Train fault remote warning method and system based on artificial intelligence

Abstract

The present invention provides a remote warning method and system for train faults based on artificial intelligence, which realizes a comprehensive and accurate characterization of the train operation status by accurately extracting the multi-dimensional state path vectors in the target train operation data and fusing them to generate the first train state feature map. Combined with the prior warning guidance knowledge data, the present invention uses the interactive focus feature analysis technology to effectively guide the train fault warning decision model to make accurate warning decisions on the target train operation data. By introducing the first characteristic flow control parameter to optimize the interactive focus feature data, a more targeted target guidance knowledge vector is generated, further improving the accuracy and reliability of fault warning. Finally, the present invention can generate train fault warning knowledge points in real time and remotely, providing a strong technical guarantee for the safe operation of trains, significantly reducing the incidence and impact range of train failures, and improving the safety and efficiency of railway transportation.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based train fault remote warning method and system. Background Art

[0002] With the rapid development of railway transportation, the safety and reliability of train operation have become crucial issues. However, during the operation of trains, various faults are inevitable due to various complex factors (such as equipment aging, environmental factors, human operation, etc.). If these faults are not discovered and handled in time, they are likely to cause serious safety accidents.

[0003] Traditional train fault detection methods usually rely on manual inspections and regular maintenance, but this method has many limitations. First, manual inspections require a lot of manpower and time, and are limited by the experience and skill level of inspectors, making it difficult to ensure the accuracy and timeliness of fault detection. Second, although regular maintenance can prevent faults to a certain extent, it is often difficult to cope with sudden faults. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a train fault remote warning method based on artificial intelligence, the method comprising:

[0005] Acquire target train operation data, extract state path vectors of multiple state dimensions from the target train operation data, and fuse the state path vectors of the multiple state dimensions to generate a first train state characteristic map of the target train operation data;

[0006] Acquire priori warning guidance knowledge data, extract a first guidance knowledge vector of the priori warning guidance knowledge data, and perform interactive focus feature analysis on the first guidance knowledge vector and the first train state feature map, wherein the priori warning guidance knowledge data is used to guide a train fault warning decision model to make a train fault warning decision on the target train operation data;

[0007] Optimizing the interactive focusing feature data according to a first feature flow control parameter to generate a target guidance knowledge vector, wherein the first feature flow control parameter is generated by performing parameter learning after locking the model parameter information of the train fault early warning decision model;

[0008] The train fault warning decision model is used to make a fault warning decision based on the target guidance knowledge vector to generate train fault warning knowledge points for the target train operation data.

[0009] On the other hand, an embodiment of the present invention also provides an artificial intelligence-based train fault remote warning system, including a processor and a machine-readable storage medium, wherein 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.

[0010] Based on the above aspects, by accurately extracting the multi-dimensional state path vectors in the target train operation data and fusing them to generate the first train state feature map, a comprehensive and accurate characterization of the train operation status is achieved. Combined with the prior warning guidance knowledge data, the present invention uses interactive focus feature analysis technology to effectively guide the train fault warning decision model to make accurate warning decisions on the target train operation data. By introducing the first characteristic flow control parameter to optimize the interactive focus feature data, a more targeted target guidance knowledge vector is generated, further improving the accuracy and reliability of fault warning. Ultimately, the present invention can generate train fault warning knowledge points in real time and remotely, providing a strong technical guarantee for the safe operation of trains, significantly reducing the incidence and impact range of train failures, and improving the safety and efficiency of railway transportation. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the execution flow of the artificial intelligence-based train fault remote warning method provided in an embodiment of the present invention.

[0012] Figure 2 It is a schematic diagram of the hardware architecture of an artificial intelligence-based train fault remote warning system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 The present invention is a flowchart of an artificial intelligence-based remote warning method for train faults provided by an embodiment of the present invention. The artificial intelligence-based remote warning method for train faults is introduced in detail below.

[0014] Step S110, obtaining target train operation data, extracting state path vectors of multiple state dimensions in the target train operation data, and fusing the state path vectors of the multiple state dimensions to generate a first train state characteristic map of the target train operation data.

[0015] In this embodiment, a high-speed train is running on a busy railway line. The train is equipped with numerous sensors to collect various data, which can be used as target train operation data after being collected.

[0016] First, identify and define multiple state dimensions in the target train operation data. For example, the state dimension of train speed is obtained through speed sensors installed on the train wheels or power system. Speed ​​data is recorded at a certain frequency (for example, once per second), and these continuous speed values ​​reflect the speed changes of the train during operation. Data of the acceleration dimension is obtained through acceleration sensors, which can reflect the speed of the train acceleration or deceleration. Data of the brake pressure dimension comes from the pressure sensor in the brake system. The magnitude of the brake pressure is crucial for the safe parking of the train. Data of the motor current dimension is collected from the motor system of the train. The change of the motor current can reflect whether the working state of the motor is normal. For example, if the motor current suddenly increases or decreases abnormally, it may indicate that the motor is faulty. Data of the car temperature dimension is collected by the temperature sensor in the car. Abnormal increase or decrease in the car temperature may indicate a fault in the air conditioning system or other related equipment. Data of the track condition dimension may be obtained through track detection equipment installed at the bottom of the train. These devices can detect the flatness of the track, the wear of the track, etc. Data in the external environmental condition dimension may include weather data (such as temperature, humidity, wind speed, etc.), which can be obtained from meteorological sensors outside the train or by receiving data from external meteorological stations.

[0017] For each identified state dimension, extract the state path data of the state dimension changing over time from the target train operation data. Take train speed as an example. In a run from station A to station B, assuming that the train runs for 2 hours and the time interval is 1 minute, 120 speed data points will be obtained. These data points arranged in chronological order constitute the state path data of the state dimension of train speed. For the acceleration dimension, data is also collected at this time interval to obtain the corresponding state path data.

[0018] Then, the state path data of each state dimension is vectorized. Taking the train speed as an example, the 120 speed data points are sequentially combined into a vector, which is the state path vector of the train speed. The same processing is performed on other state dimensions such as acceleration and brake pressure.

[0019] Next, for each state path vector, calculate the mean and standard deviation of the state path vector, and subtract the mean from each element of the state path vector and divide it by the standard deviation to get the standardized state path vector. For example, the mean of the train speed state path vector is 200 km / h and the standard deviation is 10 km / h. Then, a data point with a speed value of 210 km / h becomes (210 - 200) / 10 = 1 after standardization.

[0020] Based on the state path vectors of each state dimension after standardization, the dimensional features of each state dimension are extracted. Taking the train speed state dimension as an example, the mean value in the statistical feature may be 200 km / h, and the variance reflects the fluctuation of the speed. If the variance is large, it means that the train speed is unstable. The maximum value may be 250 km / h (the maximum speed during acceleration), and the minimum value may be 150 km / h (the minimum speed during deceleration when entering the station). The zero-crossing rate in the time domain feature can reflect the frequency of speed change. If the zero-crossing rate is high, it means that the train speed frequently switches between increasing and decreasing. The spectral density in the frequency domain feature can reveal the energy distribution of speed data at different frequencies. If the spectral density is abnormally high at a certain frequency, it may imply that the train has periodic speed anomalies. Features based on time series analysis, such as features of autocorrelation function analysis, can show the correlation of speed data at different time lags. If the autocorrelation function has a large value at a certain time lag, it may mean that there is a certain periodic law in the train speed.

[0021] The feature selection algorithm is used to evaluate the importance of each feature in the dimensional features of each state dimension, and after selecting the target feature set that is relevant to the train fault warning, the linear discriminant analysis algorithm is used to map the target feature set from the high-dimensional feature space to the low-dimensional space, and the feature vectors of each state dimension after dimensionality reduction are output. For example, after the feature selection algorithm is evaluated, it is found that the variance, maximum value, and a certain feature in the autocorrelation function analysis of the train speed are very important for train fault warning, so these features are formed into the target feature set. Then, the linear discriminant analysis algorithm is used to map this target feature set from the possible high-dimensional space (assuming that there are originally 10 relevant features, forming a 10-dimensional space) to a low-dimensional space (such as a 3-dimensional space), and the reduced-dimensional train speed feature vector is obtained.

[0022] Finally, the feature vectors of each state dimension after dimensionality reduction are fused to generate a comprehensive feature representation of the target train operation data, and weights are assigned to the feature vectors of each state dimension. Assume that the weight of the train speed feature vector in the comprehensive feature representation is 0.3, the weight of the acceleration feature vector is 0.2, the weight of the brake pressure feature vector is 0.25, etc. Based on the fused comprehensive feature representation and its weight, the first train state feature map of the target train operation data is constructed. The comprehensive feature representation is regarded as a node of the first train state feature map, the connection between the nodes represents the association relationship between the comprehensive feature representations, and the weight reflects the importance of the node in the first train state feature map. For example, if there is a strong correlation between train speed and brake pressure (because changes in train speed affect the demand for brake pressure), then in the first train state feature map, the connection between the train speed node and the brake pressure node will be relatively strong, and their weights will also affect the strength of this connection to a certain extent.

[0023] Step S120, obtaining prior warning guidance knowledge data, extracting a first guidance knowledge vector of the prior warning guidance knowledge data, and performing interactive focusing feature analysis on the first guidance knowledge vector and the first train state feature map, wherein the prior warning guidance knowledge data is used to guide the train fault warning decision model to make a train fault warning decision on the target train operation data.

[0024] In this embodiment, the server obtains prior warning guidance knowledge data. Prior warning guidance knowledge data is data summarized based on previous train operation experience and fault cases. For example, in the past train operation records, there were many train failures that were recorded in detail, including the train status before the failure, the type of failure, the train operation environment when the failure occurred, and other information.

[0025] First, search for a second train state characteristic map that has a characteristic pattern matching the first train state characteristic map, and a priori fault warning knowledge points that match the second train state characteristic map. Assume that a large amount of prior train operation data and their prior fault warning knowledge points are stored in the server database. For each set of prior train operation data, the state path vectors of multiple state dimensions in each prior train operation data are fused to generate a second train state characteristic map for each prior train operation data. Taking the high-speed train mentioned above as an example, for a set of prior train operation data that has experienced a motor failure, the state path vectors of its train speed, acceleration, motor current and other state dimensions are fused to obtain the second train state characteristic map of this set of data.

[0026] Each second train state characteristic map and the corresponding prior fault warning knowledge point are integrated and added to the index database. In the index database, the second train state characteristic map is used as the index key and the prior fault warning knowledge point is used as the corresponding value for storage. When the current target train operation data is to be analyzed, the second train state characteristic map that matches the first train state characteristic map and the prior fault warning knowledge point that matches the second train state characteristic map are searched from the index database. Assuming that the first train state characteristic map shows that the train speed fluctuates greatly and the motor current has a small abnormal fluctuation, the server searches for the second train state characteristic map with a similar characteristic pattern in the index database, and may find the second train state characteristic map in the previous motor failure case. The matching prior fault warning knowledge point may include complex logical content such as the motor failure may occur within the next 10 minutes and the motor cooling system needs to be checked.

[0027] Construct prior warning guidance knowledge data based on prior fault warning knowledge points. Integrate the trigger node and the termination node in the forward and backward nodes of the prior fault warning knowledge points respectively to generate the knowledge point representation space corresponding to the prior fault warning knowledge points. For example, for the prior fault warning knowledge points of the above motor faults, the trigger node can be the moment when the motor current first experiences abnormal fluctuations, and the termination node can be the moment when the fault is confirmed or eliminated. The time period between these two nodes contains a series of state information related to the motor fault, such as speed changes, temperature changes, etc., which constitutes the knowledge point representation space. Integrate the trigger node after the last knowledge point representation space to generate the prior warning guidance knowledge data.

[0028] The first guiding knowledge vector of the priori warning guiding knowledge data is extracted. This vector contains key information in the priori warning guiding knowledge data, such as relevant parameters of the triggering node, state characteristic parameters in the knowledge point representation space, etc.

[0029] The second train state characteristic map is integrated with the first train state characteristic map to generate an integrated train state characteristic map, and the first guided knowledge vector and the integrated train state characteristic map are interactively focused on feature analysis. The second train state characteristic map that has a matching relationship with the first train state characteristic map of the current target train operation data is integrated. For example, if the first train state characteristic map shows that the current train speed fluctuates greatly and the motor current has small fluctuations, and the second train state characteristic map found is the map in the previous motor failure case, the integrated train state characteristic map obtained by integrating the two contains more information.

[0030] Interactive focus feature analysis is performed on the first sub-vector of the knowledge point representation space in the first guiding knowledge vector and the second train state feature map corresponding to the first sub-vector in the integrated train state feature map. Different types of feature elements are extracted from the first sub-vector according to the predefined feature classification rules, and the second train state feature map is divided into multiple feature subsets corresponding to the feature classification of the first sub-vector, generating the classified feature subsets of the first sub-vector and the corresponding feature subsets of the second train state feature map. For example, according to the classification rules such as motor-related features and speed-related features, feature elements such as the abnormal fluctuation amplitude and the abnormal fluctuation duration of the motor current are extracted from the first sub-vector, and the second train state feature map is divided into motor-related feature subsets (such as motor current change trend, motor temperature change, etc.) and speed-related feature subsets (such as speed fluctuation frequency, speed maximum value, etc.).

[0031] For each corresponding feature subset pair, a local feature association relationship is established. For each feature element in each feature subset, based on the logical relationship or potential physical relationship between the feature elements, the correlation between the feature element and other feature elements in the corresponding feature subset is calculated to construct the local feature association matrix of each feature subset pair. For example, in the motor-related feature subset pair, there may be a positive correlation between the abnormal fluctuation amplitude of the motor current and the change in the motor temperature, because the abnormal current may cause the motor to heat up. The local feature association matrix is ​​constructed by calculating this correlation.

[0032] Based on the correlation threshold, each local feature correlation matrix is ​​enhanced, and the enhanced local feature correlation matrix is ​​screened to remove the feature elements with zero correlation in the enhanced local feature correlation matrix. Assuming that the correlation threshold is set to 0.3, if the correlation between a feature element and other elements is lower than this threshold, it will be removed from the local feature correlation matrix.

[0033] Based on the filtered local feature association matrix, a cross-subset association relationship is constructed between different feature subset pairs to construct a cross-subset association matrix. For example, there may be an association between the motor-related feature subset and the speed-related feature subset. For example, abnormal motor current may cause unstable train speed. The cross-subset association matrix is ​​constructed by calculating the cross-subset association degree.

[0034] All cross-subset association matrices are integrated to construct a global association matrix, which reflects the overall association relationship between the first sub-vector and the second train state feature map. Based on the global association matrix, the association features are focused and analyzed to find the combination of feature elements whose association degree meets the preset conditions in the global association matrix to generate different focus areas. Each focus area represents an interaction mode, and the combination of feature elements represents the key interaction relationship between the first sub-vector and the second train state feature map. The relative association contribution of each feature element in the focus area is analyzed, and the feature elements that do not meet the preset conditions are removed according to the relative association contribution to generate the key feature set of each focus area after refinement. The key feature sets of all focus areas are summarized to form the interactive focus feature analysis results of the interactive focus feature analysis of the first sub-vector of the knowledge point representation space in the first guiding knowledge vector and the second train state feature map corresponding to the first sub-vector in the integrated train state feature map.

[0035] Interactive focus feature analysis is performed on the second sub-vector of the last trigger node in the first guided knowledge vector and the first train state feature map in the integrated train state feature map. This process is similar to the analysis of the first sub-vector of the knowledge point representation space, except that the objects of analysis are the second sub-vector of the last trigger node in the first guided knowledge vector and the first train state feature map.

[0036] Step S130, optimizing the interactive focusing feature data according to the first feature flow control parameter to generate a target guidance knowledge vector, wherein the first feature flow control parameter is generated by performing parameter learning after locking the model parameter information of the train fault warning decision model.

[0037] In this embodiment, the server optimizes the interactive focus feature data according to the first feature flow control parameter. The first feature flow control parameter is generated by locking the model parameter information of the train fault warning decision model and then performing parameter learning. For example, in the previous training process, these parameters are learned through a large amount of sample train operation data.

[0038] Assume that the server obtains a set of sample train operation data, which includes various situations of normal operation and fault operation. Lock the model parameter information of the train fault warning decision model, and initialize each first characteristic flow control parameter and each second characteristic flow control parameter as a parameter value. Then train each first characteristic flow control parameter and each second characteristic flow control parameter based on the sample train operation data. During the training process, adjust these parameters according to the actual situation in the sample train operation data (such as known fault occurrence and normal operation) so that the parameters can better meet the needs of train fault warning.

[0039] The interactive focus feature data is optimized according to the first feature flow control parameter, and the optimized interactive focus feature data is fused with the first guide knowledge vector to generate a second guide knowledge vector. For example, the importance or weight of some features in the interactive focus feature data may change after the optimization of the first feature flow control parameter. These optimized features are fused with the corresponding features in the first guide knowledge vector according to certain rules to generate a second guide knowledge vector.

[0040] The second guided knowledge vector is transformed, and the transformation result is optimized according to the second characteristic flow control parameter to generate a target guided knowledge vector. This transformation may be a mathematical transformation of the second guided knowledge vector, such as normalization or weighted transformation, and then the transformation result is further optimized by the second characteristic flow control parameter to finally obtain the target guided knowledge vector.

[0041] Step S140, making a fault warning decision according to the target guidance knowledge vector through the train fault warning decision model, and generating a train fault warning knowledge point of the target train operation data.

[0042] In this embodiment, the train fault warning decision model includes a plurality of sequentially connected feature propagation learning units. Feature propagation learning is performed on the target guidance knowledge vector according to the plurality of feature propagation learning units, and a fault warning decision is made according to the feature propagation learning result to generate train fault warning knowledge points of the target train operation data.

[0043] The target guided knowledge vector is the data to be learned by the first feature propagation learning unit. It is assumed that the function of the first feature propagation learning unit is to perform preliminary analysis and extraction of basic features in the target guided knowledge vector. For example, the train speed related features in the target guided knowledge vector can be further refined and analyzed to extract more microscopic features such as speed change patterns.

[0044] For each feature propagation learning unit except the last feature propagation learning unit, an interactive focusing feature analysis is performed on the feature propagation learning result of the current feature propagation learning unit and the first train state feature map, the interactive focusing feature data is optimized according to the first feature flow control parameter, the optimized interactive focusing feature data is fused with the feature propagation learning result of the current feature propagation learning unit, the fusion result is transformed, and after optimizing the transformed result according to the second feature flow control parameter, it is loaded into the next feature propagation learning unit for feature propagation learning.

[0045] For example, after receiving the processed result of the first feature propagation learning unit, the second feature propagation learning unit first performs interactive focus feature analysis with the first train state feature map. If the result of the first feature propagation learning unit shows that there is an abnormal fluctuation pattern in the train speed, then when performing interactive focus feature analysis with the first train state feature map, the position of the train speed in the entire train running state (such as whether it is in the acceleration stage, deceleration stage or uniform speed stage) and other related state dimensions (such as brake pressure, motor current, etc.) will be further combined. The interactive focus feature data is optimized according to the first feature flow control parameter, such as adjusting the weight of the abnormal speed fluctuation pattern and the features associated with other state dimensions. Then, the optimized interactive focus feature data is fused with the feature propagation learning result of the current feature propagation learning unit, for example, the adjusted abnormal speed fluctuation pattern and the features associated with other state dimensions are fused into the current feature propagation learning result. The fusion result is transformed, such as data compression or feature enhancement, and after optimizing the transformation result according to the second feature flow control parameter, it is loaded into the third feature propagation learning unit for feature propagation learning.

[0046] The last feature propagation learning unit makes the final fault warning decision based on the processing results of all previous units and generates train fault warning knowledge points for the target train operation data. This train fault warning knowledge point may include the type of fault (such as motor fault, brake system fault, etc.), the possibility of the fault (such as 80% probability of motor fault in the next 30 minutes), the scope of possible impact of the fault (such as affecting the speed control of the train, car temperature adjustment, etc.), and recommended measures for the fault (such as immediately checking a certain part of the motor, adjusting the brake pressure, etc.).

[0047] Based on the above steps, the embodiment of the present application achieves a comprehensive and accurate characterization of the train operation status by accurately extracting the multi-dimensional state path vectors in the target train operation data and fusing them to generate the first train state feature map. Combined with the prior warning guidance knowledge data, the present invention uses interactive focus feature analysis technology to effectively guide the train fault warning decision model to make accurate warning decisions on the target train operation data. By introducing the first characteristic flow control parameter to optimize the interactive focus feature data, a more targeted target guidance knowledge vector is generated, further improving the accuracy and reliability of fault warning. Ultimately, the present invention can generate train fault warning knowledge points in real time and remotely, providing a strong technical guarantee for the safe operation of trains, significantly reducing the incidence and impact range of train failures, and improving the safety and efficiency of railway transportation.

[0048] For example, in a possible implementation, step S110 includes:

[0049] Step S111, identifying and defining multiple state dimensions in the target train operation data, wherein the state dimensions include train speed, acceleration, brake pressure, motor current, car temperature, track condition, and external environmental conditions, and each state dimension reflects different aspects of the train operation.

[0050] Step S112, for each identified state dimension, extract the state path data of the state dimension changing with time from the target train operation data. Specifically, for each state dimension, the continuous measurement value of the state dimension within a set time is regarded as a state path, and the state path is composed of multiple data points arranged in chronological order, and each data point contains a timestamp and a corresponding state value.

[0051] Step S113, vectorize the state path data of each state dimension, convert the state path data into a state path vector, each state path vector is composed of the state values ​​of all data points on the state path in order, and retains the information of state changes over time.

[0052] Step S114, for each state path vector, calculate the mean and standard deviation of the state path vector, and subtract the mean from each element in the state path vector and divide it by the standard deviation to obtain a standardized state path vector.

[0053] Step S115, based on the standardized state path vectors of each state dimension, extract the dimensional features of each state dimension, the dimensional features include statistical features, time domain features, frequency domain features and features based on time series analysis, the statistical features include mean, variance, maximum value, minimum value, kurtosis, and skewness, the time domain features include zero crossing rate and mean crossing rate, the frequency domain features include spectral density and energy spectrum, and the features based on time series analysis include features of autocorrelation function analysis and features of cross-correlation function analysis.

[0054] Step S116, using a feature selection algorithm to evaluate the importance of each feature in the dimensional features of each state dimension, and after selecting a target feature set that is relevant to train fault warning, a linear discriminant analysis algorithm is applied to map the target feature set from a high-dimensional feature space to a low-dimensional space, and the feature vectors of each state dimension after dimensionality reduction are output.

[0055] Step S117, the feature vectors of each state dimension after dimensionality reduction are merged to generate a comprehensive feature representation of the target train operation data, and a weight is assigned to the feature vector of each state dimension.

[0056] Step S118, based on the fused comprehensive feature representation and its weight, construct a first train state feature map of the target train operation data, wherein the comprehensive feature representation is regarded as a node of the first train state feature map, the connection between the nodes represents the association relationship between the comprehensive feature representations, and the weight reflects the importance of the node in the first train state feature map.

[0057] In this embodiment, taking a long-distance passenger train as an example, the state dimension of train speed reflects the speed of the train operation, and its data is collected by high-precision speed sensors installed on the train axles or specific parts of the power system. The acceleration state dimension reflects the speed of the train speed change, and the data comes from a special acceleration sensor. The data of the brake pressure dimension is obtained from the pressure monitoring device of the train brake system, which is crucial for the safe braking of the train. The data of the motor current state dimension comes from the current monitoring device of the train motor. Whether the motor current is stable or not is directly related to the normal operation of the motor. The data of the carriage temperature dimension is collected by temperature sensors distributed at different positions in the carriage, which can reflect the environmental conditions in the carriage and the operating effects of equipment such as air conditioners. The data of the track condition dimension is obtained by the track detection instrument installed at the bottom of the train, such as the flatness of the track, the vibration frequency of the track, etc., which can reflect whether there are defects or foreign objects in the track. The data of the external environmental conditions dimension may include multiple sources, such as temperature, humidity, wind speed and other data collected by meteorological sensors outside the train. These external environmental factors will have many effects on the operation of the train.

[0058] For each identified state dimension, the server extracts the state path data from the target train operation data. For example, for the dimension of train speed, assuming that the set time is a run from the starting station to the terminal station, the total running time is 5 hours, and the speed data is collected every 30 seconds, 600 continuous speed measurements will be obtained. Each measurement value contains an accurate timestamp and the corresponding speed state value. These data points arranged in chronological order constitute the state path of the train speed state dimension. For state dimensions such as acceleration, brake pressure, motor current, car temperature, track conditions and external environmental conditions, the same principle is followed to obtain the corresponding state path data based on the collection frequency of each sensor and the set time range.

[0059] Subsequently, the server vectorizes the state path data of each state dimension. Taking the train speed as an example, the 600 speed data points obtained previously are arranged in sequence to form a state path vector. This vector completely retains the information of the train speed changing over time during the 5-hour operation. Similarly, the state path data of other state dimensions are also converted into their own state path vectors.

[0060] Next, for each state path vector, the server calculates its mean and standard deviation, and then performs normalization. For example, for the train speed state path vector, assume that the calculated mean is 180 km / h and the standard deviation is 10 km / h. For each speed value in the vector, such as a data point with a speed value of 195 km / h, after normalization, it becomes (195 - 180) / 10 = 1.5. The state path vectors of other state dimensions are also standardized using the same calculation method, which makes the data of different state dimensions have a unified dimension in subsequent processing, which is convenient for comparison and analysis.

[0061] Based on the standardized state path vectors of each state dimension, the server begins to extract the dimensional features of each state dimension. Taking the train speed state dimension as an example, in terms of statistical features, the mean of 180 km / h reflects the average speed of the train during the entire operation process; the variance reflects the discreteness of the speed. If the variance is large, it means that the train speed fluctuates greatly; the maximum value may appear at a certain moment during the acceleration of the train, such as 220 km / h; the minimum value may be reached when decelerating when entering the station, such as 120 km / h; kurtosis and skewness describe the distribution characteristics of speed data from a more complex perspective. The zero-crossing rate in the time domain feature, if the speed data frequently changes from increasing to decreasing or vice versa, the zero-crossing rate will be higher, and the mean crossing rate reflects the conversion frequency of speed data between different mean levels. In terms of frequency domain features, the spectrum density can show the energy distribution of speed data at different frequencies, and the energy spectrum describes the frequency domain characteristics of speed data from the perspective of energy. Among the features based on time series analysis, the features of autocorrelation function analysis can reveal the correlation of speed data at different time lags, such as the correlation between the speed at a certain moment and the speed 10 minutes later; the features of cross-correlation function analysis can be used to analyze the relationship between speed and other related variables (such as acceleration). For other state dimensions, such as brake pressure, the mean in the statistical features can reflect the average brake pressure during normal operation, and the variance can reflect the fluctuation of brake pressure, etc., and various dimensional features are extracted in a similar way.

[0062] Afterwards, the server uses a feature selection algorithm to evaluate the importance of each feature in the dimensional features of each state dimension. For example, in the dimensional features of train speed, after analysis by the feature selection algorithm, it is found that a certain feature of the variance and autocorrelation function analysis has a high relevance to train fault warning, while the kurtosis and skewness are relatively less important. After such an evaluation is performed for each state dimension, the target feature set that is relevant to the train fault warning is selected. Then the linear discriminant analysis algorithm is applied to map the target feature set from the high-dimensional feature space to the low-dimensional space. Assume that the train speed state dimension initially has 8 features that are considered to have a certain correlation, forming an 8-dimensional feature space. After being processed by the linear discriminant analysis algorithm, it is mapped to a 3-dimensional space, and the feature vector of the reduced train speed state dimension is output. Other state dimensions are also reduced in the same way.

[0063] The feature vectors of each state dimension after dimensionality reduction are then fused to generate a comprehensive feature representation of the target train operation data. For example, suppose the feature vector of train speed after dimensionality reduction, the feature vector of acceleration after dimensionality reduction, the feature vector of brake pressure after dimensionality reduction, etc. are combined according to certain rules. During the fusion process, a weight is assigned to the feature vector of each state dimension. For example, the weight of the train speed feature vector is 0.3, the weight of the acceleration feature vector is 0.2, the weight of the brake pressure feature vector is 0.25, etc. These weights are determined according to the relative importance of each state dimension in train fault warning.

[0064] Finally, based on the fused comprehensive feature representation and its weight, the server constructs the first train state feature map of the target train operation data. The comprehensive feature representation is regarded as a node of the first train state feature map. For example, the comprehensive feature representation after the fusion of state dimensions such as train speed and acceleration is a node. The connection between nodes represents the association between the comprehensive feature representations. For example, there is a correlation between train speed and brake pressure. Because the change of train speed will affect the demand for brake pressure, there will be a connection between these two nodes in the first train state feature map. The weight reflects the importance of the node in the first train state feature map. Nodes with higher weights have greater influence in the map and will be given priority consideration in the subsequent analysis and decision-making process. In this way, the server constructs the first train state feature map that comprehensively reflects the operating status of the target train, providing an important basic data structure for subsequent train fault warning decisions.

[0065] In a possible implementation, step S130 includes:

[0066] Step S131, optimizing the interactive focusing feature data according to the first characteristic flow control parameter, and fusing the optimized interactive focusing feature data with the first guided knowledge vector to generate a second guided knowledge vector.

[0067] Step S132, transform the second guiding knowledge vector, optimize the transformation result according to the second characteristic flow control parameter, and generate a target guiding knowledge vector, wherein the second characteristic flow control parameter is generated by locking the model parameter information of the train fault warning decision model and then performing parameter learning.

[0068] In a possible implementation, the train fault warning decision model includes a plurality of sequentially connected feature propagation learning units, and step S140 includes:

[0069] The target guidance knowledge vector is subjected to feature propagation learning according to the plurality of feature propagation learning units, a fault warning decision is made according to the feature propagation learning result, and a train fault warning knowledge point of the target train operation data is generated.

[0070] The target guided knowledge vector is the data to be learned of the first feature propagation learning unit.

[0071] For each of the feature propagation learning units except the last one, an interactive focusing feature analysis is performed on the feature propagation learning result of the current feature propagation learning unit and the first train state feature map, the interactive focusing feature data is optimized according to the first feature flow control parameter, the optimized interactive focusing feature data is fused with the feature propagation learning result of the current feature propagation learning unit, the fused result is transformed, and after optimizing the transformed result according to the second feature flow control parameter, it is loaded into the next feature propagation learning unit for feature propagation learning.

[0072] In a possible implementation manner, each of the feature propagation learning units is respectively configured with the first feature flow control parameter and the second feature flow control parameter, and before optimizing the interactive focus feature data according to the first feature flow control parameter, the method further includes:

[0073] Step A110, obtaining sample train operation data, locking the model parameter information of the train fault warning decision model, and initializing each of the first characteristic flow control parameters and each of the second characteristic flow control parameters as parameter values.

[0074] Step A120: training each of the first characteristic flow control parameters and each of the second characteristic flow control parameters according to the sample train operation data.

[0075] In this embodiment, taking the operation data processing of a certain high-speed train as an example, the interactive focus feature data includes a lot of feature information previously obtained by analyzing the train state feature map and the guided knowledge vector, which reflects the correlation between the train operation state and the prior knowledge. The first feature flow control parameter is a parameter obtained through a specific learning process, and its optimization of the interactive focus feature data is based on a pre-set mathematical model and algorithm rules. For example, the first feature flow control parameter may specify the weight adjustment method or data screening rules for certain features. The server processes the interactive focus feature data according to these rules, highlights important features, suppresses noise or irrelevant features, and makes the interactive focus feature data more in line with the requirements of subsequent processing.

[0076] Next, the server fuses the optimized interactive focus feature data with the first guidance knowledge vector to generate a second guidance knowledge vector. The first guidance knowledge vector contains key knowledge information extracted from the prior warning guidance knowledge data, such as the train state characteristics when the prior fault occurs, the correlation between the fault type and the train state, etc. During the fusion process, the server integrates the relevant features in the optimized interactive focus feature data with the corresponding features in the first guidance knowledge vector according to the set fusion algorithm. For example, if the train speed fluctuation feature in the interactive focus feature data is associated with the speed-related fault knowledge in the first guidance knowledge vector, the server will fuse the two according to a certain weight, and integrate the new information of the speed fluctuation feature into the speed-related knowledge of the first guidance knowledge vector, thereby generating a second guidance knowledge vector. This second guidance knowledge vector contains both the core content of the prior knowledge and the new feature information obtained from the interactive focus analysis under the current train operation status.

[0077] Then, the server transforms the second guided knowledge vector. This transformation may be based on a specific mathematical transformation method, such as linear transformation or nonlinear transformation. Taking linear transformation as an example, the second guided knowledge vector can be multiplied according to a pre-set transformation matrix to change the values ​​of each element in the vector and the relationship between them. After the transformation, the transformation result is optimized according to the second characteristic flow control parameter to generate a target guided knowledge vector. The second characteristic flow control parameter is also generated by locking the model parameter information of the train fault warning decision model and then performing parameter learning. For example, the second characteristic flow control parameter may specify rules for feature selection or feature enhancement of the transformed knowledge vector. The server optimizes the transformation result according to these rules, removes possible redundant information, enhances the expressiveness of key features, and finally obtains the target guided knowledge vector.

[0078] In the case where the train fault warning decision model includes multiple sequentially connected feature propagation learning units, the server makes fault warning decisions based on the target guidance knowledge vector through the model to generate train fault warning knowledge points for the target train operation data. The target guidance knowledge vector is input into the model as the data to be learned by the first feature propagation learning unit. For example, the function of the first feature propagation learning unit may be to perform preliminary feature extraction and classification on the target guidance knowledge vector. For example, according to the predefined feature classification rules, the train speed-related features, brake pressure-related features, etc. in the target guidance knowledge vector can be classified and extracted, and the preliminary relationship between these features can be analyzed at the same time. For example, the coupling relationship between speed and brake pressure may be preliminarily quantified at this stage.

[0079] For each feature propagation learning unit except the last feature propagation learning unit, the server has to perform a series of complex operations. Taking the second feature propagation learning unit as an example, it will perform interactive focus feature analysis on the feature propagation learning results of the current feature propagation learning unit (i.e., itself) and the first train state feature map. The first train state feature map contains comprehensive information of various state dimensions of the train, such as the fusion features of state dimensions such as speed, acceleration, and brake pressure, and the correlation between them. During the interactive focus feature analysis process, the server will find the correlation features between the feature propagation learning results and the first train state feature map according to the set algorithm. For example, if the feature propagation learning results show that there is an abnormal fluctuation pattern in the train speed, when interactively focusing the analysis with the first train state feature map, the server will further combine the position of the train speed in the entire train operation state (such as in the acceleration stage or the deceleration stage) and other related state dimensions (such as brake pressure, motor current, etc.) to determine the deep relationship between this abnormal speed fluctuation and other state dimensions.

[0080] Afterwards, the server optimizes the interactive focus feature data according to the first feature flow control parameter. For example, the first feature flow control parameter may specify how to adjust the weights of the features associated with the abnormal speed fluctuation pattern and other state dimensions in this case, and the server optimizes the interactive focus feature data according to this rule. Then, the optimized interactive focus feature data is fused with the feature propagation learning results of the current feature propagation learning unit. For example, the adjusted abnormal speed fluctuation pattern and the features associated with other state dimensions are fused into the current feature propagation learning results, and the content of the feature propagation learning results is updated. The fusion result is then transformed, and this transformation may be to adapt to the input requirements of the next feature propagation learning unit, such as data compression or feature enhancement. Finally, after optimizing the transformation result according to the second feature flow control parameter, it is loaded into the next feature propagation learning unit for feature propagation learning.

[0081] Before the whole process, the server needs to obtain the sample train operation data, lock the model parameter information of the train fault warning decision model, and initialize each first characteristic flow control parameter and each second characteristic flow control parameter as a parameter value. The sample train operation data contains a large number of train operation instances, including both normal operation and operation data under various fault conditions. For example, these sample data may come from actual records of trains of different lines and different models during many years of operation. The server locks the model parameter information of the train fault warning decision model to ensure that the model structure remains unchanged during the parameter learning process. Initialize the first characteristic flow control parameter and the second characteristic flow control parameter to a preset initial value, such as initializing some weights of the first characteristic flow control parameter to 0.5, and initializing a scaling factor of the second characteristic flow control parameter to 1, etc.

[0082] Then, the server trains each first characteristic flow control parameter and each second characteristic flow control parameter based on the sample train operation data. During the training process, the server inputs the sample train operation data into the train fault warning decision model, and adjusts the first characteristic flow control parameter and the second characteristic flow control parameter according to the difference between the output result of the model and the actual fault condition in the sample data (known fault occurrence and normal operation). For example, if the fault result predicted by the model on a certain sample data does not match the actual fault condition, the server will analyze which characteristic flow control parameters cause this error according to the predefined error back propagation algorithm, and then adjust the values ​​of these parameters accordingly. Through repeated training of a large number of sample train operation data, the first characteristic flow control parameter and the second characteristic flow control parameter gradually converge to appropriate values, so that the train fault warning decision model can make fault warning decisions more accurately. Finally, after being processed by multiple feature propagation learning units, the server generates train fault warning knowledge points for the target train operation data based on the final results of the model. This train fault warning knowledge point contains complex information, such as the type of fault (it may be a motor fault, brake system fault, etc.), the possibility of a fault occurring (such as an 80% chance of a certain fault occurring within the next 30 minutes), the scope of possible impact of the fault (such as affecting the speed control of the train, car temperature regulation, etc.) and recommended measures for the fault (such as immediately checking a certain component of the motor, adjusting the brake pressure, and other operational suggestions) and other detailed content.

[0083] In a possible implementation, step S120 includes:

[0084] Step S121, searching for a second train state characteristic map having a characteristic pattern matching the first train state characteristic map, and a priori fault warning knowledge points matching the second train state characteristic map, wherein the a priori fault warning knowledge points are fault warning knowledge points of a priori train operation data, and the second train state characteristic map is generated by fusing state path vectors of multiple state dimensions in the a priori train operation data, and the dimension of the state path vector fused by the second train state characteristic map is the same as the dimension of the state path vector fused by the first train state characteristic map.

[0085] Step S122: constructing priori warning guidance knowledge data according to the priori fault warning knowledge points, and extracting a first guidance knowledge vector of the priori warning guidance knowledge data.

[0086] Step S123: Integrate the second train state characteristic map with the first train state characteristic map to generate an integrated train state characteristic map, and perform interactive focusing feature analysis on the first guided knowledge vector and the integrated train state characteristic map.

[0087] In this embodiment, taking the trains in a railway system as an example, the server stores a large amount of prior train operation data, which contains various state information of different trains in the past operation process. For each set of prior train operation data, the server will fuse the state path vectors of multiple state dimensions (such as train speed, acceleration, brake pressure, motor current, car temperature, track conditions, external environmental conditions, etc.) to generate the second train state characteristic map of each prior train operation data. This process is similar to the way of generating the first train state characteristic map before, ensuring that the dimensions of the state path vectors fused by the two are the same. For example, when processing the prior train operation data that has had a brake failure, the server fuses the state path vectors of the state dimensions such as brake pressure, train speed, motor current, etc. in a specific time period according to the prescribed algorithm to obtain the corresponding second train state characteristic map.

[0088] When searching for feature pattern matches, the server will compare each feature in the first train state feature map with many stored second train state feature maps. For example, if the first train state feature map shows that the motor current of the current train has a small abnormal fluctuation and the train speed fluctuation frequency increases, the server will search for a second train state feature map with similar motor current and train speed feature patterns in the stored prior data. At the same time, the server will also search for prior fault warning knowledge points that match the second train state feature map. These prior fault warning knowledge points contain fault warning related information in the prior train operation data, such as fault type, possible time range of fault occurrence, and other complex contents.

[0089] Next, construct the prior warning guidance knowledge data based on the found prior fault warning knowledge points. For example, if the prior fault warning knowledge point indicates that a motor failure may occur within 10-15 minutes after the train speed fluctuates and the motor current is abnormal, the server will construct the prior warning guidance knowledge data with this knowledge point as the core. The specific operation is to integrate the trigger node and the termination node in the forward and backward nodes of the prior fault warning knowledge point respectively, and generate the knowledge point representation space corresponding to the prior fault warning knowledge point. The trigger node can be the moment when the motor current first fluctuates abnormally, and the termination node can be the moment when the fault is confirmed or eliminated. Between these two nodes, a series of status information related to the motor fault is contained, such as speed changes, temperature changes, etc. Then integrate the trigger node after the last knowledge point representation space to generate the prior warning guidance knowledge data.

[0090] After that, the server extracts the first guiding knowledge vector of the prior warning guiding knowledge data. This vector contains key information in the prior warning guiding knowledge data, such as relevant parameters of the trigger node, state characteristic parameters in the knowledge point representation space, etc. For example, the amplitude value and duration of the abnormal fluctuation of the motor current corresponding to the trigger node, the fluctuation range of the train speed in the knowledge point representation space, and the change value of the carriage temperature will all be extracted into the first guiding knowledge vector.

[0091] Finally, the server integrates the second train state characteristic map with the first train state characteristic map to generate an integrated train state characteristic map, and performs interactive focus feature analysis on the first guided knowledge vector and the integrated train state characteristic map. For example, the second train state characteristic map that has a matching relationship with the current train state (such as the map related to the brake fault mentioned above) is integrated with the first train state characteristic map to obtain an integrated train state characteristic map, which contains more comprehensive information about the train state.

[0092] When performing interactive focus feature analysis, the server analyzes the first sub-vector of the knowledge point representation space in the first guiding knowledge vector and the second train state feature map corresponding to the first sub-vector in the integrated train state feature map. Different types of feature elements are extracted from the first sub-vector according to the predefined feature classification rules, and the second train state feature map is divided into multiple feature subsets corresponding to the first sub-vector feature classification, generating the classified feature subsets of the first sub-vector and the corresponding feature subsets of the second train state feature map. For example, according to the classification rules such as motor-related features and speed-related features, feature elements such as the abnormal fluctuation amplitude of the motor current and the duration of the abnormal fluctuation are extracted from the first sub-vector, and the second train state feature map is divided into motor-related feature subsets (such as motor current change trend, motor temperature change, etc.) and speed-related feature subsets (such as speed fluctuation frequency, maximum speed, etc.).

[0093] For each corresponding feature subset pair, the server establishes a local feature association relationship. For each feature element in each feature subset, based on the logical relationship or potential physical relationship between the feature elements, the correlation between the feature element and other feature elements in the corresponding feature subset is calculated to construct a local feature association matrix for each feature subset pair. For example, in the motor-related feature subset pair, there may be a positive correlation between the abnormal fluctuation amplitude of the motor current and the change in the motor temperature, because the abnormal current may cause the motor to heat up. The local feature association matrix is ​​constructed by calculating this correlation.

[0094] Based on the correlation threshold, each local feature correlation matrix is ​​enhanced, and the enhanced local feature correlation matrix is ​​screened to remove the feature elements with zero correlation in the enhanced local feature correlation matrix. Assuming that the correlation threshold is set to 0.3, if the correlation between a feature element and other elements is lower than this threshold, it will be removed from the local feature correlation matrix.

[0095] Based on the filtered local feature association matrix, a cross-subset association relationship is constructed between different feature subset pairs to construct a cross-subset association matrix. For example, there may be an association between the motor-related feature subset and the speed-related feature subset. For example, abnormal motor current may cause unstable train speed. The cross-subset association matrix is ​​constructed by calculating the cross-subset association degree.

[0096] All cross-subset association matrices are integrated to construct a global association matrix, which reflects the overall association relationship between the first sub-vector and the second train state feature map. Based on the global association matrix, the association features are focused and analyzed to find the combination of feature elements whose association degree meets the preset conditions in the global association matrix to generate different focus areas. Each focus area represents an interaction mode, and the combination of feature elements represents the key interaction relationship between the first sub-vector and the second train state feature map. The relative association contribution of each feature element in the focus area is analyzed, and the feature elements that do not meet the preset conditions are removed according to the relative association contribution to generate the key feature set of each focus area after refinement. The key feature sets of all focus areas are summarized to form the interactive focus feature analysis results of the interactive focus feature analysis of the first sub-vector of the knowledge point representation space in the first guiding knowledge vector and the second train state feature map corresponding to the first sub-vector in the integrated train state feature map.

[0097] The server also performs interactive focus feature analysis on the second sub-vector of the last trigger node in the first guide knowledge vector and the first train state feature map in the integrated train state feature map. This process is similar to the analysis of the first sub-vector of the knowledge point representation space, except that the objects of analysis are the second sub-vector of the last trigger node in the first guide knowledge vector and the first train state feature map. Through these steps, the server can comprehensively and deeply analyze the relationship between prior knowledge and the current train state, providing an important basis for subsequent train fault warning decisions.

[0098] In a possible implementation, step S122 includes:

[0099] Step S1221 , integrating trigger nodes and termination nodes in the forward and backward nodes of the priori fault warning knowledge point respectively, and generating a knowledge point representation space corresponding to the priori fault warning knowledge point.

[0100] Step S1222, integrating the trigger node after the last knowledge point representation space to generate priori warning guidance knowledge data.

[0101] In a possible implementation, step S123 includes:

[0102] Step S1231, performing interactive focusing feature analysis on a first sub-vector of the knowledge point representation space in the first guiding knowledge vector and a second train state feature map corresponding to the first sub-vector in the integrated train state feature map.

[0103] Step S1232: performing interactive focus feature analysis on the second sub-vector of the last trigger node in the first guiding knowledge vector and the first train state feature map in the integrated train state feature map.

[0104] In a possible implementation, step S1231 includes:

[0105] Step S1231-1, extract different types of feature elements from the first sub-vector according to predefined feature classification rules, and divide the second train state feature map into multiple feature subsets corresponding to the feature classification of the first sub-vector, generating a feature subset of the classified first sub-vector and a corresponding feature subset of the second train state feature map.

[0106] Step S1231-2, for each corresponding feature subset pair, a local feature association relationship is established, wherein, for each feature element in each feature subset, based on the logical relationship or potential physical relationship between the feature elements, the association degree between the feature element and other feature elements in the corresponding feature subset is calculated to construct a local feature association matrix for each feature subset pair, wherein the elements in the local feature association matrix represent the association degrees between the corresponding feature elements.

[0107] Step S1231-3, based on the correlation threshold, each local feature correlation matrix is ​​enhanced, and the enhanced local feature correlation matrix is ​​screened to remove feature elements with zero correlation in the enhanced local feature correlation matrix.

[0108] Step S1231-4, based on the screened local feature association matrix, construct a cross-subset association relationship between different feature subset pairs to construct a cross-subset association matrix, wherein the elements in the cross-subset association matrix represent the association degree of feature elements between different feature subsets.

[0109] Step S1231-5, integrating all cross-subset correlation matrices to construct a global correlation matrix, wherein the global correlation matrix reflects the overall correlation relationship between the first sub-vector and the second train state characteristic map.

[0110] Step S1231-6, based on the global correlation matrix, focus analysis is performed on the correlation features to find out the characteristic element combination whose correlation degree meets the preset conditions in the global correlation matrix, so as to generate different focus areas, each focus area represents an interaction mode, and the characteristic element combination represents the key interaction relationship between the first sub-vector and the second train state characteristic map.

[0111] Step S1231-7, analyzing the relative correlation contribution of each feature element in the focus area, and removing feature elements that do not meet preset conditions according to the relative correlation contribution, to generate a refined key feature set for each focus area.

[0112] Step S1231-8, summarizes the key feature sets of all focus areas to form an interactive focus feature analysis result of interactive focus feature analysis on the first sub-vector of the knowledge point representation space in the first guiding knowledge vector and the second train state feature map corresponding to the first sub-vector in the integrated train state feature map.

[0113] Continuing with the example of a train in a railway operation system, assume that the prior fault warning knowledge point is the warning information about the train motor failure.

[0114] The trigger node and the termination node are integrated in the forward and backward nodes of this priori fault warning knowledge point to generate the knowledge point representation space corresponding to the priori fault warning knowledge point. For example, the priori fault warning knowledge point indicates that the motor failure may occur within a period of time after the abnormal fluctuation of the train motor current. The trigger node here can be set to the moment when the motor current first fluctuates abnormally and reaches a certain amplitude (such as exceeding the normal range by 10%), which marks the beginning of the abnormal situation that may cause the failure. The termination node can be set to the moment when the motor fault is confirmed (the specific fault code of the motor is detected by the train's fault diagnosis system) or the fault is eliminated (the maintenance personnel repair the motor fault and the train resumes normal operation). The time period between these two nodes contains a series of state information related to the motor fault, such as the fluctuation of the train speed during this period, because the motor failure may cause insufficient power of the train and thus affect the speed; the change of the motor temperature, the motor failure may be accompanied by abnormal temperature increase; the adjustment of the brake pressure, in order to cope with the possible impact of the motor failure on the train speed, the brake system may be adjusted accordingly, etc. These information together constitute the knowledge point representation space corresponding to the priori fault warning knowledge point.

[0115] Then, after the last knowledge point representation space, the trigger node is integrated to generate prior warning guidance knowledge data. The purpose of this is to form a complete data structure that can be used for subsequent analysis. For example, when a similar abnormal motor current fluctuation (trigger node) occurs again, the prior warning guidance knowledge data constructed previously can be referred to for fault warning and decision-making.

[0116] The interactive focused feature analysis of the first guided knowledge vector and the integrated train state feature map includes multiple detailed steps.

[0117] First, interactive focusing feature analysis is performed on the first sub-vector of the knowledge point representation space in the first guided knowledge vector and the second train state feature map corresponding to the first sub-vector in the integrated train state feature map. Taking the situation related to train motor failure as an example, different types of feature elements are extracted from the first sub-vector according to the predefined feature classification rules. Assume that the predefined feature classification rules are classified according to different systems of the train, such as power system related features, brake system related features, environmental impact related features, etc. The power system related feature elements extracted from the first sub-vector may include the amplitude and frequency of abnormal fluctuations in motor current; the brake system related feature elements may include the number of adjustments to the brake pressure during motor failure; the environmental impact related feature elements may include the degree of influence of the external ambient temperature on the motor heat dissipation at that time.

[0118] At the same time, the second train state feature map is divided into multiple feature subsets corresponding to the first sub-vector feature classification. The feature subset related to the power system includes features such as the motor current change trend and the motor temperature change curve; the feature subset related to the braking system includes features such as the historical change data of the braking pressure and the response time of the braking system; the feature subset related to environmental impact includes data such as the external environment temperature and humidity and features such as their interaction with the internal equipment of the train. In this way, the feature subset of the classified first sub-vector and the corresponding feature subset of the second train state feature map are generated.

[0119] For each corresponding feature subset pair, a local feature association relationship is established. For example, for each feature subset pair related to the power system, for each feature element in each feature subset, based on the logical relationship or potential physical relationship between the feature elements, the correlation between the feature element and other feature elements in the corresponding feature subset is calculated to construct the local feature association matrix of each feature subset pair. In the power system, there is a logical relationship between the abnormal fluctuation amplitude of the motor current and the change in motor temperature, because the abnormal current may cause the motor to heat up. Assuming that the abnormal fluctuation amplitude of the motor current is A and the change in motor temperature is B, the correlation between them is calculated to be C through a specific algorithm (such as calculation based on a physical model or statistical analysis based on historical data). This C is an element in the local feature association matrix, indicating the correlation between the two feature elements. Similarly, the correlation between the abnormal fluctuation amplitude of the motor current and the motor torque output, the correlation between the motor temperature change and the motor torque output, etc. are calculated to construct a complete local feature association matrix of the feature subset pairs related to the power system.

[0120] Based on the correlation threshold, each local feature correlation matrix is ​​enhanced, and the enhanced local feature correlation matrix is ​​screened to remove the feature elements with zero correlation in the enhanced local feature correlation matrix. Assuming that the correlation threshold is set to 0.3, if the correlation between a feature element and other elements in the local feature correlation matrix of the power system related feature subset pair is lower than this threshold, it means that this feature element has almost no correlation with other elements in the feature subset, and it is removed from the local feature correlation matrix. For example, the operating status feature element of an auxiliary device related to a motor fault is removed after calculation, and the correlation with other power system related feature elements is lower than 0.3. This simplifies the local feature correlation matrix and highlights the feature elements with a higher correlation with the fault.

[0121] Based on the filtered local feature association matrix, a cross-subset association relationship is constructed between different feature subset pairs to construct a cross-subset association matrix. For example, there is a cross-subset association relationship between the power system feature subset and the brake system feature subset. Because motor failure (power system problem) may cause a change in train speed, thereby causing a response from the brake system. By calculating the correlation between the abnormal fluctuation amplitude of the motor current in the power system feature subset and the brake pressure adjustment amplitude in the brake system feature subset, as well as the correlation between other similar cross-subset feature elements, a cross-subset association matrix is ​​constructed. The elements in this matrix represent the correlation between feature elements in different feature subsets.

[0122] All cross-subset correlation matrices are integrated to construct a global correlation matrix. This global correlation matrix reflects the overall correlation between the first sub-vector and the second train state feature map. It integrates the information of each local feature correlation matrix and the cross-subset correlation matrix, covering the complex relationship between different train systems and between prior knowledge and the current train state.

[0123] Based on the global correlation matrix, the correlation features are focused and analyzed to find the characteristic element combinations whose correlations meet the preset conditions in the global correlation matrix, so as to generate different focus areas. Each focus area represents an interaction mode, and the characteristic element combination represents the key interaction relationship between the first sub-vector and the second train state characteristic map. For example, the preset condition is a characteristic element combination with a correlation greater than 0.5. If it is found that the correlations between the three characteristic elements of the abnormal fluctuation amplitude of the motor current, the increase amplitude of the motor temperature, and the adjustment amplitude of the brake pressure are all greater than 0.5 in the global correlation matrix, then these three characteristic elements constitute a focus area, which represents an interaction mode between the power system and the brake system in the event of a motor failure.

[0124] Analyze the relative correlation contribution of each characteristic element in the focus area, and remove the characteristic elements that do not meet the preset conditions according to the relative correlation contribution, and generate a refined key feature set for each focus area. In the above focus area, the relative correlation contribution of the three characteristic elements of motor current abnormal fluctuation amplitude, motor temperature increase amplitude and brake pressure adjustment amplitude is analyzed by a specific algorithm. Assume that after calculation, the relative correlation contribution of the abnormal motor current fluctuation amplitude is 0.6, the relative correlation contribution of the motor temperature increase amplitude is 0.3, and the relative correlation contribution of the brake pressure adjustment amplitude is 0.1. If the preset condition is that the relative correlation contribution is greater than 0.2, then the characteristic element of the brake pressure adjustment amplitude does not meet the condition, and it is removed from the focus area, and the key feature set of the refined focus area is obtained as the abnormal motor current fluctuation amplitude and the motor temperature increase amplitude.

[0125] The key feature sets of all focus areas are summarized to form the interactive focus feature analysis results of the first sub-vector of the knowledge point representation space in the first guiding knowledge vector and the second train state feature map corresponding to the first sub-vector in the integrated train state feature map. This result contains the key feature information that has been screened and refined layer by layer, which can accurately reflect the key interactive relationship between prior knowledge and the current train state, and provide an important basis for subsequent train fault warning decisions.

[0126] In addition, the server will also perform interactive focus feature analysis on the second sub-vector of the last trigger node in the first guide knowledge vector and the first train state feature map in the integrated train state feature map. This process is similar to the analysis of the first sub-vector of the knowledge point representation space mentioned above. For example, for the second sub-vector of the last trigger node, if it contains relevant feature elements of the train at a certain critical state transition moment (such as the critical moment from normal operation to the occurrence of a fault), such as the sudden change value of the motor current, the instantaneous change of the train speed, etc. These feature elements are also classified according to the predefined feature classification rules, and the first train state feature map in the integrated train state feature map is divided into corresponding feature subsets, and then the steps of establishing local feature association relationships, processing based on association thresholds, constructing cross-subset association relationships, constructing global association matrices, focusing analysis, and refining key feature sets are performed to finally obtain the interactive focus feature analysis results of this part. These results are combined with the previous analysis results of the first sub-vector of the knowledge point representation space, and can more comprehensively perform interactive focus feature analysis on the first guide knowledge vector and the integrated train state feature map, thereby providing a richer and more accurate basis for train fault warning decision-making.

[0127] In a possible implementation, step S121 includes:

[0128] Step S12111, obtain multiple prior train operation data and prior fault warning knowledge points of each prior train operation data, fuse the state path vectors of multiple state dimensions in each prior train operation data, and generate the second train state characteristic map of each prior train operation data.

[0129] Step S12112, integrate each of the second train state characteristic maps and the corresponding prior fault warning knowledge points and add them to the index database, and search the index database for the second train state characteristic map that has a characteristic pattern matching the first train state characteristic map, and the prior fault warning knowledge points that match the second train state characteristic map.

[0130] In this embodiment, first, the server obtains multiple prior train operation data and prior fault warning knowledge points of each prior train operation data. These prior train operation data are derived from the operation records of previous trains in the railway system, including various types of trains, operation lines, and different operation environments. The prior fault warning knowledge points are summaries and warning information for the fault conditions that occurred in these prior train operation data. For example, for a train that has been running on a mountain line, the prior train operation data records the data of multiple state dimensions such as train speed, acceleration, brake pressure, motor current, car temperature, track conditions, and external environmental conditions during the train's travel. The corresponding prior fault warning knowledge points may be that under specific track conditions (such as slight deformation of the track) and external environmental conditions (such as low temperature and humidity), the train's braking system may have a fault warning of braking delay, including detailed information such as the time range in which the fault may occur and the train operation conditions that may be affected.

[0131] Next, for each set of a priori train operation data, the server fuses the state path vectors of multiple state dimensions, thereby generating a second train state characteristic map of each a priori train operation data. Taking the mountain line train mentioned earlier as an example, for the state dimension of train speed, the server will integrate the speed data of the train during the entire operation process in a certain time sequence to form a state path vector of the train speed. Similarly, for other state dimensions such as acceleration, brake pressure, and motor current, their own state path vectors will be formed respectively. Then, through a specific fusion algorithm, the state path vectors of these different state dimensions are fused together to generate a second train state characteristic map of the a priori train operation data. This map can comprehensively reflect the comprehensive relationship between the various state dimensions of the train under the current operating conditions.

[0132] Afterwards, the server integrates each second train state characteristic map with the corresponding prior fault warning knowledge point and adds it to the index database. In this index database, each second train state characteristic map is closely linked to the corresponding prior fault warning knowledge point. For example, in the index database, a specific second train state characteristic map is used as the index key, and the corresponding prior fault warning knowledge point (such as the previously mentioned braking delay fault warning information) is stored as the associated value. In this way, when the server needs to find the second train state characteristic map that matches the characteristic pattern of the first train state characteristic map and the related prior fault warning knowledge point, it can easily search in this index database.

[0133] Finally, the server searches the index database for a second train state characteristic map that has a characteristic pattern matching with the first train state characteristic map, and a priori fault warning knowledge point that matches the second train state characteristic map. Taking a train currently being analyzed as an example, its first train state characteristic map contains comprehensive information of the train's current train speed, acceleration, brake pressure and other state dimensions. The server will compare this first train state characteristic map with each second train state characteristic map in the index database one by one. During the comparison process, the server will make a detailed comparison of the features of each state dimension according to the pre-set characteristic pattern matching rules. For example, if the train speed fluctuation in the first train state characteristic map is similar to the train speed fluctuation in a second train state characteristic map in the index database, and there are similar characteristic patterns in other state dimensions such as acceleration and brake pressure, then it is considered that a second train state characteristic map that has a characteristic pattern matching with the first train state characteristic map has been found. Once a matching second train state characteristic map is found, the server can directly obtain the matching priori fault warning knowledge point. These priori fault warning knowledge points contain detailed warning information for possible faults, such as the fault type (brake system fault, motor fault or other type of fault), the probability of the fault (such as the percentage of the probability of the fault), the scope of the possible impact of the fault (such as whether it will affect the speed control of the train, the temperature adjustment of the carriage, etc.), and the recommended measures for the fault (such as whether it is necessary to immediately check a certain device, adjust the parameters of a certain system, etc.). Through this search process, the server can use prior knowledge to provide valuable reference for the fault warning of the current train.

[0134] Figure 2 The hardware structure of the artificial intelligence-based train fault remote warning system 100 for implementing the artificial intelligence-based train fault remote warning method provided in an embodiment of the present invention is shown as follows: Figure 2As shown, the artificial intelligence-based train fault remote warning system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0135] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the artificial intelligence-based train fault remote warning system 100 uses to execute or use to complete the exemplary method described in the present invention.

[0136] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the artificial intelligence-based train fault remote warning method of the above method embodiment. 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 sending and receiving actions of the communication unit 140.

[0137] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned artificial intelligence-based train fault remote warning system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0138] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above-mentioned artificial intelligence-based train fault remote warning method is implemented.

[0139] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A train fault remote warning method based on artificial intelligence, characterized in that: The method comprises: Acquire target train operation data, extract state path vectors of multiple state dimensions from the target train operation data, and fuse the state path vectors of the multiple state dimensions to generate a first train state characteristic map of the target train operation data; Acquire priori warning guidance knowledge data, extract a first guidance knowledge vector of the priori warning guidance knowledge data, and perform interactive focus feature analysis on the first guidance knowledge vector and the first train state feature map, wherein the priori warning guidance knowledge data is used to guide a train fault warning decision model to make a train fault warning decision on the target train operation data; Optimizing the interactive focusing feature data according to a first feature flow control parameter to generate a target guidance knowledge vector, wherein the first feature flow control parameter is generated by performing parameter learning after locking the model parameter information of the train fault early warning decision model; The train fault warning decision model is used to make a fault warning decision based on the target guidance knowledge vector to generate train fault warning knowledge points for the target train operation data.

2. The train fault remote warning method based on artificial intelligence according to claim 1 is characterized in that: The step of optimizing the interactive focusing feature data according to the first feature flow control parameter to generate a target guidance knowledge vector includes: Optimizing the interactive focusing feature data according to the first characteristic flow control parameter, and fusing the optimized interactive focusing feature data with the first guiding knowledge vector to generate a second guiding knowledge vector; The second guiding knowledge vector is transformed, and the transformation result is optimized according to the second characteristic flow control parameter to generate a target guiding knowledge vector, wherein the second characteristic flow control parameter is generated by locking the model parameter information of the train fault warning decision model and then performing parameter learning.

3. The train fault remote warning method based on artificial intelligence according to claim 2 is characterized in that: The train fault warning decision model includes a plurality of sequentially connected feature propagation learning units, and the train fault warning decision model performs fault warning decision according to the target guided knowledge vector to generate train fault warning knowledge points of the target train operation data, including: Performing feature propagation learning on the target guidance knowledge vector according to the plurality of feature propagation learning units, making a fault warning decision according to the feature propagation learning result, and generating train fault warning knowledge points of the target train operation data; Wherein, the target guided knowledge vector is the data to be learned of the first feature propagation learning unit; For each of the feature propagation learning units except the last one, an interactive focusing feature analysis is performed on the feature propagation learning result of the current feature propagation learning unit and the first train state feature map, the interactive focusing feature data is optimized according to the first feature flow control parameter, the optimized interactive focusing feature data is fused with the feature propagation learning result of the current feature propagation learning unit, the fused result is transformed, and after optimizing the transformed result according to the second feature flow control parameter, it is loaded into the next feature propagation learning unit for feature propagation learning.

4. The train fault remote warning method based on artificial intelligence according to claim 3 is characterized in that: Each of the feature propagation learning units is respectively configured with the first feature flow control parameter and the second feature flow control parameter. Before optimizing the interactive focusing feature data according to the first feature flow control parameter, the method further includes: Acquire sample train operation data, lock model parameter information of the train fault warning decision model, and initialize each of the first characteristic flow control parameters and each of the second characteristic flow control parameters as parameter values; Each of the first characteristic flow control parameters and each of the second characteristic flow control parameters are trained according to the sample train operation data.

5. The train fault remote warning method based on artificial intelligence according to claim 1 is characterized in that: The obtaining of prior warning guidance knowledge data, extracting a first guidance knowledge vector of the prior warning guidance knowledge data, and performing interactive focusing feature analysis on the first guidance knowledge vector and the first train state feature map include: Searching for a second train state characteristic map having a characteristic pattern matching the first train state characteristic map, and a priori fault warning knowledge points matching the second train state characteristic map, wherein the a priori fault warning knowledge points are fault warning knowledge points of a priori train operation data, the second train state characteristic map is generated by fusing state path vectors of multiple state dimensions in the a priori train operation data, and the dimension of the state path vector fused by the second train state characteristic map is the same as the dimension of the state path vector fused by the first train state characteristic map; Constructing priori warning guidance knowledge data according to the priori fault warning knowledge points, and extracting a first guidance knowledge vector of the priori warning guidance knowledge data; The second train state characteristic map is integrated with the first train state characteristic map to generate an integrated train state characteristic map, and interactive focusing feature analysis is performed on the first guided knowledge vector and the integrated train state characteristic map.

6. The train fault remote warning method based on artificial intelligence according to claim 5 is characterized in that: The constructing of priori warning guidance knowledge data based on the priori fault warning knowledge points includes: Integrate the trigger node and the termination node in the forward and backward nodes of the priori fault warning knowledge point respectively to generate a knowledge point representation space corresponding to the priori fault warning knowledge point; The trigger node is integrated after the last of the knowledge point representation space to generate priori warning guiding knowledge data.

7. The train fault remote warning method based on artificial intelligence according to claim 6 is characterized in that: The interactive focusing feature analysis of the first guiding knowledge vector and the integrated train state feature map includes: Performing interactive focusing feature analysis on a first sub-vector of the knowledge point representation space in the first guiding knowledge vector and a second train state feature map corresponding to the first sub-vector in the integrated train state feature map; Interactive focusing feature analysis is performed on the second sub-vector of the last trigger node in the first guiding knowledge vector and the first train state feature map in the integrated train state feature map.

8. The train fault remote warning method based on artificial intelligence according to claim 7 is characterized in that: The interactive focusing feature analysis of the first sub-vector of the knowledge point representation space in the first guiding knowledge vector and the second train state feature map corresponding to the first sub-vector in the integrated train state feature map includes: Extracting different types of feature elements from the first sub-vector according to a predefined feature classification rule, and dividing the second train state feature map into a plurality of feature subsets corresponding to the feature classification of the first sub-vector, generating a feature subset of the classified first sub-vector and a feature subset of the corresponding second train state feature map; For each corresponding feature subset pair, a local feature association relationship is established, wherein, for each feature element in each feature subset, based on the logical relationship or potential physical relationship between the feature elements, the association degree between the feature element and other feature elements in the corresponding feature subset is calculated to construct a local feature association matrix for each feature subset pair, wherein the elements in the local feature association matrix represent the association degrees between the corresponding feature elements; Based on the correlation threshold, each local feature correlation matrix is ​​enhanced, and the enhanced local feature correlation matrix is ​​screened to remove feature elements with zero correlation in the enhanced local feature correlation matrix; Based on the screened local feature association matrix, a cross-subset association relationship is constructed between different feature subset pairs to construct a cross-subset association matrix, wherein the elements in the cross-subset association matrix represent the association degree of feature elements between different feature subsets; Integrate all cross-subset association matrices to construct a global association matrix, wherein the global association matrix reflects the overall association relationship between the first sub-vector and the second train state feature map; Based on the global correlation matrix, focusing analysis is performed on the correlation features to find out a combination of characteristic elements whose correlation degree meets a preset condition in the global correlation matrix, so as to generate different focus areas, each of which represents an interaction mode, and the combination of characteristic elements represents a key interaction relationship between the first sub-vector and the second train state characteristic map; Analyze the relative correlation contribution of each feature element in the focus area, and remove the feature elements that do not meet the preset conditions according to the relative correlation contribution, so as to generate a refined key feature set for each focus area; The key feature sets of all focus areas are summarized to form an interactive focus feature analysis result of interactive focus feature analysis on the first sub-vector of the knowledge point representation space in the first guiding knowledge vector and the second train state feature map corresponding to the first sub-vector in the integrated train state feature map.

9. The train fault remote warning method based on artificial intelligence according to claim 5 is characterized in that: The searching for a second train state characteristic map having a characteristic pattern matching the first train state characteristic map, and a priori fault warning knowledge point matching the second train state characteristic map, includes: Acquire a plurality of the a priori train operation data and a priori fault warning knowledge points of each of the a priori train operation data, fuse the state path vectors of a plurality of state dimensions in each of the a priori train operation data, and generate the second train state characteristic map of each of the a priori train operation data; Integrate each of the second train state characteristic maps with the corresponding prior fault warning knowledge points and add them to an index database, and search the index database for the second train state characteristic map that has a characteristic pattern matching the first train state characteristic map, and the prior fault warning knowledge points that match the second train state characteristic map.

10. A train fault remote warning system based on artificial intelligence, characterized in that: The artificial intelligence-based train fault remote warning 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 artificial intelligence-based train fault remote warning method described in any one of claims 1 to 9.

Images