Electronic board card anomaly detection and health evaluation method and device and medium

Abstract

The invention relates to an electronic board card anomaly detection and health evaluation method and device and a medium. The method comprises the steps that historical data and real-time collection data are collected; secondly, a lightweight machine learning technology of a Gaussian mixture model is utilized, time sequence characteristics and correlation of multi-modal sensing data are fused, accurate recognition of the abnormal state of the electronic board card and dynamic quantitative evaluation of the health degree are carried out, and anomalies in real-time data and the moment when the anomalies are generated are detected; and finally, an analysis detection result is transmitted to the intelligent operation and maintenance platform of the rail transit signal equipment for alarm display. Compared with the prior art, the method has the advantages that calculation efficiency and diagnosis performance are both considered, and accurate sensing and dynamic evaluation of the operation state of the electronic board card are achieved.

Description

Technical Field

[0001] This invention relates to train signal control systems, and more particularly to a method, device, and medium for anomaly detection and health assessment of electronic circuit boards based on multimodal sensor signals. Background Technology

[0002] With the rapid development of rail transit systems, the reliability and safety of their core signaling equipment are receiving increasing attention. Electronic circuit boards, as a key component of rail transit signaling equipment, undertake important functions such as control, communication, and data processing; their operational status directly affects the stability and safety of the entire system. Therefore, real-time anomaly detection and health status evaluation of electronic circuit boards have become an important technical means to ensure the high availability of rail transit systems.

[0003] Currently, most methods for anomaly detection and health assessment of electronic circuit boards employ threshold-based rules. Specifically, fixed upper and lower thresholds are set based on parameters such as maximum power, rated voltage, and maximum current specified in the circuit board design. When monitored signals such as voltage, current, or temperature exceed these thresholds, an anomaly is identified, and the circuit board's health is assessed using a deduction-based system. However, this method has significant limitations in practical applications. Firstly, due to differences in manufacturing processes, individual component variations, and aging processes, even electronic circuit boards of the same model may exhibit multiple discrete stable values ​​rather than a single fixed value under the same operating conditions, making it difficult to accurately identify anomalies using only static thresholds. Secondly, current and temperature signals are highly dependent on the circuit board's current operating tasks and load status, exhibiting a strong dynamic coupling relationship. Ignoring this time-varying nature and correlation, and relying solely on independent thresholds to determine anomalies, can easily lead to false alarms or missed alarms.

[0004] Furthermore, due to the strict constraints on cost, power consumption, and computing resources imposed by embedded applications in rail transit, electronic boards typically lack the ability to deploy highly complex algorithms such as deep learning. Existing threshold rule methods are also unable to effectively model the nonlinear correlations and temporal characteristics between multimodal sensing signals, resulting in insufficient accuracy in anomaly detection and health assessment.

[0005] A search revealed that CN114756000A discloses a method, device, vehicle, and storage medium for detecting abnormal control signals. Specifically, it discloses acquiring the system state of the computing unit and a first control signal; calculating a first control quantity difference of the control signal based on the first control signal and a second control signal from the previous moment; and sending a second control signal to the vehicle control unit when the first control quantity difference is greater than a first threshold and the system state is faulty. However, this existing patent is still limited to diagnosis based on preset thresholds and rules, lacking adaptive capabilities. Meanwhile, Chinese Patent Publication No. CN120288088A discloses a method and apparatus for detecting anomalies in track circuit signals. Specifically, it discloses the periodic acquisition of raw waveform array data of track circuit signals; obtaining time-domain index data and frequency-domain index data corresponding to each acquisition moment based on the raw waveform array data of the track circuit signals acquired at each acquisition moment; reconstructing the track circuit signals based on the time-domain index data and frequency-domain index data corresponding to multiple acquisition moments; and determining whether there is an anomaly based on the reconstructed track circuit signals. However, this existing patent utilizes time-frequency domain features such as signal-to-noise ratio, effective current value, and frequency to perform anomaly detection using rules, which lacks adaptive capability and relies on the effectiveness of feature extraction.

[0006] Therefore, there is an urgent need for an anomaly detection and health assessment method that balances computational efficiency and diagnostic performance, which can make full use of multimodal signals such as voltage, current, and temperature collected by pre-embedded sensors under limited resource conditions to achieve accurate perception and dynamic evaluation of the operating status of electronic boards. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art by providing a method, device and medium for abnormal detection and health evaluation of electronic circuit boards based on multimodal sensing signals, which balances computational efficiency and diagnostic performance, and realizes accurate perception and dynamic evaluation of the operating status of electronic circuit boards.

[0008] The objective of this invention can be achieved through the following technical solutions: According to a first aspect of the present invention, a method for detecting and evaluating the health of electronic circuit boards is provided. This method collects historical data and real-time acquired data. Secondly, it utilizes lightweight machine learning technology based on Gaussian mixture models to integrate the temporal characteristics and correlations of multimodal sensor data, accurately identifying the abnormal state of electronic circuit boards and dynamically quantifying their health status, thereby detecting anomalies in real-time data and the time when the anomalies occur. Finally, the analysis and detection results are transmitted to the intelligent operation and maintenance platform for rail transit signaling equipment for alarm display.

[0009] As a preferred technical solution, the method specifically includes the following steps: Step S1: Intelligent grouping of multi-source sensor signals to obtain a total of One effective sensing signal pair and Each unpaired independent sensor signal is represented as a set of modeling units, which combines all valid sensor signal pairs with the independent sensor signals. Step S2: Using the set of modeling units output in step S1 as input, construct the Gaussian mixture model offline. Step S3: Anomaly detection and online health level determination; Step S4: Send the data anomaly detection, health evaluation results and corresponding operation and maintenance suggestions to the intelligent operation and maintenance platform, and display the alarm.

[0010] As a preferred technical solution, step S1 specifically includes: Step S1.1: Intelligent grouping of multi-source sensor signals based on physical mechanisms and statistical relationships. Step S1.2: Intelligent grouping of multi-source sensor signals based on statistical relationships; Step S1.3: Save the modeling unit grouping rules obtained in steps S1.1 and S1.2. After filtering in steps S1.1 and S1.2, a total of [number] units are obtained. One effective sensing signal pair and For each unpaired independent sensing signal, all valid signal pairs and independent signals are represented as a set of modeling units, denoted as: Among them, for , For the first One effective two-dimensional sensing signal pair; for , For the first Each independent one-dimensional sensing signal.

[0011] As a preferred technical solution, step S1.1 specifically includes: Step S1.1.1: Acquire the data collected by the electronic board to be monitored under normal operating conditions. The original sensing timing data of the path is denoted as , of which Road signals include Each sampling point is denoted as . ; Step S1.1.2: Based on the mechanistic coupling relationship and data correlation between different signals, the signals are divided into multiple signal pairs; Step S1.1.3, for each group of candidate two-dimensional signal pairs The maximum information coefficient is calculated based on historical normal operation sampling data. ; Step S1.1.4: Set the adaptive threshold ; if At that time, candidate signal pair To effectively model two-dimensional signal pairs; if At that time, candidate signal pair To avoid forming an effective two-dimensional signal pair for modeling, and and Store the unpaired signal set, denoted as .

[0012] As a preferred technical solution, the maximum information coefficient in step S1.1.3 The specific calculations are as follows: in, and The maximum information coefficient; and They represent respectively to and The number of intervals to be discretized; and Given a number of bins and Below, by analyzing the data points in The maximum mutual information value obtained by optimally partitioning the grid; This represents the maximum discretized grid density.

[0013] As a preferred technical solution, step S1.2 specifically includes: For all items in step S1.1 that were not paired using a combination of physical mechanisms and statistical relationships Road signal set ,in Perform the following operations: Step S1.2.1, from Select Calculate in sequence and Maximum information coefficient ,in ; Step S1.2.2, if At that time, candidate signal pair To effectively model signal pairs, and and from Delete; Step S1.2.3, if At that time, candidate signal pair To avoid forming a valid modeling signal pair, For independent one-dimensional sensing signals Step S1.2.4: Repeat steps S1.2.1 to S1.2.3 until... All signals have been traversed.

[0014] As a preferred technical solution, step S2 specifically includes: The set of modeling units output in step S1 For each modeling unit, as input. Perform the following operations independently: Step S2.1: Standardize the data within the modeling unit, and denote the standardized modeling data unit as... ; Step S2.2: Construct a probability distribution model; Step S2.2, save each modeling unit. The corresponding standardized parameters and probability distribution model are used to synchronously reconstruct the distribution model of the affected modeling units when intelligent grouping update is triggered in step S1.

[0015] As a preferred technical solution, step S2.1 specifically includes: like For a two-dimensional sensing signal pair, denoted as Then calculate based on historical data. and mean , and standard deviation , and to and Each sampling point is Z-score normalized to obtain and The calculation formula is as follows: like A one-dimensional independent sensing signal, denoted as Then, calculations are based on historical data. mean and standard deviation and to Each sampling point is Z-score normalized to obtain The calculation formula is as follows: The standardized modeling data unit is denoted as .

[0016] As a preferred technical solution, step S2.2 specifically includes: The modeling data units standardized in step S2.1 For each unit as input The Gaussian mixture model is fitted using the expectation-maximization algorithm. ; For two-dimensional sensing signal pairs , A two-dimensional Gaussian mixture model is used to fit the joint distribution of its standardized data, and its probability density function is calculated as follows: For one-dimensional independent sensing signals ,in The probability density function is calculated as follows, using a univariate Gaussian distribution for modeling: in Indicates the number of Gaussian components in the mixture. Indicated by and The parameter is a multivariate Gaussian distribution. , , They represent the first The weights, mean, and covariance matrices of each Gaussian component distribution, when the input data is one-dimensional. Degenerate into Variance; parameters are estimated using the expectation-maximization algorithm. , , ; For the number of Gaussian components in the mixture The BIC is automatically determined using the Bayesian Information Criterion (BIC), and the BIC calculation formula is as follows: in For the model's log-likelihood, for a signal pair, For independent signals, , take minimum value As the optimal number of Gaussian components in the mixture.

[0017] As a preferred technical solution, step S3 specifically includes: for Real-time acquisition of multiple signal data that are constantly being monitored Perform the following steps, where : Step S3.1: Group the modeling units according to the grouping strategy stored in step S1 to obtain the set of units to be detected. For each unit The new sampled values ​​are converted into normalized form using the normalization parameters saved in step S2. ; Step S3.2.1, for the standardized unit to be tested Calculate its value up to the number saved in step S2. The shortest Mahalanobis distance of the component distributions in a Gaussian mixture model Construct the normalized Mahalanobis distance outlier score, denoted as ; Step S3.2.2, for the standardized unit to be tested Calculate the number of times it is stored in step S2. Log-likelihood on a Gaussian mixture model Construct the abnormal fractions using the normalized log-likelihood method, denoted as ; Step S3.2.3, for the standardized unit to be tested Mahalanobis distance outlier score Log-likelihood anomaly score The weighted sum is used to obtain the overall anomaly score. ; Step S3.3: Calculate the anomaly score of the unit set. ,in Using the "weakest link" principle, the maximum value of the abnormal scores for all units is taken as the electronic board card value. The global anomaly score at time step is denoted as ;Calculation electronic board in Health at all times ,in ; Step S3.4: Based on the preset threshold range, adjust the health level... Mapped to a five-level health status.

[0018] As a preferred technical solution, the shortest Mahalanobis distance in step S3.2.1 The specific calculations are as follows: The The specific calculations are as follows: in This represents the baseline value of Mahalanobis distance calculated based on historical data. Indicates the sensitivity of outlier scores using the Mahalanobis distance method. The larger the value, Follow The faster the rate of increase, the more... The larger the value, the more sensitive the unit to be detected. The higher the degree of anomaly in the Mahalanobis distance; The specific definition of the Mahalanobis distance baseline value is as follows: (1) If the historical data contains outliers, then select the quantile as the baseline value, such as the 90th, 95th, or 99th quantile. The more outliers the historical data contains, the lower the quantile. Therefore, the proportion of outliers in the historical data can be used as the quantile value. (2) If all historical data are normal, then use the historical maximum value as the baseline value.

[0019] As a preferred technical solution, the log-likelihood of step S3.2.2 The specific calculations are as follows: The The specific calculations are as follows: in, This represents the baseline log-likelihood value calculated based on historical data. This indicates the sensitivity of abnormal fractions in the log-likelihood method. The larger the value, Follow The faster the rate of decrease and increase, The larger the value, the more sensitive the unit to be detected. The higher the log-likelihood anomaly, the better; The specific definition of the log-likelihood baseline value is as follows: (1) If the historical data contains outliers, then select the quantile as the baseline, such as the 1%, 5%, or 10% quantile. The more outliers the historical data contains, the higher the quantile. Therefore, use 1 minus the proportion of outliers in the historical data as the quantile value. (2) If all historical data are normal, then use the historical minimum value as the baseline.

[0020] As a preferred technical solution, the comprehensive anomaly score in step S3.2.3 The specific calculations are as follows: in, and These represent the weights of the Mahalanobis distance outlier score and the log-likelihood outlier score, respectively.

[0021] As a preferred technical solution, in step S3.3 The specific calculations are as follows:

[0022] As a preferred technical solution, the five health statuses in step S3.4 are specifically as follows: Simultaneously, corresponding business handling suggestions are output: (1) Normal: No need to pay attention; (2) Minor abnormalities: Continuous monitoring is recommended; (3) Moderate abnormality: It is recommended to arrange for examination; (4) Highly abnormal: It is recommended to shut down the machine for troubleshooting; (5) Extreme abnormality: Stop the machine immediately and sound an alarm.

[0023] According to a second aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described thereon.

[0024] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0025] Compared with the prior art, the present invention has the following advantages: 1) This invention addresses the problem that existing rule-based methods based on fixed thresholds are difficult to adapt to individual differences in circuit boards, dynamic changes in operating status, and the characteristics of multi-signal coupling. Without relying on high computing resources, this patent uses lightweight machine learning technology based on Gaussian mixture models, which integrates the temporal characteristics and correlations of multimodal sensor data such as voltage, current, and temperature, based on historical data and real-time acquired data. This enables accurate identification of abnormal circuit board states and dynamic quantitative assessment of health status, detects anomalies in real-time data and the time when the anomalies occur, and transmits the analysis and detection results to the intelligent operation and maintenance platform for rail transit signaling equipment for alarm display, facilitating timely handling by maintenance personnel, thereby improving the reliability and intelligent operation and maintenance level of rail transit signaling systems. 2) This invention is based on intelligent coupling identification and adaptive variable grouping of sensor signals through the fusion of mechanism and data. By integrating prior physical mechanisms with statistical information (maximum information coefficient) for data-driven verification, this invention achieves adaptive intelligent grouping of sensor signals: it can accurately identify strongly coupled variable pairs such as voltage-current and current-temperature to support joint distribution modeling, while retaining unpaired independent signals for single-variable monitoring; the grouping threshold is dynamically set based on historical normal data and supports online updates, effectively adapting to multiple operating conditions and equipment aging scenarios, fundamentally avoiding the sparsity and irrelevant interference problems of high-dimensional modeling, and laying a structural foundation for subsequent high-sensitivity and interpretable anomaly detection; 3) This invention integrates multiple anomaly measurement methods to improve detection robustness, employing both Mahalanobis distance and log-likelihood methods for anomaly assessment. Mahalanobis distance evaluates the standardized distance between the test point and each Gaussian component of the Gaussian mixture model, effectively considering the correlation and scale differences between features; while log-likelihood captures the sparseness of data in the probability density space. These two methods complement each other, retaining the advantages of each while facilitating adjustments based on actual business needs, significantly enhancing the model's adaptability and detection accuracy under complex conditions. 4) This invention adaptively sets thresholds based on historical data, reducing manual intervention. It uses historical data to calculate the Mahalanobis distance and probability density quantiles as dynamic reference thresholds, and normalizes the score of the degree of anomaly (mapped to the [0,1] interval) accordingly. This data-driven threshold setting method avoids the failure problem of fixed thresholds in different operating environments or equipment states, and improves the generalization ability and deployment flexibility of the model. 5) The business interpretability classification and suggestions of this invention divide the final anomaly score into five clear levels: "normal", "minor anomaly", "moderate anomaly", "high anomaly" and "extreme anomaly", and provide corresponding business suggestions for each level (such as "no need to pay attention", "monitoring recommended", "immediate handling", etc.). This direct mapping from algorithm output to operation and maintenance actions greatly improves the interpretability and practicality of the results, and facilitates rapid response by front-line personnel. 6) This invention is optimized for electronic board sensor signal scenarios, taking into account both efficiency and accuracy. In a low-dimensional space considering voltage, current and temperature, the Mahalanobis distance and log-likelihood calculation based on Gaussian mixture model are efficient and stable, making it very suitable for deployment in edge computing devices or real-time monitoring systems, thus balancing algorithm accuracy and engineering feasibility. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the electronic circuit board anomaly detection and health evaluation method of the present invention. Figure 2 This is a flowchart illustrating the intelligent grouping of multi-source sensor signals according to the present invention. Figure 3 This is a flowchart illustrating the offline construction of the Gaussian mixture model of this invention. Figure 4 This is a flowchart illustrating the specific process of anomaly detection and online health level determination in this invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0028] The technical solution of this invention is developed based on the company's self-developed intelligent operation and maintenance platform for rail transit signaling equipment. This invention employs a highly efficient and low-cost multimodal sensor signal anomaly detection and health assessment method suitable for electronic boards in rail transit signaling equipment. Addressing the limitations of existing rule-based methods based on fixed thresholds in adapting to individual board differences, dynamic changes in operating status, and multi-signal coupling characteristics, this patent, without relying on high computing resources, utilizes lightweight machine learning technology based on historical and real-time acquired data, combining the temporal characteristics and correlations of multimodal sensor data such as voltage, current, and temperature to achieve accurate identification of board anomalies and dynamic quantitative assessment of health. It detects anomalies in real-time data and the time of their occurrence, and transmits the analysis and detection results to the intelligent operation and maintenance platform for rail transit signaling equipment for alarm display, facilitating timely handling by maintenance personnel, thereby improving the reliability and intelligent operation and maintenance level of the rail transit signaling system.

[0029] like Figure 1 As shown, the technical solution of the present invention specifically includes the following steps: Step S1) Intelligent grouping of multi-source sensor signals, specifically: like Figure 2 As shown, step S1.1) is intelligent grouping of multi-source sensor signals based on physical mechanisms and statistical relationships: Acquire data from the electronic board to be monitored under normal operating conditions. The original sensing timing data of the path is denoted as , of which Road signals include Each sampling point is denoted as . .

[0030] Based on the mechanistic coupling relationships and data correlations between different signals, the signals are divided into multiple signal pairs. First, candidate signal pairs are constructed from signals with strong coupling relationships (existing in a clear physical interaction mechanism), denoted as [list of pairs]. .

[0031] For each pair of candidate two-dimensional signals The maximum information coefficient is calculated based on historical normal operation sampling data to statistically determine the dependency strength. The calculation formula is as follows: in, and The maximum information coefficient; and They represent respectively to and The number of intervals to be discretized (binned); and Given a number of bins and Below, by analyzing the data points in The maximum mutual information value obtained by optimally partitioning the grid; This represents the maximum discretization grid density, used to control the upper limit of grid complexity. This is a preset constant (usually taken as...). ).

[0032] Set adaptive threshold For all valid operating conditions in historical normal data 10th percentile of the value; if At that time, candidate signal pair To effectively model two-dimensional signal pairs.

[0033] if At that time, candidate signal pair To avoid forming an effective two-dimensional signal pair for modeling, and and Store the unpaired signal set, denoted as .

[0034] Optionally, candidate signal pairs with strong coupling relationships may include “voltage and current”, “current and temperature”, and “voltage and temperature”.

[0035] Step S1.2) Intelligent grouping of multi-source sensor signals based on statistical relationships: For all items in step S1.1 that were not paired using a combination of physical mechanisms and statistical relationships Road signal set Perform the following operations: (1) From Select Calculate in sequence and Maximum information coefficient ; (2) If At that time, candidate signal pair To effectively model signal pairs, and and from Delete; (3) If At that time, candidate signal pair To avoid forming a valid modeling signal pair, It is an independent one-dimensional sensing signal; (4) Repeat (1) to (3) until All signals have been traversed.

[0036] After filtering in steps S1.1 and S1.2, a total of [number] results were obtained. One effective sensing signal pair and There are 10 unpaired independent sensing signals. All valid signal pairs and independent signals are represented as a set of modeling units, denoted as: Among them, for , For the first One effective two-dimensional sensing signal pair; for , For the first Each independent one-dimensional sensing signal.

[0037] Step S1.3) Save the modeling unit grouping rules obtained in steps S1.1 and S1.2.

[0038] Step S2) Offline construction of Gaussian mixture model The set of modeling units output in step S1 For each modeling unit, as input. Perform the following operations independently, such as Figure 3 As shown: Step S2.1) Data standardization within the modeling unit like For a two-dimensional sensing signal pair, denoted as Then calculate based on historical data. and mean , and standard deviation , and to and Each sampling point is Z-score normalized to obtain and The calculation formula is as follows: like A one-dimensional independent sensing signal, denoted as Then, calculations are based on historical data. mean and standard deviation and to Each sampling point is Z-score normalized to obtain The calculation formula is as follows: The standardized parameters are calculated based solely on the historical data of the modeling unit itself and are not shared with other units.

[0039] The standardized modeling data unit is denoted as .

[0040] Step S2.2) Constructing the probability distribution model The modeling data units standardized in step S2.1 For each unit as input The Gaussian mixture model was fitted using the expectation-maximization (EM) algorithm. .

[0041] For two-dimensional sensing signal pairs A two-dimensional Gaussian mixture model is used to fit the joint distribution of its standardized data, and its probability density function is calculated as follows: For one-dimensional independent sensing signals The probability density function is calculated as follows, using a univariate Gaussian distribution for modeling: in Indicates the number of Gaussian components in the mixture. Indicated by and The parameter is a multivariate Gaussian distribution. , , They represent the first The weights, mean, and covariance matrices of each Gaussian component distribution, when the input data is one-dimensional. Degenerate into That is, variance. Parameters are estimated using the Expectation-Maximization (EM) algorithm. , , .

[0042] For the number of Gaussian components in the mixture The BIC is automatically determined using the Bayesian Information Criterion (BIC). The BIC calculation formula is as follows: in For the model's log-likelihood, for a signal pair, For independent signals, Take or use minimum value As the optimal number of Gaussian components in the mixture.

[0043] Step S2.3) Model storage and update Save each modeling unit The corresponding standardized parameters and probability distribution model.

[0044] When step S1 triggers intelligent grouping update, the distribution model of the affected modeling unit is reconstructed synchronously to ensure system consistency and realize online evolution of the model. Finally, a probability distribution model matching its dimension is established for each modeling unit, providing a basis for subsequent anomaly scoring.

[0045] Step S3) Anomaly Detection and Online Health Level Determination like Figure 4 As shown, for Real-time acquisition of multiple signal data that are constantly being monitored Perform the following sub-steps: Step S3.1) Grouping and standardizing the signal data to be detected The units are grouped according to the modeling unit grouping strategy stored in step S1 to obtain the set of units to be detected. For each unit The new sampled values ​​are converted into normalized form using the normalization parameters saved in step S2. ; Step S3.2) Calculation of unit-level anomaly scores Step S3.2.1) Calculation of Mahalanobis distance anomaly score For the standardized test unit Calculate the number of times it is stored in S2. The shortest Mahalanobis distance of the component distributions in a Gaussian mixture model The calculation formula is as follows: Construct the normalized Mahalanobis distance outlier score, denoted as The calculation formula is as follows: in, This represents the baseline value of Mahalanobis distance calculated based on historical data. Indicates the sensitivity of outlier scores using the Mahalanobis distance method. The larger the value, Follow The larger it increases, the faster the rate of increase. (Optional) You can take the 95th percentile, 99th percentile, or maximum value of the Mahalanobis distance from historical data. The larger the value, the more sensitive the unit to be detected. The higher the degree of anomaly in the Mahalanobis distance, the better.

[0046] Step S3.2.2) Calculation of log-likelihood anomaly scores For the standardized test unit Calculate the number of times it is stored in S2. Log-likelihood on a Gaussian mixture model The calculation formula is as follows: Construct the abnormal fractions of the normalized log-likelihood method, denoted as The calculation formula is as follows: in, This represents the baseline log-likelihood value calculated based on historical data. This indicates the sensitivity of abnormal fractions in the log-likelihood method. The larger the value, Follow The faster the rate of increase, the more likely it is to decrease. (Optional) The 5th percentile, 1st percentile, or minimum value of the log-likelihood of historical data can be used. The larger the value, the more sensitive the unit to be detected. The higher the log-likelihood anomaly, the better.

[0047] Step S3.2.3) Calculation of Comprehensive Abnormal Score For the standardized test unit Mahalanobis distance outlier score Log-likelihood anomaly score The weighted sum is used to obtain the overall anomaly score. The calculation formula is as follows: in, and These represent the weights of the Mahalanobis distance outlier score and the log-likelihood outlier score, respectively.

[0048] Optional, , Weighting Explanation: Mahalanobis distance directly measures the "geometric" distance between a point and the center of the distribution, and is generally more sensitive to anomalies; therefore, a higher weight is assigned to the distance score. The log-likelihood score serves as an auxiliary measure, providing supplementary information at the probabilistic level, and therefore has a lower weight.

[0049] Step S3.3) Global-level anomaly score fusion and health calculation The unit set aggregated anomaly score calculated in step S3.2 Using the "weakest link" principle, the maximum value of the abnormal scores for all units is taken as the electronic board card value. The global anomaly score at time step is denoted as The calculation formula is as follows: Electronic board card in Health at all times The calculation formula is as follows: Health The larger it is, the more the electronic board is stuck in The higher the health level at any given time.

[0050] Step S3.4) Health Status Level Determination and Business Suggestion Generation Based on a preset threshold range, the health level is... Mapped to five health levels and recommendations provided: Simultaneously, corresponding business handling suggestions are output: (1) Normal: No need to pay attention; (2) Minor abnormalities: Continuous monitoring is recommended; (3) Moderate abnormality: It is recommended to arrange for examination; (4) Highly abnormal: It is recommended to shut down the machine for troubleshooting; (5) Extreme abnormality: Stop the machine immediately and sound an alarm.

[0051] Step S4) Send the data anomaly detection, health evaluation results and corresponding operation and maintenance suggestions to the intelligent operation and maintenance platform, and display the alarm.

[0052] The above is an introduction to the method embodiments. The following embodiments using electronic devices and storage media will further illustrate the solution of the present invention.

[0053] This invention also provides an electronic device including a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0054] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0055] The processing unit executes the various methods and processes described above, such as methods S1 to S4. For example, in some embodiments, methods S1 to S4 may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of methods S1 to S4 described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute methods S1 to S4 by any other suitable means (e.g., by means of firmware).

[0056] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0057] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0058] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0059] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An electronic board card abnormality detection and health evaluation method, characterized in that, The method collects historical data and real-time collected data; secondly, a lightweight machine learning technology of Gaussian mixture model is used to fuse the time sequence characteristics and correlation of multi-modal sensor data, accurately identify the abnormal state of the electronic board card, dynamically quantify the health degree, detect the abnormality in the real-time data and the time when the abnormality occurs; finally, the analysis and detection results are transmitted to the intelligent operation and maintenance platform of the rail transit signal equipment for alarm display.

2. The electronic board card abnormality detection and health evaluation method of claim 1, wherein, The method specifically comprises the following steps: Step S1, intelligent grouping of multi-source sensing signals, a total of effective sensing signal pairs and unpaired independent sensing signals, all effective sensing signal pairs and independent sensing signals are uniformly represented as a set of modeling units; Step S2, taking the modeling unit set output in step S1 as input, offline constructing a Gaussian mixture model; Step S3, online determining abnormality detection and health level; Step S4, sending data abnormality detection, health evaluation results and corresponding operation and maintenance suggestions to the intelligent operation and maintenance platform, and performing alarm display.

3. The electronic board card abnormality detection and health evaluation method of claim 2, wherein, The step S1 specifically comprises: Step S1.1, intelligent grouping of multi-source sensor signals based on physical mechanism and statistical relationship, Step S1.2, intelligent grouping of multi-source sensor signals based on statistical relationship; Step S1.3, save the modeling unit grouping rules obtained from steps S1.1 and S1.

2. After screening through steps S1.1 and S1.2, a total of effective signal pairs and unpaired independent signals are obtained. All effective signal pairs and independent signals are uniformly expressed as a modeling unit set, denoted as: wherein, for each of the first and second effective two-dimensional sensing signal pairs , is a first effective two-dimensional sensing signal pair; for each of the first and second effective two-dimensional sensing signal pairs , , is a first effective two-dimensional sensing signal pair; for each of the first and second effective two-dimensional sensing signal pairs , is a first independent one-dimensional sensing signal.

4. The electronic board card abnormality detection and health evaluation method of claim 3, wherein, The step S1.1 specifically comprises: Step S1.1.1, obtaining original sensing time series data of the electronic board card to be monitored under normal operation state , denoted as , wherein the first channel signal contains sampling points, denoted as ; Step S1.1.2, based on the mechanism coupling relationship and data correlation between different signals, the signals are divided into a plurality of signal pairs; Step S1.1.3, for each set of candidate two-dimensional signal pairs , calculate the maximum information coefficient based on historical normal operation sampling data ; Step S1.1.4, setting adaptive threshold ; If the candidate signal pair is effective in modeling the two-dimensional signal pair; If , the candidate signal pair is not a valid modeling two-dimensional signal pair, and the pair is stored in the unpaired signal set, denoted as . .

5. The electronic board card abnormality detection and health evaluation method of claim 4, wherein, The maximum information coefficient in the step S1.1.3 The specific calculation is as follows: wherein, with the maximum information coefficient; and denote the number of intervals for discretizing and respectively; with the maximum mutual information value obtained by optimally partitioning the data points on a grid, given the number of bins and ; and denotes the maximum discretization grid density.

6. The electronic board card abnormality detection and health evaluation method of claim 4, wherein, The step S1.2 specifically comprises: For all the pairs in step S1.1 that did not pass the joint pairing of physical mechanisms and statistical relationships Road signal set wherein the following operations are performed: Step S1.2.1, selecting from , sequentially calculating and the maximum information coefficient , wherein ;​ Step S1.2.2, if the candidate signal pair is valid, then the candidate signal pair is added to the set of valid signal pairs and the set of invalid signal pairs is updated by removing Step S1.2.3, if the candidate signal pair is not a valid modeling signal pair, is an independent one-dimensional sensing signal Step S1.2.4, repeat steps S1.2.1 to S1.2.3 until all signals in the middle are traversed.

7. The method of claim 3, wherein the method further comprises: Step S2 specifically comprises: The set of modeling units output in step S1 For each modeling unit Independently perform the following operations: Step S2.1, standardizing data within the modeling unit, the standardized modeling data unit is denoted as ; Step S2.2, constructing a probability distribution model; Step S2.2, save each modeling unit Corresponding standardized parameters and probability distribution models, when the intelligent grouping update is triggered in step S1, the distribution model of the affected modeling unit is reconstructed synchronously.

8. The electronic board card abnormality detection and health evaluation method of claim 7, wherein, The step S2.1 specifically comprises: If is a two-dimensional sensing signal pair, denoted as , then the mean and standard deviation of , and , are calculated based on historical data, respectively, and each sampling point of and is Z-score standardized to obtain and , and the calculation formula is as follows: like A one-dimensional independent sensing signal, denoted as Then, calculations are based on historical data. mean and standard deviation and to Each sampling point is Z-score normalized to obtain The calculation formula is as follows: The standardized modeling data unit is denoted as .

9. The method of claim 7, wherein the method further comprises: The step S2.2 specifically comprises: with the modeling data units standardized in step S2.1 For the input, for each cell a Gaussian mixture model is fitted using the expectation-maximization algorithm ; For two-dimensional sensing signal pairs , , the joint distribution of their normalized data is fitted with a two-dimensional Gaussian mixture model, whose probability density function is computed as follows: For one-dimensional independent sensor signals where A univariate Gaussian distribution is adopted to model the probability density function, which is calculated as follows: wherein denotes the number of Gaussian components, denotes the mean of the and multivariate Gaussian distribution with parameters , , denote the weight, mean and covariance matrix of the th Gaussian component distribution, respectively, when the input data is one-dimensional, reduces to , i.e. the variance; the parameters , , are estimated by the expectation-maximization algorithm; For the number of Gaussian components is automatically determined by the Bayesian information criterion (BIC), which is calculated as follows: where is the model log-likelihood for a signal pair, for an independent signal, , take the value of that minimizes as the optimal number of Gaussian components.

10. The method of claim 3, wherein the method further comprises: Step S3 specifically comprises: For Real-time acquisition multi-channel signal data to be detected at the moment The following steps are performed, wherein : Step S3.1, grouping according to the modeling unit grouping policy stored in step S1, obtaining a set of units to be detected ; for each unit , converting its corresponding newly sampled value into normalized form using the normalization parameters saved in step S2 ; Step S3.2.

1. For the normalized cell to be detected , compute the shortest Mahalanobis distance to each component distribution of the Gaussian Mixture Model saved in step S2 , construct the normalized Mahalanobis distance anomaly score, denoted as ; Step S3.2.

2. For the normalized cell to be detected , compute its log-likelihood on the Gaussian mixture model it saved in step S2 , construct the normalized log-likelihood method anomaly score, denoted as ;​ Step S3.2.3, for the normalized to-be-detected unit the Mahalanobis distance anomaly score the log-likelihood anomaly score weighted sum to obtain a comprehensive anomaly score ; Step S3.3, integrate the calculated unit set comprehensive anomaly score , wherein , the maximum value of all unit anomaly scores is taken as the global level anomaly score of the electronic board at moment, denoted as ; the health degree of the electronic board at moment is calculated , wherein ; Step S3.4: Based on the preset threshold range, adjust the health level... Mapped to a five-level health status.

11. The method of claim 10, wherein the method further comprises: The shortest Mahalanobis distance in said step S3.2.1 The specific calculation is as follows: The The specific calculation is as follows: in This represents the baseline value of Mahalanobis distance calculated based on historical data. Indicates the sensitivity of outlier scores using the Mahalanobis distance method. The larger the value, Follow The faster the rate of increase, the more... The larger the value, the more sensitive the unit to be detected. The higher the degree of anomaly in the Mahalanobis distance; Wherein the Mahalanobis distance baseline value is specifically defined as follows: (1) If the historical data contains abnormal data, select the quantile as the baseline value, the more abnormal data contained in the historical data, the lower the quantile, therefore, the proportion of abnormal data in the historical data is used as the quantile value; (2) If the historical data is all normal data, use the historical maximum value as the baseline value.

12. The method of claim 10, wherein the method further comprises: The log-likelihood of step S3.2.2 The specific calculation is as follows: The The specific calculation is as follows: in, This represents the baseline log-likelihood value calculated based on historical data. This indicates the sensitivity of outlier fractions in the log-likelihood method. The larger the value, Follow The faster the rate of decrease and increase, The larger the value, the more sensitive the unit to be detected. The higher the log-likelihood anomaly, the better; Wherein the log-likelihood baseline value is specifically defined as follows: (1) If the historical data contains abnormal data, select the quantile as the baseline, such as 1%, 5%, or 10% quantile, the more abnormal data contained in the historical data, the higher the quantile, therefore, the proportion of abnormal data in the historical data is used as the quantile value; (2) If the historical data is all normal data, use the historical minimum value as the baseline.

13. The method of claim 10, wherein the method further comprises: The integrated abnormality score of step S3.2.3 The specific calculation is as follows: wherein, with respectively denote the weights of the Mahalanobis distance anomaly score and the log-likelihood anomaly score.

14. The method of claim 10, wherein the method further comprises: The step S3.3 The specific calculation is as follows: 。 15. The method of claim 10, wherein the method further comprises: The five-level health state in step S3.4 is specifically: Meanwhile, output the corresponding business disposal suggestion: (1) Normal: no need to pay attention; (2) Slight abnormality: suggest continuous monitoring; (3) Moderate abnormality: suggest arranging inspection; (4) High abnormality: suggest stopping and troubleshooting; (5) Extreme abnormality: immediately stop and alarm.

16. An electronic device comprising a memory and a processor, said memory having stored thereon a computer program, characterized in that, The processor executes the program to realize the method of any one of claims 1-15.

17. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to realize the method of any one of claims 1-15.

Images