A metro catenary control method based on millimeter wave monitoring and a catenary diverter device

Abstract

The application discloses a subway contact network control method based on millimeter wave monitoring and a contact network shunt device, and the method comprises the following steps: collecting train state data, and preprocessing the train state data; analyzing the preprocessed train state data, and locking a target train; acquiring electrical parameters of the contact network, estimating the state of the contact network based on the electrical parameters; judging whether the target train is stopped in a preset forbidden stopping area based on the position of the target train; and issuing an instruction to control the main control contactor based on the state of the contact network and the positional relationship between the target train and the forbidden stopping area. The application provides a subway contact network control method based on millimeter wave monitoring, intelligently controls the opening and closing of the main control contactor, and thus avoids the problem that the non-working branch of the contact line is fused due to the parking of the electric train at the insulating anchor joint of the subway flexible contact network.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control of urban rail transit, and discloses a subway catenary control method and catenary diverter device based on millimeter wave monitoring. Background Technology

[0002] When a train stops at the joint of the insulated anchor section of the overhead contact line due to train operation requirements or a sudden emergency, if the train's pantograph is connected to the transition area between the non-working and working branches of the contact wire within the joint of the insulated anchor section, the pantograph's carbon sliding plate will short-circuit the contact wires on both sides of the insulated section (the working branch on one side of the section and the non-working branch on the other side of the section). The traction current of other trains on the line will pass through this short-circuit point along the contact wire. Since the train is stopped at this time, the transition resistance of the area where the short-circuit point contacts the non-working branch of the contact wire is relatively large (there is a certain air gap). Due to the continuous flow of current, the contact area will overheat and eventually melt the non-working branch of the contact wire.

[0003] Currently, there are no effective protective measures against line breakage caused by parking in high-altitude areas, and the risk will become increasingly higher as the operating density gradually increases. Therefore, there is an urgent need to study joint shunt devices for the insulated anchor sections of flexible contact wires. Summary of the Invention

[0004] Purpose of the invention: To provide a subway catenary control method and catenary diverter device based on millimeter-wave monitoring, so as to solve the above-mentioned problems existing in the prior art.

[0005] Technical solution: A subway catenary control method based on millimeter-wave monitoring, comprising the following steps: S1. Collect train status data and preprocess the train status data; the train status data includes the train's position, speed, distance, direction of travel, and acceleration. S2. Analyze the preprocessed train status data to locate the target train; S3. Obtain the electrical parameters of the overhead contact line, and estimate the state of the overhead contact line based on the electrical parameters; S4. Based on the location of the target train, determine whether the target train has stopped within the preset no-stopping zone; S5. Based on the status of the overhead contact line and the positional relationship between the target train and the no-stopping zone, issue instructions to control the main control contactor.

[0006] According to one aspect of this application, step S1 further comprises: S11. The millimeter-wave radar scans the surrounding environment at a predetermined frequency to obtain the range scanning matrix and the first reflection intensity matrix as train status data. S12. The median filtering algorithm is used to perform noise reduction filtering on the distance scan matrix and the first reflection intensity matrix. S13. The millimeter-wave radar transmits chirp signals, receives the echoes of the chirp signals and decodes them to obtain the range matrix, velocity matrix and second reflection intensity matrix.

[0007] According to one aspect of this application, step S2 of the millimeter-wave radar further comprises: S21. Based on the denoised and filtered distance scan matrix and the first reflection intensity matrix, the train outline is extracted using an edge extraction algorithm, and the train outline is fitted to obtain the train bounding box. S22. Perform range-Doppler domain detection on the range matrix, velocity matrix, and second reflection intensity matrix to extract potential target trains; project the potential target trains onto the millimeter-wave radar coordinate system, intersecting with the train bounding box to obtain the fused detection results; S23. Using Kalman filtering, perform multi-target tracking on the fused detection results, and update the position, speed, and acceleration of each train based on the tracking results.

[0008] According to one aspect of this application, step S3 of the millimeter-wave radar further comprises: S31. Obtain the electrical parameters of the contact network, including the first segment voltage and the second segment voltage of the contact network, and calculate the voltage difference between the first segment voltage and the second segment voltage; S32. Based on the voltage difference between the first segment voltage and the second segment voltage, obtain the voltage and current at both ends of the main control contactor, and estimate the state of the contact network through the voltage and current at both ends of the main control contactor. S321. Determine whether the voltage difference across the main control contactor is less than the preset threshold. If it is, allow the main control contactor to close; otherwise, do not allow it to close. S322. Obtain the current passing through the main control contactor. Based on the current passing through the main control contactor, the contactor's rated current, and the overload factor, determine whether the contactor is overloaded. If overloaded, issue an alarm indicating the overload risk.

[0009] According to one aspect of this application, step S4 further comprises: S41. Obtain train dispatch instructions and route topology, and determine no-stopping areas; S42. Compare the position of the target train with the no-stopping zone to determine whether the target train is located in the no-stopping zone. If it is, start the entry timer; otherwise, start the departure timer.

[0010] According to one aspect of this application, step S5 further comprises: S51. When the target train is located in a no-stopping zone, determine whether the following conditions are met simultaneously: the target train stops and the entry time exceeds the preset allowable dwell time threshold, and the voltage difference between the first segment voltage and the second segment voltage is less than the preset allowable voltage difference threshold. If the conditions are met, a closing command is issued to close the main control contactor; if the conditions are not met, no command is issued. S52. When the target train leaves the no-stopping area, determine whether the following conditions are met simultaneously: the departure time exceeds the preset delay disconnection time, and the current of the main control contactor is lower than the preset current lower limit threshold. If the conditions are met, a disconnection command is issued to disconnect the main control contactor; if the conditions are not met, no command is issued.

[0011] According to one aspect of this application, it also includes S6: acquiring historical action data of the main control contactor, and performing fault diagnosis and early warning based on the historical action data, specifically: S61. Obtain historical action data of the main control contactor, including action time, triggering conditions and parameters before and after the action, and establish a normal action mode library based on the historical action data. S62. Collect new action data of the main control contactor, and use machine learning algorithms to detect anomalies in the new action data based on the normal action mode library. S63. Based on the results of anomaly detection, determine the cause of the anomaly, including prolonged train stops, abnormal changes in overhead contact line parameters, or equipment failure. Based on the cause of the anomaly, issue timely warnings, prompt maintenance personnel to conduct repairs, and activate emergency response plans.

[0012] According to one aspect of this application, step S61 further comprises: S611. Obtain historical action data of the main control contactor, clean and preprocess the historical action data, and remove noisy and invalid data. S612. Extract time-domain and frequency-domain features from the preprocessed historical motion data; where time-domain features include motion duration, motion interval, rise / fall time, and overshoot; frequency-domain features include spectrogram, dominant frequency, spectral entropy, and center frequency; S613. Standardize the time-domain and frequency-domain characteristics to eliminate the influence of dimensions; S614. Based on the time-domain and frequency-domain characteristics after eliminating the influence of dimensions, establish a normal action mode library.

[0013] According to one aspect of this application, step S614 further comprises: S6141. Construct a deep convolutional neural network model based on the time-domain and frequency-domain features after eliminating the influence of dimensions. S6142. Obtain typical action sample data from the main control contactor. Based on the typical action sample data, use the backpropagation algorithm to optimize the parameters of the deep convolutional neural network model. S6143. In the optimized deep convolutional neural network model, historical action data is propagated forward to extract deep feature vectors. S6144. Perform cluster analysis on the deep feature vectors to automatically discover the cluster structure inside the historical action data, with each cluster corresponding to a typical action pattern. S6145. Statistical action pattern center feature vector and boundary threshold. Based on the center feature vector and boundary threshold, construct a normal action pattern library.

[0014] A subway catenary shunt device based on millimeter-wave monitoring, using the subway catenary control method based on millimeter-wave monitoring described in any of the above technical solutions, includes a DC contact module, a monitoring module, and a train status acquisition module; wherein the DC contact module includes a DC contactor, a current transmitter, and a voltage transmitter; the monitoring module includes a communication management unit, a 4G communication module, a core controller, anti-condensation equipment, and an operation panel; the train status acquisition module includes at least two sets of millimeter-wave radars. Beneficial effects: This invention provides a control method for subway contact networks based on millimeter-wave monitoring, which intelligently controls the opening and closing of the main control contactor, thereby avoiding the problem of non-working support of the contact wire melting at the joint of the insulated anchor section of the subway flexible contact network due to the stopping of the electric train. Attached Figure Description

[0015] Figure 1 This is a flowchart of the present invention.

[0016] Figure 2 This is a flowchart of step S1 of the present invention.

[0017] Figure 3 This is a flowchart of step S2 of the present invention.

[0018] Figure 4 This is a flowchart of step S3 of the present invention.

[0019] Figure 5 This is a flowchart of step S4 of the present invention.

[0020] Figure 6 This is a flowchart of step S5 of the present invention.

[0021] Figure 7 This is a simplified diagram illustrating the application scenario of the present invention.

[0022] Figure 8 This is a system structure diagram of the present invention.

[0023] The attached diagram is labeled as follows: 1. First working support; 2. Second working support; 3. First non-working support; 4. Second non-working support; 5. Pantograph. Detailed Implementation

[0024] like Figure 1 As shown, this application proposes a subway catenary control method based on millimeter-wave monitoring, the specific steps of which are as follows: S1. Collect train status data and preprocess the train status data; the train status data includes the train's position, speed, distance, direction of travel, and acceleration. S2. Analyze the preprocessed train status data to locate the target train; S3. Obtain the electrical parameters of the overhead contact line, and estimate the state of the overhead contact line based on the electrical parameters; S4. Based on the location of the target train, determine whether the target train has stopped within the preset no-stopping zone; S5. Based on the status of the overhead contact line and the positional relationship between the target train and the no-stopping zone, issue instructions to control the main control contactor.

[0025] In this implementation, real-time data on train position, speed, distance, direction of travel, and acceleration are collected. This data is then preprocessed, including filtering, noise reduction, and standardization, to ensure accuracy and reliability, providing a solid data foundation for precise control. The preprocessed data is analyzed using pattern recognition and machine learning algorithms to identify and locate specific target trains. This rapid and accurate identification of target trains supports subsequent control decisions. Electrical parameters of the overhead contact system, such as voltage, current, and impedance, are collected, and physical or data-driven models are used to estimate the contact system's condition. Real-time monitoring of the contact system prevents faults and ensures stable power supply. Based on the real-time position data of the target train, it is determined whether the train has entered a pre-defined no-stopping zone, improving safety, preventing trains from entering dangerous areas, and reducing the risk of accidents. Based on the relationship between the contact system condition and train position, the control system issues commands to adjust the state of the main contactor. Precise control of the contact system power optimizes energy use and improves the overall efficiency and reliability of the system.

[0026] This embodiment monitors the train's status and, in conjunction with the voltage monitoring of the overhead contact line, effectively prevents contact wire meltdown caused by train stopping at the insulated anchor section by closing a short-circuit switch to divert current. This not only improves the intelligence level of the subway system but also enhances safety and efficiency.

[0027] like Figure 2 As shown, according to one aspect of this application, step S1 further comprises: S11. The millimeter-wave radar scans the surrounding environment at a predetermined frequency to obtain the range scanning matrix and the first reflection intensity matrix as train status data. S12. The median filtering algorithm is used to perform noise reduction filtering on the distance scan matrix and the first reflection intensity matrix. S13. The millimeter-wave radar transmits chirp signals, receives the echoes of the chirp signals and decodes them to obtain the range matrix, velocity matrix and second reflection intensity matrix.

[0028] In further implementation, millimeter-wave radar will use f... L Frequency scan of the surrounding environment to obtain the distance scan matrix D L and reflection intensity matrix I L For the distance scan matrix D L and reflection intensity matrix I L After performing noise reduction filtering, D is obtained. L 'and I L The noise reduction filtering uses a median filtering algorithm, taking the median value of pixels within a 3x3 neighborhood to remove impulse noise and Gaussian noise; the millimeter-wave radar uses f M The chirp signal is transmitted at a certain frequency, the echo is received and decoded to obtain the distance matrix D. M Velocity matrix V M and reflection intensity matrix I M The chirp signal is a signal whose frequency changes over time. In radar technology, chirped signals are used to improve measurement resolution and range accuracy; for the range matrix D M Velocity matrix V M and reflection intensity matrix I M Range-Doppler Aspect Ratio (CFAR) detection is performed to extract potential target points. CFAR detection selects a reference window centered on the target cell on a two-dimensional range-Doppler matrix and estimates the mean value u of the background noise. t and variance σ t Set the detection threshold T d If the amplitude of the unit under test is greater than u t +T d σ t If the value is 0, it is considered the target point. The inertial navigation module collects data from the three-axis accelerometer and three-axis gyroscope, and estimates the device's own position, velocity, and attitude in real time using Kalman filtering, denoted as vectors P0, V0, and A0. The state transition equation is:

[0029] Among them, X k Let I be a 9-dimensional state vector representing position, velocity, and attitude, where I is the identity matrix, Δt is the sampling interval, and W is the position, velocity, and attitude. kFor process noise, the observation equation is:

[0030] Among them, Z k V represents the position, velocity, and attitude observations output by the inertial navigation module. k To observe noise.

[0031] In this embodiment, millimeter-wave radar accurately measures the distance and reflection intensity of objects around the train, providing high-precision spatial positioning and obstacle detection. Median filtering removes random noise from the data while preserving edge features, improving data reliability. Millimeter-wave radar simultaneously acquires target distance, speed, and reflection intensity information via chirp signals, playing a crucial role in monitoring dynamic environments. Range-Doppler detection accurately identifies and extracts information about potential target trains, which is essential for ensuring the safety and efficiency of train operation. Finally, the inertial navigation module collects data from three-axis accelerometers and three-axis gyroscopes, combined with a Kalman filter algorithm, to estimate the device's position, speed, and attitude in real time, providing important support for stable train operation and precise positioning.

[0032] like Figure 3 As shown, according to one aspect of this application, step S2 further comprises: S21. Based on the denoised and filtered distance scan matrix and the first reflection intensity matrix, the train outline is extracted using an edge extraction algorithm, and the train outline is fitted to obtain the train bounding box. S22. Perform range-Doppler domain detection on the range matrix, velocity matrix, and second reflection intensity matrix to extract potential target trains; project the potential target trains onto the millimeter-wave radar coordinate system, intersecting with the train bounding box to obtain the fused detection results; S23. Using Kalman filtering, perform multi-target tracking on the fused detection results, and update the position, speed, and acceleration of each train based on the tracking results.

[0033] The millimeter-wave radar will use the pre-processed millimeter-wave radar data range scanning matrix D L '、Reflection Intensity Matrix I L 'and millimeter-wave radar data range matrix D M Velocity matrix V M and reflection intensity matrix I M The data is fed into the joint target detection module; in the distance scan matrix D... L 'and reflection intensity matrix I LAn edge detection algorithm is applied to extract the train contour and fit a 2D bounding box. The detection results from the millimeter-wave radar range-Doppler domain are projected onto the millimeter-wave radar coordinate system and intersected with the train bounding box to obtain the fused detection results of the train under the two sensors. Multi-target tracking is performed on the fused detection results based on Kalman filtering to estimate the position, velocity, acceleration, and other state variables of each train. The state vector is defined as {x, y, v}. x v y a x a y The state transition equation is:

[0034] The observation equation is:

[0035] Among them U k For process noise, V k To observe noise.

[0036] In this embodiment, an edge extraction algorithm accurately extracts the train's outline from millimeter-wave radar data and fits a 2D bounding box. Projecting the range-Doppler domain (CFAR) detection results of the millimeter-wave radar onto the millimeter-wave radar coordinate system and intersecting it with the train bounding box yields more accurate train detection results. This fusion method leverages the advantages of both sensors, improving the accuracy and reliability of target detection. Multi-target tracking based on Kalman filtering of the fused detection results allows for real-time estimation of each train's position, speed, acceleration, and other state variables. This effectively tracks the dynamic changes of multiple targets, which is crucial for command and dispatch and safety monitoring systems. This embodiment achieves precise monitoring and tracking of train status, providing strong data support for the subway catenary control system, thereby improving the intelligence level and operational safety of the entire urban rail transit system. The application of these technologies is not limited to subway systems but can also be extended to other types of rail transit systems, further promoting the intelligent development of the transportation industry.

[0037] like Figure 4 As shown, according to one aspect of this application, step S3 further comprises: S31. Obtain the electrical parameters of the contact network, including the first segment voltage and the second segment voltage of the contact network, and calculate the voltage difference between the first segment voltage and the second segment voltage; S32. Based on the voltage difference between the first segment voltage and the second segment voltage, obtain the voltage and current at both ends of the main control contactor, and estimate the state of the contact network through the voltage and current at both ends of the main control contactor. S321. Determine whether the voltage difference across the main control contactor is less than the preset threshold. If it is, allow the main control contactor to close; otherwise, do not allow it to close. S322. Obtain the current passing through the main control contactor. Based on the current passing through the main control contactor, the contactor's rated current, and the overload factor, determine whether the contactor is overloaded. If overloaded, issue an alarm indicating the overload risk.

[0038] In this embodiment, the monitoring box measures the voltage U1 of the first segment and the voltage U2 of the second segment of the contact network, calculates the voltage difference as ΔU=|U1-U2|, and monitors the voltage U across the main control contactor. a and U b , if |U a -U b If |>ΔU0, the main control contactor is determined to be open; otherwise, it is closed. Here, ΔU0 is the set differential pressure threshold. The current I passing through the main control contactor is measured and combined with the contactor's rated current I0. N And the overload factor α, to determine whether the contactor is overloaded. If I>αI N An alarm will be triggered to indicate an overload risk.

[0039] By acquiring the voltages of the first and second segments of the overhead contact line and calculating their voltage difference, the power grid status can be monitored in real time, and voltage anomalies can be detected promptly, ensuring the stable operation of the overhead contact line. Judging the status of the main control contactor based on the voltage difference across it ensures the power grid operates within a safe voltage range. If the voltage difference exceeds a preset threshold, it indicates a potential safety hazard; disconnecting the contactor at this time can prevent further equipment damage or safety accidents. Monitoring the current flowing through the main control contactor and comparing it with the contactor's rated current and overload factor determines whether the contactor is overloaded. If an overload is detected, the system will issue an alarm, prompting maintenance personnel to handle the situation promptly, thereby preventing equipment damage or power outages caused by overload. This embodiment improves the safety and reliability of the subway overhead contact line control system through precise voltage and current monitoring, as well as intelligent status judgment and alarm mechanisms. Simultaneously, this intelligent monitoring and control method also helps reduce maintenance costs, improve energy efficiency, and promote the sustainable development of the rail transit system.

[0040] like Figure 5 As shown, according to one aspect of this application, step S4 further comprises: S41. Obtain train dispatch instructions and route topology, and determine no-stopping areas; S42. Compare the position of the target train with the no-stopping zone to determine whether the target train is located in the no-stopping zone. If it is, start the entry timer; otherwise, start the departure timer.

[0041] In this embodiment, the no-stopping zone is determined as R based on the train dispatching instructions and the line topology. s , represented as {{x} min x max}, {y min y max}; The train position {x, y} from the target tracking result is compared with the stopping area range R. s The system compares the data to determine if the train is within a no-stopping zone. If the train enters a no-stopping zone, the entry timer t is initiated. in If the train leaves the no-stopping zone, the departure timer will be activated. out .

[0042] This embodiment monitors the train's dwell time in the area by entering the timer and records the time the train leaves the area by leaving the timer. This helps to promptly identify and address potential safety issues, ensuring the standardization and safety of train operations.

[0043] like Figure 6 As shown, according to one aspect of this application, step S5 further comprises: S51. When the target train is located in a no-stopping zone, determine whether the following conditions are met simultaneously: the target train stops and the entry time exceeds the preset allowable dwell time threshold, and the voltage difference between the first segment voltage and the second segment voltage is less than the preset allowable voltage difference threshold. If the conditions are met, a closing command is issued to close the main control contactor; if the conditions are not met, no command is issued. S52. When the target train leaves the no-stopping area, determine whether the following conditions are met simultaneously: the departure time exceeds the preset delay disconnection time, and the current of the main control contactor is lower than the preset current lower limit threshold. If the conditions are met, a disconnection command is issued to disconnect the main control contactor; if the conditions are not met, no command is issued.

[0044] In this embodiment, the positional relationship and dwell time between the train and the no-stopping zone, the pressure difference ΔU, and the contactor state S are obtained. c Parameters; if the condition is met (train is in a no-stopping zone) &(t) in >T s )&( ΔU<ΔU c )&( S c If ==0), then close the main control contactor and set S. c =1, where T s ΔU is the threshold for the allowed dwell time. c The allowable differential pressure threshold; if the condition is met (train leaves the no-stopping area) &(t) out >T l ) | (I c If the main control contactor is disconnected, set S to... c =0. Where T​l To delay the disconnection time and ensure the train has completely departed, I c The lower current threshold; if S c When the state changes, a control command is output, and the timing of the action and the triggering conditions are recorded for subsequent analysis.

[0045] This embodiment improves the operational efficiency and safety of urban rail transit systems through intelligent judgment and timely instruction issuance, ensures the rational allocation and effective utilization of power supply, and also provides support for rapid response in emergency situations.

[0046] According to one aspect of this application, it also includes S6: acquiring historical action data of the main control contactor, and performing fault diagnosis and early warning based on the historical action data, specifically: S61. Obtain historical action data of the main control contactor, including action time, triggering conditions and parameters before and after the action, and establish a normal action mode library based on the historical action data. S62. Collect new action data of the main control contactor, and use machine learning algorithms to detect anomalies in the new action data based on the normal action mode library. S63. Based on the results of anomaly detection, determine the cause of the anomaly, including prolonged train stops, abnormal changes in overhead contact line parameters, or equipment failure. Based on the cause of the anomaly, issue timely warnings, prompt maintenance personnel to conduct repairs, and activate emergency response plans.

[0047] In this embodiment, historical action data of the main control contactor is recorded and analyzed, including action time, triggering conditions, and parameters before and after the action. A normal action mode library is established, and machine learning algorithms such as single-class SVM are used to detect anomalies in newly collected action data. Abnormal actions that deviate from the normal mode are promptly identified. Abnormal actions generally fall into several categories: action too early, action too late, multiple actions, and prolonged closure. The cause of the anomaly needs to be analyzed, which may be caused by prolonged train stops, abnormal changes in catenary parameters, or equipment failure. If the abnormal action is determined to be caused by equipment failure, an early warning is issued to the maintenance personnel for repair, and an emergency response plan is activated. If necessary, measures such as speed limits or shutdowns can be taken to ensure train operation safety.

[0048] In another embodiment of this application, step S61 further comprises: S611. Obtain historical action data of the main control contactor, clean and preprocess the historical action data, and remove noisy and invalid data. S612. Extract time-domain and frequency-domain features from the preprocessed historical motion data; where time-domain features include motion duration, motion interval, rise / fall time, and overshoot; frequency-domain features include spectrogram, dominant frequency, spectral entropy, and center frequency; S613. Standardize the time-domain and frequency-domain characteristics to eliminate the influence of dimensions; S614. Based on the time-domain and frequency-domain characteristics after eliminating the influence of dimensions, establish a normal action mode library.

[0049] In this embodiment, 10,000 main control contactor action data collected within one month are cleaned, and 2% of noise and invalid data are removed, resulting in 9,800 valid data. Twenty time-domain and frequency-domain features are extracted from each action data, including 10 time-domain features such as action duration, action interval, rise time, fall time, and overshoot; and 10 frequency-domain features such as the main frequency, spectral entropy, center frequency, and frequency band energy ratio on the spectrum. The max-min normalization method is used to standardize all feature data to the [0, 1] interval.

[0050] Furthermore, step S614 further comprises: S6141. Construct a deep convolutional neural network model based on the time-domain and frequency-domain features after eliminating the influence of dimensions. S6142. Obtain typical action sample data from the main control contactor. Based on the typical action sample data, use the backpropagation algorithm to optimize the parameters of the deep convolutional neural network model. S6143. In the optimized deep convolutional neural network model, historical action data is propagated forward to extract deep feature vectors. S6144. Perform cluster analysis on the deep feature vectors to automatically discover the cluster structure inside the historical action data, with each cluster corresponding to a typical action pattern. S6145. Statistical action pattern center feature vector and boundary threshold. Based on the center feature vector and boundary threshold, construct a normal action pattern library.

[0051] In this embodiment, a pre-trained deep convolutional neural network (AlexNet) is selected as the base model. The parameters of the first 5 convolutional layers and 2 fully connected layers are frozen, and the output layer is replaced with 2 neurons, corresponding to normal and abnormal respectively. 500 typical action data points are manually labeled, including 300 normal and 200 abnormal ones, for fine-tuning the network parameters and setting the training parameters. batch_size=32, learning_rate=0.001, epochs=50; The network loss decreased from the initial 0.8 to 0.1, the accuracy increased from 0.6 to 0.95, and the model converged.

[0052] All 9800 historical action data points were forward-propagated in the fine-tuned deep convolutional neural network. The output of the penultimate layer (4096 dimensions) was extracted as the deep feature vector, resulting in a 9800×4096 feature matrix. K-means clustering analysis was performed on the feature matrix, and the optimal number of clusters K=5 was determined by the elbow rule, meaning that the historical data contained 5 typical action patterns. The feature vectors and boundary thresholds of the 5 cluster centers were statistically analyzed to construct an action pattern library.

[0053] In another embodiment of this application, the deep convolutional neural network model in step S6141 specifically includes: a dual-channel input layer, a mixed convolutional layer, an attention pooling layer, a residual block, a global average pooling layer, a fully connected layer, and a softmax output layer.

[0054] The dual-channel input layer can simultaneously receive time-domain and frequency-domain feature data from the main control contactor, capturing the time-frequency characteristics of the action signal. The two parallel input channels each receive a 20-dimensional feature vector. The time-domain channel receives features such as action duration and interval; the frequency-domain channel receives features such as spectral entropy and main frequency.

[0055] The hybrid convolutional layer performs convolution operations on both the time and frequency channels separately to extract local features and achieve cross-channel information exchange. Three different sized convolutional kernels (1×1, 3×3, and 5×5) are used to convolve each channel, and then the three convolutional results are concatenated to form a hybrid feature map. The number of convolutional kernels is 32, and the activation function is ReLU. By mixing convolutional kernels of different scales, both local and global features are considered; cross-channel connections enhance information fusion in the time and frequency domains.

[0056] The attention pooling layer adaptively weights the mixed feature map through an attention mechanism, highlighting key regions and suppressing noise interference. First, a 1×1 convolution compresses the mixed feature map into 32 channels. Then, a softmax function maps it to an attention weight matrix. The weight matrix is ​​element-wise multiplied with the original mixed feature map to obtain a weighted feature map. Finally, global average pooling is performed to obtain a 32-dimensional feature vector. By introducing an attention mechanism, the importance of features is automatically learned, enhancing the network's feature extraction capability and robustness.

[0057] Residual blocks deepen the network layers, improving feature representation capabilities while avoiding gradient vanishing / exploding issues. A "bottleneck structure" is employed: first, 1×1 convolutions compress the number of channels, then 3×3 convolutions extract features, and finally, 1×1 convolutions restore the number of channels. Input features are directly added to output features via shortcut connections to form residual connections. Three residual blocks are stacked, with convolutional channels of 64, 128, and 256 respectively in each block. By introducing residual learning, the network can be effectively trained even at deeper layers; the bottleneck structure reduces the number of parameters and improves computational efficiency.

[0058] The global average pooling layer compresses the feature map output by the residual block into a fixed-length feature vector, reducing model parameters and avoiding overfitting. Average pooling is performed on the 256-channel feature map output by the residual block in the spatial dimension to obtain a 256-dimensional feature vector.

[0059] The fully connected layer and the softmax output layer perform nonlinear transformations and classification predictions on the extracted deep features, outputting the probability distribution of action patterns. The fully connected layer first maps the 256-dimensional features to 64 dimensions, then activates them through ReLU, and finally outputs two neurons (normal / abnormal), which are then normalized by softmax to obtain the probability.

[0060] In the input layer, temporal features (20-dimensional) and frequency features (20-dimensional) are input into two parallel convolutional channels respectively. In the hybrid convolutional layer, each channel is convolved with three types of convolutional kernels (1x1, 3x3, 5x5), and then concatenated to obtain a 96-channel hybrid feature map. In the attention pooling layer, a 1x1 convolution is first used to compress the feature map to 32 channels, and then attention weights are obtained through softmax. After weighted pooling, a 32-dimensional feature map is output. In the residual block, there are three residual blocks, each containing two convolutional layers (1x1 and 3x3) and a shortcut connection, with 64, 128, and 256 channels respectively. In the global average pooling layer, the 256-channel feature map output from the residual block is compressed into a 256-dimensional vector. In the fully connected layer and the output layer, the 256-dimensional feature map is first mapped to 64 dimensions, then activated by ReLU, and finally the probabilities of the two classes are output.

[0061] This embodiment employs a network architecture of "dual-channel input + hybrid convolution + attention pooling + residual block + global pooling + fully connected layers," which integrates temporal and frequency domain information. Furthermore, it enhances the network's feature extraction and generalization capabilities through attention mechanisms and residual learning. In addition, the introduction of global pooling layers and bottleneck structures helps control model complexity and improve training efficiency.

[0062] Furthermore, step S62 further comprises: S621. For newly acquired motion data, extract the same time-frequency features and standardize them, then feed them into a fine-tuned convolutional neural network to extract deep feature vectors. S622. Calculate the distance between the action data feature and the center of each action pattern. If the minimum distance exceeds the boundary threshold of the corresponding pattern, it is judged as an abnormal action; otherwise, it is judged as the nearest normal action pattern. S623. For action data that is judged as abnormal, prompt the user to manually label its true category and add it to the training set. Use active learning strategies to update the original convolutional neural network and clustering model online to continuously improve system performance. S624. Statistically track the frequency and severity of abnormal actions. If the frequency exceeds the preset alarm threshold, trigger an early warning and notify maintenance personnel to intervene.

[0063] In this embodiment, 20-dimensional time-frequency features are extracted from 100 newly collected action data points and standardized. These features are then fed into a fine-tuned deep convolutional neural network to extract 4096-dimensional deep feature vectors. The Euclidean distance between each new data point and the centers of the five patterns is calculated. If the minimum distance exceeds the corresponding threshold, the data point is judged as abnormal; otherwise, it is judged as the closest normal pattern. The results show that 70 data points are normal and 30 are abnormal. The abnormal data points are manually reviewed, and 25 are confirmed as true anomalies, including 10 cases of actions that are too early, 8 cases of actions that are too late, and 7 cases of multiple actions. The remaining 5 are false positives. The 25 true anomaly data points are added to the training set, and the deep convolutional neural network and clustering model are fine-tuned using an active learning strategy. After three rounds of iteration, the false positive rate decreases from 5% to 1%. The frequency of abnormal actions is counted every hour. If the frequency exceeds 5 times in 3 consecutive hours, an alert is triggered and maintenance personnel are notified.

[0064] Furthermore, step S63 further comprises: S631. When an abnormal action warning is triggered, the system automatically makes a preliminary judgment on possible causes of the fault based on the type of abnormal action and the trend of parameter changes, such as: long-term train stop, abnormal changes in catenary parameters, or equipment failure. S632. Generate a fault analysis report, provide relevant maintenance suggestions and emergency plans, and assist maintenance personnel in quickly locating the problem; S633. Based on the severity and scope of the fault, the system provides corresponding emergency response measures, such as speed-limited operation and shutdown for maintenance, to ensure driving safety to the greatest extent possible. S634. Feed back the fault analysis report and emergency response status to the system to enrich the knowledge base and optimize subsequent diagnosis and decision-making processes.

[0065] In a specific implementation, the system detects eight premature aberrations within three consecutive hours, along with a continuous decrease in the overhead contact line differential pressure. This is determined to be due to increased resistance caused by aging of the overhead contact line, triggering an early warning. The system automatically generates a fault analysis report, indicating that the fault is caused by aging of the overhead contact line; the recommended maintenance is to replace the conductor in the faulty section; and the emergency plan is to limit speed and reduce train traction current. After receiving the warning, maintenance personnel review the fault analysis report and remotely reduce the train's speed, effectively reducing the overhead contact line load. After troubleshooting, the fault report and handling details are fed back to the system, enriching the knowledge base. The system then summarizes its experience and optimizes the diagnostic model and emergency plan database.

[0066] like Figure 7 As shown, when a train stops at the joint of the insulated anchor section of the overhead contact line due to train operation requirements or a sudden emergency, if the train's pantograph is connected to the transition area between the non-working and working branches of the contact wire within the joint of the insulated anchor section, the pantograph's carbon sliding plate will short-circuit the contact wires on both sides of the insulated section. The traction current of other trains on the line will pass through this short-circuit point along the contact wire. Since the train is stopped at this time, the transition resistance of the area where the short-circuit point contacts the non-working branch of the contact wire is relatively large. Due to the continuous flow of current, the contact area will overheat and eventually melt the non-working branch of the contact wire.

[0067] like Figure 8 As shown, this application proposes a subway catenary shunt device based on millimeter-wave monitoring, including a DC contact module, a monitoring module, and a train status acquisition module; wherein the DC contact module includes a DC contactor, a current transmitter, and a voltage transmitter; the monitoring module includes a communication management unit, a 4G communication module, a core controller, anti-condensation equipment, and an operation panel; the train status acquisition module includes at least two sets of millimeter-wave radars.

[0068] The device has the function of remotely or locally engaging and disengaging the "automatic opening and closing of the main control contactor", and also has the function of uploading data to the metro's power monitoring (PSCADA) system. The uploaded data includes, but is not limited to: the open / closed position status of the main control contactor, the current value passing through the device, the number of times the main control contactor has opened and closed, and the time and speed of passing trains.

[0069] The monitoring module collects and transmits information such as train speed and position via millimeter-wave radar. When it determines that the train has stopped within the no-stopping zone and the voltage difference between the two contact wire sections is less than a set value, the core controller issues a command to close the main control contactor, short-circuiting the contact wires on both sides of the insulated contact wire section (the working branch on one side of the section and the non-working branch on the other side). This converts the insulated anchor section into a non-insulated anchor section, forming a current bypass. Even if the carbon sliding plate is connected between the working and non-working branches of the contact wire, most of the current flowing through the contact wire is diverted to the short-circuit bypass, thus preventing excessive current in the electrical path "working branch → carbon sliding plate → non-working branch," which could cause localized overheating and melting of the conductor. When it is determined that the train has left the no-stopping zone and after a fixed delay, and the monitored DC current is less than a set value, the main control contactor is disconnected. The device must ensure reliable contact between the two contact wire sections when a train stops in the area.

[0070] The train status acquisition module, composed of millimeter-wave radar, uploads train status data (target speed, distance, direction of travel, category, etc.) to the communication management unit. After filtering out interference data, the system locks onto the train target and then performs a logical judgment process to determine whether the train has stopped in the no-stopping area or has left the no-stopping area. Finally, it outputs opening and closing commands to control the main control contactor.

[0071] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A subway catenary control method based on millimeter-wave monitoring, characterized in that, Includes the following steps: S1. Collect train status data and preprocess the train status data; the train status data includes the train's position, speed, distance, direction of travel, and acceleration. S2. Analyze the preprocessed train status data to locate the target train; S3. Obtain the electrical parameters of the overhead contact line, and estimate the state of the overhead contact line based on the electrical parameters; S4. Based on the location of the target train, determine whether the target train has stopped within the preset no-stopping zone; S5. Based on the status of the overhead contact line and the positional relationship between the target train and the no-stopping zone, issue instructions to control the main control contactor; Step S3 further includes: S31. Obtain the electrical parameters of the contact network, including the first segment voltage and the second segment voltage of the contact network, and calculate the voltage difference between the first segment voltage and the second segment voltage; S32. Based on the voltage difference between the first segment voltage and the second segment voltage, obtain the voltage and current at both ends of the main control contactor, and estimate the state of the contact network through the voltage and current at both ends of the main control contactor. S321. Determine whether the voltage difference across the main control contactor is less than the preset threshold. If it is, allow the main control contactor to close; otherwise, do not allow it to close. S322. Obtain the current passing through the main control contactor. Based on the current passing through the main control contactor, the contactor's rated current, and the overload factor, determine whether the contactor is overloaded. If overloaded, issue an alarm indicating the overload risk. Step S5 further includes: S51. When the target train is located in a no-stopping zone, determine whether the following conditions are met simultaneously: the target train stops and the entry time exceeds the preset allowable dwell time threshold, and the voltage difference between the first segment voltage and the second segment voltage is less than the preset allowable voltage difference threshold. If the conditions are met, a closing command is issued to close the main control contactor; if the conditions are not met, no command is issued. S52. When the target train leaves the no-stopping area, determine whether the following conditions are met simultaneously: the departure time exceeds the preset delay disconnection time, and the current of the main control contactor is lower than the preset current lower limit threshold. If the conditions are met, a disconnection command is issued to disconnect the main control contactor; if the conditions are not met, no command is issued. It also includes S6, acquiring historical action data of the main control contactor, and performing fault diagnosis and early warning based on the historical action data, specifically: S61. Obtain historical action data of the main control contactor, including action time, triggering conditions and parameters before and after the action, and establish a normal action mode library based on the historical action data. S62. Collect new action data of the main control contactor, and use machine learning algorithms to detect anomalies in the new action data based on the normal action mode library. S63. Based on the results of anomaly detection, determine the cause of the anomaly, including prolonged train stops, abnormal changes in overhead contact line parameters, or equipment failure. Based on the cause of the anomaly, issue timely warnings, prompt maintenance personnel to conduct repairs, and activate emergency response plans.

2. The subway catenary control method based on millimeter-wave monitoring according to claim 1, characterized in that, Step S1 further includes: S11. The millimeter-wave radar scans the surrounding environment at a predetermined frequency to obtain the range scanning matrix and the first reflection intensity matrix as train status data. S12. The median filtering algorithm is used to perform noise reduction filtering on the distance scan matrix and the first reflection intensity matrix. S13. The millimeter-wave radar transmits chirp signals, receives the echoes of the chirp signals and decodes them to obtain the range matrix, velocity matrix and second reflection intensity matrix.

3. The subway catenary control method based on millimeter-wave monitoring according to claim 2, characterized in that, Step S2 further comprises: S21. Based on the denoised and filtered distance scan matrix and the first reflection intensity matrix, the train outline is extracted using an edge extraction algorithm, and the train outline is fitted to obtain the train bounding box. S22. Perform range-Doppler domain detection on the distance matrix, velocity matrix, and second reflection intensity matrix to extract potential target trains; The potential target train is projected into the millimeter-wave radar coordinate system and intersected with the train bounding box to obtain the fused detection result; S23. Using Kalman filtering, perform multi-target tracking on the fused detection results, and update the position, speed, and acceleration of each train based on the tracking results.

4. The subway catenary control method based on millimeter-wave monitoring according to claim 3, characterized in that, Step S4 further includes: S41. Obtain train dispatch instructions and route topology, and determine no-stopping areas; S42. Compare the position of the target train with the no-stopping zone to determine whether the target train is located in the no-stopping zone. If it is, start the entry timer; otherwise, start the departure timer.

5. The subway catenary control method based on millimeter-wave monitoring according to claim 4, characterized in that, Step S61 further includes: S611. Obtain historical action data of the main control contactor, clean and preprocess the historical action data, and remove noisy and invalid data. S612. Extract the temporal and frequency domain features from the preprocessed historical action data; The time-domain features include action duration, action interval, rise / fall time, and overshoot; the frequency-domain features include spectrogram, dominant frequency, spectral entropy, and center frequency. S613. Standardize the time-domain and frequency-domain characteristics to eliminate the influence of dimensions; S614. Based on the time-domain and frequency-domain characteristics after eliminating the influence of dimensions, establish a normal action mode library.

6. The subway catenary control method based on millimeter-wave monitoring according to claim 5, characterized in that, Step S614 further includes: S6141. Construct a deep convolutional neural network model based on the time-domain and frequency-domain features after eliminating the influence of dimensions. S6142. Obtain typical action sample data from the main control contactor. Based on the typical action sample data, use the backpropagation algorithm to optimize the parameters of the deep convolutional neural network model. S6143. In the optimized deep convolutional neural network model, historical action data is propagated forward to extract deep feature vectors. S6144. Perform cluster analysis on the deep feature vectors to automatically discover the cluster structure inside the historical action data, with each cluster corresponding to a typical action pattern. S6145. Statistical action pattern center feature vector and boundary threshold. Based on the center feature vector and boundary threshold, construct a normal action pattern library.

7. A subway overhead contact line diverter device based on millimeter-wave monitoring, characterized in that, The subway catenary control method based on millimeter-wave monitoring according to any one of claims 1 to 6 includes a DC contact module, a monitoring module, and a train status acquisition module; wherein the DC contact module includes a DC contactor, a current transmitter, and a voltage transmitter; the monitoring module includes a communication management unit, a 4G communication module, a core controller, an anti-condensation device, and an operation panel; and the train status acquisition module includes at least two sets of millimeter-wave radars.

Images