Lightweight vehicle-mounted edge calculation circuit module and lightweight vehicle-mounted edge calculation method

Abstract

The invention provides a lightweight vehicle-mounted edge calculation circuit module and a lightweight vehicle-mounted edge calculation method. The circuit module comprises a multi-source data interface which is used for being electrically connected with vehicle-mounted equipment so as to receive various types of source data of the vehicle-mounted equipment; the preprocessing unit is electrically connected with the multi-source data interface and is used for performing predetermined processing on the source data of various types to obtain predetermined processed data, and the predetermined processing comprises de-noising processing and normalization processing; the heterogeneous computing unit is electrically connected with the preprocessing unit and used for processing the preprocessed image data to obtain image feature data and processing the preprocessed time sequence data to obtain time sequence feature data; and the decision-making unit is connected with the heterogeneous calculation unit and is used for making a decision according to the image feature data and the time sequence feature data. According to the scheme, the problem that the computing power of an existing edge chip for vehicle-mounted edge computing is insufficient and needs to depend on cloud processing is solved.

Description

Technical Field

[0001] This application relates to the field of rail transit, and more specifically, to a lightweight on-board edge computing circuit module and a lightweight on-board edge computing method, a rail train, and a computer-readable storage medium. Background Technology

[0002] Onboard systems (e.g., rail transit onboard systems) need to monitor vehicle operating status in real time, such as data collected by sensors including bearing temperature, traction motor vibration, and brake air pressure. Related technologies often rely on cloud servers to process this sensor data, but cloud-based solutions have several shortcomings:

[0003] 1) High transmission latency: Data upload to the cloud server causes abnormal response delays, making it difficult to meet the needs of emergency scenarios (e.g., instantaneous warning of bearing overheating);

[0004] 2) High bandwidth pressure: When multiple vehicles communicate concurrently, the massive amount of data will exacerbate the network load;

[0005] 3) Insufficient computing power of edge devices: Existing in-vehicle edge processors (such as ARM and MCU) are unable to run complex AI models (such as deep neural networks). Summary of the Invention

[0006] The main objective of this application is to provide a lightweight vehicle-mounted edge computing circuit module, a lightweight vehicle-mounted edge computing method, a railcar, and a computer-readable storage medium, so as to at least solve the problem that current edge devices for vehicle-mounted edge computing have insufficient computing power and need to rely on cloud processing.

[0007] To achieve the above objectives, according to one aspect of this application, a lightweight automotive edge computing circuit module is provided, comprising: a multi-source data interface for electrically connecting to an automotive device to receive various types of source data from the automotive device, wherein the multi-source data interface includes an image data interface, a non-image data interface, and a bus interface; a preprocessing unit electrically connected to the multi-source data interface for performing predetermined processing on the various types of source data to obtain predetermined processed data, wherein the predetermined processed data includes preprocessed image data and preprocessed time-series data, wherein the predetermined processing includes denoising processing and normalization processing; a heterogeneous computing unit electrically connected to the preprocessing unit for processing the preprocessed image data to obtain image feature data and processing the preprocessed time-series data to obtain time-series feature data; and a decision unit connected to the heterogeneous computing unit for making decisions based on the image feature data and the time-series feature data.

[0008] Optionally, the heterogeneous computing unit includes: a lightweight convolutional array module for performing lightweight convolutional processing on the preprocessed image data to obtain the image feature data; and a temporal signal processing module for processing the preprocessed temporal data to obtain the temporal feature data.

[0009] Optionally, the lightweight convolutional array module includes a multiply-accumulate operator for performing multiply-accumulate operations in the lightweight convolutional processing to obtain the multiply-accumulate result.

[0010] Optionally, the time-series signal processing module includes: a Fourier accelerator for performing Fourier acceleration processing on the preprocessed time-series data to obtain Fourier-accelerated time-series data; and a long short-term memory network operator for performing time-domain analysis on the preprocessed time-series data to obtain time-domain analyzed data, wherein the long short-term memory network operator supports sliding window processing.

[0011] Optionally, the lightweight vehicle-mounted edge computing circuit module further includes a dynamic power management unit, which is electrically connected to the multi-source data interface, the preprocessing unit, the heterogeneous computing unit and the decision unit, respectively, and is used to dynamically adjust the voltage frequency of each unit connected to the dynamic power management unit according to the amount of source data to perform hierarchical energy consumption control.

[0012] Optionally, the decision unit includes: a multimodal fusion module, used to perform weighted fusion processing on the image feature data and the temporal feature data to obtain a weighted fusion result; and a lightweight classification module, used to perform classification processing on the source data based on the weighted fusion result to obtain the anomaly level and confidence level of the source data.

[0013] According to another aspect of this application, a lightweight vehicle-mounted edge computing method is provided, comprising: establishing a communication connection with a vehicle-mounted device to obtain various types of source data from the vehicle-mounted device; performing predetermined processing on the various types of source data to obtain predetermined processed data, wherein the predetermined processed data includes preprocessed image data and preprocessed temporal data, wherein the predetermined processing includes denoising processing and normalization processing; processing the preprocessed image data to obtain image feature data, and processing the preprocessed temporal data to obtain temporal feature data; and making a decision based on the image feature data and the temporal feature data.

[0014] Optionally, processing the preprocessed image data to obtain image feature data and processing the preprocessed temporal data to obtain temporal feature data includes: processing the preprocessed image data using a lightweight convolution algorithm including a binarization sparse processing algorithm to obtain the image feature data; and performing Fourier acceleration processing and long short-term memory network operations on the preprocessed temporal data to perform joint analysis in the time domain and frequency domain to obtain the temporal feature data.

[0015] Optionally, the method further includes: determining the data volume of the various types of source data based on the sampling frequency, data type, and data transmission protocol of the various types of source data; and dynamically adjusting the voltage frequency of each unit in the vehicle edge computing circuit module according to the data volume of the source data to perform graded energy consumption control.

[0016] Optionally, the voltage frequency of each unit in the vehicle edge computing circuit module is dynamically adjusted according to the amount of source data to perform graded energy consumption control, including: when the amount of source data is greater than or equal to a preset amount of data, controlling the voltage frequency of each unit in the vehicle edge computing circuit module to be at full voltage frequency; when the amount of source data is less than the preset amount of data, controlling the voltage frequency of the target unit in the vehicle edge computing circuit module to be at a non-full voltage frequency, wherein the target unit is an idle unit or a non-full-load working unit.

[0017] Optionally, making a decision based on the image feature data and the time-series feature data includes: performing a weighted fusion process on the image feature data and the time-series feature data to obtain a weighted fusion result, wherein the weight parameters in the weighted fusion process are related to the amount of data in the image feature data and the amount of data in the time-series feature data; and classifying the source data based on the weighted fusion result to obtain the anomaly level and confidence level of the source data.

[0018] Optionally, classifying the source data based on the weighted fusion result to obtain the anomaly level and confidence level of the source data includes: using a hybrid model of decision tree and support vector machine to classify the weighted fusion result to obtain the anomaly level and confidence level of the source data.

[0019] According to another aspect of this application, a rail train is provided, comprising: a vehicle body; a variety of on-board devices installed on the vehicle body and any one of the aforementioned lightweight on-board edge computing circuit modules, wherein the variety of on-board devices are electrically connected to the lightweight on-board edge computing circuit module via a multi-source data interface, and wherein the variety of on-board devices includes an on-board air conditioner, an on-board display device, and an on-board communication device.

[0020] According to another aspect of this application, a computer-readable storage medium is provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform any of the aforementioned lightweight vehicle edge computing methods.

[0021] According to the technical solution of this application, a lightweight automotive edge computing circuit module includes a multi-source data interface, a preprocessing unit, a heterogeneous computing unit, and a decision-making unit. The multi-source data interface is electrically connected to an in-vehicle device to receive various types of source data from the in-vehicle device. The multi-source data interface includes an image data interface, a non-image data interface, and a bus interface. The preprocessing unit is electrically connected to the multi-source data interface and performs predetermined processing on the various types of source data to obtain preprocessed data. The preprocessed data includes preprocessed image data and preprocessed time-series data. The predetermined processing includes denoising and normalization. The heterogeneous computing unit is electrically connected to the preprocessing unit and processes the preprocessed image data to obtain image feature data and processes the preprocessed time-series data to obtain time-series feature data. The decision-making unit is connected to the heterogeneous computing unit and makes decisions based on the image feature data and the time-series feature data. In this solution, the preprocessing unit, heterogeneous computing unit, and decision unit are separate circuit modules, each undertaking a portion of the computing power. Compared to the existing solutions that use a single MCU for automotive edge processors, the computing power is more sufficient, eliminating the need to rely on the cloud. This solves the problem of insufficient computing power in current edge chips for automotive edge computing, which necessitates cloud processing. Attached Figure Description

[0022] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0023] Figure 1 A schematic diagram of the structure of a lightweight vehicle-mounted edge computing circuit module provided in an embodiment of this application is shown;

[0024] Figure 2 A schematic flowchart of a lightweight vehicle edge computing method according to an embodiment of this application is shown.

[0025] Figure 3 A schematic diagram of a specific lightweight vehicle edge computing module provided according to an embodiment of this application is shown;

[0026] Figure 4 A flowchart illustrating a multimodal data fusion processing method according to an embodiment of this application is shown.

[0027] Figure 5A state transition diagram of a dynamic power management unit according to an embodiment of this application is shown;

[0028] Figure 6 A structural block diagram of a lightweight in-vehicle edge computing device provided according to an embodiment of this application is shown.

[0029] The above figures include the following reference numerals:

[0030] 01. Lightweight vehicle-mounted edge computing circuit module; 10. Multi-source data interface; 20. Preprocessing unit; 30. Heterogeneous computing unit; 40. Decision unit; 50. Vehicle-mounted equipment; 101. Image data interface; 102. Non-image data interface; 103. Bus interface; 301. Lightweight convolutional array module; 302. Timing signal processing module; 3011. Multiply-accumulate arithmetic unit; 3021. Fourier accelerator; 3022. Long short-term memory network arithmetic unit; 60. Dynamic power management unit; 401. Multimodal fusion module; 402. Lightweight classification module. Detailed Implementation

[0031] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0032] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0033] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0034] As described in the background section, existing solutions for vehicle-mounted systems that rely on cloud servers for processing have many shortcomings. To address the problem of insufficient computing power in current edge devices for vehicle-mounted edge computing, which necessitates reliance on cloud processing, embodiments of this application provide a lightweight vehicle-mounted edge computing circuit module, a lightweight vehicle-mounted edge computing method, a rail train, and a computer-readable storage medium.

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0036] Figure 1 This is a schematic diagram of the structure of a lightweight vehicle-mounted edge computing circuit module provided in an embodiment of this application, as shown below. Figure 1 As shown, the lightweight vehicle-mounted edge computing circuit module 01 includes a multi-source data interface 10, a preprocessing unit 20, a heterogeneous computing unit 30, and a decision unit 40. The multi-source data interface 10 is electrically connected to the vehicle-mounted device 50 to receive various types of source data from the device. The multi-source data interface 10 includes an image data interface 101, a non-image data interface 102, and a bus interface 103. The preprocessing unit 20 is electrically connected to the multi-source data interface 10 and performs predetermined processing on the various types of source data to obtain predetermined processed data. The predetermined processed data includes preprocessed image data and preprocessed time-series data. The predetermined processing includes denoising and normalization. The heterogeneous computing unit 30 is electrically connected to the preprocessing unit 20 and processes the preprocessed image data to obtain image feature data and processes the preprocessed time-series data to obtain time-series feature data. The decision unit 40 is connected to the heterogeneous computing unit 30 and makes decisions based on the image feature data and the time-series feature data.

[0037] The aforementioned decisions include: data anomaly decisions, predictive maintenance decisions, driver assistance decisions, passenger safety warnings, etc.

[0038] Among them, data anomaly decision-making includes: identifying abnormal situations during vehicle operation based on image feature data and time-series feature data, such as sudden system failures, abnormal driving behavior, or environmental changes.

[0039] Predictive maintenance decisions include using information from image feature data and time-series feature data, combined with historical data and machine learning algorithms, to predict potential failures of vehicle components, perform maintenance in advance, and reduce unexpected downtime.

[0040] Assisted driving decisions include making decisions such as path planning, obstacle avoidance, and speed control based on image feature data (such as the recognition of objects such as road signs, pedestrians, and other vehicles) and temporal feature data (such as the dynamic motion state of the vehicle) to assist or realize autonomous driving functions.

[0041] Passenger safety warnings include: analyzing occupant behavior patterns (such as fatigue driving detection (which can be achieved through image feature data) and distracted driving monitoring) and vehicle status (such as speed exceeding a threshold (which can be achieved through time-series feature data) and vehicle approaching obstacles (which can be achieved through image feature data)) to issue timely safety warnings and remind drivers to pay attention to driving safety.

[0042] Specifically, the multi-source data interface is electrically connected to the onboard equipment to receive various types of source data from different onboard sensors. The multi-source data interface includes interfaces specifically for receiving image data, non-image data (such as time-domain signals and frequency-domain signals), and general-purpose bus data, ensuring information can be obtained from various sensors on the train (such as infrared cameras, vibration sensors, and current sensors). For example, the image data interface conforms to the MIPI-CSI2 standard, the non-image data interface can support CAN or MPI bus communication, and the bus interface can be an Ethernet-based communication interface for receiving other non-real-time or control-related data.

[0043] The preprocessing unit is electrically connected to the multi-source data interface. It is responsible for initially processing the collected source data of different types, converting it into a format suitable for subsequent calculations. The preprocessing process mainly includes denoising, normalization, and format unification of the input data (e.g., converting vibration signals to spectrograms). Denoising eliminates noise interference during data acquisition to ensure data quality, typically involving digital filtering and signal smoothing. Normalization standardizes the data, ensuring that data from different sensors are compared and processed on a uniform scale, which is crucial for subsequent feature extraction and data analysis. Furthermore, the preprocessing unit converts image data and time-series data into preprocessed image data and preprocessed time-series data, respectively, preparing them for processing by the heterogeneous computing unit.

[0044] The heterogeneous computing unit is electrically connected to the preprocessing unit and is used to extract features from preprocessed image data and preprocessed time-series data. The heterogeneous computing unit consists of sub-modules with different functions, enabling efficient processing of different types of input data. This heterogeneous design allows the chip to process multimodal data simultaneously under limited resource conditions, improving the flexibility and efficiency of edge computing.

[0045] The decision-making unit connects to the heterogeneous computing unit, receiving and comprehensively processing image feature data and time-series feature data. Based on these two types of feature data, the decision-making unit uses pre-trained classification models (such as support vector machines, decision trees, neural networks, etc.) to perform data analysis and judgment to determine whether any abnormalities exist. This is particularly important in onboard environments, as the accuracy and timeliness of the decision-making unit directly affect the safe operation of the train. For example, it can identify whether equipment is overheating based on infrared image features, and simultaneously analyze vibration signals to determine whether the equipment has experienced mechanical failure, thereby issuing early warnings or taking immediate action.

[0046] The lightweight onboard edge computing circuit module integrates a multi-source data interface, a preprocessing unit, a heterogeneous computing unit, and a decision-making unit to achieve efficient processing and intelligent analysis of various types of sensor data from onboard equipment. The multi-source data interface is compatible with image, time-series signal, and general-purpose bus data, ensuring comprehensive data acquisition. The preprocessing unit denoises and normalizes the source data, improving the accuracy of subsequent feature extraction. The heterogeneous computing unit efficiently extracts spatial and spectral features from image and time-series data, overcoming the limitations of traditional onboard edge computing in multimodal data processing. The decision-making unit, based on image and time-series feature data, enables rapid anomaly detection, effectively ensuring the safety and efficiency of train operation while optimizing power consumption and adapting to the stringent requirements of the onboard environment. In summary, the preprocessing unit, heterogeneous computing unit, and decision-making unit in the lightweight onboard edge computing circuit module are independent circuit modules, each contributing a portion of the computing power. Compared to existing onboard edge processors that use a single MCU, the computing power is sufficient, eliminating the need for cloud reliance.

[0047] To address the burstiness and uncertainty of data streams in vehicular environments, some embodiments of this application introduce an event-driven asynchronous data processing mechanism to improve system response speed and resource utilization. Its core components include: Event-triggered preprocessing: The preprocessing unit is designed in an event-driven mode, activating only when specific types or intensities of event data are received, avoiding ineffective processing of continuous low-intensity data and reducing unnecessary power consumption. Asynchronous data stream management: Asynchronous data stream management technology ensures that sensor data can be immediately stored or processed upon arrival, without waiting for other data streams, reducing the average waiting time for data processing and improving real-time performance. The asynchronous data processing mechanism can significantly improve the real-time performance and efficiency of data processing, especially in the case of bursty data streams, enabling rapid response and reducing data backlog and processing delays.

[0048] For details on the implementation process, please refer to... Figure 1The aforementioned heterogeneous computing unit 30 includes a lightweight convolutional array module 301 and a temporal signal processing module 302. The lightweight convolutional array module is used to perform lightweight convolutional processing on the aforementioned preprocessed image data to obtain the aforementioned image feature data; the temporal signal processing module is used to process the aforementioned preprocessed temporal data to obtain the aforementioned temporal feature data.

[0049] Specifically, the lightweight convolutional array module (LCA) is designed to process image data received from the preprocessing unit. It extracts key features from images by performing convolution operations within a convolutional neural network. Unlike other complex image processing chips, the LCA features a lightweight design, meaning that while maintaining necessary computational power and accuracy, it significantly reduces required hardware resources and power consumption, making it more suitable for the compact environment and energy constraints of automotive edge computing. Specifically, the LCA employs low-precision computation (such as binarized weights), miniaturized MAC unit arrays, and efficient data flow scheduling to achieve rapid feature extraction.

[0050] Further, see Figure 1 The aforementioned lightweight convolution array module 301 includes a multiply-accumulate operator 3011, which is used to perform the multiply-accumulate operation in the aforementioned lightweight convolution processing to obtain the multiply-accumulate operation result.

[0051] Specifically, the Multiplier-Accumulator Unit (MAC unit) is a hardware unit in Convolutional Neural Networks (CNNs) that performs basic computations. It is responsible for the dot product operation between the convolutional kernels and the input data in the convolutional layers. That is, it performs element-wise multiplication of each local region of the input feature map with its corresponding convolutional kernel, and then sums all the multiplication results to obtain the output value at a specific location. In processing preprocessed image data, the MAC unit performs the convolution operation of the CNN. In this process, the image data is broken down into smaller regions, which are multiplied point-by-point with a series of convolutional kernels (also called filters) and summed to output a new feature map containing specific spatial features of the original image. For example, one convolutional kernel might be set to detect edge features in an image, while another might focus on certain color or texture patterns, which are crucial for identifying the state of train components (such as infrared image features of an overheated gearbox). Considering the power consumption and space constraints in the automotive environment, the MAC units in the lightweight convolutional array module have been designed to reduce power consumption and area footprint while maintaining sufficient computing performance. For example, traditional convolutional neural networks use floating-point numbers as weights, but in the lightweight convolutional array module, the weights are quantized into binary (0 or 1), which greatly simplifies multiplication operations and reduces computational energy consumption. Processing only non-zero input values ​​(i.e., dynamically sparse activation values) can reduce invalid computations, further improving computational efficiency and reducing power consumption. The number and size of MAC units are optimized, for example, it may be a 4x4 array, to save chip area and reduce circuit complexity. The multiply-accumulate unit is hardware-based to achieve faster parallel computing speeds than CPUs or general-purpose processors, which is especially beneficial for real-time anomaly detection scenarios.

[0052] In summary, the multiply-accumulate operator not only performs the core computational steps in CNNs, but also, through the aforementioned lightweight design, enables the lightweight convolutional array module to operate efficiently on resource-constrained automotive edge devices. This allows for the rapid and accurate extraction of features from image data, providing valuable input for subsequent anomaly detection and analysis. This improves the intelligent monitoring capabilities and response speed of the automotive system while ensuring its energy efficiency and sustainability.

[0053] The timing signal processing module processes signal data that varies over time, i.e., it preprocesses timing data. See also... Figure 1The aforementioned timing signal processing module 302 includes a Fourier accelerator 3021 and a long short-term memory network (LSTM) arithmetic unit 3022. The Fourier accelerator is used to perform Fourier acceleration processing on the aforementioned preprocessed timing data to obtain Fourier-accelerated timing data; the LTM arithmetic unit is used to perform time-domain analysis on the aforementioned preprocessed timing data to obtain time-domain analyzed data, and the aforementioned LTM arithmetic unit supports sliding window processing.

[0054] Specifically, Fourier accelerators are primarily used to transform preprocessed time-series data from the time domain to the frequency domain to extract the frequency characteristics of the signal. The Fourier transform is a mathematical tool that converts a time-varying signal into a set of frequency components. This is particularly important for analyzing time-fluctuating signals such as vibrations, sounds, and currents, as anomalies in these signals often manifest in specific frequency components. Fourier accelerators implement the Fast Fourier Transform (FFT) in hardware, a highly efficient discrete Fourier transform algorithm. In automotive applications, the accelerator's rapid processing capabilities significantly reduce the conversion time from the original signal to spectral characteristics, enabling real-time monitoring and anomaly detection. For example, for bearing vibration signals, Fourier accelerators can quickly identify fault characteristic frequencies, which is crucial for early fault warning and maintenance decisions.

[0055] Long Short-Term Memory (LSTM) network processors perform time-domain analysis on preprocessed time-series data, capturing long-term dependencies in signal sequences. LSTM is a special form of Recurrent Neural Network (RNN) used to address the vanishing or exploding gradient problems that traditional RNNs encounter when processing long sequences of data. LSTM processors process input time-series data using a sliding window mechanism, meaning they can move across the time series, progressively analyzing the data and updating their internal state at each step to reflect new information and long-term dependencies. LSTM processors can understand and capture complex patterns in signals in the time domain, such as gradually changing trends or periodic behavioral anomalies, which is extremely useful for monitoring train health and predicting potential faults.

[0056] By combining a Fourier accelerator and a Long Short-Term Memory (LSTM) network operator, the time-series signal processing module can efficiently and comprehensively analyze time-series data from different sensors on an in-vehicle edge computing device. It extracts spectral features and understands temporal patterns, thus providing more comprehensive data support for anomaly detection. This design overcomes the resource and time efficiency limitations of traditional processing methods, especially in resource-constrained in-vehicle environments, ensuring the real-time performance and accuracy of anomaly detection.

[0057] Through a lightweight convolutional array module and a temporal signal processing module, the heterogeneous computing unit can efficiently process images and temporal signals on edge devices, extract depth features, and provide data input for subsequent decision units, ultimately achieving high accuracy and low latency in vehicle anomaly detection while maintaining operation under hardware constraints of small area and low power consumption.

[0058] In some embodiments, see Figure 1 The aforementioned lightweight vehicle-mounted edge computing circuit module 01 also includes a dynamic power management unit 60, which is electrically connected to the aforementioned multi-source data interface 10, the aforementioned preprocessing unit 20, the aforementioned heterogeneous computing unit 30 and the aforementioned decision unit 40, respectively, and is used to dynamically adjust the voltage frequency of each unit connected to the aforementioned dynamic power management unit according to the amount of data of the aforementioned source data in order to perform graded energy consumption control.

[0059] Specifically, the Dynamic Power Management Unit (DPMU) is electrically connected to the multi-source data interface, preprocessing unit, heterogeneous computing unit, and decision-making unit. It monitors the data flow and processing demands of each component in real time, and then dynamically adjusts the operating voltage and frequency of these components to achieve refined and tiered energy consumption control. This design can minimize power consumption, extend equipment uptime, and reduce heat generation while ensuring system performance, thereby enhancing system reliability and security.

[0060] The Dynamic Power Management Unit (DPMU) continuously monitors the amount and type of data input from multiple data interfaces, as well as the current operating status of each processing unit. When the data volume is large or the processing task is heavy, it automatically identifies the issue and takes appropriate measures. Based on the monitored load, the DPMU can implement different energy consumption control strategies. For example, when data inflow is low or the processing task is light, it reduces voltage and frequency to put the system into energy-saving mode, reducing unnecessary power consumption; while under data-intensive or high-performance computing conditions, the DPMU increases voltage and frequency to ensure the processor has sufficient resources to complete the task in a timely manner. The DPMU achieves fine-grained control of voltage and frequency through Dynamic Voltage and Frequency Scaling (DVFS) technology. DVFS allows for flexible adjustment of supply voltage and processor frequency under different operating conditions to balance performance and power consumption. In idle or low-load conditions, reducing voltage and frequency can significantly reduce power consumption; while when the load increases, voltage and frequency are increased in a timely manner to ensure the rapid completion of computing tasks.

[0061] In summary, the Dynamic Power Management Unit (DPMU) dynamically adjusts the voltage and frequency of each unit based on the actual load, avoiding energy waste under low loads while ensuring computing performance under high loads, thus significantly improving the overall system's energy efficiency. Reducing power consumption by lowering voltage and frequency also reduces processor heat and wear, indirectly extending the lifespan of the circuit modules. Furthermore, the DPMU enables the system to flexibly respond to constantly changing operating conditions, finding the optimal energy consumption point to maintain high efficiency without sacrificing performance, whether it's a sudden surge in data volume or sustained low-load operation. In short, the DPMU provides an intelligent energy management and performance optimization mechanism for lightweight automotive edge computing circuit modules by real-time monitoring and dynamic adjustment of voltage and frequency, ensuring both real-time performance and accuracy while achieving low power consumption and long lifespan when processing variable automotive data.

[0062] In some embodiments of this application, see Figure 1 The aforementioned decision-making unit 40 includes a multimodal fusion module 401 and a lightweight classification module 402. The multimodal fusion module is used to perform weighted fusion processing on the aforementioned image feature data and the aforementioned temporal feature data to obtain a weighted fusion result; the lightweight classification module is used to perform classification processing on the aforementioned source data based on the aforementioned weighted fusion result to obtain the anomaly level and confidence level of the aforementioned source data.

[0063] Specifically, the multimodal fusion module performs weighted fusion processing on image feature data extracted from the lightweight convolutional array module and temporal feature data extracted from the temporal signal processing module to form a more comprehensive analysis result. Weighted fusion processing involves assigning different weights to different types of data features, and then combining the weighted feature data to generate a unified decision vector or result. The weights are typically set based on an assessment of the contribution of different feature data to anomaly detection. For example, in some cases, image features may more directly reveal anomalies, while temporal features may reflect the development trend of anomalies; therefore, their weight allocation needs to be determined according to the actual application scenario. The multimodal fusion module can employ various fusion strategies, including feature-level fusion and decision-level fusion.

[0064] The lightweight classification module determines the anomaly level and confidence level of the source data based on weighted fused feature data. It utilizes lightweight machine learning models, such as Support Vector Machines (SVM), decision trees, or variants, trained to adapt to the low-resource environment of in-vehicle edge computing. The module quickly processes the fused feature data, comparing it with known normal and abnormal states in the database to output a classification result indicating whether the current data represents an abnormal state and its specific level (e.g., mild, moderate, severe). By identifying anomalies and further determining their severity based on the fused feature data, the module helps in implementing different levels of response measures; for example, a minor anomaly triggers a warning, while a severe anomaly immediately initiates an emergency shutdown procedure. In addition to the classification result, a confidence score is output, reflecting the reliability of the classification. High-confidence anomaly detection results indicate a stronger need for response, while low-confidence results may require further manual review or system calibration.

[0065] Integrating the lightweight onboard edge computing circuit module of this embodiment into a single IC (integrated circuit) not only saves space but also significantly improves system performance and reliability. Through module-level optimization and system-level expansion, coupled with rigorous safety redundancy design, it can provide a highly efficient and robust onboard anomaly detection platform for rail transit, meeting the high requirements of future intelligent transportation systems.

[0066] Through the collaborative work of the multimodal fusion module and the lightweight classification module, the decision unit can achieve comprehensive analysis of image and time-series data, ensuring the comprehensiveness and accuracy of anomaly detection. Simultaneously, the lightweight design ensures real-time processing capabilities and power efficiency in the automotive environment. This allows the lightweight automotive edge computing circuit module to provide efficient and reliable services even under limited computing resources.

[0067] In addition to real-time anomaly detection, some embodiments of this application introduce a behavior recognition-based anomaly prediction algorithm for proactively assessing potential anomaly risks. The algorithm mainly includes: Behavioral pattern learning: Learning "healthy behavior patterns" of various train components from a large amount of data collected during normal operation using unsupervised learning techniques (such as autoencoders and clustering algorithms). Real-time behavior comparison: Collecting sensor data in real-time during train operation and comparing it with the learned healthy behavior patterns to assess the degree of deviation between the current behavior and the normal pattern. Anomaly risk prediction: Based on the degree of deviation and time series analysis, predicting the probability of abnormal events occurring in the future (e.g., the next few hours), providing early warnings and a basis for intelligent scheduling and maintenance. This behavior recognition-based anomaly prediction algorithm expands the chip's functionality, shifting from passive anomaly detection to proactive risk management, providing more comprehensive safety assurance for train operation.

[0068] This embodiment provides a lightweight in-vehicle edge computing method that runs on a mobile terminal, computer terminal, or similar computing device. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0069] The lightweight vehicle edge computing method of this application embodiment is implemented using the lightweight vehicle edge computing circuit module in the above embodiment. Figure 2 This is a flowchart illustrating a lightweight vehicle-mounted edge computing method according to an embodiment of this application. Figure 2 As shown, the method includes the following steps:

[0070] Step S201: Establish a communication connection with the vehicle-mounted equipment to obtain various types of source data from the vehicle-mounted equipment;

[0071] Specifically, communication connections are established with various sensors and monitoring systems on the train to acquire multiple types of source data from onboard equipment, including image data and non-image data. This data is obtained through various interfaces to ensure that information can be acquired from various sensors on the train (such as infrared cameras, vibration sensors, and current sensors). The acquisition of source data is a continuous and real-time process, ensuring the timeliness and accuracy of anomaly detection.

[0072] Establishing a communication connection with in-vehicle devices to obtain various types of source data is the first step in a lightweight in-vehicle edge computing approach. By ensuring the real-time nature and integrity of the data, it provides a solid foundation for subsequent data processing and anomaly detection.

[0073] Step S202: Perform predetermined processing on the above-mentioned source data of various types to obtain predetermined processed data. The predetermined processed data includes preprocessed image data and preprocessed time series data. The predetermined processing includes denoising processing and normalization processing.

[0074] Specifically, the pre-processing mainly includes denoising and normalization. Denoising aims to eliminate noise interference during data acquisition and ensure data quality, typically involving digital filtering and signal smoothing. Normalization standardizes the data, ensuring that data from different sensors are compared and processed on a uniform scale, which is crucial for subsequent feature extraction and data analysis. Specifically, denoising removes unnecessary noise interference from the source data, preserving useful signal information. Due to the complex operating environment of trains, sensor data is easily affected by electromagnetic interference, mechanical vibration, and other factors, resulting in noise in the data. Digital signal processing techniques, such as filtering and smoothing, can effectively identify and eliminate these noise effects, making images or time-series signals clearer and improving the accuracy of feature extraction and anomaly detection. Normalization is the process of converting data to the same scale, which is particularly important for the fusion analysis of multimodal data. By employing appropriate normalization strategies, such as minimum-maximum scaling and z-score normalization, the data can be transformed to the [0,1] range or other preset ranges. This helps to reasonably compare and weight subsequent image feature data and time-series feature data during fusion, avoiding analytical bias caused by differences in data scale.

[0075] Denoising and normalization significantly improve the quality of source data, making it more suitable for anomaly detection and analysis. Denoising reduces the masking effect of noise on signal authenticity, while normalization ensures consistency across different data types, facilitating subsequent feature extraction and model input. This not only improves the accuracy and reliability of anomaly detection but also reduces computational resource requirements, as the dimensionality and complexity of the data are effectively controlled after preprocessing.

[0076] Step S203: Process the above preprocessed image data to obtain image feature data, and process the above preprocessed time series data to obtain time series feature data;

[0077] Specifically, image feature data is obtained by processing preprocessed image data, and time-series feature data is obtained by processing preprocessed time-series data. Through a specially optimized image and time-series data processing flow, efficient and real-time feature extraction is achieved, providing a solid foundation for subsequent anomaly decision-making. Meanwhile, its low power consumption, small area, and multimodal data processing capabilities make it a core component of lightweight automotive edge computing methods, significantly improving the intelligence level and security capabilities of automotive systems.

[0078] Step S204: Make a decision based on the above image feature data and the above time-series feature data.

[0079] Specifically, based on image feature data and time-series feature data, pre-trained classification models (such as support vector machines, decision trees, neural networks, etc.) are used to perform data analysis and judgment to determine whether any abnormal situations exist. Decision-making is based on image feature data and time-series feature data. This process involves in-depth analysis of the extracted features to identify whether there are behaviors or states outside the normal range, and to provide timely warnings of potential faults.

[0080] This approach, based on the comprehensive analysis of image and temporal feature data, significantly enhances the intelligence and response speed of lightweight onboard edge computing methods. It can quickly and accurately identify abnormal changes in train operating conditions, such as bearing overheating or sudden changes in vibration modes. The real-time and high-precision decision-making directly contributes to the timely response and effective management of the onboard system in abnormal situations, significantly improving the safety and reliability of train operations.

[0081] This lightweight onboard edge computing method establishes a communication connection with onboard equipment to acquire diverse source data in real time. It then performs preprocessing such as denoising and normalization to ensure data quality and consistency. Dedicated image processing and time-series data analysis techniques are employed to extract image feature data and time-series feature data, enriching the information dimensions for anomaly detection. Based on these image and time-series features, rapid decisions can be made to identify abnormal changes in train operating status. This significantly improves the real-time performance and accuracy of onboard edge computing in anomaly detection, providing strong technical support for safe train operation while also optimizing data processing efficiency and reducing reliance on cloud resources.

[0082] In the specific implementation process, the above-mentioned preprocessed image data is processed to obtain image feature data, and the above-mentioned preprocessed time series data is processed to obtain time series feature data, including: using a lightweight convolution algorithm including a binarization sparse processing algorithm to process the above-mentioned preprocessed image data to obtain the above-mentioned image feature data; performing Fourier acceleration processing and long short-term memory network operations on the above-mentioned preprocessed time series data to perform joint analysis in the time domain and frequency domain to obtain the above-mentioned time series feature data.

[0083] Specifically, a lightweight convolutional algorithm, including a binarized sparse processing algorithm, is employed. This algorithm significantly reduces the computational complexity and memory requirements of the convolutional neural network by quantizing weights to binary (1 bit) and sparsifying activation values. This enables the efficient operation of complex image feature extraction models on resource-constrained automotive edge devices without resulting in significant increases in power consumption or latency. Preprocessed image data (such as infrared thermal images) is fed into a lightweight convolutional array (LCA), which utilizes a binarized sparse convolution algorithm to extract key features from the image, such as the location, size, and shape of hotspots, through a series of convolution and pooling operations. These features are then represented as compact image feature data for subsequent anomaly detection and classification.

[0084] Preprocessed time-series data (such as vibration signals or current changes) are analyzed using a Time-Series Signal Processor (TSP). This TSP employs Fourier transform technology to quickly extract the spectral characteristics of the time-series data, such as fault characteristic frequencies and power spectral density. Long Short-Term Memory (LSTM) networks are used to analyze the time-series characteristics of the signals, capturing long-term dependencies and short-term dynamic changes in the time-series data to identify potential anomalous patterns.

[0085] Fourier transform is used to convert time-domain signals into frequency-domain representations, which helps identify frequency-related anomalies, such as the characteristic frequencies of bearing failures. LSTM networks perform in-depth analysis in the time domain to identify anomalies in timing patterns, such as sudden current spikes. Through joint analysis in the time and frequency domains, anomalous information in the signal can be comprehensively captured, improving the coverage and accuracy of anomaly detection.

[0086] The application of binarization sparse processing and Fourier acceleration techniques significantly improves the processing speed of image and time-series data while reducing the demand for computing resources and power consumption, making it suitable for the operating environment of automotive edge computing devices. Through the synergistic effect of lightweight convolutional arrays and time-series signal processors, multi-dimensional features are extracted from image and time-series data, providing a rich and comprehensive information foundation for subsequent anomaly decision-making. In summary, in the feature extraction stage, specially designed algorithms and techniques enable efficient and comprehensive analysis of image and time-series data, providing strong support for anomaly detection and real-time response in automotive devices.

[0087] Furthermore, the above method also includes: determining the data volume of the above source data of various types based on the sampling frequency, data type, and data transmission protocol of the above source data; and dynamically adjusting the voltage frequency of each unit in the vehicle edge computing circuit module based on the data volume of the above source data to perform graded energy consumption control.

[0088] Specifically, sensors in automotive equipment generate and transmit various types of source data in real time using different sampling frequencies, data types, and data transmission protocols. To effectively process this data, the first step is to accurately calculate the data volume for each type of source data. This step is based on the sampling frequency, data type, and data transmission protocol. The sensor's sampling frequency determines the number of data points per unit time. For example, a temperature sensor sampling at 100Hz will generate 100 data points per second. Different types of data (such as images and timing signals) have different storage formats and sizes. For example, image data might be stored in JPEG format, while timing signal data might be a set of floating-point numbers. The overhead of data transmission protocols (such as CAN, MIPI-CSI2, and Ethernet) also affects the actual amount of data transmitted. For example, a CAN message includes a frame header, data segments, and checksum information; the actual data volume will be less than the number of bytes transmitted. By analyzing these parameters, the amount of data generated by each sensor per unit time can be estimated, providing a basis for subsequent data processing and resource allocation.

[0089] The in-vehicle edge computing circuit module comprises multiple units responsible for different tasks. To reduce overall power consumption while ensuring real-time computing performance, the current data processing load is assessed based on the aforementioned data volume. When the load is low, the number of active computing units can be reduced; when the load is high, all units need to be kept operational to meet processing demands. Dynamic Voltage and Frequency Scaling (DVFS) technology is used to adjust the operating frequency and supply voltage of each computing unit according to the real-time data processing load. For example, in scenarios with small data volumes and low processing loads, voltage and frequency are reduced to decrease power consumption; as data volumes increase and processing loads intensify, voltage and frequency are increased to ensure processing speed and accuracy. Energy consumption tiered management is implemented, employing different energy control strategies under different data processing loads. For example, under low load conditions, an energy-saving mode can be entered, shutting down unnecessary computing units; under high load conditions, a high-performance mode is entered, with all units operating at full capacity to ensure the real-time performance and integrity of data processing.

[0090] Based on the dynamic demands of data processing, the voltage and frequency of the computing unit are intelligently adjusted, achieving a balance between power consumption and performance and improving the resource utilization efficiency of the in-vehicle edge computing device. Dynamic power management reduces unnecessary energy consumption, helping to extend the lifespan of the in-vehicle equipment and lower maintenance costs. This mechanism can automatically adapt to changes in data volume and processing load, ensuring that the device maintains optimal operating conditions under different operating conditions. In short, by dynamically assessing the amount of source data and adjusting the power consumption of the in-vehicle edge computing circuit module accordingly, not only is efficient utilization of computing resources achieved, but the adaptability and reliability of the device are also improved.

[0091] Furthermore, the voltage frequency of each unit in the vehicle edge computing circuit module is dynamically adjusted according to the amount of the source data to perform graded energy consumption control, including: when the amount of the source data is greater than or equal to a preset amount of data, the voltage frequency of each unit in the vehicle edge computing circuit module is controlled to be the full voltage frequency; when the amount of the source data is less than the preset amount of data, the voltage frequency of the target unit in the vehicle edge computing circuit module is controlled to be a non-full voltage frequency, wherein the target unit is an idle unit or a non-full-load working unit.

[0092] Specifically, when the amount of source data is greater than or equal to a preset amount, all units in the vehicle-mounted edge computing circuit module operate at full voltage and frequency. This means that in urgent or busy scenarios with high data volumes, all computing resources are fully utilized to ensure that data can be processed quickly and effectively, achieving maximum operating efficiency. When the amount of source data is less than the preset amount, a more energy-efficient strategy is adopted, reducing the voltage and frequency to a lower frequency only for idle units or units not operating at full load. In this way, unnecessary power consumption can be reduced under low load conditions while maintaining necessary computing power, ensuring a balance between response speed and power efficiency.

[0093] Idle units refer to computing units that are currently not performing data processing tasks. In an automotive edge computing environment, due to the bursty and discontinuous nature of data flows, some computing units may be waiting for data at certain times; these units are marked as idle units. Non-full-loaded units refer to computing units that are performing data processing tasks, but whose processing capacity has not yet reached saturation. Although these units have data processing tasks, the workload is insufficient to fully utilize their computing resources.

[0094] The preset data volume is determined based on the system's maximum data processing capacity and the average data volume under normal operation. For example, if the system's maximum processing capacity is 2GB of data per second, and the average data volume under normal conditions is 500MB / s, then the preset data volume can be set around 1GB / s to ensure that the system operates in a highly efficient and energy-saving state in most situations.

[0095] Assuming the system can handle a stable data volume of 500MB / s under normal operating conditions, a preset data volume of 600MB / s is set to ensure responsiveness and computational power when data volume suddenly increases. When the data volume is below the preset 600MB / s, all units operate at full voltage and frequency. For example, assuming a full voltage and frequency of 1.2GHz and a voltage of 1.2V, the system is in "high-performance mode," ensuring optimal computational performance even before data volume increases, capable of handling brief spikes in data volume. Once the data volume reaches or exceeds 600MB / s, the voltage and frequency of some units are adjusted. For example, the preprocessing unit, primarily responsible for data preprocessing, can reduce its voltage and frequency to 1GHz and voltage to 1.0V when data volume is high to reduce power consumption, as its computational requirements are relatively low. The heterogeneous computing unit, responsible for critical feature extraction tasks, has higher computational performance requirements. When the amount of data increases, it is advisable not to reduce the voltage frequency temporarily. However, if the total power consumption of the system is close to overload, the frequency can be reduced to 1.1GHz and the voltage to 1.1V to maintain the stable operation of the system.

[0096] Here are a few example scenarios:

[0097] Scenario 1: During train operation late at night, the data volume is relatively low, averaging 300MB / s. At this time, all control units operate at full voltage and frequency.

[0098] Scenario 2: During train startup and acceleration, the data volume suddenly increases to 700MB / s, exceeding the preset threshold of 600MB / s. At this point, the voltage frequency of the preprocessing unit is reduced to 1GHz, and the voltage is reduced to 1.0V. Simultaneously, the real-time power consumption of the heterogeneous computing unit is monitored. If the power consumption approaches overload, the voltage frequency of the heterogeneous computing unit is also reduced accordingly, but it is kept at a high operating level to ensure that the extraction of key features is not affected.

[0099] Scenario 3: During emergency braking or fault detection, the data volume surges to 1.2 GB / s. To ensure the real-time nature of anomaly detection, the frequency reduction operation of non-critical units is temporarily suspended, ensuring that the decision unit operates at full voltage frequency, while other units adjust according to real-time load and power consumption.

[0100] By intelligently adjusting the voltage frequency, overall energy consumption can be significantly reduced without affecting critical system functions. This dynamic adjustment strategy is particularly important when dealing with large fluctuations in data volume. Properly controlling the voltage frequency helps maintain the device operating within a safe range, preventing hardware overheating or damage due to excessive power consumption. The dynamic voltage frequency adjustment strategy can flexibly adjust computing resources according to actual data processing needs, ensuring optimal computing performance and energy efficiency under all circumstances. In summary, this dynamic voltage frequency adjustment strategy effectively manages energy consumption and heat dissipation while maintaining computing performance, ensuring system stability and reliability, and adapting to data processing needs in various operating scenarios.

[0101] In some embodiments of this application, decision-making based on the aforementioned image feature data and the aforementioned time-series feature data includes: performing weighted fusion processing on the aforementioned image feature data and the aforementioned time-series feature data to obtain a weighted fusion result, wherein the weight parameters in the aforementioned weighted fusion processing are related to the amount of data in the aforementioned image feature data and the amount of data in the aforementioned time-series feature data; classifying the aforementioned source data based on the aforementioned weighted fusion result to obtain the anomaly level and confidence level of the aforementioned source data.

[0102] Specifically, the determination of weighting parameters is related to the amount of image feature data and temporal feature data. Feature types with larger data volumes may carry richer information, and therefore are given higher weights during the fusion process to ensure the accuracy and comprehensiveness of anomaly detection. For example, if the amount of image data (such as the resolution and frame rate of an infrared thermal image) is much larger than the amount of temporal signal data (such as vibration frequency data points) at a certain moment, then the proportion of image feature data in the fusion result will be larger. Image feature data and temporal feature data are weighted and fused according to their respective weighting parameters, and anomaly detection is performed based on the weighted fusion result to obtain anomaly detection results, including anomaly level and anomaly detection confidence. The anomaly level is divided according to the severity of the anomaly detection result. For example, the anomaly level can be divided into level 0 (normal), level 1 (minor anomaly), level 2 (moderate anomaly), and level 3 (severe anomaly), corresponding to different alarm and handling strategies. Confidence measures the model's certainty about the anomaly detection result, usually represented by a value between 0 and 1. High confidence means the detection result is more reliable, while low confidence requires further verification or review. For example, if the detected fusion features highly match known anomalous patterns, the confidence level may be as high as 0.95, and corresponding emergency measures should be triggered immediately.

[0103] By adjusting the weights of feature fusion to account for dynamic changes in data volume, key information can be captured more effectively, thus improving the accuracy of anomaly detection. Classification decisions based on the fusion results, combined with anomaly level and confidence output, enable more accurate assessment of anomaly severity during detection, allowing for timely and appropriate response measures. Dynamic adjustment of weight parameters allows the system to flexibly adapt to changes in data volume under different operating conditions, ensuring optimal anomaly detection performance in all situations. Through weighted fusion processing and classification decisions based on the fusion results, anomaly detection can be performed more intelligently, providing strong technical support for the safe operation of trains.

[0104] Furthermore, based on the weighted fusion results, the source data is classified to obtain the anomaly level and confidence level of the source data, including: using a hybrid model of decision tree and support vector machine to classify the weighted fusion results to obtain the anomaly level and confidence level of the source data.

[0105] Specifically, a decision tree (DT) is a machine learning model that classifies data through a series of conditional judgments. In automotive edge computing scenarios, decision trees can serve as a preliminary filter, quickly filtering out data points that are clearly not anomalies, thus reducing computational load. Support Vector Machines (SVMs) are supervised learning models used for classification and regression tasks, particularly adept at handling high-dimensional data and complex nonlinear classification problems. In anomaly detection, SVMs can accurately distinguish between normal and anomalous data points, further refining the classification after the initial screening by the decision tree, thereby improving the accuracy of anomaly detection.

[0106] The weighted fusion data is used as input to a hybrid model of decision trees and support vector machines (SVMs). First, the decision tree model acts as a fast filter, performing preliminary classification of the fusion results based on a predefined set of rules and feature thresholds. If a data point is determined to be "non-abnormal" in the decision tree's decision path, it is directly classified as normal without further processing; otherwise, it enters the SVM model for more detailed classification. Data points that the decision tree fails to explicitly exclude as "suspected abnormalities" are submitted to the SVM model for final classification. The SVM, based on the decision boundary constructed from the training set, accurately determines the category (normal / abnormal) of each data point. Based on the classification results from the SVM model, the severity of the abnormality is assessed, categorizing it into different levels (e.g., level 0 - normal, level 1 - slight abnormality, level 2 - moderate abnormality, level 3 - severe abnormality). In addition to outputting class labels, the SVM model also provides a confidence score during classification. This score is typically between 0 and 1; a higher value indicates greater certainty about the classification result. Based on the confidence score, additional weight is given to the detected abnormal events to determine whether to trigger an alarm and corresponding emergency measures.

[0107] The initial screening by the decision tree model significantly reduces the computational burden on the SVM model, ensuring the efficiency of the entire classification process, which is crucial for real-time onboard applications. The SVM model ensures classification accuracy, especially when handling complex, non-linear anomaly patterns. The use of a hybrid model compensates for the limitations of a single model in different scenarios. Confidence scoring provides a metric for anomaly detection, allowing for flexible adjustment of alarm thresholds based on actual needs, avoiding both over-alarms and ignoring potential risks. By employing a hybrid model combining decision trees and support vector machines, the accuracy and robustness of anomaly detection are significantly improved while maintaining real-time performance, providing strong support for the safe monitoring of train operations.

[0108] The following is a specific example of obtaining the anomaly level and confidence level of source data:

[0109] Suppose a train is traversing a complex terrain, resulting in a surge in data volume (images and time-series data). After preprocessing, the feature data obtained through weighted fusion is input into a hybrid model. In the decision tree model, data points are initially labeled as "suspected anomalies." Subsequently, in the SVM model, after decision boundary judgment, they are ultimately determined to be "moderate anomalies" with a confidence score of 0.75. Based on this score, they are marked as Level 2 anomalies, and the train control center is promptly notified to take appropriate preventative measures or maintenance plans according to the specific context.

[0110] To enable those skilled in the art to better understand the technical solution of this application, the implementation process of the lightweight vehicle edge computing method of this application will be described in detail below with reference to specific embodiments.

[0111] This embodiment relates to a specific lightweight in-vehicle edge computing method, which is implemented using a specific lightweight in-vehicle edge computing module. A schematic diagram of the specific lightweight in-vehicle edge computing module is shown below. Figure 3 As shown.

[0112] 1) Multi-source data interface:

[0113] It supports IEEE 802.3 (Ethernet), CAN bus, and MIPI-CSI2 (camera) interface to receive sensor data in real time.

[0114] Built-in preprocessing unit (PU): Denoises, normalizes, and unifies the format of input data (e.g., converts vibration signals to spectrograms).

[0115] 2) Heterogeneous computing units:

[0116] Lightweight Convolutional Array (LCA): 4×4 configurable MAC units, supporting 1D / 2D convolution operations for image feature extraction.

[0117] By employing binary weights (1 bit) and dynamically sparsifying activation values, computational energy consumption is reduced by 70%.

[0118] Timing Signal Processor (TSP): Dedicated FFT accelerator and LSTM unit for processing vibration and current signals.

[0119] It has a built-in sliding window mechanism that supports joint time-domain and frequency-domain analysis.

[0120] 3) Dynamic Power Management Unit (DPMU):

[0121] Dynamically adjust voltage frequency (DVFS) based on data load, and shut down power to inactive modules in idle state.

[0122] Tiered energy consumption mode:

[0123] Normal mode (100MHz, 3W): Full-featured operation.

[0124] Energy Saving Mode (50MHz, 1.2W): Only activates the basic detection engine.

[0125] 4) Abnormal Decision Engine (Decision Unit):

[0126] Multimodal fusion module: weighted fusion of image and time-series signal features (weights are determined by offline training).

[0127] Lightweight classifier: A hybrid model based on decision trees and support vector machines (SVM), supporting hardware acceleration.

[0128] 5) Output and feedback interfaces:

[0129] Output the anomaly level (0-3) and confidence level.

[0130] The model is fine-tuned online (incremental learning) based on continuous false alarm data.

[0131] Data processing flow:

[0132] Step 1: Multi-source data input and preprocessing: Multi-channel sensor data is converted into a unified tensor format (e.g., size 256×256×3) by the PU module.

[0133] Step 2: Parallel feature extraction. LCA extracts spatial features from the image (e.g., infrared hotspots on the gearbox transmitted from the camera). TSP extracts frequency domain features from the vibration signal (e.g., the bearing fault characteristic frequency of 32Hz).

[0134] Step 3: Feature fusion and anomaly detection, see [link / reference] Figure 4 After receiving multi-source data, the system performs vibration signal preprocessing to extract key vibration signal features and image data preprocessing to extract spatial features. Using binarization weights and sparse computation strategies, efficient convolution operations are performed on the image data to generate image feature vectors. Fast Fourier Transform (FFT) is then performed, and combined with LSTM unit analysis to analyze the dynamic characteristics of the time-series signal and extract time-series signal feature vectors. The extracted image feature vectors and time-series signal feature vectors are then fused using a weighted formula: Joint = 0.6 × Img + 0.4 × Spec, where Img and Spec are the image feature and spectral feature vectors, respectively. The fused feature vectors are then fed to an SVM classifier, which outputs anomaly probabilities. If the anomaly probability output by the classifier exceeds a set threshold (e.g., 0.9), the source data is determined to be abnormal. Further analysis of the anomaly level and confidence level triggers corresponding alarm mechanisms. For example, a level 1 alarm only records the event, a level 2 alarm notifies the operator, and a level 3 alarm immediately triggers emergency braking. Normal samples are stored, and the classifier parameters are updated through an incremental learning mechanism to adapt to signal feature drift caused by equipment aging or environmental changes.

[0135] Step 4: Dynamic Power Consumption Adjustment: The DPMU switches power consumption modes in real time according to the task load (e.g., entering power-saving mode when idle). See [link to relevant documentation]. Figure 5The initial state is power-saving mode, with the frequency set to 50MHz and the voltage set to 0.8V. In this mode, the chip only enables basic computing and communication functions to maintain the lowest power consumption level. If the load exceeds 80% and this state lasts for more than 5 milliseconds, it automatically switches to normal mode. In this mode, all chip functions are fully activated to handle higher data processing demands. All computing units and communication interfaces are active to ensure real-time data processing and transmission. If the load drops below 30% and lasts for more than 10 milliseconds, it will automatically return to power-saving mode to conserve energy. If the chip temperature rises to or above 80°C, it immediately switches to overload protection mode to prevent chip damage. In this mode, a forced frequency reduction operation is performed to reduce chip heat and protect the chip from damage. When the chip temperature drops below 75°C, it returns to power-saving mode.

[0136] The specific lightweight onboard edge computing module is deployed at the train's onboard edge computing node. It monitors the bogie status in real time through vibration sensors and infrared cameras. When an abnormality is detected (such as gearbox overheating), an alarm is triggered within 10ms and the lightweight diagnostic results are uploaded to the driver's cab.

[0137] This application also provides a lightweight in-vehicle edge computing device. It should be noted that this lightweight in-vehicle edge computing device can be used to execute the lightweight in-vehicle edge computing method provided in this application. This device is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0138] The following describes the lightweight vehicle-mounted edge computing device provided in the embodiments of this application.

[0139] Figure 6 This is a structural block diagram of a lightweight vehicle-mounted edge computing device according to an embodiment of this application. Figure 6As shown, the device includes an acquisition unit 100, a pre-processing unit 200, a processing unit 300, and an anomaly decision unit 400. The acquisition unit establishes a communication connection with the vehicle-mounted equipment to acquire various types of source data from the equipment. The pre-processing unit performs pre-processing on the various types of source data to obtain pre-processed data, which includes pre-processed image data and pre-processed time-series data. The pre-processing includes denoising and normalization. The processing unit processes the pre-processed image data to obtain image feature data and processes the pre-processed time-series data to obtain time-series feature data. The anomaly decision unit makes a decision based on the image feature data and the time-series feature data.

[0140] By establishing a communication connection with onboard equipment, diverse source data is acquired in real time and preprocessed, including denoising and normalization, ensuring data quality and consistency. Dedicated image processing and time-series data analysis techniques are employed to extract image feature data and time-series feature data, enriching the information dimensions for anomaly detection. Based on these image and time-series features, rapid decisions can be made to identify abnormal changes in train operating status. This significantly improves the real-time performance and accuracy of onboard edge computing in anomaly detection, providing strong technical support for safe train operation while also optimizing data processing efficiency and reducing reliance on cloud resources.

[0141] In the specific implementation process, the above-mentioned processing unit includes a first processing module and a second processing module. The first processing module is used to process the above-mentioned preprocessed image data using a lightweight convolution algorithm including a binarization sparse processing algorithm to obtain the above-mentioned image feature data; the second processing module is used to perform Fourier acceleration processing and long short-term memory network operations on the above-mentioned preprocessed time-series data to perform joint analysis in the time domain and frequency domain to obtain the above-mentioned time-series feature data.

[0142] The application of binarization sparse processing and Fourier acceleration techniques significantly improves the processing speed of image and time-series data while reducing the demand for computing resources and power consumption, making it suitable for the operating environment of automotive edge computing devices. Through the synergistic effect of lightweight convolutional arrays and time-series signal processors, multi-dimensional features are extracted from image and time-series data, providing a rich and comprehensive information foundation for subsequent anomaly decision-making. In summary, in the feature extraction stage, specially designed algorithms and techniques enable efficient and comprehensive analysis of image and time-series data, providing strong support for anomaly detection and real-time response in automotive devices.

[0143] Furthermore, the aforementioned device also includes a determining unit and an adjusting unit. The determining unit is used to determine the data volume of the various types of source data based on the sampling frequency, data type, and data transmission protocol of the various types of source data; the adjusting unit is used to dynamically adjust the voltage frequency of each unit in the vehicle edge computing circuit module according to the data volume of the source data to perform graded energy consumption control.

[0144] Based on the dynamic demands of data processing, the voltage and frequency of the computing unit are intelligently adjusted, achieving a balance between power consumption and performance and improving the resource utilization efficiency of the in-vehicle edge computing device. Dynamic power management reduces unnecessary energy consumption, helping to extend the lifespan of the in-vehicle equipment and lower maintenance costs. This mechanism can automatically adapt to changes in data volume and processing load, ensuring that the device maintains optimal operating conditions under different operating conditions. In short, by dynamically assessing the amount of source data and adjusting the power consumption of the in-vehicle edge computing circuit module accordingly, not only is efficient utilization of computing resources achieved, but the adaptability and reliability of the device are also improved.

[0145] Furthermore, the aforementioned adjustment unit includes a first control module and a second control module. The first control module is used to control the voltage frequency of each unit in the vehicle edge computing circuit module to be at full voltage frequency when the amount of the source data is less than a preset amount of data; the second control module is used to control the voltage frequency of at least some of the units in the vehicle edge computing circuit module to be at a non-full voltage frequency when the amount of the source data is greater than or equal to the preset amount of data.

[0146] By intelligently adjusting the voltage frequency, overall energy consumption can be significantly reduced without affecting critical system functions. This dynamic adjustment strategy is particularly important when dealing with large fluctuations in data volume. Properly controlling the voltage frequency helps maintain the device operating within a safe range, preventing hardware overheating or damage due to excessive power consumption. The dynamic voltage frequency adjustment strategy can flexibly adjust computing resources according to actual data processing needs, ensuring optimal computing performance and energy efficiency under all circumstances. In summary, this dynamic voltage frequency adjustment strategy effectively manages energy consumption and heat dissipation while maintaining computing performance, ensuring system stability and reliability, and adapting to data processing needs in various operating scenarios.

[0147] In some embodiments of this application, the aforementioned anomaly decision-making unit includes a weighted fusion module and a classification module. The weighted fusion module is used to perform weighted fusion processing on the aforementioned image feature data and the aforementioned temporal feature data to obtain a weighted fusion result, wherein the weight parameters in the aforementioned weighted fusion processing are related to the data volume of the aforementioned image feature data and the data volume of the aforementioned temporal feature data; the classification module is used to classify the aforementioned source data based on the aforementioned weighted fusion result to obtain the anomaly level and confidence level of the aforementioned source data.

[0148] By adjusting the weights of feature fusion to account for dynamic changes in data volume, key information can be captured more effectively, thus improving the accuracy of anomaly detection. Classification decisions based on the fusion results, combined with anomaly level and confidence output, enable more accurate assessment of anomaly severity during detection, allowing for timely and appropriate response measures. Dynamic adjustment of weight parameters allows the system to flexibly adapt to changes in data volume under different operating conditions, ensuring optimal anomaly detection performance in all situations. Through weighted fusion processing and classification decisions based on the fusion results, anomaly detection can be performed more intelligently, providing strong technical support for the safe operation of trains.

[0149] Furthermore, the classification module includes a classification submodule, which uses a hybrid model of decision tree and support vector machine to classify the weighted fusion result to obtain the anomaly level and confidence level of the source data.

[0150] The initial screening by the decision tree model significantly reduces the computational burden on the SVM model, ensuring the efficiency of the entire classification process, which is crucial for real-time onboard applications. The SVM model ensures classification accuracy, especially when handling complex, non-linear anomaly patterns. The use of a hybrid model compensates for the limitations of a single model in different scenarios. Confidence scoring provides a metric for anomaly detection, allowing for flexible adjustment of alarm thresholds based on actual needs, avoiding both over-alarms and ignoring potential risks. By employing a hybrid model combining decision trees and support vector machines, the accuracy and robustness of anomaly detection are significantly improved while maintaining real-time performance, providing strong support for the safe monitoring of train operations.

[0151] The aforementioned lightweight vehicle-mounted edge computing device includes a processor and a memory. The acquisition unit, pre-processing unit, processing unit, and anomaly decision-making unit are all stored as program units in the memory. The processor executes these program units stored in the memory to achieve the corresponding functions. All of the above modules reside in the same processor; alternatively, the modules may be located in different processors in any combination.

[0152] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0153] This invention also provides a rail train, comprising: a vehicle body; various on-board devices installed on the vehicle body and any one of the aforementioned lightweight on-board edge computing circuit modules, wherein the various on-board devices are electrically connected to the lightweight on-board edge computing circuit module via a multi-source data interface, and wherein the various on-board devices include an on-board air conditioner, an on-board display device, and an on-board communication device.

[0154] This invention provides a computer-readable storage medium including a stored program, wherein the program, when running, controls the device containing the computer-readable storage medium to execute the lightweight vehicle edge computing method.

[0155] This invention provides a processor for running a program, wherein the program executes the lightweight vehicle edge computing method.

[0156] This invention provides an electronic device, including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the aforementioned lightweight in-vehicle edge computing method. The device described herein can be a server, PC, PAD, mobile phone, etc.

[0157] This application also provides a computer program product that, when executed on a data processing device, is adapted to perform a program that initializes the steps of the aforementioned lightweight vehicle edge computing method.

[0158] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0159] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0161] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0162] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0163] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0164] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0165] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0166] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0167] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A lightweight vehicle-mounted edge computing circuit module, characterized in that, include: A multi-source data interface is used to electrically connect to an in-vehicle device to receive various types of source data from the in-vehicle device, wherein the multi-source data interface includes an image data interface, a non-image data interface, and a bus interface; The preprocessing unit is electrically connected to the multi-source data interface and is used to perform predetermined processing on the source data of various types to obtain predetermined processed data. The predetermined processed data includes preprocessed image data and preprocessed time series data. The predetermined processing includes denoising processing and normalization processing. A heterogeneous computing unit, electrically connected to the preprocessing unit, is used to process the preprocessed image data to obtain image feature data, and to process the preprocessed time series data to obtain time series feature data. The decision unit, connected to the heterogeneous computing unit, is used to make decisions based on the image feature data and the temporal feature data.

2. The lightweight vehicle-mounted edge computing circuit module according to claim 1, characterized in that, The heterogeneous computing unit includes: A lightweight convolutional array module is used to perform lightweight convolutional processing on the preprocessed image data to obtain the image feature data; The timing signal processing module is used to process the preprocessed timing data to obtain the timing feature data.

3. The lightweight vehicle-mounted edge computing circuit module according to claim 2, characterized in that, The lightweight convolutional array module includes: The multiply-accumulate operator is used to perform the multiply-accumulate operation in the lightweight convolution processing to obtain the multiply-accumulate result.

4. The lightweight vehicle-mounted edge computing circuit module according to claim 2, characterized in that, The timing signal processing module includes: A Fourier accelerator is used to perform Fourier acceleration processing on the preprocessed time series data to obtain Fourier accelerated time series data. The Long Short-Term Memory Network (LSTM) processor is used to perform time-domain analysis on the preprocessed time-series data to obtain time-domain analyzed data, and the LSM processor supports sliding window processing.

5. The lightweight vehicle-mounted edge computing circuit module according to claim 1, characterized in that, The lightweight vehicle-mounted edge computing circuit module also includes: The dynamic power consumption management unit is electrically connected to the multi-source data interface, the preprocessing unit, the heterogeneous computing unit, and the decision unit, respectively, and is used to dynamically adjust the voltage frequency of each unit connected to the dynamic power consumption management unit according to the amount of source data to perform hierarchical energy consumption control.

6. The lightweight vehicle-mounted edge computing circuit module according to claim 1, characterized in that, The decision-making unit includes: The multimodal fusion module is used to perform weighted fusion processing on the image feature data and the temporal feature data to obtain a weighted fusion result; The lightweight classification module is used to classify the source data based on the weighted fusion result to obtain the anomaly level and confidence level of the source data.

7. A lightweight vehicle-mounted edge computing method, characterized in that, include: Establish a communication connection with the vehicle-mounted equipment to obtain various types of source data from the vehicle-mounted equipment; The source data of various types are subjected to predetermined processing to obtain pre-processed data, which includes pre-processed image data and pre-processed time series data. The predetermined processing includes denoising processing and normalization processing. The preprocessed image data is processed to obtain image feature data, and the preprocessed time series data is processed to obtain time series feature data; Decisions are made based on the image feature data and the temporal feature data.

8. The lightweight vehicle-mounted edge computing method according to claim 7, characterized in that, Processing the preprocessed image data to obtain image feature data, and processing the preprocessed temporal data to obtain temporal feature data, includes: The image feature data is obtained by processing the preprocessed image data using a lightweight convolution algorithm that includes a binarization sparsity processing algorithm; The preprocessed time series data is subjected to Fourier acceleration processing and long short-term memory network operations to perform joint time-domain and frequency-domain analysis to obtain the time series feature data.

9. The lightweight vehicle-mounted edge computing method according to claim 7, characterized in that, The method further includes: The data volume of the various types of source data is determined based on the sampling frequency, data type, and data transmission protocol of the various types of source data. The voltage frequency of each unit in the vehicle edge computing circuit module is dynamically adjusted according to the amount of source data to perform graded energy consumption control.

10. The lightweight vehicle-mounted edge computing method according to claim 9, characterized in that, Dynamically adjusting the voltage frequency of each unit in the vehicle edge computing circuit module based on the amount of source data to perform graded energy consumption control, including: When the amount of source data is greater than or equal to the preset amount of data, the voltage frequency of each unit in the vehicle edge computing circuit module is controlled to be the full voltage frequency. When the amount of source data is less than the preset amount of data, the voltage frequency of the target unit in the vehicle edge computing circuit module is controlled to be a non-full voltage frequency, and the target unit is an idle unit or a non-full load working unit.

11. The lightweight vehicle-mounted edge computing method according to claim 7, characterized in that, Making decisions based on the image feature data and the temporal feature data includes: The image feature data and the temporal feature data are weighted and fused to obtain a weighted fusion result, wherein the weight parameters in the weighted fusion process are related to the amount of data in the image feature data and the amount of data in the temporal feature data; The source data is classified based on the weighted fusion result to obtain the anomaly level and confidence level of the source data.

12. The lightweight vehicle-mounted edge computing method according to claim 11, characterized in that, Based on the weighted fusion result, the source data is classified to obtain the anomaly level and confidence level of the source data, including: A hybrid model of decision tree and support vector machine is used to classify the weighted fusion result to obtain the anomaly level and confidence level of the source data.

13. A rail train, characterized in that, include: Vehicle body; The vehicle body is equipped with a variety of in-vehicle devices and a lightweight in-vehicle edge computing circuit module as described in any one of claims 1 to 6. The various in-vehicle devices are electrically connected to the lightweight in-vehicle edge computing circuit module through a multi-source data interface. The various in-vehicle devices include an in-vehicle air conditioner, an in-vehicle display device, and an in-vehicle communication device.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the lightweight vehicle edge computing method according to any one of claims 7 to 12.

Images