlying network, representative general features are successfully extracted from the working condition feature data; the connection between the task branches and the shared underlying network constructs a complete multi-task learning neural network model architecture; the branch network of each task branch can effectively extract specific features and complete the state prediction task.

[0116] This invention proposes a railway facility condition prediction method based on working condition identification and multi-task learning, which determines the number of task branches, including:

[0117] The total number of single working condition characteristics that exist in a strong correlation;

[0118] Calculate the frequency of occurrence and strong correlation coefficient of each working condition feature with strong correlation, and set the first coefficient for the corresponding working condition feature;

[0119] Assign a representation target to the corresponding feature combination based on the feature representation of each feature combination, and perform cluster analysis on all representation targets to obtain the number of clusters;

[0120] The comparison quantity is calculated based on all the first coefficients and the single total quantity;

[0121]

[0122] Where Nb represents the number of comparisons; N max N represents the maximum frequency of occurrence. max-1 Indicates the second highest frequency of occurrence; Indicates the rounding up symbol;

[0123] If the number of comparisons is less than or equal to the number of clusters, the number of clusters is set to the number of task branches;

[0124] If the number of comparisons is greater than the number of clusters, refine the precision of the cluster analysis, and re-cluster all representation targets, setting the new number of clusters as the number of task branches.

[0125] Preferably, the calculation formula for setting the first coefficient for the corresponding working condition characteristic is as follows:

[0126]

[0127] Where X represents the first coefficient of the corresponding working condition characteristic; Nc represents the frequency of occurrence of the corresponding working condition characteristic; Nd represents the single total quantity; g i1 It represents the strong correlation coefficient of the feature combination that appears for the i1th time in the corresponding working condition feature.

[0128] In this embodiment, the maximum frequency of occurrence N max With the second highest frequency N max-1 It reflects the frequency of occurrence of key operating condition features in strong correlations, and the total number of Nd reflects the overall scale of operating condition features with strong correlations.

[0129] In this embodiment, setting reasonable coefficients for each operating condition feature is crucial for accurately measuring the importance of the operating condition feature in a strong correlation. Dividing the two yields the relative frequency of this feature in the whole. The strong correlation coefficient of the feature combination in which the feature of this working condition appears each time is comprehensively considered, reflecting the tightness of the correlation when it appears.

[0130] The beneficial effects of the above technical solution are: it comprehensively considers multiple dimensions, avoiding the one-sidedness caused by determining the number of task branches based on only a single factor, making the determination of the number of task branches more scientific and reasonable. The calculated comparison number Nb serves as a reference value, compared with the clustering number, and the number of task branches is adjusted accordingly. The calculated first coefficient X can assign a quantitative importance index to each working condition feature. When determining the number of task branches and allocating resources, different working condition features can be treated differently based on this coefficient, highlighting the role of important features and making the model focus more on key working condition features and their correlations.

[0131] This invention proposes a railway facility condition prediction method based on operating condition identification and multi-task learning. According to key monitoring points and data requirements, and considering the operating environment characteristics of each key monitoring point, it selects matching sensor types, including:

[0132] Analyze the data types and accuracy requirements that need to be monitored at each key monitoring point;

[0133] Conduct on-site surveys and data collection of the operating environment of each key monitoring point, analyze the environmental characteristics of the corresponding key monitoring points, and collect historical environmental data of each key monitoring point to analyze the patterns of environmental change.

[0134] Fuzzy comprehensive evaluation method is introduced to handle the uncertainty factors in the evaluation indicators. Based on the data requirements, environmental characteristics and environmental change patterns of each key monitoring point, sensor types that meet the basic measurement requirements are selected from the sensor product library.

[0135] In this embodiment, the accuracy requirement is a quantitative standard for the accuracy of the acquired data. For example, the accuracy requirement for a track vibration sensor is ±0.1 m / s². 2 This ensures that subtle changes in track vibration can be accurately captured; the bridge stress-strain sensor accuracy needs to reach ±1με in order to accurately monitor changes in the stress on the bridge structure.

[0136] In this embodiment, historical environmental data refers to environmental parameter records of key monitoring points over a period of time. Examples include daily temperature and humidity data for the past year, and annual maximum wind speed and rainfall data for the past five years.

[0137] A fuzzy evaluation set of sensor performance indicators is established, with each indicator divided into different fuzzy levels (e.g., "Excellent," "Good," "Average," "Poor"). Based on the data requirements, environmental characteristics, and variation patterns of each key monitoring point, the weight of each evaluation indicator is determined; for example, in environments with strong electromagnetic interference, the anti-interference capability indicator has a higher weight. Performance parameter information of various sensors is collected from a sensor product library. For each key monitoring point, the fuzzy comprehensive evaluation method is used to evaluate candidate sensors. By calculating the membership degree of each sensor under different indicators and combining the indicator weights, a comprehensive score for each sensor is obtained. Sensor types that meet basic measurement requirements are selected according to their scores, taking into account uncertainties such as accuracy, range, and environmental adaptability.

[0138] The beneficial effects of the above technical solution are: clarifying the data types and accuracy requirements of each key monitoring point provides a clear technical basis for subsequent sensor selection; collecting the operating environment data of key monitoring points helps to gain a deeper understanding of the environmental conditions of key monitoring points, which facilitates the subsequent selection of appropriate sensor types.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the state of railway facilities based on operating condition identification and multi-task learning, characterized in that, include: Step 1: Collect multi-source data during the operation of railway facilities using sensor networks and monitoring equipment, and perform preprocessing. Step 2: Construct a working condition feature extraction model by combining convolutional neural networks and long short-term memory networks, and extract features from the preprocessed data to obtain working condition features; Step 3: Perform correlation mining on the working condition features, determine the number of task branches and the state prediction target of each task branch, and construct a multi-task learning neural network model; Step 4: Input the real-time collected railway facility operation data and current operating condition information into the trained multi-task learning prediction model, and output the state prediction results of the railway facility in different directions; Step 5: Adjust the structure and parameters of the working condition feature extraction model and the multi-task learning model based on actual working condition changes and prediction effect feedback; Step 3 includes: The association rule mining algorithm is used to analyze the working condition features, calculate the support and confidence between different working condition features, and find feature combinations with strong correlation. The number of combinations of the aforementioned feature combinations is counted, and the number of task branches and the state prediction target of each task branch are determined in combination with the actual needs of railway facility status prediction. Based on a shared underlying network, common features are extracted from all operating condition characteristics; Connect the task branch to the shared underlying network; The branch network of each task branch is filtered according to the state prediction target, the general features are processed, and specific features related to the corresponding task branch are extracted to complete the state prediction task. Determine the number of task branches, including: The total number of single working condition characteristics that exist in a strong correlation; Calculate the frequency of occurrence and strong correlation coefficient of each working condition feature with strong correlation, and set the first coefficient for the corresponding working condition feature; Assign a representation target to the corresponding feature combination based on the feature representation of each feature combination, and perform cluster analysis on all representation targets to obtain the number of clusters; The comparison quantity is calculated based on all the first coefficients and the single total quantity; ； in, Indicates comparison of quantities; Indicates the maximum frequency of occurrence; This indicates the second highest frequency of occurrence; Indicates the rounding up symbol; If the number of comparisons is less than or equal to the number of clusters, the number of clusters is set to the number of task branches; If the number of comparisons is greater than the number of clusters, refine the precision of the cluster analysis, and re-cluster all representation targets, setting the new number of clusters as the number of task branches.

2. The railway facility condition prediction method based on working condition identification and multi-task learning according to claim 1, characterized in that, Collecting and preprocessing multi-source data during the operation of railway facilities using sensor networks and monitoring equipment, including: Based on the distribution characteristics and operational features of railway facilities, a sensor network deployment plan was developed; The sensor network is activated, and each sensor collects railway facility operation data in real time according to the set sampling frequency. During train operation, the on-board sensors continuously collect data on train speed, load and route, and transmit the data to the ground data aggregation node through the railway dedicated communication network. The data aggregation node performs preliminary integration and encoding of the received multi-source data. The data collected from the data aggregation nodes is transmitted to the data processing center, where data cleaning algorithms are used to clean the collected data. For different types of data with different dimensions and numerical ranges, a normalization algorithm is used to process the cleaned data, and a one-hot encoding method is used to convert categorical data. Through statistical analysis of data, missing values ​​in the data are identified and filled in.

3. The railway facility condition prediction method based on working condition identification and multi-task learning according to claim 2, characterized in that, Based on the distribution characteristics and operational features of railway facilities, a sensor network deployment plan is formulated, including: Collect basic information on railway facilities and use data analysis tools to conduct in-depth analysis of the basic information to determine the key monitoring points and data requirements for each part of the railway facilities; Based on the key monitoring points and data requirements, and combined with the operating environment characteristics of each key monitoring point, select the matching sensor type; Based on the distribution of railway facilities and monitoring needs, geographic information system technology is used to conduct preliminary planning of sensor networks on electronic maps of railway lines. According to the key parts of the track and the characteristics of operation data, the deployment locations of sensors matched with the key monitoring points involved in the corresponding key parts are determined, thus forming a sensor network deployment scheme.

4. The railway facility condition prediction method based on working condition identification and multi-task learning according to claim 1, characterized in that, Step 2 includes: Construct a CNN network and extract local features from the preprocessed data to obtain preliminary feature representations; The preliminary feature representations are used as input to the Long Short-Term Memory Network for learning, to explore the dynamic changes of the preprocessed data in the time dimension, and to extract the temporal features and contextual information in the preprocessed data. The features output by the LSTM network are fused with some local features extracted by the CNN network in the intermediate layer to form a complete working condition feature vector containing spatiotemporal information and dynamic change patterns of the data. The working condition feature vector is then output as the final extraction result.

5. The railway facility condition prediction method based on working condition identification and multi-task learning according to claim 1, characterized in that, The formula for calculating the first coefficient for the corresponding working condition characteristic is as follows: ； in, The first coefficient represents the characteristic of the corresponding working condition; Indicates the frequency of occurrence of the corresponding operating condition characteristics; Indicates a single total quantity; It represents the strong correlation coefficient of the feature combination that appears for the i1th time in the corresponding working condition feature.

6. The railway facility condition prediction method based on working condition identification and multi-task learning according to claim 1, characterized in that, The multi-source data includes: vibration data, temperature data, and geometric dimension data of railway tracks; stress and strain data and displacement data of bridges; current and voltage data and operating frequency data of signaling equipment; train speed, load, and route.

7. The railway facility condition prediction method based on working condition identification and multi-task learning according to claim 3, characterized in that, The basic information includes: topographic data of the railway line, track design parameters, bridge structure information, and signal equipment layout diagram.

8. The railway facility condition prediction method based on working condition identification and multi-task learning according to claim 1, characterized in that, Based on key monitoring points and data requirements, and considering the operational environment characteristics of each key monitoring point, suitable sensor types are selected, including: Analyze the data types and accuracy requirements that need to be monitored at each key monitoring point; Conduct on-site surveys and data collection of the operating environment of each key monitoring point, analyze the environmental characteristics of the corresponding key monitoring points, and collect historical environmental data of each key monitoring point to analyze the patterns of environmental change. Fuzzy comprehensive evaluation method is introduced to handle the uncertainty factors in the evaluation indicators. Based on the data requirements, environmental characteristics and environmental change patterns of each key monitoring point, sensor types that meet the basic measurement requirements are selected from the sensor product library.

Images