A coordinated monitoring device and system for a traction electric drive and safety monitoring equipment

Abstract

The application provides a kind of coordinated monitoring device and system of rail train traction electric drive and safety monitoring equipment using edge computing.The edge computing device includes sensor module, positioning module, edge host module and storage module, various communication interface module;Sensor module pre-processes sensor signal to obtain normal state quantity and or abnormal quantity, which is sent to edge host module through internal signal transmission bus;Edge host module includes DCPU, MCPU, data storage array, flash memory and bus interface;DCPU extracts features for normal state, analyzes and distinguishes abnormal state and stores the results;MCPU realizes the initialization of edge computing device, various bus transmission and management of each module.The device can exchange data with other edge computing devices through Ethernet to form a coordinated monitoring system, realize the coordinated monitoring of rail train traction electric drive and safety detection equipment, and perform fault prediction and health management.

Description

Technical Field

[0001] This invention relates to the field of rail train monitoring technology, and in particular to a collaborative monitoring device and system for traction electric drive systems and safety monitoring system equipment utilizing edge computing. Background Technology

[0002] As a high-capacity public transportation tool, the safe operation of rail trains is of paramount importance. The traction drive system involves multiple fields such as power electronics, motors, and microcomputer control, and mainly includes: traction converters and control devices (TCUs), network communication and control devices, monitoring devices, and traction motors. The traction drive and its control system, to a certain extent, reflect the advanced technological level of rail train equipment.

[0003] Traction drive control typically employs a network control system integrating train monitoring, diagnostics, and control, collectively known as TCN (Train Communication Network), used for the control, detection, and diagnostics of traction drive equipment in rail trains. TCN includes WTB (Wired Train Bus) for train-level network control and MVB (Multi-function Vehicle Bus) for vehicle-level network control. WTB connects the equipment in dynamically assembleable vehicles, enabling the transmission of process and message data. MVB connects fixed equipment within the vehicle, enabling the transmission of control commands and status data. Condition monitoring of the traction drive system is also based on TCN, playing a crucial role in ensuring safe train operation.

[0004] Existing monitoring systems for rail trains fall into two main categories: condition monitoring for traction drive control and condition monitoring for operational safety. Condition monitoring for traction drive control includes monitoring transformer secondary voltage, motor current, intermediate DC voltage, and motor speed. These systems are connected to the MVB or directly to the TCU. Condition monitoring for traction drive control primarily involves monitoring voltage and current, including monitoring the input voltage, DC voltage, stator three-phase voltage, output current, and speed signal of the traction converter, estimating motor flux and torque. Direct torque control without speed may not require a speed signal. Vector control systems monitor three-phase voltage, three-phase current, and motor speed for vector control. Both the monitoring sensors and processing operate within the traction converter.

[0005] Safety-oriented condition monitoring includes smoke and fire alarm systems, bearing temperature monitoring systems, instability monitoring systems, and speed monitoring systems. The bearing temperature monitoring system deploys temperature sensors on bogie bearings, axle box bearings, and traction motor bearings to monitor bearing temperature in real time during train operation. If the temperature exceeds the limit, an automatic alarm is triggered or a train is stopped. The instability monitoring system deploys acceleration sensors on the bogies to monitor lateral vibration acceleration and determine the possibility of instability. Sensor data is transmitted to the control unit, where the acceleration is analyzed. When signs of bogie instability are detected, an alarm is issued, and deceleration measures are taken. The speed monitoring system installs speed sensors on the front end caps around the axles. These sensors provide speed signals to both the braking system and the ATP (Automatic Train Protection) system to control train speed.

[0006] Current TCN data transmission rates are relatively low, designed primarily for control commands and limited status data transmission. They haven't considered the broadband data transmission needs of intelligent and driverless train technologies. Current monitoring systems mainly detect anomalies, resulting in large data transmission volumes from sensors. The current train communication network cannot handle this volume, limiting the increase in the number of sensors deployed and leading to a limited number and types of connected sensors. Newer rail trains have adopted a TCN+Ethernet approach, with TCN used for traction drive system control and Ethernet for monitoring equipment status.

[0007] Currently, the shortcomings of existing collaborative monitoring methods between electric drive equipment and safety monitoring equipment in rail trains include:

[0008] 1) Inability to achieve collaborative monitoring: The existing monitoring system for rail trains is relatively independent, and different monitoring quantities cannot interact, making it impossible to collaboratively monitor and detect problems in the safe operation of equipment.

[0009] Current independent monitoring methods cannot promptly detect the deterioration process and faults in the traction electric drive system, nor can they coordinate monitoring of equipment operation. For example, when the motor current reaches or exceeds the rated current, and the shaft temperature approaches the fault threshold, a rapid failure will occur, which current methods cannot detect. Existing monitoring methods cannot form a collaborative monitoring system, resulting in a high rate of missed detections.

[0010] 2) Lack of suitable edge computing devices for collaborative monitoring: Current monitoring systems for rail train traction electric drive systems are designed to detect anomalies, resulting in limited host computing power, weak data processing capabilities, and few network interfaces. Onboard circuit boards have small storage capacity and a limited number of connected sensors. A large amount of sensor data needs to be transmitted through the onboard network, but due to limited bandwidth, the number and types of connected sensors are also limited. Existing monitoring systems cannot interact with each other onboard circuit boards, hindering collaborative monitoring.

[0011] 3) Lack of intelligent online analysis methods: Existing monitoring methods judge anomalies by using fixed thresholds, which cannot record the status of electric drive equipment and its evolution, and cannot realize the complex decision-making process on the equipment side during the journey, resulting in high false alarm rate and high missed detection rate.

[0012] Intelligent and driverless trains urgently require Prognostics and Health Management (PHM) systems, necessitating the deployment of additional sensors on various devices, as well as the introduction of previously unused sensor types to monitor equipment status. Currently, sensor data is transmitted directly to the onboard host via the onboard network. With the increase in sensor monitoring nodes, the amount of data transmitted increases significantly, exceeding the real-time performance and network capacity requirements. The current train communication network cannot handle the increased data transmission volume, limiting the expansion of sensor deployment.

[0013] A PHM (Property Management) system needs to understand the normal operating status data and characteristics under different operating environmental conditions, as well as the patterns of equipment status evolution. Existing monitoring systems lack these functions and are difficult to apply to condition-based equipment maintenance and repair.

[0014] Current status monitoring of electric drive systems in rail trains focuses on detecting anomalies. Fixed thresholds are used as the standard for anomaly detection; for example, axle temperature monitoring systems only measure temperature and compare it to a fixed threshold. Voltage and current systems used for train control, etc., trigger an alarm when a set threshold is exceeded. This approach results in a high false alarm rate. Summary of the Invention

[0015] Embodiments of the present invention provide a collaborative monitoring device and system for a traction electric drive system utilizing edge computing to achieve effective fault monitoring of the traction electric drive system.

[0016] To achieve the above objectives, the present invention adopts the following technical solution.

[0017] According to one aspect of the present invention, an edge computing device is provided, comprising: a power module, a sensor module, an edge control module, and a storage module interconnected via an internal signal transmission bus;

[0018] The sensor module includes a sensor, conditioning and transformation, processing and judgment, storage, and communication. Different conditioning and transformation circuits are used depending on the sensor's output signal to preprocess and threshold-determine the sensor signal, obtaining the normal state quantity and / or abnormal quantity of the monitored object. These quantities are then sent to the edge control module via an internal signal transmission bus interface. The edge control module includes a data processing central processing unit (DCPU), an edge device management central processing unit (MCPU), a data storage array, a flash memory, and a bus interface. The DCPU aggregates data from one or more sensor modules and uses a feature extraction algorithm to obtain the normal state's characteristic quantities. The system reflects the state changes that evolve under normal conditions. It learns and models normal data or normal data features to obtain the change patterns of normal states. It analyzes and further distinguishes abnormal quantities by combining normal data features and their evolutionary state change rates. It performs fault diagnosis on abnormal quantities and determines the threshold revision amount for frequently occurring faults based on the fault diagnosis results. The MCPU realizes the initialization of the edge computing device, internal bus transmission, management of various communication modules and each module, and collaborative address allocation of different edge computing devices during the initial operation of the train. The DCPU and MCPU exchange relevant data through Flash memory, and the DCPU and MCPU control and interact with other modules of the edge computing device through the internal bus interface.

[0019] The storage module is used to configure the storage chip array, manage the CPU and bus interface, store normal operation characteristics and their evolution characteristics, key data points, abnormal data, fault diagnosis models and parameters, fault diagnosis results, thresholds, data characteristics and key data points of multi-source collaborative monitoring, and interact with the edge master control module through the internal data transmission bus.

[0020] The power module is used to draw power from the vehicle's power supply, convert it to DC power through a DC-DC converter, and then supply power to other modules via the power bus.

[0021] Preferably, the device further includes: a positioning module, a multi-functional communication bus MVB interface module, an Ethernet interface module, and a wireless interface module;

[0022] The MVB interface module is used as the main interface for edge computing devices to access the train communication network (TCN), to receive control commands from the central control unit (CCU), and to report key anomalies to the CCU through the TCN.

[0023] The Ethernet interface module is used as a communication network connecting the edge computing device and the Train Monitoring Center (TMC) to enable interaction between edge computing devices.

[0024] The wireless interface module is used to control the access, networking and address allocation of the vehicle-mounted wireless sensors, transmit data with the ground monitoring data center, and send the processing results of the vehicle-mounted edge computing device to the ground monitoring center.

[0025] The positioning module is used to provide location data to the edge computing device using a satellite positioning chip, and to provide location data to the edge master control module through an internal signal transmission bus.

[0026] According to one aspect of the present invention, a collaborative monitoring system for traction electric drive and safety monitoring equipment based on the aforementioned edge computing device is provided, comprising: installing an edge computing device and a sensor module in each car of a rail train for collaborative monitoring of multi-source physical quantities of traction electric drive equipment and safety monitoring equipment; the edge computing devices of each rail train are interconnected via Ethernet and connected to the Train Monitoring Center (TMC); the TMC receives relevant data, performs model training and parameter determination, determines various thresholds, performs fault prediction and health management, and publishes results and tasks.

[0027] The edge computing device uses the MVB interface to access the TCN through the vehicle gateway, connects to the CCU, reports anomalies to the CCU, and receives control commands sent by the CCU.

[0028] Preferably, the collaborative monitoring system includes a vehicle layer, a side layer, and an end layer;

[0029] The end layer includes sensors and sensor modules, which are installed at the train traction electric transmission equipment and / or safety monitoring equipment. The type and number of sensors are selected according to the characteristics of the physical quantities of the monitored equipment. The sensors collect data on the status of the monitored equipment. The sensors are connected to the sensor modules. The sensor modules preprocess the data, perform threshold discrimination and early warning, and send the discrimination results to the edge computing device. The sensor modules also receive instructions from the edge computing device. Depending on the monitoring task and deployment location, the sensor modules are deployed together with the sensors at the monitoring equipment, or the sensors are deployed at the monitoring equipment and the sensor modules are deployed within the edge computing device.

[0030] The edge layer includes edge computing devices installed on the train. These devices receive data on traction motors, bearing temperature, vibration, and speed deployed on the bogies. They aggregate data transmitted from one or more sensor modules, use feature extraction algorithms to obtain normal state features and state change quantities reflecting the evolution of normal states, learn and model normal data or normal data features to obtain the change patterns of normal states, and store these patterns. They analyze and further distinguish abnormal quantities by combining normal data features and their evolving state change rates, perform fault diagnosis on abnormal quantities, determine threshold revision amounts for frequently occurring faults based on the fault diagnosis results, and store the results. Different edge computing devices exchange data with each other via Ethernet and receive control commands from the TMC in the train layer.

[0031] The vehicle deck includes a TMC (Traffic Management Center) that connects edge computing devices in each carriage via Ethernet to form an onboard cooperative monitoring system. This system interacts with the ground monitoring center via GPRS (General Packet Radio Service) or a mobile communication network. It performs model training and parameter determination for the same physical quantity; receives and compares monitoring and analysis results of the same physical quantity from different edge computing devices; performs model training and parameter determination for collaborative monitoring of different physical quantities; receives and compares monitoring and analysis results of different physical quantities under collaborative status monitoring; determines anomaly and dynamic thresholds for different track physical quantities; and uses algorithms related to equipment operation fault prediction and health management based on the results from the vehicle's edge computing devices to determine the equipment's status. Finally, it receives instructions from the ground monitoring center.

[0032] Preferably, the collaborative monitoring system is used to learn the features of normal axle temperature data through a Long Short-Term Memory (LSTM) network model via the vehicle floor TMC, and to establish a mapping relationship between historical axle temperature data and current data through a deep learning algorithm. The specific processing includes:

[0033] LSTM introduces a new internal state c in each cell. t This involves linear, cyclic information transfer, outputting information to the external state h of the hidden layer. t By introducing the forget gate f t Input gate i t and output gate o t Control the amount of information;

[0034] The vehicle-level TMC (Transmission Control Center) standardizes and performs sliding window processing on the normal axle temperature data. The preprocessed axle temperature sequence is then input into an LSTM (Laser-to-Shipment Module) model for training. The training dataset comes from historical normal axle temperature data stored in the TMC or from normal axle temperature data transmitted from edge computing devices in the side layers. This process establishes a mapping between historical and current axle temperature data.

[0035] temp t =f(temp) t-1 temp t-2 ,…,temp t-m (1)

[0036] In the formula, temp t represents the shaft temperature at time t; f(·) represents the nonlinear mapping relationship, referring to the black-box model without explicit expression; m is the length of the historical shaft temperature data of the model input variable.

[0037] Preferably, the DCPU in the edge computing device of the edge layer performs joint detection of multiple anomalies in the shaft temperature data by analyzing the shaft temperature residual sequence. The specific processing includes:

[0038] The edge computing device's DCPU downloads a pre-trained mapping relationship between historical axle temperature data and current data from the vehicle floor TMC. It then inputs the historical axle temperature data into this pre-established mapping relationship to obtain the predicted axle temperature value for the current moment. Compared with the actual shaft temperature data at the current moment y t The residual sequence E is obtained by taking the difference, and is expressed as:

[0039] E = [e1, e2, ..., e t-1 ,e t (2)

[0040] Among them, e t The residual at time t is expressed as:

[0041] Secondly, to calculate the anomaly detection threshold at different times, the edge computing device's DCPU divides the residual sequence into non-overlapping subsequences of length h. For time t, the residual subsequence is represented as:

[0042]

[0043] Where h represents the number of residuals used for analysis; t is an integer multiple of h;

[0044] The DCPU of the edge computing device obtains the probability density function of each residual subsequence using the KDE method. right Integrating yields the cumulative distribution function of each residual subsequence, expressed as:

[0045]

[0046] Take the sequence value corresponding to the 0.05 quantile of the desired distribution. As the lower limit, the sequence value corresponding to the 0.95 quantile is used. This serves as an upper limit and a non-parametric dynamic threshold for anomaly detection. If the residual exceeds this non-parametric dynamic threshold, the data is identified as a point anomaly. for The inverse function;

[0047] Using the detected point anomalies as the dividing points, the residual sequence is re-divided into multiple subsequences of different lengths. The DCPU of the edge computing device applies an improved Local Outlier Factor (ILOF) method to the residual subsequences. When the ILOF score is higher than the non-parametric dynamic threshold for anomaly detection, the subsequence and its dividing point are identified as collective anomalies. Conversely, the subsequence is normal, and the dividing point of the subsequence is still identified as a point anomaly.

[0048] Preferably, the DCPU in the edge computing device performs vibration signal feature learning and anomaly detection methods for the motor bearing. It uses fractional-order scattering entropy (FrFDE) to reflect changes in the working state of the traction transmission equipment and vibration signals to reflect changes in the internal structure. The specific processing steps include:

[0049] The charge quantity of the end-layer accelerometer installed near the vibration source of the equipment is signal-conditioned using a charge amplifier. The obtained acceleration detection quantity is sent to the vehicle-mounted edge master control module of the side layer through the internal bus interface module and stored in the storage array. The DCPU performs real-time sliding window processing on the data with a window length of N and a step size of L. Each sliding window produces a one-dimensional vibration signal time series x = {x1, x2, ..., x...} with a length of N. N The time series is mapped to y = {y1, y2, ..., y3} using the normal cumulative distribution function. N};

[0050]

[0051] In the formula: γ and σ are the mean and standard deviation of the time series x, respectively. The following linear transformation maps y to {1,2,...,c}:

[0052]

[0053] In the formula: round() is the rounding function; c is the number of categories; It is the i-th element in class c;

[0054] For sequence Perform phase space reconstruction with embedding dimension m and delay factor d to obtain a new reconstruction vector. as follows:

[0055]

[0056] Then each embedding vector Mapping to the scattering pattern of fluctuations

[0057]

[0058] Assigned to each time series The total number of possible scattering patterns is (2c-1). m-1 ;

[0059] For each possible scattering pattern Calculate the relevant probability of its occurrence. As shown in the following formula:

[0060]

[0061] Introducing fractional operators to capture more dynamic properties of vibration signals, namely the Riemann-Liouville operator:

[0062]

[0063] D α f(x) = D n I n-α f(x) (14)

[0064] In the formula: I α D represents the α-order RL integral. α Let Γ(·) be the α-order RL derivative, and Γ(·) be a class of functions that extend the factorial function to real and complex numbers. Calculate the fractional wave scattering entropy for different α-orders according to the definition of Shannon entropy:

[0065]

[0066] Preferably, the collaborative monitoring system is used to monitor vibration signals for anomalies based on fractional-order wave spread entropy and cumulative control charts, and the specific processing steps include:

[0067] Vibration signals of the electric drive equipment to be monitored are collected using the end-layer accelerometer and signal conditioning circuit, and transmitted to the edge master control module of the side layer via the internal bus;

[0068] The DCPU in the main control module of the edge computing device calculates the fractional order wave distribution entropy of the vibration signal under each sliding window in real time, tracks the state changes of the equipment, and stores the calculation results in the storage module.

[0069] The DCPU uses the Cumulative and Control Chart (CUSUM) algorithm to track the trajectory of the fractional-order fluctuation spread entropy sequence. It uses a threshold to determine whether the monitoring data is abnormal. When the equipment is abnormal, it promptly reports the detected abnormal points to the vehicle monitoring center. After the vehicle monitoring center confirms the abnormality, it reports it to the CCU of the vehicle level through the TCN.

[0070] The expression for the CUSUM algorithm is:

[0071]

[0072] x(t) represents the process detection data, g + and g - These are the cumulative sums of positive and negative changes, respectively. The initial value of parameter h is zero, i.e., g + (0)=g - (0) = 0, given a threshold T, if g + >T or g - >T, the CUSUM algorithm detects anomalies and promptly sends alarms to the operator.

[0073] Preferably, the collaborative monitoring system is used for multi-source collaborative monitoring of motor status based on the characteristics between different signals from multiple sources. The specific processing steps include:

[0074] Using the sensor module to measure temperature X temp Current X c Torque X tor The data is collected and transmitted to the respective edge master control modules via the bus interface module;

[0075] In the edge master control module, the DCPU first performs a preliminary judgment on the data through FFT analysis or thresholding. If no anomalies are found, the data is simply monitored and displayed. If anomalies are found, the data undergoes further detailed processing, such as X-raying the stator temperature at different sampling frequencies. temp Stator current X c Torque signal X tor The data is divided into equal-length segments. The length k of each segment of the current data is determined by comparing the diagnostic results. The time t of each data segment is determined based on the value of k. The duration t of each data segment of the temperature and torque signals is also determined by t. Thus, the number of data points contained in each data segment can be determined. The data of the three different signals input to the convolutional neural network are collected in the same time period.

[0076]

[0077] Normalize each data segment separately, where t = 1, 2, ..., k:

[0078]

[0079] In the DCPU of the edge master control module, for stator temperature X temp Stator current X c Torque signal X torThree different signals were used for feature extraction using a CNN with identical parameters except for the convolution operation parameters. The feature extraction process is as follows:

[0080] Step 1: Assuming the input signal is x, perform a convolution operation on x:

[0081] y i l =f(b k l +conv1D(W l ,S i l-1 (19)

[0082] Where i = 1, 2, ..., C represents the number of convolution kernels. This represents the convolution result of the i-th convolution kernel with respect to the input x. It is the output neuron of the (l-1)th convolutional pooling module. W represents the k-th output bias corresponding to the l-th convolutional layer. l denoted by , f(.) represents the weights of the l-th convolutional layer, f(.) is the activation function, and conv1D represents a one-dimensional convolution;

[0083] Step 2 uses max pooling for dimensionality reduction and batch normalization of the output features to obtain C features N. l =[n1 l n2 l ,...,n C l Each feature has a dimension of M×1, where max pool represents max pooling and BN represents batch normalization;

[0084] p i l =maxpool(y i l )

[0085] n i l =BN(p i l (20)

[0086] Step 3 uses global average pooling and a fully connected layer to evaluate the importance of different features. `global_avepool` represents global average pooling, `FC` represents a fully connected layer, and `fc` represents the full connection layer. 2 This represents C evaluation vectors, each corresponding to one of the C features;

[0087] ga l =global_avepool(N l )

[0088] fc1 =FC(ga) l (C / r)

[0089] fc 2 =FC(fc) 1 (C) (21)

[0090] Where r represents the dimensionality reduction ratio, and the C features are respectively compared with the evaluation vector fc 2 Multiplying the C elements in the formula yields C weighted features Y = [y1] 1 ,y2 1 ,…,y c 1 ];

[0091] In their respective edge computing devices, for the three sets of M×1 dimensional features obtained by the three convolutional neural networks, it is assumed that they are F1, F2, and F3 respectively.

[0092]

[0093] Transpose F1, F2, and F3 respectively, and then connect them using the concatenation function Conc, i.e.:

[0094] F=Conc(Trans(F1),Trans(F2),Trans(F3)) (23)

[0095] The edge control module uses the Ethernet interface module to send the extracted features to the vehicle monitoring center. The vehicle monitoring center integrates various features, uses the Softmax function to train the diagnostic model and determine the model parameters, maps the elements in the output vector to the range (0, 1) and the sum of the elements is 1, which is used as the probability of each fault category. The category with the highest probability is used as the final classification result.

[0096]

[0097] Where exp(z) i ), j = 1, 2, ..., N represents the j-th input to the Softmax function, and N is the total number of categories;

[0098] After the vehicle monitoring center completes the modeling, it transmits the model to the corresponding edge computing device through the Ethernet interface module. The edge computing device uses the model to perform fault diagnosis and then transmits the fault to the central control unit through TCN. The vehicle monitoring center sends relevant fault data to the central control unit through the vehicle gateway based on the collaborative monitoring results and fault prediction results. In the central control unit, the traction system is operated to reduce speed, stop, or switch according to the predetermined program.

[0099] Preferably, the collaborative monitoring system is used to obtain dynamic thresholds for multiple physical quantities through a collaborative learning algorithm, and to perform fault prediction and health management for monitoring equipment. Different dynamic judgment thresholds are determined according to changes in the operating environment to obtain abnormal monitoring results.

[0100] As can be seen from the technical solutions provided by the embodiments of the present invention above, the present invention provides a customizable edge computing device for collaborative monitoring and diagnosis of traction transmission and safety detection equipment for high-speed trains. It can connect to different types of sensors used for detection and, through an internal bus, forms a unified architecture and customizable functionality edge computing device with a CPU unit, storage unit, and communication unit. This enables the access, feature extraction, and anomaly detection of multiple types of sensors.

[0101] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0102] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0103] Figure 1 This is a schematic diagram of the logical structure of an edge computing device provided in an embodiment of the present invention.

[0104] Figure 2 This is a logical structure diagram of a sensor module in an edge computing device provided in an embodiment of the present invention.

[0105] Figure 3 This is a schematic diagram of the structure and logic of an edge master control module provided in an embodiment of the present invention;

[0106] Figure 4 This is a schematic diagram of the logical structure of a collaborative monitoring system for train traction transmission equipment and safety monitoring equipment provided in an embodiment of the present invention;

[0107] Figure 5 This is a schematic diagram illustrating the work tasks completed at each layer of the vehicle-edge-end system, as provided in an embodiment of the present invention.

[0108] Figure 6 A schematic diagram of an LSTM cyclic cell structure provided in an embodiment of the present invention;

[0109] Figure 7 This is a schematic diagram of a normal shaft temperature data feature learning process provided in an embodiment of the present invention;

[0110] Figure 8 A flowchart of an intelligent detection algorithm for abnormal shaft temperature data provided in an embodiment of the present invention;

[0111] Figure 9 A graph showing the change of fractional fluctuation scattering entropy under different orders α and different probabilities p is provided for an embodiment of the present invention.

[0112] Figure 10 This is a flowchart of a multi-source collaborative monitoring method provided in an embodiment of the present invention. Detailed Implementation

[0113] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0114] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0115] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0116] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0117] The research object of this invention is equipment used in traction electric drive systems and safety monitoring systems for high-speed railways, urban rail transit, intelligent trains and driverless trains.

[0118] A schematic diagram of the logic structure of an edge computing device provided in this embodiment of the invention is shown below. Figure 1 As shown, edge computing devices are the foundation for collaborative monitoring. Through circuit design, edge computing devices can connect to different numbers and types of sensors. They employ powerful central processing units, have ample storage space, and multiple unified network interfaces to enable data interaction between edge computing devices, thus forming a collaborative monitoring system. These edge computing devices, through internal buses, CPU units, storage units, and communication units, constitute a unified and customizable collaborative monitoring system, enabling the access, feature extraction, and anomaly detection of multiple sensor types.

[0119] The edge computing device includes a power module, sensor module, positioning module, edge master control module, storage module, MVB interface module, Ethernet interface module, and wireless interface module. The edge computing device adopts a modular access approach, interconnecting each other via an internal signal transmission bus. This internal signal transmission bus can be either a RS-485 bus or a CAN bus. When using a RS-485 bus, the edge master control module can employ the ModBus protocol. This edge computing device allows for flexible addition or removal of modules based on the characteristics of the monitoring equipment and user needs, making it customizable hardware.

[0120] The power module draws power from the vehicle's DC110V power supply, and after DC-DC conversion, obtains the DC power required and suitable for each part of the edge computing device, and supplies power to other modules through the power bus.

[0121] The positioning module uses a GPS chip or a Beidou chip to provide location data for the edge computing device. The positioning module contains a microcontroller that provides location data to the edge master control module through an internal signal transmission bus.

[0122] A schematic diagram of the logic structure of a sensor module provided in this embodiment of the invention is shown below. Figure 2As shown. The sensor module mainly includes sensors, conditioning and conversion, processing and judgment, storage, and communication. Different sensors are used depending on the monitoring equipment. Sensor types are classified according to the needs of the monitored object, mainly including analog output, digital or switch output, 485 bus output, and custom protocol output. Different conditioning and conversion circuits are used depending on the sensor's output signal. Analog output sensors require conditioning and analog-to-digital conversion circuits before inputting to the sensor module's CPU; digital / switch output sensors require level conversion circuits before inputting to the CPU; 485 bus output sensors require 485 interface circuits before inputting to the CPU; and sensors with custom protocol outputs either require level conversion circuits or direct input to the CPU based on their output level characteristics. The sensor module's CPU processes the sensor signals to obtain the sensor's detected values, which are then sent to the vehicle-mounted edge control module via the internal signal transmission bus interface. The sensor module's CPU can be a commonly used 51-core microcontroller, performing only preliminary preprocessing of the sensor data; the storage capacity can be relatively small. The sensor module is similar to the current monitoring system setup and can be deployed near the monitoring equipment according to the actual situation. Depending on the monitoring task and deployment location, the sensor module can be deployed together with the sensor at the monitoring equipment, or the sensor can be deployed at the monitoring equipment. The sensor module is deployed within the edge computing device.

[0123] Figure 3 This is a schematic diagram of the structural logic of an edge control module provided in an embodiment of the present invention. The edge control module includes a DCPU (Data Processing Central Processing Unit), an MCPU (Management Central Processing Unit), a data storage array, a Flash memory, and a bus interface, etc. Figure 3As shown. After preprocessing the data from the sensor module, the DCPU uses feature extraction algorithms to obtain the feature quantities of the normal state and the state change quantities reflecting the state evolution. It learns and models the normal data or normal data features to obtain the change law of the normal state. It uses thresholds to determine whether the monitoring data is abnormal, thereby determining whether the equipment is abnormal. It can also use various intelligent algorithms and variable thresholds to identify abnormal points. The normal features and their description models, as well as abnormal points, are stored in the data storage array. If the storage array of the edge master control module is full, the data will be transmitted to the extended storage module for calibration storage through the internal signal transmission bus. When collaborative monitoring is required, the DCPU interacts with other edge master control modules through a unified network interface, or interacts with the on-board central control unit (CCU) to obtain the required monitoring data. It performs collaborative anomaly judgment on the current monitoring data from multiple sources, which can realize dynamic threshold judgment or intelligent anomaly judgment. The MCPU of the edge master control module is used to manage the initialization of the edge computing device, the management of internal bus transmission signals, the management of various communications, the management of each module, and the collaborative address allocation of different edge computing devices during the initial operation of the train. The DCPU and MCPU exchange data via Flash memory. They also control and interact with other modules of the edge computing device through an internal bus interface.

[0124] The storage module is an extended storage unit of the edge computing device. It can be configured with a large number of storage chips to provide more storage space for normal operation status and its changes, anomalies, multi-source collaborative monitoring data characteristics, etc. The storage module sets up and manages the CPU, storage array and bus interface, and interacts with the edge master control module through the internal data transmission bus.

[0125] The positioning module uses a GPS chip or a Beidou chip to provide location data for the edge computing device. The positioning module contains a microcontroller that provides location data to the edge master control module through an internal signal transmission bus.

[0126] The MVB interface module is the main interface for edge computing devices to access the TCN. Since most of the data from edge computing devices is transmitted via in-vehicle Ethernet, the function of the MVB interface module is to receive control commands from the in-vehicle CCU and report critical anomalies to the CCU through the TCN.

[0127] The Ethernet interface module is the main communication network connecting the edge computing device to the TMC (Train Monitoring Center). Interaction between edge computing devices also occurs through the Ethernet interface module.

[0128] The wireless interface module supports one or more types of wireless communication, such as ZigBee, NBIoT, or GPRS, and can be configured as needed. Its primary function is to receive data from wireless sensors that may be used later. The wireless interface module also controls the networking and address allocation of the vehicle-mounted wireless sensors. Furthermore, it transmits data to the ground monitoring data center, sending the processing results from the vehicle-mounted edge computing device to the ground monitoring center.

[0129] The edge computing device is modular and flexible, allowing for different combinations or parameters to be set according to the type and characteristics of the monitoring sensors. The need for a wireless communication module can be determined based on the configuration requirements of the edge computing device. The customizable and simplified configuration of the edge computing device can meet the needs of different train carriages and monitoring equipment.

[0130] By utilizing edge computing devices to collaboratively monitor multi-source physical quantities of rail train traction drive equipment and safety monitoring equipment, a collaborative monitoring system is constructed, such as... Figure 4 As shown. Each train is equipped with an edge computing device and sensor modules based on the characteristics of the monitoring equipment. The edge computing device connects to the TCN via the vehicle gateway using an MVB interface, and is connected to the onboard central control unit (CCU) to report anomalies and receive instructions. The edge computing devices of each train are interconnected via Ethernet and connected to the TMC. The TMC receives relevant data and publishes results and tasks. The TMC can be configured according to existing systems as needed, or it can be reconfigured according to the train monitoring tasks.

[0131] The collaborative monitoring system of the rail train traction transmission and safety detection equipment in this embodiment of the invention can be described in three layers. A schematic diagram of the work tasks completed at each layer (vehicle-side-end) provided in this embodiment of the invention is shown below. Figure 5 As shown. The edge layer mainly refers to the sensor modules, which are installed at the train equipment. The sensor modules select the type and number of sensors based on the characteristics of the physical quantities of the monitored equipment, collect data on the status of the monitored equipment, preprocess the data, perform threshold discrimination and early warning, and send the discrimination results to the edge computing device. The edge layer sensor modules also receive instructions from the edge computing device, including the determination threshold and dynamic threshold for anomaly detection, and the settings for data acquisition frequency and output transmission volume.

[0132] The edge layer consists of edge computing devices. Two edge computing devices are installed in each car, receiving data on the voltage and current of the traction motors deployed on the bogies, as well as the bearing temperature, vibration, and speed of the traction motor stators and bearing gearboxes. Depending on the needs of condition monitoring, the edge computing devices may connect to one or more sensor modules. The edge computing devices aggregate and process the data transmitted from the sensor modules, extract features of normal operation patterns and their evolution, identify and analyze abnormal states based on these patterns, and diagnose any transmitted anomalies to determine if they are definitive anomalies. The edge computing devices store the analysis results in storage modules and determine the data to be transmitted to them based on the train monitoring center on the car level. The edge computing nodes receive control commands from the train monitoring center on the car level, including confirmation of fault diagnosis results and anomaly judgment thresholds, characteristics of normal state patterns, storage capacity, and management of storage modules.

[0133] The train level serves as the Train Monitoring Center (TMC). The TMC's structure is identical to that of the edge computing devices, connecting to and managing the edge computing devices across the entire train via Ethernet. The TMC receives and compares monitoring and analysis results of the same physical quantity from different edge computing devices. Based on this, it trains models and determines parameters for intelligent learning and mining of the same physical quantity. It also receives and compares monitoring and analysis results of different physical quantities under collaborative state monitoring. Based on this, it trains models and determines parameters for intelligent learning and mining of multiple physical quantities. It determines anomaly thresholds and dynamic thresholds for different track physical quantities. Based on the results from the vehicle's edge computing devices, it performs equipment fault prediction and health management, assesses equipment status and estimates lifespan, and proposes preliminary maintenance and repair plans to meet the monitoring system requirements of intelligent trains and driverless trains. The TMC interacts with the ground monitoring center via wireless GPRS (General Packet Radio Service) or 5G mobile communication networks. TMC receives instructions from the ground monitoring center, including train fault prediction results, anomaly judgment criteria, collaborative model structure and parameters, monitoring model structure and parameters, and status monitoring tasks required by the ground monitoring center.

[0134] Building a collaborative monitoring system enables simultaneous monitoring of multiple physical quantities on the same monitoring equipment, and allows for the fusion of these quantities to determine the equipment's status, including normal operation, its evolution, equipment degradation, and potential malfunctions. This improves the reliability of current rail train condition monitoring results, meets the requirements of PHM (Prognostics and Health Management), and adapts to the monitoring system needs of intelligent and driverless trains. The collaborative monitoring method is implemented in edge computing devices alongside the construction of the collaborative monitoring system. This is illustrated using the monitoring of traction motor bearing temperature and gearbox bearing temperature as examples.

[0135] The application process of the collaborative monitoring system for rail train electric transmission equipment and safety monitoring equipment provided in the above embodiments of the present invention includes:

[0136] 1: Learning the characteristics of normal shaft temperature data.

[0137] The feature learning process for normal axle temperature data is completed in the vehicle layer TMC (Transportation Management Center). The dataset required for model training comes from historical normal axle temperature data stored in the TMC or from normal axle temperature data transmitted from edge computing devices in the side layers. The feature learning algorithm uses an LSTM (long short-term memory) network model. Figure 6 This invention provides a schematic diagram of an LSTM cyclic cell structure. The LSTM introduces a new internal state c in each cell. t Perform linear, cyclic information transfer while simultaneously outputting information to the external state h of the hidden layer. t By introducing the forget gate f t Input gate i t and output gate o t Controlling the amount of information improves the memory capacity of LSTM.

[0138] Figure 7 This is a schematic diagram of a normal axle temperature data feature learning process provided by an embodiment of the present invention. The specific processing steps include: first, standardizing the normal axle temperature data; second, performing a sliding window operation to adapt its size to the input of the LSTM model; third, inputting the preprocessed axle temperature sequence into the LSTM model for training; and finally obtaining the mapping relationship between historical axle temperature data and current data. The above feature learning process is based on historical normal axle temperature data stored in the vehicle layer TMC or normal axle temperature data transmitted from the edge computing device and is completed using the computing performance of the vehicle layer TMC. The final result of axle temperature data feature learning is a black-box model reflecting the relationship between historical and current axle temperature data, i.e.:

[0139] temp t =f(temp) t-1 temp t-2 ,…,temp t-m (1)

[0140] In the formula, temp t represents the shaft temperature at time t; f(·) represents the nonlinear mapping relationship, which in this invention refers to a black-box model without explicit expression; m is the length of the model input variable (i.e., historical shaft temperature data).

[0141] 2: Intelligent detection of abnormal shaft temperature data.

[0142] Figure 8This invention provides an intelligent detection algorithm for abnormal axle temperature data, which runs on the DCPU of an edge computing device at the edge layer. The key to abnormal axle temperature data detection lies in jointly detecting various anomalies in the axle temperature data by analyzing the axle temperature residual sequence. Here, the axle temperature residual sequence refers to the difference between the current axle temperature data and the predicted axle temperature data at the current moment, obtained through a pre-established mapping relationship between historical axle temperature data and current data; the various anomalies refer to point anomalies and subsequence anomalies commonly observed in the axle temperature data. The general process of intelligent detection of abnormal axle temperature data is as follows: the DCPU of the edge computing device first downloads the pre-trained mapping relationship between historical axle temperature data and current data from the vehicle layer TMC, and then performs the diagnostic process in the DCPU of the edge computing device.

[0143] First, the edge computing device's DCPU downloads a pre-trained mapping relationship between historical axle temperature data and current data from the vehicle's TMC. It then inputs the historical axle temperature data into this pre-established mapping relationship to obtain the predicted axle temperature value for the current moment. Compared with the actual shaft temperature data at the current moment y t The residual sequence E is obtained by taking the difference, which can be expressed as:

[0144] E = [e1, e2, ..., e t-1 ,e t (2)

[0145] Among them, e t The residual at time t is expressed as:

[0146] Secondly, to calculate the anomaly detection threshold at different times, the edge computing device's DCPU divides the residual sequence into non-overlapping subsequences of length h. For time t, the residual subsequence can be represented as:

[0147]

[0148] Where h represents the number of residual values ​​used for analysis; t is an integer multiple of h. The DCPU of the edge computing device performs KDE (kernel density estimation) on each residual subsequence to nonparametrically fit its distribution, using the 0.95 quantile and 0.05 quantile of the obtained distribution as the upper and lower limits, and using them as the nonparametric dynamic threshold for anomaly detection. The KDE method is a nonparametric probability density estimation method, which is suitable for modeling the probability density function (PDF) of data with unknown distributions.

[0149] The probability density function of each residual subsequence can be obtained using the KDE method. DCPU of edge computing devices Integrating yields the cumulative distribution function (CDF) of each residual subsequence, which can be expressed as:

[0150]

[0151] Take the sequence value corresponding to the 0.05 quantile of the desired distribution. As the lower limit, the sequence value corresponding to the 0.95 quantile is used. This serves as an upper limit and a non-parametric dynamic threshold for anomaly detection. If the residual exceeds this threshold, the data is identified as a point anomaly. for The inverse function of .

[0152] Finally, using the detected point anomalies as demarcation points, the residual sequence is re-divided into multiple subsequences of varying lengths (the subsequences do not include the demarcation points). The edge computing device's DCPU applies an improved local outlier factor (ILOF) method to the residual subsequences. When the ILOF score is higher than a threshold, the subsequence and its demarcation points are identified as collective anomalies; otherwise, the subsequence is considered normal, and its demarcation points are still identified as point anomalies. Notably, when the residual subsequence contains few data points (defined as fewer than 3 data points in this invention) or no data points, the subsequence and its two demarcation points are directly identified as collective anomalies without needing to calculate the ILOF score.

[0153] The ILOF algorithm is an improvement on the traditional LOF algorithm. Like LOF, it is a density-based method that describes the degree of deviation of data from its neighbors, using this as an anomaly score to identify outliers. Unlike the traditional LOF algorithm, ILOF uses soft dynamic time warping (Soft-DTW) distance instead of the Euclidean or Manhattan distance to address the issue of varying sizes between the analyzed axial temperature residual subsequences. The proposed ILOF algorithm broadens the applicability of LOF, breaking the limitation of traditional LOF algorithms that can only be used with homogeneous data.

[0154] The goal of Soft-DTW is to find the optimal alignment between two heterogeneous sequences to overcome the limitations imposed by data heterogeneity on distance computation. Specifically, given a sequence A = [a1, a2, ..., a...], ... n ] and B = [b1,b2,…,bm Let the distances be n and m, respectively. Define an n×m distance matrix. Among them: dist i,j =||a i -b j ||2,i=1,2,...n,j=1,2,...m.

[0155] The Soft-DTW distance can be expressed as:

[0156]

[0157] In the formula, The minimum distance returned by Soft-DTW is calculated using the following recursive formula:

[0158]

[0159] in:

[0160]

[0161] Therefore, by replacing the Euclidean or Manhattan distance with the Soft-DTW distance when performing the LOF algorithm, the ILOF score of the subsequence can be calculated. When the ILOF score of the subsequence is greater than a given threshold θ... ILOF When this happens, the subsequence is determined to be abnormal. In this invention, the threshold θ ILOF Set to three times the standard deviation of the ILOF scores of all subsequences.

[0162] 3: Method for learning and identifying vibration signal characteristics of motor bearings

[0163] This invention proposes a new and effective feature extraction index based on Fluctuation Dispersion Entropy (FDE): Fractional Fluctuation Dispersion Entropy (FrFDE). FrFDE reflects changes in the working state of traction transmission equipment and uses vibration signals to reflect changes in its internal structure. The specific feature extraction process for FrFDE is as follows:

[0164] Step 1 involves using a charge amplifier to condition the charge of the end-layer accelerometer installed near the vibration source of the equipment. The resulting acceleration measurement is then sent to the vehicle-mounted edge control module in the side layer via the internal bus interface module and stored in the storage array. Simultaneously, the DCPU performs real-time sliding window processing on the data with a window length of N and a step size of L. Each sliding window generates a one-dimensional vibration signal time sequence x = {x1, x2, ..., x...} of length N. NThe time series is mapped to y = {y1, y2, ..., y3} using the normal cumulative distribution function (NCDF). N};

[0165]

[0166] In the formula: γ and σ are the mean and standard deviation of the time series x, respectively. Then, y is mapped to {1,2,...,c} using the following linear transformation:

[0167]

[0168] In the formula: round() is the rounding function; c is the number of categories; It is the i-th element in class c.

[0169] Step 2: Process the sequence Perform phase space reconstruction with embedding dimension m and delay factor d to obtain a new reconstruction vector. as follows:

[0170]

[0171] Then each embedding vector Mapping to the scattering pattern of fluctuations

[0172]

[0173] Assigned to each time series The total number of possible scattering patterns is (2c-1). m-1 .

[0174] Step 3: For each possible scattering pattern Calculate the relevant probability of its occurrence. As shown in the following formula:

[0175]

[0176] Step 4: Introduce fractional-order operators to capture more dynamic properties of the vibration signal, i.e., the Riemann-Liouville operator:

[0177]

[0178] D α f(x) = D n I n-α f(x) (14)

[0179] In the formula: I αD represents the α-order RL integral. α Let Γ(·) be the α-order RL derivative, and let Γ(·) be a class of functions that extend the factorial function to real and complex numbers. Based on the definition of Shannon entropy, calculate the fractional wave scattering entropy for different α-orders:

[0180]

[0181] As can be seen from the above process, the fractional-order wave spread entropy has the following characteristics:

[0182] 1) It can extract the state characteristics of vibration signals at different orders α. The fractional wave propagation entropy change curves at different orders α and different probabilities p are shown in the figure. Figure 9 As shown, especially when the order is greater than 0 and less than 1, the p-entropy value varies greatly under different probabilities, making it more sensitive to the weak state evolution of the signal and able to better distinguish different states of the device.

[0183] 2) The monitored quantity is the fractional wave spread entropy of the vibration signal rather than the original signal, which improves the reliability of the monitoring results.

[0184] 3) The FrFDE algorithm has low time and space complexity, which reduces the computational burden and makes it very suitable for real-time signal processing at the edge layer.

[0185] 4. Monitoring vibration signals and obtaining anomaly monitoring results through intelligent algorithms.

[0186] The cumulative sum control chart (CUSUM) algorithm is designed to accumulate and sum small offsets in the process to achieve an amplification effect, thereby improving the sensitivity of the anomaly monitoring process to small offsets. The CUSUM expression is:

[0187]

[0188] x(t) represents the process detection data, g + and g - These are the cumulative sums of positive and negative changes, respectively. To avoid detecting anomalous changes in the absence of actual change or slow drift, the algorithm also relies on a parameter h for drift correction. Their initial values ​​are zero, i.e., g + (0)=g - (0) = 0. Given a threshold T, if g + >T or g - >T, the CUSUM algorithm can detect anomalies and promptly send alarms to operators.

[0189] The vibration signal anomaly detection process based on fractional wave dispersion entropy and cumulative sum control chart is as follows:

[0190] Step 1: Use the accelerometer and signal conditioning circuit of the end layer to collect the vibration signal of the electric drive equipment to be monitored, and transmit it to the edge master control module of the side layer through the internal bus;

[0191] Step 2: The DCPU in the main control module calculates the fractional order wave distribution entropy of the vibration signal under each sliding window in real time, tracks the state changes of the equipment, and stores the calculation results in the storage module;

[0192] Step 3: The DCPU uses the cumulative sum control chart to track the trajectory of the fractional order fluctuation spread entropy sequence, uses a threshold to determine whether the monitoring data is abnormal, and promptly reports the detected abnormal points to the vehicle monitoring center when the equipment is abnormal. After the vehicle monitoring center confirms the abnormality, it reports it to the CCU of the vehicle level through the TCN.

[0193] 5: Multi-source collaborative monitoring method for motor status

[0194] Monitoring and judging a single physical quantity is sometimes inaccurate. In edge computing devices, the characteristics of different signals from multiple sources are used, and weights are assigned according to the importance of the features to improve the accuracy and reliability of equipment fault diagnosis results. Multi-source collaborative monitoring methods include feature extraction from different signals, feature stitching, and fault diagnosis. Figure 10 This is a flowchart of a multi-source collaborative monitoring method provided in an embodiment of the present invention. The specific processing steps include:

[0195] (1) Feature extraction of different signals

[0196] First, the sensor module is used to measure temperature X. temp Current X c Torque X tor Data is collected and transmitted to the respective edge control modules via the bus interface module. In the edge control module (DCPU), the data is first preliminarily assessed using FFT analysis or thresholding. If no anomalies are found, the data is simply monitored and displayed. If anomalies are found, the data undergoes further detailed processing, including the stator temperature X at different sampling frequencies. temp Stator current X c Torque signal X tor The data is divided into equal-length segments. To ensure sampling accuracy, the current sampling frequency is the highest. Therefore, the length k of each current data segment can be determined by comparing diagnostic results. The data segment duration t is determined based on the k value. The duration of each temperature and torque signal segment is also set to t, thus determining the number of data points in each segment. The three different signals input to the convolutional neural network are acquired simultaneously within the same time period.

[0197]

[0198] Normalize each data segment separately, where t = 1, 2, ..., k:

[0199]

[0200] Regarding stator temperature X temp Stator current X c Torque signal X tor Three different signals were used for feature extraction using a CNN with identical parameters except for the convolution operation parameters. The feature extraction process is as follows:

[0201] Step 1: Assuming the input signal is x, perform a convolution operation on it:

[0202] y i l =f(b k l +conv1D(W l ,S i l-1 (19)

[0203] Where i = 1, 2, ..., C represents the number of convolution kernels. This represents the convolution result of the i-th convolution kernel with respect to the input x. It is the output neuron of the (l-1)th convolutional pooling module. W represents the k-th output bias corresponding to the l-th convolutional layer. l denoted by , f(.), represents the weights of the l-th convolutional layer, f(.) is the activation function, and conv1D represents a one-dimensional convolution.

[0204] Step 2 uses max pooling for dimensionality reduction and batch normalization of the output features to obtain C features N. l =[n1 l n2 l ,…,n C l Each feature has a dimension of M×1, where max pool represents max pooling and BN represents batch normalization.

[0205] p i l =max pool(y i l )

[0206] n i l =BN(p i l (20)

[0207] Step 3 evaluates the importance of different features using global average pooling and fully connected layers. Here, `global_avepool` represents global average pooling, `FC` represents a fully connected layer, and `fc` represents... 2 Let C represent C evaluation vectors, each corresponding to one of C features.

[0208] ga l =global_avepool(N l )

[0209] fc 1 =FC(ga) l (C / r)

[0210] fc 2 =FC(fc) 1 (C) (21)

[0211] Where r represents the dimensionality reduction ratio, and the C features are respectively compared with the evaluation vector fc 2 Multiplying the C elements in the formula yields the C weighted features, resulting in the C weighted features Y = [y1...]. 1 ,y2 1 ,…,y c 1 ].

[0212] (2) Feature stitching in edge computing devices

[0213] In their respective edge computing devices, for the three sets of M×1 dimensional features obtained from the three convolutional neural networks, let's assume they are F1, F2, and F3 respectively.

[0214]

[0215] Transpose F1, F2, and F3 respectively, and then connect them using the concatenation function Conc, i.e.:

[0216] F=Conc(Trans(F1),Trans(F2),Trans(F3)) (23)

[0217] (3) Fault Classification and Identification

[0218] Extracted features are stored and backed up on the edge devices. Simultaneously, the extracted features are sent to the onboard monitoring center via an Ethernet interface module. At the onboard monitoring center, various features are fused and combined with normal data to build a collaborative diagnostic model for training and parameter determination, including thresholds and dynamic thresholds. When abnormal data is detected, the model is used to fuse and identify it, and the data is uploaded to the ground monitoring center via the onboard GPRS. The determined diagnostic model and parameters are then distributed to the corresponding edge computing devices via the Ethernet interface module. Identified faults are also transmitted to the corresponding edge computing devices via the Ethernet interface module. The edge computing devices transmit the faults to the central control unit via the TCN, where the central control unit performs operations such as speed reduction, stopping, or switching on the traction system according to a predetermined procedure.

[0219] The modeling principle in the vehicle monitoring center is as follows: In the fully connected layer of the convolutional neural network, the Softmax function is used for fault classification. It can map the elements in the output vector to the range (0, 1) respectively, and the sum of the elements is 1. Therefore, it can be regarded as the probability of each fault category, and the category with the highest probability is taken as the final classification result.

[0220]

[0221] Where exp(z) i ), j = 1, 2, ..., N represents the j-th input to the Softmax function, and N is the total number of categories.

[0222] 6. Collaborative Monitoring Methods: Collaborative monitoring systems based on edge computing devices include novel monitoring methods, primarily:

[0223] ① Monitoring of a certain physical quantity enabled the extraction of features related to its normal operating state and its state evolution.

[0224] ② For the monitoring of a certain physical quantity, anomaly monitoring results are obtained through intelligent algorithms.

[0225] ③ For multiple physical quantities, dynamic thresholds are obtained through collaborative learning algorithms. Different dynamic judgment thresholds can be determined according to changes in the operating environment to carry out anomaly detection, thus solving the problem of high false alarm and false alarm rates in the current monitoring system.

[0226] 7. Dynamic threshold

[0227] Most anomaly detection processes in existing vehicle safety equipment are based on threshold methods, such as the detection of axle temperature and vibration signals. However, this approach often fails to adapt to the actual operating conditions or health status of the equipment. This invention incorporates a rule-based method into a threshold-based anomaly detection method, transforming the anomaly detection threshold into a dynamic threshold that considers both the equipment's operating conditions and health status.

[0228] The rule-based method in this invention aims to dynamically adjust the anomaly detection threshold through pre-set rules. Specifically, when considering the actual operating conditions of the train equipment: 1) Regarding the axle temperature threshold setting, if the traction motor current exceeds the rated current, the train is running at full load. In this case, the components of the train's running gear cannot withstand higher operating temperatures, so the axle temperature threshold needs to be lowered to predict axle temperature faults in advance; 2) Regarding the vibration threshold setting, if the train speed exceeds 350 km / h, even a small deviation in the vibration signal will indicate a fault in the train's running gear, so the vibration threshold needs to be lowered. When considering the health status of the train equipment: if the equipment failure rate is low or it has been operating healthily for a long time, the fault threshold can be raised; if the equipment failure rate is judged to be high or its lifespan is close to the safety boundary, the threshold should be lowered.

[0229] The coordinated monitoring of multiple physical quantities is used to assess the operating status of electric drive equipment.

[0230] In summary, the embodiments of the present invention reduce the amount of data transmission by adding embedded edge computing devices and storage media to each type of sensor, performing data processing and feature extraction at the sensor end, and transmitting feature data only through the train communication network, which can greatly increase the number of sensors deployed.

[0231] Edge computing devices process and analyze data close to the sensor terminals, extracting feature data and transmitting feature and anomaly data, significantly reducing data transmission volume. This allows for increased sensor deployment, meeting the condition monitoring needs of intelligent trains and driverless trains. This represents a substantial feature and significant advancement.

[0232] Each type of sensor is equipped with embedded edge computing devices and storage media, enabling it to store the normal state characteristics of train operation lines and climate conditions, as well as the evolution patterns of these characteristics.

[0233] Each type of sensor incorporates embedded edge computing devices and storage media, enabling it to store normal state characteristics of train operation routes and weather conditions. The edge computing devices possess self-learning capabilities, allowing them to extract and store features of the normal train operation process, while also tracking state evolution. Current train detection devices lack this functionality, representing a substantial feature and significant advancement.

[0234] Each type of sensor adds embedded edge computing devices and storage media, utilizing the devices in edge computing to implement intelligent learning algorithms, enabling the setting of different thresholds under different conditions to improve accuracy.

[0235] By employing edge computing devices, dynamic thresholds can be obtained using learning algorithms based on recorded normal state characteristics and evolutionary characteristics, thereby reducing false alarm rates. Edge computing devices can also evaluate sensor states and distinguish between anomalies caused by sensor states and anomalies caused by device states.

[0236] By using the vehicle edge computing device, the status of the traction electric drive system can be fused and judged, which can more accurately determine the equipment status than independent system monitoring and judgment.

[0237] By employing edge computing devices to fuse and discriminate device status at the feature layer, potential anomalies in devices that cannot be detected through current independent monitoring and threshold judgment can be identified, along with accurate estimates of remaining lifespan. Current train inspection devices lack this capability, representing a substantial feature and significant advancement.

[0238] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0239] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0240] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0241] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A collaborative monitoring system for traction electric drive and safety monitoring equipment based on edge computing devices, characterized in that, include: Edge computing devices and sensor modules are installed in each car of the rail train to perform multi-source physical quantity collaborative monitoring of traction electric drive equipment and safety monitoring equipment. The edge computing devices of each rail train are interconnected via Ethernet and connected to the Train Monitoring Center (TMC). The TMC receives relevant data, performs model training and parameter determination, determines various thresholds, performs fault prediction and health management, and publishes results and tasks. The edge computing device uses the MVB interface to access the TCN through the vehicle gateway, connects to the CCU, reports anomalies to the CCU, and receives control commands sent by the CCU. The edge computing device includes a power module, a sensor module, an edge control module, and a storage module that are interconnected via an internal signal transmission bus. The sensor module includes a sensor, conditioning and transformation, processing and judgment, storage and communication. Different conditioning and transformation circuits are used according to different sensor output signals to preprocess and threshold the sensor signals to obtain the normal state quantity and / or abnormal quantity of the monitored object. The normal state quantity and / or abnormal quantity of the monitored object are sent to the edge master control module through the internal signal transmission bus interface. The edge master control module includes a data processing central processing unit (DCPU), an edge device management central processing unit (MCPU), a data storage array, a flash memory, and a bus interface. The DCPU aggregates data from one or more sensor modules, uses feature extraction algorithms to obtain the features of normal states and the state changes reflecting the evolution of normal states, learns and models normal data or normal data features to obtain the change patterns of normal states; it analyzes and further judges abnormal quantities in combination with normal data features and their evolutionary state change rates, performs fault diagnosis on abnormal quantities, and determines the threshold revision amount for frequently occurring faults based on the fault diagnosis results; the MCPU realizes the initialization of the edge computing device, internal bus transmission, management of various communication modules and each module, and collaborative address allocation of different edge computing devices during the initial operation of the train; the DCPU and MCPU exchange relevant data through the flash memory, and control and interact with other modules of the edge computing device through the internal bus interface; The storage module is used to configure the storage chip array, manage the CPU and bus interface, store normal operation characteristics and their evolution characteristics, key data points, abnormal data, fault diagnosis models and parameters, fault diagnosis results, thresholds, data characteristics and key data points of multi-source collaborative monitoring, and interact with the edge master control module through the internal data transmission bus. The power module is used to draw power from the vehicle power supply, convert it to DC power through DC-DC conversion, and then supply power to other modules via the power bus. The DCPU in the edge computing device performs characteristic learning and anomaly detection methods for motor bearing vibration signals. It uses fractional-order scatter entropy (FrFDE) to reflect changes in the working state of the traction drive equipment and vibration signals to reflect changes in the internal structure. The specific processing steps include: The charge quantity of the end-layer accelerometer installed near the equipment vibration source is signal-conditioned using a charge amplifier. The obtained acceleration detection quantity is sent to the vehicle-mounted edge master control module of the side layer through the internal bus interface module and stored in the storage array. The DCPU processes the data with a window length of... Step size is Real-time sliding window processing, each sliding window yields a one-dimensional vibration signal time series of length N. The time series is mapped to using the normal cumulative distribution function. ; （8） In the formula: and They are time series The mean and standard deviation are transformed according to the following linear transformation. Mapped to : （9） In the formula: This is the rounding function; Number of categories; It is the i-th element in class c; For sequence Embedding dimension is The delay factor is The phase space is reconstructed to obtain a new reconstruction vector. as follows: （10） Then each embedding vector Mapping to the scattering pattern of fluctuations : （11） Assigned to each time series The number of all possible scattering patterns is ; For each possible scattering pattern Calculate the relevant probability of its occurrence. As shown in the following formula: （12） Introducing fractional operators to capture more dynamic properties of vibration signals, namely the Riemann-Liouville operator: （13） （14） In the formula: express RL integral of order, for RL derivative, A class of functions that extend the factorial function to real and complex numbers, calculated according to the definition of Shannon entropy. Fractional order wave scattering entropy below order: （15） The collaborative monitoring system is used to monitor vibration signals for anomalies based on fractional-order wave spread entropy and cumulative control charts. The specific processing steps include: Vibration signals of the electric drive equipment to be monitored are collected using the end-layer accelerometer and signal conditioning circuit, and transmitted to the edge master control module of the side layer via the internal bus; The DCPU in the main control module of the edge computing device calculates the fractional order wave distribution entropy of the vibration signal under each sliding window in real time, tracks the state changes of the equipment, and stores the calculation results in the storage module. The DCPU uses the Cumulative and Control Chart (CUSUM) algorithm to track the trajectory of the fractional-order fluctuation spread entropy sequence. It uses a threshold to determine whether the monitoring data is abnormal. When the equipment is abnormal, it promptly reports the detected abnormal points to the vehicle monitoring center. After the vehicle monitoring center confirms the abnormality, it reports it to the CCU of the vehicle level through the TCN. The expression for the CUSUM algorithm is: （16） For process detection data, and These are the cumulative sums of positive and negative changes, respectively, and the parameters are... The initial value is zero, that is Given a threshold ,if or The CUSUM algorithm detects anomalies and promptly sends alarms to the operator.

2. The system according to claim 1, characterized in that, The edge computing device further includes: a positioning module, a multi-functional communication bus MVB interface module, an Ethernet interface module, and a wireless interface module; The MVB interface module is used as the main interface for edge computing devices to access the train communication network (TCN), to receive control commands from the central control unit (CCU), and to report key anomalies to the CCU through the TCN. The Ethernet interface module is used as a communication network connecting the edge computing device and the Train Monitoring Center (TMC) to enable interaction between edge computing devices. The wireless interface module is used to control the access, networking and address allocation of the vehicle-mounted wireless sensors, transmit data with the ground monitoring data center, and send the processing results of the vehicle-mounted edge computing device to the ground monitoring center. The positioning module is used to provide location data to the edge computing device using a satellite positioning chip, and to provide location data to the edge master control module through an internal signal transmission bus.

3. The system according to claim 2, characterized in that, The collaborative monitoring system includes a vehicle layer, a side layer, and an end layer; The end layer includes sensors and sensor modules, which are installed at the train traction electric transmission equipment and / or safety monitoring equipment. The type and number of sensors are selected according to the characteristics of the physical quantities of the monitored equipment. Data is collected on the status of the monitored equipment. The sensors are connected to the sensor modules. The sensor modules preprocess the data, perform threshold discrimination and early warning, and send the discrimination results to the edge computing device. Receive instructions sent by edge computing devices; Depending on the monitoring task and deployment location, the sensor module can be deployed together with the sensor at the monitoring equipment, or the sensor can be deployed on the monitoring equipment, and the sensor module can be deployed within the edge computing device. The edge layer includes edge computing devices installed on the train. These devices receive data on traction motors, bearing temperature, vibration, and speed deployed on the bogies. They aggregate data transmitted from one or more sensor modules, use feature extraction algorithms to obtain normal state features and state change quantities reflecting the evolution of normal states, learn and model normal data or normal data features to obtain the change patterns of normal states, and store these patterns. They analyze and further distinguish abnormal quantities by combining normal data features and their evolving state change rates, perform fault diagnosis on abnormal quantities, determine threshold revision amounts for frequently occurring faults based on the fault diagnosis results, and store the results. Different edge computing devices exchange data with each other via Ethernet and receive control commands from the TMC in the train layer. The vehicle deck includes a TMC, which connects the edge computing devices of each carriage via Ethernet to form an on-board collaborative monitoring system. It interacts with the ground monitoring center via General Packet Radio Service (GPRS) or mobile communication network to perform model training and parameter determination for the same physical quantity, and receives monitoring and analysis results of the same physical quantity from different edge computing devices and compares and judges them. Model training and parameter determination for coordinated monitoring of different physical quantities; receiving and comparing the monitoring and analysis results of different physical quantities under coordinated state monitoring. The system determines the anomaly detection threshold and dynamic threshold for different track physical quantities, and uses algorithms related to equipment operation fault prediction and health management based on the results of the vehicle's edge computing device to determine the equipment status; it also receives instructions from the ground monitoring center.

4. The system according to claim 3, characterized in that, The collaborative monitoring system is used to learn the features of normal axle temperature data through a Long Short-Term Memory (LSTM) network model via the vehicle floor TMC, and to establish a mapping relationship between historical axle temperature data and current data through deep learning algorithms. The specific processing steps include: LSTM introduces new internal states in each cell. This involves linear, cyclic information transfer, outputting information to the external state of the hidden layer. By introducing the forget gate Input gate and output gate Control the amount of information; The vehicle-level TMC (Transmission Control Center) standardizes and performs sliding window processing on the normal axle temperature data. The preprocessed axle temperature sequence is then input into an LSTM (Laser-to-Shipment Module) model for training. The training dataset comes from historical normal axle temperature data stored in the TMC or from normal axle temperature data transmitted from edge computing devices in the side layers. This process establishes a mapping between historical and current axle temperature data. （1） In the formula, Indicates time t Shaft temperature; Represents a nonlinear mapping relationship, referring to a black-box model without explicit expressions; m The length of the historical shaft temperature data is used as an input variable for the model.

5. The system according to claim 4, characterized in that, The DCPU in the edge computing device of the edge layer performs joint detection of multiple anomalies in the shaft temperature data by analyzing the shaft temperature residual sequence. The specific processing includes: The edge computing device's DCPU downloads a pre-trained mapping relationship between historical axle temperature data and current data from the vehicle floor TMC. It then inputs the historical axle temperature data into this pre-established mapping relationship to obtain the predicted axle temperature value for the current moment. Compared with the actual shaft temperature data at the current moment The residual sequence is obtained by subtraction. E , is represented as: （2） in, Indicates time t The residual is expressed as: ; Secondly, to calculate the anomaly detection threshold at different times, the edge computing device's DCPU divides the residual sequence into non-overlapping subsequences, with a subsequence length of [missing information]. h For time t The residual subsequence is represented as: （3） in, h Indicates the number of residuals used in the analysis; t for h Integer multiples of; The DCPU of the edge computing device obtains the probability density function of each residual subsequence using the KDE method. ,right Integrating yields the cumulative distribution function of each residual subsequence, expressed as: （4） Take the sequence value corresponding to the 0.05 quantile of the desired distribution. As the lower limit, the sequence value corresponding to the 0.95 quantile is used. This serves as an upper limit and a non-parametric dynamic threshold for anomaly detection. If the residual exceeds this non-parametric dynamic threshold, the data is identified as a point anomaly. for The inverse function; Using the detected point anomalies as the dividing points, the residual sequence is re-divided into multiple subsequences of different lengths. The DCPU of the edge computing device applies an improved Local Outlier Factor (ILOF) method to the residual subsequences. When the ILOF score is higher than the non-parametric dynamic threshold for anomaly detection, the subsequence and its dividing point are identified as collective anomalies. Conversely, the subsequence is normal, and the dividing point of the subsequence is still identified as a point anomaly.

6. The system according to claim 1, characterized in that, The collaborative monitoring system is used to perform multi-source collaborative monitoring of motor status based on the characteristics between different signals from multiple sources. The specific processing steps include: Using the sensor module to measure temperature Current Torque The data is collected and transmitted to the respective edge master control modules via the bus interface module; In the edge control module, the DCPU first performs a preliminary judgment on the data through FFT analysis or thresholding. If no anomalies are found, the data is simply monitored and displayed. If anomalies are found, the data undergoes further detailed processing, including the stator temperature at different sampling frequencies. Stator current Torque signal The data is divided into equal-length segments, and the length of each segment is determined by comparing diagnostic results. k The value of is determined according to k Value determines data period t The duration of each data segment for temperature and torque signals is also taken. t This allows us to determine the number of data points contained in a data segment, as the data from the three different signals input to the convolutional neural network are collected at the same time. （17） Each data segment is normalized separately, where t =1, 2, ..., k : （18） In the DCPU of the edge master control module, the stator temperature is... Stator current Torque signal Three different signals were used for feature extraction using a CNN with identical parameters except for the convolution operation parameters. The feature extraction process is as follows: Step 1 Assume the input signal is x ,right x Perform convolution operations: （19） in i =1, 2, ..., C represents the number of convolution kernels, Indicates the first i Each convolutional kernel pairs with the input x The convolution result, It is the first l The output neuron of the -1 layer convolutional pooling module Indicates the first l The first convolutional layer corresponding to the first k Each output bias. Indicates the first l The weights of each convolutional layer f (.) activation function, conv1D This represents one-dimensional convolution; Step 2: Max pooling is used for dimensionality reduction, and batch normalization is performed on the output features to obtain C features. Each feature has a dimension of M×1, where max pool represents max pooling and BN represents batch normalization; （20） Step 3 uses global average pooling and fully connected layers to evaluate the importance of different features. `global_avepool` represents global average pooling, and `FC` represents a fully connected layer. fc 2 This represents C evaluation vectors, each corresponding to one of the C features; （21） in r The expression represents the dimensionality reduction ratio, where C features are compared with the evaluation vector. fc 2 Multiplying the C elements in the formula yields C weighted features Y = [ y 1 1 , y 2 1 ,…, y c 1 ]; In their respective edge computing devices, for the three sets of M×1 dimensional features obtained from the three convolutional neural networks, let them be respectively... F 1. F 2. F 3; （22） Will F 1. F 2. F 3. Perform transpose on each part, then use the concatenation function. Conc To make a connection, that is: （23） The edge control module uses the Ethernet interface module to send the extracted features to the vehicle monitoring center. The vehicle monitoring center integrates various features, uses the Softmax function to train the diagnostic model and determine the model parameters, maps the elements in the output vector to the range (0, 1) respectively, and the sum of the elements is 1, which is used as the probability of each fault category. The category with the highest probability is used as the final classification result. （24） Where exp( z i ), j =1, 2, ..., N Indicates the softmax function's... j One input, N Total number of categories; After the on-board monitoring center completes the modeling, it transmits the model to the corresponding edge computing device through the Ethernet interface module. The edge computing device uses the model to perform fault diagnosis and then transmits the fault to the central control unit through TCN. Based on the collaborative monitoring results and fault prediction results, the on-board monitoring center sends relevant fault data to the central control unit through the vehicle gateway. In the central control unit, the traction system is subjected to deceleration, stopping, or switching operations according to a predetermined procedure.

7. The system according to claim 1, characterized in that, The collaborative monitoring system is used to obtain dynamic thresholds for multiple physical quantities through a collaborative learning algorithm, and to perform fault prediction and health management for monitoring equipment. It determines different dynamic judgment thresholds based on changes in the operating environment to obtain abnormal monitoring results.

Images