Intelligent operation and maintenance model for rail transit based on dynamic clustering algorithm

Abstract

The invention discloses an intelligent operation and maintenance model for rail transit based on a dynamic clustering algorithm, and relates to the technical field of rail transit, and the intelligent operation and maintenance model comprises the steps of data preprocessing and feature extraction, online real-time Kmeans dynamic clustering, intra-cluster SVMPSO classification, and fault prediction and maintenance decision making. Firstly, large-scale data is divided into a plurality of clusters through Kmeans, each cluster trains a lightweight SVM model, and the calculation complexity of a single model is reduced; sVM sub-models of different clusters can be trained in parallel, so that the overall model construction time is remarkably shortened, and the method is suitable for rail transit scenes with high real-time requirements; different clusters correspond to different operation states of equipment, the SVM is trained independently for each cluster, and the adaptability of the model to different working conditions is improved.

Description

Technical Field

[0001] The present invention relates to the field of rail transit technology, and in particular to an intelligent operation and maintenance model for rail transit based on a dynamic clustering algorithm. Background Art

[0002] Rail transit refers to a type of transportation vehicle or transportation system in which operating vehicles need to travel on specific tracks. The national standard "Common Terminology of Urban Public Transportation" defines urban rail transit as "a general term for fast, large-capacity public transportation that is usually powered by electricity and operates in a wheel-rail mode." With the diversified development of train and railway technology, rail transit has taken on more and more types. It is not only used in long-distance land transportation, but is also widely used in medium and short-distance urban public transportation.

[0003] Most of the existing rail transit operation and maintenance models are models for the normal operation of rail transit, but there is a lack of effective safety monitoring for the safety hazards existing in rail transit.

[0004] In summary, an intelligent operation and maintenance model for rail transit based on dynamic clustering algorithm is designed. Summary of the Invention

[0005] In order to overcome the above-mentioned deficiencies, the present invention provides an intelligent operation and maintenance model for rail transit based on a dynamic clustering algorithm.

[0006] The present invention achieves the above-mentioned purpose through the following technical solutions: An intelligent operation and maintenance model for rail transit based on a dynamic clustering algorithm includes the following steps: Step 1: Data preprocessing and feature extraction: Input the original time domain signal collected by the vibration sensor to eliminate dimensional differences, enhance feature separability, and extract physical features that are strongly correlated with the health status of the bearing; Step 2: Online real-time Kmeans dynamic clustering is used to classify the bearing operating status into clusters of normal, early wear, and severe faults in real time, while adaptively updating the cluster center to adapt to the bearing performance degradation process; Step 2 includes the following specific steps: S21, dynamic cluster center update, adapts to the gradual change of bearing wear, and updates the cluster center in real time to capture the evolution of equipment status. The formula is as follows: (When the sample is assigned to cluster k) in, is the center of the k-th cluster at the t-th iteration, is the current input sample, is the learning rate (usually decays over time), which controls the weight ratio of old centers to new samples; S22, clustering objective function, is used to minimize the sum of the squares of the distances from all samples to their cluster centers. The formula is as follows: in, For indicator samples Whether it belongs to cluster k (value is 0 or 1), is the center of cluster k; Step 3: Intra-cluster SVMPSO classification is used to train a high-precision classifier for each cluster to distinguish subtle fault modes and optimize SVM hyperparameters to improve classification robustness. Step 4: Fault prediction and maintenance decision-making, which is used to generate warning levels based on the integrated clustering and classification results, and dynamically adjust maintenance priorities.

[0007] Preferably, the step 1 comprises the following steps: S11. Data standardization is performed using the following formula: ,in, is the characteristic mean, is the standard deviation, which is used to eliminate sensor range differences, such as the unit difference between accelerometers and thermometers; S12, time domain feature extraction, including signal averaging , signal peak and kurtosis K, ,in, is the i-th sampling value of the time domain vibration signal, such as the bearing vibration amplitude, is the signal average value, , n is the total number of sampling points of the signal, and the kurtosis reflects the sharpness of the signal distribution. The kurtosis of normal bearing vibration is close to 3 (normal distribution), and the kurtosis increases significantly in the early stage of failure; S13, frequency domain feature extraction, the extraction value is FFT spectrum energy, the 0.12kHz frequency band is divided into 5 sub-bands to calculate the energy ratio, the energy calculation formula is , energy reflects the overall strength of the signal, and vibration energy usually increases when bearings wear.

[0008] Preferably, the step three includes the following specific steps: S31, SVM objective function, the formula is as follows: in, is the hyperplane normal vector, which determines the classification boundary direction, and C is the penalty factor, which balances the classification interval and the tolerance of misclassified samples. It is a slack variable that allows a small number of samples to cross the classification boundary; S32, PSO optimization process is used to optimize the kernel parameter γ and penalty factor C of SVM to improve classification accuracy. The specific steps are as follows: S321, particle position ( is the Gaussian kernel parameter); S322, fitness function, cross-validation classification accuracy; S323, update rule, the formula is as follows: + + ; in, is the particle velocity, controlling the search step size, is the inertia weight (usually decaying from 0.9 to 0.4), balancing global and local search, 、 is the individual and group learning factor (usually 2.0), 、 is a random number in [0,1], introducing exploration randomness, is the optimal position found by each particle in the historical iteration (i.e. the optimal solution of the corresponding objective function), is the optimal position found so far for all particles in the entire particle swarm.

[0009] Preferably, the step 4 includes the following specific steps: S41. Health score. The specific formula is as follows: in, is the baseline kurtosis under normal conditions, is the fault threshold, H , the closer it is to 0, the higher the failure risk; S42, maintenance strategy optimization, is used to dynamically adjust the maintenance cycle based on the remaining life prediction (RUL). The priority formula is as follows: + in, 、 is the weight coefficient.

[0010] Preferably, in step S42, the crack growth rate parameter is extracted by the vibration signal, and the RUL is predicted in combination with the real-time data. The calculation formula of the RUL value is as follows: Where a is the crack length, N is the number of load cycles, C and m are material constants, is the stress intensity factor range, R is the stress ratio, is the critical stress intensity factor.

[0011] Preferably, in S21, a learning rate decay strategy is adopted to balance early rapid convergence and late stability in the dynamic cluster center update. The exponential decay formula is as follows: in, is the initial learning rate, is the decay rate, which controls the decline speed, and t is the number of training iterations.

[0012] The beneficial effects of the present invention are: in the intelligent operation and maintenance model for rail transit based on the dynamic clustering algorithm: 1. First, large-scale data is divided into multiple clusters through Kmeans, and a lightweight SVM model is trained for each cluster to reduce the computational complexity of a single model; 2. SVM sub-models of different clusters can be trained in parallel, significantly shortening the overall model building time and making it suitable for rail transit scenarios with high real-time requirements; 3. Different clusters correspond to different operating states of the equipment (such as normal, early wear, and serious failure). SVM is trained separately for each cluster to improve the model's adaptability to different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The present invention will now be described by way of example with reference to the accompanying drawings, in which: Figure 1 It is a step diagram of the present invention. DETAILED DESCRIPTION

[0014] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0015] like Figure 1 As shown, an intelligent operation and maintenance model for rail transit based on a dynamic clustering algorithm includes the following steps: Step 1: Data preprocessing and feature extraction: Input the original time domain signal collected by the vibration sensor to eliminate dimensional differences, enhance feature separability, and extract physical features that are strongly correlated with the health status of the bearing; Step 2: Online real-time Kmeans dynamic clustering is used to classify the bearing operating status into clusters of normal, early wear, and severe faults in real time, while adaptively updating the cluster center to adapt to the bearing performance degradation process; Step 2 includes the following specific steps: S21, dynamic cluster center update, adapts to the gradual change of bearing wear, and updates the cluster center in real time to capture the evolution of equipment status. The formula is as follows: (When the sample is assigned to cluster k) in, is the center of the k-th cluster at the t-th iteration, is the current input sample, is the learning rate (usually decays over time), which controls the weight ratio of old centers to new samples; S22, clustering objective function, is used to minimize the sum of the squares of the distances from all samples to their cluster centers. The formula is as follows: in, For indicator samples Whether it belongs to cluster k (value is 0 or 1), is the center of cluster k; Step 3: Intra-cluster SVMPSO classification is used to train a high-precision classifier for each cluster to distinguish subtle fault modes and optimize SVM hyperparameters to improve classification robustness. Step 4: Fault prediction and maintenance decision-making, which is used to generate warning levels based on the integrated clustering and classification results, and dynamically adjust maintenance priorities.

[0016] Specifically, the step 1 includes the following steps: S11. Data standardization is performed using the following formula: ,in, is the characteristic mean, is the standard deviation, which is used to eliminate sensor range differences, such as the unit difference between accelerometers and thermometers; S12, time domain feature extraction, including signal averaging , signal peak and kurtosis K, ,in, is the i-th sampling value of the time domain vibration signal, such as the bearing vibration amplitude, is the signal average value, , n is the total number of sampling points of the signal, and the kurtosis reflects the sharpness of the signal distribution. The kurtosis of normal bearing vibration is close to 3 (normal distribution), and the kurtosis increases significantly in the early stage of failure; S13, frequency domain feature extraction, the extraction value is FFT spectrum energy, the 0.12kHz frequency band is divided into 5 sub-bands to calculate the energy ratio, the energy calculation formula is , energy reflects the overall strength of the signal, and vibration energy usually increases when bearings wear.

[0017] Specifically, the step three includes the following specific steps: S31, SVM objective function, the formula is as follows: in, is the hyperplane normal vector, which determines the classification boundary direction, and C is the penalty factor, which balances the classification interval and the tolerance of misclassified samples. It is a slack variable that allows a small number of samples to cross the classification boundary; S32, PSO optimization process is used to optimize the kernel parameter γ and penalty factor C of SVM to improve classification accuracy. The specific steps are as follows: S321, particle position ( is the Gaussian kernel parameter); S322, fitness function, cross-validation classification accuracy; S323, update rule, the formula is as follows: + + ; in, is the particle velocity, controlling the search step size, is the inertia weight (usually decaying from 0.9 to 0.4), balancing global and local search, 、 is the individual and group learning factor (usually 2.0), 、 is a random number in [0,1], introducing exploration randomness, is the optimal position found by each particle in the historical iteration (i.e. the optimal solution of the corresponding objective function), is the optimal position found so far for all particles in the entire particle swarm.

[0018] Specifically, the step 4 includes the following specific steps: S41. Health score. The specific formula is as follows: in, is the baseline kurtosis under normal conditions, is the fault threshold, H , the closer it is to 0, the higher the failure risk; S42, maintenance strategy optimization, is used to dynamically adjust the maintenance cycle based on the remaining life prediction (RUL). The priority formula is as follows: + in, 、 is the weight coefficient.

[0019] Specifically, in step S42, the crack growth rate parameter is extracted through the vibration signal, and the RUL is predicted in combination with the real-time data. The calculation formula of the RUL value is as follows: Where a is the crack length, N is the number of load cycles, C and m are material constants, is the stress intensity factor range, R is the stress ratio, is the critical stress intensity factor.

[0020] Specifically, in S21, a learning rate decay strategy is adopted to balance early rapid convergence and late stability in the dynamic cluster center update. The exponential decay formula is as follows: in, is the initial learning rate, is the decay rate, which controls the decline speed, and t is the number of training iterations.

[0021] Specific implementation case: Urban subway train bearing health monitoring system Step 1: Data preprocessing and feature extraction Sensor configuration: A three-axis vibration sensor (sampling frequency 20kHz) is installed on the train bogie.

[0022] Normalization: Perform Zscore normalization on the acceleration signal to eliminate the impact of temperature drift.

[0023] Feature Engineering: Time domain characteristics: Calculate the kurtosis value of each 2-second window. The normal value is 3.2±0.5. A value exceeding 5 triggers an alert.

[0024] Frequency domain characteristics: The 0.12 kHz spectrum is divided into five sub-bands. When the energy proportion of the third sub-band exceeds 30%, it indicates a roller defect.

[0025] Step 2: Dynamic clustering implementation Parameter setting: initial learning rate =0.3 State division: The number of cluster centers k=3, corresponding to: 1. Normal state (cluster center eigenvector: kurtosis 3.1, subband energy distribution [65%, 15%, 10%, 5%, 5%]).

[0026] 2. Early wear (kurtosis 4.8, third sub-band energy 25%).

[0027] 3. Severe failure (kurtosis 7.5, third sub-band energy 38%).

[0028] Step 3: SVMPSO classification optimization Parameter optimization: The particle swarm size was 50, and the optimal parameter combination (C = 8.2, γ = 0.15) was obtained after 100 iterations.

[0029] Classification effect: roller spalling (92% accuracy) and cage fracture (88% accuracy) can be distinguished in early wear clusters.

[0030] Step 4: Maintenance Decision Health calculation: When H=0.6×(5.0K) / 2.5+0.4×(E30.25) / 0.15, H<0.3 triggers the second-level warning.

[0031] RUL prediction: When the crack length a=2mm is detected, the predicted remaining life is 2850km.

[0032] The above description is for inspiration. Based on the above description, relevant personnel can make various changes and modifications without departing from the technical concept of this invention. The technical scope of this invention is not limited to the content of the specification, but must be determined according to the scope of the claims.

Claims

1. An intelligent operation and maintenance model for rail transit based on a dynamic clustering algorithm, characterized by: The following steps are involved: Step 1: Data preprocessing and feature extraction: Input the original time domain signal collected by the vibration sensor to eliminate dimensional differences, enhance feature separability, and extract physical features that are strongly correlated with the health status of the bearing; Step 2: Online real-time Kmeans dynamic clustering is used to divide the bearing operating status into clusters in real time, while adaptively updating the cluster centers to adapt to the bearing performance degradation process; Step 2 includes the following specific steps: S21, dynamic cluster center update, adapts to the gradual change of bearing wear, and updates the cluster center in real time to capture the evolution of equipment status. The formula is as follows: (When the sample is assigned to cluster k) in, is the center of the k-th cluster at the t-th iteration, is the current input sample, is the learning rate (usually decays over time), which controls the weight ratio of old centers to new samples; S22, clustering objective function, is used to minimize the sum of the squares of the distances from all samples to their cluster centers. The formula is as follows: in, For indicator samples Whether it belongs to cluster k (value is 0 or 1), is the center of cluster k; Step 3: Intra-cluster SVMPSO classification is used to train a high-precision classifier for each cluster to distinguish subtle fault modes and optimize SVM hyperparameters to improve classification robustness. Step 4: Fault prediction and maintenance decision-making, which is used to generate warning levels based on the integrated clustering and classification results, and dynamically adjust maintenance priorities.

2. The intelligent operation and maintenance model for rail transit based on a dynamic clustering algorithm according to claim 1 is characterized by: The step 1 comprises the following steps: S11. Data standardization is performed using the following formula: ,in, is the characteristic mean, is the standard deviation, which is used to eliminate the difference in sensor range; S12, time domain feature extraction, including signal averaging , signal peak and kurtosis K, ,in, is the i-th sampling value of the time domain vibration signal, is the signal average value, , n is the total number of sampling points of the signal; S13, frequency domain feature extraction, the extraction value is FFT spectrum energy, the energy calculation formula is .

3. The intelligent operation and maintenance model for rail transit based on a dynamic clustering algorithm according to claim 1 is characterized by: The step three includes the following specific steps: S31, SVM objective function, the formula is as follows: in, is the hyperplane normal vector, which determines the classification boundary direction, and C is the penalty factor, which balances the classification interval and the tolerance of misclassified samples. It is a slack variable that allows a small number of samples to cross the classification boundary; S32, PSO optimization process is used to optimize the kernel parameter γ and penalty factor C of SVM to improve classification accuracy. The specific steps are as follows: S321, particle position ( is the Gaussian kernel parameter); S322, fitness function, cross-validation classification accuracy; S323, update rule, the formula is as follows: + + ； in, is the particle velocity, controlling the search step size, is the inertia weight, balancing global and local search, 、 are individual and group learning factors, 、 is a random number in [0,1], introducing exploration randomness, is the optimal position found by each particle in the historical iteration, is the optimal position found so far for all particles in the entire particle swarm.

4. The intelligent operation and maintenance model for rail transit based on a dynamic clustering algorithm according to claim 1 is characterized by: The step 4 includes the following specific steps: S41. Health score. The specific formula is as follows: in, is the baseline kurtosis under normal conditions, is the fault threshold, H , the closer it is to 0, the higher the failure risk; S42, maintenance strategy optimization, is used to dynamically adjust the maintenance cycle based on the remaining life prediction (RUL). The priority formula is as follows: H+ in, 、 is the weight coefficient.

5. The intelligent operation and maintenance model for rail transit based on a dynamic clustering algorithm according to claim 4 is characterized by: In step S42, the crack growth rate parameter is extracted through the vibration signal, and the RUL is predicted in combination with the real-time data. The calculation formula of the RUL value is as follows: Where a is the crack length, N is the number of load cycles, C and m are material constants, is the stress intensity factor range, R is the stress ratio, is the critical stress intensity factor.

6. The intelligent operation and maintenance model for rail transit based on a dynamic clustering algorithm according to claim 1 is characterized by: In S21, a learning rate decay strategy is adopted to balance early rapid convergence and late stability in the dynamic cluster center update. The exponential decay formula is as follows: in, is the initial learning rate, is the decay rate, which controls the decline speed, and t is the number of training iterations.

Images